Quantum circuits can be screened for trainability before costly optimization: a representation-theory criterion plus a sampling algorithm decide with polynomially many runs
Synopsis
Gradients of variational quantum circuits can vanish exponentially with qubit count, creating barren plateaus that stall training. A new analysis splits the gradient variance over the invariant subspaces of the unitary representation of the circuit's compact Lie group and gives a sufficient condition for trainability: if the input state and the observable concentrate on the same polynomially sized invariant subspace, and the differentiated generator moves that subspace by an inverse-polynomial amount, the gradient variance has an inverse-polynomial lower bound. A companion sampling algorithm runs the circuit itself, without constructing representation-theoretic projectors, and decides whether the variance exceeds a tolerance with polynomially many runs.
Interpretation
Whether a quantum circuit is trainable is no longer decided by the circuit alone; where the input state and the measurement operator sit among the invariant subspaces also governs whether the gradient vanishes. Earlier theory attributed barren plateaus to the expressivity of the ansatz, or bounded partial-derivative variance by decomposing the Lie algebra the circuit generates, with exact variance formulas typically assuming the observable is generated by that algebra. When the circuit segments on both sides of the differentiated gate are designs, the variance is given exactly; for the standard representation the upper bound has the correct exponent and overestimates the known exact variance by a factor tending to four.
The criterion rests on how the state and observable are distributed across the tensor-square representation, and on how the differentiated generator couples those subspaces. The analysis moves from the Lie algebra to the unitary representation of the compact connected Lie group, separating the variance into contributions from the input state, the observable and the algebraic generators. If the state and observable concentrate on one irreducible invariant subspace of polynomial dimension and the generator moves it by an inverse-polynomial amount, the gradient variance is at least inverse-polynomial; if they occupy different polynomial-dimensional subspaces, the variance can be exactly zero.
The variance can be estimated without building projectors: the algorithm writes it as a central difference quotient, and each run executes the circuit twice. Evaluating the variance analytically requires decomposing the unitary representation and explicitly constructing its projectors, which is intractable at large system size. Each run costs linearly in circuit depth and number of generators; deciding whether the variance exceeds a tolerance takes a number of runs polynomial in qubit count, and a parameter-shift rule removes the finite-difference bias when the generator admits one.
Perspective
The discussion covers parameterized quantum circuits associated with compact connected Lie groups, and the exact variance requires both segments around the differentiated gate to be designs; the sampling algorithm does not rely on this convergence but only decides whether the variance exceeds a given tolerance. The authors state explicitly that the absence of a barren plateau does not imply the system is efficiently classically simulable, and they make no claim of quantum advantage.
The authors note that the analytical bounds are generally not tight, and that an inverse-polynomially large upper bound can coexist with a variance that is exactly zero, so the upper bound alone cannot rule out a barren plateau; the exact formula needs the irreducible decomposition of the representation and is available only for structured circuits. With polynomially many runs the sampling algorithm resolves the variance only to a tolerance, and resolving an exponentially small variance to relative precision still needs exponentially many runs; extension to non-compact Lie groups is not established.
