Characterizing and Optimizing the Spatial Kernel of Multi Resolution Hash Encodings

Conference: ICLR2026 arXiv: 2602.10495 Code: To be confirmed Area: Other Keywords: multi-resolution hash encoding, neural radiance field, point spread function, spatial anisotropy, Instant-NGP

TL;DR

This paper analyzes Instant-NGP's multi-resolution hash encoding (MHE) through the lens of physical systems, deriving a closed-form approximation of its point spread function (PSF). The analysis reveals that the effective resolution is governed by the geometric mean resolution \(N_{\text{avg}}\) rather than the finest resolution \(N_{\max}\), and that axis-aligned grids introduce spatial anisotropy. The paper further proposes Rotated MHE (R-MHE), a zero-overhead method that eliminates anisotropy by applying a distinct rotation to the input coordinates at each hash level.

Background & Motivation

Background: Multi-Resolution Hash Encoding (MHE) is the core innovation of Instant-NGP, providing efficient spatial parameterization for NeRF and SDF. However, its behavior is highly sensitive to hyperparameters (number of levels \(L\), growth factor \(b\), resolutions \(N_{\max}/N_{\min}\), hash table size \(T\)), which are typically selected heuristically.

Limitations of Prior Work: MHE lacks rigorous analysis from the perspective of physical systems. Fundamental questions remain unanswered: What is the shape of MHE's effective spatial kernel? What is its true resolution limit? How do hash collisions quantitatively degrade quality?

Key Challenge: The intuitive assumption that MHE resolution is determined by the finest level \(N_{\max}\) is incorrect—optimization dynamics cause substantial spatial broadening, so the true resolution is far lower than \(N_{\max}\).

Goal: To develop a rigorous physical analysis framework for understanding MHE's spatial behavior, thereby guiding hyperparameter selection and architectural improvement.

Key Insight: By analogy with Green's functions in physical systems, the spatial characteristics of MHE—resolution, anisotropy, and collision noise—are characterized by measuring its response to a point source (i.e., its PSF).

Core Idea: The effective resolution of MHE is jointly determined by \(N_{\text{avg}}\) and an empirical broadening factor \(\beta_{\text{emp}}\), not by \(N_{\max}\); grid-induced anisotropy can be eliminated through per-level coordinate rotations.

Method

Overall Architecture

The analysis proceeds in three stages: (1) deriving a closed-form approximation of the collision-free ideal PSF, revealing logarithmic decay and B-spline-induced anisotropy; (2) empirically characterizing optimization-induced spatial broadening and establishing the relationship between effective FWHM and \(N_{\text{avg}}\); (3) analyzing collision noise under finite hash capacity and quantifying SNR degradation. The R-MHE improvement is proposed based on these findings.

Key Designs

1. Ideal PSF Derivation (Collision-Free)

   • Function: Derive the spatial response function of MHE after point-source constrained optimization.
   • Mechanism: Under a linearized decoder assumption, the ideal PSF is the average superposition of \(L\) normalized B-spline kernels across levels: \(P_{\text{Ideal}}(\mathbf{x}) = \frac{1}{L}\sum_{l} \hat{B}_l(\mathbf{x})\). Approximating the sum as an integral and applying a Taylor expansion of the B-spline yields the closed form: \(P \approx \frac{1}{L\ln b}[-\ln\|\mathbf{v}\| + C_D - A_D(\mathbf{v})]\), where \(A_D\) captures the intrinsic anisotropy of the B-spline.
   • Design Motivation: PSF is the standard characterization tool in physical systems. The closed-form solution reveals two key properties: (a) logarithmic radial decay (rather than Gaussian or exponential); (b) anisotropy in which the kernel is narrower along coordinate axes than along diagonals.

2. Optimization-Induced Spatial Broadening

   • Function: Quantify how much wider the PSF is after actual training compared to the ideal PSF.
   • Mechanism: A total broadening factor is defined as \(\beta_{\text{emp}} = \beta_{\text{ideal}} \cdot \beta_{\text{opt}}\), where \(\beta_{\text{ideal}} \approx 1.18\) (intrinsic B-spline contribution) and \(\beta_{\text{opt}} > 1\) (optimization contribution). Empirical measurements with the Adam optimizer yield \(\beta_{\text{emp}} \approx 3.0\), meaning the effective FWHM is approximately 2.5 times the ideal value.
   • Design Motivation: This is the most counterintuitive finding—spectral bias (the tendency to learn low frequencies first) causes coarse levels (low \(N_l\)) to be over-weighted, broadening the effective spatial kernel. The true two-point resolvable distance satisfies \(d_{\text{crit}} \propto \beta_{\text{emp}}/N_{\text{avg}}\), not \(1/N_{\max}\).

3. Collision Noise SNR Analysis

   • Function: Quantify signal quality degradation caused by finite hash table capacity.
   • Mechanism: Collisions cause spatially distant grid vertices to share the same feature vector, producing speckle noise: \(P_{\text{Collision}} = P_{\text{Ideal}} + n(\mathbf{x})\), where noise variance increases with the collision rate. Increasing the number of levels \(L\) or the growth factor \(b\) improves SNR for a fixed \(T\).
   • Design Motivation: Provides quantitative guidance for selecting hash table size \(T\)—enabling computation of the minimum \(T\) required to maintain a target SNR for a given scene complexity.

4. Rotated MHE (R-MHE)

   • Function: Eliminate grid-induced spatial anisotropy.
   • Mechanism: A distinct rotation \(\mathbf{R}_l\) is applied to the input coordinates at each level \(l\): \(\mathbf{e}_l(\mathbf{x}) = \text{Interpolate}(\mathbf{F}^l, \mathcal{H}(\lfloor N_l \mathbf{R}_l \mathbf{x}\rceil))\). In 2D, progressive rotations \(\theta_l = l \cdot \theta\) are used; in 3D, rotations are sampled from SO(3) using icosahedral vertex directions. Critically, no additional parameters or computation are introduced—only the coordinate transformation changes.
   • Design Motivation: By using grids with different orientations across levels, the per-level anisotropies cancel upon aggregation, yielding a more isotropic PSF.

Hyperparameter Selection Guidance

Based on the PSF analysis, a theoretical growth factor \(b_{\text{theory}}\) is computed such that \(\beta_{\text{emp}}/N_{\text{avg}}\) matches the target spatial resolution (e.g., a single pixel). Experiments confirm that \(b_{\text{theory}}\) is nearly identical to the empirically optimal \(b_{\text{opt}}\), enabling principled hyperparameter selection without manual tuning.

Key Experimental Results

Main Results

Task	Method	PSNR (dB)
2D Image Regression	Standard MHE (M=1)	23.88
	R-MHE (M=2)	24.62
	R-MHE (M=4)	24.69
	R-MHE (M=8)	24.82 (+0.94)
3D NeRF (Synthetic)	Standard MHE	35.346
	R-MHE (Icosa)	35.479 (+0.13)
3D SDF	Standard MHE	0.9986 IoU
	R-MHE (any)	0.9986 IoU

Ablation Study (PSF Property Verification)

Property	Theoretical Prediction	Experimental Verification
Anisotropy ratio (axis vs. diagonal)	1.17	≈1.17 (exact match)
Total broadening factor \(\beta_{\text{emp}}\) (Adam)	—	≈3.0 (stable across configurations)
FWHM vs. \(N_{\text{avg}}\) relationship	Linear	Linear (exact match)
Two-point resolvable distance \(d_{\text{crit}}\)	\(\propto\) FWHM	Linear correlation (R²≈1)

Key Findings

• Effective resolution is far below \(N_{\max}\): \(\beta_{\text{emp}} \approx 3.0\) implies that the actual resolution is approximately 3× lower than \(N_{\max}\) suggests, explaining the diminishing returns of increasing \(N_{\max}\).
• \(N_{\text{avg}}\) is the true governing parameter: For fixed \(N_{\text{avg}}\), the FWHM remains unchanged regardless of variations in \(L\) and \(b\), greatly simplifying hyperparameter selection.
• R-MHE yields significant gains in 2D but marginal gains in 3D: The improvement is +0.94 dB in 2D and only +0.13 dB in 3D NeRF. The authors attribute this to the ray integration in volumetric rendering, which inherently averages over viewing directions and thus attenuates the effect of anisotropy.
• PSF-guided hyperparameter selection is effective: The theoretically derived \(b_{\text{theory}}\) agrees with the empirically optimal \(b_{\text{opt}}\), eliminating the need for manual tuning.

Highlights & Insights

• Physical thinking applied to neural fields: Employing PSF/Green's function—standard tools from physics—to analyze neural field encodings represents a genuinely novel methodological perspective, directly transferable to other grid-based encodings such as TensoRF and K-Planes.
• Counterintuitive core finding: The result that \(N_{\text{avg}}\), not \(N_{\max}\), governs resolution overturns the intuition that "the finest level determines accuracy" and has direct practical implications for hyperparameter selection.
• Spatial interpretation of spectral bias: The well-known spectral bias phenomenon in optimization is translated into a concrete spatial broadening effect, quantified by the factor \(\beta_{\text{opt}}\).
• Zero-cost improvement via R-MHE: A pure coordinate transformation that introduces no additional parameters or computation is particularly valuable in resource-constrained settings such as mobile rendering.

Limitations & Future Work

• Limited 3D improvement: R-MHE yields only marginal gains on standard 3D benchmarks. Validation on more challenging scenarios (sparse views, high-frequency textures) is needed.
• Linearization assumption: The PSF analysis relies on a linearized decoder assumption; its applicability to deep MLPs requires further verification, although the authors report insensitivity to MLP depth in their experiments.
• Optimizer dependence of \(\beta_{\text{opt}}\): The broadening factor is approximately 3.0 for Adam but differs for other optimizers; a systematic analysis across optimizers is lacking.
• Point-source analysis only: The PSF characterizes the response to a single-point constraint; interactions among multiple constraints in real scenes are more complex.

Related Work & Insights

• vs. Instant-NGP: The original Instant-NGP paper introduced the MHE architecture without analyzing its spatial properties. This work provides a deep theoretical complement, revealing the shape of the spatial kernel, the resolution limit, and the effect of collisions.
• vs. NTK analysis: The NTK literature analyzes the frequency bias of neural networks. This paper instantiates the NTK perspective as a spatial PSF for MHE, yielding quantitative conclusions that are directly actionable in engineering practice.
• vs. TensoRF/K-Planes: All methods based on axis-aligned grids share analogous anisotropy issues. The rotation strategy underlying R-MHE can be directly transferred to these architectures.

Rating

• Novelty: ⭐⭐⭐⭐⭐ — Analyzing neural field encodings via physical PSF/Green's function is a genuinely new methodology; the finding that \(N_{\text{avg}}\) governs resolution is counterintuitive and significant.
• Experimental Thoroughness: ⭐⭐⭐⭐ — Comprehensive validation across 2D, 3D NeRF, and SDF; PSF theory matches experiments precisely; however, 3D improvements are limited.
• Writing Quality: ⭐⭐⭐⭐⭐ — The analysis builds progressively from physical intuition, with rigorous mathematical derivations paired with corresponding experiments.
• Value: ⭐⭐⭐⭐⭐ — Establishes a physically principled analytical methodology for the neural fields community; PSF-based hyperparameter guidance has direct practical utility.

Related Papers

• [ICLR 2026] Probabilistic Kernel Function for Fast Angle Testing
• [ICLR 2026] An Information-Theoretic Framework For Optimizing Experimental Design To Distinguish Probabilistic Neural Codes
• [AAAI 2026] Structure-Aware Encodings of Argumentation Properties for Clique-width
• [AAAI 2026] Tab-PET: Graph-Based Positional Encodings for Tabular Transformers
• [ICLR 2026] Predicting Kernel Regression Learning Curves from Only Raw Data Statistics