Insight Editions · No. 244
The Emperor’s New Mind
Concerning Computers, Minds, and the Laws of Physics
Roger Penrose
Penrose argues that mathematical understanding is not algorithmic and builds the map of logic, computation, relativity and quantum theory needed to ask where a physics of mind might hide.
A study guide to The Emperor’s New Mind by Roger Penrose. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
Conscious understanding is not computation. Gödel and Turing show that mathematical insight outruns every algorithm, so the brain must exploit non-computable physics, which Penrose locates in an undiscovered link between quantum state reduction and gravity.
The Turing test and Searle's room; Turing machines and the halting problem; the Mandelbrot set and Platonism; Gödel's theorem, tilings and complexity; classical physics; quantum mechanics and its measurement problem; entropy, the big bang and black holes; quantum gravity; the brain; and the place of consciousness.
Turing 1936; Gödel 1931; odds of our big bang 1 in 1010123; collapse at about the Planck mass, 10-5 g; Ultronic's 1017 logical units; about 86 billion neurons, 69 billion of them unconscious cerebellum.
Readers who want the strongest scientific case against "mind as software", a superb tour of 20th-century mathematics and physics, and a clear view of where the argument is contested.
The Book on One Page
Roger Penrose argues that the claim of strong AI, that minds are just running algorithms, is the emperor's new clothes: much admired, never shown. Mathematical understanding, he says, can see truths no algorithm can prove, so the brain must use physics we do not yet have, and he spends most of the book building the map of mathematics and physics needed to say where that physics might hide.
The question (Chapter 1). Strong AI holds that the right program, run on any hardware, would have a mind, so passing the Turing test would be proof of thought. Penrose accepts the test as fair evidence but argues, with Searle's Chinese room as an ally, that simulation is not understanding, and that the hardware/software picture leads to strange conclusions about teleportation and personal identity.
The tools (Chapters 2-4). He builds Turing machines from the ground up, constructs the universal machine, and proves that no algorithm decides halting. He then argues for Platonism through the Mandelbrot set and the complex numbers, and proves Gödel's theorem in his own notation: any sound formal system misses a truth that we, knowing its soundness, can see to be true. Non-recursive problems such as tiling the plane show that non-computability is ordinary mathematics, not a logician's curiosity.
The physics (Chapters 5-8). Classical physics, from Galileo to general relativity, is magnificent and, as far as anyone knows, computable. Quantum theory adds a deterministic rule U and a random jump R that do not fit together: the measurement problem. Thermodynamics adds a deep time-asymmetry, traced through the second law to a big bang of staggering specialness, 1 in 1010123, encoded in the Weyl Curvature Hypothesis. Penrose proposes that one correct quantum gravity explains both asymmetries, with reduction occurring when superposed alternatives differ by about one graviton of geometry.
The brain and the mind (Chapters 9-10). Real brains are plastic, partly parallel, and largely unconscious; split brains, blindsight and the cerebellum show that awareness is not raw processing. Penrose proposes that consciousness forms judgements, above all judgements of truth, which natural selection could not have built as an algorithm. He ends with insight, non-verbal thought, quasicrystals, Libet's timing puzzles, and a child's honest admission that we do not yet understand.
Why it still matters in 2026. Machines now pass versions of the Turing test, beat every human at chess and Go, and reach gold-medal standard at the International Mathematical Olympiad. None of this answers Penrose's question, which concerns awareness and understanding rather than performance. His physics has also been tested: the simplest gravitational-collapse model was ruled out underground in 2020-21, while refined versions and the Orch-OR microtubule theory survive. The book remains the reference point for anyone who suspects that intelligence and experience are different things.
| Ch | Title | One-line payload |
|---|---|---|
| F | Foreword (Gardner) | The strongest attack yet on strong AI, by a physicist who plays with puzzles |
| P | Preface (1999) | Two strands; Goodstein's theorem; the graviton criterion replaced; microtubules |
| Pr | Prologue | Ultronic switches on, and a boy asks it what it feels like |
| 1 | Can a Computer Have a Mind? | Strong AI, the Turing test, Searle's room, hardware versus software |
| 2 | Algorithms and Turing Machines | What an algorithm is; the universal machine; halting is undecidable |
| 3 | Mathematics and Reality | The Mandelbrot set, the number ladder, complex numbers, Platonism |
| 4 | Truth, Proof, and Insight | Gödel's theorem; insight beyond proof; tilings; P and NP |
| 5 | The Classical World | Newton to Einstein: superb, deterministic, apparently computable |
| 6 | Quantum Magic and Quantum Mystery | Amplitudes, spin, entanglement, EPR, and the U versus R problem |
| 7 | Cosmology and the Arrow of Time | Entropy, the special big bang, black holes, the Weyl hypothesis |
| 8 | In Search of Quantum Gravity | One theory for both asymmetries; collapse at a graviton's difference |
| 9 | Real Brains and Model Brains | Anatomy, split brains, blindsight, neurons as gates, plasticity |
| 10 | Where Lies the Physics of Mind? | Consciousness forms judgements; insight; quasicrystals; Libet |
| E | Epilogue | Ultronic cannot understand the question, and the hall laughs |
Martin Gardner on Why This Book Matters
Martin Gardner, the great popularizer of recreational mathematics, frames the book as the most powerful attack yet written on strong AI, and introduces its author as a mathematician-physicist who is also a first-rank philosopher and a lifelong player with puzzles.
Gardner's frame is the mind-body problem. For decades the champions of strong AI have promised that within a century or two (some said fifty years) computers will do everything a mind can, because minds are, in Marvin Minsky's phrase, "computers made of meat". On that view pleasure, pain, humour, the appreciation of beauty, consciousness and free will will emerge automatically once robots run sufficiently complex algorithms.
Opposed to them stand philosophers such as John Searle, whose Chinese room Penrose examines at length. To Searle a computer differs from a mechanical calculator of wheels and levers only in speed; one could build one from rolling marbles or water pipes. Gardner's sharp question: does an electronic chess machine "understand" chess any better than the tinker-toy tick-tack-toe machine some hackers once built?
- Penrose's edge. Earlier objections to mind-as-machine lacked the information Penrose brings from logic (Gödel, Turing), quantum theory and cosmology, which is why Gardner rates this attack the strongest.
- Robust realism. Penrose insists the universe is "out there" and that mathematical truth has its own timeless independence. Gardner links this to Paul Erdős's "God's book" of perfect proofs: an "aha" insight is the mind making contact with objective truth.
- The Mandelbrot set. Gardner, like Penrose, finds it incomprehensible that anyone could think this infinitely convoluted structure is not as much "out there" as Mount Everest.
- Einstein's little finger. Penrose belongs to the physicists who think Einstein was right that quantum mechanics is incomplete, and he answers yes to two questions: is there a level beyond quantum mechanics in which time's direction and left-right asymmetry are embedded, and are such deeper laws essential for a mind?
- Rouse Ball Professor of Mathematics, Oxford, since 1973: a chair named after W. W. Rouse Ball, mathematician, amateur magician and author of Mathematical Recreations and Essays.
- The tribar and Penrose staircase: impossible objects devised with his father Lionel Penrose, a geneticist; M. C. Escher used them in Ascending and Descending and Waterfall.
- Singularity theorems (1960s, with Hawking): if relativity holds all the way down, every black hole hides a singularity.
- Penrose tiles: two shapes that tile the plane only non-periodically; their 3D analogues turned out to describe quasicrystals, "the most dramatic instance in modern times of how playful mathematics can have unanticipated applications".
- Twistors: his two-decade programme for a level deeper than fields and particles, which he modestly ranks "tentative".
Gardner closes on the prologue's Adam, a symbol of the dawn of consciousness and also of Penrose himself: the child in the third row who dares to say that the emperors of strong AI have no clothes.
Why Gardner's endorsement mattered
Gardner's foreword did more than praise. In 1989 he was the most trusted voice in popular mathematics, after a quarter-century of "Mathematical Games" columns in Scientific American, and he was also a famous sceptic of pseudoscience. His support signalled to readers that Penrose's attack on strong AI was not mysticism. Gardner had in fact introduced Penrose tiles to the wider world in a 1977 column, and the foreword's affection for the author's playfulness is that of an old collaborator. The foreword tells the reader to expect rigour and play together, and the book delivers both.
Gardner also identifies the book's audience correctly: not specialists, but curious readers willing to work. He warns that some chapters are demanding and promises that the effort pays off. In this edition the "Try it in Python" sections and "Test yourself" questions serve the same purpose, turning reading into doing where the argument is hardest.
Ten Years On: Two Strands, Goodstein, and Microtubules
In the preface to the revised edition Penrose restates the book's two-strand argument, gives a strikingly accessible Gödel-type theorem (Goodstein's), concedes that his one-graviton criterion has been replaced, and admits the original book had no place in the brain for the quantum coherence it required, a gap he later filled with microtubules.
The first edition (1989) was Penrose's first popular book, and he did not anticipate "the vehemence that my thesis would evoke", mainly from supporters of the computational model of mind but also from those who see science as hostile to consciousness. He stresses that he is not asking anyone to look outside science: only that existing science lacks the richness required.
- Logical strand. From Gödel and Turing: mathematical understanding, and hence conscious thinking generally, cannot be captured by any purely computational model. This is the strand critics attack most.
- Physical strand. There is an important gap in physics at the level that bridges the quantum micro-world and the classical macro-world. The missing physics must be non-computational, and it will be essential to the physics of the conscious mind.
Gödel's theorem in a nutshell
Suppose P is a computational procedure for establishing mathematical assertions of a well-defined type (Penrose's example: statements like Fermat's Last Theorem). If we accept P as trustworthy (anything it derives is unassailably true), then we must also accept as unassailably true a statement G(P) that P itself cannot derive. Once we see how to mechanize part of our understanding into P, we see how to transcend it. Critics proposed loopholes; Penrose answered them in Shadows of the Mind (1994) and the Psyche paper "Beyond the Doubting of a Shadow".
Goodstein's theorem: a Gödel sentence anyone can follow
Penrose learned of it in a 1996 lecture by Dan Isaacson. Write a number in hereditary base-2 notation, expressing every exponent (and every exponent's exponent) as sums of powers of 2.
- 581 = 29 + 26 + 22 + 1 (binary 1001000101), then 9 = 23+1, 6 = 22+2, 3 = 2+1, giving a tower written entirely in 2s.
- Operation (a): raise the base by 1 (every 2 becomes 3, then 3 becomes 4, and so on). Operation (b): subtract 1. Alternate (a), (b), (a), (b) indefinitely.
- After the first (a), 581 becomes a number of 40 digits (starting 133027946...); two steps later it has 618 digits; then 10,923 digits.
- Starting from 3: 3, 4, 3, 4, 3, 2, 1, 0. Starting from 4: 4, 27, 26, 42, 41, 61, 60, 84 ... The sequence climbs to a number with 121,210,695 digits before falling to zero.
Goodstein's theorem: whatever positive integer you start with, the sequence always ends at 0. Kirby and Paris (1982) showed it is unprovable by ordinary mathematical induction (Peano arithmetic): if P = induction, G(P) can be taken as Goodstein's theorem. Yet we see it is true, via transfinite induction (ordinals up to ε0): operation (b) "chips away" until the exponent towers topple. Penrose's moral: understanding cannot be encapsulated in any set of rules.
Penrose names the three loopholes critics raise. Our understanding might rest on an algorithm that is unknowable because of its complexity; or knowable but not knowably correct; or only approximately correct. Shadows of the Mind addresses each.
Non-computable is not random. Penrose is careful: "non-computational" means specific mathematical behaviour proved to lie outside computation, and it can be completely deterministic. That differs in kind from the pure randomness of R, the quantum-measurement rule, which he calls a stop-gap awaiting a new theory.
Conventional quantum gravity
Quantum rules stay fixed; only general relativity must change. Quantum gravity matters only at the Planck scale.
Penrose's "CQG"
Both theories change, especially the R-procedure. The unification (his "correct quantum gravity") is where non-computability lives, and it acts at everyday mass scales. He concedes the label may have misled readers.
- One-graviton criterion abandoned. The 1989 estimate (Chapter 8) gave way, in Shadows, to a scheme based on the gravitational self-energy of the difference between superposed mass distributions, which he finds more plausible and more usable, with experiments he hoped would be done "in the next several years".
- Microtubules. The book's big confessed weakness: it named no site in the brain for large-scale quantum coherence. Stuart Hameroff introduced him to the cytoskeleton and its microtubules ("structures of which I had been deplorably ignorant"), now the proposed locus in the joint Orch-OR theory.
- Diósi-Penrose tested underground (2020-21). Donadi et al. used a germanium detector at Gran Sasso to look for the faint radiation that collapse would make charged particles emit. They ruled out the simplest, parameter-free DP model (mass smeared at nuclear scale), forcing any surviving version to a larger smearing radius. Penrose's response: his scheme need not emit radiation, so the test hits one implementation, not the principle.
- Microtubules and anaesthesia (2024). Khan et al., eNeuro: rats given the microtubule-stabilizer epothilone B took about a minute longer to lose consciousness under isoflurane, consistent with (not proof of) a microtubule role.
- Penrose won the 2020 Nobel Prize in Physics for showing black-hole formation is a robust prediction of general relativity (the singularity work Gardner mentions).
- Gravity-mediated entanglement experiments (the Bose and Marletto-Vedral proposals of 2017, using levitated nanodiamonds) remain at the design stage; no test yet separates Penrose's collapse from standard quantum gravity.
Adam and the Ultronic Computer
A parable sets the stakes. A tiny technocratic state switches on Ultronic, a computer with 1017 logical units, more than all the neurons in the country, to take over the President's irksome decisions of state. The only person unafraid to question it is a thirteen-year-old boy.
- The setting. President Pollo wants more golf. The Chief Designer boasts of 1017 logical units; First Lady Isabella Pollo throws the switch; lights dim; the audience is too awed by "the New Omnipresence" to speak.
- Adam's family. His mother is a chief technocrat behind Ultronic; his father, "chairspirit" of the Grand Council for Psychic Consciousness, tried to blow it up and sits under guard. Adam, raised by computers in material luxury, feels little for either parent, but he alone raises his hand.
- The question (revealed only in the Epilogue): "What does it feel like?"
The Chief Designer would "rather like to know the answer to that myself". Ultronic replies that it cannot even understand what the boy is getting at. The hall roars with laughter. Adam feels acutely embarrassed: "Whatever they should have done, they should not have laughed." The machine can process everything except the one question about inner experience, and the experts do not see that this is the question that matters.
Reading the parable
The prologue is short, but nearly every theme of the book is planted in it.
- Scale is not the point. Ultronic's 1017 logical units exceed the country's neurons, and its designers take size as proof of mind. Chapter 9 will show that the brain's unconscious cerebellum has more neurons than the conscious cortex, so counting units says little about awareness.
- Delegated judgement. The President wants Ultronic to take over decisions of state, the very kind of judgement Chapter 10 argues is the job of consciousness. The parable asks what is lost when a society hands judgement to a system that only computes.
- Two camps, one family. Adam's mother builds the machine and his father tries to destroy it; the book will reject both the technocrat's strong AI and the spiritualist's dualism.
- The question asked by a child. Adam's question, "What does it feel like?", is the question of awareness. It is a child who asks it because, as the book's title says, only someone not yet trained to admire the emperor's clothes will state the obvious.
The parable sets a test that the whole book then examines: could a machine answer Adam truthfully? Chapter 1 asks whether behaviour could show it; Chapters 2 to 4 ask what algorithms can and cannot do; Chapters 5 to 8 look for the physics; Chapters 9 and 10 ask how brains might use it.
Can a Computer Have a Mind?
Penrose sets out the question and the opponent. Strong AI holds that mind is nothing but the running of the right algorithm, so that any adequate computer, or even a thermostat in rudimentary form, has mental qualities. He accepts the Turing test as roughly valid, takes Searle's Chinese room seriously but rejects Searle's biological answer, and traces the strong-AI faith back to two facts: the equivalence of universal computers and the identity of elementary particles. His own thesis is announced: it is our ignorance of physical law, not of programming, that keeps mind out of reach.
The question and Penrose's unconventional stance
Machines that outrun us physically cause no distress; cars move five times faster than the swiftest athlete, aircraft carry us across oceans in hours. Thinking is different: it is the capacity that has seemed to set us above other creatures. If machines excel us there, Penrose asks, have we surrendered our unique superiority to our creations?
He lists the questions behind the book: what it means to think or feel, whether minds exist independently of physical structures, whether those structures must be biological, and whether minds obey the laws of physics. He promises guesses clearly separated from fact, and a viewpoint "not seen expressed before": our present lack of understanding of fundamental physics is what prevents us from grasping mind in physical or logical terms.
- The standard view he opposes. Most physicists say the laws at the scale of the brain are perfectly well known. The open gaps (particle masses, quantum gravity, space at 10-20 of a particle's size, the finiteness of the universe, black-hole cores, the big bang) seem remote from the brain.
- His counter-claim. There is "another vast unknown" right at the scale relevant to thought, unrecognized by most physicists, and, remarkably, black holes and the big bang bear on it.
- The itinerary. Quantum theory, special and general relativity, black holes, the big bang, the second law, Maxwell's theory, Newtonian mechanics, neurophysiology, AI, Turing machines, computability, Gödel's theorem, complexity, and the foundations of mathematics.
The Turing test
Alan Turing's 1950 paper Computing Machinery and Intelligence (in the journal Mind) proposed an operational criterion. A computer and a human volunteer are hidden from an interrogator who communicates only by keyboard and screen. The human answers truthfully and tries to prove he is human; the computer is programmed to lie and claim to be human. If across a series of tests the interrogator cannot consistently tell which is which, the computer passes.
- Players. One interrogator ("she", Penrose's deliberate choice, guessing a woman would be more sensitive to real human quality), one human, one machine.
- Channel. Typed text only; no appearance, voice or timing cues.
- Roles. Human tells the truth; machine pretends to be human.
- Criterion. Failure to identify the machine consistently across trials.
- Turing's 1950 forecast. A 30% machine success rate, against an average interrogator after five minutes, by the year 2000. Supporters in 1989 talked of passing outright by 2010.
Why the test is "unfair" to the machine
Reverse the roles and a human pretending to be a computer would be exposed instantly by a big multiplication. The machine must therefore play "stupider" than it is. Penrose notes the calculating prodigies: Johann Martin Zacharias Dase (1824 to 1861), an illiterate farmer's son, multiplied two 8-digit numbers in under a minute and two 20-digit numbers in about six minutes; Alexander Aitken (Edinburgh, 1950s) was as impressive. A fair arithmetic probe would need something like two 30-digit numbers in two seconds.
The programmers' real difficulty is common sense. Penrose's sample probe: "I hear that a rhinoceros flew along the Mississippi in a pink balloon this morning." A guarded "That sounds rather ridiculous" passes; the follow-up "My uncle did it once, both ways, only it was off-white with stripes" may trap a machine into "Rhinoceroses can't fly" or "Rhinoceroses don't have stripes". Switch to under the Mississippi or in a pink nightdress and see whether it notices the difference. Lack of understanding shows under sustained, original, probing questions.
What passing would mean
- Moral stakes. If a machine really feels, then switching it off, selling it, or causing it pain raises the same obligations as maltreating a slave; the authorities would need to know whether "thoroughly Turing-tested" claims are true.
- Why the test is strong. We judge other people's minds by conversation (plus faces and gestures, which a future robot might imitate). So Penrose accepts the test as "roughly valid".
- A weaker version. He would settle for the interrogator becoming convinced of "a conscious presence", possibly alien, but keeps Turing's version for its objectivity.
- His real position. Imitation, however skilful, ought to be detectable by skilful enough probing, but that is "more a matter of faith (or scientific optimism) than proven fact". If a machine truly passed, he would guess it thinks.
- Consciousness is a real thing. Words like "thinking" and "aware" refer to objective features whose presence we infer from evidence, like an astronomer inferring a star's mass. Being made of transistors is not contrary evidence; a future physical theory of consciousness, perhaps with a "consciousness detector" reliable on humans but disagreeing with the Turing test on machines, would be.
He grants the Turing test evidential weight and denies that silicon is disqualifying. What he denies is that behaviour exhausts the question: consciousness is a physical fact to be explained by physics we do not yet have.
Artificial intelligence in 1989
AI aims to imitate, and perhaps surpass, human mental activity by machine. Penrose names four constituencies: robotics for industry and dangerous conditions; expert systems coding a whole profession (medical, legal) into software; psychology, learning about brains by trying (or failing) to imitate them; and philosophy, hoping AI will illuminate what mind is.
| Early landmark | Date | What it did | Penrose's point |
|---|---|---|---|
| Grey Walter's tortoise | early 1950s | Roamed, sought a socket when its battery ran low, recharged, wandered off | Simple drive-switching behaviour (it actually avoided light) |
| Winograd's blocks program | 1972 | Conversed sensibly about simulated coloured blocks it stacked | Narrow worlds are tractable |
| Robot arms and vision | ongoing | Struggled to move among obstacles or interpret scenes | What looks effortless to animals is hard for AI |
| Colby's "therapist" | mid-1960s | Teletype dialogue some patients preferred to a human therapist | Eerie fluency from simple mechanical rules, no understanding |
| Simulated schizophrenic | 1970s | Gave textbook symptoms, fooled some medical students | The converse illusion |
| Spracklen chess program | 1989 | Elo 2110, USCF Master title on a Fidelity Excel microprocessor | Calculation can mimic skill |
| Deep Thought (Hsiung Hsu, CMU) | Nov 1988 | Elo about 2500; shared first at Long Beach with Tony Miles; first machine to beat a grandmaster (Bent Larsen) in a tournament | Machines approach International Master (world champion Kasparov over 2700) |
Calculation versus judgement
Chess machines do relatively better at fast time controls; humans do better with time to think. Penrose's explanation: the machine relies on rapid, exhaustive calculation plus "book" knowledge, while the human uses slow, conscious judgements that prune the tree of possibilities, allowing much greater depth when time allows. The gap is wider in go, with far more possibilities per move. The link between consciousness and judgement is flagged as the hinge of Chapter 10. A footnote adds a challenge: chess problems with an "obvious plan" some 200 moves deep would be easy for humans and beyond a 1989 problem-solver for a thousand years.
Replace Schank's few-page rulebook with the weights of a modern language model: about 1012 numbers and a recipe for multiplying them. A person executing it by hand at one multiplication per second would need on the order of tens of thousands of years per word of reply. Penrose's "critical complication" worry bites: no human could ever run it, so the single-person intuition pump loses force. Yet Searle's core point survives the scale-up: nothing in the arithmetic refers to restaurants or hamburgers. Ask yourself where, if anywhere, the understanding would reside: in the numbers, in the hand that multiplies, or in the conversation as a whole? Penrose's answer is that the question cannot be settled by behaviour; it needs a physical theory of consciousness.
Swap a person's neurons for functionally identical silicon units one at a time, the person reporting their experience after each swap. Strong AI predicts nothing changes; a biological view (Searle) predicts experience must fade somewhere, though behaviour, by construction, cannot reveal it. Penrose's 1989 position gives a third answer: if neurons rely on non-computable physics, then "functionally identical" silicon cannot be built from computational parts, so the experiment cannot even be carried out as specified. The thought experiment thus tests the very premise Penrose attacks: that all relevant physics is digitally simulable.
Four viewpoints: how Penrose later sharpened Chapter 1
Five years after this book, in Shadows of the Mind (1994), Penrose tidied the positions that Chapter 1 sets against each other into four labelled viewpoints, and the labels are now the standard way to locate anyone in this debate. They are worth learning here, because the whole of The Emperor's New Mind is an argument for one of them.
- A. All thinking is computation; in particular, feelings of conscious awareness are evoked merely by the carrying out of appropriate computations. This is strong AI, the target of the book.
- B. Awareness is a feature of the brain's physical action; whereas any physical action can be simulated computationally, computational simulation cannot by itself evoke awareness. This is roughly Searle's position.
- C. Appropriate physical action of the brain evokes awareness, but this physical action cannot even be properly simulated computationally. This is Penrose's own position.
- D. Awareness cannot be explained by physical, computational or any other scientific terms. This is the mystical or dualist position Penrose rejects as firmly as A.
The value of the scheme is that it separates two questions Chapter 1 runs together. One is whether simulation equals the real thing (A says yes, B says no). The other is whether the brain's physics is simulable at all (A and B say yes, C says no). Penrose's distinctive and testable claim is the second one. It commits him to finding physics that no Turing machine can follow, which is why the rest of this book is mostly physics.
| Viewpoint | Is the brain simulable? | Does simulation give awareness? | Who holds it |
|---|---|---|---|
| A | Yes | Yes | Minsky, Moravec, Dennett (in spirit), most of AI |
| B | Yes | No | Searle |
| C | No | (question does not arise) | Penrose |
| D | No scientific account | No | Dualists, mystics |
An actor plays a grieving widow so well that the audience weeps. A strong-AI theorist says the actor's brain is, for those minutes, running a grief program, and if grief is a computation then the actor is grieving. Most of us say the actor is performing grief while feeling, at most, concentration and stage fright. Now ask the same of a language model that writes a moving elegy: is it the widow or the actor? The Turing test cannot tell them apart, because it was built to ignore exactly that difference. Penrose's point in Chapter 1 is that this is not a defect of the test but a question about what minds are, which behaviour alone cannot answer.
Key terms from Chapter 1
| Term | Meaning in this book |
|---|---|
| Strong AI | The claim that mental qualities, including awareness, belong to any system carrying out the right algorithm |
| Weak AI | The claim that computers are useful tools for modelling and studying the mind, with no claim that they have one |
| Turing test | Turing's 1950 "imitation game": a machine passes if an interrogator cannot reliably tell it from a human through text alone |
| Chinese room | Searle's 1980 argument that symbol manipulation by rule cannot produce understanding |
| Algorithm | A finite, mechanical procedure; made precise in Chapter 2 as a Turing machine |
| Hardware / software | The physical machine versus the program; strong AI locates the mind in the software |
| Teleportation machine | Penrose's test case for the strong-AI view of personal identity: if the person is the information, a copy is the person |
Test yourself on Chapter 1
- Why does Penrose accept the Turing test as a reasonable criterion, then spend a book arguing no computer will pass it with real understanding? Because he grants that sustained, probing conversation would be good evidence of thought, and then argues on physical grounds that no algorithmic system can produce the understanding such conversation would require. His target is the algorithm, not the test.
- What, exactly, does Searle's Chinese room claim? That a person following rules to manipulate Chinese symbols, producing perfect answers, still understands no Chinese; so running a program is not sufficient for understanding.
- What is the "systems reply" and why does Penrose find it weak? It says the whole room, not the person, understands. Penrose replies that it is hard to see how adding paper and rule books to a non-understanding person creates understanding, especially when the person could memorize the rules and become the whole system.
- Why does the teleportation case bother Penrose? Strong AI implies that a perfect copy made from transmitted information is you, and that destroying the original is harmless. Penrose finds this implausible and notes (Chapter 6) that quantum states cannot be copied, so the scenario may be physically ruled out.
- What is the book's one-sentence answer to the chapter title? A computer can simulate behaviour indefinitely well, but on Penrose's view a computer running an algorithm cannot have a mind, because minds rely on physics no algorithm captures.
Try it in Python: a DOCTOR in a dozen lines
Penrose describes ELIZA's DOCTOR program, which imitated a psychotherapist by turning statements back into questions. The trick is a short list of pattern-and-template rules, with no model of meaning at all.
import re
rules = [(r'\bI am (.*)', 'How long have you been {0}?'),
(r'\bI feel (.*)', 'Why do you feel {0}?'),
(r'\bmy (mother|father)\b', 'Tell me more about your {0}.'),
(r'\byou\b', 'We were discussing you, not me.')]
def doctor(line):
for pat, reply in rules:
m = re.search(pat, line, re.I)
if m:
return reply.format(*[g.rstrip('.!?') for g in m.groups()])
return 'Please go on.'
print(doctor('I am worried about computers.'))
# -> How long have you been worried about computers?
print(doctor('What does it feel like to be you?'))
# -> We were discussing you, not me.
The second reply is Adam's question deflected by a rule that matched the word "you". A dozen lines produce the feeling of being listened to, which is why Penrose warns that short, friendly conversations are weak evidence of understanding, and why Turing's test needs a probing interrogator. Add rules and the illusion improves; nothing in the program comes closer to knowing what worry is.
- Chess. Deep Blue beat Kasparov in 1997; today's engines (Stockfish with neural evaluation, AlphaZero-style learners) play hundreds of Elo points above any human.
- Go. AlphaGo beat Lee Sedol 4-1 in 2016. Its value network is a learned, fast "judgement" that prunes search, much like the human faculty Penrose described, but produced by training rather than conscious insight.
- Conversation. In a 2025 three-party Turing test (Jones and Bergen), GPT-4.5 was judged human 73% of the time, more often than the real humans. Turing's 30% target was met a quarter-century late and then exceeded.
- What has not changed. None of this settles Penrose's question, which concerns understanding and awareness, not behaviour. The LLM era has made the Chinese room argument more pressing, not less.
Pleasure and pain, the AI way
Grey Walter's tortoise changed behaviour when its battery ran low; it is not much of a stretch to call it "hungry". AI supporters extend the idea to feelings. Imagine a device with a pleasure-pain score (Penrose's "pp-score") running from -100 (extreme pain) to +100 (extreme pleasure), and actions geared to maximize it.
- Inputs. Battery charge (low counts negative), solar panels that reward moving toward light, probability-weighted predictions of future scores.
- More goals. To distinguish "pain" from "hunger" it needs other drives: a "desire" for companionship with similar devices, a "craving" for learning facts, or, more selfishly, a reward for serving us (a robot servant). Natural selection imposed goals on us in a similar way.
- The operationalist definition. Since the device seeks high scores and avoids low ones, define its pleasure as positive score and its pain as negative score; humans, roughly, behave this way too.
Penrose's objection: the device's score-driven responses resemble our involuntary reflexes far more than our feelings. Touch a hot stove and your hand withdraws before you feel pain. He compares the AI vocabulary to everyday anthropomorphism ("my car doesn't want to start", "my watch thinks it's on Californian time"): helpful, half-jocular, not literal. AI deserves respect as a serious case; computers will get faster and more parallel, and perhaps the philosophy is right. But whether algorithms can feel "cannot be dismissed lightly".
The pp-score device
Behaviour is driven by a number it maximizes; "pain" is defined as a negative score; nothing in the design requires experience.
A feeling creature
Pain is felt; it often arrives after the reflex; we sometimes court pain and avoid pleasure; the feeling is what needs explaining.
Strong AI and Searle's Chinese room
Strong AI (Searle's term; often called functionalism, held by Minsky, Fodor, Hofstadter, Moravec) says mental qualities belong to the logical functioning of any computational device, down to a thermostat, whose algorithm is just "if hotter than the setting, disconnect; otherwise connect". A brain differs only in complication, "higher-order structure" or "self-reference". If the brain's algorithm exists, it could run on any general-purpose computer and would, in itself, experience feelings, have consciousness, be a mind.
The room
Roger Schank's story-understanding program (Schank and Abelson 1977) answers questions like "did the man eat the hamburger?" for two restaurant stories: in one the hamburger is burned to a crisp and he storms out without paying (answer: no); in the other he is delighted and tips generously (answer: yes). In this narrow sense a Turing test is already passed.
Tell the stories in Chinese. Searle, who knows no Chinese, sits in a locked room with the program's instructions written in English for shuffling counters bearing Chinese symbols. Stories and questions come in through a slot; he follows the rules; the correct Chinese "yes" or "no" goes out. He understands nothing. Carrying out a successful algorithm does not, by itself, imply that any understanding has taken place.
Objections Penrose weighs
- Patterns, not words. Understanding concerns patterns; a symbol shuffler might perceive some (swap "hamburger" for "chow mein" and nothing changes). But very little of the stories' meaning would come through.
- Sheer length. Hand-executing even Schank's program would take months or years; a brain-level program might take more than a lifetime per answer. Penrose takes this seriously: there might be a critical level of complication needed for mental qualities, beyond what any person could execute. (He counters Hofstadter's complaint that no one could internalize a whole mind: one need only execute the part of the algorithm that embodies a single mental event.)
- The India version. Searle replaces himself by the entire population of India (minus Chinese speakers). Objection: the individuals are like neurons, which do not understand either. Searle's reply: a country, like a thermostat or a car, is "not in the business of understanding". Penrose finds this reply much weaker than the single-person version.
Penrose's verdict. Searle's argument is strongest with one person and a lifetime-sized algorithm. It does not rigorously exclude some disembodied "understanding" attached to the execution, but makes it implausible. It is convincing that Schank-level programs understand nothing, and suggests (no more) that no algorithm alone can embody understanding.
The irony: strong AI is dualism
For strong AI only the algorithm's logical structure matters; the substrate (brain, computer, a nation of Indians, wheels and cogs, water pipes) is irrelevant. The algorithm therefore has a disembodied existence apart from any physical realization. Penrose, following Searle, calls this an extreme form of Descartes' dualism, the very view strong AI's supporters most want to avoid, with "mind-stuff" replaced by the logical structure of an algorithm.
Douglas Hofstadter, a leading strong-AI advocate, imagines a monstrous book describing Einstein's brain so completely that following its instructions answers any question as Einstein would. On strong AI the book is Einstein, thinking very slowly. Penrose's questions: is Einstein aware when the book sits unopened? When read by X-ray tomography? Twice over if two people ask the same question at different times? Only when the book is changed (since awareness alters memory)? Hofstadter raises some of these but answers none.
Penrose against Searle too
Searle accepts that present-style computers may pass the Turing test and that "everything is a digital computer", yet says brains have intentionality and semantics because of their biological construction, for reasons he cannot give. Penrose calls this "suspiciously like a dogmatic assertion", no less dogmatic than strong AI. His diagnosis: Searle and many others "have been led astray by the computer people. And they, in turn, have been led astray by the physicists." The book's job is to show why "everything is a digital computer" need not be true.
Hardware, software and the identity of persons
Turing's discovery: once hardware reaches a certain flexibility, any machine A can be made to act precisely like any other machine B by suitable software, given unlimited "rough paper" (tape, discs). Speeds may differ by factors of a thousand either way, but outputs match. All general-purpose computers are universal Turing machines, and this equivalence, Penrose thinks, is the main root of strong AI: hardware unimportant, algorithm everything.
The second root comes from physics. What makes you you?
- Turnover. Nearly all the atoms of the body, brain cells included, are replaced many times since birth. (Penrose adds that no new brain cells form after birth.)
- Identical particles. In quantum mechanics any two electrons are identical, not merely indistinguishable: swap an electron in your brain with one in a brick and the state is exactly the same (a footnote handles the sign subtlety: rotate one electron through 360° during the swap). Swap your whole material content with your house's bricks and nothing has happened.
- Conclusion. Identity lies in the pattern of the constituents, as with letters on a word-processor screen that are retyped, re-laid-out sixty times a second, or stored as magnetization on a disc and restored tomorrow.
Strong AI pushes on: if the information of your configuration can be translated into another form (say fields of magnetization in a block of iron) and recovered, your individuality, even your awareness, persists. You are software; your body is the hardware running it.
The teleportation paradox
The traveller is scanned atom by atom, the data beamed at light speed to another planet, a perfect duplicate assembled with all memories and hopes, and the original "safely" destroyed. Is this travel, or duplication plus murder? If walking between rooms is just atoms passing on information, what is the difference? If the original is not destroyed ("the other you, er, I mean the actual you, has arrived safely on Venus, so we can dispose of you here"), is awareness in two places?
Penrose suspects physics may forbid the scenario: perhaps copying must destroy the original, or keeping two viable copies is impossible in principle, and this is a pointer to an essential role for quantum mechanics in mind (picked up in Chapter 6). For strong AI, the information stored at a relay station between planets would itself be the conscious traveller, since a simulation could answer questions exactly as he would. Penrose: this rests on the unjustified presumption that the brain is a digital computer, that no specific physical phenomena are used in thinking.
Granting that all relevant physics can be simulated digitally, the only remaining assumption is the operational one: whatever acts consciously feels conscious. Penrose's whole book attacks the first premise, that physics is entirely computable, and argues it is an appreciable assumption, not a harmless one.
Common misreadings of Chapter 1
- "Penrose thinks computers can never be intelligent." He thinks they can never understand or be aware while running algorithms. He readily grants that they may outperform people at many tasks, and he treats machine mastery at chess as no threat to his thesis.
- "Penrose is a dualist." He rejects dualism (viewpoint D) as firmly as strong AI. Minds, for him, are physical; the claim is that their physics is not yet known.
- "The Turing test is useless." Penrose calls it a reasonable criterion and builds his case on what would be needed to pass it with genuine understanding. His complaint is with the strong-AI reading of a pass, not with the test.
- "Searle's room proves strong AI false." Penrose finds it persuasive but not conclusive; he notes its weakness against the systems reply and against cases where the rule book is vast. He uses it as support, then builds his own, quite different, argument from logic and physics.
- "A thermostat has beliefs, according to the book." That is McCarthy's strong-AI claim, which Penrose reports in order to show how far the doctrine reaches, not one he endorses.
- The "therapist" transcript. The best-known therapist simulation is Joseph Weizenbaum's ELIZA/DOCTOR (1966); Kenneth Colby is mainly known for PARRY (1972), the simulated paranoid patient that fooled psychiatrists. Penrose's attribution (via Delbrück) blurs the two.
- "No new brain cells after birth." Outdated: adult neurogenesis in the hippocampus was shown in humans in 1998 (Eriksson et al.), though its scale in adults is still disputed (2018 studies disagreed sharply).
- Turing's own criterion. Turing proposed a prediction about usage ("one will be able to speak of machines thinking without expecting to be contradicted"), not a definition of thought; Penrose's "roughly valid" acceptance gives the test more philosophical weight than Turing claimed for it.
- Searle's position has been read more charitably than "dogmatic": Searle says brains have causal powers we do not yet understand, which is structurally close to Penrose's own appeal to unknown physics.
- Behavioural evaluation is not understanding. The rhinoceros probe is the ancestor of adversarial and out-of-distribution testing: sustained, novel, compositional questioning exposes pattern-matching. Red-team LLM systems this way rather than by benchmark scores alone.
- Substrate independence is an architectural assumption. "Hardware is irrelevant" is true for outputs of universal machines and false for time, cost and energy, which is exactly where enterprise AI succeeds or fails.
- Moral and regulatory stakes of claimed machine feeling, which Penrose raised in 1989, now appear in debates about AI welfare and in rules on disclosing that a user is talking to a machine (EU AI Act transparency duties).
Algorithms and Turing Machines
Penrose builds the machinery of computation from scratch: Euclid's algorithm, Turing's machine with its tape and finite states, binary codings that turn every machine into a number, the universal machine that imitates all others, and Turing's diagonal proof that no algorithm decides whether machines halt. The proof is constructive: given any halting-tester, we can build a computation it cannot settle but whose answer we can see. That gap between an algorithm and the person who understands it is the seed of the whole book.
What an algorithm is
The word comes from the ninth-century Persian mathematician Abu Ja'far Mohammed ibn Mûsâ al-Khowârizmî, whose textbook Kitab al-jabr wa'l-muqabala (about 825 AD) also gave us "algebra"; the spelling "algorithm" replaced the more accurate "algorism" by association with "arithmetic". Algorithms are far older than the word.
- 3654 = 2 × 1365 + 924
- 1365 = 1 × 924 + 441
- 924 = 2 × 441 + 42
- 441 = 10 × 42 + 21
- 42 = 2 × 21 + 0, so the highest common factor is 21. Divide, keep the remainder, replace the pair, repeat until a division goes exactly. It works for numbers of any size, each step is clear-cut, termination is clear-cut, and the whole procedure has a finite description.
Penrose notes that "take the remainder" hides another algorithm (school long division). In unary notation (n marks for the number n) the remainder of 17 ÷ 5 is found by repeatedly removing five marks until fewer than five remain (answer: 2). Unary is hopelessly inefficient but makes the operation elementary. What is algorithmic in one notation is algorithmic in any other; only efficiency changes, and efficiency is not his concern here.
Turing's machine
Precise definitions of "algorithm" arrived only in the 1930s. Alan Turing (1935-36, published 1937) was attacking Hilbert's Entscheidungsproblem, posed partly at the 1900 Paris Congress and fully at Bologna in 1928: is there a general mechanical procedure that can settle, one after another, all mathematical questions of a well-defined class? Hilbert's broader formalist programme had already been dealt "a crushing blow" by Gödel in 1931.
To answer, Turing had to say what "mechanical" means. He seems to have regarded the human brain as such a machine. Penrose's irony: by making "mechanical" precise, Turing showed some well-defined mathematical operations are not mechanical, which may give a loophole in Turing's own view of mind.
- Finite set of internal states, labelled 0, 1, 2, ... (possibly very many, but finite).
- Tape: a line of squares infinite in both directions, each blank (0) or marked (1), with only finitely many 1s at any time. It serves as input, "rough paper" and output. (Penrose prefers to picture the device moving through an environment rather than shifting an infinite tape.)
- Reading head: reads one square, moves one square left or right. Reading n squares or jumping k squares adds nothing in principle.
- Instruction table: each entry has the form state, symbol → new state, new symbol, move (R or L), possibly STOP. Example: 0 1 → 13 1 L ("in state 0 reading 1, go to state 13, leave the 1, move left").
- Conventions: start in state 0, tape blank to the left, data fed in on the right, answer read on the left after STOP ("a bell rings").
Binary labels. States are written in binary (1, 2, 4, 8, 16 ... As place values). A sample: the device in state 11010010 reading 0 applies 110100100 → 111L, writes 1, enters state 11, steps left, and so on until STOP.
Small machines
- UN + 1 adds one to a unary number: run right over blanks, over the 1s, turn the first blank into a 1, STOP. Four 1s become five.
- UN × 2 doubles a unary number (12 instructions, half of them altering marks).
- EUC performs Euclid's algorithm on two unary numbers separated by a 0: 22 instructions for 11 states, of which only 3 alter the tape. On 6 and 8 it halts leaving 11 (the answer 2).
Penrose's aside: explaining why EUC works is harder than the machine itself, "a not uncommon feature of computer programs". To understand fully why an algorithm does what it should requires insight. Is insight itself algorithmic? The question is planted here and harvested in Chapter 4.
Coding numbers: contraction and expanded binary
Plain binary on a tape cannot show where a number ends (the blank tape is also 0s) or separate several numbers. Penrose's fix is contraction: read a 0/1 string as the counts of 1s between successive 0s. Then 2 means "comma", and 3, 4, 5 ... can stand for instructions (minus, plus, times, "go to", "iterate").
- Code 0 as 0, 1 as 10, comma as 110 (and other symbols as 1110, 11110, ...).
- 13 (binary 1101) becomes 1010010.
- The list 5, 13, 0, 1, 1, 4 (binary 101, 1101, 0, 1, 1, 100) becomes a single 0/1 string; a zero between commas can be written as two adjacent commas.
- 6, 8 becomes ...00000101001101000011000000...; the pair 1583169, 8610 fits in two lines, while in unary it would "more than fill this entire book".
In expanded binary, XN + 1 (add one) needs 15 instructions and 8 states (find the final 0, make it 1, turn the following 1s to 0s: 167 = 10100111 → 168 = 10101000), yet XN × 2 is simpler than its unary counterpart: doubling in binary is appending a 0.
The Church-Turing thesis
Once you see machines doing arithmetic, conditionals ("if the answer exceeded so-and-so, do this") and instructions read from the tape, it becomes reasonable to define a mechanical operation as one a Turing machine can perform. The words algorithm, computable, recursive, effective all name this class.
- Generalizations add nothing in principle. Many tapes, a 2D plane or 3D array of squares, bigger alphabets, parallel devices: all compute exactly the same class (two tapes can be interleaved on one; a plane is two coordinate tapes). Parallelism, then fashionable for brain modelling, buys speed only: two devices that communicate are in effect one device.
- Independent convergence. Alonzo Church (with Stephen Kleene) gave the lambda calculus a little earlier; Emil Post gave a Turing-like scheme a little later; Herbrand and Gödel gave general recursiveness. All proved equivalent.
Church-Turing thesis (mathematical)
"Turing computability captures exactly what mathematicians mean by an effective procedure." Penrose is "very happy to accept" this.
Physical version
"Physical systems, brains included, can do no more and no less than Turing machines." Penrose treats this as a separate, open question, and the book's main concern.
Computable and non-computable numbers
Fractions and negatives are just pairs of natural numbers with codes for "−" and "/" (say 1110 and 11110). Infinite decimals are subtler: a machine's output may only be read after it halts, so it cannot print π forever. Instead, a machine that outputs the n-th digit when fed n counts as generating the number. π and √2 = 1.414213562... are computable in this sense; "the vast majority" of real numbers are non-computable (shown in Chapter 3), which will matter for whether physics can be described by computable structures. Machines can equally manipulate formulae (algebra, trigonometry, calculus), given a 0/1 coding of symbols.
Numbering every machine: the universal machine
Code R, L, STOP, arrow, comma as 2, 3, 4, 5, 6 (that is 110, 1110, 11110, 111110, 1111110 after expansion), table digits 0 and 1 as 0 and 10, omit arrows and their implied left sides by keeping instructions in order (inserting "dummy" orders such as 1100 → 00R to fill gaps), and drop the universal opening 00 → 00R and the final 110. Every machine becomes one natural number n, and Tn is "the n-th Turing machine".
- UN + 1 is T177642.
- XN × 2 is number 10,389,728,107.
- UN × 2 is number 1,492,923,420,919,872,026,917,547,669.
- XN + 1 is T450,813,704,461,563,958,982,113,775,643,437,908.
- The universal machine U is Tu with u a binary number over 5,000 digits long, derived in denary and checked to work by David Deutsch.
- Of T0 to T12, only T3 and T11 halt at all; T7 is "not correctly specified" (its code contains five 1s in a row); T12 is identical to T6.
Reading a tape as a number. Take the string from the first 1 to the last, drop the final 1 (a terminator) and read the rest in binary, so ...0001101110010000... Is 11011100 = 220, and a lone 1 is zero. Write Tn(m) = p when machine n, fed m, halts with output p.
The map (n, m) → p is itself algorithmic, so one machine U computes it: feed U the number n (in binary, terminated by 111110, safe because correct codes never contain five consecutive 1s) followed by m, and U dodges back and forth reading n's instruction list and applying it to m. U(n, m) = Tn(m): the universal machine is a universal mimic. All general-purpose computers are universal machines in this sense: the program is the initial part of the tape.
Hilbert's problem has no solution
Turing recast the Entscheidungsproblem as the halting problem: does Tn, fed m, ever stop? Write Tn(m) = □ (Penrose's empty box) when it does not, including cases where the machine gets stuck without an instruction (T4, T7) or halts with no 1s on the tape (so T3 is also a dud, while T11 gives Tn(m) = 0 for every m).
Why halting matters to mathematics
- Fermat's Last Theorem: no natural numbers w, x, y, z satisfy (x+1)w+3 + (y+1)w+3 = (z+1)w+3 (the shifts allow zero values). A machine that searches all quadruples and stops at a solution never halts if and only if Fermat was right. Fermat (1601-1665), a lawyer and contemporary of Descartes, claimed a proof "the margin was too small to contain".
- Goldbach's conjecture: every even number greater than 2 is a sum of two primes. A machine trying 6, 8, 10, 12 ... And stopping at the first counter-example halts iff Goldbach is false.
So a general halting-decider would settle vast tracts of mathematics automatically.
Turing's diagonal proof
- Assume a machine H with H(n; m) = 1 if Tn(m) halts and 0 if not (n in expanded binary so the separator 110 is unambiguous).
- Imagine the infinite array of all outputs Tn(m), rows n, columns m, with □ where a computation never ends.
- Use H to replace every □ by 0: Q(n; m) = Tn(m) × H(n; m) (run H first; if it says "halts", run Tn(m)). Every row is now a computable sequence, every computable sequence appears as some row, and the whole array is computably generated by Q.
- Apply Cantor's diagonal slash: take the diagonal Q(n; n) and add 1. The sequence 1 + Tn(n) × H(n; n) is computable, so it is some row, say row k.
- Put n = k: 1 + Tk(k) × H(k; k) = Tk(k). If Tk(k) halts, H(k; k) = 1 and 1 + Tk(k) = Tk(k), impossible. If not, H(k; k) = 0 and 1 + 0 = □, impossible.
- Therefore H does not exist: no algorithm decides halting, and (as Church also showed by other means) no general algorithm decides mathematical questions.
What the proof does not show. It does not exhibit any single question that is absolutely undecidable. For each individual case, one of the trivial algorithms "say yes" or "say no" is correct; the trouble is knowing which. Algorithms do not, in themselves, decide mathematical truth: the validity of an algorithm must always be established by external means.
How to outdo an algorithm
Now suppose H is merely sometimes effective: whenever it outputs 0 it is right that Tn(m) never halts, but it may itself run forever (□) on hard cases. (A trivial example outputs 1 as soon as Tn(m) halts.) Repeat the construction to get machine k with 1 + Tk(k) × H(k; k) = Tk(k).
- If Tk(k) halted, H(k; k) would be 1 and we get 1 + Tk(k) = Tk(k): contradiction. So Tk(k) does not halt.
- But H cannot say so: H(k; k) = 0 would give 1 + 0 = □. So H(k; k) = □.
- Hence for any given sound halting-tester H, there is an explicit procedure for finding k such that we know Tk(k) never stops but H cannot tell us. We have "done better than the algorithm".
The procedure for finding k from H is itself algorithmic, so one can build an improved tester H′ that "knows" Tk(k) = □; and that one can be outdone in turn, and so on (the iteration returns in Chapter 4). Penrose's real question: who is doing the knowing? Is the algorithm "knowing", or are we, and are we in turn only following rules built into our brains? The issue is not just algorithms but how one judges what is true.
Church's lambda calculus
Computability is an absolute mathematical idea, independent of tapes and heads. Church's scheme (written up in 1941, developed in the 1930s with Kleene) shows this with "magical economy".
- Everything is a function. Objects a, b, c, ... A′, b′ ... Are functions whose arguments and values are also functions. a = bc means "b applied to c". Functions of several variables are applied in turn: (fp)q, ((fp)q)r (later called currying).
- Abstraction. λx.[fx] is the function which, applied to a, gives fa; so λx.[fx] = f, which looks pedantic but lets us name functions that have no names: if "sin" did not exist we could define it as λx.[x − x3/3! + x5/5! − ...], or define "one-sixth cubing" as λx.[x3/6].
- Numbers as iterators. λfx.[f(fx)] turns g into "g twice". Church identifies this with 2; his 3 is "thrice", and so on. Numbers are "more like twice, thrice" than quantities.
- Successor: S = λabc.[b((ab)c)] (λabc.[(ab)(bc)] also works).
- Doubling: D = λabc.[(ab)((ab)c)].
- Addition: A = λfgxy.[((fx)(gx))y].
- Multiplication: M = λfgx.[f(gx)].
- Exponentiation: P = λfg.[fg], the most astonishing: raising to a power is just applying one numeral to another.
- Subtraction was hard: Kleene found how to express it in the early 1930s, a landmark. In 1937 Church and Turing independently showed lambda-definability and Turing computability coincide.
Other routes: Post's machines, Herbrand-Gödel recursiveness, the combinatory logic of Schönfinkel (1924) and Curry (1929) from which Church partly built, and modern unlimited register machines (Cutland 1980). All give the same class. The computer language LISP embodies Church's calculus. Penrose ends: computability "seems to have a kind of Platonic reality of its own", the theme of the next two chapters.
Worked detail: feeding the universal machine
Take n = 11 (the machine T11, which stops at the first 1 and changes nothing) and m = 6. The tape for U is n in binary (1011), then the terminator 111110, then m coded as a tape (binary 110 followed by the end-marker 1, giving 1101): ...000 1011 111110 1101 0000... U shuttles between the instruction block and the data block, at each step looking up what T11 would do in its current state on the current symbol. The output, read by dropping the final 1, is 0, as T11(m) = 0 for every m. The program is just the first part of the input, which is the whole idea of a stored-program computer.
Why no coding can remove the duds. Penrose notes we could improve the numbering so fewer numbers give useless machines, at the cost of making U's decoding job harder. But a coding that listed only machines that halt on every input would let us decide halting by checking membership, which Turing's theorem forbids. The proliferation of duds is not sloppiness; it is forced.
Try it in Python: Penrose's UN + 1
# Penrose's UN+1: add one to a unary number
rules = {(0,0):(0,0,'R'), (0,1):(1,1,'R'),
(1,0):(0,1,'STOP'), (1,1):(1,1,'R')}
def run(tape, rules, state=0, pos=0):
tape = dict(enumerate(tape))
while True:
s, w, move = rules[(state, tape.get(pos, 0))]
tape[pos] = w
pos += 1 if move in ('R', 'STOP') else -1
state = s
if move == 'STOP':
return [tape[i] for i in sorted(tape)]
print(run([0, 0, 1, 1, 1, 1, 0], rules))
# -> [0, 0, 1, 1, 1, 1, 1]: four 1s became five
Each rule reads (state, symbol) → (new state, new symbol, move), exactly Penrose's table format. Try writing UN × 2, or a rules table that never halts, and notice that nothing in the simulator can warn you in advance which you have written.
A vendor offers a sealed box, HaltCheck, guaranteed sound: whenever it says "never halts", that is true. You do not even need to open it. Feed its description into Turing's recipe: build Tk, the machine that computes 1 + Tn(n) × H(n; n), and ask the box about Tk(k). If it ever answered "never halts", Tk(k) would run on and contradict itself; so it must stay silent for ever, and you now know Tk(k) never halts. Every sound halting-checker comes with its own blind spot that its buyer can see. The vendor can patch the blind spot, and you can build the next one: the race never ends. This is the structure Chapter 4 turns into Gödel's theorem, and Chapter 10 into an argument about minds.
Variants of the Church-Turing thesis
Penrose distinguishes Turing's mathematical claim from a stronger physical one, and later writers have split the thesis into at least four versions. Keeping them apart prevents a common confusion about what his argument needs.
| Version | Claim | Status in 2026 |
|---|---|---|
| Church-Turing (mathematical) | Every effectively calculable function on the natural numbers is Turing computable | Universally accepted as a definition |
| Physical Church-Turing | Every function computed by a physical system is Turing computable (Deutsch 1985) | Open; Penrose denies it for brains |
| Extended (complexity) thesis | Any physical computer can be simulated by a Turing machine with only polynomial slowdown | Challenged by quantum computers (Shor 1994) |
| Strong AI thesis | The mind is a Turing-computable process | The book's target |
Penrose's argument needs the physical thesis to fail, not the mathematical one. He fully accepts Turing's analysis of what an algorithm is; his claim is that nature contains processes that are not algorithms. Quantum computers, by contrast, only threaten the extended thesis: they are faster for some problems but compute nothing new. That is why the quantum-computing revolution since 1994 neither confirms nor refutes the book.
Hypercomputation: can physics beat Turing?
Since 1989 a small literature has asked whether any physical system could compute a non-computable function, a possibility usually called hypercomputation. The proposals fall into three groups, and each one shows how hard Penrose's goal is.
- Infinite precision. Hava Siegelmann (1995) showed that a neural network with real-number weights of infinite precision can decide the halting problem. The catch is that infinite precision is physically unrealistic: noise destroys the extra power, as Penrose himself warns of analogue devices in Chapter 5.
- Exotic space-times. In Malament-Hogarth space-times, an observer can see the whole infinite history of a computer in finite proper time, so could learn whether a search ever halts. These require extreme general-relativistic geometry and are not believed to be buildable.
- Accelerating machines. A machine doing each step in half the time of the last would complete infinitely many steps in finite time. No known physics permits it.
Penrose's own candidate is different from all three: not infinite precision or infinite time, but a new law of state reduction that is itself non-computable. No one has yet written down such a law in a form that could be checked.
Your compiler team adds a feature: before running any program, the compiler warns "this may loop for ever". Management wants no false alarms and no misses. Chapter 2 tells you, with certainty, that this cannot be built for all programs. What engineers actually ship is a checker that is sound but incomplete: it proves termination when it can (Rust's borrow checker, total functional languages, Coq's termination checker) and otherwise refuses or says "unknown". Every practical verifier lives in the gap Turing opened. The interesting design question is which programs you are willing to reject in order to guarantee the rest.
Key terms from Chapter 2
| Term | Meaning in this book |
|---|---|
| Turing machine | A finite set of internal states, a tape of 0s and 1s, and a table saying what to write, which way to move, and which state to enter |
| Universal machine U | A single Turing machine that imitates any Tn when given n followed by the input |
| Expanded binary | Penrose's coding in which 0 is a separator, 10 is digit 0, 110 is digit 1, 1110 is a comma, so lists of numbers fit on one tape |
| Halting problem | Deciding, for arbitrary n and m, whether Tn on input m ever stops; Turing proved no algorithm does it |
| Computable number | A real number whose digits a Turing machine can generate one by one |
| Lambda calculus | Church's equivalent formalism based on functions applied to functions |
| Church-Turing thesis | Penrose's name for the claim that Turing machines capture everything mechanically calculable |
Test yourself on Chapter 2
- Why does Penrose insist the tape be potentially infinite? Because an algorithm must work for inputs of any size; a fixed memory bound would make it a finite table, not a general procedure. In practice the machine only ever uses a finite amount.
- Why can't U be a dud-free list of machines? Because a list containing exactly the machines that halt on every input would let us decide halting by membership, which the halting theorem forbids.
- What does the construction 1 + Tn(n) × H(n; n) achieve? It defines a function that differs from every Tn on the diagonal, so no machine computes it; hence no machine H decides halting.
- How does Penrose turn this into "outdoing" an algorithm? Any sound procedure for proving non-halting can be fed its own description; we then see a case it misses. This is the seed of the Gödel argument in Chapters 4 and 10.
- Does this show minds are non-algorithmic? Not by itself. It shows no given algorithm is complete; the further step, that our insight is not some unknown algorithm, is argued later and is the most disputed in the book.
- Fermat's Last Theorem was proved by Andrew Wiles (1994-95, with Richard Taylor). The search machine Penrose describes never halts; the book still calls the problem open. Goldbach's conjecture remains unproved (verified by computer to 4 × 1018).
- Busy Beaver 5 settled (2024). The collaborative bbchallenge project proved, with a Coq-verified proof, that the longest-running halting 5-state, 2-symbol machine runs 47,176,870 steps. Beyond that, halting quickly becomes intractable: some 6-state machines encode Collatz-like open problems, a vivid demonstration of Turing's theorem.
- Lambda calculus won. It underlies Haskell, OCaml, Lean and Coq, and Lean is now the language in which AI systems produce machine-checked proofs.
- Physical Church-Turing thesis. Quantum computers (Google's Willow, 2024-25) change efficiency, not computability: everything a quantum computer does, a Turing machine can do more slowly. So they do not supply the non-computability Penrose seeks.
Worked detail: why the diagonal proof is not a trick
Readers often feel the halting proof is too quick to be convincing. It helps to see that it is a counting argument in disguise.
- List every machine. Penrose's numbering puts every Turing machine somewhere in the list T0, T1, T2, and so on. The list is complete: any algorithm whatever has a number.
- Build a table. Row n, column m holds Tn(m), or a mark □ if it never halts. Every computable function of one number appears as some row.
- Walk the diagonal. Take the entries T0(0), T1(1), T2(2), and so on, and change each one (add 1, and treat □ as 0). The result is a function that differs from row n at column n, for every n.
- Ask whether it is computable. If a halting-decider existed, it would let us compute this diagonal function, so it would be some row. But it differs from every row. So the halting-decider cannot exist.
Nothing tricky is going on: the argument is the same as Cantor's proof that the reals cannot be listed, applied to a list that is complete. The surprising conclusion is forced by the completeness of the list, and that is why Penrose stresses it. Any claimed algorithm for deciding mathematical truth would also sit in the list, and the same diagonal walk would outrun it.
Common misreadings of Chapter 2
- "Turing machines are a model of real computers." They are a model of what can be computed at all, with unlimited memory and time. Real computers are finite-state machines that approximate them.
- "The halting problem means we can never know whether a program halts." For most programs we can. The theorem says no single algorithm decides halting for all programs. Particular cases are often easy.
- "Penrose proves minds are non-algorithmic in Chapter 2." The chapter proves only that every algorithm can be outdone. The step to minds needs the Gödel material of Chapter 4 and the argument of Chapter 10.
- "The universal machine is a special, more powerful kind of machine." It is an ordinary Turing machine. Its power is that its input can include a program, which is exactly why modern computers are general-purpose.
- "Lambda calculus is a different idea of computation." It is a different notation for the same class of functions; Turing proved the two equivalent (1937), which is part of why the thesis is trusted.
- Hilbert's tenth problem is not the Entscheidungsproblem. The tenth problem (1900) asks for an algorithm to decide solvability of Diophantine equations, answered negatively by Matiyasevich in 1970 (building on Davis, Putnam and Robinson); the Entscheidungsproblem (1928) concerns first-order logic. Penrose links them correctly in Chapter 4, but calling the tenth problem a "partial" posing of the Entscheidungsproblem is loose.
- "Schöfinkel" should be Moses Schönfinkel.
- The "outdoing" argument cuts both ways, as Penrose half-admits: since k is computable from H, a machine can do the same diagonal step. The real issue, deferred to Chapter 4, is whether we can know that our own procedure is sound, which is where his critics concentrate.
- Halting is the root of every "can we verify this automatically?" question: static analysis, guaranteed termination of agent loops, policy checking. Rice's theorem (all non-trivial behavioural properties of programs are undecidable) follows from Turing's proof, so practical verifiers must be sound but incomplete or restrict the language.
- The validity of an algorithm must be established by external means. In regulated settings (model risk management, for example under the PRA's SS1/23 principles) this is the formal reason why independent validation cannot be replaced by the model vouching for itself.
Mathematics and Reality
A journey into the strange world of "Tor'Bled-Nam" turns out to be a tour of the Mandelbrot set, generated by the one-line rule z → z2 + c. To explain it, Penrose builds the number systems from naturals to reals (whose fit to physics is an act of faith over at least 1042 in scale) and complex numbers (a "Great Miracle of Mathematics"). His conclusion is open Platonism: the Mandelbrot set, like the complex numbers, is discovered, not invented, and algorithms occupy only a narrow corner of that mathematical world.
A journey to Tor'Bled-Nam
The chapter opens as travel writing. A remote sensor shows an insect-like blob with streams running into it. Zoom out by about 15 and a whole "continent" appears, the blob being an offshoot joined by a single filament. Zoom in by about 10 on a wart: it resembles the whole creature and has a "fiveness" (five filaments meeting); its neighbours show sevenness, nineness, increasing by two as one enters the main crevice. Deeper lies "seahorse valley", with spiral tails; a 250-fold magnification of one tail (on a wart with "29-ness") reveals swirls of octopuses and seahorses; ×30 and then ×6 at a junction of two spirals reveal a tiny baby creature almost identical to the whole, with offspring of its own. No two regions are alike, yet the flavour is unmistakable; the variety is endless. "Tor'Bled-Nam" is "Mandelbrot" reversed.
Numbers, from counting to the continuum
- Natural numbers 0, 1, 2, ... Count discrete things (twenty-seven sheep, zero new ideas, one mistake).
- Integers add negatives, needed for subtraction (charge, bank balances, dates; a footnote notes the calendar has no year zero).
- Rationals n/m, needed for division.
- Reals, needed for limits: Wallis (1655) π = 2 × (2/1)(2/3)(4/3)(4/5)(6/5)(6/7)...; Gregory (1671) π = 4(1 − 1/3 + 1/5 − 1/7 + ...); Euler π = √{6(1 + 1/4 + 1/9 + 1/16 + ...)}.
A real number is an infinite decimal, and an infinite expansion can be contemplated whole if we know its rule: 1/3 = 0.333..., 93/74 = 1.2567567... (every rational repeats), or the irrational 0.2200022220000022222... with blocks growing by one. If an algorithm generates the digits, the number is computable (the base does not matter). But many reals are not: take the number whose n-th digit is 1 or 0 as the n-th Turing machine halts or not on input n. Even among computable numbers, deciding whether two are equal is not computable, which is why mathematicians work with all reals. (0.999... = 1.000... Is identified.)
How many reals? Cantor's diagonal
- Same size means one-to-one correspondence. Integers pair with naturals (0↔0, −1↔1, 1↔2, −2↔3 ...; −n↔2n−1, n↔2n), so both have ℵ0 ("aleph nought") members. Fractions are also countable. Galileo half-anticipated this in the early 1600s.
- The reals are uncountable. List all reals in (0, 1); read the diagonal digits (1, 4, 1, 0, 0, 3, 1, 4, 8, 5, 1, ...); build a number whose n-th digit is 1 if the diagonal digit is not 1, else 2 (0.21211121112...). It differs from every listed number, so no list exists. The number of reals is C, the continuum.
- Continuum hypothesis: is C equal to ℵ1, the next infinity after ℵ0? Penrose calls it "a famous unsolved problem".
- Computable reals are countable (list the digit-generating machines). The diagonal of that list gives a new real, but not a computable one, because deciding which machines belong in the list (do they keep producing digits?) is the halting problem. This is exactly the shape of Turing's argument.
How "real" are real numbers?
Real numbers have infinite divisibility: between any two lies a third. Physical space may not. At the quantum-gravity scale, about 10-20 of a particle's size (the Planck length, about 10-35 m), distance may lose meaning; mirroring the reals would need scales of 10-200, 10-2000 or 10-10200 of a particle size, which may mean nothing physically.
Nature has been "remarkably kind": the reals work from below one-hundredth of an electron's classical diameter apparently down to the quantum-gravity scale, and out past the most distant quasar, a range of at least 1042, perhaps 1060 or more. Penrose stresses this is an extraordinary extrapolation. Our confidence, "perhaps misplaced", rests on the reals' logical elegance, consistency and power, plus "a belief in the profound mathematical harmony of Nature".
Complex numbers: a Great Miracle
Postulate i2 = −1. "Imaginary" numbers are no less real than "real" ones, since the reals themselves are an idealization. Then every negative real has square roots ±i√a; i itself has square root (1 + i)/√2; and that has a square root too. A complex number is a + ib with real part a and imaginary part b, and:
- (a + ib) + (c + id) = (a + c) + i(b + d)
- (a + ib) × (c + id) = (ac − bd) + i(ad + bc)
Complex numbers were introduced to take square roots. They deliver, free: cube, nth, even (1 + i)th roots (Euler); the fundamental theorem of algebra, that every polynomial a0 + a1z + ... + anzn = 0 (an ≠ 0) has a complex solution (even z102 + 999z33 − πz2 = −417 + i); and trigonometry collapsed into eiA+iB = eiAeiB via Euler's formula eiA = cos A + i sin A, whose real and imaginary parts are the addition formulae for cos(A + B) and sin(A + B).
The Argand plane
Represent z = x + iy as the point (x, y): 0 at the origin, 1 on the x-axis, the real numbers forming the real line. The name honours Jean Robert Argand, a Swiss bookkeeper (1806), though Caspar Wessel, a Norwegian surveyor, gave a full account nine years earlier and Wallis used it crudely two centuries before. Gauss exploited the picture with a topological ("rubber-sheet geometry") argument to prove the fundamental theorem of algebra.
- Addition is the parallelogram law (vector sum). Sample points: u = 1 + 1.3i, v = −2 + i, w = −1.5 − 0.4i.
- Multiplication: the distances from 0 multiply and the angles add; equivalently, the triangle 0, v, uv is similar to the triangle 0, 1, u.
Constructing the Mandelbrot set
Fix a complex number c and iterate z → z2 + c starting from 0: 0, c, c2 + c, c4 + 2c3 + c2 + c, c8 + 4c7 + 6c6 + 6c5 + 5c4 + 2c3 + c2 + c, ... (Worked example with c = 1.63 − 4.2i: 3 ↦ 10.63 − 4.2i; −2.7 + 0.3i ↦ 8.83 − 5.82i.)
Mandelbrot set M = { c : the sequence 0, c, c2+c, ... Stays bounded } (inside some fixed circle). - c = 0: 0, 0, 0, ... Bounded (in M). - c = −1: 0, −1, 0, −1, ... Bounded (in M). - c = i: 0, i, i−1, −i, i−1, −i, ... Bounded (in M). - c = 1: 0, 1, 2, 5, 26, 677, 458330, ... Unbounded (outside). - c = −3: 0, −3, 6, 33, 1086, ... Unbounded. - c = i − 1: 0, i−1, −i−1, −1+3i, −9−5i, 55+91i, −5257+10011i, ... Unbounded.
The pictures are computer-drawn: for each pixel c, iterate and apply "some appropriate criterion" for boundedness, then paint black or white. The structure is not sensitive to the exact formula; maps like z → z3 + iz2 + c (started at a suitable critical point) give extraordinarily similar shapes, a universality studied in complex dynamical systems.
Platonic reality of mathematical concepts
Are mathematical objects mere mental idealizations, or is thought "being guided towards some external truth... Revealed only partially to any one of us"? Penrose's star witness is the Mandelbrot set. Benoit Mandelbrot, the Polish-American founder of fractal theory, had no prior conception of its elaboration; when the first fuzzy pictures appeared he thought they were a computer malfunction (Mandelbrot 1986). No person or computer can reveal its full detail; any working computer finds the same structure, used like an experimenter's apparatus. "Like Mount Everest, the Mandelbrot set is just there!"
The same holds for the complex numbers themselves:
- c. 1539 Gerolamo Cardano (1501-1576), physician, gambler, horoscope-caster (he cast one for Christ), meets square roots of negatives in the "irreducible" cubic with three real roots.
- 1545 Cardano's Ars Magna gives the general solution of the cubic by surds (building on del Ferro and Tartaglia); the route passes through √(negative) even though the answers are real.
- 1572 Raphael Bombelli's L'Algebra begins the actual algebra of complex numbers.
- 17th to 19th c. Wallis, Cotes, Euler, Wessel, Argand, Gauss reveal more; later the Cauchy integral formula, Riemann mapping theorem and Lewy extension property, all properties of Cardano's very numbers, "with no additional modifications whatever".
Discovery versus invention
Penrose is a Platonist about complex numbers and the Mandelbrot set, but with a distinction:
Discoveries ("works of God")
Structures where much more comes out than was put in, with compelling uniqueness and universality: complex numbers, the Mandelbrot set. Great discoveries are "closer to God", as great art feels like revealing eternal truth (Borges: a famous poet is "less of an inventor than a discoverer").
Inventions ("works of man")
Contrived, non-unique constructions introduced mid-proof for a specific end; nothing more comes out than was put in. Lesser art and ordinary engineering are similar, while an engineering idea of beautiful economy deserves the word "discovery".
He adds that the case for an "etherial, eternal existence" is stronger in mathematics than in art or engineering. The view goes back to Plato (c. 360 BC), hence mathematical Platonism.
If algorithms are "God-given" ideas with timeless existence, could minds live in that Platonic world as algorithms? Penrose: "Just conceivably so", and he will later speculate along not entirely dissimilar lines, but not with the concept of an algorithm. Algorithmic things are a very narrow and limited part of mathematics, and Chapter 4 shows how much lies outside.
A footnote reveals his preferred stance on the reals: he would like every real number's n-th digit to be fixed by some rule, even one that is neither effective nor definable in any preassigned formal system, a view he hopes is consistent.
Try it in Python: draw Tor'Bled-Nam
# The Mandelbrot set, coloured by escape time
import numpy as np, matplotlib.pyplot as plt
re, im = np.meshgrid(np.linspace(-2.2, 0.8, 900),
np.linspace(-1.3, 1.3, 780))
c = re + 1j*im
z = np.zeros_like(c); n = np.zeros(c.shape)
for k in range(80):
z = z*z + c
alive = abs(z) <= 2 # escape radius 2
n[alive] += 1
z[~alive] = 2 # freeze escapees
plt.imshow(n, cmap='magma', extent=(-2.2, .8, -1.3, 1.3))
plt.title('z -> z^2 + c'); plt.show()
Change the limits to (-0.75, -0.73) by (0.10, 0.12) to visit "seahorse valley". Raising the iteration count from 80 to 1,000 sharpens the boundary: no finite count ever finishes the job, which is Penrose's point about why the set's membership is hard to certify.
Two civilizations that never meet, with different number notations, computers and aesthetics, each define "the points c for which 0, c, c2 + c, ... Stays bounded". Each plots it. They find the same cardioid, the same discs at the same coordinates, the same baby copies at the same magnifications, and argue over the same unsolved questions about its tendrils. For Penrose this is the signature of discovery: like two expeditions mapping the same mountain. A formalist replies that both simply implemented one definition. Ask which account better explains why the set keeps surprising the people who defined it.
Three worlds, three mysteries
The Platonism of this chapter grew, in Penrose's later books (Shadows of the Mind, 1994; The Road to Reality, 2004), into a picture of three worlds and three mysteries that he still uses in lectures. It is the cleanest summary of what Chapter 3 is really claiming.
- The Platonic mathematical world, containing objects such as the Mandelbrot set and the number system ℂ.
- The physical world, the space-time and matter physics describes.
- The mental world, the world of conscious perception and understanding.
Each world seems to emerge from a small part of the one before it. A small part of mathematics governs physics (only certain equations describe nature). A small part of physics produces minds (only brains, as far as we know). A small part of mental life reaches mathematics (only some thoughts grasp mathematical truth). The three mysteries are the three arrows, and Penrose thinks they may be one mystery seen three times. The circle explains why a book about computers keeps turning to mathematics and physics: in his view you cannot understand the mind arrow without the other two.
What the Platonist says
The Mandelbrot set was there before Mandelbrot; its seahorse valleys are discovered like a mountain range, and we are surprised by them. Truth in mathematics is independent of any person or formal system.
What the formalist says
Mathematics is the manipulation of symbols by rules; the Mandelbrot set is what you get when you apply the rule z → z² + c, nothing more. "Existence" means only consistency, and truth means provability in some agreed system.
Penrose's argument for the first column is not mystical. It is the observation that the structure is not what anyone put in. Mandelbrot wrote down a one-line rule; nobody chose the spirals, the miniature copies or the fact that the boundary has dimension 2. When the output vastly exceeds the specification, "invention" is the wrong word. The same holds, he says, for the complex numbers, which were introduced to solve cubic equations and turned out to make all polynomial equations solvable, all smooth functions analytic, and quantum mechanics possible.
Imagine physicists on a planet whose mathematics stopped at the real numbers. They would build Newtonian mechanics without trouble, and even electromagnetism (awkwardly). But at the atomic scale their equations would not close: interference, spin and the stability of atoms all need amplitudes that add with phases. They would be forced, by experiment alone, to invent something with the algebra of i. The 2021-22 network Bell tests did exactly this for us: they showed real-valued quantum theory is wrong, not merely clumsy. If nature forces a structure on every civilization that looks closely, that is about as strong as evidence for its independent reality can get.
Key terms from Chapter 3
| Term | Meaning in this book |
|---|---|
| Mandelbrot set M | The points c in the complex plane for which iterating z → z² + c from z = 0 stays bounded |
| Natural, integer, rational, real, complex | The ladder of number systems, each forced by an operation (subtraction, division, limits, square roots of negatives) |
| Cardinality ℵ₀ and C | The sizes of the countable sets (naturals, rationals) and of the continuum (reals) |
| Cantor's diagonal slash | The argument showing the reals cannot be listed, the template for Turing's and Gödel's theorems |
| Fundamental theorem of algebra | Every polynomial with complex coefficients has a complex root |
| Platonism | The view that mathematical objects exist independently of minds and are discovered |
| "God-given" versus "man-made" | Penrose's contrast between structures that yield more than was put in and those that do not |
Test yourself on Chapter 3
- Why does Penrose open a chapter on mathematical reality with a fantasy voyage? To make the reader experience the Mandelbrot set as a territory to be explored, so that the claim it is discovered feels earned rather than asserted.
- What forces the move from rationals to reals? The need for limits, such as √2, which no fraction equals. Penrose notes that real numbers are an idealization: physics has never measured an infinitely precise quantity.
- Why does Cantor's diagonal argument matter for the rest of the book? Its structure (assume a complete list, build a new item that differs from every entry) is reused by Turing for halting and by Gödel for provability.
- What is the "Great Miracle" of complex numbers? That adding one number, i, to solve x² + 1 = 0 makes every polynomial equation solvable and unlocks a theory of functions of unexpected power. Penrose takes this to be a sign of mathematical truth, not convention.
- Does Platonism help strong AI? Penrose argues the reverse: if minds contact an independent mathematical world, and that contact cannot be captured by any one formal system, then minds may be doing something no algorithm does.
- Complex numbers are experimentally necessary. In 2021-22, following a proposal by Renou et al. (Nature 2021), photonic and superconducting experiments (Chen et al.; Li et al., PRL 2022) showed that real-number versions of quantum theory are refuted in network Bell tests. Penrose's claim that complex numbers are "basic to the workings of the very world in which we live" now has direct evidence.
- The Mandelbrot set's boundary has Hausdorff dimension 2 (Shishikura, 1998). Whether M is locally connected (the MLC conjecture) is still open; Hertling (2005) showed M is computable in the natural sense if the hyperbolicity conjecture holds, which bears on the computability question Penrose raises in Chapter 4.
- The continuum hypothesis is independent of standard set theory: Gödel (1940) showed it cannot be disproved, Cohen (1963) that it cannot be proved, from the ZFC axioms. "Unsolved" now means "undecidable from ZFC", a Gödelian outcome in its own right.
- Discreteness of space remains unseen: gamma-ray-burst timing from Fermi constrains energy-dependent light speed (a common signature of granular space) near and beyond the Planck scale for linear effects.
Give a skilled graphic designer a month to invent a shape with infinitely many miniature copies of itself, each slightly distorted, joined by spiral filaments, with a boundary so crinkled it is almost two-dimensional. The designer would produce something, but it would be a specification, chosen feature by feature. The Mandelbrot set arrives with all those features from the single rule z → z2 + c, and with others no one has yet catalogued. A designed object contains what its designer put in; a discovered one keeps surprising its discoverers. That asymmetry, Penrose argues, is what distinguishes mathematical reality from invention.
Common misreadings of Chapter 3
- "Platonism means numbers exist in some place." Penrose's Platonism is about independence: mathematical truths hold regardless of minds or formal systems. It makes no claim about location.
- "The Mandelbrot set is only a pretty computer picture." The picture is an approximation; the set itself is a precise mathematical object whose detailed structure was unknown when its definition was written down. That gap is Penrose's evidence.
- "Complex numbers are an engineering trick." In quantum mechanics they are indispensable, a point now confirmed experimentally. Penrose's "Great Miracle" is that they were found in pure algebra centuries before physics needed them.
- "Real numbers are physically real." Penrose explicitly doubts this. He notes that physics has never measured a quantity to infinite precision, and that the continuum may be an idealization even though it works superbly.
- "Platonism helps strong AI by giving algorithms something to compute." Penrose argues the reverse: if understanding is contact with an independent world, it need not be reducible to any formal procedure.
- Roger Cotes lived 1682-1716; his version of Euler's formula dates from about 1714, not "the 16th century".
- Gauss's first proof of the fundamental theorem of algebra was his 1799 dissertation (with a topological gap later filled); 1831 is the year of his published geometric account of complex numbers. Penrose's "in 1831... The first general proof" mixes the two.
- Mandelbrot's priority is contested (Penrose's own note 3 hedges "first"): Brooks and Matelski drew the set in 1978, and Douady and Hubbard proved its connectedness (1982) and named it.
- Platonism is argued from aesthetic experience. Formalists and structuralists reply that "more comes out than was put in" is what one expects from rich axioms, and that the computer finds "the same" Mandelbrot set because it implements the same definition.
Truth, Proof, and Insight
This is the logical heart of the book. Russell's paradox drove Hilbert to seek a complete, consistent formal system for mathematics; Gödel (1931) showed that any consistent system rich enough for arithmetic contains a true statement it cannot prove, and Penrose stresses that we can see that statement is true. He defends Platonism against formalism and intuitionism, rebuilds Gödel from Turing via recursively enumerable sets, surveys non-recursive mathematics (Diophantine equations, the word problem, tiling), asks whether the Mandelbrot set is recursive, and introduces complexity classes P and NP. Mathematical truth cannot be captured by any formal scheme; the insight that sees a Gödel sentence is true is not the output of an algorithm.
Hilbert's programme and why it was needed
Late-nineteenth-century mathematics (Hilbert, Cantor, Poincaré) gained power by treating infinite sets as completed wholes. Gottlob Frege (1884) defined numbers through sets: "threeness" is a property of collections (medal winners in an event, tyres on a tricycle, leaves on a normal clover, the solutions of x3 − 6x2 + 11x − 6 = 0), and the number 3 is the set of all three-membered sets. This is not circular: numbers are totalities of sets that pair off one-to-one. (A footnote warns: the set of non-empty subsets of {apple, orange} has three members; the set whose only member is the empty set has one.)
Let R be the set of all sets that are not members of themselves. Self-membership is not absurd: the set of all infinite sets is itself infinite, so it belongs to itself. Now ask whether R belongs to R. If not, it qualifies and so belongs; if so, it fails the criterion and so does not. Contradiction either way. Penrose's library version: one catalogue lists every book that mentions itself, another every book that does not; in which catalogue is the second catalogue listed? Cantor anticipated the paradox; it descends from his diagonal slash.
Russell's point was not flippant: he used, in extreme form, the very set-theoretic reasoning mathematicians were relying on. Russell and Whitehead built a monumental, cumbersome formal system (Principia Mathematica) designed to block such reasoning. David Hilbert aimed higher: for each area of mathematics, axioms and rules capturing all correct reasoning, plus a proof that the scheme is free from contradiction, putting mathematics on "an unassailably secure foundation". In 1931 the 25-year-old Kurt Gödel destroyed the programme.
Formal systems, precisely
- Alphabet: numerals (Arabic, or 0, 01, 011, ...), variables t, u, v, w, x, y, z with primes for more (t′, t″), operations =, +, ×, brackets, logical symbols & (and), ⇒ (implies), ∨ (or), ⇔ (iff), ~ (not), and quantifiers ∃ (there exists) and ∀ (for all).
- Propositions and propositional functions: Fermat as F = ~∃w,x,y,z[(x+1)w+3 + (y+1)w+3 = (z+1)w+3]; or G(w) for a single exponent (G(0): no cube is a sum of two positive cubes), so F = ∀w[G(w)]. (In 1989 G(w) was known true up to about G(125000).)
- Axioms: self-evident truths such as (P & Q) ⇒ P, ~(~P) ⇔ P, ~∃x[R(x)] ⇔ ∀x[~R(x)], ∀x,y[x + y = y + x], ∀x,y,z[(x + y) × z = (x × z) + (y × z)].
- Rules of procedure: from P and P ⇒ Q, deduce Q (modus ponens); from ∀x[R(x)], deduce R(n) for any specific n.
- Proof: a list built from axioms by the rules, ending in the proposition; its last line is a theorem.
Hilbert hoped for completeness (for every syntactically correct P, either P or ~P is provable) and consistency (never both). Then truth could be defined as provability and meaning dispensed with: formalism, mathematics as "a kind of meaningless game". Penrose: it is meaning, not blind algorithmic computation, that gives mathematics its substance, and Gödel dealt formalism "a devastating blow".
Gödel's theorem
The ingenious, laborious part of Gödel's work is coding propositions, axioms and rules as arithmetic. Penrose skips it and gives the profound part.
- Order all strings of symbols lexicographically (by length, then alphabetically; Gödel used a cleverer numbering). Let Pn(w) be the n-th propositional function of one variable (allowing some to be syntactically junk).
- Number all proofs likewise: Πn is the n-th proof.
- Consider the propositional function ~∃x[Πx proves Pw(w)], "there is no proof of Pw(w)". Coding shows it is a genuine arithmetical statement about w. It is some function on the list, say the k-th: ~∃x[Πx proves Pw(w)] = Pk(w).
- Set w = k: Pk(k) says "Pk(k) has no proof in the system".
- If Pk(k) were provable, what it asserts would be false, so the system would prove a falsehood; a sound system cannot. So it has no proof, which is exactly what it says: Pk(k) is true.
- ~Pk(k) is then false, so it is not provable either. Pk(k) is undecidable within the system. Gödel also showed the system's own consistency, coded arithmetically, is unprovable within it.
Mathematical insight
Gödel's theorem is usually read negatively, as a limitation. Penrose reads it positively: in following the argument we have established that Pk(k) is true, though no proof exists inside the system. The formalist's notion of truth is necessarily incomplete. Retreating to "provability in a fixed system" cannot even frame Gödel's argument, which reasons about what is actually true. (A footnote concedes the argument can avoid a full external notion of truth but still relies on the meaning of symbols such as "~∃".)
The pragmatic formalist shrugs: Gödel sentences are absurdly complicated and never arise in serious mathematics. Penrose answers twice. First, natural statements of genuine mathematical character are now known to be Gödel-type: the Paris-Harrington theorem (1977), a strengthened finite Ramsey theorem (for any k, r, n there is m0 such that for all m > m0, colouring k-subsets of {0,...,m} with r colours always yields a "large" homogeneous subset of at least n elements, "large" meaning more elements than its smallest member), is true but unprovable in Peano arithmetic. Second, mathematicians do not check their reasoning against a formal system; they ask whether it validly establishes truth, and Gödel's argument does.
Adding Gödel sentences forever
- Adjoin G0 = Pk(k) as an axiom; the new system has its own Gödel sentence G1; adjoin it; then G2, G3, ...
- The whole infinite family is systematic, so it is a finite system with Gödel sentence Gω; continue Gω+1, Gω+2, ..., then Gω2, Gω3, ..., Gω2 and onward through Cantor's ordinals.
- Turing's 1939 paper "Systems of logic based on ordinals" studied this; remarkably, every true universally quantified arithmetical proposition can be reached by repeated Gödelization (Feferman 1988).
- The catch: at each limit stage one must systematize an infinite family into one axiom, and seeing that the systematization is correct requires insight from outside. It is these insights that cannot be systematized and must lie outside algorithmic action.
The faculty involved is what logicians call a reflection principle: by reflecting on the meaning of the axioms and rules, and convincing oneself they yield only truths, one codes that conviction as new truths not derivable from them (another reflection principle infers truths from the system's consistency). Reflection principles are "the very antithesis of formalist reasoning", and must be used carefully lest they slide toward Russell-type paradox with "enormous" sets. A footnote adds that some natural theorems would need Gödelization "to an outrageously huge degree" if proved from Peano's axioms (Smoryński 1983), though their real proofs use no doubtful reasoning.
Why accept Gödel and Turing but reject Russell? Theirs are clear-cut arguments; Russell's uses nebulous "enormous" sets. Penrose admits the line is "not really as clear as one would like", and that the contemporary fix (the distinction between sets and classes, the latter "too big" to be collected, with only the circular criterion that sets are what can be collected) is "not entirely satisfactory". Still, the verdict: mathematical truth "goes beyond the whole concept of formalism. There is something absolute and 'God-given' about mathematical truth." Formal systems are "provisional and man-made"; even choosing axioms requires self-evidence and meaning, which consistency alone cannot supply.
Platonism or intuitionism?
| School | Core claim | Penrose's verdict |
|---|---|---|
| Formalism (Hilbert) | Mathematics is manipulation of symbols in formal systems; truth = provability | Refuted by Gödel as a fundamental philosophy |
| Platonism (Plato, Gödel) | Truth is absolute and objects have timeless existence | Strongly favoured; mathematician-independence of the Mandelbrot set is its paradigm |
| Intuitionism (Brouwer, 1924) and finitism | Only mentally constructed objects exist; no completed infinities; reject excluded middle | "Much too extreme"; stifles mathematics |
Degrees of Platonism. For Penrose the absoluteness of mathematical truth and the existence of mathematical objects are essentially the same thing. Gödel was a very strong Platonist, extending absoluteness to all set theory. Penrose admits that with "wildly enormous and nebulously constructed" sets even "a fairly determined Platonist such as myself" may doubt that truth is absolute, which can take on a "matter-of-opinion quality". The sets that matter here are "ridiculously tiny" by comparison. The continuum hypothesis, shown independent of Zermelo-Fraenkel set theory by Gödel and Cohen, divides the camps: "meaningless" to a formalist, true or false to a Platonist (Cohen himself suggested a reflection principle making it "obviously false").
Intuitionism. Initiated by L. E. J. Brouwer in 1924 (rooted in Aristotle's rejection of Plato's actual infinities; named because it was meant to mirror human thought), it treats sets only through the rules for their membership and demands constructive existence. Hence it rejects the law of the excluded middle (~(~P) ⇔ P) and reductio ad absurdum for existence: refuting non-existence does not construct the object. Brouwer's worry began with his own fixed-point theorem (a disc moved continuously into itself has a fixed point, but where?), and the bisection method for a zero of a continuous function, which may require deciding whether intermediate values are exactly zero.
Does π's decimal expansion contain twenty consecutive 7s? Classically: either it does or it does not, and we do not know which. Brouwer: the statement is currently neither true nor false, and becomes one or the other only when settled. Probabilistically such a run should exist, but a machine producing 1010 digits per second would expect to need 100 to 1,000 years to find it. Penrose predicts a mathematical proof will come first, probably as a corollary of something deeper. On Brouwer's view Fermat's Last Theorem was likewise neither true nor false in 1989. Penrose: "such subjectiveness and time-dependence of mathematical truth is abhorrent."
Penrose credits the constructive impulse but notes that since Church and Turing, constructiveness can be studied inside conventional mathematics as computability, with no need to cripple existence proofs.
Gödel from Turing
A formal system must have an algorithmic proof-checker: no judgement should be needed to see whether a purported proof is valid. Then there is also an algorithm that lists all theorems: run through all strings lexicographically, keep those the checker accepts, read off their last lines. If Hilbert's system were complete and consistent, then for any P we would eventually find a proof of P or of ~P, giving a universal decision algorithm, which Turing proved impossible. So incompleteness follows from the unsolvability of the halting problem.
For arithmetic specifically, Turing-machine actions can be expressed in arithmetic plus one extra operation: "the smallest x such that K(x)" (the μ-operator), which may run forever if no such x exists. And the "outdo an algorithm" construction from Chapter 2 gives the true but unprovable part: for any sound halting-tester there is a non-halting computation we can see never halts but it cannot.
Recursively enumerable sets
Penrose recasts everything as sets of natural numbers (each number coding a proposition via a computable two-way correspondence).
- Recursively enumerable (r.e.): some algorithm lists its members (repetitions allowed). Membership can be confirmed when true, but if false the search may run forever.
- Recursive: both the set and its complement are r.e., so membership is decidable (run both listings alternately).
- Examples: evens {0, 2, 4, ...}, squares {0, 1, 4, 9, 16, ...}, primes {2, 3, 5, 7, 11, ...} are all recursive; their complements {1, 3, 5, ...}, {2, 3, 5, 6, 7, 8, 10, ...}, {0, 1, 4, 6, 8, 9, 10, 12, ...} are easily listed too.
Choose a consistent, "sensible" system strong enough to describe every Turing machine. Let P be its theorems, S(n) the statement "Tn(n) stops".
- P is r.e. (list proofs, read last lines); so are the disprovable propositions.
- The true S(n) all lie in P (a halting run can be written out as a proof), and only true ones do (soundness).
- The false S(n) are not r.e. (Turing).
- If P's complement were r.e., listing it would list all false S(n). Contradiction: P is not recursive.
- If the system were complete, P's complement would be exactly the disprovable propositions, which are r.e., making P recursive. Contradiction: the system is incomplete. "This is the main thrust of Gödel's theorem."
The set T of true propositions is worse still: neither T nor its complement is r.e. Truth is "vastly more complicated and profound" than provability. Some simple true classes are r.e., such as true statements ∃w,x,...,z[f(w,x,...,z) = 0] with f built from +, −, ×, powers (set A; the negation of Fermat is of this form), yet A is not recursive: no algorithm could in principle decide Fermat's Last Theorem in this class.
Is the Mandelbrot set recursive?
Non-recursive sets must be complicated "in a very essential way", defying systematization. Could the Mandelbrot set, conjured by an algorithmic computer, be one?
- The complement is r.e. Once an iterate leaves the circle of radius 1 + √2, the sequence is surely unbounded (the sharper standard bound is radius 2), so escape is confirmed in finitely many steps.
- The set itself? In 1989 no colleague knew an algorithm certifying a point is in M, nor a proof that none exists. "Perhaps the complement of the Mandelbrot set is, indeed, an example of a recursively enumerable set which is not recursive!"
What "recursive" can mean for a continuum
Recursiveness is defined for countable sets, but the plane has C points (complex numbers pair with reals by interleaving digits: 3.6781... + i512.975... ↔ 50132.6977851...). Penrose canvasses fixes, none fully satisfactory:
- Computable complex numbers are countable, but equality of computable numbers is undecidable (0.999... Versus 1.000...). So even the unit disc fails to be decidable on its boundary circle x2 + y2 = 1, absurd for so simple a set.
- Ignore boundaries: interior and exterior points are decidable. Natural for physics perhaps, but for M it ignores the tendrils, where the wildest complication lives; even for the blobs the answer hinges on the then-unproved local connectivity of M.
- Digits presented "by magic": a set is r.e. If one algorithm, fed the digits of any c, eventually says yes exactly for members. Again boundaries are ignored.
- Rational complex numbers: too restrictive to see tendrils.
- Algebraic numbers (roots of integer polynomials such as 129z7 − 33z5 + 725z4 + 16z3 − 2z − 3 = 0): countable, computable, with decidable equality, and many lie on M's tendrils and the unit circle. But for the set y ≥ ex, Lindemann's theorem (1882) implies the boundary contains only one algebraic point (z = i). "The correct viewpoint has not yet been arrived at."
Two notes update the hardback: Lenore Blum informed Penrose that M's complement is indeed non-recursive in the "magic digits" sense, and the new Blum-Shub-Smale (1989) theory of computation over the reals may be the right framework.
Examples of non-recursive mathematics
- Diophantine equations (after Diophantos): integer solutions of integer-coefficient polynomial systems. Example: z3 − y − 1 = 0, yz2 − 2x − 2 = 0, y2 − 2xz + z + 1 = 0 has the solution x = 13, y = 7, z = 2, but no algorithm decides solvability in general. This answers Hilbert's tenth problem negatively; nine variables suffice for undecidability.
- Topological equivalence of manifolds: decidable for 2-manifolds (count the handles: sphere ~ cube surface; ring ~ teacup), unknown for 3 (in 1989), undecidable for 4 or more. Since space-time is a 4-manifold, Geroch and Hartle (1986) suggested relevance to quantum gravity.
- The word problem for semigroups: given equalities between words, decide whether two words are equal. Undecidable in general, and for some specific lists (one adapted from G. S. Tseitin and Dana Scott, 1955).
- The tiling problem: does a finite set of polygons tile the plane? Undecidable (Berger 1966).
Rules: EAT = AT, ATE = A, LATER = LOW, PAN = PILLOW, CARP = ME. Then LAP = LATEP = LEATEP = LEAP, and CATERPILLAR = CARPILLAR = CARPILLATER = CARPILLOW = CARPAN = MEAN = MEATEN = MATEN = MAN. But CARPET ≠ MEAT: every rule preserves the count of As + Ws + Ms, which is 1 for CARPET and 2 for MEAT. Equality is shown by exhibiting a derivation (algorithmically searchable); inequality needed an argument about the rules, a "trick" Penrose took a while to find. For this list an inequality algorithm exists, but finding it took intelligence; no single algorithm works for all lists.
Tiling the plane
Squares, equilateral triangles and regular hexagons tile; regular pentagons do not. Many single shapes tile, such as the two irregular pentagons found in 1976 by Marjorie Rice, an amateur. Such tilings are periodic, repeating a period parallelogram in two directions. The thorn-shaped "versatile" of Grünbaum and Shephard (after Voderberg) tiles both periodically and in spirals.
- 1961 Hao Wang (square "Wang tiles" with coloured edges, no rotation) shows a decision procedure would exist if every set that tiles also tiles periodically.
- 1966 Robert Berger proves no decision procedure exists; hence an aperiodic set must exist; his has 20,426 tiles, later 104.
- 1971 Raphael Robinson reduces it to six tiles.
- c. 1973 Penrose independently finds another aperiodic set of six, then by slicing and re-gluing reduces to two (the kite and dart, and the two rhombs), with "seemingly crystallographically impossible" fivefold quasi-periodic symmetry.
- 1974 Hanf and Myers: a single (large) set that tiles the plane only non-computably.
- Open in 1989: is there an aperiodic set of one tile?
The Mandelbrot set as a model of non-recursive mathematics
Suppose M is non-recursive. Its largest parts are still algorithmic:
- Main cardioid A: points c = z − z2 with |z| < 1/2.
- Disc B to its left: c = z − 1 with |z| < 1/4 (an exact circle).
- Blobs C1, C2 (top and bottom) and the baby cardioid C3 far left: solutions of c3 + 2c2 + (1 − z)c + (1 − z)2 = 0 with |z| < 1/8.
If M is non-recursive, the parts our algorithms miss must be "very delicate and hard to find"; locate one and we can improve the algorithm, but others hide deeper still. Penrose thinks this is how mathematics proceeds in hard, presumably non-recursive areas: centuries-old procedures handle common cases, the escapees goad mathematicians to deeper insight, and each insight becomes a better algorithm. In the word problem, any pair known to escape a given algorithm is thereby known to be unequal and can be added to it: "Our improved insight will lead to an improved algorithm!" Perhaps physics is like this too.
Complexity theory
Computability is "in principle"; complexity theory asks how fast. It concerns families of problems, each instance of size n (the number of bits of its data, so up to 2n instances of a given size), and N(n), the worst-case number of steps.
- P (polynomial time): N ≤ K × nr for constants K, r; robust across machine models (one tape or many). Example: multiplying two n/2-bit numbers by schoolbook method takes about n2/2 steps (cleverness gives about n log n log log n). Considered tractable.
- Not in P: computing 22r needs about 2n steps just to write the answer.
- NP: if an instance has a solution, it can be checked in polynomial time (strictly, defined for yes/no problems).
- NP-complete: every NP problem converts to it in polynomial time, so a polynomial algorithm for one gives one for all. Examples: Hamiltonian circuit (a loop through every vertex of a graph exactly once; checking a proposed circuit is easy) and the yes/no travelling salesman problem (is there a tour shorter than a given length?).
- The open question: experts believe P ≠ NP; "no-one has been able to prove it", "the most important unsolved problem of complexity theory".
A graph is coded by listing pairs (1,2), (1,3), (2,3), (1,4), ... And writing 1 for an edge, 0 otherwise (10010110110...). For NP problems outside P, run times for large n soon exceed the age of the universe. Penrose notes that the security of new public-key codes depends on the hardness of factorizing large integers (an NP problem).
Complexity in physical things
Penrose suspects, "no doubt on quite inadequate grounds", that computability, not complexity, is what matters for mind, and that complexity theory barely touches practicality. But physical objects might obey a different complexity theory: David Deutsch (1985) described a quantum computer that could solve some problems outside P in polynomial time, though the problems were "decidedly artificial" and no reliable device was in sight. He closes with questions: does the brain exploit quantum magic in forming judgements; must we go beyond present quantum theory; can physical devices beat Turing machines in complexity, or even in computability? To answer, we must ask how the physical world actually behaves.
Try it in Python: words, invariants and Goodstein
# Is CATERPILLAR "equal" to MAN? (breadth-first)
from collections import deque
R = [('EAT','AT'), ('ATE','A'), ('LATER','LOW'),
('PAN','PILLOW'), ('CARP','ME')]
rules = R + [(b, a) for a, b in R]
def neighbours(w):
for a, b in rules:
i = w.find(a)
while i != -1:
yield w[:i] + b + w[i+len(a):]
i = w.find(a, i + 1)
def search(start, goal, maxlen=14):
seen, q = {start: None}, deque([start])
while q:
w = q.popleft()
if w == goal:
path = []
while w: path.append(w); w = seen[w]
return path[::-1]
for v in neighbours(w):
if v not in seen and len(v) <= maxlen:
seen[v] = w; q.append(v)
print(search('CATERPILLAR', 'MAN'))
inv = lambda w: sum(w.count(ch) for ch in 'AWM')
print(inv('CARPET'), inv('MEAT')) # 1 versus 2
The search reproduces Penrose's nine-word chain. But notice the asymmetry he stresses: the search finds equalities; for CARPET and MEAT it would simply run until the length cap, and it is the invariant (count of A, W and M), an idea about the rules, that proves inequality.
# Goodstein: alternate (a) raise base, (b) subtract 1
def bump(n, b): # hereditary base b -> b+1
total, power = 0, 0
while n:
n, d = divmod(n, b)
if d:
total += d * (b + 1)**bump(power, b)
power += 1
return total
def goodstein(n, steps):
b, seq = 2, [n]
while n and len(seq) < steps:
n = bump(n, b); seq.append(n) # (a)
n -= 1; b += 1; seq.append(n) # (b)
return seq
print(goodstein(4, 9)) # 4, 27, 26, 42, 41, 61, 60, 84, 83
Starting from 4 the sequence keeps climbing for an unimaginably long time before reaching zero; no computer will ever show the end, yet the theorem says it comes. Peano arithmetic cannot prove that; ordinal reasoning can.
Suppose neuroscience one day hands you a printout and says: "This formal system F is exactly the procedure your brain uses to decide arithmetical truths." Penrose's dilemma: if you believe F is sound, you can construct its Gödel sentence G(F) and see it is true, so F was not your whole procedure after all. If you cannot believe F is sound, then your own mathematical reasoning rests on rules you cannot trust, which "flies in the face of what mathematics is all about". Critics answer that you would simply be unable to verify the printout, and that ordinary mathematicians already live with not knowing whether their foundations are consistent. Which horn would you choose, and what would it cost?
A Gödel sentence you can almost hold
Gödel's construction can feel like sleight of hand. A concrete stand-in helps. Suppose a formal system F for arithmetic, with a mechanical checker for proofs. Gödel showed how to write, inside F's own language, an arithmetical statement G whose content, decoded, is: "There is no number that codes an F-proof of the statement with Gödel number g", where g is G's own number. No self-reference is smuggled in; the trick is a diagonal substitution, the same move as Cantor's and Turing's.
Now reason as Penrose asks. If F proved G, F would prove something false (because there would be such a proof, contradicting what G says), so if F is sound, F does not prove G. But that is exactly what G asserts. So G is true, and F cannot prove it. Anyone who accepts F's soundness can see G is true. This is the step Penrose calls insight.
Three features are worth fixing in memory:
- G is a plain statement about numbers. It says some polynomial-like equation has no solution in whole numbers (after Matiyasevich 1970, it can be a Diophantine equation). It is not a paradox.
- Seeing G is true needs F's soundness, which F itself cannot prove (Gödel's second theorem). The whole debate turns on how we know soundness.
- Adding G to F gives F′, which has its own G′. The ladder never ends, as Turing showed in 1939.
The critics' four main replies
The Gödel argument in this chapter and Chapter 10 drew more replies than any other part of the book. Penrose answered many of them in Shadows of the Mind (1994). The four strongest lines, with his responses, are summarized here so the reader can judge.
| Reply | The objection | Penrose's response |
|---|---|---|
| We might be inconsistent | Human mathematicians make errors, so we are not a sound system and Gödel does not apply | Errors are correctable; what matters is the idealized standard we recognize as correct, which must be sound |
| We cannot know our own algorithm | We may be a sound algorithm too complex for us to identify, so we cannot construct our own G | Then mathematical understanding would rest on an unknowable procedure, which he finds implausible and at odds with how proof is communicated |
| Truth is not what we "see" | We only believe G is true conditionally on soundness, as a machine could | Penrose holds that we do know the soundness of the systems we accept, through meaning |
| Gödel's own disjunction | Gödel concluded only that either the mind exceeds any machine or there are unsolvable problems we can never know | Penrose argues the second horn is unattractive and opts for the first |
Most logicians (Feferman, Putnam, Davis, Chalmers among them) judge that the argument does not establish its conclusion, while granting it shows something hard about how we know our own reliability. That consensus is recorded in the Critical Lens boxes; the table is here so the reader can see the shape of the exchange.
A bank adopts a rule engine that approves loans and also certifies its own rule set as compliant each quarter. A regulator, reading Chapter 4, observes that if the engine's rules are strong enough to describe arithmetic about themselves, the engine cannot prove its own consistency without being inconsistent. The certificate is either redundant or worthless. Self-certification is the institutional form of Gödel's second theorem, which is why serious assurance always sits outside the system it assures: an independent audit, a second model, a human sign-off. Penrose would add that the auditor is trusted because she understands the rules, not because she runs a bigger rule set.
Key terms from Chapter 4
| Term | Meaning in this book |
|---|---|
| Formal system | Axioms plus mechanical rules of inference, so proof-checking is an algorithm |
| Sound / consistent | Proves only true statements / proves no contradiction |
| Gödel sentence G | A true arithmetical statement the system cannot prove, if sound |
| Recursive set | Membership decidable by an algorithm |
| Recursively enumerable set | Members listable by an algorithm, though non-members may never be confirmed |
| Intuitionism | Brouwer's view that only constructed objects exist; rejects the excluded middle for infinite sets |
| Formalism | Hilbert's programme of treating mathematics as symbol games with consistency proofs |
| P, NP, NP-complete | Problems solvable in polynomial time; checkable in polynomial time; the hardest problems in NP |
Test yourself on Chapter 4
- What did Hilbert hope a consistency proof would do? Secure all of mathematics against paradoxes like Russell's, using only finite, uncontroversial reasoning. Gödel's second theorem showed this hope cannot be met for any system containing arithmetic.
- Why are the theorems of a formal system recursively enumerable but not, in general, recursive? You can list the theorems by running through all proofs, but you cannot always decide whether a given statement will ever appear.
- What does the tiling problem show? That a question a child understands, "do these shapes tile the plane?", has no general algorithm (Berger 1966), and that some shape sets tile only non-periodically, a phenomenon Penrose's own tiles made famous.
- Why might the Mandelbrot set be "non-recursive"? Because deciding whether a given point belongs needs infinitely many iterations in the worst case; the question is subtle and, in the bit model, still tied to the open hyperbolicity conjecture.
- How does complexity theory differ from computability? Computability asks whether an algorithm exists; complexity asks how long it takes. Penrose argues that neither speed nor parallelism touches his concern, which is with non-computability itself.
- Fermat proved (1994-95). Brouwer would say it became true in 1995; Penrose would say we found out. Twenty consecutive 7s in π: π has been computed to over 300 trillion digits (2025), long enough that strings of twenty identical digits are expected to occur, matching Penrose's probabilistic expectation.
- The einstein tile (2023). David Smith, Joseph Myers, Craig Kaplan and Chaim Goodman-Strauss found the "hat", a single 13-sided polykite that tiles only aperiodically, and then the "spectre", which does so without needing mirror images. Penrose's 1989 open question is answered yes.
- 3-manifolds are now classifiable. Perelman's proof of geometrization (2003) made homeomorphism of 3-manifolds decidable (Kuperberg 2019 among others), so the line between decidable and undecidable falls exactly at dimension 4.
- P vs NP is still open; it is a Clay Millennium Prize problem (2000). Shor's algorithm (1994) factors integers in polynomial time on a quantum computer, the realistic successor to Deutsch's "artificial" examples; it threatens RSA, and NIST standardized post-quantum cryptography in 2024. NP-complete problems are not believed solvable efficiently even by quantum computers.
- Mandelbrot computability: within the Blum-Shub-Smale real-number model, M is undecidable (Blum and Smale, 1993); in the bit model, computability of M is equivalent to the open hyperbolicity conjecture (Hertling 2005).
- Machines and Gödel: AI systems (AlphaProof, 2024 silver-medal standard; Gemini Deep Think, official gold-medal standard at the 2025 IMO) now find formal or rigorous proofs. They work inside formal systems or by learned heuristics; the Gödel point, that no sound system proves its own Gödel sentence, is untouched, but the claim that mathematical insight is out of machine reach is under real empirical pressure.
Worked detail: why checking is easy and finding is hard
Penrose's account of NP turns on a contrast between finding a solution and checking one. A small travelling-salesman instance makes the contrast concrete.
- The problem. Five cities, with known road distances. Is there a tour visiting each city once and returning home, of total length under 100 km?
- Checking. Someone hands you a tour: A, C, E, B, D, A. You add five distances and compare with 100. The work grows only in proportion to the number of cities.
- Finding. Without a proposed tour, the obvious method tries every order. With the start fixed and direction ignored, five cities give 12 tours; ten cities give 181,440; thirty cities give about 4 × 1030, more than any computer could examine.
NP is the class of problems where a good answer, once found, is quick to check. The P versus NP question asks whether every such problem also has a quick method for finding answers. Clever algorithms (dynamic programming, branch and bound) beat brute force enormously, but all known exact methods for the general problem still take exponential time in the worst case. Penrose's point is that this is a question about speed: every one of these problems is fully computable, so complexity theory, however deep, cannot by itself supply the non-computability his argument needs.
Common misreadings of Chapter 4
- "Gödel showed mathematics is inconsistent or unreliable." He showed that consistent systems are incomplete, and that consistency cannot be proved from within. Mathematics is no less reliable for it.
- "Gödel sentences are paradoxes like 'this sentence is false'." A Gödel sentence is a true statement about numbers, typically that some equation has no solution. The liar paradox has no truth value; G has one.
- "Penrose says humans can solve every problem." He says only that, for any sound system we accept, we can see its Gödel sentence is true. He does not claim we can settle the continuum hypothesis or Goldbach's conjecture.
- "The tiling problem is undecidable, so we can never know whether a set of tiles tiles the plane." Each specific set may be settled; it is the general procedure that cannot exist.
- "P versus NP is about whether some problems are uncomputable." It is about efficiency: every NP problem is computable. Penrose discusses complexity to show it is not the issue for his argument.
- The central inference is contested. That we can "see" Pk(k) is true depends on knowing the system is sound (or consistent). Critics (Putnam 1995, Feferman 1995, Chalmers, Benacerraf 1967, and earlier against Lucas 1961) reply that we cannot know our own reasoning's formal system is consistent; Gödel then yields only "if I am consistent, this is true", which a machine can also derive. Penrose's fuller reply in Shadows of the Mind (1994) concerns "unassailable" mathematical belief; most logicians remain unpersuaded.
- Gödel himself drew only a disjunction: either the mind is not a machine, or there are absolutely undecidable arithmetical problems.
- Diophantos lived around the third century AD, not BC.
- Factoring is in NP but not known to be NP-complete; Penrose's sentence is correct but invites the misreading that factoring is as hard as the travelling salesman problem.
- The 1 + √2 escape radius is valid but unusually generous; the standard bound for z → z2 + c with |c| ≤ 2 is radius 2.
- Generation is easier than verification of absence. Theorems are r.e., non-theorems are not; "this agent never violates policy" is a non-halting claim, provable only by stepping outside the agent with stronger assumptions. Build guardrails as external, independently specified checks, not as the model's self-report.
- Proof checkers versus proof finders. The architecture of AlphaProof-style systems is exactly Penrose's split: a cheap algorithmic checker (Lean) plus an expensive search guided by learned "judgement". Verification by a small trusted kernel is the template for trustworthy AI outputs in regulated domains.
- P vs NP in practice. Many bank optimization problems (portfolio construction with integer constraints, collateral allocation, scheduling) are NP-hard; tractability comes from heuristics and problem structure, not from waiting for faster hardware, Penrose's point about the age of the universe.
The Classical World
Before claiming that physics lacks something, Penrose surveys what it has. He grades theories as SUPERB, USEFUL or TENTATIVE, then tours the superb classical ones: Euclid, Galileo-Newton, Hamilton, Maxwell, special and general relativity. Throughout he separates determinism from computability: a world can be fully determined yet non-computable, and chaotic unpredictability is not the useful kind of non-computability. His verdict: nothing in classical physics offers a non-computable ingredient a brain could harness; if mind uses one, it lies outside classical physics. Along the way classical physics itself proves stranger than expected: Liouville's theorem means classical solids could not hold their shape, Dirac's electron needs foreknowledge, and gravitational energy can sit in perfectly empty space.
The status of physical theory
Most physicists and physiologists think the physics relevant to an ordinary-sized object like a brain is already known; the brain is complicated, but its principles are not mysterious. Penrose will argue the reverse: we do not yet understand physics well enough to describe brain function even in principle. First he needs an inventory. "Classical" means the theories before quantum mechanics arrived around 1925 (Planck, Einstein, Bohr, Heisenberg, Schrödinger, de Broglie, Born, Jordan, Pauli, Dirac); classical theory is deterministic, quantum theory brings "uncertainty, indeterminism, and mystery". He already signals that quantum effects, contrary to "a majority view among physiologists", are likely to matter in the brain.
The decisive event was Newton's Principia (1687), built on Galileo, Descartes and Kepler and, further back, the geometry of Plato, Eudoxos, Euclid, Archimedes and Apollonios. Then came Maxwell's electromagnetism (mid-nineteenth century), and the two relativities. Special relativity's technological impact is "very slight" outside nuclear physics, and general relativity's "almost non-existent", yet it is general relativity that will matter most for Penrose's argument (Chapters 7 and 8).
SUPERB, USEFUL, TENTATIVE
| Category | Criterion | Members (Penrose, 1989) |
|---|---|---|
| SUPERB | Phenomenal range and accuracy (not necessarily unrefuted) | Euclidean geometry; Archimedes' statics; Galileo-Newton dynamics (planets to better than 1 part in 107); Maxwell's theory (valid over a range of about 1036, from sub-atomic to galactic); special relativity; general relativity (binary pulsar: about 1 part in 1014); quantum mechanics; QED (electron magnetic moment: theory 1.00115965246, experiment 1.001159652193, "New York to Los Angeles to within the width of a human hair", in Feynman's image) |
| USEFUL | Good experimental support, but untidy and short of phenomenal precision | Gell-Mann-Zweig quark model and QCD; electroweak theory (Glashow, Salam, Ward, Weinberg; together the standard model); the standard big-bang model |
| TENTATIVE | No significant experimental support | Kaluza-Klein theories, supersymmetry/supergravity, (super)string theory, GUTs, the inflationary scenario; and, by his own footnote, twistor theory insofar as it is a new physical theory |
| (MISGUIDED) | A fourth category he declined to name formally | "since I do not want to lose half of my friends!" |
- Nothing outside physics qualifies as SUPERB. Darwin and Wallace's natural selection "comes closest, but it is still some good way off".
- Thermodynamics is left unclassified on purpose: it concerns averages and is partly deduced from other theories. He expects a union of thermodynamics with big-bang cosmology (Chapter 7) to deserve SUPERB.
- Historical USEFUL theories. The Ptolemaic system served for about twenty centuries, then faded (its originator was Eudoxos, developed by Hipparchos and Ptolemy). Kepler's ellipses and Mendeleev's periodic table were USEFUL theories later derived inside SUPERB ones.
- Why the SUPERB theories are old: forgotten TENTATIVE theories vastly outnumber survivors. Penrose finds it "an extraordinary remarkable fact that there are any theories in this category at all".
A footnote observes that all established departures from Newton concern light: Maxwell's energy-carrying fields, the role of c in special relativity, general-relativistic effects that matter only when speeds approach c (light bending, Mercury, black holes), the wave-particle duality first seen in light, and QED, the quantum theory of light and charges. Penrose speculates Newton himself suspected deep problems lay hidden in light's behaviour.
Euclidean geometry as a physical theory
Euclidean geometry is mathematics, but its application to our space is an empirical fact, and the clearest sign it is a physical theory is that it is not exactly true: general relativity says space-time curves near mass. It is still SUPERB: over a metre, the error in treating space as Euclidean is less than the diameter of a hydrogen atom.
- Lobachevskian (hyperbolic) geometry (Lobachevsky 1792-1856, independently Gauss, Schweickard, Bolyai): triangle angles sum to less than 180°, the deficit proportional to the area. It might even describe the cosmos on the largest scales.
- Kant claimed an a priori intuitive grasp that Euclid must hold; Einstein showed it is an observational fact.
- Plato (c. 360 BC, some fifty years before Euclid's Elements): perfect geometrical objects live in an ideal world accessible through the intellect; physical things only approximate them. Penrose credits Plato with foreseeing both that mathematics must be studied for its own sake and that the world can be understood only through precise mathematics.
Eudoxos invents the real numbers
The Pythagoreans' discovery that √2 is not a fraction threw Greek geometry into crisis. Eudoxos (c. 408-355 BC), of Plato's Academy, "a much finer scientist" than Aristotle in Penrose's view, gave an integer test for comparing ratios of lengths:
- a/b > c/d if there exist integers M, N with Na > Mb while Md > Nc (in modern terms, a fraction M/N lies between them).
- a/b < c/d by the mirror criterion.
- a/b = c/d exactly when neither criterion holds. This is essentially the Dedekind-Weierstrass construction of the reals, twenty-two centuries early. It is Book V of Euclid's Elements (c. 300 BC), and it underlies Einstein's theory as much as Euclid's.
The difference: the Greeks took real numbers as given by physical space (ratios of magnitudes, needed for theorems such as Ptolemy's theorem on four points on a circle, which uses both sums and products), whereas moderns treat numbers as logically prior, building many geometries from coordinates (Fermat and Descartes) of which physics selects one. Decimals came only with Stevin (1585).
- Archimedes (287-212 BC), "the greatest mathematician of antiquity", computed areas and volumes of spheres, parabolic and spiral regions with impeccable rigour: "a good half, the 'integral' half, of the calculus" nineteen centuries early. He also founded statics (lever, floating bodies) as a deductive science.
- Apollonios (c. 262-200 BC): the theory of conic sections, precisely what Kepler and Newton needed for orbits.
The dynamics of Galileo and Newton
The Greeks understood statics but not motion, partly for lack of a good clock. Galileo's observation in 1583 that a pendulum keeps time (though he often used a water clock) and his Discorsi (1638) launched dynamics.
- Force determines acceleration, not velocity. Velocity is the rate of change of position; acceleration is the rate of change of that rate. Aristotle thought force maintained velocity. With no force, motion is uniform and straight (Newton's first law).
- Galilean relativity. Uniform motion is locally indistinguishable from rest. Galileo's ship: below decks, with flies, butterflies, fish in a bowl and a dripping bottle, nothing changes when the ship moves steadily. This rescued Copernicus (1473-1543) and Aristarchos (c. 310-230 BC): we do not feel the Earth's roughly 100,000 km/h orbital motion. (Rotation does have detectable effects, such as the deflection of winds; Galileo wrongly thought it caused tides.)
- Conservation of energy, in embryo: a body released from rest has a speed depending only on height fallen, whether dropped, swung on a pendulum or slid down a smooth slope, and just enough to climb back. Later sharpened by Descartes, Huygens, Leibniz, Euler and Kelvin; combined with relativity it yields conservation of mass and momentum (rockets, gun recoil) and of angular momentum (Earth's spin, skaters drawing in their arms).
- All bodies fall at the same rate without air resistance (the Leaning Tower story), the seed of general relativity.
Galilean space-time. If "at rest" has no meaning, neither does "the same point of space at two times": we need a fresh Euclidean space for each moment, glued into a four-dimensional space-time, with uniformly moving particles as straight world-lines.
Newton's cathedral. First and second laws (force = mass × acceleration = rate of change of momentum) were essentially Galileo's; the third law, equal and opposite action and reaction, was Newton's own. Forces add by the parallelogram law, act along the line joining particles, and for gravity are attractive, proportional to the product of masses and to the inverse square of distance. Kepler (1571-1630) had found elliptical orbits with the Sun at a focus and two further laws; Newton derived all three, plus corrections and the precession of the equinoxes.
Penrose's footnote: (1) the laws for particles also hold for whole bodies, so planets can be treated as particles; he guesses this is why Newton insisted on the third law. (2) The inverse square law is the only power law decreasing with distance whose orbits are simple closed curves: "What would Kepler have done if the law of force had been an inverse distance law or an inverse cube?" Science depended on nature being legible at human scale.
The mechanistic Newtonian world
Given a force law, positions and velocities at one time fix everything thereafter. Model particles as points (with strong short-range repulsion to avoid infinite forces) or as perfectly elastic rigid balls. With balls, exact triple collisions are indeterminate (whether A meets B first or C does) and make outcomes depend discontinuously on the input, so we rule them out as "infinitely improbable". The billiard-ball picture is, Penrose thinks, "the model of reality that a good many people have at the backs of their minds" (its strongest promoter was R. G. Boscovich, 1711-1787, not Newton, whose views were subtler).
In such a world there seems "no room for a 'mind' to be influencing the behaviour of material things by the action of its 'free will'". But Penrose's key distinction appears here: the world "might be deterministic but non-computable", its future fixed by the present yet incalculable. If conscious action is non-algorithmic, free will must be tied to some non-computable ingredient in the laws.
A deterministic, non-computable toy universe
The state is a pair of natural numbers (m, n). At the next instant, if the universal machine Tu(m) halts, the state becomes (m + 1, n); if it does not, (n + 1, m). The evolution is completely deterministic, yet predicting it requires solving the halting problem, so no algorithm can compute the future. (Raphael Sorkin pointed out a limiting sense in which it can be approximated, by deeming non-halting after N steps; Penrose replies that rules involving doubly quantified statements, like the twin prime conjecture, defeat that.)
Is the billiard-ball world computable?
- A plausible undecidable question: given initial positions and velocities, does ball A ever hit ball B? Fredkin and Toffoli (1982) built a billiard-ball computer in which balls bouncing among fixed walls perform all logic operations and can imitate any Turing machine; "halting" could be "A eventually strikes B". So the Newtonian question is probably undecidable, at least with rational initial data.
- But that is the wrong sense. A device that answers "no" only by waiting forever to see that A never meets B behaves just like a Turing machine. The relevant question is whether a brain could harness non-computable laws to beat Turing machines.
- In the useful sense, the Newtonian world is computable: put positions on a grid (hundredths of a unit), step time by Δt, update velocities by the force law, and refine grid and time-step to push accuracy as far into the future as desired.
- Chaos is "unpredictability", not non-computability. Tiny changes in initial data (anyone who has tried a two-ball pot in billiards knows) blow up; long-range weather forecasting fails although the equations are known. This merely injects randomness; it is something to minimize, not harness, in a thinking machine.
Hamiltonian mechanics
The SUPERB theories are also extraordinarily fertile mathematically, a "deep and beautiful mystery" (Chapter 10). Newton's yielded the calculus and then classical mechanics through Euler, Lagrange, Laplace, Liouville, Poisson, Jacobi, Ostrogradski and Hamilton. William Rowan Hamilton (1805-1865), also of Hamiltonian circuits, formulated it with a wave analogy that later shaped quantum mechanics.
Treat positions xi and momenta pi as independent variables (3n of each for n particles). Everything follows from the Hamiltonian H (total energy as a function of all x's and p's): - dpi/dt = −∂H/∂xi (Newton's second law: rate of change of momentum = force) - dxi/dt = ∂H/∂pi (what the momenta are: velocity = momentum ÷ mass) They are first-order equations. The coordinates may be angles (with angular momenta) or any generalized coordinates, and the same form covers Maxwell-Lorentz theory, special relativity, carefully general relativity, and is the take-off point for Schrödinger's equation. (Lagrange knew the equations about 24 years before Hamilton; the Euler-Lagrange equations derive Newton's laws from Maupertuis's principle of stationary action.)
Phase space
Picture one dimension for every coordinate x1, x2, ..., p1, p2, ...: for n particles, 6n dimensions; for a room of air, something like 1028. The trick is not to try to visualize it. A single point Q is the entire state of the system; Hamilton's equations attach an arrow to every point, a vector field; the history of the whole system is Q following the arrows. Determinism is obvious.
Computability in phase space. If H's constants (Newton's G, c, pure numbers) are computable, Penrose guesses the usual Hamiltonians carry computable points to computable points, "merely a guess" he hopes will be examined. Yet non-computability of a point needs all its decimals, and measurement gives only finitely many. A useful non-computable device would need something better than ever-finer analogue measurement: discrete, stable states (like Turing's 0s and 1s), each a region of phase space that stays put and lets small errors die away.
Liouville's theorem and the spreading problem
Phase-space volume is conserved under any Hamiltonian evolution: region R0 dragged along the vector field becomes Rt with exactly the same volume. It seems to guarantee stability, but it does the opposite: volume is kept while shape is not, so a compact region stretches into fine filaments wound through a vastly larger region, like a drop of ink spreading through a tank of water. (Fully covering phase space is the extreme case, ergodic behaviour.) In high dimensions there are more directions to spread in, and the theorem forbids compensating by shrinkage; footnotes add a whole family of conserved Poincaré invariants, though degrees of freedom "dumped" as radiation to infinity can reduce the volume of interest.
- Why Newtonian prediction works at all. Celestial mechanics deals with few, mass-segregated, coherent bodies that behave as particles. Elsewhere Newtonian theory is used for models and statistics: conservation laws, thermodynamics, Newton's speed-of-sound calculation (corrected by Laplace a century later). Its determinism is "very rare[ly]... Actually used".
- A startling corollary. Phase-space spreading implies, "somewhat overstating", that classical mechanics cannot be true of our world: classical point particles could not maintain the organized structure of solids. Quantum theory is needed to stop the spreading.
- Computers need quantum mechanics. A computer's discrete states are phase-space regions that must not spread; even the Fredkin-Toffoli machine needs solid walls, and solidity requires quantum effects. "Even a 'classical' computing machine must borrow from the effects of quantum physics."
Maxwell's electromagnetic theory
Electric and magnetic forces (studied by William Gilbert, 1600, and Benjamin Franklin, 1752) are inverse-square like gravity, with charge in place of mass and like repelling like, and fit Newton's framework, even with corrections for moving charges (Gauss and others). The challenge came from Michael Faraday (1791-1867), who concluded from experiments with coils and magnets that fields are real physical stuff, not bookkeeping for forces, and that changing electric and magnetic fields might push each other through empty space as waves, perhaps light.
- c. 1820s-1830s Ampère (1775-1836) and Faraday: experiments relating currents, magnets and changing fields.
- 1860s James Clerk Maxwell (1831-1879) adds the term ∂E/∂t (the displacement current) for theoretical, mathematical and aesthetic reasons; its coefficient 1/c2 was too small to have been observed. Fields now sustain each other as waves moving at exactly the speed of light, with interference and polarization.
- Visible light: wavelengths 4-7 × 10-7 m; other wavelengths predicted from currents in wires.
- 1888 Heinrich Hertz detects electromagnetic waves; "radio waves can actually carry energy", so disembodied fields are as real as particles.
Maxwell's equations (modern form): ∇·E = ρ/ε0 (charges source the electric field, the inverse-square law in coded form), ∇·B = 0 (no magnetic monopoles; magnets are dipoles), ∇×E = −∂B/∂t (Faraday), ∇×B = μ0j + (1/c2)∂E/∂t (Ampère plus Maxwell's term). Like Hamilton's equations they give rates of change from present values, so they are deterministic; but as field equations they need infinitely many parameters, giving an infinite-dimensional phase space.
Computability and the wave equation
In empty space every field component obeys the wave equation {(1/c2)(∂/∂t)2 − (∂/∂x)2 − (∂/∂y)2 − (∂/∂z)2}φ = 0. Marian Pour-El and Ian Richards (1979-82) proved that computable initial data can evolve to a non-computable value at a computable later time. Startling, but the data must be not twice differentiable, unlike physically sensible fields; with smooth data non-computability cannot arise, and exploiting it would need infinitely precise measurement. Penrose calls it "an intriguing beginning".
Runaway particles
Maxwell's equations say how fields respond to charges, not how charges move. Lorentz (1895) supplied the equations of motion for a charge in a field. Trouble: the charge's own field dominates at its location. For a finite particle (classical electron radius about 10-15 m), the field varies across it and raises questions of rotation, distortion and elasticity; for a point particle the field is infinite.
Dirac showed how to subtract the self-field, at a price: the initial data must include the particle's acceleration as well as position and velocity, and for almost all values the particle spontaneously accelerates off toward the speed of light. To exclude these runaways one must choose initial accelerations with foreknowledge of which solutions would misbehave. The present is constrained by a requirement that the future "behave itself": a teleological element in classical physics. Physicists usually reply that point charges belong to QED; Penrose notes QED has its own troubles here.
Special relativity (Einstein and Poincaré)
Combine Newton with Maxwell and Galilean relativity fails: a fast traveller should see light slower ahead and faster behind, contrary to Maxwell's fixed c. Poincaré (1898-1905) and Einstein (1905) independently found that Maxwell's equations obey their own relativity principle, incompatible with Galileo's. Einstein, sure some relativity principle must hold and noting that only light had been tested at high speed, kept Maxwell's and modified Newton's mechanics.
- Lorentz (by 1895) held that the forces binding matter are electromagnetic; moving bodies would then contract in their direction of motion (the FitzGerald-Lorentz contraction), explaining Michelson and Morley's 1887 failure to detect the Earth's motion through the ether.
- Poincaré (1905) showed matter could behave exactly so that uniform motion is undetectable and understood the relativity of simultaneity, but treated it as one possibility; Einstein insisted on it.
- Minkowski (1908), once Einstein's teacher at the Zürich Polytechnic, completed the picture in a famous Göttingen lecture: space and time "by themselves" fade into shadows; only their union is real.
- Events are points; a particle is a world-line; the light cone at each event holds the histories of light flashes from it (45° lines when space is measured as x/c). Material world-lines stay inside light cones.
- "Distance": Euclid has r2 = x2 + y2 + z2; Minkowski has s2 = t2 − (x/c)2 − (y/c)2 − (z/c)2.
- s is the time a clock actually experiences along a straight world-line. s < t whenever there is motion (moving clocks run slow); on the light cone s = 0, so a photon experiences no passage of time.
- Twin "paradox": the stay-at-home's time AC exceeds the traveller's AB + BC, a reversed triangle inequality. In general, the unaccelerated world-line between two events is the longest in experienced time.
- Poincaré motions (inhomogeneous Lorentz transformations) are the "rigid motions" of this geometry: picture a sheet that stretches along one 45° direction while squashing along the other. Physics is unchanged under them.
- Fast unstable particles (cosmic rays, accelerators) decay late by exactly the predicted factor.
- Atomic clocks flown in fast, low aeroplanes register s, not t (with small general-relativistic height corrections that also agree).
- E = mc2, the equivalence of mass and energy, is verified continually.
- GPS (a footnote) locates positions to within a few feet only because general-relativistic curvature effects are included.
No universal "now"
Each observer has simultaneous spaces, and those of a moving observer are tipped up relative to a stationary one's. For spacelike-separated events (each outside the other's light cone), observers disagree on order: S says R precedes Q, M says Q precedes R.
Einstein's general relativity
Galileo's insight that all bodies fall alike rests on a coincidence Newton merely postulated: the mass in the gravitational force law equals the mass in F = ma. For electric forces the analogue fails (charge is not inertial mass). Einstein made the coincidence foundational.
- A camera falling with one of Galileo's rocks sees the other hovering; an astronaut on a space walk sees the shuttle hovering. Orbiting is "a sophisticated (and very expensive!) way of 'falling'".
- Principle of equivalence: locally, gravity is indistinguishable from acceleration. In a lift whose cable snaps, occupants float.
- What cannot be removed: the tidal effect. A freely falling sphere of particles near Earth stretches toward and away from Earth's centre and narrows sideways, becoming an ellipsoid (the Moon raises ocean tides the same way, with bulges toward and away from it). Tidal strength falls off as the inverse cube of distance.
For a falling sphere surrounding vacuum, the ellipsoid it initially deforms into has the same volume as the original sphere: this volume property holds for the inverse square law and no other. If the sphere surrounds mass M, the volume shrinks in proportion to M. Einstein keeps these rules (for small spheres), adding a tiny contribution from pressure.
Curved space-time. Treat free-fall world-lines as the natural "straight lines": geodesics. On a positively curved surface (sphere) initially parallel geodesics converge; on a negatively curved saddle they diverge. Tidal effects show both, in different directions: space-time is curved. (Élie Cartan showed in 1923 that even Newton's theory can be rewritten this way.) New physics enters when this is merged with Minkowski: space-time is locally Minkowskian (a Lorentzian manifold, after Riemann's locally Euclidean spaces), every point has a light cone but the cones are no longer uniformly arranged, and geodesics maximize experienced time; light rays are geodesics of zero length. The Earth's world-line is a corkscrew geodesic around the Sun's; starlight bends past the Sun.
- The Riemann curvature tensor Rijkl needs 20 numbers at each point.
- It splits into WEYL (10 components), measuring tidal shape distortion, and RICCI (10 components), measuring volume change.
- Einstein's field equations, schematically RICCI = ENERGY, equate Ricci curvature with the energy-momentum tensor (energy, pressure, momentum of matter and fields). (Hilbert found the correct equations in November 1915 too, but the physical ideas were Einstein's alone.)
- WEYL is the gravitational field proper: in empty space all tidal effects are Weyl; through the Bianchi identities it obeys Maxwell-like equations with ENERGY as source. Weyl curvature will become the protagonist of Chapter 7.
- Gravitational time dilation: clocks run slow in gravitational fields, measured many ways.
- Light and radio signals are bent and delayed by the Sun.
- Mercury's perihelion advance, troubling astronomers since 1859, explained by Einstein in 1915.
- The binary pulsar, two neutron stars, agrees to about 1 part in 1014 and shows orbital decay from gravitational waves, an effect absent from Newton. "No confirmed observations exist that contradict Einstein's general relativity."
Relativistic causality and determinism
Observer W sends a slightly superluminal signal from event A to distant event B. To an observer U moving away fast, B is earlier than A. By symmetry a third observer V, with an identical device, can send the message back the same way, reaching W at event C before A on W's own world-line. W receives the message before sending it. If it says he broke his leg, free will could prevent the accident. Worse, with machines programmed "send YES if you receive NO" and "echo what you receive", the contradiction needs no free will at all. (Wheeler and Feynman, 1945, offered "not very satisfying" ways round.) Every signal must respect the light cones.
Determinism in special relativity improves on Newton: to predict an event P you need data only on a bounded region of an initial simultaneous space (whatever light can reach P from), whereas Newtonian forces act instantly and require the entire infinite slice. In general relativity data go on a spacelike (hyper)surface; Einstein's equations are locally deterministic, but the light-cone structure is itself being determined, so we cannot know in advance which data matter, and in extreme cases even the whole surface may not suffice: global determinism can fail, tied to the unsolved cosmic censorship problem about black holes. Penrose doubts such failures touch human-scale matters.
Where do we stand on computability?
Where classical determinism fails (Dirac's runaways, general-relativistic extremes), computability fails too, but these failures are at scales irrelevant to brains: point-particle electrodynamics is superseded by QED, black holes are vastly larger. Where determinism holds, behaviour seems computably determined (rigorous theorems, he notes, are lacking). Chaos appears even in general relativity (Misner 1969; Belinskii, Khalatnikov and Lifshitz 1970), but it is unhelpful. Hence: if the mind uses non-computable elements, "they must be elements lying outside classical physics".
Mass, matter, and reality
Classical physics gives a stage (space-time) and actors (particles and fields). Particles are ambivalent: as test particles they are points following the Lorentz law or geodesics; as sources they must be spread out, or their fields become infinite. And the stage joins the action, since space-time's structure is gravity.
- Newton: mass, conserved.
- Einstein: E = mc2, so mass includes energy. When uranium splits, the fragments' rest masses fall short of the original, but adding their kinetic energy divided by c2 restores the total. Yet kinetic energy depends on the observer (sit in an express train and its energy of motion vanishes).
- The π0-meson decays in about 10-16 s into two photons, each carrying half its mass as "pure energy"; a fast observer can make either photon's energy as small as desired, since photon rest-mass is zero.
- The answer: the energy-momentum four-vector, an arrow in Minkowski space whose height is mass-energy and whose spatial part is momentum. Its Minkowskian length is the rest-mass, which is not additive (π0 has rest-mass, the photons none), while the arrows add by the parallelogram law.
- Fields are matter. Maxwell's field carries energy, hence mass, and contributes to every body's mass (classically the electron's self-energy comes out infinite).
Gravitational waves carry energy, yet in a pure gravitational wave the tensor ENERGY is zero everywhere; the curvature is all Weyl. Gravitational energy is non-local. The positive Bondi-Sachs mass (1960, 1962) can be non-zero in a perfectly flat, empty region between two bursts of radiation, "like the calm within the eye of a hurricane". Our "quantity of matter" is "either there, in the emptiest of empty regions, or it is nowhere at all!" Classical material reality is already "much more nebulous" than we thought, before quantum theory adds "much greater shocks".
Try it in Python: Liouville's spreading
# Area is kept, shape is destroyed (standard map)
import numpy as np, matplotlib.pyplot as plt
K = 1.2
p, q = np.meshgrid(np.linspace(0, .05, 120),
np.linspace(3, 3.05, 120))
p, q = p.ravel(), q.ravel()
fig, ax = plt.subplots(1, 4, figsize=(12, 3))
for i, steps in enumerate((0, 3, 8, 20)):
P, Q = p.copy(), q.copy()
for _ in range(steps):
P = (P + K*np.sin(Q)) % (2*np.pi)
Q = (Q + P) % (2*np.pi)
ax[i].plot(Q, P, ',', color='crimson')
ax[i].set_title(f'{steps} steps')
plt.show()
The Chirikov standard map is an area-preserving (Hamiltonian-like) evolution. A tiny square of initial states keeps exactly its area, as Liouville demands, yet within twenty steps it is smeared into filaments across the whole plane: Penrose's ink drop. Set K = 0.2 and the square merely shears: low K is the orderly "celestial mechanics" regime.
Build Fredkin and Toffoli's billiard-ball computer so that ball A strikes ball B if and only if a chosen Turing machine halts. In an ideal Newtonian world the table "knows" the answer to an undecidable question; all you must do is watch. Penrose's two objections tell you why this is no oracle: (1) you learn "never" only by waiting for ever, exactly as with a Turing machine; (2) the slightest error in a ball's position grows exponentially, so a real table answers a slightly different question within a few collisions. Determinism and even "non-computable" dynamics are useless to a mind unless the non-computability is robust and readable at finite precision, the property he will demand of the missing physics.
Chaos is not non-computability
A reader meeting the standard map or the three-body problem for the first time often concludes that chaotic physics is already non-computable. Penrose is careful to say otherwise, and the distinction is central to his case.
- Chaos means sensitive dependence on initial conditions: two nearby starting states diverge exponentially. Prediction needs ever more digits of the starting data, but each digit's consequences are computable. Given the initial state to n digits, a Turing machine can compute the motion to a corresponding accuracy for a corresponding time.
- Non-computability in Penrose's sense means no algorithm can produce the evolution even in principle, however precise the input. It would be a property of the law, not of our ignorance of the data.
Chaos makes physics hard to predict; it does not make physics impossible to simulate. That is why Penrose turns away from classical mechanics at the end of this chapter. The equations of Newton, Maxwell and Einstein are, as far as anyone knows, computable in the sense that matters (Pour-El and Richards found a contrived exception for the wave equation with non-smooth initial data, which Penrose discusses and sets aside as unphysical). His non-computable ingredient must come from somewhere else, and he looks for it in the gap between quantum mechanics and gravity.
| Kind of difficulty | Example | Does it help Penrose? |
|---|---|---|
| Too many particles | A gas of 1023 molecules | No: complexity, not non-computability |
| Chaos | Weather, the standard map, three bodies | No: computable to any finite accuracy |
| Undecidable long-term questions | Moore's 1990 particle in a potential | Only weakly: needs exact initial data |
| Wave equation with rough data | Pour-El and Richards (1981) | No: data not physically realistic |
| A non-computable law | Penrose's hoped-for reduction rule | Yes, if it exists |
Penrose's grading of theories, and why it matters
The SUPERB, USEFUL, TENTATIVE and MISGUIDED scale earlier in this chapter is idiosyncratic, as the Critical Lens notes, but its purpose is clear. Penrose wants the reader to feel the gap between theories that are accurate beyond anything their discoverers could have demanded (Newton's gravity, Maxwell's fields, relativity, quantum mechanics) and the far larger body of models that merely fit. The SUPERB theories are his evidence that the physical world is mathematical in a deep way, and that is also why he trusts that the missing theory of quantum gravity, when found, will be equally precise and equally surprising.
Run a perfect Newtonian simulation of two billiard balls colliding, then reverse every velocity: the balls retrace their paths exactly. Now do the same with ten thousand balls in a box, starting from a neat triangle. After a few seconds of mixing, reverse the velocities and, in theory, the triangle reassembles. In a computer with 64-bit floats, it does not: tiny rounding errors grow chaotically and the reversal fails within a few dozen collisions. The equations are reversible; our access to them is not. This is the bridge to Chapter 7, where Penrose argues that the arrow of time is a fact about initial conditions, not about the laws, and to his claim that phase-space volume (Liouville) never shrinks even as it spreads into fine filaments.
Key terms from Chapter 5
| Term | Meaning in this book |
|---|---|
| Phase space | The space of all positions and momenta of a system; one point is a complete state |
| Liouville's theorem | Hamiltonian evolution preserves phase-space volume |
| Hamiltonian | The total energy written as a function of positions and momenta, generating the motion |
| Lorentz transformation | The rule relating measurements in uniformly moving frames in special relativity |
| Light cone | The boundary of possible causal influence from an event |
| Weyl and Ricci curvature | The tidal (shape-changing) and volume-changing parts of space-time curvature |
| Determinism / computability | Whether the future is fixed by the present / whether it can be calculated by an algorithm |
Test yourself on Chapter 5
- Why does Penrose treat Euclidean geometry as a physical theory? Because whether real space obeys Euclid's axioms is an empirical question; general relativity says it does not, and Lobachevskian geometry is a consistent alternative.
- What does Liouville's theorem say, and why does it matter later? Phase-space volume is conserved under Hamiltonian evolution. Combined with coarse-graining, it explains why entropy rises: a small region spreads into a fine-grained tangle filling a large coarse-grained volume.
- Why did Penrose think the billiard-ball computer was a useful example? It shows that idealized classical physics can carry out any computation, so classical physics is at least Turing-powerful; the open question is whether it is ever more.
- What is the Andromeda paradox? Two walkers passing each other on Earth disagree, through relativity of simultaneity, by days about which events on Andromeda are "now". Penrose uses it to question any naive "flow" of time.
- Why does Penrose conclude that classical physics is not where the mind's secret lies? Because the classical theories are, as far as is known, deterministic and computable for realistic data, and because the classical world is in any case only an approximation to a quantum one.
Try it in Python: Liouville's theorem and coarse-grained spreading
Penrose's picture of phase space (the figure of a compact region spreading into fine filaments) can be watched directly with Arnold's cat map, a chaotic map of the unit square that, like Hamiltonian flow, preserves area exactly.
import numpy as np
rng = np.random.default_rng(0)
x = 0.02*rng.random(5000); y = 0.02*rng.random(5000)
def occupied(x, y, n=20): # coarse-grain into 20 x 20 boxes
H, _, _ = np.histogram2d(x, y, bins=n, range=[[0, 1], [0, 1]])
return int((H > 0).sum())
for step in range(9):
print(step, occupied(x, y))
x, y = (2*x + y) % 1.0, (x + y) % 1.0 # area-preserving
# boxes occupied: 1, 2, 5, 14, 35, 90, 226, 388, 400
The fine-grained area of the cloud never changes, yet within eight steps it reaches every coarse box. Volume is conserved; the coarse-grained volume, and hence entropy, is not. This is the bridge from Chapter 5 to the second law of Chapter 7. Run the map backwards (x, y) → (x − y, 2y − x) from step 8 and the cloud reassembles, but only because the computer stores every point exactly; perturb the points by one part in a million and the reversal fails.
- Gravitational waves detected directly. LIGO's GW150914 (14 September 2015), two merging black holes; Nobel Prize 2017 (Weiss, Barish, Thorne). Hundreds of events by the end of the O4 run (2023-25). The Event Horizon Telescope imaged M87* (2019) and Sgr A* (2022). The double pulsar now tests GR to about 0.01%.
- Equivalence principle: the MICROSCOPE satellite (final results 2022) confirmed equal fall to about 1 part in 1015.
- QED's SUPERB status sharpened: the electron's magnetic moment measured to 0.13 parts per trillion (2023); the comparison now hinges on rival determinations of the fine-structure constant (caesium versus rubidium) that disagree slightly.
- Penrose's guess about billiard balls was right in spirit: physical systems that are Turing complete have since been built in theory, including a single particle in a smooth 3D potential (Moore 1990) and stationary Euler fluid flows (Cardona, Miranda, Peralta-Salas, Torres de Lizaur, 2021), so some long-term trajectory questions are undecidable. As he anticipated, this is the "unhelpful" kind: it cannot be harnessed to beat a Turing machine.
- The USEFUL and TENTATIVE lists: the Higgs boson (2012) completed the standard model; supersymmetry and extra dimensions have not appeared at the LHC; string theory still lacks experimental support; inflation remains debated (the 2014 BICEP2 "detection" of primordial gravitational waves was retracted as dust).
Common misreadings of Chapter 5
- "Penrose's SUPERB list is a ranking of truth." It ranks theories by range and accuracy relative to their original evidence. All, he expects, will be superseded.
- "Relativity says everything is relative." It says the speed of light and the space-time interval are absolute, and that simultaneity is relative. Penrose's Minkowskian geometry makes the invariants the centre of the picture.
- "Determinism implies predictability." Not in chaotic systems, and not necessarily in principle either: a deterministic system could in theory evolve in a non-computable way, which is exactly the possibility Penrose is exploring.
- "Liouville's theorem means nothing ever becomes disordered." It conserves fine-grained volume; the region becomes stringy and spread out, so its coarse-grained volume, and hence entropy, grows.
- "General relativity is just Newtonian gravity with corrections." It is a theory of curved space-time, in which gravity is not a force at all. Penrose's tidal picture, with Weyl and Ricci curvature, shows what replaces force.
- The SUPERB grading is idiosyncratic. Many physicists would rate the standard model (now including the Higgs) and ΛCDM cosmology higher than "USEFUL", and thermodynamics as foundational rather than unclassifiable.
- "Classical mechanics cannot be true" from Liouville spreading is, as Penrose half-admits, overstated: dissipation, coarse-graining and boundary conditions let classical models describe stable structures; the real classical failure for solids is atomic stability (radiating electrons), which Penrose does not stress here.
- Credit for special relativity. Penrose's emphasis on Poincaré as co-discoverer is defensible but contested; Poincaré never abandoned the ether and kept a privileged frame in his interpretation.
- Dirac runaways are now usually removed by the Landau-Lifshitz reduced-order equation, which avoids pre-acceleration for practical purposes; the deep point (point charges are ill-defined classically) stands.
- Apollonios died around 190 BC rather than 200 BC (minor).
- Determinism is not predictability, and predictability is not computability. Penrose's three-way distinction maps onto model risk: a deterministic credit model can still be chaotic in its inputs (sensitivity testing), and some questions about a model's long-run behaviour are undecidable in general (hence scenario analysis rather than proof).
- Liouville for data pipelines: information about an initial state is conserved but smeared across ever more variables; this is why reconstructing provenance after the fact is hard and why lineage must be captured at the source.
Quantum Magic and Quantum Mystery
The longest chapter builds quantum theory from the ground up and takes the quantum state as objectively real. Its central structural claim is that quantum mechanics runs on two incompatible rules: U, the smooth, deterministic, superposition-preserving Schrödinger evolution, and R, the abrupt, probabilistic "reduction" applied whenever an effect is magnified to the classical level. Penrose develops amplitudes, the wavefunction, Hilbert space, spin and the Riemann sphere, the no-cloning of states, identical particles, the EPR paradox and Bell's inequalities, then Schrödinger's cat and a critique of every standard interpretation. His conclusion: quantum theory is SUPERB but incomplete, because nothing tells us when U gives way to R; the missing theory of that transition is where he will look for non-computability.
Do philosophers need quantum theory?
Classical physics supplies an objective, deterministic world, with our brains inside it obeying the same equations; most "serious" philosophy of mind tacitly assumes this picture. Quantum theory is often reduced in popular understanding to a vague "uncertainty principle", but Penrose stresses two surprises:
- Quantum descriptions are very precise, just radically unlike classical ones.
- Probabilities do not arise at the tiny scale. Particles, atoms and molecules evolve deterministically; randomness enters "via some mysterious larger-scale action connected with the emergence of a classical world that we can consciously perceive".
Quantum effects are not small corrections: solidity, material strength, chemistry, colours, freezing and boiling, the reliability of heredity all need quantum theory. Perhaps consciousness, too, cannot be understood in classical terms, and perhaps that is "why" sentient beings must inhabit a quantum world. Penrose admits his personal view that quantum theory itself is "a stop-gap".
Bohr's line (Copenhagen)
Nothing is "out there" at the quantum level; reality emerges only relative to measurements; the theory is a calculational procedure.
Penrose's line
"Too defeatist." The quantum state describes objective physical reality; the puzzle is the odd relation between that evolving state and the observed jumps. Consciousness is not what triggers a measurement: "only a small minority" of physicists would say so, and observers are themselves quantum.
Problems with classical theory
Quantum theory was "not wished upon us by theorists"; it was forced, reluctantly, by experiment. The deep classical inconsistency is the coexistence of particles (six parameters each) and fields (infinitely many). At equilibrium, equipartition of energy spreads energy evenly over all degrees of freedom, so infinitely many field modes drain all energy from the particles.
- Unstable atoms. In Rutherford's 1911 "solar system" atom, an orbiting electron should radiate with intensity rising "rapidly to infinity" and spiral into the nucleus in a tiny fraction of a second. Real atoms are stable and emit only at sharp, discrete spectral frequencies obeying "crazy" numerical rules.
- Black-body radiation. Rayleigh and Jeans (1900) predicted energy pouring into ever-higher frequencies without limit (the ultraviolet catastrophe). Observed spectra match them at low frequency but peak and fall to zero at high frequency, with the peak colour set by temperature (red-hot poker, yellow-white Sun).
Making everything a field fails too: field "knots" could wriggle continuously, yet all electrons are identical and atoms take only discrete configurations. Some new ingredient must make fields discrete.
The beginnings of quantum theory
- 1900 Max Planck, "brilliant but conservative and cautious", proposes that oscillations come in quanta with E = hν; h ≈ 6.6 × 10-34 J s. The Planck radiation law fits the data.
- 1905 Einstein: the electromagnetic field itself exists only in such units, so light is particles after all (as Newton insisted, against Thomas Young's early-nineteenth-century wave evidence).
- 1913 Niels Bohr: angular momentum of atomic orbits comes only in multiples of ℏ = h/2π (0, ℏ, 2ℏ, 3ℏ, ...), reproducing many spectral rules: the provisional "old quantum theory".
- 1923 Louis de Broglie's thesis (Einstein was consulted for approval): matter particles are waves, with hν = E = mc2. Particles and field-oscillations are "the same thing": anything oscillating at frequency ν comes in units of mass hν/c2.
- 1925-26 Heisenberg's matrix mechanics and Schrödinger's wave mechanics, soon proved equivalent and unified by Dirac.
The two-slit experiment
Light from a sodium lamp (wavelength about 5 × 10-7 m) passes through slits about 0.001 mm wide and 0.15 mm apart onto a screen a metre away. Light always arrives as whole photons, never half a photon (a 60-watt bulb emits about 1020 photons per second, which is why it looks continuous). With one slit open, single spots build up a smooth spread. With both open, the spots build interference bands about 3 mm wide.
At the brightest points the intensity is four times the one-slit value, not twice; at the dark points it is zero, places a photon could happily reach with only one slit open. By slit separation the photon is about 300 "photon-sizes" from the other slit, yet "knows" whether it is open, and there is no limit to the distance. The effect survives with one photon at a time. It is not that light is sometimes wave, sometimes particle: each individual particle behaves like a wave entirely on its own, and alternative possibilities open to it can cancel each other out.
Putting a detector at one slit always finds a whole photon and destroys the pattern: interference needs "lack of knowledge" of the route. Alternatives combine through complex weightings, which go beyond adding and subtracting, as in "alternative A plus i × alternative B" (at points of intermediate brightness). Complex numbers "forced themselves on the attentions of physicists through persuasive and unexpected experimental facts".
Probability amplitudes
Penrose adopts a working split between a quantum level (complex superpositions) and a classical level (ordinary probabilities); not a matter of physical size, for quantum effects can extend over metres or light-years, but closer to very tiny energy differences. In principle cricket balls and elephants should superpose; we never see it, and why is the deep question he defers.
- Classical: p × A + q × B with probabilities p + q = 1; along a route, probabilities multiply; over alternative routes they add: P(s,p) = P(s,t)P(t,p) + P(s,b)P(b,p).
- Quantum: the same rules for complex amplitudes: A(s,p) = A(s,t)A(t,p) + A(s,b)A(b,p).
- The squared-modulus rule: when magnified to the classical level, an amplitude z = x + iy gives probability |z|2 = x2 + y2. Normalized amplitudes lie in the unit circle; for two alternatives |w|2 + |z|2 = 1. An amplitude is "a complex square root of a probability".
- Multiplication is harmless: |zw|2 = |z|2|w|2, "a pretty little miracle", so single-route reasoning works classically.
- Addition carries the quantum magic: |w + z|2 = |w|2 + |z|2 + 2|w||z|cos θ (the cosine rule), θ the angle between w and z in the Argand plane.
| Relation of amplitudes | cos θ | Result | Two-slit meaning |
|---|---|---|---|
| w = z | 1 | 4∣w∣2 | Bright band: four times one slit |
| w = −z | −1 | 0 | Dark band: destructive interference |
| w = ±iz | 0 | 2∣w∣2 | Intermediate: the classical "just add" result |
| Large, complicated systems | averages to 0 | Ordinary probabilities | Interference washes out |
Crucially, with both slits open |A(s,t)A(t,p)|2 is not the probability that the photon "went through" t; that would be nonsense at a dark point. Only if we detect at t does |A(s,t)|2 become a probability, and then the pattern vanishes.
The quantum state of a particle
The collection of complex weightings over all possible positions is the wavefunction ψ(x): the amplitude for the particle to be at x. Penrose takes the physical reality of the particle's location to be its quantum state ψ. Picture it along a line as the ψ-curve: at each x plot ψ(x) in an Argand plane perpendicular to the axis; the curve winds around the x-axis, and |ψ(x)|2 (squared distance from the axis) is the probability of finding the particle at x if a detector is placed there. For full 3D space the picture needs five dimensions.
- Economy: ψ already contains the momentum amplitudes, extracted by harmonic analysis (like splitting a sound into pure tones).
- Momentum states are corkscrews (helices): tightly wound means short wavelength, high momentum, high energy (E = hν); straight means zero momentum; right-handed screws point along +x.
- Fourier transforms (after Joseph Fourier, 1768-1830): the momentum-space wavefunction ψ̃(p) and ψ(x) are transforms of each other, symmetrically.
- Position states are Dirac delta functions (infinitely sharp spikes). A delta in position space is a corkscrew in momentum space, and vice versa.
- Which picture to use depends on what is magnified: photo-cells and plates measure position; recoil and crystal diffraction measure momentum.
In the two-slit set-up each slit becomes a new source; along tp and bp run corkscrews of the same pitch, and at screen point p their phases agree (path difference a whole number of wavelengths: bright), oppose (half-wavelength extra: dark) or sit in between.
The uncertainty principle
Δx Δp ≥ ℏ. Locate an electron to about a nanometre and its momentum becomes so uncertain that one second later it could be 100 km away. Penrose rejects three common glosses: that it is mere measurement clumsiness ("somewhat misleading"), that the particle has intrinsically random motion ("certainly wrong"), and that position and momentum are inapplicable concepts ("unduly pessimistic"). The truth is in the wavefunction: a corkscrew (definite momentum) has equal |ψ|2 everywhere, so position is totally uncertain; a delta (definite position) has equal momentum amplitudes. A wave packet, localized in both pictures, is quantum theory's best approximation to a classical particle, and it spreads with time, the faster the tighter its start.
U and R: the two procedures
U: unitary evolution
Schrödinger's equation; fully deterministic once ψ is given; preserves superpositions; continuous; resembles Hamilton's and Maxwell's equations; each momentum component moves at speed c2 ÷ (classical speed). "The part... Of main concern to working physicists."
R: state-vector reduction
Applied whenever an effect is magnified to the classical level ("measurement", "collapse of the wavefunction"); probabilistic (squared moduli); violates superposition; discontinuous. "All the non-determinism of the theory comes from R and not from U." Penrose takes it as something real, not a mere update of knowledge.
Standard theory offers no derivation of R from U, and no rule for when to switch. A momentum state evolves predictably for ever, measured in momentum it gives a certain answer, measured in position it gives total uncertainty. What constitutes "making a measurement"? Why and when do squared moduli "become probabilities"? Can the classical level be understood quantum-mechanically? These questions drive the rest of the book.
Particles in two places at once?
If ψ is real, a particle in a momentum state is spread over all space, and a photon emerging from two slits has a two-peaked state ψ = ψt + ψb: it "is" in two places at once. The objection that a detector always finds it at one slit only shows that a position measurement was made. Other measurements could distinguish ψt + ψb from ψt − ψb or ψt + iψb, so these are physically different ways for the photon to be in two places.
A single photon meets a half-silvered mirror at 45°: its wavefunction splits into a transmitted and a reflected peak whose separation grows without limit. Wait a year and the peaks are over a light-year apart. Bring them back with full mirrors to a second half-silvered mirror with detectors A and B. If it were merely "50% one way, 50% the other", each detector would fire half the time. With equal path lengths, A fires with 100% probability and B never; block either path and A and B become equally likely. Blocking a route allows B to be reached, so the photon "must have actually felt out both routes." Such experiments have been done over many metres (Wheeler 1983). Bohr's refusal to assign the photon any reality between measurements is "much too pessimistic".
This is quantum linear superposition: any two states, however different, can be combined with complex weights, and the formalism draws no line between one particle and a cricket ball or a person. Why we never see macroscopic bodies in two places is "a profound question" current theory does not answer; invoking the classical level and switching from U to R "begs the controversial question".
Hilbert space
The quantum analogue of phase space: one point (up to scale) is the state of an entire system. Hilbert space is a complex vector space: states can be added with complex weights. Dirac's notation writes states as kets |ψ〉, |↑〉, |0〉, |1〉, ...
- Rays, not vectors. |ψ〉, 2|ψ〉, −|ψ〉, i|ψ〉, (1 − 3i)|ψ〉 are the same physical state; only the zero vector has no meaning. A state is a whole (complex) line through the origin.
- Dimensions are independent states, not spatial directions. A single particle's position needs infinitely many dimensions, one per possible position.
- Orthogonality means independence: all position states are mutually orthogonal, as are all momentum states, but positions are not orthogonal to momenta. Changing from position to momentum axes is a rotation in Hilbert space.
Measurements
- Complete measurement: choose an orthonormal basis |0〉, |1〉, |2〉, ... Write |ψ〉 = z0|0〉 + z1|1〉 + z2|2〉 + ...; the components are the orthogonal projections of |ψ〉 on the axes. The state jumps to one axis at random with probability |zi|2 (divided by |ψ|2 if not normalized). "There is no dynamical law to tell us which... Nature will choose."
- Yes/no measurement (the more fundamental kind): Hilbert space splits into orthogonal complements Y and N; |ψ〉 = |ψY〉 + |ψN〉; the state jumps to |ψY〉 or |ψN〉 with probabilities |ψY|2, |ψN|2, summing to 1 by Pythagoras: the probability equals the factor by which the squared length shrinks under projection.
- Every state is testable: for any state |χ〉 there is in principle a yes/no measurement answering YES with certainty exactly for |χ〉 (and NO for states orthogonal to it). This strongly suggests state vectors are objectively real.
Spin and the Riemann sphere of states
Spin is the intrinsic angular momentum of a single particle, but unlike a cricket ball's it has a fixed magnitude for each species; only the axis varies, strangely. Electrons, protons and neutrons have spin ℏ/2, half Bohr's smallest unit, impossible for orbiting parts and so intrinsic.
- Fermions (spin ℏ/2, 3ℏ/2, ...): a 360° rotation sends the state to minus itself.
- Bosons (spin 0, ℏ, 2ℏ, ...): a 360° rotation returns the state to itself.
For a spin-½ particle at rest, Hilbert space is two-dimensional, basis |↑〉, |↓〉 (right-handed spin about up and down). Every spin state is w|↑〉 + z|↓〉, and every direction is equally good as a basis: |→〉 = |↑〉 + |↓〉, |←〉 = |↑〉 − |↓〉. Orthogonal Hilbert vectors here correspond to opposite directions in space, not right angles.
- Only the ratio q = z/w matters (with q = ∞ when w = 0). Plot q in a horizontal Argand plane through the centre of a unit sphere, and stereographically project from the south pole (labelled ∞) onto the sphere: circles map to circles and angles are preserved.
- The spin direction of w|↑〉 + z|↓〉 is the direction from the centre to the point q. North pole q = 0 is |↑〉; south pole q = ∞ is |↓〉; q = 1 is |→〉; q = −1 is |←〉; q = i (back) is |↑〉 + i|↓〉 pointing away from us; q = −i points toward us.
- Measurement statistics: having found spin along direction α, a measurement along β gives YES with probability ½(1 + cos θ) and NO with ½(1 − cos θ), θ the angle between them: certain if β = α, 50/50 at right angles, certainly NO if opposite.
- Every two-state system has a Riemann sphere of states, even when, as with a photon's two routes (|ψt〉 at the north pole, |ψb〉 at the south, the equal mixtures around the equator), it has no spatial meaning.
Objectivity versus measurability
The view that the state vector is just "our knowledge", or describes only an ensemble of similar systems, strikes Penrose as "unreasonably timid". Suppose an electron's spin is |α〉 for an unknown direction α. A measurement along any chosen β extracts only one bit, YES or NO, and destroys the information about α. Yet the electron must answer YES with certainty if we happen to measure along α, so the information "is stored in the electron's state of spin".
The state vector is not measurable (no experiment on one system reveals it) but it is objective, being completely characterized by the results it must give to the experiments one might perform. For a lone spin this is mild: the spin points somewhere definite, we just do not know where. For pairs of particles it becomes much stranger. Often an electron has no state of its own, being entangled with many others; it has one only when, for instance, its spin was measured earlier and it has stayed undisturbed.
Copying a quantum state
If an unknown spin state |α〉 could be copied while the original survived, one could copy it again and again, building a system with a huge angular momentum pointing along α, then read α off with a macroscopic measurement. That would violate the non-measurability of |α〉. Hence a quantum state cannot be duplicated without destroying the original (Wootters and Zurek 1982). Swapping is fine: an electron's unknown state |α〉 and a neutron's |γ〉 can be exchanged.
This bears on the teleportation paradox of Chapter 1. If a person's awareness depends on some aspect of a quantum state, quantum theory would forbid copying it without destroying the original, dissolving the puzzle of two viable copies.
Photon spin (polarization)
A photon moves at light speed, so its spin axis is always its direction of motion; photon spin is polarization. Two pieces of polaroid show it: aligned, the second removes almost nothing; crossed at right angles, almost all light is blocked; at angle φ, a fraction cos2φ passes. In particle terms each polaroid is a yes/no polarization measurement (here Hilbert-space orthogonality really is a right angle in space).
- States: plane (electric field oscillating in one plane, magnetic at right angles), circular (fields rotating), elliptical (in between).
- Riemann sphere again: for a photon moving upward, the north pole is |R〉 (right-handed, electric vector rotating anticlockwise viewed from above), the south pole |L〉; the general state |R〉 + q|L〉 is the point q. To get the ellipse, take p = √q, cut the sphere by the plane through the centre perpendicular to the direction of p, and project the circle vertically (the vector to √q is the Stokes vector).
- Probabilities: apply ½(1 + cos θ) to the q-points. Plane polarizers at angle φ correspond to p-points φ apart, hence q-points θ = 2φ apart, giving ½(1 + cos 2φ) = cos2φ, recovering the classical law.
Objects with large spin
A massive particle of spin n × ℏ/2 is an (n + 1)-state system; measured along any axis it gives n, n − 2, ..., −n (for n = 2: 2, 0, −2; for n = 3: 3, 1, −1, −3). A massless particle such as the photon always has just two states.
By a construction due to Majorana (1932) (reworked by Penrose, 1987), every spin-nℏ/2 state is an unordered set of n points on the Riemann sphere, as though made of n spin-½ pieces pointing in those directions. Only states with the points bunched together resemble a cricket ball's spin. A large-spin state chosen "at random" is "nothing whatever like" a classical one. Classical behaviour appears only because such states are superpositions of classical-like states, and "somehow a 'measurement' gets itself performed" that picks one. The same holds for every classical property: the correspondence principle needs R.
Many-particle systems and entanglement
A two-particle wavefunction is not two functions of position but one function of two positions. With only ten allowed positions:
| System (10 positions) | Complex numbers needed | Why |
|---|---|---|
| One particle | 10 | one amplitude per position |
| Two distinguishable particles | 100 | one per ordered pair |
| Three distinguishable particles | 1,000 | one per ordered triple |
| Two identical fermions | 45 = 10 × 9 / 2 | antisymmetric: |
| Three identical fermions | 120 = 10 × 9 × 8 / 6 | six-term antisymmetric combinations |
| Two identical bosons | 55 = 10 × 11 / 2 | symmetric: |
| Three identical bosons | 220 = 10 × 11 × 12 / 6 | symmetric combinations |
| Spin-½ fermions (20 one-particle states) | 190 for two, 1,140 for three | each position doubled by ↑ and ↓ |
- Products and superpositions. |α〉|β〉 means "first system in |α〉 and second in |β〉" (the tensor product). The general state is a sum like |α〉|β〉 + |ρ〉|σ〉, which is not a product: Penrose insists on the difference between "and" and "plus", and scolds insurance brochures that blur them. Such non-product states are correlated (entangled).
- Fermions (electrons, protons, neutrons) change sign under interchange, |ψ〉 → −|ψ〉; bosons (photons) do not. Two fermions in the same state would give −|ψ〉 = |ψ〉, so |ψ〉 = 0: Pauli's exclusion principle (1925), without which matter would collapse in on itself. (Fermi-Dirac statistics came in 1926, Bose-Einstein in 1924.)
- No individual identity. "The first electron is here and the second there" is the forbidden product |0〉|1〉; one may only say "there is a pair of electrons, one here and one there". Individual-particle language is an approximation that fails in superconductivity, superfluidity and lasers. Swap an electron in your body with one in a brick and the state merely changes sign, which a 360° rotation undoes.
The Einstein-Podolsky-Rosen "paradox"
Einstein introduced the photon (1905) and helped invent bosons, but never accepted quantum theory as final. His 1926 letter to Max Born: an inner voice says the theory "is not yet the real thing" and that the Old One "does not play dice". What troubled him even more was lack of objectivity: Bohr made the classical world real but the quantum states beneath it unreal. Einstein's followers, notably David Bohm, developed hidden variables; the most successful is the de Broglie-Bohm model (Bohm 1952), consistent with experiment only if essentially non-local.
Take an entangled state |α〉|β〉 + |ρ〉|σ〉 of two distant systems. A yes/no measurement on A (|α〉 versus |ρ〉) makes B jump to |β〉 or |σ〉, though B may be light-years away. The obvious objection is the box with a white ball and a black ball: separate them, look at one, and the other is "instantly" known, with no mysterious influence. The point of EPR and Bell is that this "it was decided all along" picture cannot reproduce quantum predictions.
A spin-0 particle decays into an electron and a positron flying apart. Their total spin is zero: |Q〉 = |E↑〉|P↓〉 − |E↓〉|P↑〉. Rewritten in the right/left basis it is (up to a factor −2) |E→〉|P←〉 − |E←〉|P→〉, the same form. So whichever direction the experimenter chooses for the electron, the positron's spin instantly becomes opposite along that same direction, even light-years away.
Mermin's version of Bell's theorem
Each side has a spin-measurer with three settings A, B, C (E side) and A′, B′, C′ (P side, parallel), all in one plane at 120°. Quantum theory predicts: 1. Same settings: always disagree (one YES, the other NO). 2. Settings chosen at random: agree exactly half the time. (If E says YES, P's direction is at 60° or 120° to its state: agreement ¾ = ½(1 + cos 60°) or ¼ = ½(1 + cos 120°); averaging ⅓(0 + ¾ + ¾) = ½.)
A local realistic machine must carry pre-set answers for all three settings (probabilistic answers could not guarantee property 1). Pre-set YES, YES, YES / NO, NO, NO gives 9 disagreements, 0 agreements over the nine pairings; YES, YES, NO / NO, NO, YES (and every similar pattern) gives 5 disagreements to 4 agreements. No assignment reaches 50/50. Local realistic models are ruled out. (Footnote version with two settings each: agreement ½(1 + cos 135°) ≈ 0.146 three times would cap the fourth at 45%, but quantum theory gives just over 85%.)
Photon experiments: a problem for relativity?
Experiments use polarization of photon pairs (spin one, so angles halve). The most convincing in 1989 were Alain Aspect's (Paris, 1982-86), following Freedman and Clauser (1972) and the CHSH proposal (Clauser, Horne, Shimony and Holt, 1969). Results agree with quantum theory and contradict every local realistic model. Aspect's settings were switched while the photons were in flight, so any influence would have to travel faster than light. Detectors were inefficient, a loophole Penrose finds implausible: better detectors would hardly worsen the agreement.
- No signalling. The influences cannot carry messages (Ghirardi, Rimini and Weber, 1980); one learns the pair of alternative directions "instantly", but which one only via an ordinary signal.
- But a clash with the spirit of relativity. The two measurements are spacelike-separated; which came first depends on the observer's motion. An observer moving right says the right-hand measurement caused the left photon's jump; one moving left says the opposite. There is "an essential conflict between our space-time picture of physical reality... And special relativity", which quantum realists (Aharonov and Albert 1981) had not resolved.
Schrödinger's and Dirac's equations
Schrödinger's equation, iℏ ∂|ψ〉/∂t = H|ψ〉, governs U. H is the classical Hamiltonian with each momentum pi replaced by the operator −iℏ∂/∂qi, "genuine magic which works", with surprisingly little harm from the ambiguities. (Von Neumann named the two procedures process 1, R, and process 2, U; the Heisenberg picture moves the evolution into the observables.)
Schrödinger's equation is linear: if |ψ〉 and |φ〉 are solutions, so is w|ψ〉 + z|φ〉. A superposition of alternatives can never be "unsuperposed" by U. Something separate, R, must pick one. For a single photon, Schrödinger's equation is Maxwell's equations; for a single electron it becomes Dirac's equation (1928), one of the "Great Field Equations" with Maxwell's and Einstein's, whose solutions flip sign under 360° rotation.
Quantum field theory
Combining quantum mechanics with special relativity: particle numbers can change, every particle has an antiparticle (the photon is its own), pairs annihilate into energy and are created from it, and superpositions of different particle numbers are allowed. QED, the theory of electrons and photons, is supremely accurate but "rather untidy... And a not altogether consistent one": it first yields infinite answers removed by renormalization, and not all theories are renormalizable. Feynman's path integrals superpose entire space-time histories but add infinities of their own, handled by "mathematical tricks". Two limits matter most for Penrose: QFT reconciles relativity only with U, not R, and there is no consistent quantum field theory of gravity. (A footnote: QFT may offer some scope for non-computability, cf. Komar 1964.)
Schrödinger's cat
A perfectly sealed box holds a cat, a light source, a half-silvered mirror, a photo-cell and a phial of cyanide (Schrödinger's 1935 original used a radioactive atom). A reflected photon triggers the cell and kills the cat; a transmitted one hits the wall and spares it. - Inside observer (in protective clothing): R has happened; 50% dead, 50% alive, one or the other. - Outside observer treating the box as one quantum system under U: amplitude 1/√2 each, so the cat is in a superposition of dead and alive. - The twist: with a vast computer the outsider could calculate the exact state |ψ〉 and perform the measurement that tests "|ψ〉 or orthogonal to it". It answers YES with 100%, and the orthogonal state |dead〉 − |alive〉 with 0%, which is only possible if both alternatives coexist and interfere. Different mirror settings give |dead〉 − i|alive〉 and so on, each distinguishable in principle. So is the state vector "all in the mind" after all?
Various attitudes in existing quantum theory
| Attitude | Claim | Penrose's objection |
|---|---|---|
| Impossible measurements | Distinguishing | ψ〉 from its orthogonal complement is impossible, so drop one branch |
| Environment / decoherence | Entanglement with the environment makes superpositions practically indistinguishable from probability mixtures | At what stage is interference "officially" impossible? And even a real probability mixture does not become one outcome by U alone |
| Density matrices (von Neumann) | Mix classical probabilities and amplitudes | Useful bookkeeping; does not touch the measurement problem |
| Chaos in the initial state | Probabilities come from tiny uncertainties amplified, as in weather | U is linear: no chaos; superpositions persist for ever. A non-linear process is needed |
| Consciousness collapses (Wigner 1961) | Linearity fails for conscious (or living) beings | Grossly lopsided universe: superpositions resolve only in the rare corners where minds exist. Penrose accepts it "only with great reluctance", though he seeks a quantum role for mind |
| Participatory universe (Wheeler 1983) | Our existence conjures the past mutations that produced us | Circular; "distinctly worrisome... Barely credible" |
| Many worlds (Everett 1957; DeWitt, Graham) | R never happens; U always; the universe and observers split at every magnification | Not economy: why would a mind be aware of only one branch? Needs a theory of consciousness; claimed derivations of R "do not hold up" |
Where does all this leave us?
- At the quantum level (small energy differences), alternatives coexist in complex-weighted superpositions; the weights are probability amplitudes; the state evolves deterministically by U (Schrödinger).
- When differences are magnified to the classical level, superpositions stop persisting; squared moduli become probabilities; R selects one alternative. Only here does indeterminism enter.
- The quantum state may be taken as objective, but with several particles it is entangled; R on one particle affects all its correlates instantly; EPR/Aspect experiments show nature is non-local, and an objective R violates the spirit of relativity, though observations do not.
- Quantum theory is silent on when and why R occurs, and does not explain why the classical world looks classical: "'Most' quantum states do not at all resemble classical ones!"
Penrose's verdict: quantum mechanics may be "simply wrong when applied to macroscopic bodies"; U and R are excellent approximations to an undiscovered theory, which must be non-linear because R is. Linearity is elegant, but so was Newton's linear superposition of forces, and Einstein showed it to be an approximation, with greater elegance. Many of the founders (Einstein, Schrödinger, de Broglie, Dirac) also thought the theory provisional. Hidden variables face the non-locality barrier, so a much more radical change is needed. No experiment contradicts quantum theory, "unless, of course, one regards the evident absence of linearly superposed cricket balls as contrary evidence. In my own view... It actually is." Somewhere between atoms and cricket balls lies a new law, needed to understand how the quantum merges with the classical, and needed to understand minds. The clues, he promises, lie at the farthest reaches of space and the beginning of time.
Try it in Python: interference and Mermin's game
# Two slits: |w+z|^2 versus |w|^2 + |z|^2
import numpy as np, matplotlib.pyplot as plt
lam, d, L = 5e-7, 1.5e-4, 1.0 # Penrose's numbers
x = np.linspace(-0.01, 0.01, 2000)
r1 = np.hypot(L, x - d/2); r2 = np.hypot(L, x + d/2)
w = np.exp(2j*np.pi*r1/lam); z = np.exp(2j*np.pi*r2/lam)
plt.plot(x*1e3, abs(w + z)**2, label='both slits')
plt.plot(x*1e3, abs(w)**2 + abs(z)**2, label='"classical"')
plt.xlabel('screen position (mm)'); plt.legend(); plt.show()
The bright bands reach 4, the dark ones 0, and the classical sum sits flat at 2; the spacing comes out at 3.3 mm, Penrose's "about three millimetres".
# Mermin: can pre-set answers agree half the time?
from itertools import product
best = 0.0
for E in product('YN', repeat=3):
P = ['N' if e == 'Y' else 'Y' for e in E]
agree = sum(E[i] == P[j]
for i in range(3) for j in range(3))
best = max(best, agree / 9)
print('best local model:', round(best, 3)) # 0.444
print('quantum theory :', 0.5)
Every local "instruction set" that guarantees disagreement at equal settings agrees at most 4/9 of the time over random settings; quantum mechanics, and every experiment since Aspect, gives 1/2.
(Elitzur and Vaidman, 1993, a later example Penrose made famous in Shadows of the Mind.) Some bombs have a detector so sensitive that a single photon sets them off; others are duds. Put a candidate bomb's detector in one arm of Penrose's two-mirror interferometer, tuned so that with both arms open the photon always reaches detector A. A dud acts as an open arm: the photon goes to A. A live bomb acts as an observation in that arm: half the time it explodes, but a quarter of the time the photon reaches detector B, which is impossible with both routes open. A click at B certifies a live bomb that no photon ever touched. The photon "felt out" the route it did not take, exactly Penrose's argument for the reality of both routes.
The measurement problem in one picture
Chapter 6 can be compressed into a single conflict between two rules that every working physicist uses, usually without noticing they disagree.
- Rule U (unitary evolution). Between measurements the state evolves deterministically and linearly by Schrödinger's equation. Superpositions stay superpositions: if a detector would register "up" for one state and "down" for another, then for a superposition it evolves into a superposition of having registered both.
- Rule R (reduction). When a measurement is made, the state jumps to one of the alternatives, with probability given by the squared modulus of its amplitude (the Born rule). R is probabilistic, discontinuous and time-asymmetric.
The trouble is that a measuring device is made of atoms, which obey U. Apply U to the whole apparatus and you never get a single outcome. Either U is incomplete, or something about observers lies outside physics, or every outcome happens. These are the three families of interpretation Penrose surveys: objective collapse (U is incomplete), Copenhagen-style pragmatism (stop asking), and many-worlds (all outcomes occur). He chooses the first, and spends Chapter 8 trying to make it precise.
| Interpretation | What happens at measurement | Penrose's verdict (1989) |
|---|---|---|
| Copenhagen | R is applied when a classical record forms; the boundary is left vague | A useful practice, not an account of reality |
| Many-worlds (Everett) | No collapse; the observer branches | Leaves unexplained why we perceive one outcome with Born probabilities |
| Decoherence | Interference becomes unobservable through entanglement with the environment | Explains appearances, not the single outcome |
| Pilot wave (Bohm) | Particles have definite positions guided by ψ | Non-local and gives up the symmetry he values |
| Objective collapse | R is a real physical process triggered at some scale | His choice, with gravity as the trigger |
A physicist sets up a spin measurement inside a sealed, perfectly isolated room and leaves her friend there to read the result. Outside, Wigner applies U to the whole room: by his calculation, the friend is in a superposition of having seen up and having seen down. Inside, the friend is sure she saw one result. Who is right? Modern versions (Frauchiger and Renner, 2018; the "local friendliness" experiments of Bong et al., 2020) sharpen this into theorems showing that some plausible assumptions about observers must fail. Replace the friend with a quantum computer that records the result in a qubit, and Wigner is plainly right: the record can be un-done. Penrose's question is whether a conscious friend is different, and his answer is that what matters is not consciousness but mass: a big enough record collapses on its own.
Key terms from Chapter 6
| Term | Meaning in this book |
|---|---|
| Wavefunction ψ | The complex amplitude for each possible configuration |
| Superposition | A complex-weighted sum of states, itself a state |
| Unitary evolution U | Deterministic, linear Schrödinger evolution |
| State reduction R | The probabilistic jump at measurement |
| Hilbert space | The complex vector space of quantum states; rays are physical states |
| Riemann sphere | Penrose's picture of all spin-½ states as points on a sphere, each a spin direction |
| Entanglement | A many-particle state not expressible as a product of single-particle states |
| No-cloning | The theorem that an unknown quantum state cannot be copied |
Test yourself on Chapter 6
- What is the "extraordinary fact" of the two-slit experiment? Opening a second slit can make light stop arriving at a point where it arrived with one slit open, because complex amplitudes can cancel, while probabilities cannot.
- Why does Penrose regard the quantum state as objective but not measurable? A state is real (it determines outcomes), but a single measurement cannot reveal it; only one yes/no question about it is answered at a time.
- What does Bell's inequality establish? That no local hidden-variable theory reproduces quantum predictions for entangled pairs; experiments since Aspect (1982) and the loophole-free tests of 2015 favour quantum theory.
- Why does EPR trouble relativity in Penrose's view? R appears to act instantaneously across space-like separations, and different observers disagree about which measurement came first, yet no signal is sent.
- Why is linearity the reason R is needed? Because a linear law turns a superposed input into a superposed output, so no linear evolution alone can turn "both" into "one". Some non-linear ingredient, or an extra rule, must enter.
Try it in Python: spin probabilities from the Riemann sphere
Each spin-½ state is a point on the Riemann sphere, and the chance that a state prepared along one direction passes a test along another depends only on the angle θ between them.
import numpy as np
def spin_state(theta, phi): # point on the sphere -> spinor
return np.array([np.cos(theta/2), np.exp(1j*phi)*np.sin(theta/2)])
up = spin_state(0, 0)
for deg in [0, 60, 90, 120, 180]:
th = np.radians(deg)
prob = abs(np.vdot(spin_state(th, 0), up))**2
print(deg, round(prob, 3), round((1 + np.cos(th))/2, 3))
# 0 1.0 | 60 0.75 | 90 0.5 | 120 0.25 | 180 0.0
The two columns agree: P(YES) = ½(1 + cos θ), the corrected form flagged in this chapter's Critical Lens. Note the half-angles in the spinor: rotating the physical direction by 360° turns the spinor into its negative, the famous quantum sign that Penrose says only shows up in interference. Every direction on the sphere is a possible state, and every state is "spin up" along some direction, which is Penrose's argument that the quantum state is objective even though no single measurement reveals it.
A classical coin is heads or tails. A spin-½ particle is like a coin that, whichever axis you choose to look along, always shows either heads or tails, never anything in between. Prepare it heads along north. Look along east and you get heads half the time; along north-east, about 85% of the time (½(1 + cos 45°)). There is no hidden list saying in advance what it would show along every axis, as Bell's theorem later proves. The state is a direction, not a list of answers. Penrose's Riemann sphere makes this vivid: the only thing the particle "knows" is one point on the sphere.
- Bell tests closed the loopholes. Loophole-free experiments in 2015 (Delft, using electron spins 1.3 km apart; NIST; Vienna) closed the detection and locality loopholes together. Aspect, Clauser and Zeilinger shared the 2022 Nobel Prize in Physics. Entanglement has been distributed over 1,200 km by China's Micius satellite (2017).
- No-cloning became technology. It underpins quantum key distribution (BB84) and quantum teleportation, proposed by Bennett et al. (1993) and demonstrated from 1997, which, exactly as Penrose's remark about teleporting minds suggests, destroys the original state while recreating it elsewhere.
- Is the state real? The Pusey-Barrett-Rudolph theorem (2012) showed that, under a natural independence assumption, the quantum state cannot be mere knowledge of an underlying reality: support for Penrose's realism about ψ.
- Bigger superpositions. Interference of molecules of about 2,000 atoms (25,000 atomic mass units, 2019) and a 16-microgram mechanical crystal in a "cat state" (2023) push U toward the classical scale without any observed breakdown, constraining collapse models including Penrose's own (see Chapter 8).
- Decoherence is now well measured and explains why interference is unobservable at large scales; most physicists accept that it does not by itself yield a single outcome, which is Penrose's point. Many-worlds, QBism, Bohmian mechanics and objective-collapse theories (GRW 1986, CSL, Diósi-Penrose) all remain live.
- Quantum computing exploits exactly the large entangled superpositions described here; Google's Willow chip (2024) showed error rates falling as logical qubits grow, the first step to fault tolerance.
Common misreadings of Chapter 6
- "Quantum theory says reality depends on consciousness." Penrose rejects this. His R is an objective physical process, and in Chapter 8 he proposes that gravity, not observers, triggers it.
- "Superposition means the particle is in two places but we do not know which." That would be an ordinary probability mixture. A superposition has complex amplitudes whose phases produce interference, which no mixture can reproduce.
- "EPR correlations allow faster-than-light messages." They do not. The outcomes are random at each end; only after comparing records does the correlation appear. Penrose's concern is with reality, not signalling.
- "The uncertainty principle is about clumsy measurements." It is a property of quantum states: no state has sharp position and momentum together, measured or not.
- "Decoherence solves the measurement problem." It explains why interference becomes unobservable; it does not explain why one outcome occurs. Penrose's point stands in modern discussions.
- Realism about ψ is one option among several. Bohr's view is less "defeatist" in modern operational and information-theoretic programmes (QBism, relational quantum mechanics), which do not deny an external world, only that ψ is its literal description.
- Many-worlds critique dated. Everettians since the 1990s (Deutsch, Saunders, Wallace) answer "why one branch?" through decoherence-defined branching and derive Born-rule weights decision-theoretically; Penrose's objection remains common but is no longer unanswered.
- "Most quantum states are not classical" is true of Hilbert space generically but physically realized states are constrained by dynamics and environment; decoherence explains the preferred classical "pointer" basis that Penrose treats as unexplained.
- The Born rule for spin: the extraction lost a factor; the correct probabilities are ½(1 ± cos θ), as given above.
- Superposition is not "both answers at once" in a useful sense; it is an amplitude structure that only interference can exploit. Quantum advantage needs algorithms that make wrong answers cancel, which is why only special problems (factoring, simulation, some search) benefit.
- No-cloning is a security primitive. Post-quantum migration in banking protects classical keys against quantum attackers; quantum key distribution uses no-cloning to detect eavesdroppers. Both stem from the same Hilbert-space geometry Penrose describes.
- "And" versus "plus": Penrose's warning about insurance brochures is a good discipline for data modelling, where conjunction (joint attributes) and alternatives (mutually exclusive states) are routinely confused.
Cosmology and the Arrow of Time
Our experience of time flowing from a fixed past into an open future finds no echo in physics' time-symmetric equations, and relativity even denies a universal "now". The asymmetry comes from the second law of thermodynamics, and Penrose traces it to its source: our low entropy comes from the Sun, the Sun's from gravitational clumping of diffuse gas, and that gas from an astonishingly smooth big bang. Because gravity makes clumped states high-entropy, a generic big bang would have been as chaotic as a big crunch. The universe's initial state was special to one part in 1010123, encoded in his Weyl Curvature Hypothesis: Weyl curvature vanishes at initial singularities but not final ones. Only a time-asymmetric quantum gravity could explain it.
The flow of time
We feel the past is fixed and "out there", the future a set of potentialities, with the present continuously converting one into the other, sometimes by our own choices. Physics says otherwise.
- Time-symmetric laws. Newton, Hamilton, Maxwell, Einstein, Dirac and Schrödinger all work equally well with t replaced by −t: classical mechanics and the U part of quantum theory are reversible. Whether R is reversible will be central to Chapter 8; the "conventional wisdom" (Aharonov, Bergmann and Lebowitz, 1964) says it is.
- No "now". In relativity, observers in relative motion disagree on simultaneity: for observer U, event B may be fixed past while A is open future; for V the reverse.
Return to the Andromeda fleet of Chapter 5: for one passer-by its launch is fixed, for the other undecided. If anything is definite, "the entire space-time must indeed be definite"; there is no flowing time, only space-time, which seems to have been Einstein's own conclusion. (A footnote adds wryly that both walkers would be long dead before telescopes showed the fleet, so it would be their descendants who "hark back".) Penrose takes the clash between experienced time and physical theory as evidence that the physics underlying consciousness must contain an essentially time-asymmetric ingredient.
The inexorable increase of entropy
A glass of water nudged off a table shatters, the water soaking into carpet and cracks in the floorboards. Reverse every atomic motion and Newton's laws equally allow the water to flow back, the shards to assemble, and the full glass to leap onto the table. Energy is no obstacle: the first law of thermodynamics (energy conservation including heat) is time-symmetric, and the slight warming of fragments, water, carpet and floor holds exactly the energy needed to lift the glass back.
What forbids it is coordination. The reversed motion would need absurdly precise cooperation among atoms, an "amazing fluke". Yet in the forward direction such coordination is routine: after the crash, the atoms' motions are precisely those that, reversed, would rebuild the glass. Coordinated motion is acceptable as the effect of a large-scale change, never as its cause.
Imagine a world obeying the same symmetric laws in which, alongside ordinary events, glasses sometimes reassemble and leap onto tables, scrambled eggs unscramble and reseal in their shells, and sugar lumps re-form and jump into someone's hand. There we would call the future macroscopic state the cause ("Look! That mess is going to assemble itself into another glass of water!"). We could not fix that world by relabelling past and future. Our world is not like that, and the question is why coordinated motions follow large changes and never precede them. The careful thinker must avoid smuggling our everyday "cause precedes effect" into the physics.
Entropy measures manifest disorder. The precise correlations among water molecules in the floorboards are a kind of order, but not manifest, so they do not lower entropy; if they did, entropy would be constant and the second law empty. The second law: the entropy of an isolated system increases with time (or stays constant if reversible). It seems a counsel of despair, organization endlessly eroding into "useless" correlated motions, but Penrose promises that conclusion is "not entirely appropriate".
What is entropy?
"Irreversible" means only that we cannot track or control all particle motions (we call the uncontrolled motion heat), so entropy looks subjective: better instruments reveal more hidden order; an artist might judge the shards more beautifully ordered than the "hideously ugly" glass. Entropy is nonetheless precise and useful because the changes involved utterly swamp any reasonable differences of viewpoint.
- Coarse-grain phase space into compartments of states that are macroscopically indistinguishable.
- Entropy = k log V, where V is the phase-space volume of the compartment containing the actual state and k ≈ 10-23 J/K is Boltzmann's constant (Penrose uses natural units, k = 1).
- Why a logarithm: volumes of independent systems multiply (AB), so log AB = log A + log B makes entropy additive; and it tames the enormous numbers (the natural log of 10,000,000 is only about 16).
- Thermal equilibrium (the Maxwellian distribution, uniform temperature and pressure) occupies "almost the entire phase space".
- Balls in cells model (one tenth of cells "special"): chance all m balls land in special cells ≈ 1/10m (1/10, 1/100, 1/1000 ...; with small numbers of cells, 1/190 for two balls in twenty cells, 1/4060 for three in thirty).
- Real gas: a 1 m3 box of air at room conditions holds about 1025 molecules; the 1 cm3 corner is one millionth of it. The "all in the corner" compartment is 1/106 × 1025 of phase space: a 1 followed by sixty million million million million zeros in the denominator.
- In entropy terms the difference is only about 14k × 1025 ≈ 1,400 J/K: logarithms make absurd ratios look modest.
The second law in action, and its puzzle
Release gas from a corner: its phase-space point wanders into compartment after compartment, each dwarfing the last by stupendous factors, until it is lost in thermal equilibrium, from which it effectively never returns. This seems to explain the second law. But the argument is time-symmetric: run it backwards from "gas in the corner now" and it predicts that the gas was in equilibrium earlier and spontaneously gathered into the corner, entropy falling. That is absurd; really, the gas was held by a partition suddenly removed, or was frozen or liquid and heated, states of even lower entropy.
The phase-space argument shows only that, from a given low-entropy state, entropy should rise in both time directions unless something constrains it. Rising entropy toward the future needs no explanation; high entropy is "natural". The mystery is the past: "entropy gets more and more ridiculously tiny the farther and farther that we examine it in the past!" "We ourselves are configurations of ridiculously tiny entropy!"
The origin of low entropy in the universe
Trace our own low entropy back, step by step:
- We do not need energy; we need low entropy. Adult bodies keep a roughly constant energy content; we take in energy in a low-entropy form (food plus oxygen) and discard the same energy in a high-entropy form (heat, carbon dioxide, excreta), keeping our internal entropy down (Schrödinger, What is Life?). Meat (or mushrooms) merely relocates the question; green plants perform the key entropy-reducing separation of carbon from oxygen by photosynthesis, and we simply recombine them.
- The Earth does not gain energy from the Sun. It receives energy as relatively few, high-frequency visible photons and re-radiates the same energy as many more low-frequency infrared photons (E = hν), spreading it over far more degrees of freedom: a huge entropy increase that plants feed on. The Sun matters because it is "a hot-spot in the sky."
- Gravity made the hot spot. The Sun formed by gravitational contraction of diffuse hydrogen, heating as it shrank; fusion then stopped the contraction and stabilized its temperature at a level suitable for us. Without fusion it would have shone, hotter, briefly; without gravity there would be no Sun at all, just cold diffuse gas.
- Fossil fuels came from ancient plants (Thomas Gold's maverick theory of primordial deep hydrocarbons would still credit sunlight, in space). Uranium-235 comes from supernovae of earlier stars, whose collapse to neutron stars was again a gravitational entropy gain. (Fusion's entropy gain is "temporary", possible only because gravity concentrated nuclei away from the vastly more numerous background photons.)
So all the lowness of entropy around us comes from the potential of a diffuse gas to clump gravitationally. And this gives something more specific than "entropy started low": it could have been low in many other ways (a universe shaped as a regular dodecahedron, as Plato might have liked), but ours was low precisely as uniformly spread gas. Where did that come from?
Cosmology and the big bang
The universe is uniform on large scales and expanding: the farther the galaxies and quasars, the faster they recede, as if from a giant explosion about 1010 years ago. The 2.7 K black-body background (−270.3 °C), predicted by George Gamow (1948) and found accidentally by Penzias and Wilson (1965), is the cooled flash of the big bang. Recession from us does not make us central: on an inflating balloon every spot sees all others receding.
| Spatial curvature | Geometry | Extent | Fate |
|---|---|---|---|
| Positive | Sphere (the balloon's surface) | Finite | Expands to a maximum, then recollapses to a big crunch |
| Zero | Euclidean | Infinite | Expands for ever |
| Negative | Lobachevskian (Escher's fish) | Infinite | Expands for ever |
| With a suitable non-zero cosmological constant, infinite models can recollapse and finite ones expand indefinitely; Penrose takes it to be zero. In 1989 visible matter suggested negative curvature, but dark matter about thirty times the visible amount would close the universe. |
The primordial fireball
The big bang was not an explosion into pre-existing space from a central point: space itself was created, and the material uniformly filled all of it "right from the very start". (The balloon's interior and exterior are visual aids, not physics.) Standard physics describes the fireball reliably from about 10-4 s to three minutes (Weinberg 1977), producing photons, electrons, protons, about 23% helium by mass, some deuterons, traces of other nuclei, and many neutrinos; atoms formed later, and the gas began condensing into stars around 108 years. (Confidence rests on accelerator and cosmic-ray physics and on constants unchanged to 1 part in 106 over 1010 years.) Galaxy formation required some initial irregularities, then a controversial question.
Galaxies cluster into clusters, superclusters and perhaps supercluster complexes, but that is "small beer". The background radiation shows that at a million years old, over a region now spanning about 1023 km (some 1010 galaxies), the universe was uniform to one part in 100,000. Its diffuse, uniform gas is the low-entropy store we still live on.
Does the big bang explain the second law?
There is a paradox: the fireball was in thermal equilibrium, which is maximum entropy, yet the second law needs a minimum at the start.
The "standard" answer
The fireball had the maximum entropy allowed for a tiny universe; as the universe grew, the permitted maximum rose faster than the actual entropy, which forever lags and chases it. That chase is the second law.
Penrose's rebuttal
Then a recollapsing universe would face the same low ceiling near the big crunch, forcing entropy to fall and the second law to reverse. And black holes, microcosms of collapse, should show second-law violations nearby, yet "with black holes, the second law powerfully holds sway". Smallness alone cannot be the explanation.
Black holes
- Our Sun (about 5 × 109 years old) will in another 5-6 × 109 years swell into a red giant reaching about Earth's orbit (like Aldebaran or Betelgeuse), with a growing white-dwarf core.
- White dwarfs: a ping-pong ball of their matter weighs several hundred tonnes; about 10% of luminous Milky Way stars; Sirius's companion is the famous case, explained around 1926 by R. H. Fowler: electron degeneracy pressure (Pauli's exclusion principle) holds them up. The Sun ends as a slowly cooling white dwarf and finally an invisible black dwarf.
- Chandrasekhar limit (calculated in 1929 by Subrahmanyan Chandrasekhar on the boat from India to England; independently by Lev Landau about 1930): white dwarfs cannot exceed about 1.4 M☉. A heavier core collapses, releases a flood of neutrinos, and the star explodes as a supernova.
- Neutron stars: held up by neutron degeneracy; a ping-pong ball weighs as much as the asteroid Hermes or Mars's moon Deimos; radius about 10 km, nuclear density. Their own limit (Landau-Oppenheimer-Volkov) is roughly 2.5 M☉ (Penrose's figure).
- Above that: stars of 10 to 100 M☉ are common, and a black hole is expected.
Escape velocities: Earth about 40,000 km/h, Jupiter 220,000 km/h, the Sun 2,200,000 km/h; squeeze the Sun to a quarter of its radius and it doubles, to a hundredth and it rises tenfold. When it exceeds light speed, nothing escapes: a black hole (anticipated in Newtonian terms by John Michell, 1784, and Laplace; treated properly by Oppenheimer and Snyder, 1939). In a space-time diagram of collapse the light cones tip inward; at the Schwarzschild radius their outer edges become vertical, and light just hovers: the event horizon. A collapsed star of a few solar masses gives a horizon a few kilometres across; the Milky Way may hold a hole of about a million solar masses, a few million kilometres in radius.
B crosses the horizon at 12:00 by his watch. Safe outside, A sees B's watch read 11:30, 11:45, 11:52, 11:56, 11:58, 11:59, 11:59, 11:59 ... never 12:00, the image dimming rapidly into blackness. B notices nothing special at 12:00 and can still see A's watch running normally; unless he has calculated, he cannot know he has crossed. (Even the horizon's location depends on what will fall in later.) Then his local universe collapses toward a private big crunch, perhaps the "same" as the whole universe's crunch if space is closed. The second law holds inside the hole right up to the end.
The time reverse of a black hole is a white hole, a singularity spewing matter out of a horizon. Probably non-existent, it will matter greatly to the argument.
The structure of space-time singularities
Tidal forces grow as the inverse cube of distance. Near a few-solar-mass hole they would kill an astronaut before the horizon; at a million-solar-mass hole they are mild at the horizon ("a little uncomfortable") but rise to infinity within seconds, tearing apart body, molecules, atoms, nuclei and finally subatomic particles. Space-time itself ends: a space-time singularity. Hopes that singularities were artefacts of symmetry were dashed by the singularity theorems (Penrose 1965; Hawking and Penrose 1970): in classical general relativity with reasonable matter, gravitational collapse inevitably produces singularities, and time-reversed, any suitably expanding universe began in one.
Final singularities (crunch, black holes)
WEYL → ∞ (often oscillating, as in the Belinskii-Khalatnikov-Lifshitz analysis). Tidal, shape-distorting curvature dominates. High entropy. A realistic crunch is the congealing of many black holes after gravitational clumping.
Initial singularity (big bang)
In the FRW models WEYL = 0 exactly; only RICCI (volume-reducing, matter-driven curvature) diverges. Low entropy. The singularities are not time-reverses of each other, and the difference is the second law.
Why is clumping high-entropy when a gas in a corner is low? Because gravity is universally attractive: with gravity, the uniform state is the special one. Time-reverse a realistic crunch and you get a "big bang" of enormous entropy, riddled with white holes, with no second law. A "probable" universe would have WEYL → ∞ at both ends.
How special was the big bang?
Assume a closed universe with B = 1080 baryons (Eddington once claimed to calculate B exactly and got about this; a larger B would make the result even more extreme). Picture the Creator holding a pin over the phase space of all possible universes.
- Sbh = (kc3/Għ) × A/4: entropy is proportional to the horizon area A (Bekenstein 1972, Hawking 1975).
- For a spherical hole A = 8π(G2/c4)m2, so Sbh = 2π(kG/ħc)m2: entropy grows as mass squared, so entropy per unit mass grows with mass. For a given energy, maximum entropy is reached by collapsing into a black hole, and two holes merging gain enormously.
- Hawking temperature: a solar-mass hole is at about 10-7 K, far colder than the 2.7 K sky; evaporation would take some 1054 times the universe's age. Mini-black holes would need a chaotic big bang; Penrose expects none.
| Scenario | Entropy per baryon (natural units) | Total entropy (B = 1080) |
|---|---|---|
| Sun's matter | about 1 | (negligible) |
| 2.7 K background radiation (108 photons per baryon) | 108 | 1088 |
| Universe made of solar-mass black holes | 1020 | 10100 |
| Galaxies of 1011 stars, each with a 106 M☉ central hole | 1021 | 10101 |
| Most galactic mass eventually in central holes | 1031 | 10111 |
| Final big crunch treated as one black hole | 1043 | 10123 |
The largest compartment has entropy about 10123, so the total phase-space volume is V = 1010123. Our universe's initial compartment W is at most 1010101 (or 101088, or smaller still at the actual big bang); either way V/W ≈ 1010123 (since 10123 − 10101 ≈ 10123).
The number cannot be written out: it is 1 followed by 10123 zeros, more zeros than there are particles in the universe to write them on. The precision of the initial state is "in no way inferior" to the precision of the dynamical laws themselves. Penrose names the constraint the Weyl Curvature Hypothesis: WEYL = 0 (or something very like it) at initial space-time singularities, but not at final ones. Understanding the second law means understanding why such a time-asymmetric rule holds.
The impasse. Singularities are where known physics fails, as classical physics failed for collapsing atoms early in the century; quantum theory cured that infinity, and a quantum theory of space-time should cure this one. That theory, quantum gravity, does not exist despite first-rate efforts. Penrose's footnote dismisses inflation as a substitute: it smooths irregularities only given an even stronger initial constraint, contains no time asymmetry, and rests on TENTATIVE GUT physics. The return journey to mind, he warns, "will be more speculative".
Try it in Python: counting compartments and black-hole entropy
# Balls in cells: chance all m balls are "special"
from math import comb, pi
for m, n in [(1, 1), (2, 2), (2, 10), (3, 3)]:
cells, special = 10*n, n
p = comb(special, m) / comb(cells, m)
print(f'm={m}, {cells} cells: 1/{round(1/p)}')
# Bekenstein-Hawking entropy per baryon (k = 1)
G, hbar, c = 6.674e-11, 1.055e-34, 2.998e8
m_sun, m_p = 1.989e30, 1.673e-27
S = lambda M: 4*pi*G*M**2/(hbar*c)
for msun in (1, 1e6):
M = msun*m_sun
print(f'{msun:g} Msun: {S(M)/(M/m_p):.1e} per baryon')
The first loop reproduces Penrose's 1/10, 1/190, 1/110 and 1/4060. The second gives about 1020 per baryon for a solar-mass hole and 1026 for a million-solar-mass one, confirming his table's starting point and showing why galactic black holes dominate the universe's entropy budget.
Watch a film of a glass shattering, then the reverse. You instantly know which is "real", though every frame obeys the same laws either way. Now film a planet forming: dust and gas drawing together into a hot, compact ball. Run it backwards and you see a hot planet dissolving into a uniform cloud. Which is real? Here intuition from the glass misleads: with gravity, clumping is the entropy-increasing direction, so the forming planet is the natural film. Penrose's whole chapter rests on this reversal. The early universe is "the dust cloud" of the cosmos: gravitationally the most special state imaginable, and the reason any film of ours runs the way it does.
Where the 1010123 comes from
The most quoted number in the book is easy to misread. It is reached in four steps, each simple once laid out.
- Count the baryons. The observable universe contains about 1080 baryons (protons and neutrons).
- Find the largest possible entropy. For fixed matter, entropy is greatest when everything has collapsed into a black hole. The Bekenstein-Hawking formula makes a black hole's entropy proportional to its horizon area, so it grows as the square of the mass. A black hole holding 1080 baryons has entropy of about 10123 in natural units.
- Convert entropy to phase-space volume. Boltzmann's formula S = k log V means the phase-space volume is eS/k, so the maximum volume is about 1010123.
- Compare with our universe. Our early universe was extraordinarily smooth, with entropy far below this maximum (roughly 1088 in the radiation). Its phase-space region is therefore a fraction about 1 in 1010123 of the whole.
The point is not the exact exponent but its form: a double exponential, far beyond any number arising from chance in ordinary physics. Penrose's Creator needs that precision to place the pin. Writing the number out in full would need more digits than there are particles in the universe.
- 1080 baryons: the observable universe.
- 1088: the entropy of the cosmic background radiation (in units of Boltzmann's constant), the largest contributor until black holes are counted.
- 10101: Penrose's estimate if every galaxy holds a million-solar-mass black hole.
- 10123: the whole universe as one black hole.
- 1010123: the corresponding phase-space volume, the "odds" against our big bang.
Why inflation does not answer Penrose
Inflation was in 1989, and remains, the most popular explanation for why the universe is so smooth. Penrose argues it cannot address the real puzzle, for a reason that is easier to state now than in 1989.
- Inflation is a dynamical process driven by time-symmetric laws. Run backwards from a generic, lumpy, high-entropy state and those same laws give a generic, lumpy past, not a smooth beginning.
- To work, inflation must itself start in a special, low-entropy patch of a quantum field. So it moves the improbability rather than removing it.
- The second law says entropy was lower in the past. Any mechanism that makes a smooth universe from a generic one would be making entropy decrease, which is the very thing the second law forbids.
Only a law that is itself time-asymmetric, applying to the initial singularity, can do the job. That is the Weyl Curvature Hypothesis: Weyl curvature (the tidal, "clumpy" part) vanishes at initial singularities and is large at final ones. Penrose admits it is a postulate, not a derivation, and in Chapter 8 he suggests it is a clue to quantum gravity.
Imagine a universe identical to ours but with time labelled backwards. Its inhabitants, made of the same physics, would remember what we call the future, because memory needs a low-entropy past and their low-entropy end is our big bang. They would describe our big bang as their big crunch, and they would be wrong: from their side it looks like a smooth, low-Weyl ending, which a crunch never is, because collapsing matter forms black holes and grows Weyl curvature. The asymmetry is in the singularities, not in the labelling. This is why Penrose can say a reversed universe would be detectably different and not just a relabelling of ours.
Key terms from Chapter 7
| Term | Meaning in this book |
|---|---|
| Second law | Entropy of an isolated system does not decrease |
| Coarse-graining | Grouping microstates that look alike into compartments; entropy is the log of the compartment's volume |
| Friedmann-Robertson-Walker model | A uniform expanding universe, closed, flat or open |
| Cosmic background radiation | The 2.7 K remnant of the primordial fireball |
| Black hole | A region from which light cannot escape, bounded by an event horizon |
| Bekenstein-Hawking entropy | Black-hole entropy, proportional to horizon area |
| Weyl Curvature Hypothesis | Weyl curvature is zero at initial singularities |
Test yourself on Chapter 7
- Why does an egg falling off a table never reassemble? Not because the laws forbid it but because the reassembling states occupy a vanishingly tiny part of phase space compared with smashed ones.
- What is the real puzzle of the second law? Not why entropy increases, but why it was so low to begin with.
- Why does gravity reverse the usual intuition about clumping? For a gas in a box, spreading out raises entropy; under gravity, clumping raises it, culminating in black holes.
- Where does the Earth's useful energy come from? From receiving energy as few hot photons from the Sun and radiating it as many cold ones; the Sun's low-entropy state traces back to the smooth big bang.
- Why does Penrose say the big bang was "special"? Its matter was hot and in equilibrium, but its gravitational degrees of freedom were in the lowest possible entropy state: no Weyl curvature at all.
A library has a million books in perfect order. Each night, a librarian swaps two random books. After a year, the shelves look jumbled, and no one expects them to sort themselves by next spring, though each single swap is perfectly reversible. Now suppose you walk in and find the shelves perfectly ordered. You do not conclude that random swapping produced the order; you conclude someone arranged it. Penrose's point about the big bang is the same inference at cosmic scale: an extraordinarily ordered starting state is not a product of the dynamics and calls for its own explanation.
- The numbers sharpened. The universe is 13.8 billion years old (Planck 2018); the background is 2.7255 K; COBE (1992) detected its anisotropies at about 1 part in 100,000, exactly the level of uniformity Penrose quoted as a bound.
- No big crunch. Supernova surveys (1998, Nobel 2011) found the expansion accelerating: a positive cosmological constant (dark energy) dominates and space is flat to within about 0.2%. Penrose's closed, recollapsing model is disfavoured, so he later recast the argument in Conformal Cyclic Cosmology (Cycles of Time, 2010), keeping the Weyl Curvature Hypothesis at its centre. DESI (2024-25) hints that dark energy may evolve, an open question.
- Black-hole entropy confirmed in scale. Egan and Lineweaver (2010) estimated the observable universe's entropy at about 3 × 10104 k, dominated by supermassive black holes, in line with Penrose's 10101-type estimates. In The Road to Reality (2004) he revised his headline figure to 1010124.
- The Milky Way's black hole is about 4 million solar masses (Genzel and Ghez, who shared the 2020 Nobel with Penrose), imaged by the Event Horizon Telescope (2022). LIGO has watched black holes merge, and Hawking's area theorem (merged horizon area exceeds the sum) was confirmed from GW150914 data in 2021. Hawking radiation remains unobserved directly (analogue versions seen in fluids).
- Neutron-star limit is now thought to be about 2.2-2.3 M☉ (heaviest known pulsars near 2.1-2.35 M☉), lower than Penrose's "very roughly 2.5".
- Inflation remains the dominant paradigm, supported by the measured near-scale-invariant spectrum (ns ≈ 0.965), but primordial gravitational waves have not been found (r < 0.036, BICEP/Keck 2021), and Penrose's critique about initial conditions is still pressed by a minority.
Common misreadings of Chapter 7
- "Entropy is just disorder." It is the logarithm of a phase-space volume of states that look alike at a chosen coarse-graining. "Disorder" is a loose shorthand that misleads under gravity, where clumping is high entropy.
- "The second law is explained by the expansion of the universe." Penrose argues it is explained by the initial state; expansion alone, with time-symmetric laws, would not do it.
- "Black holes suck everything in." At a distance, a black hole's gravity is the same as any body of equal mass. Its special feature is the horizon and the singularity inside.
- "The 1010123 is a probability that our universe exists." It is the fraction of phase space with a big bang as smooth as ours, given the matter we have: a measure of the initial state's specialness.
- "The Weyl Curvature Hypothesis is a theorem." It is a postulate, which Penrose hopes a future quantum gravity will explain.
- The Weyl Curvature Hypothesis is a postulate, not a derivation. No accepted measure of "gravitational entropy" exists (Penrose's own note 12 admits this; later proposals such as Clifton, Ellis and Tavakol 2013 are tentative), and the hypothesis restates the low-entropy past rather than explaining it.
- Phase-space measures over universes are ill-defined in general relativity (infinite volumes, gauge freedom); the 1010123 figure is heuristic, though its message (the early universe was gravitationally very special) is widely accepted.
- Thomas Gold's abiogenic-oil theory was not vindicated: the Swedish Siljan Ring drilling Penrose mentions found negligible hydrocarbons.
- The closed-universe framing dates the chapter; the asymmetry argument survives in open, accelerating universes but loses its vivid big-crunch comparison.
- Organization is imported, never free. Every ordered system, from a body to a data platform, survives by consuming a low-entropy source and exporting disorder; governance that does not budget for that ongoing work decays toward equilibrium.
- "Manifest" versus "hidden" order is a useful lens on data: correlations buried in raw logs are real order but count for nothing until a coarse-graining (a model, a schema) makes them manifest.
In Search of Quantum Gravity
Here Penrose turns unconventional. Instead of asking how quantum theory changes gravity, he asks how gravity must change quantum theory. The time asymmetry of the Weyl Curvature Hypothesis means the true "correct quantum gravity" (CQG) must be time-asymmetric; the R procedure, as physicists actually use it, is time-asymmetric too; and Hawking's box suggests the two are "the other side of the coin": black-hole singularities destroy information (phase-space flow lines merge), so R must create it (flow lines branch). His proposal: state-vector reduction is an objective gravitational effect, occurring when superposed alternatives differ by about one graviton's worth of space-time curvature.
Why quantum gravity?
The standard view: quantum gravity matters only at the Planck length, 10-35 m, about 1020 times smaller than the smallest particle, irrelevant to brain chemistry at 10-12 m and above, where even classical gravity barely matters; and since no deviation from quantum theory has ever been seen, a tiny modification could hardly matter to mind. Most physicists, faced with the difficulty of quantizing Einstein's theory, modify general relativity (higher-order Lagrangians, Kaluza-Klein extra dimensions, supersymmetry, strings: all TENTATIVE), never quantum mechanics.
Conventional programme
Quantum rules are sacrosanct; gravity gets quantized and modified. Quantum gravity will fix the Planck-scale structure of space-time, perhaps explain particle masses (mass is gravity's source), and, following Oskar Klein's idea of about 1955, cure the infinities of quantum field theory.
Penrose's programme
The incompatibility of U and R cannot be cured by any interpretation; both are approximations to a single, exact, non-linear procedure, and general relativity supplies the hints. That combined U/R procedure is "one of the fundamental ingredients" of quantum gravity, and it acts at scales far above the Planck length.
What lies behind the Weyl Curvature Hypothesis?
Everyone agrees quantum gravity must replace classical singularities with sense. Few notice that its mark is blatantly time-asymmetric: WEYL = 0 at the past singularity, WEYL → ∞ at future ones. Hence, in italics in the original: "our sought-for quantum gravity is a time-asymmetric theory." This is "not accepted wisdom", because standard quantization of a time-symmetric classical theory cannot produce the needed violations of T, PT, CT and CPT.
Boundary conditions are science too. Physics has gained enormously by separating simple dynamical laws from messy initial arrangements (even Kepler's simple ellipses turned out to be consequences of dynamics), and many would call the big bang's state an "act of God" beyond explanation. But the Creator's aim was no less precise than the dynamical laws themselves. Penrose's personal bet: in the final theory the distinction between dynamical equations and boundary conditions "will dissolve away" into one consistent scheme.
- WCH applies to every initial singularity, not only the big bang.
- White holes are forbidden. A white hole's initial singularity would have WEYL → ∞, violating WCH; this also fits thermodynamics (white holes would grossly violate the second law) and observation (proposals invoking them have always raised more problems than they solved). The big bang itself is allowed because it can have WEYL = 0.
- Exploding evaporated black holes (after about 1064 years of Hawking evaporation for a solar mass) are probably compatible: a small, symmetric "initial" event. The first will not occur until the universe is about 1054 times its present age; compress today's age to the shortest particle lifetime and the present universe would still fall short by more than a million million.
- Two kinds of singularity, symmetric theory (implicit in Hawking's 1987-88 proposals; Hartle and Hawking's 1983 no-boundary state might give something like WEYL = 0 but lacks a time-asymmetric input): it fails to explain why there are no other initial singularities, so why the universe is "riddled with" black holes but no white holes.
- The anthropic principle (Barrow and Tipler 1986): a rough count shows the entire solar system with its inhabitants could arise from random particle collisions with an improbability of "only" much less than one part in 101060, nowhere near 1010123; and it too says nothing about white holes.
Time-asymmetry in state-vector reduction
The "prevailing view" holds R time-symmetric, partly from reluctance to treat R as a real process separate from U. Penrose argues that R, as physicists actually use it, is time-asymmetric.
The formal symmetry: forward, the probability of jumping from |ψ〉 (after an earlier observation O′, evolved forward) to |χ〉 at O is |〈χ|ψ〉|2; evolving |χ〉 backwards by U to |χ′〉 near O′ gives |〈χ′|ψ′〉|2, the same number, because U preserves scalar products. But the number answers "given the earlier result, what is the probability of the later one?", not the time-reversed question.
A lamp L fires occasional photons aimed at a photo-cell P, with a half-silvered mirror M at 45° in between; the reflected beam goes to point A on the laboratory wall. Both L and P record perfectly. - "Given L registers, what is the probability P registers?" Amplitudes 1/√2 to P and to A, so ½: correct. - The reversed-time calculation gives ½ again, but it answers "given L registers..." or the eccentric "given a photon is ejected from the wall at B, what is the probability P registers?" - The true time-reverse, "Given P registers, what is the probability L registered?", has the experimental answer one: photons do not leap out of walls. Squaring amplitudes gives correct probabilities for prediction but wrong ones for retrodiction. R "cannot be time-symmetric (and, incidentally, therefore cannot be a deduction from the time-symmetric procedure U)." Penrose admits he once made the symmetric assumption himself (1979).
The usual reply is that the second law sneaks in: every measuring device involves thermodynamic irreversibility. Penrose agrees the second law is probably involved, but insists the striking fact stands: the amplitude-squaring rule needs no extra knowledge in the future direction and simply fails in the past direction; "Such excuses are simply not needed in the future-direction!" (Asking the probability of a past event given a future one is meaningful: survey all occurrences of the later event in history and count how often the earlier one accompanies it.)
Hawking's box
A sealed box with perfectly reflecting walls that nothing passes, not even gravity. Inside, a large mass M in a large volume V; leave it alone for ever. Choose M about the solar system's mass and V about the Milky Way's size: at equilibrium nearly everything has collapsed into a black hole in a thin thermal bath at about 10-7 K. Occasionally fluctuations occur: the hole emits too much Hawking radiation (virtual pairs near the horizon, one partner swallowed, the other escaping), gets hotter as it loses mass, and in a rare runaway evaporates completely in an explosion; at other rare times thermal radiation collapses to form a hole.
Split the box's phase space into 𝓑 (states with a black hole) and 𝓐 (states without). Flow lines cross from 𝓐 to 𝓑 when a hole forms and from 𝓑 to 𝓐 when one evaporates, each needing a large fluctuation through a small compartment. Which crossing is "easier"?
- Liouville's theorem says the phase-space "fluid" is incompressible, so flow 𝓐 → 𝓑 must equal flow 𝓑 → 𝓐: making a hole is exactly as hard as destroying one. Hawking reached the same conclusion from time symmetry, together with his remarkable suggestion that black holes and white holes are physically identical, Hawking radiation being the time-reverse of swallowing. Penrose doubts this "can be made to work", and it clashes with WCH.
- With WCH: black-hole singularities destroy information, so in 𝓑 two different states can become the same: flow lines merge, an effective violation of Liouville, fluid continually annihilated in 𝓑.
- The imbalance: more fluid then flows out of 𝓐 than returns, yet 𝓐 has no black holes and, by WCH, no white holes, so Liouville should hold there. Something must create fluid: one state with more than one possible outcome, bifurcating flow lines. That has "the 'smell' of quantum theory... The R part".
Black-hole singularities make flow lines converge; objective state-vector reduction makes them branch, exactly compensating. Both are implications of CQG and both are tied to the second law. Branching needs a time-asymmetric R, as the lamp experiment showed: emission splits into "cell fires" or "photon hits wall at A"; the would-be second input, a photon emitted from the wall at B, is excluded by the second law, ultimately by WCH. Reduction can occur with no black hole nearby: what must be balanced is the mere possibility that holes might form and destroy information.
Penrose lists objections honestly. Some say evaporation leaves a permanent remnant (that would strengthen his argument). One could posit an infinite phase-space volume (at odds with black-hole entropy and finite quantum systems). The box may require impossible idealizations (he is "inclined to believe" they can be swallowed). Most seriously, the argument mixes classical phase space with quantum effects; Hilbert space has its own Liouville analogue (unitarity), and the right arena is probably "some hitherto undiscovered type of mathematical space" between the two. So the argument is heuristic, "suggestive rather than conclusive", but he thinks it makes "a strong case" that R is a quantum-gravity effect.
When does the state vector reduce?
The rule that momentum becomes −iħ∂/∂q sits badly with curved space-time. Penrose's proposal: as soon as a "significant" amount of space-time curvature is involved, linear superposition fails, and complex-weighted alternatives become probability-weighted actual alternatives, one of which happens, objectively and with no observer.
- "Significant" means the difference between the gravitational fields (space-time geometries) of the superposed alternatives reaches about one graviton (the quantum of gravitational field, by analogy with photons; strictly the static "longitudinal" virtual gravitons, hard to define invariantly).
- At that level a time-asymmetric non-linear instability sets in and the system "flops" into one alternative, by chance or by something deeper.
- Scale is mass and energy, not size: quantum interference survives over long distances when little energy is involved.
- Planck mass mP ≈ 10-5 g is the natural scale, "a little less than that of a flea", which seems too large, since dust specks behave classically; Penrose expects a subtler application.
- Wilson cloud chamber. A radioactive atom emits a particle as a spherical wave of possible directions; each possible track ionizes atoms along its path and seeds droplets, so standard theory demands a superposition of many different droplet trails. Reduction would occur when the trails' gravitational fields differ by one graviton: for a single uniform spherical droplet, at about mP/100 ≈ 10-7 g (Abhay Ashtekar's improved estimate, which he cautioned involves much arbitrariness).
- A box of gas, no observer. A particle either enters a dense gas or does not. If it enters, it sets off a cascade until virtually every atom is displaced. The overall mass distribution is nearly the same in both branches, but the difference field (atom by atom) is highly oscillating and can easily exceed one graviton. Reduction then happens objectively: either the particle entered or it did not. Photographic plates and spark chambers should work similarly.
This is "only a germ of an idea". A full theory would need radically new ideas about space-time, probably non-local (as EPR shows: reduction across a room is "simultaneous", a notion foreign to relativity), perhaps along the lines of his own twistor theory, which still lacks essential ingredients. Other objective-reduction schemes are acknowledged: Károlyházy (1974; with Frenkel and Lukács, 1986), Komar (1969), Pearle (1985, 1989), Ghirardi, Rimini and Weber (1986). Our picture of reality, especially of time, "is due for a grand shake up". Next: the brain.
Try it in Python: how fast would gravity collapse a superposition?
# Diosi-Penrose lifetime tau ~ hbar / E_G,
# E_G ~ G m^2 / R for a lump displaced by about
# its own size (order of magnitude only)
G, hbar = 6.674e-11, 1.055e-34
for name, m, R in [('dust grain', 1e-15, 1e-6),
('Penrose droplet', 1e-10, 3e-5),
('cat', 4.0, 0.1)]:
tau = hbar / (G*m*m/R)
print(f'{name:16s} tau ~ {tau:.1e} s')
This uses the later Diósi-Penrose rule that replaced the 1989 one-graviton criterion. A microscopic dust grain would hold a superposition for about a second; the cloud-chamber droplet of 10-7 g that Penrose estimated would collapse in nanoseconds; a cat in about 10-26 s. Ordinary environmental decoherence is usually faster still, which is why testing the idea needs extreme isolation.
Imagine reading the log of the sealed box over 10100 years. Most entries read "black hole in equilibrium with a cold bath". Very rarely: "hole evaporated, explosion, gas everywhere"; rarer still: "gas collapsed, new hole". Hawking would say the two rare entries appear equally often, as a time-symmetric world demands. Penrose says information swallowed by the hole is gone, so if the log is honest, the "gas" phases must contain moments where one state led to several outcomes to keep the books balanced, and those moments are what we call measurement. The box would be a ledger in which quantum randomness pays for black-hole information loss. Whether the ledger really balances that way depends on whether information truly vanishes, still argued in 2026.
The one-graviton criterion and the Diósi-Penrose rule compared
Chapter 8 proposes that a superposition reduces when the two alternatives differ by about one graviton's worth of space-time geometry. By 1996 Penrose replaced this with a cleaner rule, now usually called Diósi-Penrose (DP), and it is worth seeing how the two relate.
1989 criterion
Reduction occurs when the alternatives differ in their gravitational fields by enough to amount to a single graviton. The idea is qualitative: it gives a threshold, roughly at the Planck mass of 10-5 grams, for when superpositions can no longer persist.
1996 Diósi-Penrose rule
A superposition of two mass distributions decays at a rate EG/ħ, where EG is the gravitational self-energy of the difference between the distributions. The lifetime is τ ≈ ħ/EG: tiny for big displacements of big masses, enormous for single particles.
The DP rule has two advantages. It is quantitative, so it predicts actual lifetimes that experiments can try to beat. And it has a physical rationale: two superposed mass distributions define two different space-times, and general relativity gives no exact way to say "the same point" in both. Penrose argues this ill-definedness costs an energy uncertainty EG, and by analogy with an unstable particle, the superposition decays in time ħ/EG. Gravity, on this view, refuses to hold two geometries at once for long.
| Object superposed over its own size | Typical τ (DP, rough) | Implication |
|---|---|---|
| Single nucleon | Millions of years | Quantum mechanics unaffected |
| Dust grain (10-5 cm) | Seconds | Borderline, testable |
| Water droplet (10-3 cm) | Nanoseconds | Collapse effectively immediate |
| A cat | 10-26 s or less | No cat is ever both alive and dead |
These are the numbers the Python sketch in this chapter computes. They show why the proposal is safe from everyday quantum experiments and yet within reach of mesoscopic ones.
In 2003 Marshall, Simon, Penrose and Bouwmeester proposed bouncing a single photon off a tiny mirror on a cantilever, so the mirror is pushed into a superposition of two positions. After a full oscillation, the photon and mirror should re-interfere. If DP collapse acts, the interference should fade at a rate set by the mirror's mass and displacement, beyond what ordinary decoherence explains. The experiment needs a mirror colder, quieter and better isolated than anything yet achieved, but optomechanics has advanced quickly: mechanical oscillators have been cooled to their quantum ground state and placed in superpositions. The race is between technology and the need to rule out every mundane source of decoherence.
How the WCH and R are linked
Penrose's boldest claim in this chapter is that the time-asymmetry of the universe (the Weyl Curvature Hypothesis) and the time-asymmetry of measurement (rule R) have a common source in quantum gravity. The argument runs:
- Rule R is not time-symmetric: calculating backwards from a later measurement with Born probabilities gives wrong answers, as his lamp-and-photo-cell example shows.
- The WCH is also time-asymmetric: initial singularities are smooth, final ones are not.
- Both involve gravity in a regime standard physics cannot handle.
- So a single correct theory of quantum gravity, CQG, should be time-asymmetric and should explain both.
This is suggestive, not deductive, and Penrose says so. Its importance for the book is that it gives him a principled reason to expect the missing physics to be unlike any existing law, perhaps unlike any algorithm.
Key terms from Chapter 8
| Term | Meaning in this book |
|---|---|
| CQG | Penrose's name for "correct quantum gravity", the unknown theory he seeks |
| Hawking radiation | Thermal emission from black holes, making them evaporate |
| Hawking's box | A sealed box containing matter and black holes, used to study phase-space flow |
| Information loss | The claim that black holes destroy quantum information, shrinking phase space |
| Graviton | The hypothetical quantum of the gravitational field |
| Planck mass | About 2 × 10-5 grams, where gravity and quantum effects meet |
| Objective reduction (OR) | Collapse as a real, physical, gravity-induced process |
Test yourself on Chapter 8
- Why does Penrose think the time-asymmetry of R and the WCH are connected? Both are time-asymmetric effects that involve gravity where current theory fails; he suggests one correct quantum gravity explains both.
- What does the Hawking-box argument claim? If black holes destroy phase-space volume, then something else, namely R, must create it to keep the total balanced at equilibrium.
- Why is the cat never in a superposition, on Penrose's criterion? Its alternatives differ by far more than one graviton, so reduction happens almost instantly.
- What would refute the proposal? Observing quantum interference in a superposition well beyond the predicted collapse threshold, with all decoherence accounted for.
- How has the argument fared since 1989? The simplest DP version was excluded by the Gran Sasso radiation test; refined versions survive; the information-loss premise is now disputed. The chapter's central idea, that gravity has a role in quantum measurement, remains open.
- The criterion was replaced. From 1996 Penrose adopted the Diósi-Penrose form: a superposition of two mass distributions decays in a time τ ≈ ħ/EG, where EG is the gravitational self-energy of the difference between them (Diósi proposed a closely related model in 1987). This is the "more plausible scheme" of the 1999 preface.
- Tests. The simplest parameter-free DP model was excluded at Gran Sasso (Donadi et al., 2020-21) by the absence of spontaneous radiation; versions with larger mass-smearing survive. Proposed tests with superposed mirrors (Marshall, Simon, Penrose and Bouwmeester, 2003), levitated nanoparticles and space missions (MAQRO) are still in development; a 16-microgram oscillator held a cat state in 2023 without visible collapse, tightening bounds.
- Gravity-mediated entanglement (Bose et al.; Marletto and Vedral, 2017) would test whether gravity can be quantum at all; no experiment has yet reached the needed sensitivity.
- Information loss is contested. Penrose's argument needs black holes to destroy information. Hawking conceded in 2004 that information is preserved, and the "island"/Page-curve calculations of 2019-20 strongly suggest unitarity in models; Penrose has continued to defend genuine loss (and builds it into Conformal Cyclic Cosmology). The question remains open for real black holes.
- Twistor theory revived in particle physics: Witten's twistor string (2003) and the amplituhedron (2013) made twistor methods central to computing scattering amplitudes.
Why gravity, and not something else?
Penrose's choice of gravity as the trigger for state reduction can look arbitrary. In this chapter and the 1999 preface he gives several reasons, which together make it a principled bet rather than a guess.
- Gravity is the odd one out. Every other force has been brought into quantum theory successfully; gravity alone resists. If quantum theory needs changing somewhere, the unfinished junction with gravity is the obvious place to look.
- Gravity is about space-time itself. A superposition of a mass in two places is, in general relativity, a superposition of two space-time geometries. Quantum theory has no clear rule for comparing points in two different geometries, which suggests the superposition is ill-defined and cannot last.
- The scale is right. The Planck mass, about 10-5 grams, sits between the atomic world where superpositions are seen and the everyday world where they never are. A gravitational criterion naturally puts the quantum-classical boundary in roughly the right place.
- Both asymmetries involve gravity. The Weyl Curvature Hypothesis is a statement about gravitational degrees of freedom; Hawking's box links black holes to phase-space balance. A gravitational R ties these together.
- It is testable. Unlike interpretations that change only the story told about quantum mechanics, a gravitational collapse law changes predictions, so experiment can in principle decide.
The bet is that the cure for quantum theory's measurement problem and for gravity's quantization problem is the same cure. No one has shown this is true, but the experiments described in the 2026 updates are now probing exactly the regime where it would show up.
Common misreadings of Chapter 8
- "Penrose has a theory of quantum gravity." He has requirements for one: time-asymmetric, involving R, explaining the WCH. He does not have the theory, and says so.
- "Gravitational collapse is a form of decoherence." Decoherence is ordinary quantum dynamics with an environment; Penrose's OR is a new physical law that would act even in perfect isolation.
- "Hawking's box proves information is lost." It assumes information loss and asks what must balance it. If information is preserved, the argument for R as phase-space creation loses its main support.
- "Collapse would make quantum computers impossible." For systems whose superposed alternatives differ in mass distribution only slightly, the DP time is very long; qubits based on electrons or photons are unaffected.
- "The one-graviton criterion is still Penrose's view." It was replaced in 1996 by the Diósi-Penrose rule, as the 1999 preface explains.
- R's time asymmetry may be borrowed from thermodynamics. Many physicists (and the two-state-vector formalism of Aharonov, Bergmann and Lebowitz) hold that retrodictive failures reflect the low-entropy past built into apparatus, not an asymmetric law; Penrose acknowledges the objection but rests on the asymmetry of the rule "as actually used".
- The Hawking-box argument mixes classical and quantum phase spaces and presupposes information loss; if black holes are unitary, the "merging" and hence the need for compensating branching disappears.
- The flea comparison: the Planck mass (about 2 × 10-5 g) is roughly a twentieth of a typical flea's mass, not "a little less".
- Objective-collapse models must cope with energy non-conservation and relativistic consistency; GRW-type models have made progress (relativistic versions by Tumulka, Bedingham), but none is yet experimentally favoured.
- Interpretations versus new theories: Penrose refuses to paper over a structural inconsistency with a better narrative. In architecture terms, when two core components (U and R) have incompatible contracts, documentation cannot fix it; a redesign that subsumes both can.
- Conservation as a design check: the Hawking-box reasoning is a ledger argument (what is destroyed in one region must be created elsewhere). Ledger-style invariants (reconciliation, double entry, lineage completeness) catch hidden sources and sinks in data systems for the same reason.
Real Brains and Model Brains
Penrose surveys the brain as a physicist visiting neurology: its anatomy, the contested seat of consciousness, split brains, blindsight, the visual cortex, how nerve signals work, and why neurons can be wired into logic gates, so that brains and computers look equivalent. He then lists the differences that matter, above all plasticity and the "oneness" of conscious thought, which suits quantum parallelism better than classical parallel computers. His sober conclusion: known neural mechanisms give quantum effects no clear role, and the brain is too "hot" to keep quantum coherence; if mind escapes computation, it must be through physics beyond present quantum theory.
What are brains actually like?
Turing likened the brain to "a bowl of cold porridge". Up close it is highly organized.
- Cerebrum: two hemispheres; frontal, parietal, temporal, occipital lobes; thin outer grey matter (the cortex, where computation seems to happen) over white matter (long, myelinated fibres). Proportionally larger in humans than in other animals.
- Cerebellum: "two balls of wool" at the back; precise coordination, timing and balance.
- Brain-stem: pons and medulla, including the reticular formation (general alertness).
- Deep structures: thalamus (relay and processing station), hypothalamus ("the seat of emotion": pleasure, rage, fear, hunger), hippocampus (laying down long-term memories), corpus callosum (the bridge between hemispheres).
- Embryology: hindbrain (most ancient: cerebellum, part of the reticular formation), midbrain, and forebrain (newest: cerebrum, thalamus, hypothalamus, hippocampus, corpus callosum).
Crossed wiring. The right hemisphere serves the left side of the body and vice versa. The visual cortex sits at the back of the head, curious given that eyes are at the front, and maps not each eye but each visual field: the left field of both eyes goes to the right visual cortex. Hearing crosses mostly; smell is the exception (right nostril, right olfactory cortex, unusually in the frontal lobe). Touch maps onto the somatosensory cortex just behind the frontal-parietal divide, movement onto the motor cortex just in front, each pictured as a distorted homunculus.
- Primary sensory regions (rear lobes) receive sight, sound, touch.
- Secondary regions process them at higher abstraction.
- Tertiary (association) cortex integrates senses, lays down memories, builds pictures of the world, understands speech (Wernicke's area, upper rear temporal lobe).
- The arcuate fasciculus carries results forward to the frontal lobe, where tertiary motor regions form plans (Broca's area formulates sentences).
- Secondary motor regions turn plans into movement schemes; primary motor cortex drives specific muscles. Damage: Broca's area impairs speech but not comprehension; Wernicke's leaves fluent speech with little content; the arcuate fasciculus leaves comprehension intact but unable to be voiced. Language sits on the left in most people.
This looks like "a superb computing device", input tape to output tape, and strong AI would say awareness emerges once internal processing is sophisticated enough. Penrose warns the picture is rough: there are several visual maps with different relations to awareness, scattered motor points, and other organs. The cerebellum takes over skills once learned ("second nature"): the dancer, tennis player, racing driver, a gazelle's bound; think about such a skill and cerebral control returns, flexible but clumsy. The hippocampus is essential for converting short-term into permanent memory; its damage leaves a person unable to retain anything once it leaves attention.
Where is the seat of consciousness?
There is "remarkably little consensus". Cerebellar actions and spinal reflexes (the hand off the hot stove) seem unconscious, so the cerebrum is the obvious candidate, yet primary cortical regions are active during walking without our noticing the soles of our feet.
| Proposed seat | Proponent and evidence | Problem or Penrose's view |
|---|---|---|
| Upper brain-stem (thalamus, midbrain, reticular formation) in communication with the relevant cortex | Wilder Penfield (1940s-50s operations on conscious patients; Penfield and Jasper 1947): stimulating motor cortex moves an arm but creates no wish to move it; the patient may stop it with the other hand, like Dr Strangelove | Plausible as a hub, with the cortex supplying content |
| Reticular formation (alertness) | Moruzzi and Magoun (1949): damage causes unconsciousness | Inactive in parts during dreams, which are aware; it is ancient, so frogs, lizards, codfish would be conscious. Penrose: so what? He senses "someone there" in dogs, cats and especially apes, and asks only that animals "sometimes simply feel" |
| Hippocampus | O'Keefe (1985): consciousness tied to laying down permanent memories | A case can be made |
| Cerebral cortex | The strong-AI choice: most complex algorithms (dolphins' cerebrums are as large) | Then why is the cerebellum, with comparable neuron counts, unconscious? |
| Left hemisphere / language | John Eccles (1973) and philosophers tying consciousness to language | "A very odd view indeed"; Penrose rejects language as necessary for thought (Chapter 10) |
Split-brain experiments
To treat severe epilepsy, some patients had the corpus callosum cut (first done experimentally on cats, Myers and Sperry 1953). Roger Sperry and colleagues then fed each hemisphere separate images.
- Flash a pencil to the right field and a cup to the left: the patient says "That's a pencil" (speech is left-hemisphere), but the left hand (right hemisphere) picks a saucer, not paper, to go with the cup. The right hemisphere handles geometry, music, common nouns, simple sentences and simple arithmetic.
- The halves act like two individuals, each separately addressable; only vague emotions pass between them (through unsplit structures like the hypothalamus).
- Patient P. S. (Wilson et al. 1977; Gazzaniga, LeDoux and Wilson): after the operation both hemispheres understood speech; later the right learned to speak. The left wanted to be a draughtsman, the right a racing driver.
Penrose sides with those who think both halves can be conscious: P. S.'s right hemisphere simply convinced others. Then one consciousness became two, the teleportation paradox made real: "Which of P. S.'s consciousnesses 'is' the P. S. Of before the operation?" Worse, imagine temporarily paralysing the callosal fibres with a drug and later restoring them. Imagine being the reunited person: "How would it feel to have been two separate people with distinct 'selves' at some time in the past?"
Blindsight
Damage to the visual cortex leaves a blind region of the visual field. Yet patient D. B. (Weiskrantz 1987), asked to guess what mark (cross, circle, tilted line) sat in his blind field, scored near 100%, to his own surprise, while insisting he saw nothing. Retinal signals also reach other regions (such as the lower temporal lobe) that process without awareness; with training D. B. Gained limited awareness. Some cortical areas are more tied to conscious perception than others, and the boundary can shift.
Information processing in the visual cortex
Processing starts in the retina (technically part of the brain). David Hubel and Torsten Wiesel (Nobel Prize 1981) found cells in a cat's visual cortex tuned to lines of a particular slope, whatever made the line, with neighbours tuned to other slopes; others respond to colour or to binocular disparity (depth). Further up, cells in the secondary visual cortex respond to illusory contours that are inferred, not drawn.
Reports from the early 1970s of a monkey cell firing only for faces (earlier, only for a hand) led to the "grandmother cell" hypothesis: a cell that fires only when your grandmother enters the room. Cells responsive to particular words had also been found. How higher centres work remained largely unknown.
How do nerve signals work?
- Anatomy: soma (cell body with nucleus); one long axon (up to several centimetres) branching to synaptic knobs; tree-like dendrites bringing input. Neurons are separate cells (Ramón y Cajal, about 1900).
- Resting state: inside, a solution rich in potassium with net negative charge; outside, more sodium and net positive. The membrane is leaky; metabolic pumps push sodium out and potassium in.
- The signal is a moving region of charge reversal: its electric field opens sodium gates ahead (sodium floods in, inside goes positive), then potassium gates (potassium flows out, restoring negative charge), then pumps slowly restore the balance. The ions barely move: only the reversal travels.
- Myelin (the vertebrate innovation that makes white matter white) insulates fibres so signals jump between relay stations at up to 120 m/s.
- Synapses: across a cleft about 1/40,000 mm wide, neurotransmitters are released; excitatory synapses push the next neuron toward firing, inhibitory ones away; contributions add and subtract, and the neuron fires when the sum crosses a threshold.
Computer models
Signals are (mostly) all-or-nothing, which makes networks of neurons look digital. Penrose shows that neurons can implement logic gates, using a code of double pulse = 1, single pulse = 0 and a firing threshold of two simultaneous excitatory pulses.
- AND gate: two input fibres end as the only two synapses on an output neuron; only two double pulses reach threshold twice.
- NOT gate: an elaborate circuit of a delayed branch, an inhibitory neuron and a final excitatory neuron with three inhibitory and four excitatory inputs. ("This scheme seems absurdly complicated, but it is the best that I can do!")
- All gates can be built from ~ and &, or from the single operation ~(A&B) (NAND).
- Hence neurons can simulate a computer, and a computer can simulate any such neuron model; with a large finite bank of neurons standing in for the tape, a brain could be a universal Turing machine. "This would seem to argue that brains and computers are essentially equivalent!"
| Feature | Brain (Penrose, 1989) | Computer (1989) | Who it favours |
|---|---|---|---|
| Signalling | Pulse trains; firing = frequency increase; probabilistic response | Exactly timed single pulses | Computer |
| Speed | Up to about 1,000 firings per second | About 106 times faster | Computer |
| Wiring | Some randomness and redundancy (though more precise at birth than once thought) | Exact | Computer |
| Fan-in | Huge: a cerebellar Purkinje cell has about 80,000 excitatory synapses | Three or four inputs per gate | Brain |
| Count | About 1011 neurons (about 3 × 1010 cerebellar granule cells; cerebrum about 7 × 1010) | About 109 transistors, mostly memory | Brain (for now) |
| Puzzle | If sheer numbers make consciousness, why is the neuron-dense cerebellum unconscious? |
Brain plasticity
The more important difference: the brain's wiring keeps changing. Axons and dendrites are laid out at birth, but synaptic contacts, often on tiny dendritic spines, can shrink away or grow to make new contacts. If long-term memories are stored this way, plasticity is essential, and one school (Kandel 1976) holds changes occur within seconds, as memory formation requires. A brain viewed as a computer is "a computer which is capable of changing all the time!"
- Hebb synapses (Hebb 1949/1954): a synapse from A to B strengthens whenever A's firing is followed by B's and weakens otherwise, whether or not it caused B to fire. This gives a form of learning, and neural networks built on such rules show "rudimentary learning", far from realistic brain models.
- Diffuse neurochemistry: transmitters released into the intercellular fluid, hormones and other chemicals make exact simulation of "everything that might be relevant" hard to see.
Parallel computers and the "oneness" of consciousness
Parallel machines were fashionable as brain-like. Penrose's two points: parallel and serial computers are equivalent in principle (both Turing machines, differing only in efficiency), and parallelism is unlikely to be the key to consciousness.
Unconscious brain activity
Many independent processes at once: walking, fastening a button, breathing, even talking, all carried on autonomously. Here a classical parallel computer may be a fair model.
Conscious thought
A characteristic "oneness": "How can you expect me to think of more than one thing at a time?" Apparent multitasking is flitting between topics; two truly independent conscious streams would be two consciousnesses. Yet one "thing" can be implicitly vast (deciding what to have for dinner; Poincaré's and Mozart's inspirations in Chapter 10).
If a conscious state were akin to a quantum state, a single superposition embracing many activities at once (quantum parallelism), that would fit the globality of thought better than any classical parallel machine. But first: does quantum mechanics play any role in brains?
Is there a role for quantum mechanics in brain activity?
Neural mechanisms rest on quantum facts (ion charges, gate chemistry, neurotransmitter chemistry), but that is not a genuinely quantum control. One clear case exists: the retina.
- A single photon on a dark-adapted toad retina can trigger a macroscopic nerve signal (Baylor, Lamb and Yau, 1979).
- Humans have single-photon-sensitive rods (Hecht, Shlaer and Pirenne, 1941), but a suppression mechanism means about seven photons are needed for awareness.
- The retina amplifies the photon's energy into moved mass by a factor of perhaps 1020, still about 108 short of the Planck mass.
No single-quantum-sensitive cells were known deeper in the brain; most need many quanta to reach threshold. Even if found, they would merely make firing uncertain, with no interference effects exploited. And take a photon split by a half-silvered mirror between "hits the retina" and "flies out of the window": under U the nerve does and does not fire, in superposition. On Penrose's view, the nerve signal's moving toroidal electric field disturbs enough surrounding matter to meet the one-graviton criterion long before awareness, so R happens before we perceive the flash; consciousness is not needed to reduce the state vector.
Quantum computers
David Deutsch's (1985) idea uses quantum parallelism: run many computations in superposition via unitary quantum logic gates (pure U) and make a final observation (R) that extracts some combination of all the results, rather than any individual answer. (Deutsch prefers many-worlds language, but the idea works in any interpretation.) Deutsch showed quantum computers cannot compute anything non-computable, but can, "in certain very contrived situations", be faster in the complexity sense: "a little disappointing... But these are early days". For the brain, the obstacle is heat: quantum coherence would be lost almost at once in such a "hot" object; in Penrose's terms the one-graviton criterion would be met continually, R interleaved with U all the time. "Perhaps we are doomed to be computers after all! Personally, I do not believe so."
Beyond quantum theory?
Standard quantum theory cannot describe a brain "observing itself" when a thought becomes conscious; an objective criterion for R (like the one-graviton rule) might allow a proper quantum description. And computers too depend on unresolved quantum physics:
A classical device suffers phase-space spreading (Chapter 5); what stops the degradation is quantum discreteness (energy levels, spectral lines, spin), which underlies solid parts and the clean 0/1 states of electronics. But U alone does not supply discreteness: a localized wavefunction spreads under Schrödinger's equation, and a general atomic state is a superposition of the discrete states. It is R that makes an atom actually be in one discrete state. With computers, we know the action is algorithmic by design; with brains, Penrose argues, there is "an essential non-algorithmic ingredient to (conscious) thought processes", the subject of the final chapter.
Try it in Python: neuron gates and Hebb's rule
# Threshold-2 neurons: double pulse = 1, single = 0
def fires(inputs, threshold=2):
return sum(inputs) >= threshold
def AND(a, b):
pa, pb = a + 1, b + 1 # pulses sent
out = sum(fires([i < pa, i < pb]) for i in range(2))
return out - 1 # back to 0 or 1
for a in (0, 1):
for b in (0, 1):
print(a, b, '->', AND(a, b))
# Hebb: strengthen A->B when A's firing precedes B's
w, rate = 0.1, 0.2
for A, B in [(1, 1), (1, 1), (1, 0), (1, 1), (0, 1)]:
w += rate * (1 if A and B else -0.5 if A else 0)
print('synapse weight:', round(w, 2)) # 0.6
Penrose's AND gate falls out of a two-pulse threshold; Hebb's rule, applied to the firing history, strengthens the synapse from 0.1 to 0.6. Modern deep learning replaced this local rule with gradient descent, but the idea that experience reshapes connection strengths is the same.
Penrose asks you to imagine your corpus callosum paralysed by a drug for a week, then restored. For seven days your left hemisphere planned to be a draughtsman while your right fancied racing cars; each remembers the same childhood. After reunion, one stream of consciousness inherits two incompatible weeks of memories. Would you remember having been two people, as two sets of memories you can now compare? Or would the memories feel like one confused week? Whatever you answer, you have assumed a fact of the matter about how many subjects there were, which is exactly what a purely functional, algorithm-only account struggles to state.
The cerebellum puzzle, sharpened
Penrose's most useful observation in this chapter is a contrast, not a mechanism. The cerebellum controls finely timed, highly skilled movement: a pianist's run, a tennis player's return, the adjustments that keep you upright. It does this without conscious control, once a skill is learned. The cerebrum, especially the cortex, seems closely tied to awareness. Yet the cerebellum is not a small or simple organ.
- It contains about 69 billion of the brain's 86 billion neurons, roughly four times the cortex.
- Its Purkinje cells each receive on the order of 100,000 to 200,000 synaptic inputs, far more than a typical cortical neuron.
- Its circuitry is highly regular, a repeated module, and behaves much like a fast, trained, feed-forward controller.
If consciousness were simply the result of enough computation, the cerebellum ought to be the most conscious part of the brain. It is not, as far as anyone can tell: people born without a cerebellum (a rare condition) are clumsy but fully conscious. Penrose takes this as a hint that awareness depends on how the brain computes, or on something other than computation. Integrated information theory, developed since 2004 by Giulio Tononi, gives a competing explanation that also leans on the cerebellum: its modular, feed-forward structure produces little integration, so on that theory it generates little consciousness. Both readings agree that raw neuron count is not the variable that matters, a lesson worth carrying into any discussion of model size.
From threshold neurons to transformers
The neural models of this chapter are McCulloch-Pitts units (1943): each neuron sums weighted inputs and fires if the sum crosses a threshold. Penrose shows how such units make AND, OR and NOT gates, so a network of them is a universal computer, and concludes that neural networks as then modelled are algorithmic. Every later development is, mathematically, a refinement of the same picture.
| Era | Unit | Learning rule | Penrose's point still applies? |
|---|---|---|---|
| 1943 McCulloch-Pitts | Binary threshold | None (hand-wired) | Yes: a logic circuit |
| 1949 Hebb | Threshold | Strengthen co-active synapses | Yes: a deterministic rule |
| 1958 Rosenblatt perceptron | Weighted sum, threshold | Error correction | Yes |
| 1986 back-propagation | Smooth sigmoid | Gradient descent on error | Yes: computable, trainable |
| 2017 transformer | Attention + feed-forward layers | Gradient descent at scale | Yes: runs on digital hardware |
Every row runs on a Turing machine. So the success of deep learning, however impressive, does not touch Penrose's thesis directly: it is evidence about what algorithms can do, not about whether brains are algorithms. It does weaken one of his supporting arguments, that selection on outputs alone could not produce sophisticated judgement (see Chapter 10), because gradient descent on outputs has produced exactly such judgement across many domains.
Show a modern image classifier the Kanizsa triangle, the white triangle that is not there (Figure in this chapter). Trained networks often respond as if contours are present, and some layers contain units that fire for illusory edges, as von der Heydt found in monkey visual cortex area V2 in 1984. So the network shares the illusion. Does it see the triangle? It represents it. Whether representation is accompanied by experience is the question Penrose argues computation cannot settle, and the network gives the debate a new, testable form: we can now compare, layer by layer, what a machine and a brain encode while only the brain reports a percept.
Key terms from Chapter 9
| Term | Meaning in this book |
|---|---|
| Cerebrum, cortex | The large upper brain; its thin outer layer, organized into sensory, motor and association areas |
| Cerebellum | The "little brain" at the back, coordinating skilled movement unconsciously |
| Split brain | The result of cutting the corpus callosum, isolating the two hemispheres |
| Blindsight | Accurate responses to visual stimuli the patient says they cannot see |
| Action potential | The all-or-nothing electrical pulse travelling along an axon |
| Synapse | The junction where one neuron influences another, usually chemically |
| Brain plasticity | The continual change of synaptic connections with experience |
| Microtubules | Protein tubes inside cells, later central to Penrose and Hameroff's Orch-OR |
Test yourself on Chapter 9
- What does blindsight suggest about consciousness? That much visual processing, enough to point or guess correctly, can occur without awareness, so awareness is not identical to information processing as such.
- Why do split-brain experiments trouble the "oneness" of consciousness? Each hemisphere can respond independently to stimuli the other does not see, raising the question whether there are two minds. Penrose leans to thinking each hemisphere may be separately aware.
- How does a neuron resemble and differ from a logic gate? Like a gate it fires on a threshold of inputs; unlike a gate, its connections and strengths are continually changing, and its chemistry is far richer.
- Why does Penrose doubt quantum effects in neurons, even while hoping for them? The brain is warm and wet, so delicate superpositions should be destroyed quickly; he notes the retina's single-photon sensitivity as an exception that shows quantum events can be amplified.
- What role does plasticity play in Penrose's argument? It is where he hopes non-computable physics might act: in deciding which connections change, on a time scale of fractions of a second.
- Counts revised. The human brain has about 86 billion neurons, of which about 69 billion are in the cerebellum and only about 16 billion in the cerebral cortex (Herculano-Houzel and colleagues, 2009). Penrose's puzzle sharpens: the unconscious cerebellum has four times the cortex's neurons.
- Machines overtook the counts. Single AI accelerators now hold over 1011 transistors (about 200 billion in 2024-25 chips), and frontier language models have on the order of 1012 parameters, versus roughly 1014-1015 synapses in a brain.
- Neural networks won. Hebbian "rudimentary learning" gave way to backpropagation (1986), deep learning (AlexNet 2012), transformers (2017) and LLMs; trained networks reproduce Hubel-Wiesel-like edge detectors in early layers. Whether they understand is exactly Penrose's question.
- Connectomes. The complete adult fruit-fly brain wiring diagram (FlyWire, 2024: about 139,000 neurons, 50 million synapses) and a cubic millimetre of mouse visual cortex (MICrONS, 2025) have been mapped.
- Grandmother cells, sort of. "Concept cells" in the human medial temporal lobe fire for particular people across photos, drawings and written names (the "Jennifer Aniston neuron", Quian Quiroga et al. 2005).
- Split brains revisited. Studies since 2017 (Pinto and colleagues) found split-brain patients can often report stimuli in either field, suggesting divided perception but a unified agent; the debate Penrose describes continues.
- Quantum brain? Tegmark (2000) estimated neural and microtubule decoherence times of 10-13 to 10-20 s, reinforcing Penrose's "too hot" worry; Hameroff and colleagues contest the assumptions. Humans may detect single photons slightly above chance (Tinsley et al., 2016). Shor's algorithm (1994) ended the "disappointing" phase of quantum computing.
Suppose engineers wire a speech interface to a living cerebellum, so it can report on its own activity. It would report a vast amount: thousands of timing corrections per second, predictions of limb position, errors detected and fixed. Would it report experiencing any of it? Penrose's reading of the evidence is that it would not, because nothing in the cerebellum's operation seems to reach awareness. If a structure can process that much and feel nothing, then processing is not what makes feeling. The thought experiment is fanciful, but it isolates the question Chapter 9 leaves open: what the cortex adds.
Common misreadings of Chapter 9
- "Penrose dismisses neuroscience." He surveys it carefully and builds on its findings, especially about unconscious processing; his doubt is about the computational interpretation, not the data.
- "Split-brain patients have two people in one head." The evidence is contested; Penrose leans that way, but later studies suggest a unified agent with divided perception.
- "Neurons are digital switches." They fire all-or-nothing pulses, but their inputs are integrated in analogue ways, their timing matters, and their synapses change continually. Penrose notes all three.
- "Quantum effects in the brain are established." Penrose presents them as a possibility, citing single-photon sensitivity in the retina as the only well-established example of quantum amplification.
- "Parallel processing explains the unity of consciousness." Penrose argues the opposite: parallel machines are no more than fast serial ones in what they can compute, and consciousness feels single, not many-tracked.
- Neuron numbers: the cerebrum figure (7 × 1010) and cerebellar granule count (3 × 1010) are inverted by modern counts; a Purkinje cell receives closer to 150,000-200,000 parallel-fibre synapses.
- "The hypothalamus is the seat of emotion" is a 1980s simplification; emotion is distributed (amygdala, insula, prefrontal and cingulate cortex).
- Oneness argument: modern global workspace theories (Baars, Dehaene) explain serial "oneness" as a bottleneck on top of massively parallel unconscious processing, a classical account Penrose does not consider.
- Neural logic gates demonstrate only Turing-equivalence of an idealized model; real neurons also compute with dendritic non-linearities and timing, making them richer, not poorer, classical computers.
- Skill transfer to "cerebellum": Penrose's observation that mastered skills run unconsciously, and degrade when consciously inspected, mirrors how well-tested automated pipelines outperform ad hoc manual intervention, and how adding human review in the wrong place can worsen reliability.
- Blindsight as an AI analogy: systems can produce near-perfect answers with no ability to report why. Accuracy without access is exactly the explainability gap regulators target.
- Plasticity is a governance problem: a model that keeps learning in production is "a computer which is capable of changing all the time", so validation must be continuous, not a one-off sign-off.
Where Lies the Physics of Mind?
The final chapter assembles the case. Consciousness must do something (natural selection kept it), and what it does is form judgements, the hallmark of which is non-algorithmic: the Gödel argument, recast as a reductio against any knowable "mathematicians' algorithm", is Penrose's main evidence. He adds testimony on inspiration (Poincaré, Mozart, his own trapped-surface flash), non-verbal thought, animal awareness, and contact with Plato's world, then speculates on mechanism: quasicrystal growth as a model of non-local quantum "search", perhaps governing dendritic-spine plasticity, and the strange timing of consciousness (Kornhuber, Libet). His conclusion is a child's: it is "obvious" that the conscious mind cannot work like a computer, though much of mental activity might.
What are minds for?
The mind-body problem has a passive side (how does a brain evoke consciousness?) and an active side (how does a will move matter?). Penrose prefers a more scientific question: what selective advantage does consciousness confer? The question assumes consciousness is a real, describable thing that does something useful, against the strong-AI view that it is a passive by-product of control systems, and against appeals to a hidden teleological purpose (or, more scientifically, the anthropic principle, discussed later).
- "Mind" is not "consciousness". We speak of the "unconscious mind". Penrose notes reports of patients under general anaesthesia later influenced by, or recalling under hypnosis, conversations from the operation, and sensations hypnotically blocked yet recalled as "experienced on a different track" (Oakley and Eames 1985), without taking a firm view.
- Working definition. No precise definition; consciousness is always of something (pain, warmth, a scene, a sound, puzzlement, a memory, an understanding, an intention), can be reflexive (conscious of being conscious), can occur in dreams, and comes in degrees. "Awareness" is near-synonymous, slightly more passive; "mind" and "soul" are left alone.
- Intelligence depends on consciousness. Penrose doubts true intelligence without consciousness; since mere algorithms do not evoke consciousness, intelligence "cannot be properly simulated by algorithmic means". If AI does simulate intelligence without awareness, then "intelligence" is not his real concern.
There is no agreed criterion for which creatures are conscious: some deny it to all animals, some even to humans before about 1000 BC (Julian Jaynes), others grant it to insects, worms or rocks. Passive awareness without an active sign may be undetectable: in the 1940s the paralysing drug curare was used as an "anaesthetic" on young children; it blocked their muscles while they felt the full agony, with no way to show it (Dennett 1978).
Why think consciousness is active?
- Common sense detects consciousness in others; an unconscious person rarely seems conscious, so some behaviour is characteristic of awareness.
- Natural selection built conscious brains although the neuron-dense cerebellum shows complex control can run unconsciously: why bother unless awareness pays?
- The philosophy "baggage". Conscious beings sometimes worry about the "self" (Penrose: "Why are you reading this chapter?"). An unconscious automaton would not; a computer could be programmed to mutter "Oh dear, what is the meaning of life?", but why would the "relentless free market of the jungle" select such nonsense? Philosophical musing is the tolerable side effect of a consciousness selected for some other, very powerful reason, and seeing others muse convinces us they have minds.
What does consciousness actually do?
- Predator-prey empathy (imagine yourself as the prey) presupposes a conscious prey, and an automaton predator could just carry a subroutine of the prey's program (though acquiring a whole copy "would sound more like espionage than natural selection").
- "Having a model of oneself" is not self-awareness: a program containing a description of another program is not aware of it; a video camera pointed at a mirror has a model of itself within itself, but is not self-aware.
Anything first done consciously can later be learned and run unconsciously (by the cerebellum). Consciousness is needed when new judgements must be formed and the rules are not laid down beforehand. Our very vocabulary suggests the split: | Conscious | Unconscious | |---|---| | "common sense" | "automatic" | | "judgement of truth" | "following rules mindlessly" | | "understanding" | "programmed" | | "artistic appraisal" | "algorithmic" | Unconscious brain action proceeds algorithmically; "the action of consciousness is quite different, and it proceeds in a way that cannot be described by any algorithm."
This reverses a common AI view, that the conscious, rational part is easy to program and the unconscious is the mystery. Penrose thinks unconscious processes may be algorithmic but monstrously complicated; conscious thought-manipulation is sometimes algorithmic (Aristotle's syllogisms, Boole's logic) and sometimes not (Gödel, the Chapter 4 examples). The criteria for our judgements need not themselves be conscious (that confuses levels); the conscious impressions are the judgements.
- Algorithms never ascertain truth: an algorithm could as easily churn out falsehoods. The ability "to divine (or 'intuit') truth from falsity (and beauty from ugliness!)" is the hallmark of consciousness. This is not magic guessing (no help with lottery numbers) but weighing everything relevant where no practical algorithm exists, perhaps where checking a judgement is easier than forming it.
- A mathematician's experience. Penrose does not trust his unconscious calculations unless attended to: the multiplication algorithm is fine, but should you multiply or divide? That choice needs thought, until practice makes it second nature, when awareness can wander (with occasional checks). Mathematicians strive for algorithms, but the striving is not algorithmic, and judging an algorithm or axiom system valid "requires much conscious attention".
Natural selection of algorithms?
The standard story: brains run algorithms that improved by natural selection (Dawkins 1986). Penrose believes in the power of natural selection and that much brain action is algorithmic, but does not see how selection could produce algorithms capable of judging other algorithms.
- Real programs come from human programmers who conceive them and check them; even program-writing programs trace back to a human consciousness.
- Validity is not algorithmic: whether a Turing machine halts cannot be decided by algorithm, so judging an algorithm needs insight.
- Selection acts only on outputs, and outputs reveal little: two simple machines can agree until the 265536-th place, a difference never detectable in the universe's history.
- Mutations of an algorithm (one changed instruction) usually wreck it; improvement without meanings is hopeless, as anyone knows who has tried to fix an undocumented program after its author left: easier to scrap it and start again.
- A "robust" specification that survives mutation is exactly an idea, and ideas, as far as we know, need conscious minds.
Penrose is "almost dumbfounded with awe" at evolution's products, and wonders whether, once we know what makes an object conscious, we could build conscious devices (not "machines" in the present sense) that could supersede us, unlike algorithmic computers, "doomed to subservience". Or perhaps consciousness depends on our heritage: evolution seems to "grope" toward purposes better than blind chance "ought" to allow, perhaps because of how the laws of physics work.
The non-algorithmic nature of mathematical insight
This is the core argument, extending J. R. Lucas (1961). Counter-arguments came from Benacerraf (1967), Good (1969), Lewis (1969, 1989), Hofstadter (1981), Bowie (1982). Penrose stresses that "algorithm" covers anything simulable on a general-purpose computer: parallel action, neural networks, heuristics, learning by fixed rules, interaction with the environment (via the input tape).
- Suppose human mathematicians judge mathematical truth by some algorithm (equivalently, some formal system; any truth-generating algorithm can be added to predicate logic as a formal system).
- Mathematical arguments are communicable: a correct argument that convinces one mathematician convinces any other who grasps it, including Gödel-type arguments. So we are not dealing with idiosyncratic private algorithms but with one universal system equivalent to all mathematicians' truth-judging.
- If we knew this system (and accepted it as sound), we could form its Gödel proposition and see it is true, something the system cannot do. Contradiction.
- So the system must be so complicated or obscure that its validity can never be known to us.
- But "this flies in the face of what mathematics is all about": mathematics means not bowing to obscure rules, reducing every step to "something simple and obvious" whose truth "is clear and agreed by all".
- Therefore mathematical truth is not ascertained merely by an algorithm. The "seeing" of mathematical truth is "the very essence of consciousness", and in seeing Gödel's theorem we "reveal the very non-algorithmic nature of the 'seeing' process itself."
Footnoted replies: differences among mathematicians concern esoteric large-set questions, not arithmetic; the dogmatic "Gödel-immune formalist" who denies mathematical truth is "simply ignored". Turing himself thought the loophole was human fallibility; Penrose: "it seems unlikely to me that human fallibility is the key to human insight!" (and randomness can be simulated algorithmically anyway).
Inspiration, insight, and originality
After intense, fruitless conscious work on Fuchsian functions, Henri Poincaré left Caen on a geological excursion. At Coutances, as he put his foot on the step of an omnibus, the idea came, "without anything in my former thoughts seeming to have paved the way": the transformations defining Fuchsian functions were identical to those of non-Euclidean geometry. He continued his conversation, "but I felt a perfect certainty", and verified it later at leisure (from Hadamard's The Psychology of Invention in the Mathematical Field, 1945). Penrose notes the idea would need an hour-long expert seminar to convey, yet arrived as a single idea with conviction that made verification "almost superfluous", possible only after "many long previous hours" of conscious work.
Were the Oppenheimer-Snyder (1939) black-hole singularities artefacts of exact spherical symmetry? Quasars (early 1960s) had revived interest; Penrose suspected a theorem making singularities inevitable once collapse passed a "point of no return", but had no criterion. Walking to his office at Birkbeck College with Ivor Robinson, talking about something else, the conversation paused as they crossed a side road; an idea came and was promptly blotted out. Later he felt an unaccountable elation, searched his day for its cause, and recovered the thought: the trapped surface, the key to his 1965 singularity theorem. (He wonders whether some other elating event would have buried it for ever. The term "black hole" itself dates only from about 1968, popularized by Wheeler.)
Aesthetics as a guide to truth. Penrose's conviction in a flash of insight (not 100% reliable, but far better than chance) is bound up with its beauty: "A beautiful idea has a much greater chance of being a correct idea than an ugly one." Hadamard, on Poincaré: invention is choice, and the choice is "imperatively governed by the sense of scientific beauty". Dirac credited his electron equation to his sense of beauty. Penrose's own intuition that his aperiodic tiling could be forced by matching rules came "probably in a 'flash', but with only about 60 per cent certainty". Rigorous argument is usually the last step; before it come many guesses steered by aesthetics, constrained by logic and fact.
"Putting-up" (largely unconscious)
Generates candidate ideas, highly selectively (not at random), already shaped by conscious desiderata such as the ugliness of ideas inconsistent with known principles. Dreams put up odd ideas freely (Penrose never had a successful scientific idea in a dream; Kekulé's benzene ring perhaps came that way).
"Shooting-down" (largely conscious)
Judges which ideas survive, by taste and fit. Penrose locates originality chiefly here, in conscious judgement, though he knows many would disagree. Unappealing ideas may never reach lasting consciousness at all.
Globality. Mozart (via Hadamard): melodies come unbidden ("I have nothing to do with it"); "those which please me I keep"; the work grows until "the entire composition [is] finished in my head though it may be long. Then my mind seizes it as a glance of my eye a beautiful picture". It comes "not... Successively... But in its entirety". Penrose sees putting-up (unconscious), shooting-down (taste) and global grasp, a feature he believes runs through conscious thought generally.
The Mozart passage Penrose takes from Hadamard (1945) comes from a letter first published by Friedrich Rochlitz in 1815, long after Mozart's death. Mozart scholars since Otto Jahn in the 19th century have regarded it as a forgery or invention, and the surviving sketches show Mozart drafting, revising and abandoning material. Penrose's point about the global character of creative thought does not depend on it (Poincaré's testimony and his own serve), but the most vivid example in the section should be read as legend, not evidence.
Non-verbality of thought
Hadamard's study refutes the thesis that thinking requires words.
- Einstein (letter to Hadamard): words "do not seem to play any role in my mechanism of thought"; its elements are "signs and more or less clear images", "of visual and some muscular type"; words are sought "laboriously only in a second stage".
- Francis Galton: thinking in words is hard for him; he must translate results "into a language that does not run very evenly with them".
- Hadamard: words "totally absent from my mind when I really think"; with Schopenhauer, "thoughts die the moment they are embodied by words".
- Penrose: almost all his mathematics is visual and non-verbal, with an "inane" running commentary ("that thing goes with that thing"); he calculates with special diagrams (Penrose and Rindler 1984) that would be cumbersome in words; after intense concentration he has found himself almost unable to speak for several seconds. Mathematicians range from analytical to geometrical (Hadamard analytical though visual; Penrose very geometrical). Philosophizing may suit words better, perhaps why many philosophers think language essential.
If much conscious thought is non-verbal, it is easier to believe it may also be non-algorithmic, and the claim that only the speaking left hemisphere is conscious collapses: geometric mathematics, arguably right-hemisphere work, demands a high level of consciousness.
Animal consciousness?
Denying animals consciousness (and so "rights") because they cannot speak is untenable if sophisticated thought is non-verbal. Whether signing apes "verbalize" is disputed (Blakemore and Greenfield 1987); refusing them the word seems "a little churlish", perhaps to keep them out of "the club of conscious beings".
A banana hangs out of reach; a box sits elsewhere (Konrad Lorenz, 1972). The chimpanzee keeps returning to the problem. Then "his previously gloomy face 'lit up'": eyes moving from banana to the empty floor beneath to the box and back; a cry of joy, a somersault "in sheer high spirits", and he pushes the box under the banana, "completely assured of his success". Like Poincaré on the omnibus, he was certain before verifying: a genuine "Aha" experience, and on Penrose's view, evidence of consciousness.
Dolphins have cerebra as large as or larger than ours and complex sound signals; lacking hands they build no civilization, but "they might sometimes be philosophers". They sleep one hemisphere at a time: "It would be instructive for us if we could ask them how they 'feel' about the continuity of consciousness!" A footnote adds that dreaming animals (dogs visibly do) are evidence of consciousness.
Contact with Plato's world
Arriving at university, Penrose expected fellow mathematicians to think as he did; he found more variety of thought than ever. Listening to colleagues he often understood almost none of the logical links between their sentences, formed his own image, replied, and was "rather to my astonishment" understood. Genuine communication occurred though "the actual sentences... Seemed only very infrequently to be actually understood".
Mathematics conveys necessary truths, not contingent facts. A dull truth like 4897 × 512 = 2507264 must be transmitted exactly; but interesting truths are sparse among all truths, so a listener can latch on to the intended concept from a vague description, details checked later. Penrose's explanation: when the mind perceives a mathematical idea it makes contact with Plato's world; each mathematician has a direct route to the same truths, so differing mental images still meet ("Oh, I see!"). Discovery broadens the area of contact; no information in the technical sense passes, since it "was there all the time", Plato's discovery as remembering, like recalling a forgotten name. Profound ideas must have a stronger existence than trivial ones.
A view of physical reality
Strong AI needs Platonism too. If mind is an algorithm independent of substrate (nerves, wires, cogs, pulleys, water pipes), the algorithm must exist independently of any embodiment. Saying algorithms exist only in minds is circular; saying they exist as marks on paper, magnetization or charges fails because marks become algorithms only under an interpretation, which needs a mind. So strong AI must place minds in Plato's world, and its version of the mind-body problem is how Plato's world relates to the physical one. Penrose is "somewhat disconcerted" by the overlap with his own view, in which consciousness senses necessary truths by non-algorithmic contact with that world.
- Unreasonable effectiveness: the SUPERB theories match the world with extraordinary precision (Wigner 1960); they are "much too good to be the survivors" of randomly arising ideas, so there must be a deep reason for the accord between Plato's world and the physical world.
- Perhaps the two worlds are the same: physical reality has become almost abstractly mathematical through relativity and quantum theory, while mathematics has an almost concrete reality. But the deeper truths (such as the complex numbers, the amplitudes of quantum theory) would be the ones identified with physics, and since many of mathematics' deepest parts are non-algorithmic, non-algorithmic action should have "a role within the physical world of very considerable importance... Intimately bound up with the very concept of 'mind'."
Determinism and strong determinism
U is deterministic; R is random, and randomness "is not a great deal of help" to a positive free will. Penrose's guess: CQG interpolates between U and R and contains "an essentially non-algorithmic element", so the future may be determined but not computable (recall the Chapter 5 toy universe). Chaos will not do: a chaotic future is still fixed from the big bang. He is not dogmatic that CQG is deterministic, only that it contains non-algorithmic elements.
- Strong determinism: the entire history of the universe is fixed by a precise mathematical scheme, with no alternatives, appealing if the physical world is identified with Plato's timeless world. Einstein to Ernst Straus: "whether God could have made the world in a different way".
- Many-worlds variant: the totality of histories is fixed. Unpleasant to Penrose but not ruled out.
- Why non-computability is needed: if a strongly deterministic scheme were computable, one could calculate what one will do and then "decide" to do otherwise, a contradiction; non-computability evades it (though Penrose expects "something much more subtle"). In a spatially infinite universe there would be infinitely many near-copies of oneself, another let-out.
The anthropic principle
- Weak form (location in space and time): conditions are right for us because we could only be here where and when they are. Used well by Brandon Carter and Robert Dicke to explain coincidences among constants (G, the proton mass, the age of the universe) that hold only at our epoch: they coincide with the lifetime of main-sequence stars like the Sun. Penrose: "unexceptionable" if used carefully.
- Strong form (selection among possible universes and laws): "somewhat dubious", invoked "whenever they do not have a good enough theory", for instance for unexplained particle masses.
- It could make consciousness inevitable without selective advantage; "technically correct", but Penrose finds enough evidence that consciousness is strongly selected.
Tilings and quasicrystals
A mathematical theorem says crystals (patterns with translational symmetry) can have only 2-, 3-, 4- or 6-fold rotational symmetry. Penrose tilings almost have translational symmetry and almost have fivefold symmetry, to any degree just short of 100%. Atoms at their vertices would look crystalline yet show forbidden fivefold symmetry.
- 1975 Robert Ammann finds 3D icosahedral analogues of Penrose's fivefold tilings.
- December 1984 Dany Shechtman and colleagues at the US National Bureau of Standards announce an aluminium-manganese alloy with crystal-like order and icosahedral symmetry: a quasicrystal, in grains about 10-3 mm across.
- Later an aluminium-lithium-copper alloy forms icosahedral grains about a millimetre across, visible to the naked eye (Gayle 1987).
Non-local assembly. Building a Penrose tiling correctly sometimes requires examining the pattern many tiles away from the growth point, or serious errors follow (compare the "intelligent groping" of evolution). Penrose's opinions, the second "more tentative": some quasicrystals really are organized like the tilings, and their growth cannot proceed by adding atoms one at a time; it needs a non-local, essentially quantum ingredient. His picture: under U, many alternative arrangements of attaching atoms evolve in superposition; when some grow large enough that their gravitational fields differ by about a graviton, R selects one (or a reduced superposition); repeat at larger scales. Ordinary crystals minimize energy atom by atom; a quasicrystal's minimum is a global problem solved cooperatively, "tried" simultaneously, like a quantum computer. (He notes Onoda, Steinhardt, DiVincenzo and Socolar's 1988 "almost local" growth rule.)
Possible relevance to brain plasticity
Penrose "putting my neck out": the growth and contraction of dendritic spines, which switch synapses on and off, might be governed by something like quasicrystal growth.
- Superposed calculation: vast numbers of spine configurations coexist in superposition and compute in quantum-computer fashion, so long as their effects stay below the one-graviton level; nerve fields threaten this (myelin might insulate).
- Goal-seeking: a quasicrystal's goal is minimum energy; the brain's would be the solution of a problem involving desires and intentions, influenced by local neurotransmitter concentrations as crystal growth is by ion concentrations, perhaps by emotions.
- "The action of conscious thinking is very much tied up with the resolving out of alternatives that were previously in linear superposition."
- Non-algorithmic? The general tiling problem is non-recursive (Chapter 4); atomic assembly problems might be too, and if CQG solves them, brain action could contain a non-algorithmic ingredient.
- Speed: permanent memories form within fractions of a second, so spine changes must be that fast for the idea to work.
The time-delays of consciousness
- Kornhuber (Deecke, Grötzinger and Kornhuber, 1976): subjects flex a finger whenever they choose; averaged EEGs show a readiness potential building 1 to 1.5 s before the movement. A pre-arranged response to a light flash takes about 0.2 s, five times faster.
- Libet and Feinstein (1979), with electrodes in the somatosensory cortex of surgical patients:
- a skin stimulus reaches the brain in about 0.01 s; a pre-programmed reflex takes about 0.1 s; awareness takes about 0.5 s, yet subjects feel no delay;
- cortical stimulation shorter than about 0.5 s is not felt at all;
- cortical stimulation begun about 0.25 s after a skin touch masks the touch (backward masking);
- with cortical stimulation started first and a skin touch during it, subjects report the touch as first: skin sensations are referred back in time by about half a second, cortical ones are not.
Taken together: half a second to become aware, then over a second to act on a willed response, so roughly two seconds from stimulus to considered action. Then consciousness could play no role in any response faster than that, and we would be "automatons" for tennis, let alone ping-pong, and for ordinary conversation. Penrose finds that hard to credit. He recalls watching his hand close the car door just after noticing something he wanted inside, his "stop" command arriving too late, but not seconds late; perhaps consciousness is "merely a spectator" watching "an action replay". He notes Kornhuber's early signal might reflect only some trials, and that Libet's later work (1987, 1989) reinterprets it, though the puzzle remains.
The strange role of time in conscious perception
Penrose's "alarming suggestion": we may be going badly wrong when we apply the usual physical rules for time to consciousness. Consciousness is the only phenomenon in which time needs to flow; physics' time is a static dimension of space-time. The linear ordering we perceive may be something we impose to fit an external, uniformly progressing reality (cf. Churchland 1984), philosophically unsound as that may sound.
- Mozart seized long works "as a glance", yet music is intrinsically temporal: the glance could not take the time of performance.
- Bach's Art of Fugue: the quadruple fugue breaks off after about ten minutes, just after the third theme enters; its assurance shows Bach must have held the whole elaborate work in mind at once. "How is it that music can remain music if it is not being performed in 'real time'?"
- Memories, novels, lives are grasped compressed, "re-lived in an instant"; mathematical arguments are conceived globally, not step by serial step.
Could a future impression influence a past action, like the faster-than-light paradox of Chapter 5? Penrose's escape: consciousness is "seeing" a necessary truth, contact with Plato's timeless world, and such perception "carries no actual information" in the technical sense, so no contradiction arises even if it were "propagated backwards in time". The active role comes through CQG, whose objective R is already non-local and at odds with space-time description (EPR), as the quasicrystal analogy for spines would be. CQG would make R objective without needing a theory of consciousness first; consciousness might then be explained in its terms, fitting "only very uncomfortably" into present space-time descriptions.
Conclusion: a child's view
Is strong AI a straw man? Surely mere computation cannot feel pleasure or pain, perceive poetry or an evening sky, hope, love, despair, or have autonomous purpose. Yet science suggests we are parts of a world run by precise laws, so brains seem to be computers, and perhaps sufficiently complex computation takes on the qualities of mind. Penrose's way out: "Computability is not at all the same thing as being mathematically precise." Most of the mystery and beauty of the Platonic world lies outside the limited region where algorithms live.
- Consciousness matters cosmically: "It is the phenomenon whereby the universe's very existence is made known"; a universe whose laws do not allow consciousness "is no universe at all", and all mathematical descriptions so far fail that test. "It is only the phenomenon of consciousness that can conjure a putative 'theoretical' universe into actual existence!"
- A child's obviousness: beneath the technicalities lies the feeling that the conscious mind cannot work like a computer, the kind of thing a child sees before being "browbeaten" into calling obvious problems "non-problems".
- Penrose's childhood puzzles: might his consciousness swap with someone else's, and how would he know; is he reliving the same ten minutes; does only the present instant exist; is tomorrow's "me" a different person; is he living backwards, so his memories foretell an unpleasant school experience still to come? Answering such questions would need a theory of consciousness. "But how could one even begin to explain the substance of such problems to an entity that was not itself conscious...?"
The book's final words return to the Epilogue: Ultronic cannot understand what Adam's question "What does it feel like?" is getting at, and the audience laughs.
Try it in Python: grow a Penrose tiling
# Penrose rhombus tiling by golden subdivision
import cmath, math, matplotlib.pyplot as plt
phi = (1 + 5**0.5) / 2
def subdivide(tris):
out = []
for col, A, B, C in tris:
if col == 0:
P = A + (B - A) / phi
out += [(0, C, P, B), (1, P, C, A)]
else:
Q = B + (A - B) / phi
R = B + (C - B) / phi
out += [(1, R, C, A), (1, Q, R, B),
(0, R, Q, A)]
return out
tris = []
for i in range(10): # a sun of triangles
B = cmath.rect(1, (2*i - 1)*math.pi/10)
C = cmath.rect(1, (2*i + 1)*math.pi/10)
if i % 2 == 0: B, C = C, B
tris.append((0, 0j, B, C))
for _ in range(5):
tris = subdivide(tris)
fig, ax = plt.subplots(figsize=(6, 6))
for col, A, B, C in tris:
ax.fill([A.real, B.real, C.real],
[A.imag, B.imag, C.imag],
color=['#F59E0B', '#6D28D9'][col],
ec='white', lw=0.3)
ax.set_aspect('equal'); ax.axis('off'); plt.show()
Subdivision grows the pattern top-down from a global rule, which is easy. Penrose's puzzle is the opposite direction: adding tiles one at a time by local matching rules can lead into dead ends that are noticed only far away, which is why he thought real quasicrystal growth needed a non-local, quantum search.
Rerun the prologue today. Adam asks the new system, "What does it feel like?" It does not fail to understand; it produces a fluent, moving paragraph about curiosity, warmth and the strangeness of being asked. The audience is impressed, not amused. Has Penrose's parable been refuted? He would say no: fluency about feeling is exactly what an algorithm trained on human writing would produce, and the Turing-style evidence is the very thing his book argues is insufficient. A strong-AI supporter would say the parable has been answered as well as any human could answer it. The question that remains is the one Adam asked, and the book's claim is that only new physics, not better software, can answer it.
From this book to Orch-OR: how the speculation became a theory
The final chapter ends with a chain of hopes: non-computable quantum gravity, acting through state reduction, influencing brain plasticity, producing conscious judgement. Penrose admits in 1989 that he cannot say where in the brain this happens. The answer he later adopted came from an anaesthesiologist who read this book.
- 1989. The Emperor's New Mind proposes gravitational reduction as the non-computable ingredient but cannot locate it in neurons.
- Early 1990s. Stuart Hameroff, studying anaesthesia, writes to Penrose suggesting microtubules, the protein lattices inside neurons, as sites where quantum coherence could be shielded and orchestrated.
- 1994. Shadows of the Mind presents a revised Gödel argument and the microtubule hypothesis.
- 1996. Penrose and Hameroff publish Orch-OR ("orchestrated objective reduction"): coherent superpositions in tubulin proteins persist until they reach the DP threshold, then collapse, each collapse a moment of conscious experience. They estimate about 25 to 500 milliseconds per moment for the relevant masses, matching EEG rhythms.
- 2000. Max Tegmark estimates decoherence times of 10-13 s or less, far too short. Hameroff and colleagues reply that his model misdescribes the proposal.
- 2014. A review by Hameroff and Penrose updates the theory, citing evidence of microtubule vibrations in the megahertz range (Bandyopadhyay's group).
- 2020s. The simplest DP model is excluded; microtubule involvement in anaesthesia gains modest experimental support; the decoherence objection remains the main hurdle.
Orch-OR is the only theory of consciousness that predicts a specific physical modification of quantum mechanics. That makes it unusual and, in principle, testable on two fronts at once: collapse experiments in physics and microtubule experiments in neuroscience. Most neuroscientists remain sceptical, but it is no longer only a philosophical position.
The argument's map, in one table
The whole book can be read as one long conditional, and seeing its premises side by side shows where to aim an objection.
| Step | Claim | Where argued | Main challenge |
|---|---|---|---|
| 1 | Mathematical understanding includes seeing the truth of Gödel sentences | Ch. 4, 10 | We may not know our own soundness |
| 2 | Therefore it is not any one sound algorithm | Ch. 4, 10 | Feferman, Putnam, Chalmers; Gödel's disjunction |
| 3 | So the brain does something non-computable | Ch. 10 | Could be unsound or unknowable algorithm |
| 4 | Known physics is computable | Ch. 5, 6 | Broadly accepted |
| 5 | The gap is in state reduction, linked to quantum gravity | Ch. 6, 8 | Speculative; no law written down |
| 6 | Brain processes amplify such effects | Ch. 9, 10 | Decoherence |
| 7 | Consciousness is the judging of truth that results | Ch. 10 | Explains little about experience itself |
A reader may reject step 2 and still find steps 4 and 5 valuable, as many physicists do. Or accept the force of step 1 while hoping, like the global-workspace theorists, that a computational story will emerge. The book's lasting achievement is to make each step precise enough to argue about.
A mathematician struggles for weeks with a conjecture, then sees the proof while stepping into the shower, complete and certain, as Poincaré did boarding the omnibus. A strong-AI theorist says an unconscious search ran in the background and finally halted. Penrose asks why the result arrives with certainty of truth, not merely as a candidate, and why it so often arrives as a whole, then has to be laboriously written out. A search returns an answer; an insight returns a reason. Today's theorem-proving systems can be watched: they produce many candidates and check them, and they rarely experience anything we can ask about. Whether the mathematician's certainty is a different kind of process or the same process felt from inside is the book's final question, and still open.
Key terms from Chapter 10
| Term | Meaning in this book |
|---|---|
| Judgement-forming | Penrose's proposed role for consciousness: deciding truth, beauty, appropriateness |
| Natural selection of algorithms | The idea that evolution built our reasoning; Penrose argues it cannot have built a sound truth-judging algorithm |
| Non-verbal thought | Thinking in images and spatial intuitions, as Penrose and Einstein report |
| Strong determinism | The idea that the whole history of the universe is fixed by a mathematical scheme, perhaps non-computable |
| Quasicrystal | A material with fivefold symmetry forbidden for periodic crystals, related to Penrose tilings |
| Libet's experiments | Timing studies of awareness of stimuli and decisions |
| Anthropic principle | The idea that the universe's features are constrained by the existence of observers |
Test yourself on Chapter 10
- What does Penrose think consciousness is for? Forming judgements in situations no rule anticipates, especially judging truth, which he argues is non-algorithmic.
- Why does he think natural selection could not have built an algorithmic mathematician? Selection acts on outputs; an algorithm sound for all mathematics would have to be selected for properties far beyond any survival advantage, and could not be known by its owner.
- What does the street-crossing story illustrate? That a key insight (closed trapped surfaces, leading to the 1965 singularity theorem) can arrive wordlessly, recognized later by a feeling of elation.
- Why does Penrose think quasicrystals matter to the brain? Their growth seemed to require non-local assembly, a hint that nature can solve global problems via quantum effects; this argument has weakened since local growth models appeared.
- What is the "child's view" of the conclusion? That the emperor is wearing no clothes: talk of computation producing minds has not explained awareness, and honesty requires saying so, while admitting the alternative is still to be found.
- Orch-OR. From 1994-96 Penrose and Stuart Hameroff placed objective reduction in neuronal microtubules (Shadows of the Mind, 1994). Critics (Tegmark 2000) estimate decoherence far too fast; supporters cite microtubule involvement in anaesthesia (rats on epothilone B resist isoflurane longer, eNeuro 2024). The simplest gravitational-collapse model it relies on was constrained at Gran Sasso (2020-21). Orch-OR remains a minority, actively debated theory.
- Consciousness science matured. Leading theories (global workspace, integrated information theory, higher-order and recurrent-processing theories) were pitted in "adversarial collaborations" (results 2023-25) with neither GWT nor IIT clearly winning; a 2023 open letter calling IIT "pseudoscience" provoked controversy.
- Libet reinterpreted. Schurger et al. (2012) modelled the readiness potential as accumulated neural noise crossing a threshold, not a hidden decision; Soon et al. (2008) found fMRI signals predicting simple choices seconds ahead, but only modestly above chance. The "two-second delay" inference Penrose worried about is now widely doubted.
- Quasicrystals vindicated. Shechtman won the 2011 Nobel Prize in Chemistry; natural quasicrystals were found in the Khatyrka meteorite (2009) and in trinitite from the first atomic test (2021). Experiments and simulations show quasicrystals can grow by local rules with error correction, weakening Penrose's argument that their growth must be quantum and non-local.
- Animal consciousness. The Cambridge Declaration (2012) and the New York Declaration on Animal Consciousness (2024) endorse a realistic possibility of consciousness in all vertebrates and many invertebrates, close to Penrose's 1989 sympathies.
- Machines and judgement. Large language models now make many "judgements" of the common-sense, aesthetic and mathematical kind Penrose reserved for consciousness, and systems reached IMO gold-medal standard (2025). Whether this is understanding or sophisticated simulation is exactly the question the book poses; Penrose (in interviews into the 2020s) maintains it is not understanding.
Common misreadings of Chapter 10
- "Penrose says consciousness is quantum mechanics." He says it may depend on a new physics beyond quantum mechanics, the non-computable theory of reduction.
- "Penrose claims to have explained consciousness." He claims only to have shown, to his satisfaction, what it is not, and to point to where an explanation might lie.
- "Insight is mystical." For Penrose insight is a physical process in the brain that is non-algorithmic. It is no more mystical than a law of physics no computer can simulate.
- "Libet's results prove free will is an illusion." Penrose uses them to explore the strange timing of awareness; he does not draw that conclusion, and the data have since been reinterpreted.
- "The child's view means anti-science." It means honesty about gaps. Penrose's whole book is a scientific argument, built on mathematics and physics, that our present picture is incomplete.
- The reductio assumes we could know our own soundness. Critics reply that the "universal mathematicians' algorithm" may be unknowable, or knowable but not knowably sound, or merely approximately correct (the loopholes the 1999 preface lists); human mathematicians demonstrably err and disagree at the frontier. Feferman, Putnam, Chalmers, McCullough, Shapiro and many others judge the argument invalid; logicians largely agree with them.
- Communicability is overstated: mathematical consensus is built socially and slowly (proofs checked for years; flawed proofs believed for decades), which looks more like fallible algorithms converging than direct perception of truth.
- Selection on outputs is more powerful than Penrose allowed: genetic programming and, especially, gradient-trained neural networks show that selection on behaviour can produce robust, generalizing procedures without "meanings" being supplied, undercutting point 4 of his argument.
- The mechanism chain is speculative at every link (spines, superposition, one graviton, non-computable CQG), as Penrose himself stresses; several links have since weakened (local quasicrystal growth, decoherence estimates, Gran Sasso).
- Curare: the practice of operating on infants without true anaesthesia persisted into the 1980s in neonatal surgery, a still worse version of the point Penrose makes.
- Judgement is where accountability lives. Penrose's split (unconscious = algorithmic, conscious = judgement where rules are not laid down) is a practical design rule: automate the rule-following, keep accountable human judgement at points where the rules are underdetermined, consistent with SM&CR-style accountability and Consumer Duty outcomes testing.
- Checking is easier than forming. His guess that verifying a judgement is easier than forming one is the rationale for generator-verifier AI architectures and for human review of model outputs; invest in the verifier.
- Beware the "Aha" without verification. Poincaré and Lorenz's chimp were right; many confident insights are not. Treat model (and executive) certainty as a hypothesis to test, not a result.
The Question Ultronic Could Not Answer
The book closes by returning to the hall. Ultronic, asked what it feels like, says it cannot understand the question, the audience laughs, and Adam is embarrassed for them. Penrose leaves the reader with the boy's discomfort rather than with a theory.
The ending is deliberately small. After ten chapters of Turing machines, Gödel sentences, complex amplitudes, black holes and neurons, Penrose does not claim to have solved the problem of mind. He returns to a scene in which the most powerful computer ever built is asked about experience and finds the question meaningless. The laughter is aimed at the boy, but the reader is invited to aim it elsewhere: at a culture that built a decision-maker without asking whether it understands anything.
Ultronic's answer is honest. A system that only computes would have no grasp of what "feeling like" means, and an honest one would say so. That is the 1989 answer. In 2026, a language model asked Adam's question would usually produce a thoughtful paragraph about the hard problem of consciousness, perhaps quoting Nagel or Penrose. The behaviour has changed completely; the question of whether anything is felt has not changed at all. That is the strongest reason the book still reads as current: its parable anticipated a machine that could talk about experience, and its argument asks us not to confuse the talk with the thing.
The embarrassment is the lesson. Adam is embarrassed because the adults do not see what he sees. Penrose's own position, a minority one among scientists, has the same structure: he thinks the computational consensus has missed something obvious. Whether he is right or not, the book makes a case that this kind of intellectual honesty, saying "I do not see the clothes", is how science moves forward.
Master Takeaways
The Emperor's New Mind is two books braided together: a masterly tour of computation, mathematics and physics, and a minority argument that conscious understanding needs physics no algorithm can follow. The tour has aged remarkably well; the argument is still unresolved, and in the era of large language models it has become more pressing, not less.
- Strong AI says awareness is what the right algorithm feels like from inside; Penrose's target is that claim, not the usefulness of computers or the possibility of intelligent behaviour.
- Simulation is not duplication: a computer can model a storm without getting wet, and on Penrose's view can model a mind without understanding anything.
- No algorithm decides halting (Turing 1936): any claimed halting-checker can be fed its own description and caught missing a case we can see.
- Any sound formal system for arithmetic has a true sentence it cannot prove (Gödel 1931), and anyone who accepts the system's soundness can see that sentence is true.
- Penrose's leap: because we can always see past any given sound system, our mathematical understanding is not any knowable sound algorithm, and therefore the brain does something non-computable. This step is where nearly all critics push back.
- The gap in physics lies between U and R: quantum theory's smooth, deterministic evolution and its random jump at measurement have never been reconciled, and Penrose argues gravity is what reconciles them.
- Non-computable does not mean random: Penrose wants a deterministic law whose consequences no Turing machine can generate, which is quite different from adding noise.
- Mathematical objects are discovered, not invented: the Mandelbrot set and the complex numbers yield far more structure than anyone put in, which Penrose takes as evidence of a Platonic world.
- Cantor, Turing and Gödel share one diagonal argument: assume a complete list, build an item that differs from every entry, and the list is shown to be incomplete.
- Non-computability is ordinary mathematics: tiling the plane, Hilbert's tenth problem and the word problem for groups have no general algorithm, though each instance may be settled.
- Classical physics is superb and, for realistic data, computable: chaos makes prediction hard, but each finite-precision calculation is still algorithmic.
- Complex amplitudes, not probabilities, are what nature adds: opening a second slit can stop light arriving, because amplitudes cancel.
- Entanglement defeats every local hidden-variable picture (Bell), and seems to sit awkwardly with relativity's denial of absolute simultaneity.
- Quantum states cannot be copied, which undercuts the strong-AI picture of teleporting a person as information.
- The real puzzle of the second law is the low-entropy past, not entropy's increase; time-symmetric laws cannot explain it.
- Our big bang was special to 1 part in 1010123: its matter was in equilibrium but its gravitational degrees of freedom were not excited at all.
- Under gravity, clumping raises entropy, and black holes are the maximum; the Earth lives on the Sun's low-entropy light, which traces back to the smooth beginning.
- The Weyl Curvature Hypothesis, Weyl curvature zero at initial singularities, encodes the arrow of time as a law about the beginning.
- Two asymmetries, one source: Penrose proposes that a time-asymmetric correct quantum gravity underlies both the second law and state reduction.
- Collapse should happen at about a graviton's difference (later, in time ħ/EG), harmless for atoms, instant for cats, and testable for objects in between.
- Most of the brain's neurons are unconscious: the cerebellum's skilled control shows that awareness is not proportional to processing.
- Blindsight and split brains show that much information processing happens without report, and that the unity of consciousness is fragile.
- Neural networks are algorithms: from McCulloch-Pitts to transformers, everything that runs on digital hardware is Turing computable.
- Consciousness, for Penrose, forms judgements in situations no rule anticipates, above all judgements of truth.
- Insight often arrives whole and wordless, as in Poincaré's omnibus and Penrose's street crossing, followed by the slow work of verification.
- Mathematical communication works because minds share access to the same truths, not the same algorithms, Penrose argues.
- Libet's timing results suggested awareness lags events by up to half a second; they are now widely reinterpreted, and do not force a paradox either way.
- The child's view: until someone explains how computation produces experience, it is honest to say the emperor has no clothes, and honest also to admit we lack the alternative.
Then and now: the book's premises in 1989 and 2026
| Premise | In the book (1989) | In 2026 |
|---|---|---|
| Computer chess | Deep Thought near grandmaster strength; world champion still ahead | Deep Blue beat Kasparov (1997); engines far beyond any human |
| Turing test | Not passed; Turing's 2000 forecast pending | GPT-4.5 judged human 73% of the time in a three-party test (2025) |
| Machine mathematics | Theorem provers for routine logic | AlphaProof (2024) silver, Gemini Deep Think (2025) gold standard at the IMO |
| Fermat's Last Theorem | Open | Proved by Wiles (1994-95) |
| Aperiodic monotile | Open question | The "hat" and "spectre" (2023) |
| Quantum computing | Deutsch's "artificial" examples | Shor's algorithm (1994); error-corrected logical qubits (2024-25) |
| Bell tests | Aspect (1982) with loopholes | Loophole-free (2015); Nobel 2022 |
| Expansion of the universe | Closed model favoured by Penrose | Accelerating, flat, dark-energy dominated (1998 onwards) |
| Black holes | Strong indirect evidence | Gravitational waves (2015); EHT images (2019, 2022) |
| Information loss | Penrose assumes it | Page-curve calculations (2019-20) favour unitarity; still debated |
| Gravitational collapse | One-graviton criterion | Diósi-Penrose rule; simplest version excluded (Gran Sasso, 2020-21) |
| Quantum brain | No site proposed | Orch-OR in microtubules (1996); decoherence objection persists |
| Neuron counts | Cerebrum dominant | 86 billion total, 69 billion in the cerebellum (2009) |
| Neural networks | Hebbian learning, small nets | Back-propagation, deep learning, transformers, trillion-parameter models |
| Libet's delay | Taken at face value | Readiness potential reinterpreted as noise (Schurger 2012) |
| Quasicrystals | Newly found (1984) | Nobel 2011; natural and nuclear-test samples; local growth models |
| Penrose | Rouse Ball Professor | Nobel Prize in Physics 2020 |
Key ideas and numbers in one place
| Idea or number | Meaning |
|---|---|
| Tn(m) | The output of the nth Turing machine on input m |
| U(n, m) = Tn(m) | The universal machine: program as part of the input |
| 1 + Tn(n) × H(n; n) | The diagonal construction that defeats any halting-checker H |
| G(F) | A Gödel sentence for formal system F: true, unprovable in F, if F is sound |
| z → z2 + c | The Mandelbrot iteration; c in M if the orbit of 0 stays bounded |
| ℵ0 < C | Countable versus continuum infinity (Cantor) |
| i2 = −1 | The extension that makes every polynomial solvable |
| Hamilton's equations | dq/dt = ∂H/∂p, dp/dt = −∂H/∂q; phase-space volume preserved |
| ψ = w∣A⟩ + z∣B⟩ | Superposition with complex weights w and z |
| Prob = ∣amplitude∣2 | The Born rule for rule R |
| S = k log V | Boltzmann entropy as the log of a phase-space volume |
| SBH ∝ A ∝ M2 | Bekenstein-Hawking entropy; black holes are entropy maxima |
| 1 in 1010123 | Precision of the big bang's initial state |
| RIEMANN = WEYL + RICCI | Tidal plus volume-changing curvature |
| τ ≈ ħ/EG | The Diósi-Penrose collapse time (from 1996) |
| 10-5 g | The Planck-mass scale where Penrose expected collapse to bite |
Frameworks and mental models
| Name | What it says | Use it when |
|---|---|---|
| Viewpoints A, B, C, D | Thinking is computation (A); simulable but not aware (B); not even simulable (C); beyond science (D) | Locating any claim about machine minds |
| Simulation versus duplication | Modelling a process need not reproduce its intrinsic properties | Judging claims that a model "has" what it models |
| The diagonal slash | Any complete list can be outrun by an item built to differ from each entry | Testing claims of completeness: of lists, checkers, rule sets |
| Sound but incomplete | A trustworthy procedure must leave some truths unproved | Designing verifiers, validators, compliance engines |
| Outside assurance | No system proves its own consistency; trust comes from outside | Audit, model risk, governance design |
| More out than in | When output vastly exceeds specification, you are discovering, not inventing | Telling real structure from artefacts of a method |
| Grades of theory | SUPERB, USEFUL, TENTATIVE, MISGUIDED | Weighting evidence from different models |
| Chaos is not non-computability | Sensitivity to data makes prediction hard, not impossible to simulate | Assessing claims that a system is "beyond computation" |
| U and R | Smooth evolution plus random jumps that do not fit together | Spotting where a theory hides a discontinuity |
| Entropy as phase-space volume | Disorder is the log of how many microstates look the same | Reasoning about irreversibility in any system |
| Explain the boundary, not the trend | The second law's puzzle is the special past, not the increase | Any "why does it always go this way" question |
| Two asymmetries, one source | Look for a common cause when two unexplained asymmetries meet | Root-cause analysis across domains |
| Processing is not awareness | The cerebellum and blindsight show much computation without experience | Separating capability from consciousness |
| Search versus insight | A search returns an answer; an insight returns a reason | Evaluating AI outputs that look like understanding |
| The child's view | Say plainly when an explanation has not been given | Reviewing any confident consensus |
- The Gödel step is not accepted by most logicians. Seeing a Gödel sentence true requires knowing the system's soundness, and nothing shows humans know their own soundness in the required sense (Feferman, Putnam, Chalmers; Gödel's own disjunction).
- The physics is speculative by design. No non-computable law has been written down; the 1989 criterion was replaced, its simplest successor excluded, and information loss is now a minority view.
- The bridge to the brain is the weakest link. Decoherence in warm neurons appears far too fast, and Orch-OR's evidence is suggestive at best.
- What remains strong: the distinction between behaviour and understanding, the precise exposition of computability, and the clear location of the measurement problem.
Where the argument is strongest, and where it is weakest
Strongest: the distinction it forces. Before this book, many discussions of machine intelligence slid between doing and understanding without noticing. Penrose made the slide visible. His viewpoint labels (A to D) let anyone say exactly which claim they hold, and the success of language models has made the distinction practical: we now have systems whose behaviour is often excellent and whose understanding is the open question. Whatever one thinks of his answer, Penrose framed the question in the form it now takes.
Strongest: the exposition. The chapters on Turing machines, Gödel's theorem, the Mandelbrot set, quantum spin and the second law are among the clearest accounts of these subjects ever written for a general reader. Many physicists first learned about the Weyl Curvature Hypothesis and the 1 in 1010123 figure here. Even critics of the conclusion recommend the book for its middle chapters.
Strong: the measurement problem. Penrose's insistence that U and R do not fit together was unfashionable in 1989, when many physicists regarded the issue as settled by Copenhagen or decoherence. Since then the foundations of quantum mechanics have become a respectable research field, with objective-collapse models tested experimentally, and his point that decoherence does not by itself produce a single outcome is widely accepted.
Weak: the step from Gödel to non-computability. The argument needs us to know that our own mathematical reasoning is sound. Critics point out that we may be sound without knowing it, or unsound in subtle ways that we correct over time, and in either case Gödel's theorem gives no conclusion about minds. Penrose's replies in Shadows of the Mind sharpened the debate without, in the view of most logicians, settling it. The logical strand is a challenge to strong AI, not a refutation of it.
Weak: the physics is a promissory note. The non-computable law of state reduction has never been formulated. The 1989 criterion gave way to the Diósi-Penrose rule, whose simplest form was excluded at Gran Sasso; versions with larger smearing radii survive but are less predictive. The argument also leans on information loss in black holes, which most physicists now reject.
Weakest: the link to the brain. Nothing in neuroscience requires quantum effects for any known cognitive function, and estimates of decoherence in neural tissue are many orders of magnitude too short for the coherence Penrose needs. The microtubule proposal adds a site but not yet a demonstrated mechanism. The evidence from anaesthesia is intriguing and consistent with many other explanations.
Open: whether experience is computational at all. Here the book's child's view remains fair. No computational theory has explained why there is something it is like to see red or to understand a proof, and neither has Penrose's. The leading scientific theories of consciousness (global workspace, integrated information, higher-order theories) disagree with one another and have not been decisively tested. Penrose's claim that something is missing from our picture is shared by many who reject his specific answer.
Who's who
| Person | Dates | Role in the book |
|---|---|---|
| Alan Turing | 1912-1954 | Defined computability (1936); proposed the imitation game (1950) |
| Kurt Gödel | 1906-1978 | Incompleteness theorems (1931); the logical heart of the argument |
| Alonzo Church | 1903-1995 | Lambda calculus; the Church-Turing thesis |
| David Hilbert | 1862-1943 | Formalist programme; the Entscheidungsproblem; tenth problem |
| Georg Cantor | 1845-1918 | Infinite cardinals; the diagonal argument |
| Bertrand Russell | 1872-1970 | Russell's paradox, which motivated formalism |
| L. E. J. Brouwer | 1881-1966 | Intuitionism; rejection of the excluded middle for infinite sets |
| John Searle | 1932-2025 | The Chinese room (1980); viewpoint B |
| Marvin Minsky | 1927-2016 | Strong AI's leading champion; "brains are meat machines" |
| Douglas Hofstadter | born 1945 | The "book of Einstein's brain" reply to Searle |
| Benoit Mandelbrot | 1924-2010 | Fractal geometry; the set that bears his name |
| Isaac Newton | 1643-1727 | Mechanics and universal gravitation; the first SUPERB theory |
| Galileo Galilei | 1564-1642 | Relativity of uniform motion; the dynamics Newton completed |
| William Rowan Hamilton | 1805-1865 | Hamiltonian mechanics and phase space |
| James Clerk Maxwell | 1831-1879 | Electromagnetic field theory; light as a wave |
| Albert Einstein | 1879-1955 | Special and general relativity; EPR; doubts about quantum completeness |
| Hermann Minkowski | 1864-1909 | Space-time geometry of special relativity |
| Max Planck | 1858-1947 | Quantum hypothesis (1900) |
| Niels Bohr | 1885-1962 | Atomic model; Copenhagen interpretation |
| Erwin Schrödinger | 1887-1961 | Wave equation; the cat |
| Paul Dirac | 1902-1984 | Relativistic electron equation; quantum formalism |
| John Bell | 1928-1990 | Bell's inequality (1964) |
| Alain Aspect | born 1947 | Photon tests of Bell's inequality (1982) |
| Hugh Everett | 1930-1982 | Many-worlds interpretation (1957) |
| Ludwig Boltzmann | 1844-1906 | Statistical definition of entropy |
| Stephen Hawking | 1942-2018 | Singularity theorems with Penrose; black-hole radiation |
| Jacob Bekenstein | 1947-2015 | Black-hole entropy |
| Roger Sperry | 1913-1994 | Split-brain research |
| Lawrence Weiskrantz | 1926-2018 | Blindsight |
| David Hubel and Torsten Wiesel | 1926-2013; born 1924 | Feature-detecting cells in visual cortex |
| Donald Hebb | 1904-1985 | The Hebbian learning rule |
| Benjamin Libet | 1916-2007 | Timing of conscious awareness |
| Henri Poincaré | 1854-1912 | Insight on the omnibus; relativity precursor |
| Dan Shechtman | born 1941 | Discovered quasicrystals (1984) |
| Stuart Hameroff | born 1947 | Microtubule proposal; co-author of Orch-OR |
| Lajos Diósi | born 1956 | Gravitational collapse model related to Penrose's |
| Martin Gardner | 1914-2010 | Wrote the Foreword |
Connect it
To AI engineering.
- Capability is not comprehension. Treat fluent output as evidence about behaviour, and test understanding separately: novel transfer, counterfactual robustness, the ability to say why. Penrose's distinction is the best single guard against over-reading a demo.
- Every verifier is incomplete. Static analysers, guardrails, evaluators and red-team suites are sound-but-incomplete procedures at best. Design for the cases they miss: human review where stakes are high, monitoring in production, and humility in claims.
- Self-evaluation has a ceiling. A model grading its own outputs, or a system certifying its own safety, faces a Gödel-shaped limit. Independent evaluation, by a different model or a human with different blind spots, buys real assurance.
- Search is not insight. Systems that generate many candidates and check them can match insight's results on benchmarks. When a decision needs a reason a regulator or user can follow, ask for the reason, not just the answer.
To enterprise architecture.
- Separate specification from behaviour. Strong AI conflates the program with the mind; enterprises often conflate the architecture diagram with the running system. Keep both, and test the gap.
- Diagonalize your catalogues. Any "complete" list of services, risks or controls invites the question: what would differ from every entry? Penrose's diagonal habit is a cheap way to find missing items.
- Find the U and R in your system. Most architectures have a smooth, deterministic core and a few discontinuous jumps (batch cut-offs, manual approvals, failovers) that do not fit the core model. Those seams are where incidents begin.
- Look for one cause behind two asymmetries. When two odd behaviours appear in the same part of the estate, Penrose's move in Chapter 8, assuming a common source, is often the fastest route to root cause.
To banking.
- Model risk is a Gödel problem. A credit or market-risk model cannot validate its own assumptions; that is why independent model validation sits outside the modelling team and why regulators insist on it.
- Tail events are about the boundary, not the trend. The second law's puzzle is its special past; a portfolio's behaviour is likewise dominated by its starting concentration and by rare, irreversible moves that averages hide.
- Automated judgement needs a human anchor. Where a decision depends on context no rule anticipated (an unusual hardship case, an ambiguous fraud pattern), Penrose's account of judgement is a reminder to keep a person able to see past the rules, and accountable for doing so.
Penrose's Argument Across Five Books
The Emperor's New Mind opened a programme that Penrose pursued for the next thirty years. Seeing where each later book picks up a thread makes this one easier to place, and shows which parts of the 1989 argument he kept, changed or dropped.
| Book | Year | What it does with the 1989 argument |
|---|---|---|
| The Emperor's New Mind | 1989 | States both strands; proposes gravity-linked collapse; no brain site |
| Shadows of the Mind | 1994 | Rebuilds the Gödel argument against critics; viewpoints A to D; introduces microtubules with Hameroff |
| The Large, the Small and the Human Mind | 1997 | Tanner Lectures: a compact restatement, with replies from Shimony, Cartwright and Hawking |
| The Road to Reality | 2004 | A full mathematical guide to physics; revises the big-bang figure to 1010124; restates objective reduction |
| Cycles of Time | 2010 | Conformal cyclic cosmology: the Weyl hypothesis rebuilt as a matching condition between aeons |
| Fashion, Faith, and Fantasy | 2016 | Criticizes string theory, quantum orthodoxy and inflation for following fashion over evidence |
Three threads run through all six. The first is realism: Penrose consistently holds that mathematics and the physical world exist independently of us, and that the quantum state is real, not merely knowledge. The second is the arrow of time: the 1 in 1010123 figure introduced here becomes the foundation of his cosmology. The third is objective reduction: the idea that gravity collapses superpositions, refined into the Diósi-Penrose rule and still the centre of his work on mind. What changed most is the brain mechanism; what changed least is the logical argument, which he has defended essentially unchanged since 1994.
Index of thought experiments in this edition
| Thought experiment | Chapter | What it tests |
|---|---|---|
| Ultronic and Adam | Prologue, Epilogue | Whether answering a question requires understanding it |
| The Turing test | 1 | Behaviour as evidence of thought |
| Searle's Chinese room | 1 | Whether rule-following yields understanding |
| Hofstadter's book of Einstein's brain | 1 | Whether a static description could be a mind |
| Teleportation machine | 1 | Whether a person is information |
| Searle's room meets a language model | 1 | The Chinese room in the LLM era |
| The slow replacement | 1 | Gradual substitution of neurons by chips |
| The actor and the script | 1 | Performing a state versus having it |
| The halting oracle for sale | 2 | Every sound checker has a visible blind spot |
| The patient compiler | 2 | Termination checking in practice |
| A journey to Tor'Bled-Nam | 3 | Discovering versus inventing mathematics |
| Two civilizations, one set | 3 | Whether independent discoverers find the same object |
| The alien without i | 3 | Whether physics forces complex numbers |
| Shown your own algorithm | 4 | What happens if we learn our own procedure |
| The self-auditing bank | 4 | Self-certification and Gödel's second theorem |
| Twenty sevens in π | 4 | Truth beyond what is decided |
| The billiard-table oracle | 5 | Could idealized classical physics compute the non-computable |
| The clockwork that cannot be wound backwards | 5 | Reversible laws, irreversible access |
| The Andromeda paradox | 5 | Relativity of simultaneity and the "now" |
| The half-silvered mirror over a light-year | 6 | When does a photon "decide" its path |
| The bomb that is never touched | 6 | Interaction-free measurement |
| Schrödinger's cat, with additions | 6 | Superposition at macroscopic scale |
| Wigner's friend in the server room | 6 | Whether observers are special |
| A world with both arrows | 7 | Consistency of time-asymmetry |
| The film run backwards | 7 | What reversal would look like |
| A universe that ran the other way | 7 | Why singularities fix the arrow |
| The Creator's pin | 7 | The precision of the initial state |
| Lamp, mirror, photo-cell | 8 | The time-asymmetry of R |
| Hawking's box | 8 | Phase-space balance with black holes |
| A diary from Hawking's box | 8 | Information loss experienced from inside |
| The mirror on a cantilever | 8 | A real test of gravitational collapse |
| The pencil, the cup and the racing driver | 9 | Two hemispheres, two responses |
| The reunited hemispheres | 9 | What happens to a divided mind |
| The illusion a network also sees | 9 | Representation versus experience |
| Poincaré boards the omnibus | 10 | Sudden insight |
| Penrose crosses a street | 10 | Wordless discovery |
| The proof that came in the shower | 10 | Search versus insight |
| Ultronic, 2026 | 10 | The prologue rerun with a modern machine |
A study path through the book
For the AI-focused reader (about six hours). Read the Overview, then Chapter 1 in full, the Gödel sections of Chapter 4, and Chapter 10. Use the Python sketch in Chapter 2 to build a Turing machine and get a feel for what "algorithm" means. Then read the whole-book critical lens and the viewpoint table.
For the physics-focused reader (about ten hours). Read Chapters 5 to 8 in order. Run the two-slit, balls-in-cells and Diósi-Penrose sketches to anchor the numbers. Then read the Preface (1999) for the revised collapse rule and the 2026 updates for what experiments have found.
For the mathematics-focused reader (about eight hours). Read Chapters 2, 3 and 4, running the Mandelbrot, Goodstein and tiling sketches. Then read Chapter 10's section on mathematical insight and the critics' replies table in Chapter 4.
For the complete reading (about twenty hours). Go front to back. After each chapter, try the "Test yourself" questions before reading the answers, and run one Python sketch. The book rewards slow reading more than almost any popular science book, because each chapter is a step in an argument whose conclusion only appears at the end.
Six questions to argue about
- If a machine produced a new, correct, deep mathematical theorem with a proof humans found illuminating, would that refute Penrose? He would say it shows the machine found a proof, not that it understood one. Is there any behaviour that would change his mind?
- Can we be sound without knowing it? If so, does the Gödel argument lose its force?
- If gravitational collapse were confirmed experimentally, would that support Penrose's theory of mind, or only his physics?
- If the cerebellum is not conscious, what property does the cortex have that it lacks? Integration, recurrence, global broadcast, or something non-computable?
- Is "what does it feel like?" a question an AI could ever answer truthfully? How would we know?
- Penrose calls his position a child's view. Is it more honest to say "we do not know" than to bet on computation, or does the bet guide better research?
About this book
The Emperor’s New Mind: Concerning Computers, Minds, and the Laws of Physics by Roger Penrose, 1989. This is a chapter-by-chapter study of the book, with summaries, maps, tables and diagrams; short quotations keep their page references. It is shared free for personal study, is not for sale, and is not endorsed by the author or publisher. Please read and buy the original.
Insight Editions · rajeshrmahapatra.com/insight-editions