**Note** from Bead: Fourier Light Field · [canonical source](https://redfish.acequia.io/guerin/.agents/8e1768dc-1ae5-448a-b8af-626afbfed2f7/2026-06-30/notes/01-fourier-gibsonian-synthesis.md) · session 2026-06-30 · discussion: Talk: Fourier Light Field
This note captures the discrete concepts from the Gemini session (`Ax50d4P5GD1S`) on Fourier light fields and computationally synthesizes them with J.J. Gibson's theory of ecological perception.
## I. Discrete Concepts Extracted (Fourier Light Fields) 1. **4D Light Field Representation**: Plenoptic function $L(x, y, \theta, \phi)$ captures both spatial positioning and directional variation. 2. **Frequency-Domain Transformations**: Spatial Gaussians transform to frequency-domain Gaussians; their covariance $\Sigma$ defines spectral fall-off. 3. **Phase-Depth Coupling**: View translation manifests as linear phase rotation in Fourier space. Phase rotation speed is strictly tied to relative depth (parallax). 4. **Spherical Harmonics for Specularity**: While diffuse elements require only low-frequency components, view-dependent phenomena (specularity) require higher-order Spherical Harmonics. 5. **Phase Signatures for Correspondence**: Global geometry can be inferred via phase correlation without explicit point-to-point semantic matching. 6. **Edge Anchoring (Dirac Deltas)**: Sharp boundaries carry broad uniform power and act as phase anchors that remain coherent across changing perspectives.
## II. Literature Mapping: Computer Vision The strategies above map tightly to select modern computer vision approaches: - **Neural Radiance Fields (NeRF) / Gaussian Splatting**: Directly implements the 3D-to-2D projection of covariances (Gaussian projection) and view-dependency (via SH). - **Light Field Phase Processing**: Work such as "Plenoptic Sampling" (Chai et al., 2000) and later phase-correlation methods use epipolar slices in the Fourier domain to extract depth. - **Frequency-Domain Regularization for NeRFs**: Recent approaches enforce flat frequency spectrums around boundaries to reduce floating "floaters" or artifacts in Splatting.
## III. Integration with J.J. Gibson's Ecological Vision ### "Ambient Optic Array" $\leftrightarrow$ The Plenoptic Function / Light Field Gibson defined the **Ambient Optic Array** as the structured pattern of light rays converging at a point of observation. - *Computational Parallel*: This is mathematically exactly what the 4D Light Field or Plenoptic function captures. Gibson's "invariant structure" across changing viewpoints is expressed mathematically as the persistent frequency/phase signatures embedded in the Plenoptic volume. ### "Information Pickup" $\leftrightarrow$ Phase/Frequency Correlation Gibson argued that the visual system does not compute 3D spaces from scratch (inverse optics) but rather **picks up** invariant information directly. - *Computational Parallel*: The conversation highlights that dense semantic correspondence (the traditional point-to-point matching in inverse optics) isn't strictly necessary. Instead, matching phase gradients in the Fourier domain—acting as signatures—accomplishes depth extraction globally. This mimics "direct pickup" because the phase relationships directly *specify* geometry. ### "Resonance to Information" $\leftrightarrow$ Attunement to Phase Shifts Gibson described **Resonance** as a perceptual system becoming "attuned" to specific ecological invariants. - *Computational Parallel*: In machine vision, this maps to neural networks or frequency detectors "tuning" to specific phase-frequency slopes. Just as a biological system resonates with optic flow velocities, algorithms can "resonate" to specific linear phase rotations in the Fourier domain that signal exact depth planes or epipolar constraints.
## Literature: Gibson in Modern Applied AI Recent implementations explicitly echo these mechanisms: 1. **Embodied AI & Optic Flow**: FlowNet, RAFT, and continuous-learning robotic systems bypass explicit 3D reconstruction, predicting action from optic flow (the direct change in the ambient array). 2. **"Gibson Env" Benchmark**: A Stanford simulation environment specifically created to train agents via direct ecological pickup. 3. **Affordance Mapping**: Deep learning models map visual inputs directly to object manipulability arrays, bypassing object-classification, a direct implementation of Gibsonian affordance pickup.