The three ways of coming to rest

Every oscillator in the world stops. A pendulum left alone comes to rest hanging straight down, a plucked string goes quiet, a car settles after a bump, and a galvanometer needle that has swung to its reading eventually holds still there. What separates these cases is not whether they stop but how: the pendulum swings back and forth many times while dying away, the car settles in about one movement, and a needle in thick oil creeps back so slowly it can look jammed. All three are the same equation with one number changed.

The equation is the one every minimum in physics produces with a loss term added:

$$m\ddot x = -kx - b\dot x,$$

which is tidier written as

$$\ddot x + 2\zeta\omega_0 \dot x + \omega_0^2 x = 0, \qquad \omega_0 = \sqrt{k/m}, \qquad \zeta = \frac{b}{2\sqrt{km}}.$$

Two constants went in and one dimensionless number came out. The natural frequency $\omega_0$ sets the clock; the damping ratio $\zeta$ decides the shape of everything that happens on it, and nothing else does. Two systems with the same $\zeta$ behave identically after their time axes are rescaled, whether they are a pendulum in air, a mass on a spring in oil, or a needle in a magnetic field.

What the undamped case leaves out #

Without the loss term the motion is a level set of a conserved energy, and the phase portrait is a family of closed curves that go round for ever.

Drawn as paths in the plane of displacement against velocity, the undamped case is a family of closed curves — each one a level set of an energy that does not change, so a system placed on one stays on it and the motion never ends. Every statement in the rest of this essay is about what happens when that conservation is broken by a term proportional to velocity, which is the smallest possible way to break it and still enough to change the answer to every question that follows.

The loss term is what makes an oscillator a physical object rather than a mathematical one. It is also the least justified part of the model: a force strictly proportional to velocity is what a body moving slowly through a fluid feels, and it is not what dry contact between two surfaces produces. That distinction is taken up at the end, because it decides where all of this applies.

Two exponents, and the moment they collide #

Every solution of a linear equation with constant coefficients is a sum of exponentials $e^{st}$, and substituting one gives

$$s^2 + 2\zeta\omega_0 s + \omega_0^2 = 0, \qquad s = \omega_0\left(-\zeta \pm \sqrt{\zeta^2 - 1}\right).$$

The whole subject is in that square root. Below $\zeta = 1$ the discriminant is negative and the two roots are complex conjugates; above it they are two distinct real numbers; exactly at it they are the same number twice.

The semicircle says something a trace of displacement against time does not. As damping is added, the ringing frequency falls — the damped oscillator rings at $\omega_0\sqrt{1-\zeta^2}$, not at $\omega_0$ — while the decay rate $\zeta\omega_0$ rises, and the two are the legs of a right triangle whose hypotenuse never changes. At $\zeta = 0.1$ the frequency is 0.5 per cent low, which is why light damping is usually ignored; at $\zeta = 0.5$ it is 13 per cent low, which is not ignorable at all.

The collision at $\zeta = 1$ is a genuine mathematical event. A repeated root means the two exponentials are the same function, so the second solution is not another exponential but $t\,e^{-\omega_0 t}$ — a factor of $t$ appears from nowhere. That extra factor is why the critically damped trace can be described as just reaching zero rather than crossing it: the linear growth and the exponential decay balance for exactly one instant.

Nothing physical happens there. The number of times the motion crosses zero in a fixed interval falls continuously to zero as $\zeta$ approaches one, and a trace at $\zeta = 0.999$ cannot be told from one at $\zeta = 1.001$ by looking. What is discontinuous is the arithmetic used to describe it, which is a different thing — the same distinction as a caustic being a property of the ray picture rather than of the light.

The envelope, and the measurement it hands over free #

Below critical damping the motion is a sine confined between two exponentials.

That constancy is worth more than it looks. It means the damping of a real system can be measured from two numbers read off a recording — one peak, the next peak, one division, one logarithm — with no need to know the mass, the stiffness, the driving force, or anything else about the apparatus. The logarithmic decrement

$$\delta = \ln\frac{x_n}{x_{n+1}} = \frac{2\pi\zeta}{\sqrt{1-\zeta^2}}$$

is one of the small number of quantities in physics that can be got out of an experiment with a ruler. It is the reason a bell founder can grade a casting by listening to how long it rings, and the reason a structural engineer can find the damping of a bridge by hitting it and watching the trace decay.

What it cannot do is tell where the loss went, or whether the model is right. A decrement measures the model’s one parameter on the assumption that the model applies; it never says whether a different loss mechanism would have produced the same fading with a different shape underneath.

The number engineers actually quote #

The damping ratio is the natural variable for the mathematics and it is almost never the number written on a datasheet. That is the quality factor, and the two are one division apart:

$$Q = \frac{1}{2\zeta}.$$

Everything in this essay can be restated in it, and several statements become easier to read. The amplitude falls by a factor of $e$ in $Q/\pi$ cycles. The fraction of the stored energy lost per cycle is $2\pi/Q$. The resonance curve’s width, measured between the frequencies where the power has halved, is $\omega_0/Q$. And critical damping is $Q = 1/2$, which is the awkward-looking price of the convention.

The range of values across ordinary objects is worth having in mind, because it spans eight decades. A car suspension is around 1.7 — deliberately low, so that a bump is absorbed in about one movement. A plucked guitar string is a thousand or so, which is why it rings for a second. A quartz crystal in a watch is a hundred thousand, which is what makes it a clock rather than a resonator. A suspended mirror in a gravitational-wave detector is ten million or more, achieved by hanging it from fused silica fibres that lose almost nothing per cycle.

The inversion is the useful part. Anything designed to stop wants a small $Q$, and the design problem is adding loss without adding stiffness. Anything designed to keep time or to select a frequency wants a large one, and the design problem is removing every path by which energy can leave — clamping, air, internal friction in the material, and radiation of sound.

Which is why the same physical quantity is called damping in one industry and loss in another and quality in a third. All three are $\zeta$, and which name is used says what the object is for.

The trade nobody makes deliberately #

Here is the question actually asked of a damped system, and it is not the question the textbook regimes answer. A galvanometer, a door closer, a car suspension and a printer head all need to reach a position and be there — meaning inside some tolerance, and staying inside it. How much damping does that fastest?

Critical damping is not the answer. For a band of two per cent the fastest damping ratio is about 0.79, and the saving over $\zeta = 1$ is real: an under-damped system overshoots, but if the overshoot is smaller than the tolerance then nobody cares, and the faster approach is free. The tighter the tolerance, the closer the optimum creeps to one, because there is less and less room for an overshoot to hide in.

So the familiar statement needs its qualifier restored. Critical damping is the fastest return that never overshoots. Whether that is what is wanted depends entirely on what is being built. A needle that must not swing past its reading — because a reader would write down the wrong number — genuinely wants $\zeta = 1$. A car suspension is deliberately built at about $\zeta = 0.3$, well under, because a slight float is more comfortable than the harsh arrival a critically damped spring gives and passengers are not a tolerance band. A door closer is set near one, so the door does not bounce back off the latch.

The same number decides how sharply it responds #

Damping is usually introduced as the thing that makes free motion stop, which makes it look like a nuisance. Its more useful role is the opposite: it decides how a system responds when something drives it.

The connection is not a coincidence but an identity: a system that takes a long time to forget an impulse is a system that responds over a narrow band of frequencies, because a long memory in time is a narrow window in frequency. That trade is the same one a wave packet makes between its length and its spread of wavelengths and the same one a spectral line makes between its width and how long its source keeps in step. Three subjects, one Fourier relation, and the damping ratio is where it enters mechanics.

Where the energy went, and the direction that makes #

Nothing so far has mentioned energy, and it is the part that makes damping different in kind from everything else in the equation.

In the undamped exchange, kinetic and potential energy trade back and forth with their sum unchanged. Damping breaks the sum, and it is worth being clear about where the difference goes: not into either column, but out of the system entirely, as heat in the surrounding fluid or in the material of the spring. It does not come back, which is the whole reason the motion has a direction in time when the undamped version does not.

A damped oscillator loses energy at a rate $b\dot x^2$, which is positive whenever it is moving. That is the whole reason the motion has a direction in time. The undamped equation is unchanged by reversing $t$, so a film of it run backwards is another legal motion. The damped one is not: run it backwards and the amplitude grows, which no arrangement of springs and oil will produce. Damping is where the arrow of time enters mechanics, and it enters not as a new law but as a bookkeeping device for energy that has gone somewhere the model refuses to track — a fluid’s molecules, a metal’s lattice, the enormously larger number of ways there are of being warm.

The phase portrait makes the classification honest. Under-damped paths wind around the origin; critical and over-damped paths approach it along a direction without winding. That is the real distinction, and it is topological rather than dynamic: the question is how many times the path goes round, and the answer changes from some to none.

Where the model stops: the loss that is not proportional to speed #

Everything above assumed the loss is $-b\dot x$. That is a modelling choice, and it is right for a body moving slowly through a fluid, for a conductor moving in a magnetic field, and for a resistor in a circuit. It is wrong for the commonest loss of all.

Dry friction between two solid surfaces does not resemble a term proportional to velocity at all. The force is whatever it needs to be up to a limit, then drops to a slightly lower value, and its magnitude barely depends on speed. An oscillator damped that way decays linearly rather than exponentially, stops in a finite time rather than approaching zero for ever, and stops wherever it happens to be rather than at the equilibrium — three qualitative differences, from one change in the form of the loss.

The differences are not small. Under Coulomb friction the peaks fall by a fixed amount each swing instead of a fixed ratio, so the envelope is a straight line, the logarithmic decrement is not constant, and the motion terminates at a definite moment with the mass generally not at $x = 0$ — anywhere within a dead band where static friction can hold it. A pendulum in air is well described by the linear model at small amplitude and badly at large, where the drag on the bob goes as the square of the speed instead.

So the domain of validity is: speeds low enough that the fluid drag is linear, no sliding contact anywhere, and amplitudes small enough that the restoring force is still the parabola at the bottom of the well rather than the full nonlinear return. Inside that box, one number describes everything. Outside it, the equation is not merely inaccurate but the wrong shape.

The loss that does not know the frequency #

The dry-friction case is one departure from the model and there is a second, less dramatic and far more common in engineering practice.

The viscous term $-b\dot x$ makes the energy lost per cycle proportional to the frequency: the faster the oscillation, the faster the dashpot is pushed and the more it dissipates. That is right for a fluid and it is not what a solid does. Bending a metal beam back and forth loses a fraction of the stored energy on each cycle that is very nearly independent of how fast the cycling is — the loss comes from internal rearrangements in the material, and each cycle costs the same whether it takes a millisecond or a second.

Modelling that requires a loss proportional to the displacement rather than to the velocity, with a quarter-cycle phase shift built in — which is what a complex stiffness supplies, and which structural dynamics calls hysteretic or structural damping. It is characterised by a loss factor rather than a damping ratio, and the two agree at one frequency and diverge either side of it.

The distinction matters wherever a structure has several modes at different frequencies. Fitted with a viscous model, a beam’s high modes appear far more heavily damped than its low ones, simply because the model says so; measured, they are damped about equally. So a viscous fit made at one frequency mispredicts every other, in a direction that looks like a material property and is a modelling artefact.

It also causes a genuine difficulty of principle that is worth knowing about. A loss independent of frequency is not something a system of masses, springs and dashpots can produce, and taken literally it implies a response that begins slightly before its cause. The usual resolution is to treat it as a frequency-domain description only, valid over the band of interest, and to avoid asking it what happens to a transient.

Why a large Q is worth so much #

There is a reason beyond ringing time that a resonator’s loss is worth minimising, and it connects this essay to a quite different subject.

Any oscillator in contact with something at a temperature is being jostled by it, and the jostling and the damping are two faces of one thing: whatever mechanism removes energy from the oscillator also delivers energy to it at random. A low-loss oscillator is one weakly coupled to its surroundings, so it both keeps its energy longer and is disturbed less.

The consequence is quantitative. The total thermal energy an oscillator carries is fixed by the temperature alone and does not depend on the damping — a stiff spring’s mean squared displacement is $k_BT/k$ whatever its loss. What the damping decides is how that energy is distributed in frequency: a low-loss resonator concentrates its thermal motion into a very narrow band about its own frequency, and leaves the rest of the spectrum quiet.

For an instrument measuring at a frequency away from the resonance, that is everything. Raising the $Q$ by a factor of a hundred does not reduce the total thermal motion at all; it moves that motion out of the band being watched and piles it into a spike nobody is looking at. Which is why a gravitational-wave detector’s suspensions are made of fused silica rather than steel, and why the figure of merit quoted is a loss angle rather than a settling time.

What the pictures cannot show #

The traces here are all releases from rest, which is one initial condition out of a two-parameter family. Starting with a velocity as well as a displacement changes the shape of the early motion — an over-damped system given a large enough initial push toward equilibrium does cross zero once, exactly once, which the figures never show because none of them starts that way.

Nor do they show what happens to the lost energy. The heat produced by a decaying oscillation is genuinely there and genuinely warms something, but a figure of displacement against time has no axis for it, and the amounts involved — microjoules, for a laboratory pendulum — would be invisible on any shared scale.

And the phase portraits are drawn in one plane, which works only because the equation is second order in one variable. A system with two coupled degrees of freedom, like the pair of pendulums that trade their motion, needs four dimensions for the same picture, and the projections into two are not level sets of anything.

Where the ladder goes next #

The damping ratio has now appeared three times: as the shape of a decay, as the width of a resonance, and as a settling time. Each rung of this ladder has added something the previous one had no room for — the small-angle approximation and where it fails, then the exact period and its dependence on amplitude, and now the loss.

Two further rungs are visible. One is the driven damped oscillator followed through its transient rather than in steady state, where the two behaviours here superpose and the system rings at its own frequency while settling into the driver’s. The other is what happens when the damping is made negative — when a system takes energy from a supply instead of giving it up — which turns the decaying exponential into a growing one and is the whole of how a modulated parameter sets a swing going.

The habit worth taking away is smaller than either. Three regimes were presented, and there were never three of anything: there was one equation, one number, and a square root that changes sign. When a subject arrives pre-divided into cases, the useful question is whether the cases are in the world or in the arithmetic used to describe it.