Geometric Reciprocity Theorem (GRT)

Updated 7 July 2026

GRT is a theoretical framework that equates one-sided local geometric counts with reciprocal global cancellations, serving as a foundation for analytic mask computation in DIBR.
Under nearest-neighbor warping, GRT shows that the disocclusion mask equals the lost-pixel mask, ensuring train-inference consistency in self-supervised stereo view synthesis.
Extensions of GRT span lattice-point geometry, algebraic K-theory, and scattering theory, revealing its broad applicability in encoding reciprocity through geometric and algebraic structures.

Geometric Reciprocity Theorem (GRT) denotes a class of reciprocity statements in which a geometric construction makes two ostensibly different descriptions equivalent, typically by converting a one-sided or local count, warp, or symbol into a reciprocal count, inverse warp, or global cancellation law. In the supplied literature, the term is used explicitly for a theorem in monocular-to-stereo conversion: under nearest-neighbor Depth-Image-Based Rendering (DIBR), the disocclusion mask for synthesizing a target stereo view equals the mask of pixels lost when warping back from target to source, making self-supervised learning from monocular videos possible (Lu et al., 6 Jul 2026). Closely related geometric reciprocity theorems also appear in lattice-point geometry, algebraic $K$ -theory, projective graphic statics, and scattering theory, where reciprocity is encoded by region decompositions, flag refinements, projective dualities, or transfer-operator identities (Rehberg, 2021, Musicantov et al., 2014, Baranyai, 2019, Loran et al., 23 Aug 2025).

1. DIBR reciprocity and the contemporary GRT formulation

In the stereoscopic video setting, the theorem is formulated in the standard DIBR stereo setup, with a monocular frame treated as the source view and one stereo eye synthesized as the target view. In the main exposition, the left view is the source $I_L$ and the right view is the target $I_R$ . Under nearest-neighbor DIBR, rectified stereo geometry, and the standard “one pixel maps to one discrete target coordinate” warping rule, the theorem states

$M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$

where $M_{\text{dis}^{L\to R}}$ is the disocclusion mask required when synthesizing the right view from the left view, and $M_{\text{lost}^{R\to L}}$ is the mask of pixels lost when warping from the right view back to the left view (Lu et al., 6 Jul 2026).

The DIBR preliminaries are standard. Depth-to-disparity conversion is written as

$d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$

and for rectified stereo the left-to-right warp is

$x' = x - d_L(x,y), \qquad y' = y.$

The warped image and disocclusion mask are produced by $W_{L\to R}(I_L,d_L)$ , and an inpainting network $G$ fills the missing pixels to synthesize the target view (Lu et al., 6 Jul 2026).

The proof idea proceeds through three reductions. First, left-view inpainting is unnecessary, because pixels disoccluded in the $I_L$ 0 warp have no physical correspondence in $I_L$ 1, so they do not affect which pixels become disoccluded when warping back. Second, left-view depth estimation is unnecessary, because disparity can be transferred together with appearance during the forward warp. Third, the reverse-direction disocclusion mask can be replaced by analytic lost-pixel detection. A pixel is lost either by boundary violation,

$I_L$ 2

or by depth occlusion under a z-buffer style rule,

$I_L$ 3

with

$I_L$ 4

The complete lost-pixel mask is

$I_L$ 5

and this equals the target-view disocclusion mask (Lu et al., 6 Jul 2026).

A common misconception is that the theorem is a generic property of all stereo rendering pipelines. The stated result is exact only under nearest-neighbor warping, rectified stereo geometry, standard DIBR disparity-based projection, and a depth ordering rule where closer pixels win in collisions. The same paper gives the symmetric counterpart

$I_L$ 6

but also notes that bilinear interpolation introduces a small approximation gap (Lu et al., 6 Jul 2026).

2. Self-supervision, cycle consistency, and train-inference consistency

The practical significance of the theorem is that it enables analytical computation of test-time disocclusion masks directly from monocular images. Instead of executing a full right-left-right or left-right-left synthesis cycle, one may treat a monocular image as the target view, estimate its disparity, compute the lost-pixel mask analytically, and use that mask as the disocclusion mask required by stereo inpainting. The construction is summarized by

$I_L$ 7

with $I_L$ 8 itself used as supervision (Lu et al., 6 Jul 2026).

The underlying self-supervised signal is a cycle consistency objective. In the right-left-right cycle, the model starts from $I_L$ 9, estimates $I_R$ 0, warps to the left, inpaints, re-estimates or reuses disparity, warps back to the right, and reconstructs $I_R$ 1. The cycle loss is

$I_R$ 2

The theorem makes this cycle analytically collapsible by showing that the decisive disocclusion structure can be extracted from the reciprocal warp geometry without running the full cycle (Lu et al., 6 Jul 2026).

This yields train-inference consistency for the stated warping formulation. If a model is trained on masks obtained from GRT, then at inference the same kind of masks are produced by the same geometric process. The paper explicitly contrasts this with regimes where training masks arise from synthetic corruption or stereo-matching noise while test masks arise from real DIBR disocclusions. In that sense, the theorem is the mechanism supporting the paper’s claim of the first self-supervised framework learning from monocular videos via cycle consistency (Lu et al., 6 Jul 2026).

The same work places GRT inside a concrete stereo inpainting pipeline with depth estimation, GRT mask computation, and stereo inpainting. It constructs datasets including ImageNet-GRT, Kinetics-GRT, and DAVIS-GRT; uses LaMa for image stereo inpainting and ProPainter for video stereo inpainting; and describes a main training loss

$I_R$ 3

The appendix says the image model uses L1 and perceptual components, instantiated as L1 plus LPIPS, while the video model uses L1 reconstruction plus a temporal adversarial term (T-PatchGAN) in the ProPainter-based setup (Lu et al., 6 Jul 2026).

The theorem does not remove DIBR failure modes. The stated limitations are equally geometric: inaccurate disparity or depth leads to incorrect masks, transparent or reflective surfaces can break assumptions, depth discontinuity errors can hurt mask quality, and bilinear interpolation introduces a small approximation gap. The paper therefore prefers nearest-neighbor warping because it gives exact consistency and better matches the theorem statement (Lu et al., 6 Jul 2026).

3. Lattice-point reciprocity: pruned inside-out polytopes and generalized permutahedra

A distinct geometric reciprocity theorem appears in the theory of pruned inside-out polytopes. Classical inside-out polytopes study lattice points in a polytope $I_R$ 4 after removing a hyperplane arrangement $I_R$ 5, namely

$I_R$ 6

The extension replaces the hyperplane arrangement by the codimension-one skeleton of a complete fan. If $I_R$ 7 is a complete fan in $I_R$ 8, then

$I_R$ 9

and a pruned inside-out polytope is

$M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 0

with connected components called regions (Rehberg, 2021).

The theory introduces two counting functions for a rational pruned inside-out polytope. The inner pruned Ehrhart function counts lattice points in the open pruned polytope,

$M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 1

while the cumulative pruned Ehrhart function counts lattice points in closed regions with multiplicity according to how many closed full-dimensional cones contain the point,

$M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 2

The main geometric reciprocity theorem is

$M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 3

with both sides agreeing with quasipolynomials of degree $M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 4 (Rehberg, 2021).

The proof is explicitly not a new Ehrhart theory. The open pruned polytope is decomposed into open regions $M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 5, the inner count becomes a sum of classical Ehrhart counting functions $M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 6, and Ehrhart–Macdonald reciprocity is applied region by region: $M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 7 Summation over regions yields the global reciprocity statement (Rehberg, 2021).

The paper then applies this framework to generalized permutahedra, whose normal fan is a coarsening of the braid fan, and recalls the equivalent submodular description

$M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 8

with $M_{\text{dis}^{L\to R}} = M_{\text{lost}^{R\to L}},$ 9 submodular and $M_{\text{dis}^{L\to R}}$ 0. For $M_{\text{dis}^{L\to R}}$ 1, if

$M_{\text{dis}^{L\to R}}$ 2

then $M_{\text{dis}^{L\to R}}$ 3 is a polynomial of degree $M_{\text{dis}^{L\to R}}$ 4, and

$M_{\text{dis}^{L\to R}}$ 5

This extends the $M_{\text{dis}^{L\to R}}$ 6 reciprocity results associated with Aguiar–Ardila, Billera–Jia–Reiner, and Karaboghossian (Rehberg, 2021).

A major application is to hypergraphic polytopes

$M_{\text{dis}^{L\to R}}$ 7

which are generalized permutahedra. For a hypergraph coloring $M_{\text{dis}^{L\to R}}$ 8, proper means that every hyperedge has a unique maximal-colored node. The paper shows that

$M_{\text{dis}^{L\to R}}$ 9

is polynomial, and proves the reciprocity theorem

$M_{\text{lost}^{R\to L}}$ 0

In particular, $M_{\text{lost}^{R\to L}}$ 1 counts acyclic headings of $M_{\text{lost}^{R\to L}}$ 2 (Rehberg, 2021).

4. Flag-based reciprocity in algebraic $M_{\text{lost}^{R\to L}}$ 3-theory

A further geometric reciprocity theorem arises in algebraic $M_{\text{lost}^{R\to L}}$ 4-theory from the incidence geometry of flags in an irreducible $M_{\text{lost}^{R\to L}}$ 5-dimensional variety $M_{\text{lost}^{R\to L}}$ 6 over a field $M_{\text{lost}^{R\to L}}$ 7. If $M_{\text{lost}^{R\to L}}$ 8 is the function field and

$M_{\text{lost}^{R\to L}}$ 9

is a full flag, then the construction associates a symbol map

$d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 0

A partial flag omits one codimension level $d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 1, and a full flag refines it by restoring the missing term (Musicantov et al., 2014).

The symbol map is built by iterating codimension filtration in $d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 2-theory, boundary maps from localization sequences, projection to the piece determined by the flag, and pushforward to the base. In the paper’s notation, the codimension filtration involves $d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 3 and $d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 4, and the construction passes through a chain

$d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 5

For the collection $d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 6 associated with the flag, the symbol is $d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 7 (Musicantov et al., 2014).

The abstract reciprocity law is the vanishing of the total local contribution over all refinements of a fixed partial flag. If

$d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 8

then

$d_L = \mathcal{D}(I_L), \qquad d_L = \frac{bf}{Z_L},$ 9

hence

$x' = x - d_L(x,y), \qquad y' = y.$ 0

The paper presents this as the abstract geometric reciprocity theorem: local symbols attached to full flags cancel globally when summed over all refinements of a partial flag (Musicantov et al., 2014).

Passing to homotopy groups gives maps $x' = x - d_L(x,y), \qquad y' = y.$ 1, recovering classical reciprocity laws in low dimensions. For a regular connected proper curve over $x' = x - d_L(x,y), \qquad y' = y.$ 2, the $x' = x - d_L(x,y), \qquad y' = y.$ 3-level yields the degree formula for principal divisors,

$x' = x - d_L(x,y), \qquad y' = y.$ 4

At $x' = x - d_L(x,y), \qquad y' = y.$ 5, the symbol becomes the tame symbol and yields Weil reciprocity. After a dual-number base change, the same framework recovers the residue theorem

$x' = x - d_L(x,y), \qquad y' = y.$ 6

and for a local Artinian base ring $x' = x - d_L(x,y), \qquad y' = y.$ 7 it specializes to the Contou-Carrère symbol. In higher dimensions, on $x' = x - d_L(x,y), \qquad y' = y.$ 8, the construction identifies with the Parshin symbol in the regular-flag case and produces Parshin reciprocity (Musicantov et al., 2014).

The same paper also interprets the target $x' = x - d_L(x,y), \qquad y' = y.$ 9 through the delooping property of Tate categories,

$W_{L\to R}(I_L,d_L)$ 0

suggesting an $W_{L\to R}(I_L,d_L)$ 1-Tate interpretation of the symbol map. This suggests a geometric content similar in spirit to other reciprocity theorems: the reciprocity law is derived from the geometry of the codimension filtration and localization sequences rather than from an ad hoc explicit formula (Musicantov et al., 2014).

5. Projective and scattering formulations of geometric reciprocity

Several additional reciprocity theorems in the supplied literature are geometric in the sense that reciprocity is encoded by projective duality, transfer operators, or modal overlaps rather than by direct symmetry of observables.

Domain	Geometric object	Reciprocity statement
Graphic statics	Projective duality $W_{L\to R}(I_L,d_L)$ 2 and projections	$W_{L\to R}(I_L,d_L)$ 3 is a central-axial collineation
Potential scattering	Fundamental transfer matrix $W_{L\to R}(I_L,d_L)$ 4	$W_{L\to R}(I_L,d_L)$ 5
Waveguide scattering	Generalized overlap matrix $W_{L\to R}(I_L,d_L)$ 6 and scattering matrix $W_{L\to R}(I_L,d_L)$ 7	$W_{L\to R}(I_L,d_L)$ 8

In projective graphic statics, the main thesis is that Maxwell-style reciprocal diagrams are not limited to one special self-dual geometry. The paper proves: Any projective duality can be used to get force and form diagrams and any of them has infinitely many pairs of projections to do so. Its formal reciprocity theorem states that the reciprocity condition is equivalent with $W_{L\to R}(I_L,d_L)$ 9 being a central-axial collineation. Polarities are a special case, because they satisfy $G$ 0. The result reframes Maxwell reciprocity as a general projective phenomenon in $G$ 1, with force and form diagrams produced from arbitrary dualities plus suitable projections (Baranyai, 2019).

In multidimensional potential scattering, reciprocity is recast as a property of a fundamental transfer matrix $G$ 2, an operator-valued generalization of the one-dimensional transfer matrix. For a short-range potential, the scattering amplitude satisfies

$G$ 3

and the paper identifies this with the anti-pseudo-unitarity of $G$ 4 with respect to

$G$ 5

namely

$G$ 6

This holds for both real and complex potentials and is ներկայացված as the multidimensional analogue of the one-dimensional $G$ 7 structure behind $G$ 8 (Loran et al., 23 Aug 2025).

In waveguide scattering, the Lorentz reciprocity theorem does not generally imply that the scattering matrix itself is symmetric. For arbitrary waveguide modes, including propagating, evanescent, complex, and degenerate modes, the central statement is instead

$G$ 9

where

$I_L$ 00

is the generalized orthogonality matrix. Only in the special orthogonal case $I_L$ 01 does this reduce to $I_L$ 02, and for propagating modes with $I_L$ 03 it becomes $I_L$ 04. The paper explicitly emphasizes that the reciprocity symmetry lives in $I_L$ 05, not necessarily in $I_L$ 06 itself (Svendsen et al., 2013).

A recurring misconception across these formulations is that reciprocity always appears as direct symmetry of the most obvious observable. The supplied papers collectively show otherwise: reciprocity may be a statement about a projective iterate $I_L$ 07, an anti-pseudo-unitary transfer operator, or a weighted scattering matrix, with the observable symmetry emerging only after the correct geometric or operator-theoretic structure is fixed (Baranyai, 2019, Loran et al., 23 Aug 2025, Svendsen et al., 2013).

The acronym GRT also denotes General Recognition Theory, a different framework in mathematical psychology. In that literature, the relevant result is not a theorem named “Geometric Reciprocity Theorem,” but the paper describes a core geometric reciprocity phenomenon: in fully parameterized $I_L$ 08 Gaussian GRT models, perceptual and decisional structure can be traded off through invertible linear transformations, so one cannot uniquely infer one from the other without fixing some aspect of the coordinate system (Silbert et al., 2016).

General Recognition Theory is a two-stage model in which a stimulus generates a noisy point in an unobserved perceptual space and deterministic decision bounds partition that space into response regions. For a bivariate Gaussian model,

$I_L$ 09

Its three classic notions are perceptual independence (PI), perceptual separability (PS), and decisional separability (DS). In the $I_L$ 10 Gaussian case, PI corresponds to zero covariance or correlation between dimensions, PS means the marginal distribution on one dimension does not depend on the level of the other stimulus dimension, and DS means the decision bounds are parallel to the coordinate axes (Silbert et al., 2016).

The geometric reciprocity result is that failures of DS are not generally identifiable in fully parameterized models. If the angle between a “horizontal” bound and the $I_L$ 11-axis is $I_L$ 12, and the angle between the two decision bounds is $I_L$ 13, then one can apply a rotation and a shear,

$I_L$ 14

which preserve response probabilities because they are invertible linear reparameterizations. Geometrically, the rotation aligns one bound with an axis and the shear aligns the other, so a DS failure can be moved into the perceptual parameters. The paper therefore states that, in the fully general $I_L$ 15 Gaussian model, DS failure is not testable (Silbert et al., 2016).

The same issue persists in GRT with Individual Differences (GRTwIND). Under the assumption of universal perception, GRTwIND posits shared group-level means and covariances with subject-specific scaling parameters $I_L$ 16 and $I_L$ 17. The paper’s main theoretical result is that universal perception plus subject-specific failures of DS is mathematically, and therefore empirically, equivalent to a model with DS holding but universal perception failing. It concludes that if one fits GRTwIND assuming universal perception and finds evidence of DS, PS, or PI violations, one has only rejected the conjunction “universal perception + DS + PS + PI” (Silbert et al., 2016).

The same paper also proves that means and marginal variances are not, in general, simultaneously identifiable in $I_L$ 18 Gaussian GRT, including GRTwIND. After affine rescaling by

$I_L$ 19

one can normalize marginal variances while shifting the means relative to the criteria and preserving response probabilities. The authors therefore argue that GRT models require more than the usual fixing of location and scale: some subset of parameters must also fix the orthogonality of the modeled perceptual dimensions. This is a distinct use of “GRT,” but it is directly relevant to geometric reciprocity as a trade-off between coordinate geometry, perception, and decision (Silbert et al., 2016).