Flatness Distance: Concepts & Applications

• Flatness distance is a multifaceted measure that quantifies how closely an object or space resembles a flat model, with interpretations ranging from intrinsic flat distances in geometric measure theory to lattice-width in integer geometry.
• It is applied to assess convergence and stability in settings such as intrinsic flat convergence of manifolds, local approximations via beta and theta numbers, and concentration analysis in Fourier measures.
• Emerging applications leverage flatness distance in federated learning, optimization landscapes, and surface metrology to ensure consistent geometric behavior and precise measurement in both theoretical and practical scenarios.

Flatness distance is a polysemous term rather than a single invariant. Across the literatures represented here, it names several distinct distance-like constructions: the intrinsic flat distance between integral current spaces in geometric measure theory, normalized local distances from sets to affine planes or Euclidean model balls in Reifenberg-type analysis, lattice-width-based flatness constants in integer geometry, a model-dispersion quantity in federated learning, and the ISO minimum-zone separation between two parallel planes in surface metrology (Sormani et al., 2010, Violo, 2021, Celaya et al., 2022, Liu et al., 27 Feb 2026, Thang, 2011). The common theme is quantitative comparison with a flat model, but the underlying objects, admissible deformations, and geometric content vary sharply by field.

1. Intrinsic flat distance in geometric measure theory

The most established mathematical use of the term is the intrinsic flat distance $d_{\mathcal F}$ introduced by Sormani and Wenger for oriented compact Riemannian manifolds with boundary and, more generally, for integral current spaces. It is defined by embedding two spaces isometrically into a common complete metric space $Z$, measuring the classical Federer–Fleming flat distance between the pushed-forward currents there, and taking the infimum over all choices of $Z$ and embeddings. For $M_i=(X_i,d_i,T_i)\in\mathcal M^m$,

$$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$

equivalently,

$$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$

where

$\varphi_\# T_M-\psi_\# T_N = U+\partial V.$

Here $U$ is an $m$-dimensional residual current and $V$ is an $Z$0-dimensional filling current (Sormani et al., 2010).

This distance is an intrinsic analogue of the classical flat distance in much the same way that the Gromov–Hausdorff distance is an intrinsic analogue of the Hausdorff distance. The difference from Gromov–Hausdorff theory is structural: $Z$1 compares spaces as currents, so orientation matters, multiplicity matters, and cancellation can occur. For compact oriented Riemannian manifolds,

$Z$2

For precompact integral current spaces, zero distance is equivalent to the existence of a current preserving isometry. The distance is finite on integral current spaces of finite mass, satisfies the triangle inequality, and obeys the estimate

$Z$3

Intrinsic flat limits are again integral current spaces, hence countably $Z$4-rectifiable, and mass is lower semicontinuous: $Z$5 A sequence may converge in Gromov–Hausdorff sense to a nontrivial space while converging intrinsically flat to the zero current space, because thin regions can disappear or oppositely oriented sheets can cancel (Sormani et al., 2010).

The later paper "Properties of the Intrinsic Flat Distance" develops geometric tools for controlling $Z$6, notably sliced filling volume and interval sliced filling volume, and proves compactness theorems showing that lower sliced-filling bounds or the tetrahedral property prevent collapse and force agreement of Gromov–Hausdorff and intrinsic flat limits (Portegies et al., 2012). This suggests that, within geometric measure theory, flatness distance is best understood as a filling-based metric on oriented rectifiable geometric content rather than on the underlying metric skeleton alone.

2. Pointed and local variants

For noncompact spaces with only local mass bounds, the intrinsic flat framework is localized. The pointed intrinsic flat distance $Z$7 is defined on pointed locally integral current spaces $Z$8 by requiring small flat error on large bounded balls around the base points after isometric embedding into a common complete metric space. Its definition uses an infimum of $Z$9 such that the embedded base points are $Z$0-close and

$Z$1

with

$Z$2

for $Z$3 (Takeuchi, 2018).

This pointed construction metrizes the Lang–Wenger notion of pointed convergence for locally integral current spaces: convergence in $Z$4 is equivalent to the existence of isometric embeddings into a single complete metric space in which base points converge and the pushed-forward currents converge in the local flat topology. Zero pointed intrinsic flat distance identifies spaces up to isometry on $Z$5 that preserves both the base point and the current (Takeuchi, 2018).

The local theory clarifies an issue already present in the compact setting: under intrinsic flat convergence, points can disappear from the settled completion of the limit. The sliced filling estimates of Portegies and Sormani give explicit criteria ensuring that Cauchy sequences of points do not disappear, by imposing lower bounds on averaged sliced filling quantities around those points (Portegies et al., 2012). In this local noncompact regime, flatness distance becomes a pointed, scale-dependent control on current-theoretic agreement near chosen origins.

3. Distance to planes, hyperplanes, and Euclidean model balls

A second large family of meanings quantifies flatness by comparing a set to an affine plane or to a Euclidean ball at a point and scale. In "Flatness, Menger curvature, and parametrization", the central quantities are the unilateral beta numbers

$Z$6

and the bilateral theta numbers

$Z$7

The first measures how close $Z$8 is to lying in a plane; the second measures Hausdorff closeness in both directions. Always $Z$9, but the reverse inequality can fail badly. On closed linearly locally contractible $M_i=(X_i,d_i,T_i)\in\mathcal M^m$0-manifolds, however, small-scale unilateral flatness controls bilateral flatness: $M_i=(X_i,d_i,T_i)\in\mathcal M^m$1 under the small-$M_i=(X_i,d_i,T_i)\in\mathcal M^m$2 hypotheses of Theorem 3.2 (David et al., 12 Jun 2026).

For zero sets of harmonic polynomials, flatness is measured by a normalized local Hausdorff distance to hyperplanes,

$M_i=(X_i,d_i,T_i)\in\mathcal M^m$3

A point is flat exactly when $M_i=(X_i,d_i,T_i)\in\mathcal M^m$4. The central dichotomy states that for degree-$M_i=(X_i,d_i,T_i)\in\mathcal M^m$5 harmonic polynomials there exists $M_i=(X_i,d_i,T_i)\in\mathcal M^m$6 such that

$M_i=(X_i,d_i,T_i)\in\mathcal M^m$7

while if $M_i=(X_i,d_i,T_i)\in\mathcal M^m$8 at one scale, then

$M_i=(X_i,d_i,T_i)\in\mathcal M^m$9

Thus flat points coincide with regular points $$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$0 (Badger, 2011).

A related intrinsic–extrinsic comparison appears in "A remark on two notions of flatness for sets in the Euclidean space". There the extrinsic Reifenberg numbers

$$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$1

are compared with the intrinsic Gromov–Hausdorff numbers

$$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$2

The main result is that, in the Reifenberg regime, $$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$3 behaves like the square of $$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$4; in particular,

$$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$5

while conversely

$$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$6

This makes rigorous the David–Toro phenomenon that intrinsic flatness can be second-order compared with extrinsic Hausdorff flatness (Violo, 2021).

Taken together, these constructions show that one major meaning of flatness distance is scale-normalized distance to a flat local model. The model may be an affine plane, a hyperplane, or a Euclidean $$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$7-ball, and the chosen distance may be directed, Hausdorff, or Gromov–Hausdorff.

4. Quantitative flatness as concentration and obstruction

In Fourier analysis, the paper "Quantitative flatness and obstructions in Fourier analysis" uses neither planes nor fillings as the primary object. Instead, it defines flatness through concentration of mass in shrinking cuboids. A compactly supported finite Borel measure $$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$8 on $$d_{\mathcal F}(M_1,M_2)=\inf_{Z,\varphi_1,\varphi_2} d_F^Z(\varphi_{1\#}T_1,\varphi_{2\#}T_2),$$9 is $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$0-flat if there exists a sequence of cuboids $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$1 such that

$$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$2

It is $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$3-flat if, in addition,

$$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$4

where $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$5 is the shortest side length of $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$6 (Fraser, 11 Jun 2026).

These are not distances in the metric sense, but they are quantitative surrogates for distance-to-flatness. They encode how much mass can be packed into anisotropic boxes and serve as Knapp-type obstructions to restriction estimates, $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$7-improving estimates, Fourier decay, and $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$8-flattening. For instance, if $$d_{\mathcal F}(M,N):=\inf \{\mathbf M(U)+\mathbf M(V)\},$$9 is $\varphi_\# T_M-\psi_\# T_N = U+\partial V.$0-flat and an extension estimate $\varphi_\# T_M-\psi_\# T_N = U+\partial V.$1 holds, then

$\varphi_\# T_M-\psi_\# T_N = U+\partial V.$2

Flatness is thus detected by an exponent governing concentration rather than by a direct comparison with a plane (Fraser, 11 Jun 2026).

This suggests a broader interpretation: in analysis, flatness distance may denote any quantitative mechanism that records how strongly an object resembles a low-complexity or anisotropic model at small scales, even when no actual metric on models is introduced.

5. Combinatorial and arithmetic flatness

In integer optimization and geometry of numbers, flatness is measured by lattice width rather than by Hausdorff or Gromov–Hausdorff distance. For a polyhedron

$\varphi_\# T_M-\psi_\# T_N = U+\partial V.$3

the width in direction $\varphi_\# T_M-\psi_\# T_N = U+\partial V.$4 is

$\varphi_\# T_M-\psi_\# T_N = U+\partial V.$5

and the lattice width is

$\varphi_\# T_M-\psi_\# T_N = U+\partial V.$6

For lattice-free polyhedra, small width means containment in few integer hyperplane layers. The paper "Proximity and flatness bounds for linear integer optimization" proves that if $\varphi_\# T_M-\psi_\# T_N = U+\partial V.$7 is full-dimensional, lattice-free, and every row of $\varphi_\# T_M-\psi_\# T_N = U+\partial V.$8 is facet-defining, then some row $\varphi_\# T_M-\psi_\# T_N = U+\partial V.$9 of $U$0 satisfies

$U$1

This is derived from a refined proximity bound and interpreted as a flatness bound linear in $U$2 and in the largest absolute $U$3 minor $U$4 (Celaya et al., 2022).

A related generalized flatness constant is

$U$5

the largest lattice width of a convex body in $U$6 with at most $U$7 interior lattice points. The planar paper "Exact Flatness Constant for One-Point Convex Bodies and the Discrete Isominwidth Problem: The Planar Case" proves

$U$8

with equality iff $U$9 is unimodularly equivalent to $m$0. This extends the classical flatness problem from hollow bodies to few-point bodies and feeds into the discrete isominwidth inequality

$m$1

for planar convex bodies with interior lattice points (Averkov et al., 29 Apr 2026).

In graph theory, flatness is again distance-based but combinatorial. A class of graphs is flip-flat if for every radius $m$2 there exist $m$3 and $m$4 such that every sufficiently large vertex set contains a large subset $m$5 that is distance-$m$6 independent in a graph $m$7 obtained by at most $m$8 flips, where a flip toggles adjacency between two chosen vertex subsets. The main theorem is

$m$9

Here flatness means the ability, after bounded global edge complementation, to extract vertices that are pairwise far apart in graph distance (Dreier et al., 2022).

6. Optimization, learning, and local geometric flatness

Recent machine-learning work uses the term more explicitly. In heterogeneous federated learning, the paper "FedNSAM: Consistency of Local and Global Flatness for Federated Learning" defines the flatness distance

$V$0

where $V$1 are the final client models after $V$2 local steps and

$V$3

is the aggregated model. The quantity is not a curvature measurement; it is the mean squared distance between each client’s terminal local model and the global average. The paper interprets it as a proxy for whether client-local flat regions are geometrically aligned, and proves upper bounds on $V$4 for FedSAM and FedNSAM in terms of stochastic variance, data heterogeneity, and optimization parameters (Liu et al., 27 Feb 2026).

A different optimization-theoretic meaning appears in "On the geometry of flat minima". There flatness is defined by the maximal variation

$V$5

and the dual distance to neighboring level sets

$V$6

These satisfy the duality

$V$7

and flatness can be ordered either by smaller $V$8 for small $V$9 or, equivalently, by larger $Z$00 for small $Z$01. In this usage, flatness distance is literally the distance to neighboring level sets, and higher-order derivatives enter through asymptotics such as

$Z$02

when lower derivatives vanish (Josz, 14 Sep 2025).

A related but distinct local metric statement occurs in contact geometry: the Legendrian spectral distance is proved to be $Z$03-locally flat, meaning that in Weinstein coordinates

$Z$04

for sufficiently $Z$05-close Legendrians. Here flatness means exact local norm behavior of a distance rather than distance to a flat model (Allais et al., 2024).

7. Surface metrology and engineering usage

In surface metrology, flatness distance has a direct ISO meaning: the minimum separation between two parallel enveloping planes containing the measured surface. The report "A novel simple and accurate flatness measurement method" states that flatness of a granite plate is the distance between two tolerance planes surrounding the surface such that the two planes form the minimum zone. In angle-based measurements, the paper writes

$Z$06

with

$Z$07

It then decomposes the apparent angular flatness as

$Z$08

where $Z$09 is the true flatness contribution and the second term is contamination from overall plate inclination. The proposed measurement patterns use straight, parallel lines so that the inclination term cancels (1112.0211).

The companion uncertainty paper "Measurement uncertainty in surface flatness measurement" formulates the same ISO quantity as

$Z$10

the perpendicular separation between the two minimum-zone parallel planes. It models the measured inclination by a sum of instrument, Earth-curvature, repeatability, thermal-gradient, support-system, closure, and humidity contributions, converts angle to height via

$Z$11

and propagates uncertainty to the final flatness value as

$Z$12

In this engineering usage, flatness distance is not a metric on abstract objects; it is a physically realized normal separation between two parallel planes (Thang, 2011).

In these metrological papers, the phrase is closest to its everyday geometric meaning. The quantity is a minimum-zone plane-to-plane distance, and the main technical issue is not defining flatness but measuring it without tilt contamination and with a defensible uncertainty budget.