PaperlayerView exact revision on arXiv

Source-linked AI summary

Maximum Entropy Encoding of Energy-Weighted Spherical Moments

Jiaze Sun

arXiv:2608.20429v1cs.GRcs.CV

TL;DR

The paper asks how non-negative Monte Carlo directional energy can be compressed into finite moments while supporting irradiance reconstruction. It derives a maximum-entropy closure and specialized four- and five-parameter variants, then evaluates them against QZH and SH representations. On 981 Poly Haven HDRIs, five-parameter MaxEnt wins against stored QZH on 78.7% of scenes while retaining positivity, although SH-2 remains best overall and coaxial models show closure error on non-coaxial scenes.

  • Problem

    The paper addresses compression and reconstruction of angular energy from non-negative Monte Carlo path samples using finite moments for irradiance.

  • Method

    The paper derives a maximum-entropy closure from 1+3+5 energy-weighted moments, with analytic pure-dipole formulas and a LUT-oriented coaxial five-parameter reconstruction.

  • Results

    78.7% paired win rate against stored QZH is achieved by five-parameter MaxEnt on 981 Poly Haven HDRIs, while full SH-2 remains superior overall.

  • Takeaways & Limitations

    MaxEnt-5 provides a positive nonlinear alternative to finite-order QZH truncation, with its advantage increasing as lighting directionality increases.

  • Takeaways & Limitations

    The evaluation uses static environment lighting and Lambert cosine response, while the coaxial assumption causes systematic closure error on non-coaxial scenes.

Abstract

from arXiv · show

We study how angular energy signals composed of non-negative Monte Carlo path samples can be compressed and reconstructed for irradiance using finite moments. Writing each sample as an energy-weighted directional feature $x = r u$, we adopt total energy, the first directional moment, and the traceless second moment as $1+3+5$ linearly additive, rotationally covariant statistics. Under a fixed Lebesgue reference measure, the maximum-entropy closure yields $p(r,u) \propto \exp(-βr g(u))$, where $g(u) = 1 - b \cdot u + u^T Q u$, whose directional probability and angular energy density are proportional to $g^{-3}$ and $g^{-4}$, respectively. When $g_{\min} > 0$ the closure is normalizable and the reconstruction is strictly positive. We further provide analytic moment matching, variance, inverse sampling, and closed-form diffuse response for the pure-dipole four-parameter subfamily, as well as the realizability domain, partition function, azimuthal algebraic integral, and LUT-oriented reconstruction form for the dipole-second-moment coaxial five-parameter subfamily. Experiments cover 981 Poly Haven HDRI 2K scenes and three Debevec probes. Five-parameter MaxEnt achieves a 78.7% per-scene win rate against stored QZH, with mean luminance RMSE reduced by 15.8%; the advantage is more pronounced in scenes with strong directionality. Both MaxEnt variants maintain zero negative irradiance across all scenes. Full second-order SH-2 yields the lowest overall error, while five-parameter MaxEnt ranks second and outperforms SH-2 in the high-directionality bucket; the coaxial subfamily shows systematic closure error on non-coaxial multi-source scenes.

1 Problem Definition, Contributions, and Notation

The paper compresses non-negative directional energy into linearly additive, rotationally covariant moments and derives a maximum-entropy reconstruction that preserves low-order angular structure. It develops pure-dipole and coaxial five-parameter variants alongside comparisons with SH and QZH representations.

  • Main Contributions: The MaxEnt closure derived from independent 1+3+5 constraints has directional probability proportional to g^-3 and energy density proportional to g^-4.The closure also provides normalizability conditions and a relationship to SH projection.
  • Main Contributions: The pure-dipole subfamily supplies analytic moment matching and cosine response, while the coaxial five-parameter subfamily provides projection, partition-function, and LUT-oriented reconstruction tools.The pure-dipole family coincides with Levermore’s M1 closure; the paper’s contributions there are graphics-oriented encodings and runtime formulas.
  • Experiments and Related Work: The work compares MaxEnt accuracy and negative-value behavior with QZH and SH representations on a 981-scene HDR benchmark and analytic stress tests.The paper also relates the five-parameter model to stored QZH and distinguishes it from spherical-Gaussian approaches through independent axial second-moment storage.
  • Notation: The paper uses SH-1 and SH-2 to denote maximum spherical harmonic degrees l ≤1 and l ≤2, respectively.The ZH3 paper uses SH2 and SH3 for the corresponding band counts.

3 First Principles

The first-principles formulation treats each sample as a non-negative energy-weighted direction and imposes moment-preserving, additive, rotational, and minimal-dimensional encoding requirements. Maximum entropy then selects the distribution under the specified constraints and Lebesgue reference measure.

  • Sample Model: Each Monte Carlo path sample is represented as a direction multiplied by a non-negative scalar energy weight.The four-dimensional feature unifies the scalar weight with its directionally weighted quantity.
  • Sample Model: Signed estimators or negative filter weights require decomposition into positive and negative measures before applying the non-negative-measure model.The cone constraints cannot be applied directly to signed inputs.
  • Encoding Constraints: The encoding is required to preserve moments through second order, support anisotropic shapes, and remain first-order homogeneous, linearly additive, and rotationally covariant.These constraints define the structural requirements for filtering, moment matching, and irradiance reconstruction.
  • Encoding Constraints: The representation uses nine real parameters as the dimensional lower bound for any linear, rotationally covariant encoding that fully preserves spherical l ≤2 moments.The minimum-dimensionality claim applies only under the specified representation constraints.
  • Maximum-Entropy Principle: Maximum entropy selects the distribution that introduces no unobserved information under given moment constraints and a Lebesgue reference measure.This is a conditional optimality statement, not a claim that no better scheme exists under another encoding, loss, or reference measure.

4 Derivation of a Four-Dimensional Tensor Representation

The section derives a nine-degree-of-freedom, linearly additive tensor representation that preserves energy, first directional moments, and traceless second-order angular information.

  • Tensor construction: The aggregate tensor is obtained by algebraic accumulation over samples and is linearly equivalent to the complete real l ≤2 spherical-harmonic coefficients.This representation retains rotational covariance and second-order completeness.
  • Tensor construction: The four-dimensional energy–direction feature yields a symmetric tensor with 9 independent degrees of freedom after its spatial-trace identity is enforced.The decomposition is 1 scalar + 3 vector + 5 traceless symmetric-tensor components.
  • Rotational structure: The scalar, vector, and traceless tensor blocks correspond respectively to the l = 0, 1, and 2 irreducible components of SO(3).Under the stated covariance and homogeneity requirements, the minimal representation must carry information invertibly equivalent to the energy–stress tensor.
  • Why the stress moment is required: The required stress moment is E[RUUT], not the classical second moment E[R2UUT], because radiant-energy scaling differs between the two statistics.The classical second moment scales quadratically with sample-energy rescaling, whereas the required moments scale linearly.
  • Why the stress moment is required: No universal mapping from classical mean and covariance to the first two angular projections exists, even when total energy is additionally supplied.The constructed counterexamples share classical moments while differing in energy-weighted angular moments.
  • Alternative closures: A single-lobe spherical Gaussian cannot independently represent the general traceless second moment because its second moment is coaxial and axisymmetric with the dipole.It also is not closed under linear mixing without nonlinear refitting.

5 Derivation of the MaxEnt Distribution

The section applies maximum entropy to the independent 1+3+5 moment constraints, producing a radial–directional density governed by a scalar angular function g.

  • Maximum-entropy formulation: Maximum entropy selects the density satisfying the 1+3+5 moment constraints under a fixed Lebesgue reference measure.The independent constraints remove the tensor-trace multiplier redundancy.
  • Closed-form density: The Cartesian closure has the form pX(ru) = C exp(−βr g(u)), with g(u) = 1 − b · u + uTQu.Here β is the radial scale, b is a vector parameter, and Q is traceless symmetric.
  • Closed-form density: Integrating over radius gives directional probability proportional to g^-3, while angular energy density and irradiance reconstruction are proportional to g^-4.The extra inverse power arises because each feature contributes energy proportional to its radius.
  • Moment matching: For interior realizable moments, strict convexity of the log-partition function makes the moment-matching natural parameters unique after gauge elimination.The realizable-domain boundary is approached by weak-limit measures with natural parameters tending to infinity.
  • Partition function: When Q = 0 and b = κn, the angular partition integral is ZΩ = 4π/(1 −κ2)2.The full spatial partition function separates into radial and angular factors.
  • Degrees of freedom: The normalized spatial distribution has 9 degrees of freedom, while its spherical shape has 8 and requires one additional scalar to carry total radiant energy.The normalization coefficient is determined by shape parameters and is not itself the total energy.

6 Properties of the Complete Second-Order Distribution

The complete second-order MaxEnt family is positive and rotationally covariant throughout its normalizable domain, while its nonlinear form generates higher-frequency structure beyond SH2.

  • Relation to SH2: Near the uniform distribution, MaxEnt has SH2 as its tangent space, but the full family is not globally equivalent to SH2.Higher powers of the perturbation generate progressively higher spherical-harmonic degrees, and the complete density generally contains infinitely many bands.
  • Relation to SH2: The second-order nonlinear mapping from MaxEnt parameters to SH2 coefficients includes quadratic interactions between the dipole and tensor parameters.This distinguishes local first-order agreement from the full nonlinear closure.
  • Normalizability and positivity: When gmin > 0, the reconstructed density is smooth and strictly positive over the sphere, unlike a general SH2 function.The nonlinear g^-3 form also generates higher-frequency bands from the low-order sufficient statistics.
  • Geometry: Distribution peaks occur at minima of g, and the peak direction is generally neither parallel to b nor an eigenvector of Q.The stationary-point condition couples the dipole and tensor parameters.
  • Geometry: Under rotations, b and Q transform covariantly, while antipodal asymmetry is introduced solely by b.The tensor term itself has antipodal symmetry.
  • Conditional radial law: Conditioned on direction, the radius is Gamma-distributed with mean 3/(βg(u)) and variance 3/(β2g(u)2).The radial statistics therefore vary with direction through g(u).

7 First-Order Reduction: Analytic Closure of the Pure-Dipole Model

The pure-dipole reduction sets Q = 0 and provides an analytically tractable four-parameter closure for moment matching, sampling, covariance, and diffuse response.

  • Model and realizability: The pure-dipole model retains total energy and the first directional moment, targeting constrained budgets or settings with weak second-order anisotropy.Its state is the augmented vector L = (v, ω), with realizability condition ω ≥ |v|.
  • Model and realizability: In the normalizable domain, the dipole parameter is b = κn with 0 ≤ κ < 1, where n is the normalized first-moment direction.The anisotropy is ρ = |v|/ω.
  • Conditional radial law: The conditional radius remains Gamma-distributed with shape 3 and rate β(1 −κn · u).This specializes the general radial law to the pure-dipole angular factor.
  • Covariance interpretation: The reported covariance formulas describe the inferred MaxEnt distribution rather than empirical sample variances.They should not be interpreted as substitutes for sample variances.
  • Diffuse response: The closed-form raw diffuse response separates into symmetric and antisymmetric parts, with eraw,A = 2κµ0/(3 + κ2) = ρµ0/2.The limiting cases include isotropy as κ → 0 and max(0, µ0) as κ → 1.
  • Diffuse response: The theoretical diffuse reconstruction is non-negative for 0 ≤ κ < 1; the implementation’s max operation only suppresses floating-point negatives.Under normalized cosine convention, isotropic output is ω/(4π).
  • Inverse sampling: The exact inverse-CDF sampler targets the directional probability proportional to g^-3, whereas diffuse reconstruction uses g^-4 and therefore requires a different cube-root inverse.The two sampling laws cannot be reused interchangeably.

8 Coaxial Five-Parameter MaxEnt Subfamily

The coaxial five-parameter MaxEnt subfamily aligns the dipole and second-moment axes, retaining second-order information while enabling analytic reconstruction and LUT-oriented evaluation.

  • Model and representation: The coaxial restriction reduces the full 9-DOF model to five parameters by sharing one axis between the dipole natural parameter and stress tensor.The parameters are β, κ, α, and an axis n ∈ S2.
  • Model and representation: Five scalar moments—total energy, axial momentum magnitude, axial second moment, and axis—represent the coaxial state.When the dipole is nonzero, the state is equivalently represented by (ω, v, ζ).
  • Limitations: Mixing coaxial states with different axes generally leaves the subfamily, so filtering should accumulate in the linear 9-DOF tensor before projecting back.This is the principal closure boundary for practical filtering.
  • Moment mapping: The realizable domain is P2(ρ) ≤ η ≤ 1, with the zero-dipole branch satisfying −1/2 ≤ η ≤ 1.Inverse recovery uses the dipole direction when |v| > 0 and a principal stress-tensor eigenvector when v = 0.
  • Moment mapping: The partition function is Z(β, κ, α) = (4π/β^3)Q(κ, α), and derivatives of ln Z provide closed-form moment and covariance quantities.The inverse mapping generally requires a two-dimensional Newton solve within the exact normalizable domain.
  • Diffuse response: The azimuthal integral reduces to scalar arithmetic and two real square roots, yielding a one-dimensional outer integral without complex roots or logarithms.The algebraic form also has continuous limits recovering the pure-dipole and isotropic cases.
  • Diffuse response: A LUT implementation converts the outer integral into offline table construction and uses one trilinear texture sample at runtime.The shader performs no quadrature or root-finding; R16F tables at sizes 323, 483, and 643 occupy 64, 216, and 512 KiB.

9 ZH3/QZH as the Second-Order Truncation of Five-Parameter MaxEnt

The paper identifies stored ZH3/QZH as the exact l ≤ 2 orthogonal truncation of the coaxial MaxEnt energy density. MaxEnt retains a strictly positive full-density closure, whereas the truncation can develop negative lobes when its positivity conditions fail.

  • Representation: The coaxial representation contains a DC term, full linear SH, and one axisymmetric quadratic coefficient, totaling five scalars per channel.The axis n corresponds to the axisymmetric second-order direction.
  • Orthogonal projection: ZH3/QZH is the exact l ≤ 2 orthogonal truncation of the coaxial MaxEnt angular energy density, not a separate asymptotically matched probability family.Its stored second-order coefficient corresponds to ωρ2 after basis-function normalization conversion.
  • Orthogonal projection: The truncation discards high-frequency components that remain nonlinearly constrained by the MaxEnt natural parameters.Full MaxEnt cannot recover arbitrary unknown high frequencies, but supplies a positive closure consistent with the five stored moments.
  • Positivity: When gmin > 0, the full MaxEnt density remains strictly positive, unlike the truncated polynomial, which is not guaranteed to be non-negative.The two positivity conditions constrain different reconstructions: gmin bounds the untruncated MaxEnt density, while the quadratic condition bounds ZH3.
  • Positivity: When the truncation positivity condition fails, ZH3 produces negative lobes and ringing, whereas the untruncated MaxEnt reconstruction remains positive.This contrast follows from applying separate positivity constraints to the two representations.

10 Experimental Comparison: Irradiance Reconstruction Accuracy of MaxEnt vs. SH/QZH

Across 981 Poly Haven HDRIs and three Debevec probes, SH-2 achieves the lowest overall irradiance error, while five-parameter MaxEnt is the strongest compact alternative and preserves non-negativity.

  • Overall accuracy: 78.7% of scenes favor MaxEnt-5 over stored QZH, with mean luminance RMSE difference −1.29 × 10−3.The bootstrap 95% confidence interval excludes zero, and both the Wilcoxon signed-rank and sign tests are highly significant.
  • Scene-property dependence: MaxEnt-5’s advantage over stored QZH increases with directionality, reaching a 91.3% win rate for ρ ∈[0.6, 1).In this bucket, MaxEnt-5 median RMSE is 0.0045, approximately 0.53× stored QZH and lower than SH-2 at 0.0073.
  • Non-negativity: MaxEnt-4 and MaxEnt-5 produce zero negative irradiance across all 981 scenes, unlike SH-1, SH-2, and both QZH variants.SH-1 has negatives on 579 scenes, SH-2 on 120 scenes, and the QZH variants on approximately 52–60 scenes each.
  • Probe controls: On three Debevec probes, SH-2 has the lowest RMSE, while five-parameter MaxEnt and stored QZH are numerically close.Among four-DOF methods, curve-fit QZH is slightly lower than MaxEnt-4, and both outperform SH-1.

11 Limitations

The closure is limited by its reliance on moments up to l ≤2, non-negative measures, and a coaxial five-parameter restriction. Numerical inversion becomes ill-conditioned near the realizability boundary, while LUT reconstruction introduces approximation errors.

  • 11.1 Model Scope: The closure cannot recover arbitrary high-frequency structure because it is determined solely by moments up to l ≤2.
  • 11.1 Model Scope: The coaxial five-parameter model may incur systematic closure error for multi-peak, biaxial, or generally non-coaxial light fields.
  • 11.1 Model Scope: Signed path weights, control variates, and negative-kernel filters cannot directly use the non-negative measure formulation.
  • 11.2 Numerical Implementation: Natural-to-moment inversion is unique in the realizable interior but becomes ill-conditioned near its boundary.
  • 11.2 Numerical Implementation: LUT reconstruction shifts the outer diffuse-response integral offline but introduces grid, interpolation, and quantization errors.
  • 11.3 Dataset and Evaluation Scope: The benchmark uses static environment lighting and Lambert cosine response, while non-coaxial scenes produce systematic error under the five-parameter assumption.

12 Conclusion

The paper builds a nine-scalar, moment-preserving MaxEnt representation and specialized four- and five-parameter reconstructions. On 981 Poly Haven HDRIs, five-parameter MaxEnt beats stored QZH in paired comparisons while remaining non-negative, although SH-2 is superior on non-coaxial multi-source scenes.

  • 12 Conclusion: The method preserves angular information through l ≤2 moments and uses a maximum-entropy closure with g−3 directional probability and g−4 energy density.
  • 12 Conclusion: The pure-dipole subfamily supports analytic reconstruction and inverse sampling, while the coaxial five-parameter subfamily uses an offline LUT for reconstruction.
  • 12 Conclusion: The five-parameter state matches stored QZH’s low-order axisymmetric moments but replaces polynomial truncation with a positive nonlinear closure.
  • 12 Conclusion: 78.7% paired win rate: five-parameter MaxEnt outperforms stored QZH across 981 Poly Haven HDRIs, with the advantage increasing with lighting directionality.
  • 12 Conclusion: Zero negative values: both MaxEnt variants remain non-negative across all scenes, while full second-order SH-2 remains superior on non-coaxial multi-source scenes.

Statement on AI-Assisted Tools

The authors used generative AI tools for manuscript drafting, implementation, and verification design, and reviewed the resulting arguments, analyses, references, and wording.

  • Generative AI tools assisted with manuscript drafting, program implementation, and verification design.
  • The authors reviewed all mathematical arguments, experimental analyses, bibliographic references, and final wording and take responsibility for the work.
Loading 2608.20429v1…