OctoInformBeta

OpenAI preprint, September 2026 · §4. Matrix variation along the filter · §4.1

§4.1 (setting): cutoff, potentials, scales and the energy E_θ (scalar-potentials), (energy-definition): The exponent a=3/2a=3/2, cutoff f0f_0, weights w,φw,\varphi and potentials p,p1p,p_1; a run of moves costs the summed trace losses and gains of h=p+p1h=p+p_1

Weights, potentials and the energy of a run of moves.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

The filter is a process that updates a real coefficient matrix RR one step at a time, and for nonnegative weights qq the state Rdiag⁡(q)RTR\operatorname{diag}(q)R^{T} of RR is a positive semidefinite matrix. This item fixes scalar functions of the matrix, two matrix operations that go with them, and an energy that charges each update of a matrix for the trace it loses and the trace it gains.

Scalar Functions

The cutoff f0f_0 is the continuous piecewise linear function on the reals given by f0(x)={0,x≤1/2,2x−1,1/2<x<1,1,1≤x≤2,2−x/2,2<x<4,0,x≥4.f_0(x)= \begin{cases} 0, & x\le 1/2,\\ 2x-1, & 1/2<x<1,\\ 1, & 1\le x\le 2,\\ 2-x/2, & 2<x<4,\\ 0, & x\ge 4. \end{cases}(1)✦

The powers below are taken with the exponent, the real number a=32.a=\frac{3}{2}.(2)✦ With it, the weight ww and the size weight φ\varphi are the real functions w(x)=min⁡{xa,x−1},w(x)=\min\{x^{a},\,x^{-1}\},(3)✦ φ(x)=min⁡{xa−1,x−1}.\varphi(x)=\min\{x^{a-1},\,x^{-1}\}.(4)✦

From the weight and the exponent come two potentials: the potential pp and the gain potential p1p_1 are the real functions p(x)=∫0xw(u)udu,p(x)=\int_0^x\frac{w(u)}{u}\,du,(5)✦ p1(x)=∫0xua−1(1+u)adu.p_1(x)=\int_0^x\frac{u^{a-1}}{(1+u)^{a}}\,du.(6)✦

Matrices and Moves

Matrices are real and indexed by finite sets ι\iota and κ\kappa. The lifted matrix of an ι×ι\iota\times\iota matrix AA along an ι×κ\iota\times\kappa matrix RR is the κ×κ\kappa\times\kappa matrix RTAR.R^{T}AR.(7)✦ The weighted square of an ι×κ\iota\times\kappa matrix AA against a κ×κ\kappa\times\kappa matrix WW is the real number tr⁡(AWAT),\operatorname{tr}\bigl(AWA^{T}\bigr),(8)✦ which for positive semidefinite WW is the squared Hilbert-Schmidt norm ∥AW1/2∥HS2\|AW^{1/2}\|_{\mathrm{HS}}^{2}.

A move on ι×ι\iota\times\iota matrices is a pair π=(C,G),C,G∈ℝι×ι,\pi=(C,G),\qquad C,\,G\in\mathbb{R}^{\iota\times\iota},(9)✦ of a contraction CC and a gain GG. Under π\pi, an ι×ι\iota\times\iota matrix MM has the contracted state CMCTCMC^{T}✦ and the state after the move π(M)=CMCT+G.\pi(M)=CMC^{T}+G.(10)✦

A run is a finite list π0,…,πn−1\pi_0,\dots,\pi_{n-1} of moves, with πk=(Ck,Gk)\pi_k=(C_k,G_k). The pair that the run leads to from a pair (Z,V)(Z,V) of ι×ι\iota\times\iota matrices is (Zn,Vn)(Z_n,V_n), where (Z0,V0)=(Z,V),(Zk+1,Vk+1)=(CkZkCkT,πk(Vk))(0≤k<n).(Z_0,V_0)=(Z,V),\qquad (Z_{k+1},V_{k+1})=\bigl(C_kZ_kC_k^{T},\ \pi_k(V_k)\bigr)\quad(0\le k<n).(11)✦

The Definition

Definition 1. Let π0,…,πn−1\pi_0,\dots,\pi_{n-1} be a run of moves on ι×ι\iota\times\iota matrices, with πk=(Ck,Gk)\pi_k=(C_k,G_k). The final state of the run from an ι×ι\iota\times\iota matrix DD is the matrix DnD_n, where D0=D,Dk+1=πk(Dk)(0≤k<n).D_0=D,\qquad D_{k+1}=\pi_k(D_k)\quad(0\le k<n).(12)✦ The potential trace and the gain trace of an ι×ι\iota\times\iota matrix AA are tr⁡p(A)\operatorname{tr}p(A) and tr⁡p1(A)\operatorname{tr}p_1(A), where p(A)p(A) and p1(A)p_1(A) are given by the continuous functional calculus; with h=p+p1h=p+p_1, the trace of h(A)h(A) is the sum of the potential trace and the gain trace of AA: tr⁡h(A)=tr⁡p(A)+tr⁡p1(A).\operatorname{tr}h(A)=\operatorname{tr}p(A)+\operatorname{tr}p_1(A).(13)✦ The cost of a move π=(C,G)\pi=(C,G) at a matrix MM is the trace of hh lost under the contraction plus the trace of hh added by the gain: cπ(M)=[tr⁡h(M)−tr⁡h(CMCT)]+[tr⁡h(π(M))−tr⁡h(CMCT)].c_\pi(M)=\bigl[\operatorname{tr}h(M)-\operatorname{tr}h(CMC^{T})\bigr] +\bigl[\operatorname{tr}h(\pi(M))-\operatorname{tr}h(CMC^{T})\bigr].(14)✦ The energy cost of the run from DD is the sum of the costs of its moves at the states (12) it passes through: E(D;π0,…,πn−1)=∑k=0n−1cπk(Dk).E(D;\pi_0,\dots,\pi_{n-1})=\sum_{k=0}^{n-1}c_{\pi_k}(D_k).(15)✦

In the paper, the result reads: a=3/2a=3/2, f0f_0 is the piecewise linear cutoff, w(x)=min⁡{xa,x−1}w(x)=\min\{x^{a},x^{-1}\}, p(x)=∫0xw(u)/udup(x)=\int_0^x w(u)/u\,du, p1(x)=∫0xua−1/(1+u)adup_1(x)=\int_0^x u^{a-1}/(1+u)^{a}\,du and φ(x)=min⁡{xa−1,x−1}\varphi(x)=\min\{x^{a-1},x^{-1}\}; with RtR_t and LtL_t the coefficient matrix and the state at time tt and θ\theta a scale, Xt=Lt/θX_t=L_t/\theta, Ft=RtTf0(Xt)RtF_t=R_t^{T}f_0(X_t)R_t and kt=tr⁡φ(Xt)k_t=\operatorname{tr}\varphi(X_t), and Eθ(s,t)E_\theta(s,t), for times ss and tt, is the sum of the trace losses and gains of h=p+p1h=p+p_1. In the library the scale θ\theta is absorbed into the weights qq, and the energy EE carries no scale.

What Is Checked

The energy is stated over the potential trace and the gain trace, which apply pp and p1p_1 to a matrix by the continuous functional calculus, and over moves, through the contracted state CMCTCMC^{T} and the state after the move π(M)\pi(M); the other definitions are stated over real numbers and real matrices indexed by finite sets.