A steam engine receives driving power, loses power to its load and changes speed whenever those quantities fail to balance. One may attach a device whose resistance increases with speed. Such a device pushes back against acceleration and reduces the size of a disturbance. Maxwell calls it a moderator. Yet after a permanent increase in load, the engine must settle at a different speed so that the speed-dependent resistance can reach a new balance. The reaction is useful, but the original speed is not recovered.

A governor has another state: a sleeve, valve, brake, liquid level or comparable regulating piece that can continue to move while speed is wrong. Its displacement changes the driving or resisting force. Once the speed returns to the target, the error signal vanishes and the regulator can rest at a new position suited to the new load. In present language, the added state accumulates error. Maxwell does not use that vocabulary, but his equations make the distinction exact.

This difference defines the paper’s ambition. The question is no longer whether centrifugal balls rise or a valve closes. It is whether the entire coupled machine returns after a small disturbance. That involves inertia, gain, damping and the timing implicit in several moving parts. A mechanism can have the correct direction of action and still fail dynamically. Negative feedback as a diagram is not yet stable regulation as behaviour.

Maxwell therefore studies deviations from a state of steady motion. He writes differential equations, assumes the departures are small enough for coefficients to be treated as constant and asks what forms the solutions can take. The answer will depend on characteristic roots. This move: from mechanism to perturbation equation: is the founding act of the paper.

Let `x` denote the accumulated motion of the engine and `v = dx/dt` its speed. Let `M` be the effective inertia, `P` the driving power expressed as force, `R` the load resistance and `F(v−V)` the moderator’s extra resistance relative to a reference speed `V`. The motion is:

With constant `P` and `R`, the transient is first order. Setting acceleration to zero gives `v∞ = V + (P−R)/F`. The coefficient `F` sets both the response rate `F/M` and the steady sensitivity to load. Increasing `F` shortens the time constant and reduces the offset, but no finite `F` makes the offset vanish for every non-zero imbalance.

That trade-off is easy to miss in a purely mechanical description. A strong moderator can appear impressively “governed” because it responds promptly and permits only a small shift. Mathematically, however, its reference is not an exact equilibrium under arbitrary load. The deviation itself must remain present to generate the compensating force. This is the signature of proportional action without an independently stored correction.

The time solution reinforces the point. If the load changes abruptly and then stays fixed, speed approaches the new equilibrium exponentially. A negative exponent means the transient is stable; yet stability alone says only that the motion approaches some equilibrium. Accuracy asks whether that equilibrium is the desired one. Maxwell’s contrast between moderators and governors separates these two questions decades before their modern names became standard.

Maxwell’s idealised governor has a piece `B` that moves under a force proportional to the difference between actual speed `v` and normal speed `V`. In the stripped equation, `B d²y/dt² = F(v−V)`. Integrating once and writing `x` for the engine’s accumulated motion gives `B dy/dt = F(x−Vt)` up to the integration constant. The regulator’s velocity therefore measures accumulated phase error relative to uniform motion.

This ideal equation is not yet a complete practical governor. With no damping or restoring loads on the regulating piece, it can retain neutral motions and phase offsets. Its conceptual work is nevertheless decisive: if the regulator eventually rests, `dy/dt=0`, then `x=Vt` in the chosen reference, so `v=V`. A changed engine load is carried by a changed value of `y`, not by a continuing speed error.

The governor thus couples two balances. The engine equation includes a term such as `−G y`: regulator position changes the available drive or brake force. The regulator equation includes `F(v−V)`: speed error accelerates the regulator. Close the loop and a disturbance passes from speed to regulator and back to engine. The same extra state that removes static error also raises the order of the dynamics. It makes oscillation possible.

That is the central bargain of feedback. A mechanism with memory can correct a sustained disturbance, but its correction arrives through motion governed by inertia and friction. If the stored action arrives too late or overshoots too strongly, the regulator and engine exchange energy instead of dissipating it. Exact steady regulation does not guarantee a satisfactory approach to the steady state.

A linear differential equation with constant coefficients admits modes proportional to exponentials. For a real characteristic root `n`, the mode is `e^(nt)`. A negative root decays; a positive root grows. For a complex conjugate pair `a ± ib`, the real motion can be written `e^(at) cos(bt+β)`. The imaginary part `b` gives angular frequency, while the real part `a` governs the envelope.

Maxwell’s nineteenth-century terminology needs translation. He calls complex roots “impossible roots” and the real component their “possible part”. No metaphysical impossibility is meant. The terms come from older algebraic language. Reading them literally would obscure the modern content: he is locating roots in what we now call the complex plane and using the sign of their real part as a stability test.

The four cases are not four different mechanisms. They are four shapes produced by roots. A positive real root yields monotonic divergence. A negative real root yields monotonic return. A complex pair with positive real part yields an oscillation whose envelope expands; with negative real part, it contracts. If the real part is exactly zero, the linear model predicts an undamped oscillation of constant amplitude: marginal, not asymptotically stable.

This classification changes how instability is diagnosed. A machine that repeatedly crosses the set point need not be unstable; the peaks may be shrinking. Conversely, a non-oscillatory trajectory that creeps away is unstable even though it never “hunts”. The decisive measure is the envelope or exponential rate, not whether the trace looks busy.

To find the roots, assume each small displacement varies as `e^(nt)` and substitute that form into the linearised equations. Differentiation becomes multiplication by `n`, turning differential equations into simultaneous algebraic equations. A non-zero disturbance mode exists only when their determinant vanishes. The resulting polynomial in `n` is the characteristic equation.

The device’s order is therefore not a bookkeeping nuisance. It counts the independent dynamical states and determines how many roots must be controlled. A first-order moderator supplies one root. The Jenkin governor model produces a cubic because engine speed, regulator position and regulator velocity all matter. More elaborate coupled machinery can yield a fifth-order equation. Stability is a collective property of the entire root set.

In the modern complex plane, the imaginary axis is the boundary. Roots strictly to the left correspond to negative real parts and decaying modes. Roots to the right grow. A root on the boundary supplies a non-decaying mode, so the usual standard of asymptotic return requires strict separation. A system can be close to the boundary: formally stable yet slow to settle and easily altered by neglected effects.

Maxwell does not merely solve particular polynomials numerically. He seeks restrictions on coefficients that guarantee where the roots can be. That ambition anticipates algebraic stability criteria. His cubic calculation is particularly crisp; his fifth-order discussion extracts necessary inequalities while explicitly declining to claim sufficiency. This restraint is part of the result.

Use `e=v−V` for speed error, `y` for regulator displacement relative to equilibrium and `q=dy/dt` for regulator velocity. A convenient equilibrium-shifted reconstruction of Maxwell’s equations is:

`M` is the engine’s effective inertia. `B` is the regulator’s effective inertia. `F` couples speed error to the regulator and also appears as a direct speed-sensitive resistance in the engine balance. `G` is governing power: how strongly regulator position changes engine force. `Y` is resistance proportional to regulator velocity: what Maxwell calls viscosity, whatever its physical source.

Set the disturbance `d(t)` to zero and substitute `e,y,q ∝ e^(nt)`. Eliminating amplitudes gives:

Each positive coefficient helps exclude a positive real root, but positivity alone does not exclude a conjugate pair with positive real part. For a cubic `a₃n³+a₂n²+a₁n+a₀`, the relevant Routh condition is `a₂a₁ > a₃a₀` when coefficients are positive. Substitution yields `(MY+FB)FY > MBFG`.

Dividing by positive terms gives Maxwell’s compact inequality:

The left side combines engine and governor rates with damping; the right side prices governing power against regulator inertia. The inequality says that a governor can act in the correct direction and still oscillate increasingly if position gain outruns the damping-and-inertia budget.

Define the stability margin `S=(F/M+Y/B)(Y/B)−G/B`. Positive `S`, with positive physical coefficients, is equivalent to the strict cubic inequality. Zero places the model at the boundary. Negative `S` indicates that the Jenkin linearisation has a growing mode. The formula is not dimensionless in an arbitrary unit system, so its magnitude should be compared only within a consistently scaled model; its sign is the immediate verdict.

The expression shows why “turn up the gain” is not a universal remedy. Increasing `G` strengthens the effect of regulator displacement on the engine, apparently promising faster correction. Yet it subtracts directly from the margin. The governor moves harder, the engine responds harder, and inertia carries both beyond equilibrium. Unless velocity resistance `Y` removes the exchanged energy quickly enough, the cycle grows. Maxwell’s memorable description is that the governor and machine “dance”.

Damping enters twice. The factor `Y/B` is multiplied by the sum `F/M+Y/B`; for positive parameters, increasing `Y` raises the margin nonlinearly. Increasing regulator inertia `B` generally slows that damping rate and makes a given governing power harder to tame. Engine inertia `M` weakens the `F/M` contribution. The equation is therefore a design relation, not a slogan: power, inertia and resistance are commensurable through the characteristic dynamics.

Maxwell also explains how the coefficients can be measured. If two steady engine speeds `V` and `V′` balance powers `P` and `P′`, then `F=(P−P′)/(V−V′)`. If two regulator positions `y` and `y′` produce those power differences, `G=(P−P′)/(y−y′)`. The abstract symbols connect to deliberate tests on the mechanism. Stability analysis becomes engineering because the required quantities can be estimated rather than guessed.

The regulating piece may experience axle friction, fluid drag or another speed-dependent resistance. Maxwell denotes the proportional coefficient by `Y`. What matters in the linear model is that the resisting force changes sign with regulator velocity and removes energy on each pass. The historical word “viscosity” should not be confined to one literal material; it names the mathematical relation relevant to the dynamics.

Dry friction behaves differently near zero velocity and can introduce sticking. Backlash, valve dead bands and mechanical travel limits also violate the smooth linear approximation. Maxwell’s criterion does not deny these effects; it brackets them. Near a working equilibrium and for sufficiently small motions, proportional resistance supplies a tractable first model. When the machine reaches its stops, the growing linear solution predicts failure of the local regime, not motion through infinite displacement.

For the Jenkin governor, Maxwell recommends adding a viscous resistance: potentially by making part of the mechanism move through a liquid: when axle friction is insufficient. This recommendation follows directly from the inequality. It is more precise than telling an operator to slow the machine: preserve the regulator’s corrective leverage `G`, but increase the term opposing its velocity.

Damping is not cost-free. Excess resistance can delay valve motion and lengthen recovery, especially outside the simplified coefficient structure. The objective is not maximum `Y`; it is adequate stability with useful response. A decaying oscillation may be acceptable if its peaks shrink quickly, while a heavily overdamped monotonic return may be too sluggish for a changing load. Root placement, not one coefficient in isolation, carries the design judgement.

His first class keeps the centrifugal pieces at a fixed radius and uses their pressure to act through a movable surface. Fleeming Jenkin’s governor supplies the detailed cubic already examined. Its geometry is translated into effective inertia, speed coupling, governing power and viscous resistance. The mechanical particulars matter because they determine those coefficients.

The second class lets a centrifugal piece change position as speed changes. Maxwell associates it with arrangements by William Thomson and Léon Foucault. Here the engine and governor can each have direct velocity resistance, denoted by coefficients such as `X` and `Y`, while a coupling `K` links their motions. The characteristic cubic takes the form:

The corresponding condition is `(X/A+Y/B)(XY+K²)>GK`. Its architecture resembles the Jenkin result: inertial rates and damping-like terms must exceed a demand set by governing action and coupling. The letters differ because the mechanical partition differs; the feedback logic persists.

The third class is the Siemens liquid governor. A rotating vessel establishes a liquid surface; overflow or discharge changes with speed and supplies regulation. A fluid mechanism may look unlike the flyball governor, yet it still possesses stored motion, resistance, coupling and delay. Maxwell’s analytical programme is to identify the variables, linearise their relations and examine the resulting roots.

This is a methodological unification, not a claim that the devices are interchangeable. The same cubic degree can hide different nonlinearities, wear patterns and practical limits. The value of the common language is that it specifies what must be measured and what constitutes a stability proof within the local model. A picturesque mechanism becomes one instance of a class of dynamical systems.

When more engine and governor motions are coupled, Maxwell reaches a fifth-order characteristic polynomial:

From the assumption that all roots have negative real parts, he derives restrictions including `pq>r` and `ps>t`. Crucially, he says he cannot prove that these conditions are sufficient. Passing them does not establish stability; failing them rules it out under the model. This asymmetry should be preserved. Turning necessary tests into a complete criterion would make the companion tidier and the history false.

Maxwell also considers differential gearing arranged so that a load disturbance can affect the engine without delivering the same immediate impulse to the governor. Under particular coefficient relations, the governor responds to the resulting speed error rather than being kicked directly by the external event. In modern terms, the mechanism reshapes disturbance paths. Feedback design includes deciding what the sensor and controller are physically exposed to.

The analysis is local. Coefficients are treated as constant around steady motion; products of small departures are neglected. Valve saturation, finite sleeve travel, impacts, backlash, dry friction, boiler dynamics, measurement noise and delayed combustion can matter in an actual engine. A positive local margin does not certify arbitrary disturbances or every operating point. A negative margin, however, is a clear warning that even small deviations grow within the stated model.

The paper also supplies no evidence about intelligence, agency or consciousness. A governor has a target-like reference, senses a mechanical variable and acts through feedback. Those functional resemblances make it historically important for theories of purpose, especially in Paper 0062. They do not show subjective experience, flexible goal selection or representation. The proper bridge is conceptual: regulation makes apparently purposive behaviour mathematically analysable without settling what mentality requires.