From a Euclidean tangent to coordinates with controlled distortion

← All digests

Family 357 · Differential geometry
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-26

1. The question

If a space looks increasingly like $\mathbb R^n$ when magnified around a point, does a whole neighborhood admit Euclidean coordinates that preserve distances within a fixed factor?

A tangent space only records limiting geometry at one point. A coordinate chart must compare every pair of points in a neighborhood, including pairs much closer to one another than either is to the original center. Small errors across infinitely many scales may accumulate, so a Euclidean tangent does not automatically produce a quantitatively good chart.

Family 357 claims that such coordinates do exist at every regular point of a noncollapsed RCD space. These are metric spaces with a synthetic lower Ricci-curvature bound and their natural dimensional volume measure. The chart is defined on an open neighborhood, lands in an open subset of $\mathbb R^n$, and distorts distances by at most a constant depending only on $n$.

The neighborhood may be very small and depend on the point. The uniform part of the claim is the distortion factor, not the radius of the chart.

2. Background and essential definitions

Curvature without smooth coordinates

A metric measure space $(X,d,m)$ consists of a space, a distance and a reference measure. The condition $\mathrm{RCD}(K,n)$ is a synthetic analogue of a Ricci-curvature lower bound $K$ and dimension bound $n$. It combines a curvature-dimension condition expressed through optimal transport with a quadratic energy for weak gradients.

That quadratic structure makes the calculus Riemannian: weak gradients have inner products, the heat equation is linear, and weak Hessians and Laplacians can be defined. A smooth $n$-dimensional Riemannian manifold with Ricci curvature at least $K$, equipped with its volume measure, is a basic example. Singular limits of manifolds motivate the broader theory.

Here noncollapsed has the exact meaning

$$m=\mathcal H^n,$$

where $\mathcal H^n$ is $n$-dimensional Hausdorff measure. It is not an arbitrary weighted measure. The spaces are complete, separable and geodesic, with full-support measure finite on bounded sets. Full support means every nonempty open ball has positive measure.

Regular points and tangent spaces

A tangent space at $p$ is obtained by rescaling distances by $1/r_i$, for some $r_i\downarrow0$, and taking a pointed Gromov–Hausdorff limit. This convergence compares bounded metric balls while keeping track of their distinguished center; it does not require a common ambient Euclidean embedding.

The point $p$ is $n$-regular if every such tangent is pointed-isometric to $\mathbb R^n$ with its usual distance. Thus the infinitesimal limiting shape is unambiguously Euclidean.

A map $F:U\to V\subset\mathbb R^n$ is $L$-bi-Lipschitz when

$$L^{-1}d(x,z)\le |F(x)-F(z)|\le Ld(x,z) \qquad(x,z\in U).$$

The right inequality prevents excessive stretching, and the left prevents excessive compression. A bi-Lipschitz homeomorphism is a one-to-one correspondence with this control in both directions.

This is stronger than a bi-Hölder homeomorphism, where powers of distances may appear. For example, a bound involving $d(x,z)^\alpha$ with $\alpha<1$ allows the ratio to the original distance to grow without bound as the points approach each other.

Which distance is used inside a neighborhood?

The manuscript uses the original ambient distance $d$ restricted to $U$. It does not replace it by the length of the shortest path required to stay inside $U$. These distances can differ because an ambient shortest path may leave the chart domain. The all-pairs estimate is stated for the original distance.

3. Development history and previous partial results

Cheeger and Colding developed the structure theory of noncollapsed limits of manifolds with lower Ricci curvature. They obtained manifold neighborhoods with bi-Hölder parametrizations and raised the possibility of a bi-Lipschitz improvement. This is a historical antecedent of the synthetic regular-point problem.

The RCD theory subsequently extended weak differential geometry beyond spaces explicitly presented as smooth limits. De Philippis and Gigli established important noncollapsed volume-convergence and rigidity results, providing quantitative Euclidean approximation from the geometric hypotheses.

Mondino and Naber proved countable bi-Lipschitz rectifiability for RCD spaces. Roughly, almost all of the space can be covered by measurable pieces that are bi-Lipschitz equivalent to measurable subsets of Euclidean spaces. Such pieces need not be open, and an arbitrary prescribed regular point need not be covered by one open chart. This distinction is precisely relevant here.

Kapovitch and Mondino showed that the regular set of a noncollapsed RCD space lies in an open manifold region with bi-Hölder charts. That gave neighborhood topology, but left the stronger distance control unresolved.

Other results came close from different directions. Colding and Naber obtained bi-Lipschitz embeddings of sufficiently regular quantitative subsets into a Euclidean space that may have higher dimension. They also constructed examples showing that Euclidean tangents do not guarantee coordinates with a Hölder-continuous metric tensor. A bi-Lipschitz chart into the same dimension is therefore a specific intermediate target, not a route to unrestricted smoothness.

Honda and Zhang’s 2026 work recalls the regular-point conjecture and proves stronger geometric regularity under positive injectivity-radius assumptions. Their analytic criteria also use additional uniform control. Family 357 claims the open bi-Lipschitz chart under the noncollapsed RCD assumptions alone, without those extra conditions.

4. What the new manuscript claims

For each integer $n\ge2$, the theorem asserts a constant $L_n\ge1$. For any real $K$, any noncollapsed $\mathrm{RCD}(K,n)$ space $(X,d,\mathcal H^n)$, and any $n$-regular point $p$, there exist:

• an ambient-open neighborhood $U$ of $p$;
• an open set $V\subset\mathbb R^n$;
• a homeomorphism $F:U\to V$;

such that

$$L_n^{-1}d(x,z)\le |F(x)-F(z)|\le L_n d(x,z) \qquad\text{for every }x,z\in U.$$

The constant depends only on dimension. The domain and its size may depend on the space, the curvature bound and the chosen point. No quantitative radius valid for all examples is asserted.

The conclusion holds at every prescribed regular point, rather than only outside an unspecified measure-zero exceptional set. The image is open in $\mathbb R^n$, not just a measurable set and not a subset of a higher-dimensional target.

The paper also states an atlas consequence. There is an open set $\mathcal U\subset X$ containing every $n$-regular point and satisfying

$$\mathcal H^n(X\setminus\mathcal U)=0.$$

It has a countable atlas of $L_n$-bi-Lipschitz charts. On overlaps, transition maps have bi-Lipschitz constant at most $L_n^2$, by composing one chart with the inverse of another.

The open set $\mathcal U$ may include points outside the exact regular locus. The corollary does not assert that the set of points with Euclidean tangents is itself open. Being a point in a bi-Lipschitz manifold and having an exactly Euclidean metric tangent are different conditions.

5. Why the advance matters

Importance: 5/5, conditional on the claim. This would settle the noncollapsed regular-point bi-Lipschitz conjecture and upgrade an open topological description to quantitative metric coordinates.

For analysis, an open chart is more useful than a collection of measurable pieces: it describes an entire local region and compares all nearby pairs. For geometry, a dimension-only distortion constant separates the strength of the coordinate system from the scale at which it becomes available.

The result would also clarify how much regularity a lower Ricci bound provides. It claims genuine control of distances without claiming a smooth or Hölder-continuous metric tensor. This places the conclusion between weak measurable rectifiability and classical smooth differential geometry.

The proposed mechanism is of broader interest because it handles scale-dependent corrections that may accumulate without bound. The essential control is directional and summable, rather than a blanket assumption that every normalization remains uniformly mild.

6. The main idea

Start with harmonic coordinates that may degenerate

At a regular point, rescaling makes the geometry increasingly Euclidean. Volume comparison propagates quantitative flatness to sufficiently small balls around nearby centers. Harmonic functions—weak solutions of $\Delta u_j=0$—then approximate Euclidean coordinate functions.

Write the resulting open coordinate map as

$$u=(u_1,\ldots,u_n).$$

Its gradient inner products form the Gram matrix

$$G_{jk}=\langle\nabla u_j,\nabla u_k\rangle.$$

If this matrix were uniformly close to the identity at every relevant scale, metric control would be straightforward. The difficulty is that averages of the Gram matrix remain invertible but can become increasingly poorly conditioned at small scales. A harmonic coordinate direction can become compressed.

Simply renormalizing separately at each scale need not produce one limiting chart. The transformations can accumulate, fail to commute, and introduce a Hölder loss.

Separate cumulative stretching from directional error

The manuscript uses a matrix-valued Bochner measure. After rescaling so that the curvature lower bound is $-\kappa$, it takes the form

$$\mathsf M=\tfrac12\Delta G+\kappa G\,m,$$

with the Laplacian interpreted entry by entry as a measure. Positivity records control of second derivatives of the harmonic functions. Normalized local averages of this measure produce positive matrices governing how much stretching is needed at each dyadic radius $r_i=2^{-i}$.

There are two distinct estimates. One compares the accumulated positive matrices to the total normalization needed by the coordinates. This accumulated quantity may be unbounded.

The second evaluates the local matrices only on actual coordinate differences $u(x)-u(z)$. Schematically, the resulting error is bounded by

$$C r_i^2\bigl(E_i(x)+E_i(z)\bigr), \qquad \sup_x\sum_{i\ge0}E_i(x)\longrightarrow0$$

as the initial rescaling becomes flat. The functions $E_i$ are nonnegative defects; their sum, rather than the sum of their square roots, is controlled.

The paper obtains this directional estimate from Poisson replacements of squared distance, radial derivative approximations and a two-center projection estimate. A Poisson replacement solves a prescribed Laplace equation with boundary data chosen to imitate the distance function while providing better analytic control.

The distinction between the two estimates is the main geometric idea. A coordinate system may require large accumulated correction even though its errors on the pairs of points being compared have a small total budget.

Correct coordinates in their value space

Instead of differentiating a possibly irregular metric on $X$, the manuscript works in the open Euclidean domain of coordinate values $\Omega=u(B_1(p))$, in rescaled units. It constructs smooth maps $h_i:\Omega\to\mathbb R^n$ with invertible ordinary Jacobians.

The iteration is

$$h_0(P)=P,\qquad h_{i+1}(P)=h_i(P)+Dh_i(P)^{-\mathsf T}v_i(P).$$

Here $P$ is a coordinate value, $Dh_i$ is the derivative matrix, and $v_i$ is a covector field built from the positive local matrices and smooth cutoffs. The inverse transpose arranges that the leading change in squared lengths supplies the required positive stretching.

Differentiating the update creates further terms involving second and higher derivatives. The manuscript controls the diagonal second-derivative terms through a recurrence tied to the directional defects. Separate derivative estimates close the smooth iteration despite a loss of one derivative at each update.

This is designed to sum errors using $\sum_iE_i$, even when $\sum_i\sqrt{E_i}$ diverges. The elementary distinction is substantial: for example, squares of $1/i$ have a finite sum, while $1/i$ itself does not. Taking an unnecessary square root can destroy the entire estimate.

Obtain one chart, not separate maps at separate scales

The limiting map is

$$F=\lim_{i\to\infty}h_i\circ u.$$

The cumulative estimate supplies enough correction to prevent compression, and the directional estimate controls the total metric error. Together they yield both sides of the all-pairs bi-Lipschitz inequality.

Finally, the original open coordinate description and invariance of domain give openness of the image. Invariance of domain says that a continuous injective map between open subsets of $\mathbb R^n$ has open image. This topological step is needed because metric bounds on an arbitrary subset alone would not establish the claimed open-neighborhood chart.

7. Scope, limitations, and open questions

The theorem assumes integer dimension $n\ge2$, exact Hausdorff reference measure, the full RCD condition and Euclidean tangents at the chosen point. It is not a theorem for arbitrary metric spaces with Euclidean tangents, or for every point of every collapsed RCD space.

The chart is bi-Lipschitz, but its components are not asserted to have bounded Laplacians. Nor is the metric tensor in these coordinates asserted to be Hölder-continuous. Restrictions proved for analytically more regular coordinate maps do not automatically apply to the limiting maps claimed here.

There is no uniform chart radius, no statement that the exact regular set is open, and no elimination of singular points elsewhere in the space. The atlas region has full measure while its complement may still have important geometry.

Natural extensions concern collapsed settings, quantitative control of chart size and additional assumptions that improve coordinate regularity. The manuscript’s estimates and construction are explained as stated; they have not been independently verified by this digest.

8. Sources and further reading

The source’s Theorem 1.1 gives the open chart, and Corollary 1.2 gives the atlas. Its introduction distinguishes rectifiable pieces, bi-Hölder neighborhoods, higher-dimensional embeddings and analytically regular maps.

Primary background includes Mondino–Naber, Kapovitch–Mondino, Colding–Naber, and Honda–Zhang. These explain the different prior regularity conclusions and the extra hypotheses used for stronger structures.

Source manuscripts

• M1. OpenAI, Bi-Lipschitz Coordinates at Regular Points of Noncollapsed RCD Spaces (September 25, 2026).

Sources consulted on 26 October 2026 at repository revision 1581d4c. Download the manuscript citations.