The technological setting was changing. Digital computers, sampled control and multivariable plants created problems that the familiar single-loop, frequency-domain toolbox did not answer cleanly. Kalman does not dismiss that applied theory. He asks for a prior layer: intrinsic properties of an unalterable plant and constructive tests that say what kind and how much information a desired form of control needs.

He calls this a “pure theory of control”, while warning that it cannot ignore physical realisability. Unlike a law of nature discovered by inspecting physical objects, a general law about human-made systems is reached by abstract analysis across models. The intended theory resembles mathematics in method, yet remains answerable to whether its mathematical plants and feedback laws can be built.

The paper’s declared scope is narrower than its title. The complete results concern finite-dimensional linear systems, then chiefly stationary, single-input, single-output plants. Both continuous and discrete time appear, with sampling used to translate one into the other. Gaussian random excitation enters the filtering dual. Nonlinear, time-varying and multivariable extensions are discussed, but the clean rank tests and theorems rely on the linear setting.

The constructive emphasis matters. Kalman says the note grew from attempts to understand computation methods already used in control synthesis. Definitions of controllability and observability are not labels pasted onto familiar calculations; they organise algorithms for dead-beat control, quadratic regulation, minimum-time observation and Wiener filtering. A property of the plant becomes a gate on what any controller or observer can accomplish.

Kalman defines state as the smallest collection of numbers that must be specified at an initial time to predict future behaviour under future inputs. It is a minimal record of relevant history, not necessarily a vector of directly readable instrument values. Position and velocity may form a mechanical state; stored charge, concentration or temperature fields can play the same role after an appropriate finite-dimensional model is chosen.

The transition function maps an initial state and later inputs to a later state. It satisfies identity at the initial time and a composition rule: advance to an intermediate time, then advance again, and the result equals a single advance over the whole interval. By causality, later behaviour depends on the past through the current state. Every output must consequently be a function of state.

For the developed theory, state space is an n-dimensional real vector space. Continuous dynamics use a linear differential equation; discrete dynamics use a linear difference equation. Inputs form an m-dimensional vector, outputs a p-dimensional vector, and state-transition and input/output matrices encode their geometry. The paper later simplifies to stationary, single-input, single-output plants and assumes the discrete transition matrix is nonsingular.

This representation differs from a transfer function. An input–output formula records how applied signals appear at measured terminals, often assuming zero initial state. A state model also represents autonomous internal motion and nonzero initial conditions. Two state models can generate the same transfer function while containing different invisible or unreachable coordinates. Minimality therefore becomes a substantive structural question.

For a stationary discrete-time single-input plant, Kalman builds a sequence of vectors from the input direction and repeated inverse transitions. Complete controllability holds exactly when n such directions are linearly independent. In the now-standard forward-time convention, the equivalent reachability matrix collects B, AB, …, Aⁿ⁻¹B and asks whether its rank is n. Each column is the state displacement produced by an input applied at a different step.

The rank test has a geometric form. Start with the input direction and close it under the state transition. If the smallest invariant subspace containing that direction is the whole state space, every direction is reachable. If it is smaller, the controller is confined to a subspace no matter how cleverly it varies the signal.

Continuous time receives the analogous condition: input vector d, Fd, …, Fⁿ⁻¹d must be independent. Kalman notes that Pontryagin had used the independence requirement in another context, but the new definition supplies its intuitive control meaning. The test is constructive and finite; one does not need to enumerate every possible waveform.

Controllability is relative to the declared actuators and model. A car may be controllable in planar position with steering and acceleration, but not in vertical position. A failed actuator changes B and can collapse rank without changing the free dynamics A. Numerical near-dependence also matters in practice: exact rank says possible in ideal arithmetic, while a poorly conditioned matrix warns that reaching some directions demands enormous, noise-sensitive input.

The construction uses a basis made from successive input images and its dual basis. At each step, feedback cancels the coefficient associated with the direction that must disappear first. The state moves from the full controllable subspace into a nested subspace of one lower dimension, then repeats until nothing remains. The proof is short because the controllability definition has already exposed the right geometry.

Feedback is not assumed as a stylistic preference. The command must depend on the present state because the correction needed to remove the appropriate coordinate depends on where the plant actually is. Kalman emphasises that the feedback form is forced by the minimum-time requirement. This is the structural heart beneath sampled-data “dead-beat” response: after finitely many samples, the error is identically zero.

Continuous time creates a warning. If an arbitrary sampling interval is allowed, a controllable continuous plant can be driven to the origin in an arbitrarily short time by shrinking the interval and applying the discrete construction. The implied control signals can become unbounded and physically impossible. Mathematical reachability without input constraints is not a promise of safe, inexpensive or realisable control.

Relax minimum time and other objectives become available. Allow p≥n steps and one can reduce control energy, spread actuator movement or honour bounds. The theorem therefore separates possibility, fastest ideal action and practical design. Controllability opens the door; it does not choose the preferred path through it.

His example places two identical first-order systems in parallel. The total output makes a common factor cancellable. In state space, however, one combination of the two internal outputs responds to control while an independent combination does not. Replacing two coordinates by their sum deletes the difference coordinate. Control of the visible sum is not control of the full plant.

For stationary single-input, single-output systems, complete controllability corresponds, in engineering language, to the input exciting every natural frequency: no pole can cancel in the transfer function. The same warning applies symmetrically at the output. A mode that reaches no sensor can disappear from the external relation while continuing to evolve internally, perhaps unstably.

Sampling can create a new loss. Distinct continuous-time oscillatory modes can map to the same discrete eigenvalue when their imaginary parts differ by an integer multiple of the sampling frequency. The sampled model then admits cancellation or repeated directions that were separate in continuous time. Kalman gives explicit resonant periods where a previously controllable plant becomes uncontrollable after sampling.

This is not an argument against transfer functions or sampling. It is a demand to audit the realisation and the chosen period. Signal-flow simplification is safe for a specified zero-state input–output calculation but may be unsafe for initial-condition response, internal stability or state regulation. State-space rank tests show what algebraic cancellation hides.

Bellman’s principle of optimality supplies the recursion: whatever first command is chosen, the remaining commands must be optimal for the state that results. Because dynamics are linear and the state cost is quadratic, the finite-horizon value is a positive-definite quadratic form. Iterating the recursion produces a sequence of matrices; the proof shows that controllability bounds this sequence and its monotonic growth converges.

The general theorem says that, with positive-definite state weighting, the infinite-horizon optimal regulator exists for every positive weight if and only if the plant is completely controllable. The optimal law is unique, stationary and linear. Under the stated parameter condition μ>1, the closed loop is asymptotically stable and the value function itself is a Lyapunov function.

The necessity argument is revealing. If an uncontrollable subspace contains an unstable direction, no input can change it and the accumulated positive-definite cost diverges for some initial states. Optimisation cannot rescue a state component outside the input’s reach. The sufficiency argument uses the finite-time controllability law as one feasible policy with finite cost, proving the optimum is bounded above.

Kalman lists extensions: terminal control gives time-varying laws; control-energy cost can be added; several controls and nonstationary systems can be treated; nonquadratic criteria generally yield nonlinear laws and difficult computation. He warns that weakening positive definiteness changes conclusions. Penalising only one state combination can allow unpenalised unstable motion, so an apparently finite optimum need not stabilise the whole plant.

The language of costates avoids a hidden coordinate choice. A state vector lives in X; a costate is a linear functional in the dual space X*. Coordinate values are evaluations of dual-basis costates. If every costate is observable, every state coordinate in any basis can be recovered. If one costate is not observable, that aspect of the plant’s behaviour cannot be inferred from the measurements.

For a stationary discrete single-output plant, repeated output covectors form a finite test. In standard notation, stack C, CA, …, CAⁿ⁻¹. Complete observability holds exactly when the resulting matrix has rank n. Each row asks how one initial-state direction appears at a different measurement time. A vector in the null space produces zero output across the whole test window and is therefore indistinguishable from zero.

Observability depends on both dynamics and sensor placement. A sensor can read only one state directly, yet coupling may carry every other coordinate into that measured channel over time. Conversely, many samples of the same projection add no new information. More data points do not repair a structurally blind measurement.

The theorem assumes exact linear dynamics and noiseless observation for finite-time reconstruction. In real systems, numerical conditioning governs sensitivity to noise and modelling error. A matrix can have full mathematical rank while one singular value is so small that reconstructing its associated direction amplifies measurement error catastrophically. Exact observability is possibility; practical estimation also needs a scale.

The scheme begins with an arbitrary estimate. The free predictor advances that estimate through the same transition as the plant. At each sample, measured output minus predicted output supplies a scalar error. A correction vector maps that error into state space. With coefficients chosen from the observability basis, every costate becomes exact in the minimum number, at most n, of measurements.

The proof mirrors dead-beat control. Instead of driving a physical state through nested reachable subspaces, the observer drives estimation error through nested invisible subspaces. Subtracting predictor dynamics from plant dynamics shows that the error itself obeys a closed linear system. The correction law annihilates one remaining uncertainty coordinate per sample.

Kalman draws plant and predictor side by side. The observer is physically realisable with an analogue or digital computer because a full sampling period is available for the next estimate. This architecture is the ancestor of the modern state observer: model prediction plus innovation feedback. It also connects directly to the recursive filter introduced in Canon 0064, where noise prevents finite exactness and covariance controls the gain.

The paper derives a striking control consequence. If a plant is completely observable but its input is constrained, every state variable can still be reduced to zero in minimum time precisely when the output-based closed loop uses the corresponding feedback law. Observation is not a passive afterthought; it determines when state feedback can be implemented from available measurements.

Under this transformation, an input vector becomes an output covector and an output covector becomes an input vector. The controllability matrix of one system is the transpose of the observability matrix of its dual, apart from the time convention. Therefore a plant is completely controllable exactly when its dual is completely observable, and conversely. The minimum-time observation problem becomes the minimum-time control problem for the dual.

Kalman then dualises quadratic regulation. A positive semidefinite quadratic form in costate space can be represented through orthogonal state directions. Those directions become independent Gaussian random excitations whose covariance is the quadratic weight. After reversing time, the regulator’s accumulated quadratic cost becomes the mean-square state-estimation error of a randomly driven plant.

The conclusion is the paper’s major bridge: the optimal regulator problem in the dual space is the Wiener filtering problem in the original space. Calculation methods developed for deterministic optimal feedback immediately generate the equations of the optimal linear filter. Stability of the optimal predictor is tied to complete observability of the stochastic process model.

This is not Feldbaum’s “dual control” from Canon 0068. Feldbaum names the two operational roles of an action that both directs and learns. Kalman’s principle is an algebraic correspondence between two systems and two problems. The terms share a word but answer different questions; conflating them would erase both contributions.

If the full state is measured directly, controllability is the decisive gate for state feedback. When control must be built from outputs, observability is also required to recover the relevant state. A realisation that is both completely controllable and completely observable contains no internal direction disconnected from its terminals. Later work would formalise this as a minimal realisation of the transfer behaviour.

Kalman is candid about boundaries. Discrete time and single input/output can be relaxed, though explicit tests become more complicated. Nonlinear systems remain largely open; he hopes at least for topological descriptions. Exact yes/no properties are not enough for engineering scale, leading him to ask whether controllability and observability can become quantitative measures connected to information and achievable control quality.

The published discussion reinforces the caution. L. I. Rozonoer praises the natural concepts and duality but questions calling the account a general theory while nonlinear cases lack a clear apparatus. Kalman agrees that the paper does not pretend to present the completed development; he defends the heading as a programme and points to connections with Pontryagin’s nonlinear results.

The durable lesson is a disciplined separation. State describes internal predictive possibility; controllability asks what interventions can change; observability asks what measurements can distinguish; a particular controller or estimator is constructed only after those gates are checked. None of these ranks establishes autonomous purpose, intelligence in the broad sense or conscious experience. They describe access to modelled state directions under declared inputs and outputs.