The two theories, side by side

Two solutions to the same problem, drawn together, are worth more than either drawn alone. Here is the same cylinder in the same stream, solved twice.

Ahead of the body they agree closely. Behind it they part company entirely, and everything the subject argues about lives in that divergence.

What the exact theory gets right #

More than its reputation suggests, and it is worth listing before criticising it.

The front half. Over the forward-facing surface the two solutions are close. The stagnation point is in the same place, the acceleration round the shoulder has nearly the same magnitude, and the pressure distribution over the front is a good approximation.

The far field. Well away from the body the real flow is very nearly inviscid, because there is no surface nearby to generate vorticity.

The lift, when there is any. Circulation gives lift correctly, and for an attached flow the ideal prediction is within a few percent.

The speeds. Outside the thin layer, the velocities are close enough for engineering.

That is a large fraction of the flow field, which is why the ideal theory is a starting point rather than a historical curiosity.

What it misses #

One thing, and it costs everything.

The ideal solution assumes the pressure recovers completely over the rear of the body. The real flow cannot manage it, separates, and leaves a wake at roughly constant low pressure.

From that single failure follow: all of the drag, the wake, the unsteadiness at higher Reynolds numbers, stall, and the fact that the rear of a body matters more than the front.

Where exactly they part #

It is worth locating the divergence precisely rather than saying “behind the body”, because the location is the useful part.

Follow the surface from the front stagnation point. Over the first quarter the two solutions are nearly indistinguishable — the flow is accelerating, the pressure is falling, and a boundary layer in a favourable gradient is thin and well behaved.

Round the shoulder they still agree closely. The flow reaches its maximum speed at about the widest point in both.

Past that point the ideal solution begins to decelerate the flow and raise the pressure back up. The real flow follows it for a while — and then, somewhere between about eighty and a hundred and forty degrees round depending on the Reynolds number and the state of the layer, it stops following.

From the separation point onwards the two solutions have nothing to say to each other. The ideal one continues to a rear stagnation point that does not exist; the real one has a wake.

So the divergence is not gradual. The two agree well and then, at an identifiable station on the surface, one of them stops being a description of anything.

Which quantities survive #

A more useful way to compare than picture against picture is quantity by quantity.

The pattern is that everything decided ahead of the widest point survives, and everything decided behind it does not. Which is a compact statement of why the paradox happens and why streamlining works: a streamlined body is one whose important business is all conducted at the front.

Telling which will happen #

The useful question is not “is ideal flow right” but “will this flow separate”, and there is a reasonable answer in advance.

Attached flow, ideal theory good. Streamlined bodies at small incidence: aerofoils below the stall, slender fairings, anything whose pressure recovery is spread over a long tapering rear. Here ideal flow plus a boundary layer correction is excellent.

Separated flow, ideal theory useless for forces. Bluff bodies at any appreciable Reynolds number: cylinders, spheres, lorries, buildings, aerofoils past the stall. The ideal solution still describes the front and says nothing useful about the total.

The dividing question is whether the pressure recovery is gentle enough for the boundary layer to follow, which is a statement about the body’s afterbody shape and about the Reynolds number.

The trap at low Reynolds number #

That last figure hides a trap worth naming.

At Reynolds number 10 the real flow looks rather like the ideal one: attached, closing up behind, no wake to speak of. It would be easy to conclude that ideal flow is a good model there.

It is not, and the resemblance is a coincidence of appearance. At low Reynolds number the flow is dominated by viscosity, and it is nearly fore-and-aft symmetric because Stokes flow is symmetric — which is a completely different reason from the inviscid theory’s.

The forces are quite different. Ideal flow gives zero drag; creeping flow gives a drag that is large and proportional to speed, not to speed squared. The pictures agree and the physics does not.

Which is a good illustration of the site’s recurring caution: two flow pictures can look alike and be governed by different terms of the same equation, and a picture cannot tell anybody which.

What the solver computed #

The left panel of the comparison is a closed-form expression evaluated on the drawing grid. The right panel is a numerical solution of the incompressible equations, stepped until the flow settles.

Both are checked. The ideal field’s divergence is $3.3 \times 10^{-8}$ and its surface tangency is $7 \times 10^{-15}$. The viscous field is divergence-free by construction, since it is the curl of a streamfunction.

What the viscous solver resolves, and therefore what the essays may claim: attached flow at Reynolds number 1 and 10, separated flow at 40 and 100, and a recirculating region that grows with Reynolds number. Those are measured from the solved field.

What it does not resolve: the internal structure of the boundary layer, transition to turbulence, and the unsteady shedding that begins in reality above Reynolds number about 47. The computed recirculation length is also shorter than the laboratory value — 0.85 diameters at Reynolds 40 against about 2.2 — because the grid is coarse.

The site says so wherever it matters, and no figure claims a vortex street.

The forces, not the pictures #

Comparing pictures is persuasive and comparing numbers is decisive, so it is worth putting the forces next to each other.

For a circular cylinder, ideal flow gives a drag coefficient of exactly zero. Measurement gives about 1.2 over a wide range of Reynolds numbers, falling to around 0.3 above the drag crisis.

There is no sense in which zero is an approximation to 1.2. The theory is not out by a factor; it is reporting the absence of a quantity that dominates the problem.

For a streamlined strut the comparison is kinder. Ideal flow still says zero; measurement gives about 0.06. Still infinitely wrong as a ratio, and now small enough in absolute terms that the ideal solution plus a friction estimate is a usable design tool.

That contrast — same theory, same error in principle, wildly different usefulness — is what the separation question decides. It is not that ideal flow is more correct for the strut. It is that the quantity it omits is small there.

An honest statement of the model’s range #

Putting it as a designer would, rather than as a physicist.

Use ideal flow alone for the outer field, the lift of an attached aerofoil, induced velocities and anything happening well away from a surface.

Use ideal flow plus a boundary layer for the drag of a streamlined body at small incidence, and for predicting whether separation will occur at all.

Use neither for the forces on a bluff body, for anything past the stall, or for a wake — and reach instead for measurement or a full numerical solution.

The middle case is where most of aerodynamics was done for most of the twentieth century, and it works because the check “did the layer separate” is cheap and reliable. It is the model knowing its own range, which is the most that can be asked of one.

A note on comparing pictures #

Putting two solutions side by side is a persuasive figure and it needs a caution attached.

The two panels were produced by different methods at different resolutions. The left is exact to machine precision; the right is a coarse numerical approximation. A reader comparing them is comparing a solution with a simulation, and the differences include some that are numerical rather than physical.

The differences that are physical are the ones this essay is about — separation, the wake, the failure of pressure recovery — and they are qualitative and large. The differences that are numerical are quantitative and smaller: exactly where the separation point sits, how long the bubble is.

Being clear about which is which is the difference between a comparison figure and a rhetorical one.

What the front half is good for #

One practical consequence of the two solutions agreeing over the front is worth drawing out, because it is how a great deal of measurement is actually done.

If the ideal theory is reliable ahead of the widest point, then a measurement of the pressure there should match it — and a departure indicates something wrong with the experiment rather than the theory. Wind tunnel work uses exactly this: the forward pressure distribution is a check on the model installation, the tunnel corrections and the instrumentation.

The same logic works in reverse. A measured pressure that departs from the ideal prediction ahead of where separation would be expected is evidence of something unmodelled — a leak, a support interference, an unintended trip on the surface.

So the ideal solution earns its keep as a diagnostic even in flows where its forces are useless. A theory that is confidently right about part of a problem is a tool for finding faults in measurements of the rest, and that is a use quite separate from prediction.

The two-theory habit #

The pattern this essay describes is not unique to fluids and is worth naming as a habit.

A theory that is excellent nearly everywhere and wrong in a small region is not a theory to be discarded. It is a theory to be bounded — and once the boundary is known, both halves become useful: the good theory where it applies, and something else in the region it does not.

That was Prandtl’s contribution, and it is why the subject has two theories rather than one. Ideal flow was not replaced. It was given a domain.

The alternative — patching a single theory until it fits everywhere — was tried for a century and produced nothing, which is a useful data point about how to respond when a model fails in one place and works in all the others.

What a designer takes from each #

The two theories are not competitors in practice; they are used together, and the division of labour is worth stating.

Ideal flow gives the pressure distribution, which is what a boundary-layer calculation needs as input. It also gives the lift, the induced velocities and the far field. This is the outer solution and it is fast — fast enough that a design loop can run it thousands of times.

Boundary-layer theory gives the friction and the separation point, taking the ideal pressure distribution as given. If separation is predicted, the ideal solution must be recomputed for the effective body including its wake, and the loop iterated.

Neither gives what happens after a large separation, and at that point the honest options are an experiment or a full numerical solution of the complete equations.

So the practical workflow is: solve the easy problem, check whether its assumptions survived, and escalate only if they did not. That is the same structure as the two-region argument itself, applied to a working day.

Why both are kept #

A closing thought on why this subject teaches a theory it knows to be wrong.

It would be possible to teach only the viscous equations, which are correct, and skip the inviscid ones entirely. Nobody does, and the reason is not tradition.

Ideal flow is solvable. It gives closed-form answers, it can be reasoned about, and its structure — flows adding up, circulation, the Kutta condition — supplies the concepts that the viscous treatment then modifies. A student who meets only the full equations meets a system that cannot be solved and has no vocabulary for talking about what the solutions contain.

It is also, and this matters, nearly right. A theory that fails everywhere is discarded. A theory that fails in one identifiable place is the more instructive object, because the failure has a location and a cause and can be studied.

That is the case this comparison makes. Not that the exact theory should be believed, and not that it should be abandoned, but that knowing exactly where it stops is worth more than either.

Where the model stops #

The viscous solver is coarse, and the comparison shows a qualitatively correct wake with a quantitatively short recirculation.

Only a cylinder is compared. A streamlined body would show the two theories agreeing far better, which is the point of streamlining and deserves its own figure.

Steady only. The real flow at Reynolds 100 is unsteady and this solver settles to a steady state.

No transition. Whether the layer is laminar or turbulent is not modelled, and it changes the separation point substantially.

The same comparison, one regime down #

Worth seeing the pair at a lower Reynolds number, because it shows the divergence appearing rather than being present.

At Reynolds number 10 the real flow is attached and the picture resembles the ideal one. At 40 a small recirculating pair has appeared. At 100 it is large. The ideal solution is identical in all three cases, because it has no Reynolds number in it at all — which is itself worth noticing.

A theory whose prediction does not depend on a parameter that changes the answer is a theory missing a mechanism. The absence of Reynolds number from ideal flow is the mathematical signature of the absence of viscosity, and it is visible as a prediction that stays put while reality moves.

The ladder from here #

Nearby: the same comparison for a streamlined body; pressure distributions compared quantitatively rather than as pictures; the drag budget split into friction and pressure; and creeping flow, where neither panel applies.

Then back to the paradox that forced the split and the layer that resolved it.