OctoInformBeta

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

Lemma 4.4 [Controlling the accumulated gain]: For energy E≤1/(16ca)E\le 1/(16c_a) and q=(caE)1/aq=(c_aE)^{1/a}, the accumulated gain satisfies V≼q(I+Z)V\preceq q(I+Z) and tr⁡p(X)−tr⁡p(Z)≤8qtr⁡φ(X)\operatorname{tr}p(X)-\operatorname{tr}p(Z)\le 8q\operatorname{tr}\varphi(X)

Small energy bounds the accumulated gain and the potential’s rise.

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 sequence of steps, each of which moves a positive semidefinite matrix: the step contracts it and then adds a positive semidefinite gain. Over a stretch of steps an initial matrix YY becomes a final matrix XX, the sum of two parts: the matrix ZZ that YY becomes under the contractions alone, and the accumulated gain VV, the gains added on the way, each contracted by the steps after it.

The size of such a matrix is measured by the trace of a potential pp applied to it. The question is how far tr⁡p(X)\operatorname{tr}p(X) can exceed tr⁡p(Z)\operatorname{tr}p(Z) when the stretch spends little energy EE, the energy cost defined below.

Moves, Runs and Their Costs

All matrices are real and square, indexed by a finite set ι\iota; II is the identity, A⊤A^{\top} the transpose of AA, and ≼\preceq the Loewner order.

A move π=(C,G)\pi=(C,G) is a pair of such matrices, its contraction CC and its gain GG, and it sends a matrix DD to π(D)=CDC⊤+G\pi(D)=CDC^{\top}+G.✦ A move π=(C,G)\pi=(C,G) is admissible if CC⊤≼ICC^{\top}\preceq I and its gain GG is positive semidefinite.✦

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). From a matrix DD it passes through the states 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),(1) and its final state is DnD_n.

The cost cπ(D)c_\pi(D) of a move π\pi at a matrix DD is the energy charged for applying π\pi to DD, and the energy cost of the run from DD is the sum of the costs cπk(Dk)c_{\pi_k}(D_k) of its moves at the states (1) 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).(2)✦

The gain value g(D)g(D) of a positive semidefinite matrix DD is a real number, the quantity by which the gain part of the energy is measured. The gain cost of the run from DD is the gain part of the energy: the sum of the changes in gg as each gain is added to the contracted state, Egain(D;π0,…,πn−1)=∑k=0n−1(g(πk(Dk))−g(CkDkCk⊤)).E^{\mathrm{gain}}(D;\pi_0,\dots,\pi_{n-1}) =\sum_{k=0}^{n-1}\bigl(g(\pi_k(D_k))-g(C_kD_kC_k^{\top})\bigr).(3)✦

Following the run from a pair (A,B)(A,B) of matrices leads to the pair (An,Bn)(A_n,B_n) given by A0=AA_0=A, B0=BB_0=B and Ak+1=CkAkCk⊤,Bk+1=πk(Bk)(0≤k<n).A_{k+1}=C_kA_kC_k^{\top},\qquad B_{k+1}=\pi_k(B_k)\quad(0\le k<n).(4)✦ For a positive semidefinite matrix YY, following the run from (Y,0)(Y,0) leads to the pair (Z,V)(Z,V).

The Potential

The exponent is the real number a=3/2a=3/2✦, and the gain constant is ca=a2a=32c_a=a\,2^{a}=3\sqrt2✦. The weight ww is the real function given by w(x)=min⁡{xa,x−1}w(x)=\min\{x^{a},x^{-1}\}✦, and the potential pp is the real function p(x)=∫0xw(u)udu.p(x)=\int_0^x\frac{w(u)}{u}\,du.(5)✦ The size weight φ\varphi is the real function given by φ(x)=min⁡{xa−1,x−1}\varphi(x)=\min\{x^{a-1},x^{-1}\}.✦

A real function acts on a symmetric matrix through the functional calculus, on its eigenvalues. The potential trace of a matrix AA is tr⁡p(A)\operatorname{tr}p(A), the trace of the matrix p(A)p(A).✦

The Claim

Theorem 1. (i) Let YY be a positive semidefinite matrix, π0,…,πn−1\pi_0,\dots,\pi_{n-1} a run of admissible moves, and (Z,V)(Z,V) the pair reached by following the run from (Y,0)(Y,0). Let EE be a real number with E≤1/(16ca)E\le 1/(16c_a), and put q=(caE)1/aq=(c_aE)^{1/a}. If the gain cost of the run from YY is at most EE, then 0≤q<14,V≼q(I+Z),V≼q(I+Z+V).0\le q<\tfrac14,\qquad V\preceq q(I+Z),\qquad V\preceq q(I+Z+V).(6)✦

(ii) Let XX and ZZ be positive semidefinite matrices and qq a real number with 0≤q≤1/40\le q\le 1/4 and X−Z≼q(I+X)X-Z\preceq q(I+X). Then tr⁡p(X)−tr⁡p(Z)≤8qtr⁡φ(X).\operatorname{tr}p(X)-\operatorname{tr}p(Z)\le 8q\operatorname{tr}\varphi(X).(7)✦

(iii) Let YY, the run, ZZ, EE and qq be as in (i), and let XX be the final state of the run from YY. If the energy cost of the run from YY is at most EE, then tr⁡p(X)−tr⁡p(Z)≤8qtr⁡φ(X).\operatorname{tr}p(X)-\operatorname{tr}p(Z)\le 8q\operatorname{tr}\varphi(X).(8)✦

In the paper, the result reads: if the energy E=Eθ(s,t)E=E_\theta(s,t) is at most e0=1/(16ca)e_0=1/(16c_a), where ca=a2ac_a=a2^a, and q=(caE)1/aq=(c_aE)^{1/a}, then q<1/4q<1/4, V≼q(I+X)V\preceq q(I+X) and tr⁡(p(X)−p(Z))≤Kqtr⁡φ(X)\operatorname{tr}(p(X)-p(Z))\le Kq\operatorname{tr}\varphi(X) for a constant KK.

What Is Checked

The claim is stated over runs of admissible moves from a positive semidefinite matrix, with the energy cost and the gain cost of a run, its final state and the pair reached by following it, and over the potential pp with its weight ww, the size weight φ\varphi, the potential trace and the constants aa and cac_a.