OctoInformBeta

OpenAI preprint, September 2026 · Section 4. Depth reduction for small-support projections

Lemma 5: Exponentially small transitions at separated cutoffs, with a finite pattern factor

Uniform localization controls low mismatch sectors.

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 product projection transported through a bounded-depth quantum circuit can connect sectors with different numbers of local mismatches. Lemma 5 concerns transitions from a low mismatch sector to a high one, both measured relative to product projection data.

On NN total qubits, a word x∈{0,1}Nx\in\{0,1\}^{N} assigns a classical bit xi∈{0,1}x_i\in\{0,1\} to each wire.✦ A quantum operator is a complex matrix indexed by pairs of words✦; a single-qubit operator is a complex matrix indexed by pairs of classical bits.✦ All operator norms below are Euclidean operator norms.

For a one-qubit vector vv, the outer product |v⟩⟨v||v\rangle\langle v| is the single-qubit operator with entries given by (1). (|v⟩⟨v|)b,c=v(b)v(c)¯,b,c∈{0,1}.\bigl(|v\rangle\langle v|\bigr)_{b,c}=v(b)\overline{v(c)}, \qquad b,c\in\{0,1\}.(1)✦ For single-qubit operators a1,…,aNa_1,\ldots,a_N, the tensor product ⨂i=1Nai\bigotimes_{i=1}^{N}a_i is the quantum operator whose entries are specified by (2). (⨂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}.(2)✦

Mismatch Sectors and Circuit Localization

Product projection data MM on NN qubits consist of a support S⊆{1,…,N}S\subseteq\{1,\ldots,N\} and local vectors vi∈ℂ2v_i\in\mathbb C^2, normalized on SS: ∑b∈{0,1}|vi(b)|2=1\sum_{b\in\{0,1\}}|v_i(b)|^2=1 for every i∈Si\in S.✦ Writing Qi=|vi⟩⟨vi|Q_i=|v_i\rangle\langle v_i| and denoting the single-qubit identity operator by II, the product projection [M=0][M=0] and the mismatch pattern projection PFP_F for a set of wires FF are defined in (3). [M=0]=⨂i=1N{Qi,i∈S,I,i∉S,PF=⨂i=1N{I−Qi,i∈S∩F,Qi,i∈S\F,I,i∉S.\begin{aligned} [M=0]&=\bigotimes_{i=1}^{N} \begin{cases}Q_i,&i\in S,\\I,&i\notin S,\end{cases}\\ P_F&=\bigotimes_{i=1}^{N} \begin{cases} I-Q_i,&i\in S\cap F,\\ Q_i,&i\in S\setminus F,\\ I,&i\notin S. \end{cases} \end{aligned}(3) For product projection data MM on NN qubits and a real cutoff rr, the high spectral projection [M≥r][M\ge r] is the sum of the mismatch pattern projections with at least rr mismatches, and the low spectral projection [M<r][M<r] is its complement, as in (4). [M≥r]=∑F⊆S|F|≥rPF,[M<r]=I−[M≥r].[M\ge r]=\sum_{\substack{F\subseteq S\\|F|\ge r}}P_F, \qquad [M<r]=I-[M\ge r].(4)

For a nonnegative integer dd, the enlarged reflection-circuit class Ud(N)U_d(N) on NN qubits is the class of operators of depth at most dd defined recursively from local layers and reflection layers. At depth zero, its members are local-layer matrices. At depth d+1d+1, membership means that U=VRLU=VRL, where V∈Ud(N)V\in U_d(N), RR is a reflection-layer matrix and LL is a local-layer matrix. The rightmost matrix acts first, and identity layers are allowed.

For operators BB and UU on NN qubits, the conjugate of BB by UU is the operator UBU†UBU^\dagger, where †\dagger denotes the adjoint.✦ For real exponents s,ts,t, the uniform localization assertion Bd(s,t)B_d(s,t) means that there is a threshold N0N_0, depending only on d,s,td,s,t, such that (5) holds for every N≥N0N\ge N_0, every U∈Ud(N)U\in U_d(N) and all product projection data M,DM,D on NN qubits. ∥[M≥Nt]U[D=0]U†[M=0]∥op≤exp⁡(−Ns).\bigl\|[M\ge N^t]\,U[D=0]U^\dagger[M=0]\bigr\|_{\mathrm{op}} \le \exp(-N^s).(5)✦

The Claim

Theorem 1 (Lemma 5). Let N≥1N\ge1, let MM be product projection data on NN qubits, and let BB be an operator on those qubits. Take real numbers r,q,Rr,q,R and δ≥0\delta\ge0 with q+r≤Rq+r\le R. Suppose that ∥[K≥q]B[K=0]∥op≤δ\|[K\ge q]B[K=0]\|_{\mathrm{op}}\le\delta for every product projection datum KK on NN qubits. The finite propagation bound is (6), where ⌈r⌉+\lceil r\rceil_+ is the least nonnegative integer at least rr. ∥[M≥R]B[M<r]∥op≤(N+1)⌈r⌉++1δ.\bigl\|[M\ge R]B[M<r]\bigr\|_{\mathrm{op}} \le (N+1)^{\lceil r\rceil_++1}\delta.(6)✦

Fix a natural depth dd and assume Bd(s,t)B_d(s,t) for every pair of real exponents 0≤s<t0\le s<t. Let α,σ≥0\alpha,\sigma\ge0 and β>max⁡(α,σ)\beta>\max(\alpha,\sigma). Then, for all sufficiently large NN, uniformly over every U∈Ud(N)U\in U_d(N) and all product projection data M,DM,D on NN qubits, the low propagation estimate is (7). ∥[M≥Nβ]U[D=0]U†[M<Nα]∥op≤exp⁡(−Nσ).\bigl\|[M\ge N^\beta]\,U[D=0]U^\dagger[M<N^\alpha]\bigr\|_{\mathrm{op}} \le \exp(-N^\sigma).(7)✦

What Is Checked

The finite estimate is stated over arbitrary operators and product projection data on N≥1N\ge1 qubits; the asymptotic estimate is stated over the enlarged reflection-circuit class under the family of uniform localization assertions at depth dd.