Alexandrov curvature: from distributional tensors to every geodesic

Source and verification

This article explains family 356 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.

🏷️ Metric curvature and weak differential calculus

A lower sectional-curvature bound has two familiar expressions on a smooth Riemannian manifold. One uses the curvature tensor; the other compares geodesic triangles with triangles on a surface of constant curvature. Family 356 asks how to recover this equivalence when the space has singularities. Its companion paper isolates the passage from an integrated Hessian inequality to a statement on every geodesic.

A constant-speed minimizing geodesic is a map $\sigma :[0,1]\to M$ satisfying

$$\begin{array}{c}d(\sigma (s),\sigma (t))=|s-t|\ell ,\\ \ell =d(\sigma (0),\sigma (1)).\end{array}$$

An Alexandrov lower bound $\mathrm{CBB}(\kappa )$ means that sufficiently small geodesic triangles are at least as fat as triangles with the same side lengths in the constant-curvature $\kappa$ model. For a vertex $p$ and a point on the opposite side, the actual distance from $p$ is at least the corresponding model distance. Completeness and an appropriate local comparison theorem then yield global comparison, subject to the usual perimeter restriction when $\kappa >0$.

The analytic setting is a metric-measure space $(M,d,m)$. Its Sobolev energy is obtained by relaxing the squared local slopes of Lipschitz functions in $L^2(m)$. When that energy is quadratic, weak gradients have an inner product, written

$$\begin{array}{c}\Gamma (f,g)=⟨\nabla f,\nabla g⟩,\\ \Gamma (f)=\Gamma (f,f).\end{array}$$

The condition $\mathrm{RCD}(K,N)$ combines this Hilbert-space property with the synthetic curvature-dimension condition $\mathrm{CD}(K,N)$. The latter uses convexity inequalities for entropy along optimal transport of probability measures, with quadratic transport cost. On an unweighted smooth $N$-manifold, it corresponds to a Ricci lower bound $K$. Thus it supplies a background calculus but does not by itself assert a sectional-curvature bound. Here the manuscripts use unreduced $\mathrm{RCD}(K,N)$.

The Laplacian is the nonpositive generator characterized by integration by parts. The original global test class consists of bounded globally Lipschitz functions $g$ in the Laplacian domain with $\Delta g\in W^{1,2}$. It does not require $\Delta g$ to be bounded. Test vector fields are finite sums $a\nabla b$ with test functions $a,b$. Their weak covariant derivatives allow one to define curvature by integrating the usual second-covariant-derivative formula by parts. The result $R(X,Y,Y,X)(f)$ is a functional of a test weight $f$, rather than a curvature value at each point. (OpenAI, 2026)

The characterization stated in the main manuscript

Fix an integer $n\ge 2$ and $\kappa \in \mathbb{R}$. A complete separable metric space is an $n$-dimensional Alexandrov space of curvature at least $\kappa$ if and only if its exact reference measure $m={\mathcal{H}}_d^n$ has full support, the space is $\mathrm{RCD}((n-1)\kappa ,n)$, and its distributional curvature satisfies the sectional lower bound below for all original global test fields and all nonnegative original global test weights.

Write the squared area of the plane spanned by $X,Y$ as

$$|X\wedge Y{|}^2=|X{|}^2|Y{|}^2-⟨X,Y{⟩}^2.$$

The required inequality is

$$\begin{array}{rl}& R(X,Y,Y,X)(f)\\ & \ge \kappa \int f|X\wedge Y{|}^{2}dm.\end{array}$$

The sign convention makes $R(X,Y,Y,X)$ positive on positively curved spheres. The reference measure is not an arbitrary measure equivalent to Hausdorff measure. Likewise, a tensor inequality only on a smooth regular subset, or only for compactly supported fields, would not state the full claimed theorem. (OpenAI, 2026)

🏷️ A Hessian inequality with only first derivatives of the function

For a bounded globally Lipschitz function $F$, an $L^2$ Hessian need not exist. Nevertheless, for a compactly supported test function $g$, set $Y=\nabla g$. The smooth identity for $\mathrm{Hess}F(Y,Y)$ suggests the distribution

$$\begin{array}{rl}& H_F(Y,Y)(h)\\ & =-\int \mathrm{div}(hY)⟨\nabla F,Y⟩dm\\ & -\int h⟨\nabla F,{\nabla }_{Y}Y⟩dm.\end{array}$$

Here $h$ is a compactly supported Lipschitz weight. Both terms use only the first derivative of $F$; the derivatives of $Y$ come from the test-function calculus. In particular,

$$\begin{array}{rl}\mathrm{div}(h\nabla g)& =\Gamma (h,g)+h\Delta g,\\ {\nabla }_{Y}Y& =\nabla (\Gamma (g)/2).\end{array}$$

The companion's every-geodesic conclusion

On a complete, separable, full-support $\mathrm{RCD}(K,N)$ space, with $1<N<\infty$ and measure finite on bounded sets, let $F$ be bounded and globally Lipschitz, and let $G$ be bounded and continuous. Suppose $H_F(\nabla g,\nabla g)(h)\le \int hG\Gamma (g)dm$ for every compactly supported test function $g$ and every nonnegative compactly supported Lipschitz $h$. Then every constant-speed minimizing geodesic satisfies

$$(F\circ \sigma {)}^{\prime \prime }\le {\ell }^{2}G\circ \sigma$$

in distributions on $(0,1)$, including constant geodesics. (OpenAI, 2026b)

Concretely, pair the inequality with a nonnegative smooth function compactly supported in $(0,1)$ and integrate the left side twice by parts. No second derivative of $F\circ \sigma$ is assumed beforehand. If $G=c$ is constant, the conclusion says that $F(\sigma (t))-c{\ell }^2{t}^2/2$ is concave. Therefore

$$\begin{array}{rl}F(\sigma (t))\ge & (1-t)F(\sigma (0))\\ & +tF(\sigma (1))\\ & -\frac{c{\ell }^2}{2}t(1-t).\end{array}$$

This is an upper-Hessian bound: after subtracting the allowed quadratic bending, the function lies above its chords. The theorem permits collapsed spaces and does not assume compactness or a globally $L^2$ function $F$. A selected geodesic may lie inside an $m$-null set, so an almost-everywhere statement for transport paths does not establish this conclusion.

🏷️ Exponential changes of measure

The companion’s main device is to change the measure while keeping the distance fixed:

$$d{m}_{\lambda }=e^{\lambda F}dm,\lambda >0.$$

For each fixed $\lambda$, the old and new measures are uniformly equivalent because $F$ is bounded. Weak gradients agree. The generator changes according to

$${\Delta }_{\lambda }u=\Delta u+\lambda \Gamma (F,u).$$

In smooth geometry this changes the Ricci term in the Bochner formula from $\mathrm{Ric}$ to $\mathrm{Ric}-\lambda \mathrm{Hess}F$. The weak Hessian upper bound should therefore give the variable lower bound

$$k_{\lambda }(x)=K-\lambda G(x).$$

Making this heuristic rigorous requires care about domains. The paper proves equality of the two generator domains and equivalence of their graph norms. It does not identify their test classes. For example, a bounded Lipschitz potential agreeing with $-|x_1|$ near the origin gives a drift with a sign discontinuity. A smooth test function agreeing with $x_1$ there can then have a weighted Laplacian which is not $W^{1,2}$.

The necessary multiplier regularity comes from the original heat semigroup $P_t$. Its resolvent is

$$R_{a}h={\int }_0^{\infty }{e}^{-at}{P}_{t}hdt.$$

Heat estimates bound its gradient by a constant times $a^{-1/2}$ in the relevant $L^2$ and bounded-function norms. For a bounded Lipschitz potential $V$, the map

$$z⟼R_a(h+\Gamma (V,z))$$

is a contraction for sufficiently large $a$. Energy uniqueness identifies its fixed point with the weighted resolvent. Consequently, a bounded function in the weighted generator domain with bounded weighted Laplacian is Lipschitz. This argument uses the original heat flow before establishing any weighted curvature bound.

The manuscript next integrates the weighted Bochner expression against such multipliers. Its difference from the original expression is precisely the negative weak Hessian of $V$. A compact test core, graph-norm approximation, and the preceding multiplier estimate extend the inequality to the actual weighted generator domain. Singular Bochner measures are retained in this passage; approximation of their weights uses capacity and suitable representatives.

The resulting constant lower bound first gives a weighted $\mathrm{RCD}(\inf k_{\lambda },\infty )$ space. Imported variable-curvature results then provide entropy convexity with $k_{\lambda }(x)$. Finite dimension is not asserted to survive the measure change. It remains available for the geometry of the original space, whose metric has not changed. (OpenAI, 2026b)

🏷️ Concentrating entropy inequalities on a prescribed geodesic

For a probability measure $\mu =\rho m$, define its relative entropy by $E(\mu )=\int \rho \log \rho dm$. Changing the reference measure gives the exact identity

$${\mathrm{Ent}}_{m_{\lambda }}(\mu )=E(\mu )-\lambda \int Fd\mu .$$

This turns weighted entropy convexity into an inequality for averages of $F$. To concentrate those averages on a chosen geodesic, the paper uses metric nonbranching of finite-dimensional RCD spaces, imported from Deng’s theorem. A strict interior subsegment of any minimizing geodesic has a unique minimizing realization: an alternative could be spliced to the unchanged outside pieces, creating two minimizing paths which agree initially and then branch.

Fix such a subsegment, reparametrized as $\tau :[0,1]\to M$. Approximate each endpoint by normalized restrictions of $m$ to shrinking balls. Full support makes the normalizations positive; finiteness on bounded sets makes them finite. Their entropies are

$$\begin{array}{c}{E}_i^{(j)}=-\log m(B_{{\epsilon }_j}(\tau (i))),\\ i=0,1.\end{array}$$

Choose ${\lambda }_j$ growing faster than both endpoint entropy magnitudes. Apply the weighted transport inequality for this value of ${\lambda }_j$. All the transport geodesics lie in a common compact set $Q$ and have uniformly bounded lengths. Properness and the uniqueness of $\tau$ imply concentration of these plans on $\tau$, even though the weighted measures vary with $j$.

The Dirichlet Green kernel on the unit interval is

$$g(s,t)=\min (s,t)-st.$$

Write $A_t=\int Fd{\mu }_t$. Let $C_t$ be the transport-plan average of $\int g(s,t)|\dot{\gamma }{|}^{2}ds$, and let $D_t$ be the same average with the additional factor $G(\gamma (s))$. Rearranging weighted entropy convexity gives

$$\begin{array}{rl}{A}_t\ge & (1-t)A_0+t{A}_1-D_t\\ & +{\lambda }_j^{-1}(E_t+K{C}_t)\\ & -{\lambda }_j^{-1}((1-t)E_0+t{E}_1).\end{array}$$

Jensen’s inequality gives $E_t\ge -\log m(Q)$. Also $C_t$ is uniformly bounded by the squared length bound times $1/8$. Thus the combined entropy correction has nonnegative lower limit. Continuity of $F$ and $G$ permits passage to the concentrated plan. Writing $u=F\circ \tau$ and $v=G\circ \tau$, this yields

$$\begin{array}{rl}u(t)\ge & (1-t)u(0)+tu(1)\\ & -|\dot{\tau }{|}^2{\int }_0^{1}g(s,t)v(s)ds.\end{array}$$

Apply this on every strict interior subinterval. Subtract a twice continuously differentiable function whose second derivative is ${\ell }^{2}G(\sigma (t))$. The remainder satisfies the chord inequality for concavity, proving the distributional conclusion. The original geodesic need not be the unique path between its original endpoints. (OpenAI, 2026b)

🏷️ Recovering the tensor bound from triangle comparison

The forward direction of the characterization begins with the known RCD structure of Alexandrov spaces equipped with Hausdorff measure. Its new task is to extract distributional sectional curvature from small metric configurations. (OpenAI, 2026a)

Curvature occurs at fourth order in expansions of squared distances. The paper uses short Sobolev flows and optimal transport between their endpoint densities. First-order density expansions must be strong enough to pair with fixed $L^2$ second-derivative coefficients. Symmetric choices of gradient fields cancel the lower-order terms, leaving a curvature expression plus a nonnegative square error.

To indicate the localization mechanism, take independent fields $A,B$ from a finite sign-symmetric law of gradients. The defect involves terms of the form

$$\int f{\left|\nabla q_{AB}+{\nabla }_A(B-A)\right|}^{2}dm.$$

Finitely many local gradient choices approximate the negative connection term with arbitrarily small weighted squared error. A partition of the weight lets their inequalities be summed. Because the curvature functional is linear in its weight, the sum reconstructs its original evaluation, including the derivatives of that weight. No separate uncontrolled partition-derivative term remains. Taking the law supported on $U,-U,V,-V$ then isolates the sectional expression for $U,V$.

The order of approximation matters. The small-flow limit precedes the approximation of endpoint corrections; localization follows the extension to Lipschitz weights. Finally, strong local approximation of fields and an exhaustion of the space recover the original global test classes. The exhaustion errors vanish through the integrability of the tensor’s defining terms, rather than an assumption of compact support.

🏷️ Index forms and weak Hessian comparison

For the reverse direction, the smooth prototype is the index form along a geodesic $\gamma$ with velocity $v$:

$$I(Y)=\int \left(|D_{t}Y{|}^2-\kappa |Y\wedge v{|}^2\right)dt.$$

It measures the second-order cost of varying the geodesic, with curvature replaced by its lower bound. The manuscript establishes a distributional analogue by regularizing transport potentials with forward and backward heat factors. Their product obeys a continuity equation; their logarithmic potentials enter a Riccati identity for the Hessian.

Two limiting issues carry substantial work. Acceleration errors need a uniform one-sided bound up to the terminal time, including singular contributions from Bochner measures. Moreover, the index expression is quadratic in velocity, so weak convergence alone does not pass it to the limit. An exact action identity yields strong convergence of velocities weighted by the evolving densities. It also controls pairings with fixed $L^2$ tensor coefficients.

The resulting index inequality initially applies to global trials which are finite time-dependent combinations of test gradients. Classical Jacobi fields are not automatically members of this class. The paper constructs countably many harmonic frames, controls exceptional sets along entire paths, and solves finite-dimensional Sobolev equations for parallel transport. A single countable catalogue of global trials approximates each desired field along almost every individual path, with its endpoint constraints exact. Taking a countable infimum then avoids assuming a measurable simultaneous choice of optimal trials. These are substantive source arguments, not consequences of formally substituting weak derivatives into the smooth proof. (OpenAI, 2026a)

For $r=d(p,\cdot )$ and $q=r^2/2$, define the model function by

$$\begin{array}{c}{s}_{\kappa }^{\prime \prime }+\kappa s_{\kappa }=0,\\ s_{\kappa }(0)=0,s_{\kappa }^{\prime }(0)=1.\end{array}$$

Thus $s_0(r)=r$; for positive curvature it is $\sin (\sqrt{\kappa }r)/\sqrt{\kappa }$, and for negative curvature the corresponding hyperbolic sine. Decompose $Z$ into radial and perpendicular components. On a sufficiently small ball, the claimed Hessian comparison is

$$\begin{array}{rl}{H}_q(Z,Z)\le & |Z_{\parallel }{|}^{2}m\\ & +a_{\kappa }(r)|Z_{\perp }{|}^{2}m,\\ a_{\kappa }(r)& =\frac{r{s}_{\kappa }^{\prime }(r)}{s_{\kappa }(r)}.\end{array}$$

The coefficient has limit one at the center. The left side is a signed Radon measure whose singular part is nonpositive.

The coefficient can be checked directly from the Jacobi trial. Along a radial path of speed $r$, on $[\delta ,1]$ use a linear coefficient for the parallel part and

$$j(t)=\frac{s_{\kappa }((t-\delta )r)}{s_{\kappa }((1-\delta )r)}$$

for the transverse part. Since $j^{\prime \prime }+\kappa r^{2}j=0$,

$$\begin{array}{rl}& {\int }_{\delta }^1(j^{\prime 2}-\kappa r^2{j}^2)dt\\ & =[j{j}^{\prime }{]}_{\delta }^1\\ & =\frac{r{s}_{\kappa }^{\prime }((1-\delta )r)}{s_{\kappa }((1-\delta )r)}.\end{array}$$

The parallel contribution is $1/(1-\delta )$. Letting $\delta \downarrow 0$ gives the displayed coefficients. Positivity of the difference between an initial trial-energy bound and the Hessian functional produces the measure representation and its nonpositive singular part before the infimum is taken.

🏷️ Bounded distance modifications and triangle comparison

The squared distance is generally unbounded, whereas the companion theorem requires bounded $F$. A radial modification resolves this without losing the inequality. Put

$$\begin{array}{c}{m}_{\kappa }(r)={\int }_0^r{s}_{\kappa }(u)du,\\ c_{\kappa }(r)=1-\kappa m_{\kappa }(r).\end{array}$$

Here $c_{\kappa }=s_{\kappa }^{\prime }$. Choose radii $0<a<b$ inside the comparison ball and a smooth nonincreasing cutoff $\theta (q)$ which is one for $r\le a$ and zero for $r\ge b$. Define

$$\begin{array}{rl}{F}_p& =\chi (q),\chi (0)=0,\\ {\chi }^{\prime }(q)& =\theta (q)\frac{s_{\kappa }(r)}{r}.\end{array}$$

This function is bounded and globally Lipschitz, and agrees with $m_{\kappa }(r)$ on the inner ball. Its radial and transverse Hessian upper coefficients are respectively

$$\begin{array}{rl}A(r)& =\theta (q)c_{\kappa }(r)+r{\theta }^{\prime }(q)s_{\kappa }(r),\\ B(r)& =\theta (q)c_{\kappa }(r).\end{array}$$

Because ${\theta }^{\prime }\le 0$, one has $A\le B$. Since ${\chi }^{\prime }\ge 0$, the original nonpositive singular Hessian part keeps its sign under the chain rule. Consequently $H_{F_p}\le G_{p}m$ in the required quadratic sense, where $G_p=B$ is bounded and continuous, extended by zero outside the outer ball.

On every geodesic contained in the inner ball, the companion theorem now gives, with $y=m_{\kappa }(d(p,\sigma (t)))$,

$$y^{\prime \prime }+\kappa {\ell }^{2}y\le {\ell }^2.$$

The corresponding model-triangle function $\bar{y}$ has the same endpoint values and satisfies equality. Thus $h=y-\bar{y}$ has zero boundary values and satisfies $-h^{\prime \prime }-\kappa {\ell }^{2}h\ge 0$. Testing against its negative part gives

$${\int }_0^1|(h_{-}{)}^{\prime }{|}^{2}dt\le \kappa {\ell }^2{\int }_0^1{h}_{-}^{2}dt.$$

For $\kappa \le 0$, this forces $h_{-}=0$. For $\kappa >0$, the same follows from the Dirichlet Poincaré inequality when $\kappa {\ell }^2<{\pi }^2$. Therefore $y\ge \bar{y}$; monotonicity of $m_{\kappa }$ on the selected radius range gives the correct Alexandrov distance inequality.

Small-perimeter triangles keep their entire sides inside the required ball by the triangle inequality. The manuscript then uses local four-point comparison and Alexandrov globalization, rather than assuming small metric balls are geodesically convex. Completeness supplies the global conclusion. (OpenAI, 2026a)

🏷️ Reading connections and scope

The Cartan–Hadamard family develops comparison geometry for upper curvature bounds and fillings; the present family concerns lower curvature bounds and their analytic detection. The scalar-curvature and Ricci-flow family provides a contrasting use of weak geometric information, where scalar curvature controls smooth evolution only with additional structural arguments.

The source review checks the two theorem statements, the companion’s measure-change and concentration proof, and selected forward, heat-index, frame, Hessian-comparison, and globalization passages in the characterization manuscript. It does not independently verify all fourth-order transport expansions, heat estimates, exceptional-set estimates, or imported theorems. The post attributes the research conclusions to the manuscripts throughout.

🏷️ Key steps and proof outlines

The drift resolvent is constructed before weighted curvature is known

For bounded Lipschitz $V$, the measures $m$ and $e^{V}m$ have equivalent compression bounds, so they define the same weak gradients and Sobolev spaces. Testing the integration-by-parts formula with $e^V\eta$ and then with $e^{-V}\eta$ gives both inclusions of generator domains and the identity ${\Delta }_{V}u=\Delta u+\Gamma (V,u)$.

For the original heat resolvent $R_a=(a-\Delta {)}^{-1}$, heat gradient estimates give $\Vert \nabla R_{a}h\Vert \le C{a}^{-1/2}\Vert h\Vert$ in both the needed $L^2$ and bounded-function norms. Hence

$$z⟼R_a(h+\Gamma (V,z))$$

is a contraction in the space controlling the function and its gradient in those two norms, once $a$ is sufficiently large relative to $\Vert \nabla V{\Vert }_{\infty }$. Its fixed point solves the weighted resolvent equation. Weighted energy uniqueness identifies any bounded generator-domain solution with bounded Laplacian with this Lipschitz fixed point. This supplies the admissible multipliers for the later Bochner argument without assuming the weighted curvature conclusion in advance. (OpenAI, 2026b)

Entropy concentration converts an averaged statement into a prescribed-path statement

Endpoint balls shrinking onto a strict interior subsegment have finite entropy but that entropy may diverge. Choose the exponential weight parameter ${\lambda }_j$ after choosing the balls, large enough that both endpoint entropies divided by ${\lambda }_j$ tend to zero. The interpolating measures remain in one compact set, whose finite reference measure bounds their entropy below by Jensen.

Thus the entropy terms disappear in the favorable direction after division by ${\lambda }_j$. Uniqueness of the strict interior subsegment forces the transport plans to concentrate on that specific path. Continuity of the potential and Hessian bound permits passage to the Green-kernel chord inequality. Repeating on every interior subinterval gives the distributional second-derivative bound along the original geodesic, even when its original endpoints admit several minimizing paths.

The index-form passage requires strong velocity control

In the tensor characterization, heat-regularized transport gives a Riccati inequality and an index bound for finite combinations of test gradients. The curvature term is quadratic in velocity, so weak convergence of transport velocities would not suffice. The exact action identity supplies strong weighted convergence, allowing fixed $L^2$ tensor coefficients and the quadratic index terms to pass to the limit.

Countably many controlled frames and trials then approximate the desired Jacobi field while keeping endpoint constraints exact. The countable catalogue permits a pathwise infimum without an unproved measurable choice of optimizers. The model transverse trial has derivative energy equal to its endpoint term because it solves $j^{\prime \prime }+\kappa r^{2}j=0$. This produces $r{s}_{\kappa }^{\prime }(r)/s_{\kappa }(r)$ and the radial coefficient one. The measure inequality retains a nonpositive singular Hessian part, so the subsequent bounded distance modification can use the companion’s every-geodesic result and recover triangle comparison. (OpenAI, 2026a)

🏷️ Manuscripts and verification scope

• Gigli’s distributional curvature characterization of Alexandrov spaces — pinned manuscript and companion source.
• Weak Hessian bounds along every geodesic in RCD spaces — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. Despite its characterization heading, the documented selection is only the companion theorem converting a weak Hessian upper bound into an inequality along every minimizing geodesic. It does not select the full Alexandrov characterization. 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. Gigli’s distributional curvature characterization of Alexandrov spaces.

🐻  OpenAI 2026b. Weak Hessian bounds along every geodesic in RCD spaces.