Unverified 2026

Resolvent response basis for conditioned modules

Implementation & benchmark of arXiv:2608.15805 — Spectral duality structures and the Fisher--Rao geometry of reset distributions

Usefulness5/10

Difficulty6/10

Novelty7/10

Architecture Memory Speedup Scalability Diffusion Diffusion sampling Moe routing Inference speedup Spectral theory Linear algebra Approximation theory

Source paper: Spectral duality structures and the Fisher--Rao geometry of reset distributions arXiv:2608.15805 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Compress a module whose output changes with a scalar condition such as diffusion time, temperature, or compute budget by representing its response in a low-rank basis generated by resolvent-like functions. Distinct spectral modes produce rational factors $(1-\tau\lambda_k)^{-1}$, allowing a small number of learned components to approximate a large hypernetwork or condition-dependent parameter table.

Formulas

$$R(\tau)=\sum_k(1-\tau\lambda_k)^{-1}P_k.$$

$$\operatorname{span}\{R(\tau)\xi:\tau\in J\}=\bigoplus_{k:\,P_k\xi\neq0}\operatorname{span}\{P_k\xi\}.$$

$$M_{ij}=(1-\tau_i\lambda_{k_j})^{-1},\qquad i,j=1,\ldots,s.$$

$$y(\tau)\approx Bc(\tau),\qquad c_j(\tau)=\frac{\alpha_j}{1-\tau\widehat\lambda_j},\qquad \widehat\lambda_j\notin[\tau_{\min}^{-1},\tau_{\max}^{-1}].$$

Mathematical statement

Let $P_k$ be spectral projectors of an operator with pairwise distinct active eigenvalues $\lambda_k$, let $\xi$ be a seed vector, and define the resolvent $R(\tau)=\sum_k(1-\tau\lambda_k)^{-1}P_k$ for scalar conditions $\tau$ away from poles. Lemma A.1 gives $\operatorname{span}\{R(\tau)\xi:\tau\in J\}=\bigoplus_{k:P_k\xi\ne0}\operatorname{span}\{P_k\xi\}$. Therefore, if there are $s$ active distinct eigenvalues, responses at $s$ distinct conditions are sufficient to recover the active response span when the Cauchy matrix $M_{ij}=(1-\tau_i\lambda_{k_j})^{-1}$ is invertible. The ML adaptation replaces the unknown projectors by an empirical SVD basis $B$, and uses learned rational coefficients $c_j(\tau)=\alpha_j/(1-\tau\widehat\lambda_j)$, where $\widehat\lambda_j$ are constrained so that no pole lies in the operating condition interval.

Implementation notes

Apply this to a scalar-conditioned adapter, diffusion denoiser block, temperature-conditioned attention projection, or MoE router. Choose $K$ calibration conditions $\tau_1,\ldots,\tau_K$ over the deployment interval and collect flattened original outputs or parameter deltas $y_i=y(\tau_i)\in\mathbb R^d$. Stack them as $Y=[y_1,\ldots,y_K]$, compute $Y=U\Sigma V^{\mathsf T}$, and retain the smallest rank $r$ explaining 99 percent of calibration variance. Store $B=U_{[:,1:r]}$. Replace the original conditioner with coefficients $c_j(\tau)=\alpha_j/(1-\tau\widehat\lambda_j)$, form $\hat y=Bc(\tau)$, and feed $\hat y$ into the downstream layer. Parameterize $\widehat\lambda_j$ so its pole remains outside $[\tau_{\min},\tau_{\max}]$, for example by setting $\widehat\lambda_j=\tau_{\max}^{-1}+\operatorname{softplus}(b_j)$ when conditions are positive. Fine-tune with the original task loss plus reconstruction loss $\|y(\tau)-\hat y(\tau)\|_2^2$. The exact paper contribution is the span-recovery mechanism; empirical SVD and learned poles are adaptations. Start with a small diffusion U-Net adapter or temperature-conditioned MLP. Compare against the full conditioner and ordinary rank-r linear interpolation. Report parameter count, latency, unseen-condition reconstruction error, and task quality. Success means at least 2x fewer conditioner parameters or lower latency with under 1 percent relative task degradation.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.