Geometric Shadowing: Concepts & Applications

• Geometric shadowing is defined by geometric constraints that govern exclusion, visibility, or orbit tracing across fields such as computer graphics, astrophysics, and topology.
• It employs methodologies like differentiable rendering, explicit geometric computations, and projection techniques to simulate and compute shadow boundaries accurately.
• Its applications range from enhancing 3D reconstruction and black-hole imaging to validating dynamical systems and improving wireless propagation models.

Searching arXiv for the supplied papers and closely related entries to ground the article. Geometric shadowing denotes a family of geometry-driven notions whose common structure is exclusion, visibility, or orbit-tracing determined by an underlying geometric configuration. In computer graphics, vision, remote sensing, and scattering, it refers to the blocking of light or sensing rays by scene geometry; in black-hole optics, it refers to an optical-geometric reformulation of circular photon orbits and shadow radii; in dynamical systems, it refers to the relation between pseudo-orbits and true orbits; and in geometric topology it can denote orthogonal projections of embedded curves or pursuit-type “shadowing curves.” The term is therefore not univocal: its precise meaning is domain-specific, even when the underlying idea remains geometric constraint or geometric visibility (Luo et al., 4 Jan 2026, Qiao et al., 2022, Kawaguchi, 2017, Dobbins et al., 2017, Wei et al., 2022).

1. Visibility-based meanings in graphics and vision

In rendering-oriented literatures, geometric shadowing is a visibility phenomenon. In satellite 3D reconstruction, it is “the reduction of direct illumination at a surface point caused purely by scene geometry occluding light paths from the sun,” with a binary or continuous solar visibility term $S_{\text{sun}} \in [0,1]$ encoding whether a ray from a scene element toward the sun is blocked by other geometry (Luo et al., 4 Jan 2026). In image restoration, a simple Lambertian model writes

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$

where $V(x)$ is the visibility term, $L(x)$ the direct-light term, $R(x)$ the diffuse reflectance, and $A(x)$ the ambient or indirect term; geometric shadowing is precisely the effect of $V(x)$ (Wang et al., 20 Apr 2026). In inverse rendering, the same role appears in

$$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$

where $V(\omega,\mathbf{x})$ captures illumination visibility, including self-shadowing and cast shadows (Lyu et al., 2021).

This visibility-based usage is technically narrower than “dark region” detection. ShadowGS separates geometric shadowing from albedo, skylight, near-surface reflection, and camera visibility; attached-shadow detection distinguishes attached shadows, which form “on the surface of the occluder where light cannot reach because of self-occlusion,” from cast shadows “projected onto external surfaces” (Luo et al., 4 Jan 2026, Hu et al., 5 Dec 2025). Rough-surface scattering uses the same idea probabilistically: its combined shadowing/masking function $S(\theta_i,\theta_e,\Delta\phi)$ is “the probability that a surface point is both visible and not shadowed,” or equivalently the ratio of irradiated projected area to total projected area (Parviainen et al., 2020).

A common misconception is that geometric shadowing is interchangeable with all illumination variation. The cited formulations separate it from cosine falloff, BRDF effects, ambient occlusion, skylight, atmospheric terms, and material color. This suggests that “geometric shadowing” is most precise when reserved for line-of-sight obstruction by geometry, not for shading in the broader radiometric sense (Luo et al., 4 Jan 2026, Wang et al., 20 Apr 2026).

2. Differentiable rendering and explicit geometric computation

Several works make geometric shadowing explicit and differentiable. “Efficient and Differentiable Shadow Computation for Inverse Problems” approximates occluding geometry with spheres, represents visibility and lighting in spherical harmonics, and rewrites diffuse radiance as

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$0

with visibility constructed from per-sphere blocker functions and a spherical-harmonic exponentiation trick (Lyu et al., 2021). The resulting visibility field is low-frequency, soft, and differentiable with respect to pose and deformation parameters, and is used for texture, illumination, pose, and geometry recovery.

“Face Relighting with Geometrically Consistent Shadows” inserts a hard-shadow term directly into a Lambertian intrinsic-image pipeline. It defines

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$1

so that the directional component is multiplied by a geometry-derived visibility mask $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$2 (Hou et al., 2022). The mask is obtained by sampling points $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$3 along a shadow ray and computing

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$4

with a Sigmoid used to approximate the binary visibility test differentiably (Hou et al., 2022). In that setting, “geometrically consistent shadows” means that shadow boundaries move with the estimated face geometry and the target light direction rather than being synthesized as a purely 2D effect.

A different explicit formulation appears in image-space shadow synthesis via pixel height. “Controllable Shadow Generation Using Pixel Height Maps” introduces a 2.5D representation in which a point $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$5 has a footpoint $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$6 on the ground, with pixel height $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$7, and a point light has image position $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$8 and pixel height $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$9 (Sheng et al., 2022). Projective geometry yields the hard-shadow point

$V(x)$0

The method therefore computes hard shadows by geometry and uses a learned soft-shadow generator only for softness control (Sheng et al., 2022).

3. Geometry-conditioned learning from shadows

A major contemporary use of geometric shadowing is as supervision for 3D reconstruction and inverse rendering. “Towards Learning Neural Representations from Shadows” defines neural shadow fields as volumetric density fields learned only from binary shadow masks and known camera/light poses (Tiwary et al., 2022). The renderer computes ray termination

$V(x)$1

then constructs camera and light z-buffers and predicts shadow masks via depth comparison in light space (Tiwary et al., 2022). In that formulation, geometric shadowing is the differentiable implementation of the statement that a point is in shadow if another point along the light ray is closer to the light.

ShadowGS places geometric shadowing inside 3D Gaussian Splatting for satellite imagery. It computes solar visibility for each Gaussian by ray marching through a BVH, using

$V(x)$2

and builds a simplified remote-sensing rendering equation

$V(x)$3

with explicit direct sun, skylight, near-surface reflection, and albedo terms (Luo et al., 4 Jan 2026). It also introduces a shadow consistency loss

$V(x)$4

based on collinear sun–camera configurations, and reports improvements from MAE $V(x)$5 to $V(x)$6 and PSNR $V(x)$7 to $V(x)$8 across an ablation sequence that adds the rendering equation and shadow consistency (Luo et al., 4 Jan 2026).

Physics-grounded shadow generation from monocular geometry priors uses dense point maps and an approximate directional light to compute a hard geometric shadow support before diffusion refinement (Hu et al., 5 Dec 2025). For an occluder–receiver pair it sets $V(x)$9, $L(x)$0, $L(x)$1, and treats a receiver pixel as shadowed if

$L(x)$2

This is a narrow-cone alignment test around the flow direction $L(x)$3, yielding a hard geometric support that a diffusion model refines into a realistic shadow (Hu et al., 5 Dec 2025). A related iterative detector for attached shadows estimates a global light direction $L(x)$4, combines it with monocular normals $L(x)$5, and derives a partial attached-shadow map from the sign of $L(x)$6; this closed loop reports attached-shadow BER reduction from $L(x)$7 to $L(x)$8 relative to the retrained baseline SILT† (Hu et al., 5 Dec 2025).

Geometry-conditioned shadow removal pushes the same logic in the inverse direction. CFSR uses depth, point clouds, and normals as geometric priors, injects them into attention via a planar correlation matrix

$L(x)$9

and frames shadow removal with a visibility term $R(x)$0 inside a physics-constrained restoration model (Wang et al., 20 Apr 2026). This suggests that, in learning-based systems, geometric shadowing has shifted from a nuisance variable to an explicit geometric cue.

4. Optical geometry and black-hole shadows

In gravitational physics, geometric shadowing has a distinct but structurally related meaning. “Geometric Approach to Circular Photon Orbits and Black Hole Shadows” reformulates circular photon orbits and black-hole shadow radii in optical geometry rather than by an effective potential (Qiao et al., 2022). For static, spherically symmetric spacetimes

$R(x)$1

the equatorial optical metric is

$R(x)$2

Null geodesics of the spacetime become geodesics of this 2D optical metric, with $R(x)$3 as arc length (Qiao et al., 2022).

Within this framework, a circular photon orbit is a circle of zero geodesic curvature: $R(x)$4 Stability is encoded by the Gauss curvature

$R(x)$5

with $R(x)$6 corresponding to an unstable circular photon orbit and $R(x)$7 to a stable one, motivated through Hadamard’s theorem (Qiao et al., 2022). The black-hole shadow radius for an observer at infinity is then

$R(x)$8

The paper proves complete equivalence with the conventional effective-potential method. For Schwarzschild, the geometric approach yields $R(x)$9, $A(x)$0 outside the horizon, and $A(x)$1; for Reissner–Nordström it gives

$A(x)$2

and the corresponding closed-form shadow radius (Qiao et al., 2022). In this literature, “geometric shadowing” does not refer to radiometric visibility but to the intrinsic geometry of the optical manifold from which the shadow boundary is derived.

5. Rough surfaces, radar, and wireless propagation

In rough-surface optics, geometric shadowing and masking are properties of self-affine random rough surfaces $A(x)$3 studied in the ray-optics approximation (Parviainen et al., 2020). The combined function

$A(x)$4

is defined as the probability that a surface point is both visible and not shadowed, with $A(x)$5 (Parviainen et al., 2020). The paper develops horizon mapping and marching methods to accelerate first-order scattering simulations, and emphasizes the azimuthal rough-surface shadowing effect. Small-scale roughness dominates: decreasing Hurst exponent $A(x)$6 for fractional Brownian surfaces or decreasing correlation length $A(x)$7 for Gaussian-correlation surfaces strengthens shadowing, and the azimuthal dependence is “roughly linear in $A(x)$8” over a substantial range for many roughness regimes (Parviainen et al., 2020).

A sensor-specific form appears in FMCW radar. “Beyond a Shadow of a Doubt” exploits the chassis-induced shadow of a roof-mounted rotating radar as a geometric cue for 3D reconstruction (Boisguezennec et al., 24 Jun 2026). If the radar is at height $A(x)$9 and the chassis half-width is $V(x)$0, the opening angle is

$V(x)$1

and a slender vertical object intersecting the shadow yields a return with inner and outer bounds $V(x)$2 (Boisguezennec et al., 24 Jun 2026). Defining $V(x)$3, the in-plane inclination $V(x)$4 is recovered by the branch-wise closed form

$V(x)$5

Under ideal simulation conditions the reported mean errors are below $V(x)$6; in real radar experiments the method achieves approximately $V(x)$7 standard deviation (Boisguezennec et al., 24 Jun 2026).

Wireless communication uses yet another specialized meaning. In “Wireless Link Capacity under Shadowing and Fading,” geometric path-loss is extended by time-invariant random variations of pathloss, so received powers remain distance-based in expectation but become stochastic under shadowing (Halldorsson et al., 2017). The paper shows that for $V(x)$8-smooth shadowing distributions,

$V(x)$9

and, for co-located equal-length links under Log-Normal Shadowing,

$$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$0

even though the deterministic geometric capacity is $$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$1 in that configuration (Halldorsson et al., 2017). This is a distinct usage: “shadowing” there means stochastic path-loss variability rather than optical occlusion, though it still modifies a geometric transmission model.

6. Shadowing in dynamical systems and flows

In dynamical systems, shadowing concerns approximate trajectories rather than optical visibility. For a continuous map $$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$2 on a compact metric space, a $$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$3-pseudo-orbit is a sequence $$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$4 with

$$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$5

and $$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$6 has the shadowing property if every such pseudo-orbit is $$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$7-shadowed by some true orbit $$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$8 (Kawaguchi, 2017). Limit shadowing replaces uniform error control by

$$B(\mathbf{x}) = a(\mathbf{x}) \int_{\Omega} L(\omega)\,V(\omega,\mathbf{x})\,H(\omega,\mathbf{x})\,d\omega,$$9

and requires

$V(\omega,\mathbf{x})$0

for some $V(\omega,\mathbf{x})$1 (Kawaguchi, 2017). Kawaguchi proves that if $V(\omega,\mathbf{x})$2 has the limit shadowing property, then

$V(\omega,\mathbf{x})$3

and $V(\omega,\mathbf{x})$4 satisfies the shadowing property; for equicontinuous maps, shadowing, limit shadowing, and $V(\omega,\mathbf{x})$5 are equivalent (Kawaguchi, 2017).

The literature on shadowing maps studies not only existence but structure of the assignment from pseudo-orbits to shadowing points. A shadowing map $V(\omega,\mathbf{x})$6 is a continuous pseudo-orbit map satisfying a shift-compatibility condition, and Artigue develops a hierarchy

$V(\omega,\mathbf{x})$7

(Artigue, 31 Mar 2025). For expansive homeomorphisms with canonical coordinates, Bowen’s bracket construction yields a “self-tuning” shadowing map from a hyperbolic bracket; conversely, a shadowing map that is both shift-invariant and dynamically invariant is equivalent to topological hyperbolicity (Artigue, 31 Mar 2025).

Flows require reparametrizations. For a continuous flow $V(\omega,\mathbf{x})$8, Komuro-type pseudo-orbit tracing is weakened to average, limit, and asymptotic average shadowing using increasing homeomorphisms $V(\omega,\mathbf{x})$9 with slope uniformly close to $S(\theta_i,\theta_e,\Delta\phi)$0 (Arbieto et al., 2013). For geometric Lorenz flows, however, Araújo and coauthors show that, under the mild condition $S(\theta_i,\theta_e,\Delta\phi)$1 or $S(\theta_i,\theta_e,\Delta\phi)$2 for the induced map on the Poincaré foliation, the flow has neither the $S(\theta_i,\theta_e,\Delta\phi)$3-average shadowing property, nor the $S(\theta_i,\theta_e,\Delta\phi)$4-limit shadowing property, nor the $S(\theta_i,\theta_e,\Delta\phi)$5-asymptotic average shadowing property for any $S(\theta_i,\theta_e,\Delta\phi)$6 (Arbieto et al., 2013). In this branch of mathematics, “geometric shadowing” therefore means geometrically constrained orbit-tracing, not literal shadow formation.

7. Projection and pursuit formulations

Geometric topology uses “shadow” in the sense of orthogonal projection. For $S(\theta_i,\theta_e,\Delta\phi)$7 and nonzero $S(\theta_i,\theta_e,\Delta\phi)$8, the shadow of $S(\theta_i,\theta_e,\Delta\phi)$9 in direction $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$00 is the orthogonal projection onto $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$01,

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$02

“Shadows of a Closed Curve” proves that if $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$03 is a simple closed curve, then at most two coordinate shadows can be simple paths, equivalently no three linearly independent directions can yield shadows homeomorphic to $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$04 (Dobbins et al., 2017). The proof is topological and passes through relations on $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$05, degree-like arguments on $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$06, and the impossibility of a point whose three coordinate shadows are all endpoints of path-shadows (Dobbins et al., 2017).

A distinct pursuit-geometric use appears in planar “shadowing curves.” If an escaper follows a closed planar curve $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$07 and a shadower always stares at the escaper while keeping fixed distance $I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$08, then the shadower satisfies

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$09

(Wei et al., 2022). Writing

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$10

one gets the reduced equation

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$11

whose rotation number depends only on the geometry of the escaping curve, not on its parametrization (Wei et al., 2022). The paper introduces the critical shadowing distance

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$12

and the turning shadowing distance

$I(x) = V(x)\,L(x)\,R(x) + (1 - V(x))\,A(x),$13

which separate periodic, subharmonic, ergodic, and cusp-forming regimes of the shadower’s trajectory (Wei et al., 2022). Here again, the operative idea is geometric constraint rather than radiometric shadow.

Across these literatures, geometric shadowing is best understood as an umbrella term for geometry-governed exclusion or tracing. In rendering and sensing it is a visibility function; in black-hole optics it is an intrinsic optical-geometric mechanism determining light rings and shadow radii; in dynamical systems it is the realization of approximate trajectories by true ones; in topology and pursuit it is projection or constrained following. A plausible implication is that the term retains coherence only at this abstract level: each field specializes it into a mathematically distinct object tied to its own geometry (Luo et al., 4 Jan 2026, Qiao et al., 2022, Kawaguchi, 2017, Dobbins et al., 2017, Wei et al., 2022).