Strong diffeomorphic approximation in three dimensions

Source and verification

This article explains family 368 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

Smoothing a deformation is easy if self-intersections are allowed. Family 368 claims strong Sobolev approximation by smooth diffeomorphisms in three dimensions, preserving the exact target domain. Its two manuscripts use different constructions below and above the exponent two.

🏷️ Preliminaries: topology and Sobolev convergence

A homeomorphism $f:\Omega \to \Lambda$ is a continuous bijection with continuous inverse. A diffeomorphism requires both directions to be smooth. In nonlinear elasticity, global injectivity represents noninterpenetration of material, while the matrix $Df$ records local deformation.

The space $W^{1,p}(\Omega ;{\mathbb{R}}^3)$ consists of maps and weak first derivatives in $L^p$. Strong convergence asks for

$$\begin{array}{c}{\int }_{\Omega }(|f_j-f{|}^p+|D{f}_j-Df{|}^p)\\ ⟶0.\end{array}$$

Convolution controls this norm locally but does not automatically preserve injectivity or the image. Even uniform convergence of maps is too weak to control derivatives: on an interval, $f_j(x)=x+\sin (jx)/(2j)$ converges uniformly to $x$, whereas $f_j^{\prime }(x)=1+\cos (jx)/2$ does not converge strongly to one in $L^p$. The positive derivative in this example even preserves injectivity, isolating the missing analytic requirement.

🏷️ The combined approximation claim

Family 368, combining its two manuscripts

Let $\Omega ,\Lambda \subset {\mathbb{R}}^3$ be bounded domains and let $f:\Omega \to \Lambda$ be a homeomorphism in $W^{1,p}$, $1\le p<\infty$. There exist smooth diffeomorphisms $f_j:\Omega \to \Lambda$ converging strongly to $f$ in $W^{1,p}$. The first paper treats $1\le p\le 2$; the second treats $2<p<\infty$. (OpenAI, 2026a; OpenAI, 2026b)

Both domains may have irregular boundaries. Smoothness is asserted on their open interiors, with no prescribed boundary extension. Every approximant is onto the same target. No inverse Sobolev hypothesis or positive lower bound for the Jacobian is assumed.

Smoothness of each inverse does not imply Sobolev convergence of those inverses. Nor does strong convergence of $D{f}_j$ automatically control elastic energies containing singular penalties as $\det D{f}_j\to 0$.

🏷️ Why rank-deficient derivatives cannot be discarded

A matrix of rank one or two may have a large norm even though its determinant vanishes. Thus a subset where $Df$ loses rank can carry substantial $\int |Df{|}^p$. Treating the whole rank-deficient set as negligible would lose the desired strong approximation.

The source retains compact pieces on which the first-order behavior is controlled and separates them by rank. Full-rank pieces admit locally invertible affine models. The difficult pieces retain variation in one or two directions while compressing the remaining directions. The replacement must recover those nonzero derivatives, not merely preserve a coarse energy bound.

For $1<p<\infty$, uniform convexity explains one useful principle. If gradients converge weakly and their $L^p$ norms converge to the norm of the target gradient, then convergence is strong. The high-exponent construction aims to pin the mean gradients while obtaining nearly sharp energy estimates. At $p=1$, this principle is unavailable, so the endpoint needs direct geometric control.

🏷️ The two constructions

For $p\le 2$, the manuscript measures images of selected curves and, at the critical exponent, surfaces. Intersections with dual pieces of a target triangulation supply length and area budgets. Reparametrization converts these counts into quantitative estimates for a locally bi-Lipschitz replacement. Rank-specific local models recover the derivative on protected pieces, while trace-controlled fillings complete the remaining cells.

The exponent two is significant because surfaces are two-dimensional: radial fillings that are inexpensive below the critical exponent need a sharper disk argument at the endpoint. The source uses additional information about planar slices and the inverse at $p=2$.

For $p>2$, the proposed construction represents retained rank-one and rank-two data through two scalar coordinates on a polyhedral surface complex. Carefully routed source surfaces realize their levels. Reference paths constrain the realized labels, and controlled parameter slopes give the energy estimates. Nearly sharp norms together with pinned mean derivatives yield strong approximation.

Both constructions first produce homeomorphisms with quantitative local control, then apply a separate smoothing argument. Locally finite tolerances permit errors tending to zero toward irregular boundaries. Image-preserving local replacements keep surjectivity onto $\Lambda$. A qualitative topological replacement alone would not establish any of these derivative bounds.

🏷️ Key steps and proof outlines

The subcritical cell estimate explains the exponent split

The low-exponent construction compresses most of a source cell toward its boundary while keeping a tiny interior core to preserve injectivity. In an $m$-dimensional cell, let $\alpha$ be the core radius and let the radial reparametrization expand that core to almost the whole original cell. Its derivative is $O({\alpha }^{-1})$ on volume $O({\alpha }^m)$, so the core energy is

$$O({\alpha }^{m-p}).$$

Outside the core, send radii into a thin neighborhood of the original boundary. The radial derivative tends to zero; the tangential derivative tends to the boundary derivative divided by radius. Its energy is controlled by

$$\left({\int }_{\alpha }^1{r}^{m-1-p}dr\right){\int }_{\partial D}|D_{t}F{|}^p.$$

Both costs are controlled precisely when $p<m$. Restoring cell size $\ell$ yields the estimate ${\int }_D|DF{|}^p\le C\ell {\int }_{\partial D}|D_{t}F{|}^p+\epsilon$. Boundary values are unchanged, so adjacent cells glue exactly. This calculation works on three-dimensional cells for $p\le 2$, but on two-dimensional faces only for $p<2$. The endpoint face treatment consequently requires the additional disk parametrization and trace-budget arguments in the source; it is not obtained by substituting $p=2$ into a divergent radial integral.

Collapsing onto a lower-dimensional stratum needs a tangential chain rule

Suppose Lipschitz maps $q_j$ converge uniformly to $q_0$, with derivatives converging almost everywhere and uniformly bounded. The limit $q_0$ may have deficient rank. One cannot pull back the ambient exceptional set of an arbitrary Lipschitz outer map and assume it remains null.

The source instead arranges that the outer map $v$ agrees with finitely many $C^1$ branches on closed pieces. On the preimage of each piece, locality of weak derivatives gives $D(v\circ q_j)=D{v}_{\nu }(q_j)D{q}_j$ almost everywhere. Any branch used infinitely often contains the limiting point by closedness. All branches there give the same tangential product with $D{q}_0$. Finite branching and dominated convergence then prove strong convergence of $D(v\circ q_j)$ in every finite $L^r$. This is why piecewise smooth compatibility is retained throughout the construction. (OpenAI, 2026a s “Subcritical collapse onto boundary data” and “Composition at a collapsed stratum”)

Routing preserves parameter length without accumulating derivative factors

For $p>2$, thin walls encode output coordinates and their derivatives. At an inessential circular crossing of two wall collars of widths $a,b$, use the diagonal coordinate

$$d=a-u+v,(u,v)\in [0,a]\times [0,b].$$

Its regular level segments match incoming and outgoing boundary intervals by affine maps of absolute slope one. Repeated routing therefore introduces no product of large derivative factors. Essential arc crossings retain both coordinates and are handled separately.

The resulting assembly can be infinite, so local finiteness needs proof. Each reached state has a unique predecessor. Two distinct routing itineraries cannot arrive at the same local parameter of the same piece. Their image intervals are disjoint, and length preservation gives

$$\sum _j{\int }_0^{w_j}{N}_P(j,\lambda )d\lambda \le w_{k(P)},$$

where $N_P$ counts visits to piece $P$ and $w_{k(P)}$ is its collar width. Summing over the finitely many original collars meeting a compact set proves that almost every root parameter visits only finitely many of them. This makes the routed surfaces proper and permits finite truncation and capping when checking their prescribed boundary crossing data.

Sharp energy and the mean gradient force derivative accuracy

A coarse derivative bound alone would not prove strong approximation. On a calibrated patch, the construction normalizes the desired rank-two derivative to a matrix $D_0$ of Frobenius norm $\sqrt{2}$; for rank one its norm is close to one. The wall estimates give an upper bound for $|Z|$ close to that same value, where $Z$ is the actual normalized derivative. Label pinning controls its average: integrating a derivative difference over the cube converts it to a boundary value error divided by the cube’s side length.

For $p>2$, uniform convexity supplies

$$c_p|Z-D_0{|}^p\le |Z{|}^p-|D_0{|}^p-p|D_0{|}^{p-2}{D}_0:(Z-D_0).$$

After integration, the sharp energy bound controls the first difference and the mean-gradient estimate controls the linear term. Hence $Z$ is close to $D_0$ in $L^p$, including at rank-deficient points. The physical-coordinate conversion is calibrated so its leading factor is bounded by $C|Df|$, rather than an uncontrolled inverse singular value. Summing patches therefore charges the error to the original finite Sobolev energy.

The source finally replaces collapsed output parameters by strictly increasing ones, pays the remaining collar and exceptional-region budgets, and smooths the resulting homeomorphism. Every operation preserves its prescribed target image. These topology, wall-realization, and smoothing inputs are substantial separate lemmas; the estimates above explain how they deliver strong derivative convergence without discarding deficient-rank sets. (OpenAI, 2026b s “Diagonal routing and conservation of periods” and “Saturation and physical gradient accuracy”)

See Also

• Family 372 introduces the strain tensor of a deformation in linear elasticity.
• Family 374 studies quantitative dependence of another distinguished class of maps on their target measures.

🏷️ Manuscripts and verification scope

• Strong diffeomorphic approximation in three dimensions for 1≤p≤2 — pinned manuscript and companion source.
• Strong diffeomorphic approximation in three dimensions for p>2 — pinned manuscript and companion source.

No family-specific scope page was available at lean/docs/368.md in this snapshot. This records the catalogue lookup, not a claim that no related Lean code exists. No Lean build or axiom audit was performed.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. Strong diffeomorphic approximation in three dimensions for 1≤p≤2.

🐻  OpenAI 2026b. Strong diffeomorphic approximation in three dimensions for p>2.