The scalability of variational quantum algorithms is limited by barren plateaus, where gradients vanish exponentially in system size. This worst-case result raises fundamental questions about the trainability of quantum AI models. We present a major advance towards resolving this trainability question by answering whether a given quantum circuit has a barren plateau. We do this using an analytical framework that bounds the gradient variance for arbitrary input states ρ and observables M by decomposing the unitary representation of the associated compact Lie group. When the circuit segments on both sides of a gate form 2-designs, we obtain the variance exactly and prove a sufficient condition for trainability: if ρ⊗2 and M⊗2 are sufficiently concentrated on the same irreducible invariant subspace of polynomial dimension, and the differentiated gate moves this subspace by an inverse-polynomial amount, then the gradient variance is at least inverse-polynomial. We further present a sampling algorithm that estimates the gradient variance from runs of the ansatz, without constructing the representation-theoretic projectors. Our algorithm decides whether the variance exceeds a tolerance τ with the number of runs polynomial in 1/τ, including cases in which the bounds are inconclusive. This resolves the question of the trainability of the circuit before expensive training procedures.
Figures & tables
Figure 1: An illustrative example of the conditions for a non-vanishing upper bound. The tensor product space H⊗2 decomposes into five irreducible representations Vj , where Vj,k denotes the k -th isomorphic copy of Vj ; the dimensions of V1 and V3 are assumed to scale polynomially with system size. The bars encode the normalized magnitude of the inner product ⟨∑a=1mλ∗[…],⋅⟩ , with deep red indicating strictly dominant values and blue indicating negligible ones.
Figure 2: Conditions for a non-vanishing lower bound under the concentrated support assumptions. The supports of ρ⊗2 and M⊗2 are concentrated on a common polynomial-dimensional sector mλ∗Vλ∗ , so the isolated contribution dominates and the cross-sector interference is negligible.
Figure 3: Exact variance (solid line), lower bound of Equation 12 (squares), upper bound of Equation 11 (circles) and trivial bound 4(∥H∥∞S)2(∥M∥∞S)2 (dashed) on a logarithmic scale, for 2⩽n⩽20 . Both bounds are evaluated after compression to the irreducible sub-representation that carries ρ ( Remark 2.3 (ii)); the formulas below are derived in Supplementary Sections S8.3 and S8.4.1 . (a) Free-fermion example with ρ=∣0⊗n⟩⟨0⊗n∣ , M=Z(1)=−iγ1γ2 and H=21γ1γ2 , so that ∥H∥∞S=21 . The exact variance is n(2n−1)24(n−1) . The lower bound matches it for n⩾8 , is n(2n−1)2(n−1)(n−4) for n<8 , and is not positive for n⩽4 . The upper bound, evaluated on the even-parity sector, is 2n−11[1+n2∑j=1⌊n/2⌋8j2−n(n−2j2n)/(n2n)] ; on the full Fock space, the same bound is larger ( 1.54 instead of 0.44 at n=4 ). Variance and upper bound decay polynomially, as n−2 and n−1/2 , respectively. (b) SU(2) collective-spin example with ρ=∣λ∗⟩⟨λ∗∣ , λ∗=n/2 , M=SX+SY+SZ , and H=−iSZ , so that ∥H∥∞S=n/2 . The exact variance is 32(λ∗)2=6n2 . The lower bound equals it for n⩾4 , is 43 at n=3 and vanishes at n=2 . The upper bound, evaluated on the spin- n/2 sector, is 22n+1n2∑λ=0n2λ+1λ(λ+1)−2n(n+2) ; it lies above the trivial bound 43n4 for all n⩾5 .
Figure 4: Variance of ∂θℓ,1C for the SU(n) compound ansatz, n=4,…,20 . Setting. G=SU(n) acts on H=(C2)⊗n≅⨁m=0nΛm(Cn) through the compound representation Π(V)=⨁m=0nCm(V) , where the m -th compound matrix Cm(V) acts on the Hamming-weight- m subspace; equivalently π∗(X)=∑a,bXabaa†ab with Jordan–Wigner operators aa=Z1⋯Za−1∣0⟩⟨1∣a . The ansatz ( 4 ) is U(θ)=∏l=1L∏j=1n2−1exp(θl,jπ∗(gj)) with θl,j∼Unif[−π,π] and the basis {gj}={−i(Epq+Eqp),Epq−Eqp}p<q∪{i(Err−Er+1,r+1)}r=1n−1 of su(n) , i.e. the gates are generated by −2i(XpZpqXq+YpZpqYq) , −2i(YpZpqXq−XpZpqYq) and −2i(Zr−Zr+1) , where Zpq=∏p<k<qZk . We take ρ=∣+⟩⟨+∣⊗n , M=X1X2 , and the generator of the differentiated gate H=π∗(g1)=−2i(X1X2+Y1Y2) ( ∥H∥∞=∥M∥∞=1 ), with U± Haar-distributed on Π(SU(n)) (Case 3). Curves. Solid line: exact variance Var(n)=n2(n−1)(n+1)22n3−3n−2≈2/n2 of Equation 21 , derived in Supplementary Section S8.5 . Dashed line: its first term (n−1)(n+1)2n , the contribution of the component of M inside the algebra, which is all that the Lie-algebraic formula accounts for. Squares: Case-3 lower bound Equation 12 . Circles: output of Algorithm 1 in its exact form, with the four-term shift rule for H ( γ=2 ), δ=10−3 and ε=0.999/n2 , using N=⌈32ln(2/δ)/ε2⌉ runs ( 6.2×104 at n=4 to 3.9×107 at n=20 ; the finite-difference form needs 4.6×108 to 1.8×1014 runs for the same guarantee). Every run was simulated, with the value of the cost function at each shift computed in a time polynomial in n as described in Supplementary Section S8.5 . Error bars: the certified accuracy ±ε=±0.999/n2 , guaranteed with probability at least 1−δ=0.999 .
Whether parameterized quantum circuits (PQCs) can be systematically constructed to be both trainable and expressive remains an open question. Highly expressive PQCs often exhibit barren plateaus, while several trainable alternatives admit efficient classical simulation. We address this question by deriving a finite-sample, dimension-independent concentration bound for estimating the variance of a PQC cost function, yielding explicit trainability guarantees. Across commonly used ansätze, we observe an anticorrelation between trainability and expressibility, consistent with theoretical insights. Building on this observation, we propose a property-based ansatz-search framework for identifying circuits that combine trainability and expressibility. We demonstrate its practical viability on a real quantum computer and apply it to variational quantum algorithms. We identify quantum neural network ansätze with improved effective dimension using over 6× fewer parameters, and for VQE on H2 we achieve UCCSD-like accuracy at substantially reduced circuit complexity.
Peter Röseler, Dennis Willsch, Kristel Michielsen
Jülich Supercomputing Centre, Forschungszentrum Jülich, 52425 Jülich, Germany · Universität zu Köln, 50923 Köln, Germany · Faculty of Medical Engineering and Technomathematics, University of Applied Sciences Aachen, 52428 Jülich, Germany +1
Variational quantum circuits have been central to many proposed near-term applications of quantum computing, but a growing body of evidence suggests that trainability and quantum advantage are fundamentally at odds: ansätze expressive enough to resist efficient classical simulation tend to exhibit barren plateaus, while structures that provably rule out barren plateaus typically render them classically simulable. We propose a stacked linear combination of unitaries (S-LCU) as a variational ansatz which provides a tunable trade-off between barren plateaus and classical simulability. Using a diagrammatic analysis, we bound the loss-landscape variance of the Free Fermion S-LCU, whose elements are fermionic Gaussian unitaries. We prove a variance lower bound of Ω(1/(nk3l)), with a simulation cost of O(k2ln3) using the best known classical algorithm, compared to a quantum gate complexity of only O(lkn2). The number of layers l serves as a single dial that trades computational complexity against the rate of cost concentration. This offers practitioners a systematic method for constructing ansätze with a complexity-trainability trade-off that best suits their application and hardware.
Nikhil Khatri, Stefan Zohren, Gabriel Matos
Machine Learning Research Group, University of Oxford, United Kingdom · 2Quantinuum, London, United Kingdom
Distributed quantum algorithms offer a promising pathway to scale variational quantum algorithms beyond the constraints of noisy intermediate-scale quantum hardware. However, existing approaches implicitly assume a trusted entanglement-sharing layer across quantum processors. We show that this assumption introduces a fundamental vulnerability: adversarial perturbations of shared entanglement induce structured gate-level noise that directly impacts quantum learning. We develop a framework that maps entanglement-level perturbations to gate-level noise via an explicit Kraus representation. To quantify their impact, we introduce Kraus expressibility, a metric that generalizes unitary expressibility to noisy quantum channels. We then establish a trade-off between Kraus expressibility and trainability of noisy quantum circuits through gradient variance analysis. Our analysis reveals that an adversary can manipulate Kraus expressibility to maintain sufficiently large cost gradients (avoiding barren plateaus) while systematically biasing optimization toward incorrect solutions. We validate these findings through numerical simulations, demonstrating adversarial degradation of expressibility and trainability.
Abhishek Sadhu, Sharu Theresa Jose
School of Computer Science, University of Birmingham · Edgbaston, Birmingham B15 2TT, United Kingdom