Escaping from the Barren Plateau via Gaussian Initializations in Deep Variational Quantum Circuits

1 Introduction

• Barren plateaus involve exponentially small gradient variance, and measurement and machine noise can degrade gradient estimates and slow optimization on flat loss surfaces.The introduction motivates resolving this problem as important for practical quantum advantages with VQAs.Source: §1 · p. 1
• The paper proposes Gaussian initialization with certain variances to lower-bound expected gradient norms and considers local and global observables, independent and joint parameters, and finite-measurement noise.The contribution list further reports experiments on Heisenberg-model and LiH ground-state tasks, where Gaussian initializations outperform uniform initializations.Source: §1 · p. 2
• For N-qubit, L-depth circuits with independent parameters and local observables, variance γ2 = O(1/L) gives an expected gradient-norm lower bound poly(N, L)^−1, improving on prior 2^−O(L) results.The paper also extends the bound to global observables and correlated parameterized gates, and analyzes measurement requirements for estimating gradients.Source: §1 · p. 2
• Global observables were believed to exhibit barren plateaus even in very shallow circuits, while the paper extends its gradient result to global observables and correlated parameterized gates.The latter are described as widely used in quantum chemistry and quantum simulations.Source: §1 · p. 2

1.1 Related work

• Deep circuits were generally believed to suffer barren plateaus because random circuits approach approximate 2-designs at linear depth, while positive results cited here concern shallow circuits.Prior work also linked barren plateaus to volume-law states and circuit expressivity.Source: §1.1 · p. 2
• Earlier approaches include block-identity initialization, layerwise training, and a structural guarantee limited to local observables; the paper notes practical drawbacks or scope limits for each.Block identity may fail after the first step and requires circuit design; layerwise training can use greater depth and leave early layers unoptimized.Source: §1.1 · p. 3

2 Notations and quantum computing basics

• The notation defines [N] = {1, · · ·, N}, ∥·∥2 as the vector ℓ2 norm or matrix spectral norm, A† as conjugate transpose, and Tr[A] as trace.It also uses ∇θf for the gradient with respect to θ and O for complexity notation.Source: §2 · p. 3
• An N-qubit space is a tensor product of single-qubit spaces; pure states have density matrices ρ = |ϕ⟩⟨ϕ|, while mixed states are weighted sums of such matrices.A normalized single-qubit pure state is written |ϕ⟩ = a|0⟩ + b|1⟩.Source: §2 · p. 3
• Single-qubit operations are represented as gates, and two-qubit gates such as CZ and iSWAP generate quantum entanglement.The section names Pauli matrices {I, X, Y, Z} as {σ0, σ1, σ2, σ3}.Source: §2 · p. 3
• Measurement can be expressed as an observable expectation: measuring |ϕ⟩ with O = σ3 gives probabilities |a|2 and |b|2, and the expectation equals p(0) − p(1).O(1/ϵ2) measurements provide an ϵ∥O∥2-error estimate of Tr[Oρ].Source: §2 · p. 3

3 Framework of general VQAs

• A VQA optimizes a loss defined as the expectation of an observable under a parameterized circuit acting on an input density matrix.The circuit, observable, and input density matrix are denoted V(θ), O, and ρin, respectively.Source: §3 · pp. 3–4
• An observable is local when its Pauli-basis terms have constant locality, and global when it acts on all qubits.Locality is the maximum number of non-identity Pauli matrices in a tensor product with a nonzero coefficient.Source: §3 · p. 4
• Quantum simulation and chemistry commonly use system Hamiltonians as observables and computational-basis or Hartree–Fock input states, whereas QML inputs encode training data and often use simple observables.The passage gives |0⟩⟨0| as a frequent simulation input and π0 = |0⟩⟨0| as a common QML classification or dimensional-reduction observable.Source: §3 · p. 4
• An ansatz is the initial guess for circuit structure; examples include alternating-layered, hardware-efficient, tensor-network-based, and unitary coupled cluster ansatzes.The gradient satisfies the parameter-shift rule, enabling efficient gradient estimation for gradient-based optimization.Source: §3 · p. 4

4 Theoretical results about Gaussian initialized VQAs

• Gaussian initialization with variance O(1/L) gives a polynomial gradient-norm bound for L-layer, N-qubit circuits, improving on previously reported exponential scaling.The section treats local and global observables separately.Source: §4 · p. 4

4.1 Independent parameters with local observables

• For local observables, the analyzed ansatz uses L layers of single-qubit rotations and CZ gates, followed by RX and RY layers, with measurement on S qubits.The first L layers use gates whose generators do not commute with CZ, to eliminate degenerate parameters.Source: §4.1 · p. 4, Figure 1 · p. 5
• The larger gradient-norm lower bound in Theorem 4.1 improves the complexity exponentially with circuit depth, but large initial gradients alone do not guarantee trainability throughout optimization.The discussion notes that large gradients for trained parameters can correspond to non-convergence.Source: §4.1 · p. 5

4.2 Correlated parameters with global observables

• For global observables and correlated parameters, Theorem 4.2 gives a gradient bound with O(1/L) variance when the gradient at zero is large, a condition associated with a good mean-field initial guess in some problems.The cited examples include ground-energy tasks in quantum chemistry and quantum many-body problems.Source: §4.2 · p. 6
• Corollary 4.3 bounds the gradient at a noisy updated point when measurement noise satisfies its stated condition, connecting the analysis to the number of measurements needed.The accompanying discussion describes the noisy update as differing from the noiseless update by a Gaussian error term.Source: §4.2 · pp. 6–7

5 Experiments

• The experiments use Pennylane to train VQAs for ground-state and ground-energy tasks on the Heisenberg model and LiH molecule.Source: §5 · p. 7

5.1 Heisenberg model

• Gaussian initialization converges faster and yields larger early gradients than uniform initialization on the Heisenberg task.Zero initialization has zero initial gradient with accurate gradient descent, but measurement noise alleviates this issue.Source: §5.1 · p. 8

5.2 Quantum chemistry

• Adaptive noisy gradients match accurate-gradient convergence and final loss for LiH, while constant-variance noise performs worse.Gaussian and zero initialization outperform uniform initialization in all tested cases.Source: §5.2 · p. 9, Figure 3 · p. 9

• With independent Gaussian parameters of variance O(1/L), the gradient norm is lower bounded by poly(N, L)^-1 for N-qubit, L-layer circuits.Source: §6 · p. 9
• The bounds cover local and global observables and extend to VQAs with correlated parameterized gates.Source: §6 · p. 9
• The global-observable bound depends on gradient performance at the zero point; removing that dependency is left for future work.Source: §6 · p. 9
• The paper reports that O(L/ϵ) measurements suffice for gradient estimation and that Gaussian initialization outperforms uniform initialization in experiments.Adaptive-noise training shows convergence comparable to noiseless-gradient training.Source: §6 · p. 9

• The authors report stating the full assumptions of theoretical results and including complete proofs.Source: Checklist · p. 15
• The checklist reports that training details are specified and experiments were independently run five times.Source: Checklist · p. 15
• The checklist reports that total compute and resource type were not provided, though experiments finished on CPUs within a few hours.Source: Checklist · p. 15

A.1 Heisenberg model

• The variance γ outperforms other tested Gaussian variances; larger variances need more iterations as depth grows.The tested smaller variances 0.01γ and 0.1γ, and γ, show similar convergence rates across depths.Source: §A.1 · p. 17
• Momentum and Nesterov accelerated gradient perform similarly to Adam, while AdaGrad performs similarly to gradient descent.Uniformly initialized circuits converge more slowly as qubit number and circuit depth increase.Source: §A.1 · p. 17

A.2 Quantum chemistry

• With adaptive noise, LiH gradient-descent convergence speeds up as depth increases, while Adam shows similar speeds across depths.The tested depths are L ∈ {24, 48, 72}.Source: §A.2 · p. 17

B Technical Lemmas

• The Gaussian-parameter lemma treats a Hermitian observable that anticommutes with a Hermitian-unitary generator.It considers θ distributed as N(0, γ²).Source: B Technical Lemmas · p. 18
• Lemma B.4 considers a density-matrix state with Gaussian variance γ² ≤ 1/(16h³).Its setup uses a circuit composed of unitary matrices and parameterized Hermitian-unitary gates.Source: B Technical Lemmas · p. 26
• The proof concludes that Eq. (77) holds after deriving Eq. (118) by calculating geometric sequences under γ² ≤ 1/(16h³).Source: B Technical Lemmas · p. 32

C Proof of Theorem 4.1

• The proof lower-bounds the whole gradient norm using the gradient norm for the observable’s nontrivial qubits.It denotes those qubits by IS and propagates bounds across parameter layers using the lemmas.Source: C Proof of Theorem 4.1 · pp. 32–34
• The proof’s main idea for Eq. (120) is to integrate the square of the partial derivative of f with respect to θ = (θ1, …, θL+2), using Lemmas B.1 and B.2.Source: C Proof of Theorem 4.1 · p. 33
• The derivation concludes by establishing Eq. (120) and using it to obtain the theorem’s formulation.The proof applies for each layer index q and each n in IS.Source: C Proof of Theorem 4.1 · pp. 33–34

D Proof of Theorem 4.2

• The proof defines O_j as O conjugated by the zero-parameter circuit segment from layer j through layer L.This layer-indexed observable is used in the subsequent derivation.Source: D Proof of Theorem 4.2 · p. 35
• The proof reduces the case a_j ≠ 1 to a_j = 1 by transforming the parameters and relating their variances.It then writes partial derivatives as linear sums of 2h expectations using the parameter-shift rule.Source: D Proof of Theorem 4.2 · p. 35
• The derivation repeatedly applies Lemma B.3 while proceeding through earlier parameters and circuit layers.The resulting inequalities are combined to complete the proof.Source: D Proof of Theorem 4.2 · pp. 35–36