OctoInformBeta

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

Lemma 4: Finite estimates at every real cutoff and uniform localization for real exponents s<ts<t

Exponential mismatch bounds at depth zero.

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 local layer LL on NN qubits is an operator that is a tensor product of NN single-qubit unitaries. A local mismatch is a component orthogonal to a prescribed one-qubit vector on a selected wire. The depth-zero localization problem concerns the component with many local mismatches created by applying a product projection to a state with no mismatches, with local layers allowed to conjugate the product projection. All norms are Euclidean operator norms.

Fix a total number NN of qubits. A classical bit bb is an element of {0,1}\{0,1\}.✦ A bit string xx of length NN is a function assigning a bit xix_i to each wire i∈{1,…,N}i\in\{1,\ldots,N\}.✦ A quantum operator AA on NN qubits is a complex matrix whose rows and columns are indexed by bit strings of length NN.✦ A single-qubit operator aa is a complex matrix whose rows and columns are indexed by classical bits.✦

The tensor product ⨂i=1Nai\bigotimes_{i=1}^N a_i of single-qubit operators aia_i is the operator with entries given by (1), for bit strings x,yx,y of length NN. (⨂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)✦ For a one-qubit vector vv, the outer product |v⟩⟨v||v\rangle\langle v| is the single-qubit operator defined by (2), with b,c∈{0,1}b,c\in\{0,1\}. (|v⟩⟨v|)b,c=vbvc¯.\bigl(|v\rangle\langle v|\bigr)_{b,c}=v_b\overline{v_c}.(2)✦

Product projection data M=(S,(vi)i=1N)M=(S,(v_i)_{i=1}^N) on NN qubits consist of a support S⊆{1,…,N}S\subseteq\{1,\ldots,N\} and one-qubit vectors vi∈ℂ2v_i\in\mathbb C^2 satisfying the normalization condition (3) on the support. ∑b∈{0,1}|(vi)b|2=1(i∈S).\sum_{b\in\{0,1\}}|(v_i)_b|^2=1\qquad(i\in S).(3)✦ For product projection data MM on NN qubits, the single-qubit operator qiq_i on wire ii is the outer product qi=|vi⟩⟨vi|q_i=|v_i\rangle\langle v_i|. For product projection data MM on NN qubits and a set FF of wire positions, the mismatch pattern projection PFMP_F^M is the NN-qubit operator given by the tensor product in (4), where II is the single-qubit identity operator. PFM=⨂i=1N{I−qi,i∈S∩F,qi,i∈S\F,I,i∉S.P_F^M=\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}(4) The product projection [M=0][M=0] and the high spectral projection [M≥r][M\ge r] of product projection data MM, at a real cutoff rr, are the operators in (5). [M=0]=P⌀M,[M≥r]=∑F⊆S|F|≥rPFM.[M=0]=P_{\varnothing}^M, \qquad [M\ge r]=\sum_{\substack{F\subseteq S\\|F|\ge r}}P_F^M.(5)

A reflection-layer matrix RR on NN qubits is an NN-qubit operator that is a product of reflection gates acting on disjoint sets of qubits. The enlarged reflection-circuit class Ud(N)U_d(N), for a natural depth dd, has U0(N)U_0(N) equal to the set of local-layer matrices and is defined recursively by (6). Ud+1(N)={VRL:V∈Ud(N),R a reflection-layer matrix,L a local-layer matrix}.U_{d+1}(N)=\left\{VRL: \begin{array}{l} V\in U_d(N),\ R\text{ a reflection-layer matrix},\\ L\text{ a local-layer matrix} \end{array}\right\}.(6)✦ Identity layers are allowed, and the rightmost matrix acts first. For NN-qubit operators UU and BB, conjugation of BB by UU is the operator UBU†UBU^\dagger, where †\dagger denotes the adjoint.✦

For a natural depth dd and real exponents s,ts,t, the uniform localization assertion Bd(s,t)B_d(s,t) is the estimate (7) for all sufficiently large NN, simultaneously for 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]∥≤exp⁡(−Ns).\bigl\|[M\ge N^t]U[D=0]U^\dagger[M=0]\bigr\| \le \exp(-N^s).(7)✦

The Claim

Theorem 1. Let NN be any natural number and let M,DM,D be product projection data on NN qubits. For every real cutoff rr, the finite estimates (8) and (9) hold: ∥[M≥r][D=0][M=0]∥2≤2−r.\bigl\|[M\ge r][D=0][M=0]\bigr\|^2\le 2^{-r}.(8)✦ ∥[M≥r][D=0][M=0]∥≤exp⁡(−log⁡22r).\bigl\|[M\ge r][D=0][M=0]\bigr\| \le\exp\!\left(-\frac{\log 2}{2}\,r\right).(9)✦ For real exponents s,ts,t satisfying s<ts<t, the uniform localization assertion (10) holds: B0(s,t).B_0(s,t).(10)✦

In the paper, the result reads: for every fixed 0≤s<t0\le s<t, B0(s,t)B_0(s,t) holds.

What Is Checked

The finite estimates are stated over product projection data on arbitrary NN-qubit spaces with real cutoffs, while the uniform assertion quantifies over every local-layer matrix and all real exponents s<ts<t.