Quantum computers are often proposed as solvers for differential equations. The usual recipe prepares a trial solution as a quantum state from a tunable circuit and adjusts the circuit until the state satisfies a discretized equation $Au=b$. When such training stalls, the standard explanation is the barren plateau: in large random-looking circuits, gradients shrink exponentially with the number of qubits. This note argues that quantum PDE solvers carry a second, independent source of flatness that the barren-plateau literature does not account for. That source is the conditioning of the differential operator itself.
On a grid of $N=2^n$ points, the discrete Laplacian's condition number grows like $N^2 = 4^n$. The residual-squared losses used by variational linear solvers square it again, to about $16^n$. This is exponential in qubit count even for shallow, non-random circuits. We propose to measure this operator-induced plateau separately from the Hilbert-space one. We then remove it using two ideas imported from classical physics-informed training. The first replaces the squared operator with an energy functional. The second preconditions the parameter update with a metric weighted by a cheap, spectrally equivalent operator. Finally, we test whether the gain survives honest shot accounting.
Core Claim
The claim has three parts. First, in variational quantum solvers for linear PDEs, the number of measurements needed to converge factors roughly into two parts: a Hilbert-space concentration factor, which depends on the ansatz (the circuit family) and its initialization, and an operator factor, which depends on the condition number of the operator inside the loss. Second, the operator factor dominates the final convergence phase at the 6–12 qubit scales used today. Third, it can be driven to $O(1)$ for variable-coefficient elliptic operators. None of the existing mitigations touch it, because they all act on the circuit or the optimizer, not the operator.
The mechanism comes from An operator preconditioning perspective on training in physics-informed machine learning. That paper shows that classical physics-informed networks stall because gradient descent sees the Hermitian square $\mathcal{D}^*\mathcal{D}$ of the differential operator $\mathcal{D}$. For Poisson this means a condition number growing as $K^4$ in frequency $K$. The paper also shows that preconditioning cures this, reaching residual $10^{-10}$ in 100 epochs where unpreconditioned training stalls at $10^{-2}$ to $10^{-3}$. Its guarantees hold only for linear operators. We accept that restriction as a design constraint rather than fight it: the quantum subproblem is kept linear, and any nonlinearity is handled by an outer classical iteration.
The contrast between the two losses can be written directly. Take a normalized circuit state $|x(\theta)\rangle$ representing $u$ up to scale, where $\theta$ are the circuit parameters. The residual loss $C_{\mathrm{res}}(u)=\|Au-b\|^2$ has Hessian $2A^\top A$. The energy (Ritz) functional $E(u)=\tfrac12 u^\top A u - b^\top u$ has Hessian $A$. For the 1D Dirichlet Laplacian,
$$\kappa(A)\approx \frac{4N^2}{\pi^2},\qquad \kappa(A^\top A)=\kappa(A)^2,$$
which follows from the eigenvalues $\lambda_k=(4/h^2)\sin^2(k\pi h/2)$ on spacing $h=1/N$. Minimizing $E$ over the scale of $u$ in closed form gives a cost that depends only on the state:
$$C_E(\theta)=-\frac{|\langle b|x(\theta)\rangle|^2}{\langle x(\theta)|A|x(\theta)\rangle}.$$
This cost needs expectation values of $A$, not $A^2$. That also shrinks the measurement breadth, which Solving 1D Poisson problem with a Variational Quantum Linear Solver had to fight with a $2^{n-1}+n$-term gate decomposition.
Research Program
The first component is a diagnostic protocol that separates the two plateaus. For a fixed hardware-efficient circuit, meaning layers of single-qubit rotations and entangling gates native to the device, we measure two things. At random initialization we record the gradient variance versus $n$. This is the quantity Barren plateaus in quantum neural network training landscapes showed decays as $O(1/2^n)$ up to 24 qubits. Near the optimum we record the Hessian spectrum and the shots-to-precision. These are the quantities the operator factor should control. Measuring both on the same circuits, while only the loss operator changes, isolates the operator contribution. Varying only the circuit depth and initialization, while the operator stays fixed, isolates the Hilbert-space contribution.
The second component is the energy cost $C_E$. It is estimated with Hadamard tests: circuits with one control qubit whose measurement reveals an overlap. We write the Laplacian as $2I-S-S^\dagger$ plus a boundary correction, where $S$ is a quantum increment (shift) circuit. This needs a constant number of terms rather than an exponential decomposition. Variable coefficients $a(x)$ enter through $A=D^\top \mathrm{diag}(a) D$, where $D$ is the difference operator, and $\mathrm{diag}(a)$ is expanded in a few low-order Pauli-Z terms for smooth $a$.
The third component is operator-weighted metric preconditioning. We update $\theta \leftarrow \theta-\eta M^{-1}\nabla C_E$, with
$$M_{ij}=\mathrm{Re}\,\langle \partial_i x|L_0|\partial_j x\rangle .$$
Here $L_0$ is the constant-coefficient Laplacian at the mean coefficient. This is the quantum analogue of the "gradient modification by inverse operator" preconditioner of De Ryck et al. Setting $L_0=I$ recovers ordinary quantum natural gradient. Near the optimum, the effective condition number becomes that of $L_0^{-1}A$, which is bounded by $\max a/\min a$ independent of $N$ (standard spectral equivalence). The cost is $O(p^2)$ extra Hadamard tests per step for $p$ parameters. We therefore also study block-diagonal and diagonal approximations, and charge every shot to the method.
The fourth component extends the method to nonlinear problems without breaking the linear restriction. We take PDEBench's 1D diffusion-reaction problem and step it semi-implicitly: $(I-\Delta t\,\Delta)u^{k+1}=u^k+\Delta t\,R(u^k)$, where $R$ is the nonlinear reaction term. Each step is a linear, well-posed solve for the quantum routine, and the nonlinear term is evaluated classically between steps. The operator $I-\Delta t\,\Delta$ also gives a tunable conditioning knob, since $\kappa\approx 1+\Delta t\cdot 4^{n+1}/\pi^2$.
This route is the opposite of Physics-Informed Quantum Machine Learning: Solving nonlinear differential equations in latent spaces without costly grid evaluations. That work builds quantum multiplier circuits for $f^2$ terms and reports average deviation $10^{-3}$, but concedes susceptibility to barren plateaus from the added depth. We keep circuits shallow and move the nonlinearity outside the quantum loop. That choice is in the spirit of the linear/nonlinear split used by Origo and of the Picard-iteration view in Quantitative Approximation for Neural Operators in Nonlinear Parabolic Equations.
We do not claim quantum advantage. Classical fast Poisson solvers are trivially better on these problems. Does provable absence of barren plateaus imply classical simulability? warns that circuits made provably trainable often become classically simulable. Our contribution is the operator side: a failure mode that any future quantum PDE solver inherits whatever its circuit, and a fix that is independent of the circuit. Whether the preconditioned regime coincides with a simulable ansatz will be reported as a finding, not assumed away.
Baselines
The naive reference is the standard variational linear solver: a residual-squared global cost, hardware-efficient ansatz, uniform random initialization, and Adam with parameter-shift gradients. This is the configuration of Solving 1D Poisson problem with a Variational Quantum Linear Solver, which reports simulator fidelity 0.96 on the $4\times4$ (2-qubit) Poisson system and near-1.0 hardware fidelity on $2\times2$. We rerun it on our grids with identical circuits, so the comparison changes only the loss and the update rule.
The strongest competitor on the Hilbert-space side is the residual cost combined with depth-scaled Gaussian initialization. This initialization comes from Escaping from the Barren Plateau via Gaussian Initializations in Deep Variational Quantum Circuits, which proves gradient decay becomes polynomial. We also combine the residual cost with a local cost variant. These two are the best available tools against gradient vanishing at initialization.
The strongest optimizer competitor is NFT-with-EMICoRe from Physics-Informed Bayesian Optimization of Variational Quantum Circuits. It fits a sinusoid-structured Gaussian-process surrogate and reports 98% ground-state fidelity on a 5-qubit Ising model within 4,000–6,000 observations. We run it on both the residual and the energy cost.
The key ablation is ordinary quantum natural gradient, that is, $M$ with $L_0=I$, on the energy cost. It spends the same metric shots but carries no operator information. If it matches our method, the operator weighting is not what helps.
All methods receive the same total shot budget, counting metric and surrogate evaluations. They also receive the same hyperparameter grid (eight learning-rate or step settings, three seeds) and identical circuits. Fidelity is reported relative to the best fidelity each ansatz can reach under exact, noise-free optimization, so that limits of the circuit's expressivity are not mistaken for trainability limits. Our method beats a baseline if two conditions hold. First, it reaches 0.99 relative fidelity with fewer total shots. Second, the fitted exponent of shots-versus-$n$ is smaller by a margin outside the seed spread. At 10–12 qubits, the prediction is roughly a $\kappa(A)$-sized gap over the residual cost and an $O(1)$ versus $4^n$ gap over the energy-only cost.
Why not just…?
Why not just use Gaussian or identity-block initialization? These methods, from the Gaussian-initialization papers and from Wossnig et al., keep the circuit near a low-dispersion region so gradients are large at the start. They say nothing about the curvature of the loss near the solution, which is set by $A^\top A$. Our diagnostic predicts that they improve early-phase gradient variance but leave the final-phase shot count growing with $\kappa$. If they instead close the gap, the operator factor is not dominant at these scales and our claim fails. The experiment is designed to show this either way.
Why not just use a gradient-free or Bayesian optimizer? NES, particle swarm, and EMICoRe all avoid tiny gradient vectors. However, they must still resolve cost differences between nearby points. On an ill-conditioned cost, the differences along low-curvature directions are $\kappa$-times smaller, so the shot requirement returns in another form. EMICoRe is a real threat on small problems because its three-point rule is very sample-efficient. We therefore run it on both costs. Its advantage should shrink on the energy cost if the operator factor is the bottleneck.
Why not just solve the PDE classically, or with a preconditioned PINN? For these problems that is the right engineering answer, and we say so. De Ryck et al. already reach $10^{-10}$ residuals classically, and QuanONet reports classical DeepONet MSE 0.027 against 0.039 for its quantum model on a nonlinear ODE. The question here is scientific: whether the conditioning obstacle transfers to quantum solvers, and whether it can be removed there. Classical results serve only as ground truth.
Why not just switch to the energy cost and skip the metric? This is the most serious alternative. The energy cost alone may capture most of the gain, since it removes the squaring. The energy-only arm is therefore a full experimental condition. The metric preconditioner earns its place only if the $4^n$ residual growth of the energy cost is measurable within the budget at 10–12 qubits.
Proposed Experiments and Decisive Checks
We will use noiseless statevector and shot-noise simulation at 3–12 qubits. The test problems are the 1D Dirichlet Poisson problem, a Helmholtz problem, and variable-coefficient $-(a u')'=f$. The coefficient $a$ is drawn from Gaussian random fields, following QuanONet's function sampling, with contrast $\max a/\min a$ swept from 2 to 100. We add the semi-implicit diffusion-reaction steps from PDEBench, scored with its RMSE and Fourier-spectral error. A depolarizing-noise sweep tests interaction with noise-induced barren plateaus. The metrics are relative fidelity, relative $L^2$ error, total shots to 0.99 relative fidelity, gradient variance at initialization, and the Hessian condition number at the optimum.
Three checks are decisive. The first holds the circuit fixed and sweeps only the operator. If the final-phase shot counts scale with $\kappa$ of the operator in the loss ($16^n$, $4^n$, and bounded, for the three losses), the two-factor picture holds. The second tests whether the operator-weighted preconditioner's shot count is flat in $\max a/\min a$ while plain natural gradient's is not. This isolates the operator weighting as the active ingredient. The third runs the diffusion-reaction rollout with $\Delta t$ as the conditioning knob and tests whether per-step accuracy tracks $\kappa(I-\Delta t\,\Delta)$ for the baselines but not for the preconditioned solver. A negative result on the first check, where operator conditioning is invisible behind Hilbert-space concentration, would itself be informative. It would mean the circuit-side plateau is the only obstacle worth attacking at current scales.
References
- An operator preconditioning perspective on training in physics-informed machine learning
- Solving 1D Poisson problem with a Variational Quantum Linear Solver
- Barren plateaus in quantum neural network training landscapes
- Physics-Informed Quantum Machine Learning: Solving nonlinear differential equations in latent spaces without costly grid evaluations
- Origo: Interpretable Multi-physics PDE Foundation Model through Neural Operator Splitting
- Quantitative Approximation for Neural Operators in Nonlinear Parabolic Equations
- Does provable absence of barren plateaus imply classical simulability?
- Escaping from the Barren Plateau via Gaussian Initializations in Deep Variational Quantum Circuits
- Physics-Informed Bayesian Optimization of Variational Quantum Circuits
- An initialization strategy for addressing barren plateaus in parametrized quantum circuits
- QuanONet: Quantum Neural Operator with Application to Differential Equation
- PDEBench: An Extensive Benchmark for Scientific Machine Learning
- Noise-induced barren plateaus in variational quantum algorithms