High-Frequency Corrections, Reynolds Stress Cancellation, and Low-to-High Blowup Feedback

Good to continue. In the previous lesson, we transferred the Boussinesq wave-amplitude mechanism to axisymmetric Euler near $r=1$: a transported scalar develops a large physical gradient through deformation, that gradient generates reduced vorticity, and the induced velocity regenerates the scalar perturbation against a lower-frequency background gradient.

This lesson puts that mechanism beside a different, equally important use of high frequencies in fluid PDE: convex integration. Both constructions involve carefully designed oscillations at ever higher frequencies. But their logic is nearly opposite. Convex integration uses high-frequency velocity perturbations so that their quadratic averages cancel a pre-existing low-frequency defect. The blowup constructions use a low-frequency background to drive a high-frequency wave packet toward ever larger derivatives.

---

1. Two appearances of “high-frequency corrections”

It is easy to see why the two methods are sometimes mentally grouped together. In both cases, one encounters:

• a sequence of increasingly fine spatial scales;
• oscillatory velocity fields;
• nonlinear interactions between oscillations;
• a careful management of residual errors.

But the question is: what is the oscillation for?

In the Boussinesq and axisymmetric Euler blowup constructions studied so far, the high-frequency packet is a dynamical component of a smooth solution. Its purpose is to undergo amplification. The central local feedback was schematically

$$\dot{A}\approx -i\frac{q^{\perp }\cdot G}{|q{|}^2}B,\dot{B}\approx i{q}_{\parallel }A,$$

where $A$ is a scalar-wave amplitude, $B$ is a vorticity-wave amplitude, $G$ is a background scalar gradient, and $q_{\parallel }$ denotes the derivative direction appearing in the vorticity source. In Boussinesq, $q_{\parallel }=q_x$; in the local Euler correspondence, it is $r^{-4}{q}_z$.

This is a dynamical feedback law. A scalar gradient makes vorticity; vorticity induces velocity; velocity displaces scalar level sets relative to a background gradient; this recreates the scalar perturbation. Along with deformation of $q$, this produces increasingly large derivative norms.

Convex integration begins from a different object: not an exact Euler solution that one wants to amplify, but an approximate solution with a controlled defect that one wants to remove.

---

2. The Euler-Reynolds system: exact conservation law plus a defect

The incompressible Euler equations are

$${\partial }_{t}v+\nabla \cdot (v\otimes v)+\nabla p=0,\nabla \cdot v=0.$$

A convex-integration scheme instead first permits an error tensor $R_q$:

$${\partial }_t{v}_q+\nabla \cdot (v_q\otimes v_q)+\nabla p_q=\nabla \cdot R_q,\nabla \cdot v_q=0.$$

This is the Euler-Reynolds system. The symmetric tensor $R_q$, usually taken trace-free after absorbing its trace into pressure, records precisely how far $(v_q,p_q)$ is from satisfying Euler.

It is useful to read $R_q$ operationally:

$R_q$ is the unresolved momentum-flux correction required to make the coarse velocity $v_q$ obey the momentum balance.

If $R_q=0$, the approximate solution is an exact Euler solution. Convex integration constructs a sequence

$$(v_q,R_q),$$

with two simultaneous properties:

1. $R_q$ becomes smaller at each stage.
2. $v_q$ converges to a limiting velocity $v$.

The limit solves Euler weakly because the defect tends to zero in a suitable norm or distributional sense.

The new velocity is written

$$v_{q+1}=v_q+w_{q+1},$$

where $w_{q+1}$ is concentrated at a much higher frequency ${\lambda }_{q+1}$ than the frequency scale ${\lambda }_q$ of $v_q$ and $R_q$.

The key design problem is not “how can the new wave grow?” It is:

How can one choose $w_{q+1}$ so that its quadratic self-interaction cancels $R_q$ at coarse scales?

---

---

3. How high frequencies create a low-frequency stress

The basic algebra can be seen with one idealized oscillation. Let

$$w(x)=ab\sin (\lambda \xi \cdot x),$$

where $a$ is an amplitude, $\xi$ is a wavevector, and $b\cdot \xi =0$, so the wave is divergence-free when $a$ and $b$ are constant.

Its quadratic tensor is

$$w\otimes w=a^2(b\otimes b){\sin }^2(\lambda \xi \cdot x).$$

Using

$${\sin }^{2}s=\frac{1}{2}-\frac{1}{2}\cos (2s),$$

we obtain

$$w\otimes w=\frac{a^2}{2}b\otimes b-\frac{a^2}{2}\cos (2\lambda \xi \cdot x)b\otimes b.$$

Even though $w$ oscillates at frequency $\lambda$, its self-product has:

• a component at frequency $2\lambda$;
• a constant, zero-frequency component.

Crucially, the constant term has size $a^2$, independent of $\lambda$. So arbitrarily fine oscillations can exert an order-one coarse-grained quadratic effect.

For a real convex-integration perturbation, one uses several divergence-free building blocks $W_{\xi }$, with slowly varying amplitudes $a_{\xi }(x,t)$. Schematically,

$$w_{q+1}^{(p)}\approx \sum _{\xi }{a}_{\xi }(R_q)W_{\xi }.$$

The amplitudes and directions are chosen so that the low-frequency part satisfies

$$P_{\le {\lambda }_q}\left(w_{q+1}^{(p)}\mathring{\otimes }{w}_{q+1}^{(p)}\right)\approx -R_q.$$

Here $\mathring{\otimes }$ denotes the trace-free part of the tensor product, and $P_{\le {\lambda }_q}$ means “retain the coarse modes at or below the previous frequency scale.”

There is a subtle point behind the trace-free notation. A single tensor $w\otimes w$ is positive semidefinite, whereas a general Reynolds stress need not be. The freedom to adjust pressure resolves the apparent contradiction. Since

$$\nabla \cdot (\rho I)=\nabla \rho ,$$

an isotropic tensor $\rho I$ can be moved into the pressure. Thus one seeks a representation of the form

$$⟨w_{q+1}\otimes w_{q+1}⟩\approx \rho I-R_q,$$

with $\rho$ chosen large enough that the right-hand side is positive definite. After taking trace-free parts, this gives the desired cancellation of $R_q$.

So the underlying convex-integration move is:

1. Inspect the coarse defect $R_q$.
2. Select oscillation directions and amplitudes that encode $-R_q$ in a quadratic average.
3. Add a much finer velocity perturbation.
4. Let its low-frequency quadratic component neutralize the old defect.

This is why it is reasonable to call the effect a high-to-low spectral transfer or an “infinite-range backward cascade”: two modes at frequencies of size ${\lambda }_{q+1}$ can produce a mode at frequency zero.

That phrase should not be overinterpreted. It refers to the engineered algebra of the product $w_{q+1}\otimes w_{q+1}$, rather than to a physical statement that ordinary fluid energy usually moves from high to low frequencies.

---

4. What remains after the main cancellation

The principal cancellation is not the whole proof. Substitute

$$v_{q+1}=v_q+w_{q+1}$$

into the Euler-Reynolds equation and subtract the equation for $v_q$. Up to pressure rearrangements, the new stress must account for

$$\nabla \cdot \left(w_{q+1}\otimes w_{q+1}+R_q\right),$$ $${\partial }_t{w}_{q+1}+v_q\cdot \nabla w_{q+1},$$

and

$$w_{q+1}\cdot \nabla v_q.$$

These are conventionally called:

Term	Meaning
Oscillation error	The part of $w_{q+1}\otimes w_{q+1}+R_q$ not removed by the intended low-frequency cancellation
Transport error	The fact that a spatially oscillatory perturbation must also evolve consistently in time and under the coarse flow
Nash error	The interaction between the new fine perturbation and the gradient of the old coarse velocity

A suitable inverse-divergence operator converts these residual vector fields into a new stress tensor $R_{q+1}$. High frequency is useful here too: solving a divergence equation for an oscillation at frequency ${\lambda }_{q+1}$ generally gains one factor comparable to ${\lambda }_{q+1}^{-1}$.

The actual estimates are delicate because differentiating $w_{q+1}$ loses factors of ${\lambda }_{q+1}$, while its amplitudes vary on the lower scale ${\lambda }_q$. But conceptually the inductive contract is simple:

$$\Vert R_{q+1}\Vert \ll \Vert R_q\Vert .$$

The new high-frequency oscillation is therefore a defect-cancellation instrument. It is not introduced because the Euler dynamics have naturally amplified it into a growing packet.

---

5. Contrast with the Boussinesq and Euler blowup mechanism

The distinction can now be stated precisely.

Question	Convex integration	Boussinesq / axisymmetric Euler blowup construction
Starting object	An Euler-Reynolds subsolution with defect $R_q$	A smooth forced solution assembled to exhibit a particular dynamical instability
Main role of the high-frequency field	Cancel a pre-existing coarse momentum-flux defect	Participate in a scalar-vorticity amplification loop
Key quadratic relation	$P_{\le {\lambda }_q}(w_{q+1}\mathring{\otimes }{w}_{q+1})\approx -R_q$	$\dot{A}$ is driven by induced velocity against a background gradient, while $\dot{B}$ is driven by a scalar derivative
Role of low frequencies	A low-frequency error to be eliminated	A background gradient and strain field that drive the packet
Dominant spectral direction	High-frequency products deliberately produce a low-frequency contribution	Background structure and deformation generate progressively larger wavevectors and derivatives
Endpoint	A weak Euler solution with vanishing Reynolds stress	Smooth evolution for every $t<T$, with derivative norms diverging as $t$ approaches $T$
Nature of the iteration	Approximate equations converge toward the exact PDE	Layers are assembled to make an exact forced PDE solution display a finite-time singularity mechanism

The most compact contrast is this:

Convex integration sends the effect of high-frequency oscillations back to coarse scales, where it cancels an error. The blowup construction transfers the influence of a coarse background into increasingly fine scales, where derivatives grow.

In the Boussinesq mechanism, the lower-frequency background scalar gradient $G$ is not a defect. It is an active ingredient in the amplification law:

$$\dot{A}\approx -V\cdot G.$$

Likewise, the derivative forcing

$$\dot{B}\approx i{q}_{x}A$$

or, in the Euler correspondence,

$$\dot{B}\approx i{r}^{-4}{q}_{z}A,$$

becomes stronger as the physical wavevector becomes more favorable or larger in the relevant direction.

The high-frequency component does induce a velocity that is smaller in amplitude by one power of frequency,

$$|V|\sim \frac{|B|}{|q|},$$

but its velocity gradient has size

$$|\nabla V|\sim |B|.$$

Thus it can still alter deformation, scalar geometry, and the next stage of the feedback. This is fundamentally different from asking its self-product to reproduce an externally specified coarse tensor.

---

6. Do not confuse “correction” in the two settings

Both literatures use language such as perturbation, correction, ansatz, and error. The words conceal a major structural difference.

In convex integration, $w_{q+1}$ is chosen after examining $R_q$. Its amplitudes depend directly on the defect one intends to cancel. The iteration intentionally exploits weak convergence:

$$w_{q+1}\rightharpoonup 0$$

at coarse scales, while its quadratic term has a nontrivial coarse effect.

In the blowup construction, a new high-frequency layer is chosen to follow a desired approximate wave dynamics. One must also correct residuals and arrange smooth forcing, but there is no sequence of Euler-Reynolds stresses that is being cancelled by the coarse component of a quadratic self-product.

Instead, the desired outcome is that:

• velocity and scalar amplitudes remain controlled;
• the forcing remains smooth through the terminal time;
• successively finer scales fit into a finite interval;
• selected derivatives become unbounded as $t$ approaches $T$.

That is why the construction can produce a finite-time derivative singularity while retaining considerable control over lower-order quantities. Its central architecture is low-to-high amplification, not relaxation of a PDE constraint through weak oscillations.

There is one qualified commonality. Both methods must prevent unwanted interactions among many oscillations. In convex integration, this is largely about suppressing undesirable low modes, transport errors, and cross interactions. In the blowup construction, it is about isolating the intended amplitude feedback, preserving phase cancellations at principal order, and making all assembly errors summable. But the error-control technology serves different mathematical objectives.

---

7. A compact mental model

When reading either type of paper, ask the following.

For convex integration

• What is the current defect $R_q$?
• What coarse tensor must the new perturbation reproduce?
• Which high-frequency building blocks have quadratic averages spanning that tensor?
• Why are the remaining oscillation, transport, and cross-interaction errors smaller?

The governing slogan is:

$${\text{fine oscillations}}^2\text{produce a designed coarse stress.}$$

For the Boussinesq and Euler blowup constructions

• What lower-frequency scalar gradient and strain act on the current packet?
• How does the scalar derivative generate vorticity?
• How does the induced velocity regenerate the scalar perturbation?
• How do deformation and layer scheduling make physical gradients increasingly large?

The governing slogan is:

$$\text{coarse background}\text{drives increasingly fine dynamical structure.}$$

Neither slogan says everything. But they prevent the central conceptual error: treating every high-frequency PDE construction as a version of convex integration.

---

Takeaways

Convex integration for Euler works with the Euler-Reynolds system

$${\partial }_t{v}_q+\nabla \cdot (v_q\otimes v_q)+\nabla p_q=\nabla \cdot R_q.$$

At stage $q+1$, it adds a high-frequency perturbation $w_{q+1}$ whose amplitudes are selected so that the low-frequency part of

$$w_{q+1}\mathring{\otimes }{w}_{q+1}$$

cancels the old Reynolds stress $R_q$. The characteristic high-to-low effect comes from the elementary fact that the product of high-frequency waves contains a constant or low-frequency component.

The Boussinesq and axisymmetric Euler blowup constructions instead exploit a dynamical scalar-vorticity loop. Their lower-frequency background fields are not defects to eliminate; they are the source of the amplification that transfers activity toward progressively finer scales and ultimately unbounded derivatives.

Next, we will use this distinction as part of the final paper-navigation task: annotating Tao’s post claim by claim and locating the relevant theorem, heuristic, and technical estimates in the Boussinesq and Euler papers.