Frequency-Dependent Correction Orders for Mixed Derivative Convergence

Good to continue from the localized layers. The previous lesson explained why a spatial cutoff creates annular residuals, and why a finite correction hierarchy can cancel their oscillatory parts without disturbing the affine cores needed by later layers.

We now address the construction’s decisive regularity problem. The layers occur at ever higher frequencies, so differentiating their residual forces appears dangerous: each spatial derivative of an oscillation can cost a factor comparable to its frequency. Yet the theorem requires the total prescribed forcing to be smooth through the finite terminal time, in every mixed space–time derivative. The solution is a diagonal strategy: as layer frequency grows, the construction also increases the number of correction levels. This pushes the uncancelled oscillatory residual to a sufficiently high “grade” that its smallness beats every derivative loss required at that layer.

---

What must actually be proved

Write the force increment generated by layer $q$ as

$$(f_{\theta ,q},f_{u,q}).$$

For a mixed derivative order $k$, use the compact notation

$$\Vert h{\Vert }_{C_{x,t}^k}=\underset{a+|\gamma |\le k}{\max }\Vert {\partial }_t^a{\partial }_x^{\gamma }h{\Vert }_{L^{\infty }}.$$

The desired layerwise estimate has the form

$$\Vert f_{\theta ,q}{\Vert }_{C_{x,t}^k}+\Vert f_{u,q}{\Vert }_{C_{x,t}^k}\le {\lambda }_q^{-1/2},0\le k\le k_q,$$

where:

• ${\lambda }_q$ is the rapidly increasing frequency of layer $q$;
• $k_q$ is the derivative budget assigned to that layer;
• $k_q$ itself increases without bound as $q$ increases.

This is deliberately not a bound for every derivative order on every layer with one uniform constant. Such a claim would be far stronger than needed and is not the construction’s strategy.

Instead, fix any derivative order $k$. Since $k_q$ eventually exceeds $k$, all sufficiently late layers satisfy the estimate at order $k$. Their bounds are summable because the frequencies grow extremely rapidly:

$$\sum _{q\text{ sufficiently large}}{\lambda }_q^{-1/2}<\infty .$$

The finitely many early layers are individually smooth, so they cause no issue at the chosen fixed order. Thus the series

$$f_{\theta }=\sum _q{f}_{\theta ,q},f_u=\sum _q{f}_{u,q},$$

converges in $C^k$. Since this reasoning works for every fixed $k$, the total force is $C^{\infty }$ in both space and time, including through the terminal time.

This is a familiar diagonal principle in analysis:

1. At stage $q$, guarantee only finitely many derivative estimates, up to $k_q$.
2. Make $k_q$ tend to infinity.
3. Make the corresponding bounds summable across layers.
4. For any particular derivative order, discard the finite initial segment before applying the uniform tail estimate.

The main work lies in making the third point true despite high-frequency differentiation.

---

Three indices that should not be conflated

The paper has three separate notions of “order.”

Quantity	Meaning	Why it grows
$q$	Layer index	Later layers operate on smaller spatial and temporal scales.
$n$	Correction grade	Each correction cancels another level of oscillatory localization residual.
$k$	Physical derivative order	Counts mixed derivatives in $t$ and $x$ required of the force.

The key design choice is:

$$d=k_q,J_q=2d+8.$$

Here $J_q$ is the number of correction levels retained in layer $q$. Thus, if a layer is required to have a force estimate through derivative order $d$, the construction takes roughly $2d$ correction levels before leaving an oscillatory remainder.

Why roughly twice as many correction levels as derivatives? A physical derivative can bring back a large factor of ${\lambda }_q$. The remainder must therefore be pushed far enough into a smallness hierarchy that its accumulated gain in powers of a small parameter overcomes those losses.

---

The correction hierarchy creates powers of a small parameter

Recall the correction profiles from the preceding lesson:

$$V_{q,n}(s,a,t),{\Psi }_{q,n}(s,a,t),$$

where $n=0$ is the leading localized wave and $n\ge 1$ are corrections. The variable $s$ is the periodic fast phase and $a$ is the material coordinate.

The paper packages the relevant layer scale into

$${\delta }_q={\Pi }_q^5{\lambda }_q^{-7/8}.$$

Here ${\Pi }_q$ is a comparatively slow lifetime/deformation scale, while ${\lambda }_q$ is the dominant fast frequency. The frequency schedule is chosen so that ${\delta }_q$ is very small; schematically,

$${\delta }_q\lesssim {\lambda }_q^{-3/4}.$$

The leading amplitude is also small:

$$Y_q=\underset{t\in I_q}{\sup }|y_q^0(t)|\lesssim {\lambda }_q^{-7/8}.$$

The exact exponents are less important than the architecture:

• the principal layer remains large enough in its gradient to seed the next stage;
• its field amplitude is small;
• every correction level gains one further factor of ${\delta }_q$.

In an appropriate weighted phase/material norm, the level estimates take the schematic form

$$[V_{q,n}{]}_{r,q}+[{\mathcal{W}}_{q,n}{]}_{r,q}\lesssim C(n,r)Y_q{\delta }_q^n.$$

The constants $C(n,r)$ depend on the finite correction grade and finite derivative order, but crucially not on $q$ or ${\lambda }_q$.

The source driving the grade-$n$ correction has a companion bound

$$[S_{q,n}{]}_{r,q}\lesssim s(n,r)Y_q{\delta }_q^n{\Pi }_q^{-2}.$$

The factor ${\Pi }_q^{-2}$ offsets the finite-time cost of solving the inhomogeneous amplitude ODE. So when the source is propagated through the same two-dimensional amplitude system used for the leading wave, the resulting correction still has the expected size $Y_q{\delta }_q^n$.

Conceptually, each correction does the following:

1. It takes the nonzero-fast-phase part of the residual left by earlier profiles.
2. It solves a phasewise forced version of the same amplitude system.
3. It cancels that oscillatory residual at its assigned grade.
4. It leaves only phase averages and terms of still higher grade.

Because phase inversion acts only in $s$, it does not spread the correction in the material coordinate $a$. Thus every higher correction remains supported in the cutoff annulus and vanishes on the plateau, preserving the nested affine cores.

---

---

Why physical derivatives are expensive

A profile becomes a physical field through the evaluation

$$A^{♮}(x,t)=A({\lambda }_q{\zeta }_q(t)\cdot x,M_q(t{)}^{-1}x,t).$$

A spatial derivative therefore acts schematically as

$${\nabla }_x\sim {\lambda }_q{\zeta }_q{\partial }_s+M_q^{-T}{\nabla }_a.$$

The first term differentiates the fast phase and costs a factor of ${\lambda }_q$. Time differentiation is also nontrivial because the phase, material map, and coefficients evolve. The paper proves the conservative schematic estimate

$$\Vert A^{♮}{\Vert }_{C_{x,t}^k}\lesssim ({\lambda }_q{\Pi }_q^2{)}^k[A{]}_{k,q}^{♯}.$$

For a genuinely oscillatory remainder, one should therefore fear a factor approximately ${\lambda }_q^k$.

This is precisely why stopping after a fixed number of corrections would fail. Suppose the remainder were merely of size ${\delta }_q^N$ for some fixed $N$. For a sufficiently high derivative order $k$, the factor ${\lambda }_q^k$ would eventually dominate that fixed gain. One would obtain smoothness only up to a finite order, not $C^{\infty }$ forcing.

The construction instead sets the correction depth proportional to the desired derivative order:

$$J_q=2k_q+8.$$

The first oscillatory remainder then has grade

$$J_q+1=2k_q+9.$$

At physical derivative order $k\le k_q$, its dominant scaling is schematically bounded by

$$({\lambda }_q{\Pi }_q^2{)}^k{\lambda }_q{Y}_q{\delta }_q^{2k_q+9}.$$

The extra factor ${\lambda }_q$ reflects the vorticity-side normalization; the scalar estimate is slightly easier. Ignore the harmless ${\Pi }_q$ factors momentarily and use ${\delta }_q\lesssim {\lambda }_q^{-3/4}$. Then the dangerous part is controlled by

$${\lambda }_q^k{\delta }_q^{2k_q}\lesssim {\lambda }_q^k{\lambda }_q^{-3k_q/2}.$$

Since $k\le k_q$,

$${\lambda }_q^k{\lambda }_q^{-3k_q/2}\le {\lambda }_q^{-k_q/2}.$$

So the very correction depth chosen to protect derivatives through order $k_q$ creates a substantial remaining negative power of frequency. The additional ${\delta }_q^9$, the small amplitude $Y_q$, and the controlled ${\Pi }_q$-powers provide still more margin.

This is the core mechanism:

$$\boxed{\text{Each extra derivative budget requires extra correction grades,}\text{and those grades produce enough frequency decay to pay for differentiation.}}$$

---

The terms that are not canceled oscillatory remainders

The final residual of a layer is not just one high-grade oscillatory error. It has three qualitatively different pieces:

$$\text{force increment}=\text{activation contribution}+\text{phase means}+\text{high-grade remainder}.$$

Each is controlled differently.

Activation terms

The activation cutoff $w_q(t)$ turns a layer on smoothly. Its derivative produces terms such as

$${\dot{w}}_q{\Theta }_{q}Fg,{\dot{w}}_q{\Omega }_q{F}^{\prime }g.$$

These are not spatial-localization errors and need not vanish in the affine core during activation. But activation occurs while the layer’s seed amplitude is chosen exceptionally small, schematically of size

$${\lambda }_q^{-k_q-6}.$$

That exponent is intentionally tied to the derivative budget. Even after up to $k_q$ expensive physical derivatives, the activation residual remains tiny.

Nonoscillatory phase averages

At every correction stage, the phase average is retained rather than inverted:

$$⟨S_{q,n}⟩.$$

The decisive benefit is that a phase average is independent of $s$. Therefore the fast part of ${\nabla }_x$,

$${\lambda }_q{\zeta }_q{\partial }_s,$$

annihilates it. Its physical derivatives cost only powers of ${\Pi }_q^2$, not powers of ${\lambda }_q$:

$$\Vert A^{♮}{\Vert }_{C_{x,t}^k}\lesssim {\Pi }_q^{2k}[A{]}_{k,q}^{♯},A_s=0.$$

This is why the construction can leave phase means in the prescribed forcing rather than canceling them.

Parity gives additional help. The scalar phase means begin only at grade $3$, while the vorticity phase means begin only at grade $4$. Hence they already contain factors such as ${\delta }_q^3$ or ${\delta }_q^4$, with no compensating ${\lambda }_q^k$ from physical differentiation.

High-grade oscillatory remainder

This is the residual that does incur fast-derivative losses. It is exactly the component handled by choosing

$$J_q=2k_q+8.$$

It starts so late in the grading hierarchy that its factor ${\delta }_q^{2k_q+9}$ overwhelms the worst physical differentiation cost through order $k_q$.

---

Constants also need a diagonal treatment

There is one more subtlety. The constants in repeated product rules, phase inversions, Duhamel estimates, coordinate changes, and compact inverse-curl recovery become worse as the target derivative order $d$ rises.

The paper collects all constants needed up to order $d$ into a finite majorant $K(d)$. It then imposes the condition

$$K(d)\le \log {\lambda }_q.$$

At first this can look cosmetic, but it captures the logic of the parameter selection. For every fixed $d$, the correction algebra requires only finitely many operations and produces a finite $K(d)$. By making ${\lambda }_q$ sufficiently large, even that finite constant can be absorbed into a small negative power of ${\lambda }_q$:

$$K(d){\lambda }_q^{-3/2}\le (\log {\lambda }_q){\lambda }_q^{-3/2}\le {\lambda }_q^{-1/2}.$$

The precise exponent gap is not important. What matters is that a logarithm is negligible compared with any positive power of frequency.

This explains why the argument does not need a derivative estimate uniform in $d$. It only needs:

• $K(d)<\infty$ for every fixed $d$;
• a frequency schedule sufficiently large to dominate $K(d)$;
• an increasing sequence $k_q$ of derivative budgets.

That is the analytical meaning of “increase correction order with frequency.”

---

---

From curl-form residual to the actual vector force

The correction calculation naturally produces a vorticity-form residual $E_q$, which is the desired curl of the vector-force increment. To recover a compactly supported divergence-compatible vector force, the paper applies a fixed compact inverse-curl operator:

$$f_{u,q}={\mathcal{R}}_c{E}_q.$$

This step matters for the final smoothness statement. The inverse-curl recovery is designed so that it preserves the relevant derivative order:

$$\Vert {\mathcal{R}}_c{E}_q{\Vert }_{C^k}\lesssim \Vert E_q{\Vert }_{C^k}.$$

It therefore does not undo the smallness gained by the correction hierarchy. The zero-integral property required for this recovery is checked for the complete vorticity residual, not independently for every displayed subterm.

Thus the construction controls both components of the physical forcing:

$$f_{\theta ,q}\text{and}{f}_{u,q}.$$

---

The final summation argument

Once Theorem 7.3 has produced

$$\Vert f_{\theta ,q}{\Vert }_{C_{x,t}^k}+\Vert f_{u,q}{\Vert }_{C_{x,t}^k}\le {\lambda }_q^{-1/2},k\le k_q,$$

the endpoint smoothness argument is short.

Fix $k\ge 0$. Choose $q_0$ so that

$$k\le k_q\text{for all }q\ge q_0.$$

Then

$$\sum _{q\ge q_0}\left(\Vert f_{\theta ,q}{\Vert }_{C_{x,t}^k}+\Vert f_{u,q}{\Vert }_{C_{x,t}^k}\right)\le \sum _{q\ge q_0}{\lambda }_q^{-1/2}<\infty .$$

The tail converges uniformly together with every mixed derivative of total order at most $k$. The finite prefix $q<q_0$ is smooth by the fixed-layer estimates. Therefore the full forcing is $C^k$. Since $k$ was arbitrary, it is $C^{\infty }$.

This is stronger than saying that the force is smooth for every $t<T_{\ast }$. The uniform convergence of each fixed mixed derivative allows the force to extend smoothly through the accumulation time $T_{\ast }$, even though the solution state itself develops an unbounded temperature gradient and vorticity before that time.

---

Takeaways

The correction hierarchy is a regularity device, not just a way to tidy up cutoff errors.

• A layer has frequency ${\lambda }_q$, so physical derivatives of oscillatory terms can cost powers of ${\lambda }_q$.
• Grade-$n$ corrections are suppressed by approximately ${\delta }_q^n$, with ${\delta }_q$ a negative power of ${\lambda }_q$.
• To control force derivatives through order $k_q$, the construction takes correction depth

$$J_q=2k_q+8.$$

• The remaining oscillatory error then begins at grade $2k_q+9$, whose smallness defeats all differentiation losses through order $k_q$.
• Activation terms are controlled by very small seed amplitudes; phase means are cheaper because they have no fast-phase dependence.
• Although the bounds are only uniform through a layer-dependent finite derivative order, $k_q$ increases without bound. This diagonal structure makes the force series converge in every fixed $C^k$ norm.

The next lesson begins the transfer to axisymmetric three-dimensional Euler flow with swirl: we will derive the circulation and reduced-vorticity coupling in the meridional variables, identifying the Euler analogue of the Boussinesq amplification mechanism.