OctoInformBeta

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

Lemma 6: Preceding-depth localization controls an additional reflection layer above the support and decay exponents

Small-support projections localize at the next 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

Conjugation by a shallow circuit is measured here through transitions from a product projection’s zero-mismatch subspace to sectors with many mismatches. Lemma 6 concerns such transitions when the conjugated product projection has a restricted support.

For NN total qubits, a classical bit belongs to {0,1}\{0,1\}✦, and a bit string xx of length NN assigns a classical bit xix_i to each wire.✦ A quantum operator on these qubits is a complex matrix indexed by pairs of bit strings✦; a single-qubit operator has its rows and columns indexed by classical bits.✦ For a family aa of single-qubit operators aia_i and bit strings x,yx,y of length NN, the tensor product T(a)T(a) has the entries in (1).✦ For quantum operators U,BU,B, conjugation Ad⁡U(B)\operatorname{Ad}_U(B) is given by (1), where U†U^\dagger denotes the adjoint.✦ All norms are Euclidean operator norms. T(a)x,y=∏i=1N(ai)xi,yi,Ad⁡U(B)=UBU†.T(a)_{x,y}=\prod_{i=1}^{N}(a_i)_{x_i,y_i}, \qquad \operatorname{Ad}_U(B)=UBU^\dagger.(1)

Product projection data MM on NN qubits consist of a support SM⊆{1,…,N}S_M\subseteq\{1,\ldots,N\} and local vectors vi∈ℂ2v_i\in\mathbb C^2, normalized on SMS_M as in (2).✦ The outer product qi=|vi⟩⟨vi|q_i=|v_i\rangle\langle v_i| is the single-qubit operator with the matrix entries in (2); for i∈SMi\in S_M, it is the rank-one projector onto viv_i. ∑b=01|vi(b)|2=1(i∈SM),qi=|vi⟩⟨vi|,(qi)b,c=vi(b)vi(c)¯.\sum_{b=0}^{1}|v_i(b)|^2=1\quad(i\in S_M), \qquad q_i=|v_i\rangle\langle v_i|, \qquad (q_i)_{b,c}=v_i(b)\overline{v_i(c)}.(2)

The mismatch pattern projection PFMP_F^M, for a set FF of wire indices, uses the complementary local projector on the supported wires in FF, as specified in (3).✦ The product projection [M=0][M=0] selects the empty mismatch pattern.✦ PFM=⨂i=1N{I−qi,i∈SM∩F,qi,i∈SM\F,I,i∉SM,[M=0]=P⌀M.P_F^M=\bigotimes_{i=1}^{N} \begin{cases} I-q_i,&i\in S_M\cap F,\\ q_i,&i\in S_M\setminus F,\\ I,&i\notin S_M, \end{cases} \qquad [M=0]=P_{\varnothing}^M.(3)

For a real cutoff rr, the high spectral projection [M≥r][M\ge r] sums the mismatch patterns of cardinality at least rr✦, and the low spectral projection [M<r][M<r] is its complement, as in (4).✦ [M≥r]=∑F⊆SM|F|≥rPFM,[M<r]=I−[M≥r].[M\ge r]=\sum_{\substack{F\subseteq S_M\\ |F|\ge r}}P_F^M, \qquad [M<r]=I-[M\ge r].(4)

Circuit Layers and Localization

A local unitary layer L=(Li)i=1NL=(L_i)_{i=1}^{N} assigns a one-qubit unitary to each wire.✦ A disjoint reflection layer R=(A1,…,Am)R=(A_1,\ldots,A_m) is an ordered list of product projection data with pairwise disjoint supports.✦ Their layer matrices are those in (5); the rightmost matrix acts first. ℒ(L)=⨂i=1NLi,ℛ(R)=∏j=1m(I−2[Aj=0]).\mathcal L(L)=\bigotimes_{i=1}^{N}L_i, \qquad \mathcal R(R)=\prod_{j=1}^{m}\bigl(I-2[A_j=0]\bigr).(5)

For a nonnegative integer dd, the enlarged reflection-circuit class Ud(N)U_d(N) consists of operators of depth at most dd, with identity layers allowed. Its recursive definition is (6), where kk ranges over nonnegative integers, LL over local unitary layers and RR over disjoint reflection layers. U0(N)={ℒ(L):L a local unitary layer},Uk+1(N)={Vℛ(R)ℒ(L):V∈Uk(N),R,L as above}.\begin{aligned} U_0(N)&=\{\mathcal L(L):L\text{ a local unitary layer}\},\\ U_{k+1}(N)&= \{V\mathcal R(R)\mathcal L(L):V\in U_k(N),\ R,L\text{ as above}\}. \end{aligned}(6)✦

Uniform localization Bd(s,t)B_d(s,t), for real exponents s,ts,t, is the assertion that (7) holds for all sufficiently large NN, simultaneously for all U∈Ud(N)U\in U_d(N) and all product projection data M,DM,D.✦ ∥[M≥Nt]Ad⁡U([D=0])[M=0]∥≤exp⁡(−Ns).\bigl\|[M\ge N^t]\operatorname{Ad}_U([D=0])[M=0]\bigr\| \le \exp(-N^s).(7)

For an ordered list GG of mm product projection data, its reflection product expansion ℰ(G)\mathcal E(G) is a list of pairs (c,D)(c,D) with real coefficient cc and product projection data DD. The reflection product expansion of the empty gate list is the single-term list (1,E)(1,E), where EE is product projection data with empty support. Its coefficient list is indexed by selections JJ of gate positions as in (8); for disjoint gate supports, the term indexed by JJ uses the selected gates’ local vectors on the union of their supports. cJ=(−2)|J|,J⊆{1,…,m}.c_J=(-2)^{|J|},\qquad J\subseteq\{1,\ldots,m\}.(8)

The Claim

Theorem 1. Let dd be a nonnegative integer, and assume Bd(s,q)B_d(s,q) for every pair of real exponents 0≤s<q0\le s<q. Let u,w≥0u,w\ge0 and t>max⁡{u,w}t>\max\{u,w\}. For all sufficiently large NN, every U∈Ud+1(N)U\in U_{d+1}(N) and all product projection data M,AM,A with |SA|≤Nu|S_A|\le N^u satisfy (9).✦ ∥[M≥Nt]Ad⁡U([A=0])[M=0]∥≤exp⁡(−Nw).\bigl\|[M\ge N^t]\operatorname{Ad}_U([A=0])[M=0]\bigr\| \le \exp(-N^w).(9)✦

Let VV be a unitary operator on NN qubits, RR a disjoint reflection layer, LL a local unitary layer, and M,AM,A product projection data. Let r,r1,r2r,r_1,r_2 be real cutoffs and δ≥0\delta\ge0. Suppose that every product projection datum DD satisfies the three estimates in (10). ∥[M≥r1]Ad⁡V([D=0])[M=0]∥≤δ,∥[M≥r2]Ad⁡V([D=0])[M<r1]∥≤δ,∥[M≥r]Ad⁡V([D=0])[M<r2]∥≤δ.\begin{aligned} \bigl\|[M\ge r_1]\operatorname{Ad}_V([D=0])[M=0]\bigr\|&\le\delta,\\ \bigl\|[M\ge r_2]\operatorname{Ad}_V([D=0])[M<r_1]\bigr\|&\le\delta,\\ \bigl\|[M\ge r]\operatorname{Ad}_V([D=0])[M<r_2]\bigr\|&\le\delta. \end{aligned}(10) Then the finite layer estimate is (11). ∥[M≥r]Ad⁡Vℛ(R)ℒ(L)([A=0])[M=0]∥≤9|SA|(3δ).\bigl\|[M\ge r] \operatorname{Ad}_{V\mathcal R(R)\mathcal L(L)}([A=0])[M=0]\bigr\| \le 9^{|S_A|}(3\delta).(11)✦

For every ordered list GG of mm product projection data, the absolute coefficient weight of its reflection product expansion is (12), with the sum taken over list entries.✦ ∑(c,D)∈ℰ(G)|c|=3m.\sum_{(c,D)\in\mathcal E(G)}|c|=3^m.(12)✦

What Is Checked

The statements are over finite quantum operators, normalized product projection data and layers with disjoint reflection supports, with preceding-depth uniform localization assumed in (9) and the three transition estimates (10) assumed in (11).