Distance Mapping Distortion (DMD)

• Distance Mapping Distortion (DMD) is the discrepancy between distances in a source space and the mapped target space, quantified using differential metrics.
• It employs local Jacobians, singular value analysis, and quasi-isometry constants to measure and control mapping errors.
• DMD is applicable in diverse fields such as computational geometry, Ricci flow, imaging systems, and cartography to ensure accurate spatial reconstructions.

Distance Mapping Distortion (DMD) denotes the discrepancy between distances in a source space and the distances induced after mapping into a target space. The term is not standardized across the arXiv literature considered here: several papers do not use it explicitly, yet analyze the same underlying object under names such as quasi-isometry constant, distance distortion estimate, hyperspectral imaging distortion, tangent distance, or metrical distortion (Garanzha et al., 2022, Huang, 2018, Chen et al., 2022, Zhao et al., 2019, Voellinger, 6 Mar 2026). This suggests that DMD is best understood as a family of distortion notions indexed by the geometry of the source and target spaces, the admissible mapping class, and the metric or measure chosen for comparison.

1. Conceptual scope and problem settings

Across the cited literature, DMD appears in several distinct but structurally related settings. In computational deformation and parameterization, the basic object is a mapping

$$\vec x(\vec \xi): \Omega \subset \mathbb{R}^d \to \mathbb{R}^d,\quad d=2,3,$$

with Jacobian $J=\partial \vec x/\partial \vec \xi$, and distortion is encoded by singular values, determinants, and derived energies (Garanzha et al., 2022). In Ricci flow, the relevant map is the identity

$\mathrm{id}:(M,g(t))\to(M,g(s)),$

and DMD is the change of the Riemannian distance function between time slices (Huang, 2018). In map projection, the source is the Earth’s surface and the target is a planar map, so DMD is the difference between geodesic distances on the sphere and distances on the projection plane (Yang, 2021). In imaging systems and camera calibration, DMD is the discrepancy between physical positions or wavelengths and the corresponding pixel coordinates after passage through optical and sensor distortions (Chen et al., 2022, Pandat et al., 19 Dec 2025).

This breadth produces a common misconception: that DMD is a single universally fixed scalar. The papers instead support a metric-dependent view. Depending on context, one measures distortion by a quasi-isometry constant $\Gamma$, a local Jacobian-based density $f(J)$, a hyperbolic or quasihyperbolic displacement, a smile/keystone residual in pixels, a disparity-induced depth error, or a tangent-distance matrix (Garanzha et al., 2022, Vuorinen et al., 2012, Zhao et al., 2019).

2. Differential and metric descriptors of distortion

A recurring local description starts from the Jacobian. In practical lowest distortion mapping, Garanzha et al. define a distortion density $f(J)$ as a convex combination of a shape distortion term, based on $\operatorname{tr}(J^\top J)$ normalized by $(\det J)^{2/d}$, and a volumetric distortion term

$\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$

for $\det J>0$. For admissible orientation-preserving $J=\partial \vec x/\partial \vec \xi$0, one has $J=\partial \vec x/\partial \vec \xi$1, with $J=\partial \vec x/\partial \vec \xi$2 corresponding to a perfectly isometric map (Garanzha et al., 2022).

The same paper makes the connection to worst-case distance change explicit through the quasi-isometry constant $J=\partial \vec x/\partial \vec \xi$3, defined by

$J=\partial \vec x/\partial \vec \xi$4

where $J=\partial \vec x/\partial \vec \xi$5 are singular values. In this formulation, DMD is the worst relative stretch or shrink of lengths. A bounded-distortion threshold

$J=\partial \vec x/\partial \vec \xi$6

implies explicit upper bounds for $J=\partial \vec x/\partial \vec \xi$7, and in the $J=\partial \vec x/\partial \vec \xi$8-dimensional case with $J=\partial \vec x/\partial \vec \xi$9,

$\mathrm{id}:(M,g(t))\to(M,g(s)),$0

so $\mathrm{id}:(M,g(t))\to(M,g(s)),$1 as $\mathrm{id}:(M,g(t))\to(M,g(s)),$2 (Garanzha et al., 2022).

A related but higher-dimensional representation is given by the 3D quasiconformal framework. For a diffeomorphism $\mathrm{id}:(M,g(t))\to(M,g(s)),$3, the polar decomposition

$\mathrm{id}:(M,g(t))\to(M,g(s)),$4

yields principal dilations $\mathrm{id}:(M,g(t))\to(M,g(s)),$5 and principal directions $\mathrm{id}:(M,g(t))\to(M,g(s)),$6. The paper packages these into a 3D quasiconformal representation

$\mathrm{id}:(M,g(t))\to(M,g(s)),$7

and a direct scalar distortion measure is

$\mathrm{id}:(M,g(t))\to(M,g(s)),$8

A normalized anisotropy tensor is

$\mathrm{id}:(M,g(t))\to(M,g(s)),$9

with $\Gamma$0 (Chen et al., 2023). This provides a tensorial DMD descriptor rather than a single scalar.

In geometric function theory, DMD also appears as a lower bound problem. For open discrete ring $\Gamma$1-mappings at the origin with respect to the $\Gamma$2-modulus, $\Gamma$3, and

$\Gamma$4

one obtains

$\Gamma$5

Here the distortion function $\Gamma$6 controls modulus distortion, and the conclusion states that distances from the origin cannot collapse too fast (Salimov et al., 2017).

For quasiconformal self-maps with identity boundary values, the relevant metric descriptor is often the distance ratio metric

$\Gamma$7

which measures displacement relative to boundary proximity. The maximal dilatation $\Gamma$8 then serves as a global DMD parameter constraining $\Gamma$9, $f(J)$0, or $f(J)$1 (Vuorinen et al., 2012).

3. Geometric evolution and Riemannian distance distortion

In Ricci flow, DMD is the evolution of the Riemannian distance itself. Under a closed Ricci flow with bounded scalar curvature and sufficiently small initial global volume ratio $f(J)$2, the main estimate states that for scales $f(J)$3, $f(J)$4, and times

$f(J)$5

one has

$f(J)$6

for all pairs with $f(J)$7 (Huang, 2018). The identity map between time slices is therefore bi-Lipschitz on appropriate scales, with distortion factor $f(J)$8.

A central point of that paper is that the estimate is obtained without assuming a uniform lower bound on Perelman’s $f(J)$9-entropy. Instead, the analysis uses renormalized Sobolev inequalities, renormalized heat kernel bounds, and renormalized volume ratios based on the collapsing factor $f(J)$0 (Huang, 2018). This addresses a standard limitation of earlier non-collapsing theories associated in the details with Hamilton, Simon, Tian–Wang, Chen–Wang, and Bamler–Zhang.

A different Riemannian notion of DMD is developed under the name metrical distortion. The Levi-Civita metrical distortion

$f(J)$1

is constructed as an isometry from a synthetic Euclidean space to the manifold and is characterized by geodesically radial volume preservation, whereas the Riemannian exponential map is characterized by geodesically radial length preservation (Voellinger, 6 Mar 2026). Along geodesic rays, the radial contraction factor satisfies

$f(J)$2

and the paper identifies the associated reparametrization factor $f(J)$3 as a differential slip (Voellinger, 6 Mar 2026). This suggests a DMD interpretation in which distance distortion is the scalar gauge relating Euclidean radial length in the synthetic space to Riemannian radial length on the manifold.

4. Mesh deformation, bounded distortion, and reconstruction

In computational geometry, DMD is often a design objective rather than only an analytic observable. Garanzha et al. formulate a polyconvex energy for mean distortion,

$f(J)$4

and a stiffened bounded-distortion functional

$f(J)$5

whose infinite barrier enforces the pointwise condition

$f(J)$6

The discrete continuation problem then minimizes

$f(J)$7

while preserving orientation implicitly through the barrier (Garanzha et al., 2022). The resulting quasi-isometric stiffening (QIS) targets worst-element distortion rather than only average distortion.

The same paper emphasizes that no explicit inequality projection is used: orientation preservation and bounded distortion are enforced by the energy itself. Numerically, this produces “smallest known distortion estimates (quasi-isometry constants)” and stable quasi conformal parameterizations for very stiff problems (Garanzha et al., 2022).

The 3D quasiconformal solver provides a complementary perspective. There the representation $f(J)$8 is used to build the tensor

$f(J)$9

and the mapping components satisfy

$\operatorname{tr}(J^\top J)$0

On tetrahedral meshes, this yields sparse linear systems for reconstruction from a prescribed distortion field (Chen et al., 2023). Distortion control is then performed directly in the singular values by flipping negative eigenvalues for folded tetrahedra and clamping the ratio $\operatorname{tr}(J^\top J)$1 to a user-chosen threshold, which is a direct bounded-DMD regularization (Chen et al., 2023).

5. Optical, spectral, and photogrammetric mappings

In optical instrumentation, DMD has a terminological ambiguity: in “DMD-based Multi-Object Spectrograph,” DMD denotes a digital micromirror device, specifically a TI DLP7000 used as a programmable slit array (Chen et al., 2022). The relevant mapping distortion is hyperspectral imaging distortion in the chain

$\operatorname{tr}(J^\top J)$2

The paper treats smile and keystone as the principal distortions. Smile is modeled by parabolic fits

$\operatorname{tr}(J^\top J)$3

while wavelength is modeled as a cubic function of $\operatorname{tr}(J^\top J)$4, and keystone is analyzed through ellipse fits to the wavelength-dependent centroid locus of a fixed micromirror (Chen et al., 2022). A preliminary global model $\operatorname{tr}(J^\top J)$5 achieved a mean absolute error of $\operatorname{tr}(J^\top J)$6 pixels, which the paper uses as a quantitative measure of mapping discrepancy.

In long-range CCTV photogrammetry, DMD arises from imperfect lens models. The paper starts from a 14-parameter OpenCV-type distortion model and extends it to a 20-parameter model by adding higher-order radial and thin-prism terms. It then adds a neural residual correction model in a hybrid scheme

$\operatorname{tr}(J^\top J)$7

trained with a forward-and-inverse mean squared error loss (Pandat et al., 19 Dec 2025). The practical motivation is that depth from disparity obeys

$\operatorname{tr}(J^\top J)$8

so a small disparity error $\operatorname{tr}(J^\top J)$9 induces a large long-range depth error. The paper reports that at $(\det J)^{2/d}$0 km, a disparity error of $(\det J)^{2/d}$1 pixels yields about $(\det J)^{2/d}$2 m depth error, while $(\det J)^{2/d}$3 pixels gives about $(\det J)^{2/d}$4 m (Pandat et al., 19 Dec 2025).

A common misconception is that calibration quality should be judged only by checkerboard reprojection error. The CCTV study instead validates in GIS space and shows that standard 14-parameter distortion was reliable only to about $(\det J)^{2/d}$5 m, while the hybrid model moved $(\det J)^{2/d}$6 km points closer to ground truth on GIS maps and was capable of estimating 3D positions up to $(\det J)^{2/d}$7 kilometres (Pandat et al., 19 Dec 2025). In this setting, DMD is cumulative: lens distortion, correspondence error, triangulation error, and GIS transformation error all contribute.

6. Cartographic and manifold-learning viewpoints

Cartography provides one of the clearest explicit distance-distortion analyses. For the equidistant cylindrical projection,

$(\det J)^{2/d}$8

the local scale factors are

$(\det J)^{2/d}$9

North–south distances along meridians are preserved, and east–west distances are preserved only along the equator or, in the general Snyder form, along the chosen standard parallel (Yang, 2021). The local distance distortion function for small displacements is

$\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$0

As $\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$1, the horizontal distortion blows up because $\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$2 (Yang, 2021).

The same paper uses Tissot’s indicatrices to visualize local distortion and derives an angular distortion bound

$\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$3

with maximum about $\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$4 at $\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$5 (Yang, 2021). This directly corrects another misconception: “equidistant” does not mean globally distance-preserving.

In nonlinear dimensionality reduction, TDPM changes the metric being preserved. The tangent distance between $\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$6 and the tangent space at $\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$7 is defined by

$\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$8

where $\frac{1}{2}\left(\det J+\frac{1}{\det J}\right),$9 is estimated by local PCA from $\det J>0$0-nearest neighbors (Zhao et al., 2019). Classical MDS is then applied to the tangent-distance matrix so that Euclidean distances in the embedding approximate tangent distances rather than Euclidean or geodesic distances. The paper contrasts this with ISOMAP and LLE, which attempt to unfold the manifold, whereas TDPM tries to preserve the nonlinear folded structure (Zhao et al., 2019). This suggests that DMD in manifold learning is inseparable from the choice of source metric: preserving geodesic distance and preserving tangent distance are different objectives.

7. Recurrent themes, ambiguities, and representative regimes

Several themes recur across these otherwise disparate literatures. First, DMD is almost always controlled through local differential data: singular values and determinants in mesh maps, heat-kernel and volume-ratio estimates in Ricci flow, polynomial or neural residuals in camera models, scale factors in projections, or tangent spaces in manifold learning (Garanzha et al., 2022, Huang, 2018, Pandat et al., 19 Dec 2025, Yang, 2021). Second, the most useful global summaries are worst-case rather than average: quasi-isometry constants, worst-element condition numbers, maximal displacement in $\det J>0$1 or $\det J>0$2, or edge-case depth errors.

Representative numerical regimes illustrate how application-specific the notion is. In practical lowest distortion mapping, the wrench deformation example reports $\det J>0$3 and $\det J>0$4 after QIS, and the half-sphere flattening example gives a QIS quasi-isometry constant $\det J>0$5 against the analytical best $\det J>0$6 (Garanzha et al., 2022). In the spectrograph calibration study, preliminary residual mapping error is at the level of $\det J>0$7 pixels (Chen et al., 2022). In the CCTV study, the hybrid model reduces the maximum reprojection error on seven long-range test points from $\det J>0$8 pixels for the 14-parameter model to $\det J>0$9 pixels, while enabling usable localization up to about $J=\partial \vec x/\partial \vec \xi$00 km (Pandat et al., 19 Dec 2025).

The principal ambiguity is terminological. Some papers explicitly state that “Distance Mapping Distortion” is not their term, and one paper uses “DMD” for digital micromirror device rather than distance distortion (Garanzha et al., 2022, Chen et al., 2022). A plausible implication is that DMD is best treated as an editorial umbrella for a set of rigorously defined distortions rather than as a standardized invariant. Under that umbrella, the common core is stable: DMD is the controlled mismatch between source-space and target-space distances, quantified locally by differential data and globally by worst-case displacement, bi-Lipschitz bounds, or reconstruction error.