**Note** from Bead: Fourier Light Field · [canonical source](https://redfish.acequia.io/guerin/.agents/8e1768dc-1ae5-448a-b8af-626afbfed2f7/2026-06-30/notes/02-cv-literature-mapping.md) · session 2026-06-30 · discussion: Talk: Fourier Light Field
This note fulfills the requirement to explicitly document the discrete concepts extracted from the Gemini chat transcript (`Ax50d4P5GD1S`) about Fourier Light Fields, and maps them to their closest counterparts in the state-of-the-art Computer Vision (CV) and Machine Vision (MV) literature.
## I. Discrete Concepts Extracted from the Chat The chat transcript decomposes the mechanics of Fourier Light Fields into the following discrete conceptual blocks: **1. Foundational Light Field Mechanics** * **The Plenoptic Function / 4D Light Field:** Capturing scene radiance as $L(x, y, \theta, \phi)$, decoupling spatial position from viewing angle. * **Gaussian Spectral Duality:** A spatial Gaussian analytically transforms to a frequency-domain Gaussian. The covariance matrix $\Sigma$ completely dictates the spectral fall-off. * **Spherical Harmonics (SH) for Appearance:** Using SH as basis functions to model view-dependent variations (specularity), conceptually acting as "angular frequencies." **2. Geometry, Depth, and Phase** * **Fourier Shift Theorem for Camera Translation:** A physical translation of the camera manifests purely as a linear phase rotation in the continuous frequency domain. * **Depth-Coupled Parallactic Phase:** The speed of phase rotation during camera movement is strictly proportional to the object's depth relative to the focal plane. * **Epipolar Slices in Frequency:** Epipolar geometry (traditionally a spatial constraint) manifests as slanted 2D manifolds or trajectories within the 4D Fourier light field. **3. Edges, Boundaries, and Anchoring** * **Dirac Delta Constraints:** Sharp occlusion boundaries act as Dirac delta functions, producing flat (uniform) power spectrums across all high frequencies. * **Phase-Locking:** The phase of high-frequency components at these boundaries remains locked geometry, acting as invariant "anchors" regardless of view angle. **4. Advanced Rendering & Splatting Behaviors** * **3D to 2D Covariance Projection:** 3D semantic elements (like Gaussian Splats) project to 2D image planes; in the frequency domain, this is equivalent to taking a specific 2D central slice of the 3D spectral ellipsoid (the Central Slice Theorem). * **Frequency-Domain Regularization:** Using the expected spectral signatures (like the flat power of edges) to regularize how primitives (like Gaussian splats or neural fields) align themselves during optimization.
## II. Mapping to Computer Vision & Machine Vision Literature These concepts do not exist in a vacuum; they map clearly to historical and cutting-edge vision research. ### A. Light Field Sampling & Fourier Tomography **Closest Literature:** *Plenoptic Sampling (Chai et al., Siggraph 2000)* and *Light Field Rendering (Levoy & Hanrahan, 1996)*. * **Mapping:** The concept of modeling epipolar constraints as slices in the frequency domain originates here. Chai et al. demonstrated that continuous 4D light fields have a bounding spectral support bounded by the minimum and maximum depths of the scene. The chat's focus on "depth-dependent parallactic phase" is the phase-domain equivalent of their Plenoptic spectral bounding box. ### B. Phase-Based Correspondence & Optical Flow **Closest Literature:** *Computation of component image velocity from local phase information (Fleet & Jepson, 1990)* and modern extensions like *PhaseNet*. * **Mapping:** The chat mentions using phase coherence signatures for global alignment without dense semantic matching. Fleet & Jepson established that phase is far more robust to lighting and contrast changes than amplitude. Tracking the linear phase progression in the Fourier domain is exactly how phase-correlation algorithms (like sub-pixel image registration) establish translation matrices. ### C. Frequency Domain Analysis of Neural Rendering (NeRFs) **Closest Literature:** *Mip-NeRF (Barron et al., 2021)* and *Fourier PlenOctrees (Yu et al., 2021)*. * **Mapping:** Mip-NeRF specifically addresses the aliasing generated by sharp boundaries (edges acting as delta functions). By casting 3D conical frustums instead of 1D rays and pre-filtering the positional encodings (which are Fourier basis functions), Mip-NeRF acknowledges that sampling sharp edges requires handling their infinite-frequency bands to prevent rendering artifacts (floaters). ### D. 3D Gaussian Splatting (3DGS) Regularization **Closest Literature:** *3D Gaussian Splatting for Real-Time Radiance Field Rendering (Kerbl et al., 2023)* and very recent preprints focusing on anti-aliased 3DGS. * **Mapping:** The chat's discourse on "Splat Needle Aggregation" and "Projected 2D Covariance" is the mathematical engine of 3DGS. 3DGS evaluates the 2D projection of a 3D Gaussian covariance matrix $\Sigma$. When modeling sharp edges (delta functions), standard Gaussians fail because their exponential fall-off inherently acts as a low-pass filter. The chat's realization that you need "dense clusters of thin Gaussians" to approximate the flat frequency power of an edge is a recognized optimization challenge in current 3DGS literature, leading to regularization techniques that penalize overlapping, highly-elongated Gaussians at occlusion boundaries. ### E. Spherical Harmonics for Specularity Separation **Closest Literature:** *Spherical Harmonic Lighting: The Gritty Details (Green, 2003)* and modern NeRF/Splatting pipelines. * **Mapping:** The dichotomy between Lambertian (constant low-frequency) and Specular (high-frequency view dependence) surfaces via SH is ubiquitous in graphics. Modern inverse rendering pipelines use this exact separation to decouple scene geometry (the spatial Gaussian structure) from its appearance (the SH coefficients) allowing for relighting and novel view synthesis.