Variational quantum algorithms train a small parameterized quantum circuit with a classical optimizer. They stall on barren plateaus: as circuits get wider or deeper, the gradient vanishes exponentially, so no affordable number of measurements can tell the optimizer which way to go. The known cures (shallow circuits, local cost functions, restricted entanglement) share a catch, proved in Does provable absence of barren plateaus imply classical simulability?. They keep the quantum state in a small corner of the exponentially large state space, and an ordinary computer can simulate that corner. Trainable circuits so far have been classically simulable ones.
We propose to exploit this catch rather than avoid it. First, train a shallow circuit almost entirely on a classical computer from one batch of classical shadows, which are compact classical records built from randomized single-qubit measurements that predict many local expectation values at once. Then use the same shadows to decide which entangling gates to add. Hand training back to the quantum device only when the circuit's measured observables stop being local enough to simulate. The scientific question is whether a circuit grown this way keeps resolvable gradients after it crosses into the regime where classical simulation gets expensive. That is a claim about the optimization trajectory, not about random initialization, and the Cerezo et al. taxonomy leaves it open. Either answer is informative. A positive answer gives a concrete candidate for their "trainable but not simulable" region. A negative answer shows that warm-started growth quietly stays inside the simulable class.
Core Claim
A shallow circuit whose cost depends on local observables has a computable "locality budget": each Hamiltonian term, pulled back through the circuit, touches at most $k$ qubits. Shadow estimation of such a term costs roughly $3^k$ times more samples than a single-qubit term. As long as $k$ stays below a threshold $k^\star$, one shadow ensemble supports thousands of classical optimization steps. This is the regime exploited by ALSO, which prepares shadows once and optimizes classically for 10,000+ steps.
We claim three things. The same locality budget is a principled switch for when to stop simulating. The same shadows can choose which entangling gates to add next at no extra quantum cost. And gates added at the identity on top of a shadow-optimized state begin with gradients set by the current state, not by a random circuit. For a candidate gate $e^{-i\phi G}$ inserted at $\phi=0$, the initial energy gradient is
$$\left.\frac{\partial E}{\partial \phi}\right|_{\phi=0} = i\,\langle \psi | [H, G] | \psi \rangle ,$$
where $H$ is the Hamiltonian, $G$ is the gate's generator, and $|\psi\rangle$ is the current state. When $G$ and the terms of $H$ are local, the commutator is a local observable. Its expectation for every candidate in the pool can therefore be read off one shadow ensemble simultaneously. This replaces the gate-by-gate pool screening of adaptive VQE, and the classical DMRG pre-computation (a classical tensor-network ground-state solver) used by Mutual information-assisted adaptive VQE.
Research Program
The protocol runs in cycles.
Classical phase. The device prepares the current shallow circuit's state and collects random-Pauli shadows. The classical co-processor then optimizes all parameters in the Heisenberg picture, as ALSO does.
Growth step. From the same shadows, the co-processor estimates two quantities for every candidate gate: the commutator gradients above, and two-qubit mutual information from Rényi-2 purities of one- and two-qubit reduced states. It adds the top candidates, initialized at the identity.
Handoff check. It recomputes the locality budget. If every pulled-back cost term still has weight at most $k^\star$, the cycle repeats. Otherwise the protocol hands off. The circuit is frozen as a warm start, and the newest layers are trained on the device with shot-based gradients: parameter-shift, or shadow gradients in the style of Quantum Shadow Gradient Descent where locality still permits.
Measurements of trainability versus simulability. Along each trajectory we record two quantities. The first is a gradient signal-to-noise ratio: the norm of the true gradient divided by the shot-noise standard deviation at a fixed per-step shot budget. The second is the classical price of what was reached: the smallest matrix-product-state bond dimension $\chi$ that reproduces the energy and the measured two-point correlators of the circuit's output within tolerance. The central plot shows both quantities against qubit number $n$, before and after handoff.
Analysis. We will bound the post-handoff gradient variance for identity-initialized layers acting on a state with bounded correlation length. The expectation is that the variance is controlled by the norm of $[H,G]$ restricted to the gate's light cone, rather than by $2^{-n}$. That would explain why the warm start avoids the average-case plateau. It would also make explicit that the guarantee is local in parameter space, consistent with the Cerezo et al. argument about small parameter regions.
Several prior results play distinct roles. ALSO supplies the classical phase but fixes the ansatz and stops at depth $O(\log n)$; we add growth and an exit. QSGD supplies one-pass shadow gradients but does not analyze where locality fails; our handoff criterion is exactly that analysis. The entanglement results of Ballarin et al. and the static depth choice of entropy matching motivate growing entanglement gradually instead of fixing depth in advance. Cerezo et al. constrains the whole project: it predicts that, absent something new, the grown circuit will remain simulable. That prediction is the hypothesis we test against.
Noise-induced barren plateaus are a separate threat, and we do not claim to cure them. We therefore run every experiment under both noiseless and depolarizing-plus-readout noise models at realistic gate error rates. We report where noise, rather than the landscape, ends training.
Baselines
Naive reference. A hardware-efficient ansatz at the final depth our protocol reaches, randomly initialized and trained with Adam on parameter-shift gradients. It receives the same total shot budget, the same number of optimizer steps, and a matched learning-rate sweep. This measures the plateau we claim to avoid. On the Heisenberg and TFIM models we found no reported result on Lacuna for final energy error; the classical-initialization study Can Classical Initialization Help... reports only "marginal" gains without numbers.
Strongest existing quantum-resource method: ALSO run to its limit. We run ALSO at the largest depth its locality budget allows, with the same shadow budget. Its reported result is "2–3 orders of magnitude fewer quantum resources" than standard VQAs on 8-qubit state preparation. Beating it means reaching lower energy error at comparable total shots, specifically on instances where ALSO's depth-limited circuit plateaus above chemical accuracy.
Strongest geometric optimizer. QN-SPSA from Simultaneous Perturbation Stochastic Approximation of the Quantum Fisher Information, run on the same final-depth circuit. It reports convergence with 2,100 circuit executions per iteration, where analytic natural gradient needed about 2.6 million, on a 22-qubit, 132-parameter problem. We give it the same shot budget.
Strongest classical competitor. DMRG, a tensor-network ground-state solver, on the same Hamiltonians, allowed the same bond dimension that our $\chi$ diagnostic says is needed to reproduce our circuit's output. Beating it at matched $\chi$ is not the goal. The honest claim is that our circuit reaches a given accuracy at a point where DMRG needs a growing $\chi$ with $n$. For Max-Cut on random 3-regular graphs, the reference is Benchmarking variational quantum algorithms for combinatorial optimization in practice. There, shallow CVaR-VQE beat random sampling in about 95% of instances at $N=31$, but its margin over greedy search shrank by $N=61$ and exact success probability fell toward zero. We compare against greedy local search and against that VQE under their cap of $10^6$ function evaluations.
Ablations isolating the key ingredients. Each ablation keeps everything else fixed. Random gate growth replaces commutator-gradient growth. DMRG-computed mutual information, as in the MI-assisted ADAPT paper, replaces shadow-estimated mutual information; that paper reports chemical accuracy for H$_2$ with the pool reduced to 1.5%. A fixed handoff depth replaces the locality-budget handoff. The deep circuit initialized with identity blocks (Barren Plateaus in Variational Quantum Computing) or small Gaussian angles (Escaping from the Barren Plateau via Gaussian Initializations) replaces the shadow-optimized warm start.
Why not just…?
Why not just use identity-block or Gaussian initialization? This is the most serious threat. Both methods also start near the identity and avoid the average-case plateau for free. The difference is where they start. Our warm start sits at an optimized low-energy state with a data-chosen gate layout, not at $|0\rangle$. The initialization ablation above settles the question directly: if identity blocks match our energy error and gradient signal-to-noise ratio at equal shots, the shadow phase adds nothing, and we will say so. The empirical record for initialization alone under noise is weak (Can Classical Initialization Help...), but that study did not test Gaussian scaling at our sizes, so we do not treat the question as settled.
Why not just run DMRG or matrix-product-state simulation and skip the hardware? For 1D spin chains this probably wins, and we expect it to. MPS simulation already handled 50-qubit QNNs in Ballarin et al.. That is why the bond-dimension diagnostic is a primary outcome, not an afterthought, and why we include 2D lattices and a molecular Hamiltonian, where DMRG's cost grows. If $\chi$ stays flat in $n$ after handoff, the method has collapsed into the simulable class. We report that as a confirmation of Cerezo et al., not as a success.
Why not just stay in the shadow-simulable regime, as ALSO does? ALSO is cheap and provably trainable, but its depth is capped at $O(\log n)$, and a fixed brickwork circuit cannot represent long-range correlations. The ALSO-at-its-limit baseline measures exactly the accuracy left on the table. If ALSO already reaches chemical accuracy on every instance, the handoff is unnecessary for those problems.
Why not just prune features or fix depth by entropy matching? ATP and entropy matching control entanglement before training, for classical-data classifiers. Our targets are Hamiltonian ground states, which have no input features to prune, and the right depth there depends on the trajectory. The fixed-depth ablation tests whether adapting depth during training matters.
Proposed Experiments
Testbeds. We use three families at 8 to 30 qubits with statevector simulation, extended by MPS emulation where possible:
- 1D and 2D transverse-field Ising and Heisenberg models near criticality;
- LiH and BeH$_2$ under Jordan-Wigner and Bravyi-Kitaev mappings, with success defined as chemical accuracy of $1.6\times10^{-3}$ Hartree;
- Max-Cut on random 3-regular graphs, $N=11$ to $61$, using the 25-instance-per-size protocol of the benchmarking paper.
Chemistry is deliberately adversarial for our protocol. Its non-local terms exhaust the locality budget early, so the handoff comes sooner there, and we expect the smallest benefit.
Metrics. We report final energy error or approximation ratio, total shots, the gradient signal-to-noise ratio trajectory, and the bond dimension $\chi(n)$ needed to reproduce the output.
Three decisive checks. First, gradient signal-to-noise ratio in the first 50 steps after handoff must decay at most polynomially in $n$, compared with the exponential decay of the naive baseline at the same depth. If it decays exponentially, the warm start does not transfer. Second, $\chi(n)$ for the post-handoff states must grow with $n$ on 2D and chemistry instances at fixed accuracy. If it does not, we have not left the simulable class. Third, the protocol must beat the identity-block ablation at matched shots. If it does not, the shadow phase is unnecessary.
Only the conjunction of all three would support the core claim. A failure of the second check with success on the first and third would still be a useful result: shadow-guided growth would then be an efficient classically-assisted solver, but not a quantum-advantage candidate.
References
- Does provable absence of barren plateaus imply classical simulability?
- Alternating Layered Variational Quantum Circuits Can Be Classically Optimized Efficiently Using Classical Shadows
- Mutual information-assisted adaptive variational quantum eigensolver
- Quantum Shadow Gradient Descent for Quantum Learning
- Entanglement entropy production in Quantum Neural Networks
- Problem-dependent Quantum Circuit Design Based on Entropy Matching
- Can Classical Initialization Help Variational Quantum Circuits Escape the Barren Plateau?
- Simultaneous Perturbation Stochastic Approximation of the Quantum Fisher Information
- Benchmarking variational quantum algorithms for combinatorial optimization in practice
- Barren Plateaus in Variational Quantum Computing
- Escaping from the Barren Plateau via Gaussian Initializations in Deep Variational Quantum Circuits
- ATP: Adaptive Threshold Pruning for Efficient Data Encoding in Quantum Neural Networks