SIGReg (Sketched-Isotropic-Gaussian Regularizer)

Summary: SIGReg is a single-scalar regularization loss that enforces latent embeddings to follow an isotropic Gaussian distribution at each timestep. It uses random projections (sketching) combined with the Epps-Pulley statistic (characteristic function distance) — replacing PLDM's multiple regularization terms (variance, covariance, temporal smoothness) with one robust, hyperparameter-light objective.

Overview

Prior JEPA world models (PLDM, DINO-WM) required multiple regularization terms: variance/covariance (VICReg-style), temporal smoothness, and sometimes contrastive terms. SIGReg observes that an isotropic Gaussian latent distribution naturally encourages:

  • Variance → each dimension active (diagonal covariance ≈ I)
  • Covariance → dimensions independent (off-diagonal ≈ 0)
  • Temporal structure → left unconstrained (emerges implicitly)

The "Sketch" part uses random projections to estimate the high-dimensional Gaussian matching efficiently.

Key Components

1. Isotropic Gaussian Target

At each timestep t, latent embeddings z_t ∈ ℝ^D should follow 𝒩(0, I).

2. Epps-Pulley Distance (Characteristic Function)

Measures distance between distributions via characteristic functions:

𝒟_EP(P, Q) = ∫ |φ_P(ω) - φ_Q(ω)|² dμ(ω)

where φ_P(ω) = 𝔼_{x~P}[exp(i ω·x)] is the characteristic function.

For Q = 𝒩(0, I), φ_Q(ω) = exp(-||ω||²/2).

3. Random Projection Sketching

High-D characteristic function integral is intractable. Sketch:

  • Draw K random projection vectors R_k ∈ ℝ^D from 𝒩(0, I)
  • Project latents: u_k = R_k·z ∈ ℝ
  • Compute 1D Epps-Pulley on projected scalars
  • 𝒟_EP(Proj(z), 𝒩(0,I)) ≈ (1/K) ∑_k 𝒟_EP_1D(R_k·z, 𝒩(0,1))

1D characteristic function for 𝒩(0,1): φ(ω) = exp(-ω²/2).

4. Numerical Integration (Knots)

Approximate integral over ω with M quadrature points (knots):

𝒟_EP_1D(P, Q) ≈ ∑_{m=1}^M w_m |φ_P(ω_m) - exp(-ω_m²/2)|²

Paper finds K (projections) and M (knots) are insensitive hyperparameters — robust across wide ranges.

Mathematical Formulation

In plain English: SIGReg asks: "Do the random 1D projections of my latents look like standard Gaussians?" If yes, the full D-dimensional distribution is close to 𝒩(0,I). This is checked by comparing characteristic functions (Fourier transforms of density) at a few points.

Total loss:

ℒ = ℒ_pred + λ * ℒ_SIGReg

ℒ_SIGReg = λ * 𝔼_t[ (1/K) ∑_{k=1}^K ∑_{m=1}^M w_m | φ_{R_k·z_t}(ω_m) - exp(-ω_m²/2) |² ]

Where:

  • φ_{R_k·z_t}(ω_m) = (1/B) ∑_{b=1}^B exp(i ω_m (R_k·z_{t,b})) (empirical CF)
  • K = number of projections (default ~64, robust)
  • M = number of knots (default ~16, robust)
  • λ = only effective hyperparameter (robust in [0.01, 0.2])

Variants / Extensions

  • Analytic SIGReg: If using moment matching instead of CF (but CF captures all moments)
  • Conditional SIGReg: Match 𝒩(0, I) per action? (Paper: per-timestep only)
  • Learned Projections: Instead of random R, learn discriminative projections

Ablation Results (LeWM Paper)

Component Push-T SR ↑
PLDM (multi-term) 78%
LeWM + SIGReg (λ=0.09) 96%
No SIGReg (pred only) ~78% (collapse)
SIGReg λ ∈ [0.01, 0.2] >80% (robust)
Vary K (projections) Flat
Vary M (knots) Flat

Comparison to PLDM Regularizers

Regularizer PLDM LeWM (SIGReg)
Variance (diag cov) Explicit term Implicit in isotropic
Covariance (off-diag) Explicit term Implicit in isotropic
Temporal smoothness Explicit term None (emerges)
Hyperparameters 3+ 1 (λ)
Precision sensitivity High Low

Applications

  • LeWorldModel (end-to-end JEPA from pixels)
  • VL-JEPA (vision-language JEPA)
  • Any latent-space SSL where isotropic Gaussian is a good prior
  • Replacement for VICReg/SimSiam-style regularizers in JEPA

Historical Context

Year Work Regularization Approach
2021 VICReg Variance + Covariance + Invariance
2022 Barlow Twins Cross-correlation matrix → I
2023 I-JEPA / V-JEPA Masked prediction + teacher EMA (no explicit reg)
2024 PLDM (LeCun et al.) VICReg-like + temporal smoothness
2024 DINO-WM DINOv2 features + PLDM-style
2026 LeWM (SIGReg) Single sketched CF matching to 𝒩(0,I)

Related Concepts

Sources

  • leworldmodel-paper: Section 3.1, Appendix C.1 — SIGReg enforces isotropic Gaussian via sketched Epps-Pulley distance; single λ replaces multiple PLDM regularizers
  • leworldmodel-paper: Section 3.1, Equations 3-4 — ℒ_SIGReg = λ 𝔼_t[𝒟_EP(Proj(z_t), 𝒩(0,I))]; random projections + CF matching