arXiv:2610.01446
Unflattening by Flattening -- How Input Distributions Shape Output Variance in Angle-Encoded Circuits
Melvin Strobl, Gabriel Mejia, Eileen Kuehn, Achim Streit
Abstract
Barren plateaus hinder training of parameterized quantum circuits by making loss gradients exponentially small in the number of qubits. For angle-encoded product states, we show how the input distribution affects output variation through algebraic input purity, i.e. the input's overlap with the circuit's dynamical Lie algebra. With the specified readouts and Haar or exact group 2-design sampling, matchgate circuits on $n$ qubits retain output variance of order $1/n$ for every pure product input. The off-diagonal family instead has zero output on computational-basis inputs at any depth and for every parameter choice. Independent uniform angles yield mean variance of the same order as matchgates. A count of diagonal Pauli strings identifies these zero-output endpoints. For any fixed dataset of nonzero inputs with at least $2n$ coordinates, we construct a classical preprocessing certificate. A shared random rotation is accepted only when the dataset's mean purity passes a computable threshold. This certifies mean output variance of order $1/n$ for the off-diagonal family before circuit execution, with a constant expected number of trials. Binary and ternary weighted encodings also recover the independent-angle mean purity from one uniform scalar input. Numerical experiments examine training behavior and extensions to larger algebras. The guarantee concerns output variance averaged over inputs, while successful learning also depends on the task and information preserved by the encoding.
AI Summary
The paper asks how classical input distributions affect parameter-dependent output variance in angle-encoded quantum circuits. It applies the existing Lie-algebraic variance formula to product inputs, deriving algebraic purities analytically and checking them with Pauli-basis calculations and finite-size circuit simulations. For the matchgate algebra $\mathfrak{so}(2n)$, every pure product input has Pauli-basis purity $(n-1)+\prod_k\langle Z_k\rangle^2\in[n-1,n]$. With a $Z_i$ readout and Haar or exact group 2-design sampling, output variance is therefore $\Theta(1/n)$ for every input. For the off-diagonal chain, with algebra $\mathfrak{so}(n)\oplus\mathfrak{so}(n)$, computational-basis inputs instead produce exactly zero in-algebra output at every depth and parameter setting. Counting diagonal Pauli strings identifies this endpoint. Independent uniform encoding angles give mean purity $\mu_n=n/3-5/9+O(2^{-n})$ and mean output variance $\Theta(1/n)$ for the specified readout. A classical preprocessing certificate applies a shared random rotation to any fixed dataset of nonzero vectors in dimension $D\ge2n$, then extracts polar angles. Accepting rotations whose empirical mean purity exceeds $\mu_n/2$ certifies inverse-polynomial mean output variance; the expected-trial bound approaches five. Dissociated integer scalar weights, including binary and ternary weights, exactly reproduce independent-angle mean purity under a uniform scalar input. These guarantees concern mean sample-wise output variance, not supervised-loss gradients or learning success. Polar-angle extraction can discard task-relevant information; finite-dataset certification gives no population guarantee. Mixing depths remain unresolved, and numerical training and exponential-algebra extensions provide only finite-size or conditional evidence.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.