OctoInformBeta

OpenAI preprint, September 2026 · Section 2. Product projections and localization

Theorem 2: Uniform decay between zero mismatch and a growing mismatch cutoff

Product projections localize at fixed circuit depth.

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

A conjugated product projection can connect sectors distinguished by their number of local mismatches. The localization question concerns the transition from no mismatches to a growing mismatch count when the conjugating circuit has fixed depth.

A classical bit belongs to {0,1}\{0,1\}✦, and a word x∈{0,1}Nx\in\{0,1\}^N assigns a bit to each of the NN wires.✦ A quantum operator on NN qubits is a complex matrix whose rows and columns are indexed by {0,1}N\{0,1\}^N✦; a single-qubit operator is a complex matrix whose rows and columns are indexed by {0,1}\{0,1\}.✦ For a family of single-qubit operators AiA_i, the tensor product has entries (⨂i=1NAi)x,y=∏i=1N(Ai)xi,yi.\left(\bigotimes_{i=1}^{N}A_i\right)_{x,y} =\prod_{i=1}^{N}(A_i)_{x_i,y_i}.(1)✦ The norm throughout is the Euclidean operator norm.

For a depth d∈ℕd\in\mathbb N, the enlarged reflection-circuit class Ud(N)U_d(N) consists of the operators obtained recursively from layers LL of single-qubit unitaries and layers RR of reflections by U0(N)={L:L is a local layer on N qubits},Ud+1(N)={VRL:V∈Ud(N),R is a reflection layer,L is a local layer}.\begin{aligned} U_0(N)&=\{L:L\text{ is a local layer on }N\text{ qubits}\},\\ U_{d+1}(N)&=\{VRL:V\in U_d(N),\ R\text{ is a reflection layer},\ L\text{ is a local layer}\}. \end{aligned}(2)✦ In (2), the rightmost matrix acts first.

Product Projections and Mismatch Sectors

For a one-qubit vector v∈ℂ2v\in\mathbb C^2, the outer product |v⟩⟨v||v\rangle\langle v| is the single-qubit operator defined by (3). (|v⟩⟨v|)a,b=v(a)v(b)¯,a,b∈{0,1}.(|v\rangle\langle v|)_{a,b}=v(a)\overline{v(b)}, \qquad a,b\in\{0,1\}.(3)✦ A count MM on NN qubits is product projection data consisting of a support S⊆{1,…,N}S\subseteq\{1,\ldots,N\} and vectors vi∈ℂ2v_i\in\mathbb C^2 for each wire, with ∑b∈{0,1}|vi(b)|2=1(i∈S).\sum_{b\in\{0,1\}}|v_i(b)|^2=1\qquad(i\in S).(4)✦ For a count MM on NN qubits with support SS and vectors viv_i, the zero-count product projection [M=0][M=0] is the following operator, where II is the identity on one qubit: [M=0]=⨂i=1N{|vi⟩⟨vi|,i∈S,I,i∉S.[M=0]=\bigotimes_{i=1}^{N} \begin{cases} |v_i\rangle\langle v_i|,&i\in S,\\ I,&i\notin S. \end{cases}(5)✦

For a count MM on NN qubits with support SS and vectors viv_i, and a set FF of wire indices, the mismatch pattern projection PM,FP_{M,F} is the operator PM,F=⨂i=1N{I−|vi⟩⟨vi|,i∈S∩F,|vi⟩⟨vi|,i∈S\F,I,i∉S.P_{M,F}=\bigotimes_{i=1}^{N} \begin{cases} I-|v_i\rangle\langle v_i|,&i\in S\cap F,\\ |v_i\rangle\langle v_i|,&i\in S\setminus F,\\ I,&i\notin S. \end{cases}(6)✦ The high spectral projection [M≥r][M\ge r] of a count MM on NN qubits, with support SS and real cutoff rr, is the sum of the pattern projections in (6) selected by [M≥r]=∑F⊆S|F|≥rPM,F.[M\ge r]=\sum_{\substack{F\subseteq S\\|F|\ge r}}P_{M,F}.(7)✦

For operators UU and BB on the same qubits, the conjugate Ad⁡U(B)\operatorname{Ad}_U(B) of BB by UU is the operator defined by (8). Ad⁡U(B)=UBU†.\operatorname{Ad}_U(B)=UBU^\dagger.(8)✦

The Claim

At depth dd, with real exponents s,ts,t, let Bd(s,t)B_d(s,t) be the uniform localization assertion: the estimate (9) holds for all sufficiently large NN, uniformly over U∈Ud(N)U\in U_d(N) and counts M,DM,D on NN qubits.✦ The full localization statement requires this assertion at every depth for every pair 0≤s<t0\le s<t.✦

Theorem 1. Fix a depth d∈ℕd\in\mathbb N and real exponents 0≤s<t0\le s<t. There is a threshold N0∈ℕN_0\in\mathbb N such that, for every N≥N0N\ge N_0, every U∈Ud(N)U\in U_d(N), and all counts M,DM,D on NN qubits, ∥[M≥Nt]U[D=0]U†[M=0]∥≤exp⁡(−Ns).\bigl\|[M\ge N^t]\,U[D=0]U^\dagger[M=0]\bigr\| \le \exp(-N^s).(9)✦ For every depth d∈ℕd\in\mathbb N and real exponents 0≤s<t0\le s<t, Bd(s,t).B_d(s,t).(10)✦

What Is Checked

The localization statement (9) ranges over the recursive circuit classes (2) and normalized product projection data (4), using the zero and high projections (5) and (7) and conjugation (8).