The length the limit invents

Prandtl’s boundary-layer equations are parabolic. That is not a modelling convenience or an approximation to something else; it is the structure of the equations, and it follows from the pressure being handed down by the outer flow rather than being part of the solution.

A parabolic problem marches. Start at the leading edge with a profile, step downstream, and each station is determined by the one before it. Nothing at station $x$ can know about anything at $x + \Delta$.

And every experiment ever done on a shock impinging on a boundary layer, or a step, or a trailing edge, shows the pressure beginning to rise before the disturbance — over a distance of a few boundary-layer thicknesses. The classical theory does not say that distance is small. It says it is zero.

What has to change, and what the change costs #

The way out is to stop prescribing the pressure.

The body the outer flow sees is the first half of it: a boundary layer lets less fluid past than an inviscid one would, and the outer flow can be given the same reduced flow rate by leaving it inviscid and moving the wall out by the displacement thickness. So the potential flow that matters is the flow past the body plus a thickness the layer computes.

That is a one-way coupling in the classical account: the layer computes a displacement, the outer flow is corrected, and nothing goes back. Making it two-way — letting the corrected pressure feed back into the layer that produced it — is what the triple deck does, and the price is that the problem is no longer parabolic anywhere the coupling matters.

The scales are solved for, not chosen #

The eighths in the triple deck have a reputation for being arbitrary, and they are the solution of a linear system.

Write the interaction region’s length as $L\,Re^{-a}$, the innermost layer’s thickness as $L\,Re^{-b}$, its velocity as $U\,Re^{-c}$ and the pressure perturbation as $\rho U^2 Re^{-d}$. Four statements fix the four exponents.

The inner layer sits in the classical layer’s shear. The classical layer is $L\,Re^{-1/2}$ thick with a velocity rising linearly across it, so a sublayer of thickness $Re^{-b}$ has velocity $Re^{-b}/Re^{-1/2}$: $c = b - 1/2$.

Its inertia balances its own viscous stress. $u\,u_x \sim \nu u_{yy}$ gives $a - c = 2b - 1$.

The pressure is of the order of that inertia. $d = 2c$.

And the displacement it makes produces that pressure. A displacement of $Re^{-b}$ over a length $Re^{-a}$ turns the outer flow through a slope $Re^{-b+a}$, which by thin-aerofoil theory is a pressure of that order: $d = b - a$.

Gaussian elimination on those four gives

$$a = \tfrac38, \quad b = \tfrac58, \quad c = \tfrac18, \quad d = \tfrac14,$$

to $10^{-12}$, with nothing about eighths put in anywhere. They are what four physical balances force.

The eigenvalue is an Airy function #

With the pressure unknown the problem admits something a parabolic problem cannot: a free interaction, a solution proportional to $e^{\kappa x}$ that exists with no disturbance causing it.

Linearising the lower deck about a shear $u = \lambda y$ and eliminating the pressure leaves, for the shear perturbation $S = u_y$,

$$\lambda y\,S_x = S_{yy},$$

whose decaying solution is $S = C\,\mathrm{Ai}(\eta)\,e^{\kappa x}$ with $\eta = (\lambda\kappa)^{1/3}y$. Two conditions close it. The momentum equation at the wall gives $\kappa p = S_y(0)$. And the displacement the layer makes gives the pressure back through the interaction law, which for a supersonic outer flow is Ackeret’s $p = -dA/dx$.

Combining them, and using $\int_0^\infty \mathrm{Ai} = 1/3$, gives

$$\kappa = \big(3|\mathrm{Ai}'(0)|\big)^{3/4}\lambda^{5/4}.$$

Both Airy constants are computed rather than looked up. Integrating $y'' = xy$ inward from a three-term asymptotic start at $\eta = 14$ gives $\mathrm{Ai}(0) = 0.355028054$ and $\mathrm{Ai}'(0) = -0.258819404$, each to 1.8 parts in a thousand million, and the integral of $\mathrm{Ai}$ from zero to infinity comes out at 0.3333333341 — one third, which is the other constant the eigenvalue needs.

The eigenvalue is then 0.827158, against Stewartson and Williams’ 0.8272, and its dependence on the wall shear is measured at $\lambda^{1.2500000}$.

The starting condition that has to be right #

One detail of that computation is worth recording because it is the kind of thing that produces a plausible wrong answer.

Shooting $y'' = xy$ inward needs a starting value, and the natural one is the leading term of the asymptotic expansion, $\mathrm{Ai} \sim e^{-\zeta}/(2\sqrt\pi x^{1/4})$ with $\zeta = \tfrac23 x^{3/2}$. Started that way at $x = 9$ the shot returns $\mathrm{Ai}(0) = 0.35640$ against the true 0.35503 — 0.39 per cent out.

0.39 per cent is exactly $5/(72\zeta)$, which is the first correction to the expansion. A starting condition is a boundary condition, and an asymptotic boundary condition truncated too early is a wrong one; the error does not decay with refinement, because it is not a discretisation error.

The signs of the corrections are the second trap. The two series are $\sum(-1)^k u_k/\zeta^k$ and $\sum(-1)^k v_k/\zeta^k$ with $v_1$ and $v_2$ negative, so $\mathrm{Ai}'$ carries $+7/(72\zeta)$ and $-455/(10368\zeta^2)$. Writing the second with a plus, by analogy with $\mathrm{Ai}$'s, left the derivative wrong by nine parts in ten thousand while the function itself was right to one in a hundred thousand — an asymmetry that is invisible unless both are checked.

With three terms and the right signs, both come out at $10^{-9}$.

Why three decks and not two #

The name is worth explaining because it is a description of the structure rather than a label.

The lower deck is the sublayer computed above: thickness $Re^{-5/8}$, velocity $Re^{-1/8}$, viscous and inertial in balance, and the only place viscosity matters. It is where the no-slip condition lives and where the shear can reverse — which is to say where when the flow lets go is happening.

The main deck is the classical boundary layer itself, thickness $Re^{-1/2}$. Over the short interaction length it behaves inviscidly: the flow there is simply displaced sideways by whatever the lower deck is doing, carrying its own profile with it, because there is no time for viscous diffusion across it in a distance $Re^{-3/8}$.

The upper deck is the outer potential flow, over a region of the same length as the interaction, where the displacement produced by the other two is converted into a pressure.

Three regions, three different balances, one problem. The reason two would not do is the middle one: a theory with only a viscous sublayer and an outer flow would have no mechanism for the sublayer’s displacement to reach the outside, since the classical layer stands between them and has to be described.

That layered structure is the same idea the thin layer introduces at the level of the whole flow, applied a second time inside its own answer.

What all this is worth at a Reynolds number #

The exponents and the eigenvalue are abstract until they are evaluated, so here they are at $Re = 10^6$ with a Blasius wall shear.

The interaction region is $Re^{-3/8}$ of a chord — 0.56 per cent. The pressure perturbation is $Re^{-1/4}$, or 3.2 per cent of the dynamic head. And the upstream influence, which is the interaction length divided by the eigenvalue, is 27 boundary-layer thicknesses.

Twenty-seven thicknesses is the number every measurement of a shock–boundary-layer interaction has ever reported, give or take — and it is the scale on which the price of a gradient is paid in a real interaction. The classical theory’s answer is zero, and the gap between zero and twenty-seven is not a quantitative disagreement; it is the difference between a phenomenon existing and not existing.

The two ratios that refuse to vanish #

Now the question these essays share, which is what survives the limit.

Every length in the triple deck goes to zero as $Re \to \infty$. The interaction length vanishes as $Re^{-3/8}$, the lower deck as $Re^{-5/8}$, the classical layer as $Re^{-1/2}$, and the pressure perturbation as $Re^{-1/4}$. A reader could conclude that the whole structure disappears in the limit and that the classical theory is recovered.

Two ratios say otherwise.

$$\frac{\text{interaction length}}{\text{layer thickness}} = \frac{Re^{-3/8}}{Re^{-1/2}} = Re^{1/8} \longrightarrow \infty.$$

$$\frac{\text{pressure perturbation}}{\text{interaction length}} = \frac{Re^{-1/4}}{Re^{-3/8}} = Re^{1/8} \longrightarrow \infty.$$

Both exponents are measured back out of the sweep at $Re^{0.125000000}$. So the limit that makes the boundary layer infinitely thin makes the interaction region infinitely long compared with it, and makes the pressure gradient the layer must survive infinitely large — while the pressure perturbation itself goes to zero.

That is as clean a statement of the rule these essays run on as this collection contains. The limit removes the pressure perturbation and leaves behind a pressure gradient that diverges.

What the exponents mean in ordinary units #

It is worth converting once, because the powers are hard to feel.

At $Re = 10^6$ on a chord of a metre, the boundary layer near the trailing edge is about a millimetre thick. The interaction region is 5.6 millimetres long. The lower deck inside it is about 35 micrometres thick. And the pressure perturbation is about 3 per cent of the dynamic head.

Those four numbers are the triple deck. A region five millimetres long and thirty-five microns deep decides whether a laminar layer separates cleanly at the trailing edge or forms a bubble, and it is smaller in one direction and larger in the other than anything the classical theory contains.

Raise the Reynolds number to $10^9$ and the interaction region shrinks to 42 micrometres — and grows to thirteen boundary-layer thicknesses, and its pressure gradient grows by a factor of five and a half. The two statements are the same computation.

Where else a new power appears #

The pattern — a limit generating a length that was not in the leading-order problem — is not unique to this case, and recognising it elsewhere is the point of naming it.

Where the straight line stops is a boundary-layer essay about another such region. A wall that is not quite there is the kinematic version: a boundary condition applied at a surface that is not the surface, with the offset being a new length. And in the compressible half of the collection the shock structure itself is a region whose thickness is a new power of the mean free path, invisible to the equations that predict the shock.

In each the leading-order theory is right and incomplete in the same specific way: it is right everywhere except in a region whose size it cannot express.

What a computation has to do about it #

There is a practical consequence and it is worth stating for anybody who runs a boundary-layer code.

A standard marching code cannot represent upstream influence at all, because it marches. Feeding it a prescribed pressure and asking it what happens near a trailing edge or a shock is asking a question the scheme cannot answer, and the answer it gives — a singularity, or a solution that simply stops — is the scheme reporting that correctly.

The fixes in use are all versions of restoring the coupling. Inverse methods prescribe the displacement thickness and solve for the pressure. Semi-inverse methods iterate between the two. Interacting boundary-layer methods solve the layer and the outer flow simultaneously. All three work, all three march through separation without a singularity, and all three are doing what the triple deck says has to be done: letting the pressure be an unknown.

Why the classical limit is singular rather than merely approximate #

The word for this is that the limit is non-uniform, and it is worth distinguishing from the ordinary situation.

An ordinary asymptotic expansion is uniformly valid: the error is small everywhere, and refining the small parameter makes it smaller everywhere. Prandtl’s expansion is not like that. It is uniformly valid over a length of order $L$ and fails over a length of order $L\,Re^{-3/8}$, and the region where it fails shrinks — but shrinks more slowly than the layer it sits in.

So the failure does not go away. It becomes a smaller and smaller fraction of the body and a larger and larger multiple of the layer, and every question about what happens near a shock, a step, a trailing edge or a separation is a question about that region.

How thick is thin is the essay about the layer’s own ambiguity of thickness — three conventions measuring three different things, one of which is not a height at all — and this is the companion statement in the streamwise direction: there is a length in the problem that the leading-order theory does not contain and cannot be given.

What it fixes downstream #

The immediate application is the one the next essay is about.

The singularity a layer makes for itself marches Prandtl’s equations into an adverse pressure gradient and finds that the wall shear vanishes with an infinite slope at a finite station, and that the solution cannot be continued past it. That singularity is Goldstein’s, it is real, and it belongs to the boundary condition: prescribing the pressure is what forbids the layer from relieving itself.

In the triple deck the pressure is unknown, the layer’s displacement is what determines it, and the flow goes through separation with no singularity anywhere. The same equations, a different thing prescribed, and a different answer.

That is why the interaction region’s length matters practically. It is the distance over which a separating layer negotiates with the outer flow, and it is the reason a laminar separation bubble has the length it has. How much uphill a layer can take measures the pressure rise a layer survives with the pressure prescribed; the triple deck is what happens when it is not.

The other interaction law, and what it changes #

The eigenvalue quoted here is for a supersonic outer flow, where the pressure is Ackeret’s $-dA/dx$: a local relation, which is what makes the algebra close in half a page.

Subsonically the interaction law is a Hilbert transform — the pressure at a station depends on the displacement everywhere — and the free-interaction problem becomes an integral equation rather than an algebraic condition. The exponents are the same, because the four balances that produced them did not care about the sign of $M^2-1$. The eigenvalue is different, and the character of the upstream influence is different: it is algebraic rather than exponential, which is a weaker decay.

That difference is not computed here, and it should not be assumed away. What is computed is the supersonic case, in full.

Limits recorded rather than smoothed over #

The eigenvalue is linear. The free interaction is an infinitesimal disturbance growing exponentially, and the whole apparatus above is a linearisation about a shear. It says a disturbance can appear upstream; it does not say how large the pressure rise will be, which is a nonlinear question.

The wall shear $\lambda$ is an input. $\kappa$ goes as $\lambda^{5/4}$, so the upstream influence depends on how healthy the layer is when the interaction starts. The figures here use the Blasius value 0.332, and a layer already near separation has a much smaller $\lambda$ and a much longer influence — which is the right behaviour and is not further explored.

And nothing here solves a triple-deck problem. The exponents are derived, the eigenvalue is computed, and the scales are evaluated. Actually solving an interaction — a shock impinging, a step, a trailing edge — means solving the lower-deck equations with the interaction law as a boundary condition, and that is a numerical problem this essay does not attempt.

The residue, stated as a length #

The classical boundary-layer theory is the leading term of an expansion in $Re^{-1/2}$, and it is excellent. What the expansion does not contain is a length: nothing in it is $Re^{-3/8}$ long, because $3/8$ is not a power the leading-order scaling produces.

So the limit invented one. Taking $Re \to \infty$ produced a region whose length is a new power of the Reynolds number, which is shorter than the body and longer than the layer, and in which the whole question of upstream influence, separation and the reattachment of a bubble lives.

That is the residue: not a quantity that failed to vanish, but a scale that was not there before the limit was taken.