Characterizing and Optimizing the Spatial Kernel of Multi Resolution Hash Encodings¶

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

TL;DR¶

Analyzes the Multi-Resolution Hash Encoding (MHE) of Instant-NGP from a physical system perspective, deriving a closed-form approximation of its Point Spread Function (PSF). It reveals that the effective resolution is determined by the average resolution \(N_{\text{avg}}\) rather than the finest resolution \(N_{\max}\), identifies grid-induced anisotropy, and proposes a zero-overhead Rotated MHE (R-MHE) to eliminate anisotropy by rotating input coordinates per layer.

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 dependent on hyperparameters (number of layers \(L\), growth factor \(b\), resolution \(N_{\max}/N_{\min}\), and hash table size \(T\)), which are typically selected through heuristics.

Limitations of Prior Work: MHE lacks rigorous analysis from a physical system perspective. No prior work has addressed: What is the shape of the equivalent spatial kernel of MHE? What is its true resolution limit? How do hash collisions quantitatively affect quality?

Key Challenge: Intuitively, it was believed that MHE resolution is determined by the finest layer \(N_{\max}\). In practice, however, optimization dynamics lead to significant spatial widening, causing the real resolution to be far below \(N_{\max}\).

Goal: Establish a rigorous physical analysis framework to understand the spatial behavior of MHE and guide hyperparameter selection and architectural improvements.

Key Insight: Characterize the spatial properties—resolution, anisotropy, and collision noise—of MHE by measuring its response to a point source constraint (PSF), analogous to the Green's function in physical systems.

Core Idea: The effective resolution of MHE is jointly determined by \(N_{\text{avg}}\) and an empirical optimization widening factor \(\beta_{\text{emp}}\), rather than \(N_{\max}\) alone; grid anisotropy can be eliminated through per-layer rotation.

Method¶

Overall Architecture¶

This paper does not propose a new model but instead answers a long-ignored question: what does the equivalent spatial kernel of Instant-NGP's Multi-Resolution Hash Encoding (MHE) look like, how high is the real resolution, and how do hash collisions degrade quality. The authors treat MHE as a physical system probe—measuring its response to a single point source constraint, i.e., the Point Spread Function (PSF), much like using a Green's function in optics or physics to characterize a system. The analysis progresses in three steps: first, a closed-form PSF approximation is derived under an ideal collision-free setting; second, the extent to which optimization widens this kernel is measured; third, collision noise from finite hash capacity is incorporated into an SNR framework. The final conclusion yields a counter-intuitive judgment—resolution is determined by the average resolution \(N_{\text{avg}}\) rather than the finest layer \(N_{\max}\)—which informs the proposal of the zero-overhead Rotated MHE (R-MHE).

flowchart TD
    IN["Treat MHE as physical system probe<br/>Optimize to fit single point source → Read PSF"] --> D1["Closed-form derivation of ideal PSF<br/>No collisions · Linearized decoder"]
    D1 --> D2["Optimization-induced spatial widening<br/>β_emp≈3.0 (spectral bias)"]
    D2 --> D3["SNR analysis of hash collisions<br/>Finite table T → speckle noise"]
    D3 --> CONC["Core conclusion: Effective resolution<br/>∝ β_emp/N_avg rather than N_max<br/>+ Grid anisotropy"]
    CONC -->|"Back-solve hyperparameters"| HP["Set β_emp/N_avg = Target Resolution<br/>Solve for growth factor b_theory"]
    CONC -->|"Eliminate anisotropy"| RMHE["Rotated MHE<br/>Per-layer coordinate rotation R_l·x"]

Key Designs¶

1. Closed-form derivation of ideal PSF: Clarifying the spatial response of MHE without hash collisions

To determine the "shape of the equivalent spatial kernel," the problem is simplified by assuming a linearized decoder and no collisions in the hash table. In this case, the response of MHE after optimization for a single point source is equivalent to the average superposition of normalized B-spline kernels across \(L\) layers: \(P_{\text{Ideal}}(\mathbf{x}) = \frac{1}{L}\sum_{l} \hat{B}_l(\mathbf{x})\). By replacing the summation with an integral approximation and applying a Taylor expansion to the B-splines, a closed-form is obtained:

\[P \approx \frac{1}{L\ln b}\left[-\ln\|\mathbf{v}\| + C_D - A_D(\mathbf{v})\right]\]

where \(A_D(\mathbf{v})\) represents the anisotropy term inherent to the B-spline. This closed-form reveals two properties: the PSF exhibits logarithmic radial decay (neither Gaussian nor exponential) and is narrower along the coordinate axes than along diagonals—indicating that grid-encoded kernels are naturally anisotropic.

2. Optimization-induced spatial widening: Real trained PSF is much wider than the ideal

The ideal PSF represents only a lower bound; the kernel after actual training is significantly widened. This is a primary counter-intuitive finding. The total widening factor is split into two components \(\beta_{\text{emp}} = \beta_{\text{ideal}} \cdot \beta_{\text{opt}}\): where \(\beta_{\text{ideal}} \approx 1.18\) is inherent to the B-spline, and \(\beta_{\text{opt}} > 1\) arises from the optimization process. Empirically, with the Adam optimizer, \(\beta_{\text{emp}} \approx 3.0\), meaning the effective FWHM is approximately 2.5 times the ideal value. The root cause is spectral bias—the preference for low-frequency learning causes coarse layers (low \(N_l\)) to be over-weighted, widening the spatial kernel. The direct consequence is that the critical distance \(d_{\text{crit}} \propto \beta_{\text{emp}}/N_{\text{avg}}\) for resolving two points is controlled by the average resolution \(N_{\text{avg}}\), not \(N_{\max}\). This explains why simply increasing \(N_{\max}\) yields diminishing returns in practice.

3. SNR analysis for hash collisions: Finite hash tables mix distant vertices

The first two steps assume an infinite hash table, but real scenarios use a finite size \(T\). Collisions cause grid vertices far apart in space to share the same feature vector, superimposing speckle noise onto the PSF: \(P_{\text{Collision}} = P_{\text{Ideal}} + n(\mathbf{x})\), where the noise variance increases with the collision rate. This framework allows the choice of \(T\) to be a calculable problem: for a fixed \(T\), increasing the number of layers \(L\) or the growth factor \(b\) can improve the SNR. This allows for estimating the \(T\) required to maintain a target SNR for a given scene complexity.

4. Rotated MHE (R-MHE): Eliminating anisotropy by rotating input coordinates per layer

Design 1 exposed that the PSF is narrower along coordinate axes. R-MHE is a zero-cost fix for this anisotropy. It applies a different rotation \(\mathbf{R}_l\) to the input coordinates for each layer \(l\) before the lookup: \(\mathbf{e}_l(\mathbf{x}) = \text{Interpolate}(\mathbf{F}^l, \mathcal{H}(\lfloor N_l \mathbf{R}_l \mathbf{x}\rceil))\). In 2D, incremental rotation \(\theta_l = l \cdot \theta\) is used, while in 3D, orientations are sampled via SO(3) using vertices of regular polyhedra. Since each layer has a different grid orientation, the anisotropies cancel out during multi-layer superposition, synthesizing a PSF closer to isotropy. Crucially, this requires no additional parameters or computation, merely a change in coordinate transformation, making it highly valuable for resource-constrained scenarios like mobile rendering.

Hyperparameters can also be calculated directly using this PSF analysis: by setting \(\beta_{\text{emp}}/N_{\text{avg}}\) equal to the target spatial resolution (e.g., single pixel size), the theoretical growth factor \(b_{\text{theory}}\) can be solved. Experiments show \(b_{\text{theory}}\) aligns closely with the empirical optimal value \(b_{\text{opt}}\).

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¶

Property	Theoretical Prediction	Experimental Verification
Anisotropy Ratio (Axis vs Diagonal)	1.17	≈1.17 (Exact match)
Total Widening Factor \(\beta_{\text{emp}}\) (Adam)	-	≈3.0 (Stable across configs)
FWHM vs \(N_{\text{avg}}\) Relationship	Linear	Linear (Exact match)
Resolvable Distance \(d_{\text{crit}}\)	\(\propto\) FWHM	Linear correlation (R²≈1)

Key Findings¶

Effective resolution is far lower than \(N_{\max}\): \(\beta_{\text{emp}} \approx 3.0\) implies the actual resolution is about 3 times lower than what \(N_{\max}\) suggests. This explains diminishing returns when increasing \(N_{\max}\).
\(N_{\text{avg}}\) is the true control parameter: After changing \(L\) and \(b\), the FWHM remains identical as long as \(N_{\text{avg}}\) is the same—significantly simplifying hyperparameter selection.
R-MHE is significant in 2D but marginal in 3D: Gain of +0.94 dB in 2D, but only +0.13 dB in 3D NeRF. The authors explain that ray integration in 3D volume rendering acts as a viewing average, naturally mitigating anisotropy.
PSF-guided hyperparameter selection is effective: The theoretically calculated \(b_{\text{theory}}\) matches the empirical \(b_{\text{opt}}\), eliminating the need for manual tuning.

Highlights & Insights¶

Physical thinking for neural fields: Analyzing neural field encoding using standard physical tools like PSF and Green's functions provides a fresh methodology. This approach can be transferred to other grid encodings like TensoRF and K-Planes.
Counter-intuitive core discovery: \(N_{\text{avg}}\), not \(N_{\max}\), determines resolution—overturning the intuition that the finest layer dictates accuracy and guiding practical hyperparameter selection.
Spatial interpretation of spectral bias: Translates the well-known spectral bias in optimization into specific spatial widening, providing a quantified widening factor \(\beta_{\text{opt}}\).
Zero-cost R-MHE improvement: A coordinate transformation improvement that adds neither parameters nor computation—especially valuable in resource-constrained scenarios like mobile rendering.

Limitations & Future Work¶

Limited 3D improvement: R-MHE shows marginal gains on standard 3D benchmarks; verification in more challenging scenarios (sparse views, high-frequency textures) is needed.
Linearization assumption: The PSF analysis assumes a linearized decoder; its applicability to deep MLPs requires further validation (though experiments suggest it is insensitive to MLP depth).
\(\beta_{\text{opt}}\) depends on the optimizer: The widening factor is approximately 3.0 for Adam but differs for other optimizers—a systematic analysis of various optimizers is missing.
Point source response only: The PSF reflects the response to single-point constraints; multi-constraint interactions in real scenarios are more complex.

vs. Instant-NGP: While the original paper introduced the MHE architecture, it did not analyze its spatial characteristics. This work serves as a deep theoretical supplement, revealing kernel shape, resolution limits, and collision impacts.
vs. NTK analysis: NTK literature analyzes frequency bias in neural networks. This work concretizes the NTK perspective into a spatial PSF for MHE, providing usable quantitative engineering conclusions.
vs. TensoRF/K-Planes: All axis-aligned grid methods suffer from similar anisotropy problems. The rotation concept of R-MHE can be directly transferred.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ Using physical system PSF analysis for neural field encoding is a new methodology; the \(N_{\text{avg}}\) discovery is counter-intuitive and important.
Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive verification across 2D, 3D NeRF, and SDF; PSF theory matches experiments exactly, though 3D improvements are marginal.
Writing Quality: ⭐⭐⭐⭐⭐ Analysis proceeds logically from physical intuition with rigorous mathematical derivation and corresponding experiments.
Value: ⭐⭐⭐⭐⭐ Establishes a physics-based analysis methodology for the neural field community; PSF guided hyperparameter selection has direct practical value.