Constant-gap brickwork unitary initialization
Implementation & benchmark of arXiv:2607.20919 — Random unitary circuits with constant spectral gap
Source paper: Random unitary circuits with constant spectral gap arXiv:2607.20919 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Initialize a unitary feature-mixing layer with a shallow brickwork circuit of independent random SU(4) gates instead of sampling or factorizing a dense Haar-random unitary. Stack enough layers to obtain a target contraction of non-Haar components, using the paper's constant spectral-gap principle to make the required depth essentially independent of the number of qubits. The resulting layer is local, parameter-efficient, exactly norm-preserving, and should provide Haar-like scrambling at substantially lower initialization and memory cost.
Formulas
Mathematical statement
For a compact group G, a probability measure ν on G, and a finite-dimensional unitary representation ρ:G→U(V), the paper defines the moment operator M(ν,ρ,G)=E_{g∼ν}[ρ(g)]. Haar averaging is Π=M(μ,ρ,G), the orthogonal projector onto the G-invariant subspace of V. The essential norm is g(ν,ρ,G)=||M(ν,ρ,G)−Π||∞ and the spectral gap is Δ(ν,ρ,G)=1−g(ν,ρ,G); the global gap is the infimum over finite-dimensional unitary representations. The paper proves a constant lower bound Δ(ν,G)≥c>0 independent of the qubit count n for brickwork random SU(4) circuits on SU(2^n). Consequently, for k independent brickwork layers, ||M(ν^k,ρ,G)−Π||∞≤(1−c)^k. Here ν^k is the distribution of the product of k random circuit layers, and c is the theorem's dimension-independent gap constant. We exploit this contraction as a practical depth rule: choose k≥log(ε)/log(1−c) to suppress non-Haar representation components below ε. The guarantee concerns group moments, not directly task loss, so covariance, gradient statistics, and downstream accuracy must be measured empirically.
Implementation notes
1. Integration point: use this as a fixed or trainable unitary mixing layer in a unitary RNN, complex-valued MLP, quantum-inspired embedding module, or residual feature mixer. Represent a width d=2^n feature vector as n qubits and implement each circuit layer with two-qubit SU(4) gates on pairs (1,2),(3,4),... followed by gates on (2,3),(4,5),.... Parameterize each SU(4) gate as exp(i∑_a θ_a P_a), where P_a are the 15 non-identity two-qubit Pauli matrices, or initialize a random 4×4 complex matrix and project it to SU(4) by QR decomposition followed by determinant normalization. 2. Pseudocode: sample independent gates G[l,j] in SU(4); set h=x; for l=1,...,k, apply G[l,j] to every pair in the first brickwork matching and then every pair in the shifted matching, setting h←U_l h; optionally train the gate parameters afterward. Use k=ceil(log(ε)/log(1−c)); if the theorem's explicit c is unavailable, sweep k∈{1,2,4,8}. 3. The mathematical input is the contraction prediction (1−c)^k and exact norm preservation ||h||_2=||x||_2. Do not claim to estimate c from one sample: estimate empirical mixing by drawing B circuit products, computing covariance deviation ||B^{-1}∑_b (U_bx_b)(U_bx_b)^*−I/d||_F, or by constructing the empirical second-moment operator on a small width and fitting its decay versus k. 4. First experiment: compare dense Haar-unitary initialization, QR-Gaussian unitary initialization, and brickwork circuits with k=1,2,4,8 in a width-64 complex unitary RNN on sequential MNIST or a copying-memory task. Match parameter count and multiply-add budget where possible. Measure activation covariance anisotropy, gradient-norm variance, loss after a fixed number of updates, final accuracy, peak memory, and wall-clock time. Success means brickwork reaches Haar-like covariance and gradient statistics with lower cost, preserves gradients without exploding or vanishing, and matches or exceeds dense-unitary accuracy at equal wall-clock time.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.