Unverified 2026

Laguerre-Optimal Positive Delay Filter

Usefulness6/10

Difficulty6/10

Novelty7/10

Architecture Accuracy Stability Scalability Ssm Rnn Training dynamics Approximation theory Spectral theory Linear algebra Probability

Source paper: Least Variability in a Polynomial-Square Class of Rational Kernels arXiv:2608.29143 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace an Erlang delay or exponential smoothing cascade in a recurrent or state-space layer by a positive rational kernel of the form $\kappa(u)=C e^{-a u}p(u)^2$. Choose the degree-$m$ polynomial by deleting the adjacent pair of Laguerre zeros with smallest relative gap from $L_{m+2}$, then rescale the resulting density to unit mean. This preserves a nonnegative impulse response while reducing temporal jitter relative to Erlang filters.

Formulas

$$\kappa_N^{\mathrm{Er}}(u)=\frac{N^{N}}{(N-1)!}u^{N-1}e^{-Nu},\qquad K_N^{\mathrm{Er}}(s)=\left(\frac{N}{N+s}\right)^N,\qquad \operatorname{Var}(U)=\frac{1}{N}.$$

$$\kappa(u)=C e^{-a u}p(u)^2,\qquad \int_0^\infty \kappa(u)\,du=1,\qquad \int_0^\infty u\kappa(u)\,du=1.$$

$$r_1<\cdots<r_{m+2},\qquad g_j=\frac{r_{j+1}-r_j}{r_j},\qquad j^*=\arg\min_{1\le j\le m+1}g_j,\qquad q(x)=\prod_{i\notin\{j^*,j^*+1\}}(x-r_i).$$

$$\operatorname{Var}_{\mathrm{opt}}(m)=\min_{1\le j\le m+1}\frac{r_{j+1}-r_j}{r_j},\qquad \operatorname{Var}(U)=\int_0^\infty (u-1)^2\kappa(u)\,du.$$

Mathematical statement

The paper's benchmark is the mean-one Erlang kernel $\kappa_N^{\mathrm{Er}}(u)=\frac{N^N}{(N-1)!}u^{N-1}e^{-Nu}$, with Laplace transform $K_N^{\mathrm{Er}}(s)=\left(\frac{N}{N+s}\right)^N$ and variance $1/N$. The proposed class is a nonnegative kernel $\kappa(u)=C e^{-a u}p(u)^2$, where $u\ge 0$, $p$ is a real polynomial of degree $m$, $a>0$, and $C$ normalizes the integral to one. Positivity follows from the square. After imposing unit mean, the paper states that the minimum variance equals the smallest relative gap between adjacent zeros of the Laguerre polynomial $L_{m+2}$, and every optimizer is obtained by deleting the corresponding pair of zeros. Let $r_1<\cdots<r_{m+2}$ be the positive zeros of $L_{m+2}$, define $g_j=(r_{j+1}-r_j)/r_j$, and choose $j^*=\arg\min_j g_j$. Form $q(x)=\prod_{i\notin\{j^*,j^*+1\}}(x-r_i)$, then determine the exponential scale $a$ and normalization $C$ so that the integral and mean equal one. The resulting variance is $\operatorname{Var}(U)=\int_0^\infty (u-1)^2\kappa(u)\,du$.

Implementation notes

Integrate this at the impulse-response level of a causal SSM or recurrent smoothing layer, rather than in the token-mixing projection. For sequence step size $\Delta t$, choose polynomial degree $m$, compute the roots of the generalized Laguerre polynomial $L_{m+2}$ using a symmetric tridiagonal Jacobi-matrix eigensolver, and calculate all relative gaps $g_j=(r_{j+1}-r_j)/r_j$. Delete the pair attaining the minimum gap and construct $q(x)=\prod_{i\ne j^*,j^*+1}(x-r_i)$. Numerically find $a$ by bisection so that the normalized function $e^{-au}q(u)^2$ has mean one; obtain $C$ from the unit-integral constraint. Sample the kernel at $u_k=k\Delta t$, set $w_k=e^{-a u_k}q(u_k)^2$, truncate at a fixed horizon $K$, and renormalize with $w_k\leftarrow w_k/\sum_{k=0}^K w_k$. The forward operation is y[t] = sum_k w[k] * x[t-k]; for an SSM implementation, fit a stable finite-dimensional realization to this impulse response or retain it as a depthwise convolution. The paper provides the positivity guarantee and Laguerre root-selection rule; quadrature, discretization, truncation, and realization are estimated numerically. Begin with degrees 4, 8, and 16 on Copying Memory, Adding Problem, and sequential MNIST. Compare against equal-state Erlang cascades and learned unconstrained FIR filters at matched receptive field, parameter count, and FLOPs. Record impulse-response variance, validation loss, gradient norm, and recurrent rollout stability. The key falsifiable signal is lower validation loss or longer effective memory at equal state budget, together with variance decreasing faster with degree than the Erlang baseline and no negative or oscillatory impulse weights.

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.