The shutter is part of the answer

What a photograph shows asks which quantity each visualisation technique records — smoke a streakline, tufts a direction, oil the skin friction, a tap a scalar — and finds that four techniques pointed at one flow record four different things. Streamlines are not paths separates the three families of curve that a picture can contain.

Both are about what is being recorded. This one is about for how long, which is a parameter nobody states and which is not a detail.

The exposure is an operator #

A photograph of a flow is not a sample of the flow at an instant. It is the integral of whatever the technique responds to, over the time the shutter was open, divided by that time.

Written that way it is obviously an operator applied to the flow, with a width, and the width is chosen by whoever took the picture. Two exposures of one flow are two different quantities in the same sense that a streakline and a streamline are two different curves.

The subject flow here is a Kármán street: two rows of vortices convecting downstream at a spacing ratio of 0.281, which is the canonical unsteady wake and is what half the pictures in this subject are of. The rows convect at 0.646 of the free stream and induce 0.354 on themselves, giving a period of 1.548 in units of the spacing over the free stream. It is evaluated in closed form — an infinite row of point vortices sums to a cotangent — so that small residues can be measured rather than lost in a truncation.

The exposure that is exactly right #

Averaging a periodic signal over exactly one period gives its mean, exactly. That is not an approximation and it does not improve with longer averaging; it is already right.

The computation confirms it to the integrator’s own precision: an exposure of one shedding period lands 2.8 parts in a hundred million million from the infinite-exposure mean.

And every other exposure is wrong. An exposure of 1.42 periods carries the mean plus the average of the leftover 0.42 of a cycle, and that leftover does not average to anything. The residue is 5.1 per cent of the velocity scale.

The residue has exact zeros at every whole number of periods and crests between them. Between the crests it falls as one over the exposure — the leftover is a bounded quantity divided by a growing window — which the computation checks by multiplying each crest by its own exposure and finding the product constant to 2.3 per cent across five of them.

The residue against exposure, in units of the shedding period:

A quarter of a period is wrong by 14.5 per cent; one whole period is wrong by 0.11; six whole periods, which is a whole number, is wrong by 3·10⁻¹⁴. One over the exposure is a bad rate. Halving the error costs twice the exposure, and the error is already small. So the practical statement is not “expose for longer”; it is that a long exposure buys accuracy slowly while an exposure matched to the period buys it exactly, and nobody matches it because the period is rarely known while the picture is being taken.

The picture of a flow that never happens #

The more serious consequence is not the residue. It is what the converged answer looks like.

Take the velocity across the wake three vortex spacings downstream. The long-exposure profile is smooth and symmetric about the axis. The instantaneous profiles at four phases a quarter period apart are each strongly asymmetric, because at any instant there is a vortex on one side of the axis and not on the other.

The mean is not any of them and it is not near any of them. It is a real quantity — it is what a long exposure records, and it is what a force balance would report — but it is not a picture of the flow at any moment, and the smoothness that makes it look like a well-behaved flow field is an artefact of the operator.

This is the same failure that what a mean profile cannot tell anybody computes in a different setting: two flows with identical mean profiles and completely different stresses, because averaging is a projection and projections lose things.

The difference in this essay is that the width of the average is adjustable, which makes the information loss a decision rather than a fact.

A velocity measurement is the same operator, twice #

The point generalises past photographs, and the case where it does real damage is quantitative.

Particle-image velocimetry takes two images separated by a short interval, finds how far the particles moved, and divides by the interval. That is a chord, and the particle went round an arc.

In solid-body rotation the answer is exact: the measured speed is the true speed times the sinc of half the swept angle. A tenth of a radian costs four parts in ten thousand. Half a radian costs one per cent. Two radians costs sixteen per cent. Half a turn between pulses reads 2/π of the truth, which is 64 per cent, and the reading is not noisy — it is systematically and reproducibly low.

The bias is the sinc function of the swept angle, and it is exact rather than fitted: a particle sweeping 0.2 radians of its orbit between the two frames is measured 0.17 per cent slow, one sweeping 0.6 is 1.5 per cent slow, one sweeping 1.2 is 5.9 per cent slow, and one sweeping half a turn is 36.3 per cent slow — 2/π exactly. The rule that follows is the one PIV practice already has, arrived at from the physics rather than from experience: keep the particle displacement small compared with the local radius of curvature, not merely small compared with the interrogation window. The first is about the flow and the second is about the algorithm, and it is the first that sets the bias.

Note also which way the error goes. It is always an underestimate, and it is worst where the curvature is highest, which is inside vortex cores. A PIV field of a vortical flow is systematically weak exactly where the vorticity is, and averaging many frames does not help because the bias is not random.

What the solver computed, and how it was checked #

The street is two infinite rows of point vortices, evaluated through the closed-form row sum, convecting at the speed the two rows induce on each other. Exposures are quadratures over stated windows.

It is the sinc of half the swept angle to twelve figures, and the error is a systematic underestimate rather than noise.

Against an exact statement. A whole-period exposure must return the infinite-time mean exactly, and it does to 2.8·10⁻¹⁴. This is the check that the period is right, the convection speed is right and the quadrature is converged, all at once — any error in any of them would show up here as a residue orders of magnitude larger.

Against a rate. The residue crests must fall as one over the exposure, checked by requiring the product of crest and exposure to stay within a factor of two across five crests. It stays within 2.3 per cent.

Against a closed form. The two-pulse reading must equal the sinc of half the swept angle, and it matches to twelve figures at eleven angles — which is a check on the arithmetic rather than on the physics, and is there because a factor of two in the half-angle is the obvious way to get this wrong.

And a refusal. The exposure at exactly one period is required to be small; setting the tolerance to zero makes the assertion fire, which is how the check is proved to be capable of failing.

Smoke has two windows, not one #

A smoke photograph has the exposure discussed above and a second window that is easier to miss, and the second is usually the larger.

A streakline is the locus of all the particles that have passed through one point since the smoke was turned on. Its length is therefore set by how long the smoke has been running, and every part of it records the flow at a different earlier time — the far end recording the oldest. So a smoke picture is an accumulation even at zero exposure, and the accumulation window is the release duration rather than the shutter time.

Streamlines are not paths is the essay that separates the curves; what this one adds is that the streakline’s extent is a parameter of the experiment. A short puff of smoke gives a short streak that is nearly a pathline segment; smoke running for a minute gives a curve whose downstream end is a record of a minute ago and whose shape mixes sixty seconds of history into one line.

That is the general statement of a scalar is a record of where its fluid was, applied to the one scalar this subject photographs most. The dye is not showing the flow; it is showing an integral of the flow, over a window the experimenter set by opening a valve.

The instrument that has no shutter #

Nothing above requires a camera, which is the point at which the argument becomes general.

Every instrument has a response time and therefore an averaging window: a pressure transducer with a tube in front of it, a thermocouple with a bead, a hot wire with a thermal lag, an anemometer with a rotor. Each of them convolves the flow with its own kernel, and the kernel’s width is a property of the hardware rather than of the flow.

The case that matters for the velocimetry above is the seeding particle, which is itself a low-pass filter: it follows the flow only up to a frequency set by its own relaxation time. A particle is a low-pass filter computes the response — a Stokes number of one gives 0.71 of the fluctuation and 45 degrees of lag, and one per cent fidelity needs a Stokes number below 0.14.

So a PIV measurement carries three windows stacked: the particle’s own response, the pulse separation’s chord bias, and whatever averaging is applied afterwards. Each has a stated remedy and none of them is visible in the resulting vector field, which arrives looking like a measurement of the flow.

Reading a published picture #

The practical value of all this is a short list of questions to ask of any flow photograph, and none of them is usually answerable from the caption.

How long was the shutter open, in units of the flow’s own period? If it is much less than one, the picture is an instant and shows structures. If it is much more, it is a mean and shows none. If it is near one, it is a mean with a residue, and if it is between one and about ten it is a mean with a visible residue — which is the range that produces pictures of structures that are not there.

Was the exposure the same in the pictures being compared? Two photographs of the same rig at different Reynolds numbers, taken at the same shutter speed, are averaged over different numbers of periods, because the shedding frequency moved. A change in apparent structure between them may be a change in the operator rather than in the flow.

And for a quantitative field, what was the pulse separation against the local curvature? That is the sinc question, and it decides whether the vorticity in a published field is a measurement or a lower bound.

The window as a free parameter #

There is a general form of this, and this collection has met it twice already.

A boundary that only exists over a window computes finite-time structures in a flow map: ridges that are sharp when the integration window is one length and absent when it is another, with nothing in the flow changing. The window there is an analyst’s choice exactly as the exposure here is a photographer’s.

The same loss in a spectrum rather than an exposure is the essay after this one, on the same solver.

How far downwind a surface is remembered computes the atmospheric version, where an instrument on a mast averages over a footprint whose size depends on the averaging interval.

In all three the observable is a functional of the flow and of a window, and reporting the observable without the window is reporting half a measurement. The habit of stating the window is normal in turbulence statistics, where nobody would quote a variance without saying over what, and it is not normal in flow photography, where the same information is thrown away routinely.

What this does not say #

It does not say that long-exposure pictures are wrong. A mean is a legitimate object and it is often the one that matters — a time-averaged force is what a structure feels, and the mean field is what a Reynolds-averaged computation predicts, so a long exposure is the right comparison for it.

It does not say that structures in pictures are artefacts. A short exposure shows real structures. The failure mode is the middle, where an exposure of a few periods produces a smeared but still structured image and the smearing is read as the shape of something.

And it does not say the technique matters less than the exposure. It matters more, which is the previous essay’s subject: the exposure decides how long the recording lasted, and the technique decides what was being recorded. Both have to be stated and usually neither is.

What the picture cannot show #

The street here is a point-vortex model. It has the right kinematics and the wrong core: real vortices have finite cores, viscous decay and a spacing ratio that drifts downstream, and the residues computed here are for an exactly periodic flow.

Real shedding is not exactly periodic. It jitters, and the jitter breaks the exact zeros — an exposure of nominally ten periods is ten periods plus a random remainder, so the residue never quite reaches zero and instead settles to a floor set by the jitter. That makes the practical advice weaker than the arithmetic: matching the exposure to the period helps a great deal and does not give fourteen figures.

And the profiles are drawn as velocity, which is not what any of the four techniques in the previous essay records. A smoke picture’s brightness is a scalar concentration and its relation to velocity is another convolution again.

Where the model stops #

Two dimensions, one wake. Everything here is a plane wake behind a bluff body. A three-dimensional flow adds spanwise decorrelation, which is a second averaging that a picture also performs — over the depth of field — and which this model has no analogue of.

One probe location for the residue. The residue’s size depends on where in the wake it is measured; the exact zeros do not, because they are a property of the period rather than of the point.

The PIV calculation is solid-body rotation. For a general flow the correction is the same to second order in the displacement, with the local curvature in place of the rotation rate, and it is not exact.

And nothing here is about resolution. A picture also averages over space, with its own width, and the spatial and temporal averaging interact through the convection speed. That is one reason a fast convecting structure is smeared even at a short exposure.

Who found it, and when #

The mean-versus-instantaneous distinction is as old as the subject and is in every turbulence text. Its application to visualisation is where it is repeatedly forgotten, and the standard reference for that is Hama’s 1962 note on streaklines, which showed that a streakline picture of a wave can look like a much larger disturbance than the wave itself.

The sinc bias in two-pulse velocimetry is standard in the PIV literature — it appears in Adrian’s and Westerweel’s treatments as the curvature or “acceleration” error, and the usual remedy is multi-frame tracking rather than a correction. It is stated there as a rule of thumb about displacement, and the exact sinc form is what makes the rule of thumb a number.

What is not standard is stating the exposure at all, and that is the practice this essay would change.

Limits recorded rather than smoothed over #

The residue is measured at one point in one wake. Its size is not universal; the exact zeros and the one-over-exposure envelope are.

The crests are found numerically on a 240-point sweep, so their positions are quantised to about 0.025 of a period and their heights are slightly under the true maxima. The envelope check is a ratio, which makes it insensitive to that.

The spacing ratio is the classical stable one, 0.281, which is a property of the idealised street rather than a measurement of any wake. Changing it moves the convection speed and therefore the period, and changes none of the conclusions.

And no technique is modelled. The exposure is applied to the velocity field itself, which no instrument records directly. Every real technique adds its own transfer function on top, and those are the previous essay’s subject.