Every mode decays and it grows anyway

Pipe flow is linearly stable at every Reynolds number. That is not a modelling shortcut or an unresolved question: the eigenvalue problem for the parabolic profile has been solved, and every mode decays, at $\mathrm{Re} = 2000$ and at $\mathrm{Re} = 10^7$ alike.

Pipes become turbulent at a few thousand.

There is something wrong with the question, and what is wrong with it can be shown in a two-by-two matrix.

What modal analysis actually promises #

Linearise the equations about a steady flow, look for solutions of the form $\mathbf{x}(t) = \mathbf{v}e^{\lambda t}$, and ask whether any $\lambda$ has a positive real part. If none does, every mode decays and the flow is called stable.

That is a statement about the limit $t \to \infty$. It says the disturbance eventually vanishes; it says nothing whatever about what it does first.

The gap between those two statements is empty only when the eigenvectors are orthogonal. If they are, a disturbance decomposes into perpendicular pieces, each decaying independently, and the total energy — the sum of their squares — falls monotonically from the first instant. That is the case everybody pictures, and it is the case of a normal operator.

Fluid stability operators are not normal, and the gap is enormous.

The smallest object that has the property #

Here is a matrix with the behaviour, from Trefethen, Trefethen, Reddy and Driscoll’s 1993 argument:

$$\mathbf{A} = \begin{pmatrix} -1/\mathrm{Re} & 0 \\ 1 & -2/\mathrm{Re}\end{pmatrix}$$

Its eigenvalues are $-1/\mathrm{Re}$ and $-2/\mathrm{Re}$, both negative for every positive $\mathrm{Re}$. A modal analysis pronounces this system stable, always, exactly as it pronounces pipe flow stable.

Its eigenvectors are $(0, 1)$ and $(1, \mathrm{Re})$, and at $\mathrm{Re} = 40$ those are $1.4^\circ$ apart. That is the whole of the mechanism. A vector written as the difference of two nearly parallel large vectors is small; as the two large parts decay at slightly different rates, the cancellation between them fails, and their difference grows.

What the solver computed, and how it was checked #

The energy gain is $G(t) = \|e^{\mathbf{A}t}\|_2^2$ — the largest factor by which any initial disturbance’s energy can be multiplied by time $t$ — computed as the square of the largest singular value of the matrix exponential, at nine hundred times.

Four checks, and the third is the one that stops this being a trick.

Both eigenvalues are negative, computed from the matrix rather than asserted.

The maximum gain rises as $\mathrm{Re}^2$, fitted over four decades of Reynolds number rather than quoted: the slope comes out at $1.991$, and the closed form for this model is $G_{\max} = \mathrm{Re}^2/16$ at $t = \mathrm{Re}\ln 2$. The swept maximum lands on that closed form to a part in ten thousand by $\mathrm{Re} = 200$, and not at $\mathrm{Re} = 50$ — the correction is of order $1/\mathrm{Re}$, so what is asserted is that the gap shrinks as the Reynolds number rises.

A normal operator with the same eigenvalues never grows at all. The same routine, run on the diagonal matrix $\mathrm{diag}(-1/\mathrm{Re}, -2/\mathrm{Re})$, returns $G(t) \le 1$ at every one of four hundred times. Without that control the argument would be consistent with the growth coming from the eigenvalues after all.

And the departure from normality is measured: $\|\mathbf{A}\mathbf{A}^{\mathsf{T}} - \mathbf{A}^{\mathsf{T}}\mathbf{A}\|$ is $1.41$ for the model and exactly zero for the control. That number, and not the spectrum, is what separates them.

What the off-diagonal one is, in a fluid #

The model’s coupling term has a physical name, and it is the reason the toy is worth taking seriously.

In a shear flow, a weak streamwise vortex — a swirl about the flow direction — is itself a decaying structure: nothing sustains it and viscosity erodes it. But while it lives it moves fluid across the mean shear, carrying slow fluid up and fast fluid down, and building a streak: a long region of velocity excess or deficit aligned with the flow.

The streak can be far stronger than the vortex that made it, because the vortex works on it for as long as it survives, and the mean shear it is stirring is large. That is the lift-up mechanism, and it is exactly the off-diagonal $1$ in the matrix: a term by which one decaying quantity feeds another.

The energy in the streak eventually decays too — the eigenvalues were negative — but by then it may be large enough for the nonlinear terms this linear analysis discarded to matter, and at that point the linear question has stopped being the relevant one.

Where the energy comes from #

A disturbance whose energy rises by a factor of six hundred is being fed by something, and it is worth saying what, because “linearly stable” can suggest there is nothing available.

The energy comes from the mean flow. A linear disturbance in a shear flow exchanges energy with the mean profile through the production term $-\langle u'v'\rangle\,\mathrm{d}U/\mathrm{d}y$ — the same term that keeps a turbulent flow going — and a disturbance arranged to have the right sign of $u'v'$ extracts from it. The mean shear is an enormous reservoir, and the linear problem is a question about how efficiently a small disturbance can tap it, not about whether anything is available.

Modal analysis asks whether a single structure can tap it fast enough to outpace its own viscous decay, forever. Transient growth asks how much can be extracted before the decay wins, which is a different question with a much larger answer — and, in a flow where the eventual answer is turbulence, the more relevant one.

What it says about transition #

The picture that replaces “unstable above a critical Reynolds number” is a threshold in amplitude rather than in Reynolds number.

If a disturbance is amplified by $G \sim \mathrm{Re}^2$ before decaying, then the initial energy needed to reach a fixed level falls as $\mathrm{Re}^{-2}$. The flow does not become unstable; it becomes ever more sensitive, and the disturbance required to trip it becomes ever smaller while every eigenvalue continues to insist that it is stable.

That is a much better description of what a pipe actually does. The transition Reynolds number depends on how quiet the apparatus is, ranges over two orders of magnitude between careful and careless experiments, and has no fixed value — which is exactly what an amplitude threshold predicts and exactly what a critical Reynolds number cannot explain.

The exponent here is the model’s own and is not a pipe’s. Experiments on pipe flow report threshold amplitudes falling as something closer to $\mathrm{Re}^{-1}$, and which exponent the real problem has is an open question with a substantial literature. What the toy establishes is the shape of the argument, not the number.

What replaces the spectrum when the spectrum will not do #

The two-by-two example shows that eigenvalues are the wrong answer and does not say what the right one is. There is a right one, it is the object the 1993 paper’s title is pointing at, and it makes the argument quantitative for an operator of any size.

Instead of asking where the eigenvalues are, ask where they could be moved to by a small perturbation. For a given $\varepsilon$, collect every complex number that is an eigenvalue of $\mathbf{A} + \mathbf{E}$ for some perturbation with $\|\mathbf{E}\| \le \varepsilon$. That set is the $\varepsilon$-pseudospectrum, and it is equivalently the region where the resolvent $(z\mathbf{I} - \mathbf{A})^{-1}$ is larger than $1/\varepsilon$ — a set that can be computed without finding any eigenvalue at all.

For a normal operator the pseudospectrum is exactly what intuition expects: a disc of radius $\varepsilon$ around each eigenvalue, and nothing more. Perturb the operator a little and its eigenvalues move a little.

For a non-normal one it is not. The pseudospectrum can extend enormously further than $\varepsilon$ — and, crucially, it can protrude across the imaginary axis into the right half plane while every actual eigenvalue sits comfortably in the left. That protrusion is the diagnosis: the operator is not unstable, and it is within a very small perturbation of being unstable, and the distance to instability is a far better description of its behaviour than the position of its eigenvalues.

It also bounds the growth, which is what makes it more than a picture. If the pseudospectrum reaches a distance $d$ into the right half plane at level $\varepsilon$, the transient amplification is at least $d/\varepsilon$ — so a spectrum that stays put while its neighbourhood bulges is precisely the signature of a system that grows before it decays. The two-by-two matrix’s factor of $\mathrm{Re}^2/16$ is that bound with the numbers filled in.

Two practical readings follow and both are worth carrying.

A stability calculation on a shear flow is ill-conditioned by construction. The eigenvalues are sensitive to perturbations of the operator — that is what a protruding pseudospectrum means — so they are sensitive to the discretisation, the truncation and the rounding. Spectra of the Orr–Sommerfeld operator computed at high Reynolds number are notoriously mesh-dependent for exactly this reason, and the sensitivity is physics rather than sloppiness.

And a model error is as dangerous as a disturbance. If a perturbation of size $\varepsilon$ to the operator would make it unstable, then a modelling approximation of that size — a neglected term, a linearisation about a slightly wrong base flow, a boundary condition applied a little way from where it belongs — has the same effect as a disturbance of that size. The pseudospectrum measures how much the answer deserves to be trusted, which is a question the spectrum cannot even be asked.

Why the eigenvalue habit is so hard to break #

It is worth being fair to modal analysis, because it is not a mistake — it is the right tool for a class of problems that happens not to include this one.

Where the operator is normal, or nearly so, the eigenvalues tell the whole story, and a great deal of physics is like that: an oscillator, a wave on a string, a self-adjoint quantum system. In fluid mechanics the buoyancy-driven problems are close to normal too, which is why the Rayleigh–Bénard threshold is a genuine critical value that experiments reproduce to a few per cent — and why the contrast with pipe flow is so stark.

What makes the shear problem different is that the linear operator contains a term coupling one component to another in one direction only: the shear tilts a vortex into a streak, and a streak does not tilt back into a vortex. That one-way coupling is precisely what makes a matrix non-normal, and it appears in the equations for the same reason the inflection-point criterion does — because a mean shear is present and it has a preferred direction.

So the diagnosis is not that stability theory was done badly. It is that the standard question, asked of an operator with a one-way coupling in it, returns an answer about a limit nobody was interested in.

What the picture cannot show #

Nothing here is a fluid. The matrix has no velocity field, no geometry, no boundary and no viscosity except as a name attached to a parameter. It is the smallest object with the property, and every figure in this essay says so in its model note.

The nonlinearity is missing, and it is what actually causes transition. Transient growth makes a disturbance large; what happens next — whether it breaks down, self-sustains, or decays — is a nonlinear question, and the whole modern picture of a self-sustaining cycle of streaks and vortices lives in that gap.

Two dimensions of state is not two dimensions of space. The matrix’s two components stand in for “a vortex” and “a streak”, and a reader should resist reading them as coordinates. The real operator acts on a velocity field with a wavenumber in each direction, and the optimal disturbance in a pipe is a specific three-dimensional structure — long in the flow direction, periodic round the circumference — that this model has no way to represent.

The optimal disturbance is optimal, which is to say unlikely. $G(t)$ is a worst case over all initial conditions, and a real disturbance is not the worst case. What the bound says is that the possibility exists, not that a laboratory will find it — although a laboratory with enough noise in it will find something close enough.

The same shape, in a place with no fluid in it #

The non-normal argument is not a fluid-mechanical special case, and it is worth knowing where else it turns up, because that is what makes it trustworthy rather than convenient.

Any system with a one-way coupling has it: a quantity that feeds another without being fed back. The matrix above is the smallest example, and the same structure appears in the transient response of chemical reaction networks, in the amplification of noise in a signal chain, and in the sensitivity of an eigenvalue computation to rounding — where the departure from normality is exactly what makes an eigenvalue ill-conditioned.

That last connection is not a coincidence. The reason a non-normal operator’s eigenvalues fail to describe its behaviour is the same reason they are hard to compute accurately: the eigenvectors are nearly parallel, so the basis they define is nearly singular, and small perturbations move things a long way in it.

So the finding transfers. Wherever a linear analysis reports decay and the object of study insists otherwise, the first question is not whether the analysis was done correctly but whether the operator is normal — and if it is not, the eigenvalues have answered a question about infinity while the phenomenon lives in the first few time units.

Where the model stops #

The genuine stability operators of shear flows are infinite-dimensional and their non-normality is far more elaborate than a two-by-two example. What the two-by-two case gets right is the mechanism and the qualitative scaling; what it gets wrong is every specific number, and there is no reason its $\mathrm{Re}^2$ should be a pipe’s.

The linearisation itself has a domain: it holds while the disturbance is small compared with the mean flow. Transient growth is precisely a process that takes a disturbance from small to not-small, so the theory predicts its own invalidation — which is an honest thing for a theory to do and an awkward one to build on.

It is also worth noting what the model shares with the rest of this field’s honesty about itself. The Lorenz truncation keeps three numbers of a convection problem and is not a fluid either; both are kept because they have a property the full problem has, and both are labelled at every appearance so that the property is never mistaken for a prediction.

And the energy norm is a choice. “How much has it grown?” needs a measure, the kinetic energy is the usual one because it is what the nonlinear terms conserve, and a different norm gives a different $G$. The Reynolds-number scaling survives the choice; the constants do not.

Who found it, and when #

Reynolds’ own 1883 experiments established that the transition Reynolds number depends on the care taken, which is the observation this essay is ultimately about. Orr and Sommerfeld set up the modal problem in 1907 and 1908, and by the mid-twentieth century it was clear that its predictions and the experiments disagreed badly for pipes and for Couette flow, both of which are stable to every mode and turbulent in practice.

The resolution was assembled in the late 1980s and early 1990s. Farrell’s work on optimal perturbations, Butler and Farrell’s computation of the optimal growth in shear flows, and the 1993 Science paper by Trefethen, Trefethen, Reddy and Driscoll — “Hydrodynamic stability without eigenvalues” — put the non-normality at the centre, and the two-by-two example above is theirs. Waleffe and others then supplied the nonlinear half, the self-sustaining process that keeps the streaks and vortices regenerating once they are large.

Where the ladder goes next #

This field’s three rungs on transition now say that the laminar solution stays a solution, that the number it stops being taken at is not a property of the fluid, and that the linear operator’s eigenvalues were asking the wrong question. What is still missing is what the disturbance becomes, and that is the nonlinear problem this site cannot solve — the same boundary it has met in the wake and in the cascade.