Metric and connection recovery from one boundary patch

Source and verification

This article explains family 365 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

Boundary measurements can conceal changes of coordinates and changes of a vector bundle’s frame. Family 365 claims recovery of smooth geometric coefficients up to these unavoidable ambiguities, even from one boundary patch. A separate companion claims nonuniqueness for bounded measurable scalar conductivities, emphasizing the role of regularity.

🏷️ Preliminaries: boundary energy measurements

Begin with a scalar conductivity $\gamma >0$ in a domain. Applying boundary voltage $f$ produces the weak solution of

$$\mathrm{div}(\gamma \nabla u_f)=0,u_f{|}_{\partial \Omega }=f.$$

The Dirichlet-to-Neumann map records the outgoing flux. It can be defined without pointwise derivatives through the energy pairing

$$⟨{\Lambda }_{\gamma }f,h⟩={\int }_{\Omega }\gamma \nabla u_f\cdot \nabla u_h.$$

For a Riemannian metric $g$, the corresponding energy is $\int ⟨d{u}_f,d{u}_h{⟩}_{g}d{V}_g$. The geometric problem asks whether this boundary information determines the interior metric.

A unitary connection on the trivial rank-two complex bundle adds a matrix-valued one-form $A$ with skew-Hermitian coefficients. Its covariant derivative is $d_A=d+A$, and its connection Laplacian is $d_A^{\ast g}{d}_A$. This only requires understanding differentiation of a two-component vector field in a varying orthonormal frame; no classification of groups or bundles is needed here.

🏷️ Smooth recovery and rough nonuniqueness

Joint recovery claimed in the newest manuscript

Let $M$ be compact, connected, smooth, of dimension $n\ge 3$, with smooth nonempty boundary. On the trivial Hermitian rank-two bundle, suppose two smooth metrics and smooth unitary connections have identical zero-frequency boundary energy pairings for all inputs and observations supported in the same nonempty proper open patch $\Gamma$. Then they differ by a diffeomorphism and a unitary change of frame, both equal to the identity on $\Gamma$. (OpenAI, 2026)

The scalar geometric companion claims the analogous same-patch recovery of a smooth metric up to diffeomorphism, as well as smooth scalar conductivity uniqueness. (OpenAI, 2026b) The data are the energy pairing, including the boundary density; one must not silently assume a known boundary metric while interpreting them.

The rough companion instead claims two distinct uniformly positive bounded measurable scalar conductivities on $B(0,3)\subset {\mathbb{R}}^3$, equal to one near the boundary, with identical full weak Dirichlet-to-Neumann maps. (OpenAI, 2026b) This is a different regularity class, not a contradiction between the stated smooth and rough conclusions.

🏷️ A direct calculation of the gauge ambiguity

For a smooth unitary matrix field $U$, put

$$\begin{array}{c}{A}^{\prime }=U^{-1}AU+U^{-1}dU,\\ u^{\prime }=U^{-1}u.\end{array}$$

The product rule gives

$$d_{A^{\prime }}{u}^{\prime }=U^{-1}{d}_{A}u.$$

Indeed, the derivative of $U^{-1}$ cancels the $U^{-1}dU$ term. Unitarity preserves the fiber inner product, so the energy is unchanged. If $U=I$ on the measured patch, the input and output labels there are unchanged too. A diffeomorphism fixing the patch preserves the same energy by change of variables.

Consequently, literal equality of the coefficients is too strong a target. The claimed theorem says precisely that these transformations exhaust the ambiguity:

$$\begin{array}{rl}{g}_2& ={\Phi }^{\ast }{g}_1,\\ A_2& =U^{-1}({\Phi }^{\ast }{A}_1)U+U^{-1}dU.\end{array}$$

🏷️ The proposed continuation mechanism

Boundary symbol analysis first recovers coefficient derivatives on the patch after a normal-coordinate and frame choice. Attaching a small exterior region converts the boundary data into agreement of Green kernels on an open set. Source solutions there identify points by their evaluations and identify fibers by their vector values.

The difficult step is to continue a local identification across its possible frontier. The manuscripts use Green kernels in two variables: quantitative unique continuation in each variable enlarges the region of agreement. Transfers of harmonic functions across shrinking spheres are represented by densities whose angular behavior must stay controlled as the radius tends to zero.

For scalar functions, coordinate normalization removes first moments. For vector-valued functions, a matrix first moment remains. The joint-recovery manuscript derives an overdetermined differential system for its nonscalar part. Controlling that system improves the error in normalized inverse metrics from order $r$ to order $r^{3/2}$. The improved power enables the proposed limiting continuation and subsequent global identification.

These are the principal analytic claims being explained, not consequences of the elementary gauge calculation. In particular, the rough counterexample’s formalization entry cannot serve as verification of smooth recovery.

🏷️ Key steps and proof outlines

Shrinking harmonic moments produce a superlinear metric estimate

The smooth scalar manuscript first converts equality of boundary responses into corresponding harmonic functions near a possible frontier of the matched region. A finite family of their gradients and Hessians determines the relevant normalized principal tensor. Let $h$ be its difference and $H_8$ a norm of derivatives through order eight. The shrinking-sphere argument works under the provisional bound $H_8\le \epsilon r_{\delta }$, where $r_{\delta }$ is a positive radius profile.

Its output is an $L^2$ estimate $\Vert V{\Vert }_{L^2(B(y,s))}\le C{r}^2$ for the renormalized harmonic differences, with $r=r_{\delta }(y)$ and $s=r^{\eta }$. To recover derivatives, expand $V(y+sz)$ into a Taylor polynomial on the unit ball. Equivalence of norms on this finite-dimensional polynomial space gives

$$|{\partial }^{j}V(y)|\le C\left(s^{-j-n/2}\Vert V{\Vert }_{L^2(B(y,s))}+s^{m-j}\Vert V{\Vert }_{C^m}\right).$$

Choose $D=10+n/2$, $\eta =1/(2D)$, and $m\ge \lceil 10+3D\rceil$. For $j\le 10$, both resulting powers of $r$ are at least $3/2$. The finite harmonic-jet system therefore gives $H_8\le C{r}^{3/2}$, a strict improvement at small radius.

A single spatial dilation makes the coefficient difference small at all remaining radii. Fix that dilation before letting $\delta \downarrow 0$. The set of positive $\delta$ satisfying the provisional bound is nonempty, closed, and open by the strict improvement. It is therefore the whole interval. On the fixed neighborhood where $r_{\delta }\to 0$, the same fixed tensor $h$ must vanish. This order avoids changing the unknown metric pair every time the observation radius shrinks. The uniform shrinking-sphere density estimate remains a major analytic input to this argument. (OpenAI, 2026c“From harmonic differences to local metric extension”)

The matrix extension uses a finite-type differential system

For the unitary-connection companion, scalar interpolation alone does not recover matrix displacement. After separating scalar and trace-free matrix parts, the source obtains a perturbed conformal-Killing equation for a vector field $W$ with values in a finite-dimensional real matrix space. Its principal part is the trace-free symmetric gradient.

Why can that equation control all higher derivatives? A tensor in the kernel of its twice-differentiated principal part defines a homogeneous cubic vector field $Z$ satisfying

$${\partial }_i{Z}_j+{\partial }_j{Z}_i=2{\delta }_{ij}\sigma ,\sigma =\mathrm{div}Z/n.$$

Permuting differentiated indices gives ${\partial }_{ij}{Z}_k={\delta }_{jk}{\partial }_i\sigma +{\delta }_{ik}{\partial }_j\sigma -{\delta }_{ij}{\partial }_k\sigma$. Taking divergence yields

$$(n-2){\partial }_{ij}\sigma =-{\delta }_{ij}\Delta \sigma .$$

For $n\ge 3$, tracing forces $\Delta \sigma =0$ and then $D^2\sigma =0$. Thus $D^{3}Z=0$, so the cubic field vanishes. This injectivity supplies a bounded finite-dimensional left inverse for third derivatives. Along a segment from the already matched region, derivatives of $W$ through order two obey a first-order linear system with zero initial data. Grönwall transfers the $O(r^{3/2})$ forcing bound to $W$, after which differentiation recovers the remaining orders.

The first conclusion is conformal equality, $g_2=e^{2\omega }{g}_1$. In harmonic frames let $C_j$ be the connection matrices and $\varphi =(n-2)\omega /2$. Equality of normalized first-order coefficients gives $C_2=C_1-d\varphi I$. Equality of the zeroth-order terms then gives

$$\Delta \varphi +|d\varphi {|}^2=0,\Delta e^{\varphi }=0.$$

Since $e^{\varphi }=1$ on the known side, unique continuation removes the conformal factor. Parallel transport shows that the resulting fiber identification is unitary, although the intermediate harmonic frames need not be. (OpenAI, 2026a s “Finite type and the improved estimate” and “Removing the conformal factor”)

Rough nonuniqueness is an exact transformation of solutions

The rough construction has a different mechanism. Its branching block produces a bounded nonzero $w\in H_0^1$, supported away from the boundary, and a common average $u_{\ast }$ for every bounded $\gamma$-harmonic function $u$. The crucial weighted identity is

$$\int \gamma \nabla w\cdot \nabla ((u-u_{\ast })\phi )=0.$$

This does not assert that $w$ is harmonic. Its nonzero source annihilates these particular products. At finite branching depth, subtract the constant value of $\phi$ on each small cell; equal averages cancel the constant part and the remaining error is bounded by the cell diameter. Strong $H^1$ convergence then gives the identity above.

Choose $\rho =1+\delta w$ with $1/2\le \rho \le 3/2$, and set $\tilde{\gamma }={\rho }^2\gamma$, $\tilde{u}=u_{\ast }+(u-u_{\ast })/\rho$. Direct differentiation gives

$$\tilde{\gamma }\nabla \tilde{u}=\rho \gamma \nabla u-(u-u_{\ast })\gamma \nabla \rho .$$

Testing the original equation with $\rho \phi$ and using the weighted identity cancels the divergence of this flux. Near the outer boundary, both coefficients and both solutions agree. Their weak Dirichlet-to-Neumann pairings therefore coincide. Density extends the equality from smooth boundary data to all $H^{1/2}$ traces. The difficult branching-block and scalarization construction supplies $w$; this calculation explains exactly how its output hides a nontrivial coefficient change from every boundary measurement. (OpenAI, 2026b“An exact transformation of all boundary measurements”)

See Also

• Family 372 studies boundary recovery for a vector-valued equation whose measurements are physical displacements and tractions.
• Family 071 also distinguishes an existence problem from uniqueness modulo geometric transformations.

🏷️ Manuscripts and verification scope

• Determination of a metric and a unitary connection from one boundary patch — pinned manuscript and companion source.
• Smooth anisotropic uniqueness in the Calderón problem from one boundary patch — pinned manuscript and companion source.
• Nonuniqueness for Bounded Measurable Scalar Conductivities in Three Dimensions — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. Despite its broad family heading, the scope page selects only the bounded measurable scalar conductivity counterexample. It does not document formalization of either smooth same-patch uniqueness manuscript or joint metric–connection recovery. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

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

🐻  OpenAI 2026a. Determination of a metric and a unitary connection from one boundary patch.

🐻  OpenAI 2026b. Nonuniqueness for Bounded Measurable Scalar Conductivities in Three Dimensions.

🐻  OpenAI 2026c. Smooth anisotropic uniqueness in the Calderón problem from one boundary patch.