Rough Riemannian Metrics

• Rough Riemannian metrics are measurable, low-regularity analogues of smooth metrics that relax differentiability while maintaining fundamental analytic structures.
• They enable robust Lᵖ and Sobolev theories as well as variational formulations of the Laplacian despite the absence of classical tools like Levi–Civita connections.
• Different frameworks—including bounded and space-of-metrics approaches—offer varied geometric controls, supporting applications in spectral theory, approximation, and analysis of PDEs.

Rough Riemannian metrics are low-regularity analogues of classical Riemannian metrics on a smooth manifold, in which the differentiable structure remains smooth but the metric coefficients are only measurable, typically with local boundedness and ellipticity assumptions rather than continuity or differentiability. In this setting one still retains substantial analytic structure—Radon measures, $L^p$ and Sobolev theories, variational Laplacians, divergence-form operators, and, in several frameworks, intrinsic distances and metric-space structures—while classical pointwise constructions such as Levi–Civita connections, curvature tensors, or geodesic equations may cease to be available in any direct sense (Bandara et al., 2018, Bandara, 2014, Bandara et al., 14 Jul 2025).

1. Definitions and competing formulations

The basic measurable-coefficient definition used in spectral and PDE work starts from a smooth manifold $M$ and a symmetric tensor field $g$ that is only measurable in charts, subject to a local comparability condition. In the form used for compact manifolds with smooth boundary, $g$ is a rough metric if for every $x\in M$ there exists a chart $(U_x,\psi_x)$ and a constant $C(U_x)\ge 1$ such that for almost every $y\in U_x$ and all $u\in T_yM$,

$$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$

On compact $M$0, this local condition is equivalent to global comparison with a fixed smooth metric $M$1,

$M$2

almost everywhere. In the later formulation on connected manifolds without boundary, rough metrics are equivalently characterized by

$M$3

again in the measurable sense (Bandara et al., 2018, Bandara et al., 14 Jul 2025).

A closely related but stricter notion is the bounded rough Riemannian metric. There one assumes that for each $M$4, $M$5 is a positive definite symmetric bilinear form on $M$6, and that in coordinates the components of $M$7 are bounded above, uniformly bounded away from zero, and measurable. This is explicitly stronger than the Bandara-style rough metric formalism, because it rules out the exceptional singular set of measure zero and is designed to support a length-space theory based on piecewise smooth curves (Allen et al., 5 Mar 2026).

The 2014 functional-analytic framework allows degeneration on sets of measure zero. It defines the singular set

$M$8

with regular set $M$9. Outside $g$0 the metric is finite and positive definite, while $g$1 has null measure for the induced volume measure (Bandara, 2014).

Framework	Regularity hypothesis	Primary output
Rough metric	measurable, locally bounded, locally elliptic a.e.	measure, $g$2/Sobolev theory, variational operators
Bounded rough metric	positive definite at every point; bounded, uniformly nondegenerate, measurable	piecewise-smooth length and induced distance
Space-of-metrics viewpoint	all rough metrics with finite or infinite mutual comparability distance	complete length extended metric space of metrics

A persistent terminological caution is necessary. In the measurable-coefficient literature, “rough metric” refers to low-regularity tensor fields. In a distinct large-scale sense, “rough similarity” refers to coarse comparison of left-invariant Riemannian metrics on Lie groups; that notion concerns quasi-isometric geometry rather than measurable-coefficient Riemannian analysis (Donne et al., 2022).

2. Measure, Sobolev theory, and variational operators

Even when $g$3 is only measurable, the density

$g$4

is well defined in charts and transforms correctly under coordinate changes, yielding a Radon measure $g$5. The resulting $g$6-theory of tensors is robust: under uniform comparability of two rough metrics, the corresponding $g$7 spaces and Sobolev spaces coincide as sets, with explicit norm equivalences. In particular, if $g$8 and $g$9 are $g$0-close, then for tensors of type $g$1 and $g$2,

$g$3

and similarly for $g$4, $g$5, and $g$6 spaces (Bandara et al., 2018, Bandara, 2014, Bandara et al., 14 Jul 2025).

Because the Levi–Civita formula differentiates the metric coefficients, it is unavailable in the merely measurable regime. The standard replacement is variational. On a compact manifold with smooth boundary, one begins with

$g$7

shows that $g$8 is closable, and defines

$g$9

For an admissible boundary condition $x\in M$0 with

$x\in M$1

the energy form

$x\in M$2

is closed, symmetric, and non-negative. Standard form theory then yields a unique non-negative self-adjoint operator $x\in M$3, which plays the role of the Laplace–Beltrami operator. On compact manifolds, Rellich–Kondrachov gives compact resolvent, hence discrete spectrum (Bandara et al., 2018).

The same analytic architecture supports divergence-form operators and Hodge-theoretic constructions. Under a uniform change of metric, divergence operators transform by

$x\in M$4

with $x\in M$5 the pointwise change-of-metric endomorphism and $x\in M$6. On compact manifolds, the Hodge–Dirac operator associated with a rough metric is self-adjoint, and its kernel in degree $x\in M$7 is identified with $x\in M$8 (Bandara, 2014, Bandara et al., 14 Jul 2025).

3. Spectral theory and Weyl asymptotics

A central theorem for rough metrics is the Weyl law for weighted Laplace equations on compact manifolds with smooth boundary. For an admissible boundary condition $x\in M$9 and a real-valued weight $(U_x,\psi_x)$0 with $(U_x,\psi_x)$1 and $(U_x,\psi_x)$2, one considers the generalized eigenvalue problem

$(U_x,\psi_x)$3

on the closed subspace $(U_x,\psi_x)$4, where

$(U_x,\psi_x)$5

The nonzero eigenvalues split into positive and negative branches because $(U_x,\psi_x)$6 is not assumed to have fixed sign. Writing

$(U_x,\psi_x)$7

the eigenvalues $(U_x,\psi_x)$8 satisfy

$(U_x,\psi_x)$9

The same framework yields the classical Weyl law for the unweighted rough Laplacian: $C(U_x)\ge 1$0 Thus the leading Weyl coefficient is unchanged by passage from smooth to rough metrics; it depends only on dimension and the rough volume measure (Bandara et al., 2018).

The proof does not rely on a compatible geodesic distance or on heat-kernel asymptotics. Instead it uses a purely functional-analytic and operator-theoretic strategy adapted from Birman–Solomyak: localization to finitely many coordinate patches mapped to Lipschitz Euclidean domains, reduction to weighted elliptic model problems, and a Dirichlet–Neumann bracketing argument. In this sense the principal Weyl term survives even when classical metric geometry is unavailable (Bandara et al., 2018).

This result also clarifies a common misconception. The absence of smoothness does not force a loss of high-frequency spectral asymptotics. What is lost first is the classical pointwise differential-geometric apparatus; the leading spectral term remains stable under $C(U_x)\ge 1$1-rough perturbations and sign-changing weights (Bandara et al., 2018).

4. Intrinsic distances and the geometry of the space of metrics

The status of distance depends strongly on which rough-metric formalism is used. In the compact-with-boundary variational framework, the authors explicitly emphasize that one cannot in general build a good metric-space structure from a rough metric: the curve-length expression may be ill defined, so there is no robust canonical distance, and hence no useful geodesic theory or curvature in the classical sense. This is a statement about that low-regularity measurable framework, not about all later definitions of rough metric (Bandara et al., 2018).

Later work supplies two distinct remedies. For bounded rough metrics one defines, for any piecewise smooth curve $C(U_x)\ge 1$2,

$C(U_x)\ge 1$3

and then

$C(U_x)\ge 1$4

Uniform bounds against a smooth background metric immediately give bi-Lipschitz comparison of distances. In a different direction, the intrinsic metric associated with a general rough metric on a smooth connected manifold is defined by

$C(U_x)\ge 1$5

and this metric is finite and intrinsic; for smooth $C(U_x)\ge 1$6, $C(U_x)\ge 1$7 agrees with the classical Riemannian distance (Allen et al., 5 Mar 2026, Bandara et al., 14 Jul 2025).

The metric geometry of the space of rough metrics itself is organized by the extended distance

$C(U_x)\ge 1$8

Here $C(U_x)\ge 1$9 means

$y\in U_x$0

for all tangent vectors and almost every $y\in U_x$1. The resulting space $y\in U_x$2 is a complete length extended metric space; the finite-distance components

$y\in U_x$3

are path-connected; and $y\in U_x$4 is connected if and only if $y\in U_x$5 is compact (Bandara et al., 14 Jul 2025).

Within a component, geometry on $y\in U_x$6 is quantitatively stable. If $y\in U_x$7, then

$y\in U_x$8

and the corresponding $y\in U_x$9 and Sobolev norms are equivalent with explicit exponents. This provides an $u\in T_yM$0-type moduli theory for rough geometry (Bandara et al., 14 Jul 2025).

Another misconception is that a rough metric should be determined by the induced metric-measure structure $u\in T_yM$1. This fails in general: there exist distinct rough metrics on $u\in T_yM$2 with the same intrinsic distance and the same volume measure, so the space of rough metrics is strictly richer than the space of induced metric-measure structures (Bandara et al., 14 Jul 2025).

5. Approximation, closures, and probabilistic models

The closure of smooth metrics depends on which topology one places on the ambient space of metrics. For the extended metric $u\in T_yM$3, the closure of $u\in T_yM$4 is exactly $u\in T_yM$5: $u\in T_yM$6 Thus continuous metrics are precisely the $u\in T_yM$7-limits of smooth ones, whereas more singular measurable rough metrics generally are not (Bandara et al., 14 Jul 2025).

A different completion arises from the $u\in T_yM$8 Riemannian metric on the Fréchet manifold $u\in T_yM$9 of smooth metrics. The induced path metric $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$0 has a completion identified with equivalence classes of measurable finite-volume positive-semidefinite semimetrics modulo degeneracies. The topology induced by $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$1 is equivalent to an $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$2-type topology defined by an integrated pointwise metric $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$3, and convergence to a smooth metric $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$4 is characterized by convergence in measure of the tensor coefficients together with $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$5-convergence of the volume densities. This places rough or degenerate metrics naturally at the boundary of the smooth $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$6-geometry of $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$7 (Clarke, 2010).

Within the smooth category, spectral approximation theory offers another perspective. The “Riemannian Bergman metrics” of degree $$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$8 are pullbacks of Euclidean metrics under spectral embeddings

$$C(U_x)^{-1} |u|_{\psi_x^*\delta(y)} \le |u|_{g(y)} \le C(U_x) |u|_{\psi_x^*\delta(y)}.$$9

built from eigenspaces of a fixed Laplacian. The union of the Bergman spaces is dense in $M$00 in the $M$01-topology, and the natural approximation maps $M$02 recover a given smooth metric asymptotically. The paper proves this only in the smooth setting, but the construction isolates a finite-dimensional spectral parametrization that suggests one route toward approximation schemes for lower-regularity metrics (Potash, 2013).

A probabilistic counterpart appears in Gaussian models on the space of metrics with fixed volume form. There, random metrics are generated by exponentiating Gaussian fields in the diagonal directions of the symmetric-space fibers. The $M$03 Ebin distance to a reference metric has an explicit characteristic function, and a Lipschitz-type distance

$M$04

controls diameter, Laplace eigenvalues, and volume entropy. Since the construction depends on spectral decay coefficients, it suggests a route to random rough metrics by lowering regularity while preserving positivity via exponentiation, although the stated theorems are proved for smooth metrics (Clarke et al., 2013).

6. Analytic applications, stability, and sharp metric bounds

One of the strongest applications of rough metrics is to the Kato square root problem. The central stability principle is that quadratic estimates in the Axelsson–Keith–McIntosh first-order framework persist across a change of inner product induced by a uniformly close metric. As a consequence, if a rough metric is uniformly close to a continuous or smooth reference metric for which Kato estimates are known, then the Kato square root problem for functions and forms transfers to the rough metric. On compact manifolds this yields solutions of the Kato problem for continuous metrics, and the paper further shows that a lower bound on injectivity radius is not necessary for such results (Bandara, 2014).

The same componentwise stability under $M$05 has broader consequences. Within $M$06, generalised local Poincaré inequalities are preserved, heat-kernel Hölder regularity transfers from one metric to another, and divergence-form operators $M$07 can be rewritten as weighted Laplacians for another rough metric in the same component. This framework also supports compact and noncompact Hodge theory for rough metrics, with cohomological invariance of the Hodge–Dirac kernel (Bandara et al., 14 Jul 2025).

For bounded rough metrics, recent work isolates sharp conditions under which comparison of tensor fields yields Lipschitz or uniform comparison of the induced distances. A lower Lipschitz bound follows if $M$08 on a set $M$09 whose complement has $M$10-measure zero: $M$11 A lower uniform bound holds when the complement has small but nonzero $M$12-measure, giving

$M$13

Upper bounds are easier: if $M$14 on an open set of full $M$15-volume, then

$M$16

and more refined hypotheses on the decomposition of the bad set into components of small total diameter yield uniform additive error bounds. The paper supplies conformal examples on $M$17 showing that these assumptions cannot be weakened further (Allen et al., 5 Mar 2026).

Taken together, these developments show that “rough Riemannian metric” is not a single theorem or technique but a family of closely related low-regularity frameworks. In all of them, the smooth manifold structure is retained while metric regularity is relaxed; what varies is the geometric output one demands. For variational spectral theory, measurable ellipticity suffices. For intrinsic length geometry, stronger boundedness or dual-Lipschitz formulations are required. For moduli questions, the extended metric $M$18 yields a complete length space of rough metrics whose smooth closure is exactly the continuous stratum.