Noise-Aware Optimization

Everything so far assumed perfect circuits. On real NISQ hardware every gate leaks a little error, measurements misreport, and qubits decohere. A noise-aware VQE accepts that the energies it measures are biased and noisy, and adapts both its optimizer and its measurement pipeline accordingly.

Two distinct problems: bias and variance

Noise corrupts an energy estimate in two different ways, and they need different cures.

Statistical variance comes from finite shots: even a noiseless device only samples the distribution. The standard error of an expectation value falls as $1/\sqrt{N_{\text{shots}}}$ , so halving the error costs four times the shots.
Systematic bias comes from gate and readout errors: they push the measured energy away from the true value in a reproducible direction, no matter how many shots you take. More shots shrink variance but never touch bias.

A flat, noisy cost landscape stalls gradient-based optimizers, and a biased landscape can move the location of the minimum, so $\boldsymbol{\theta}^\star$ itself becomes wrong.

Robust optimizers

Gradient methods that assume clean derivatives misbehave when each evaluation is noisy. Two families cope better:

SPSA (Simultaneous Perturbation Stochastic Approximation). Estimates the whole gradient from just two energy evaluations using a single random perturbation direction, regardless of the number of parameters. It is cheap and surprisingly resilient to shot noise, which is why it is a default for hardware VQE.
Gradient-free methods such as COBYLA or Nelder–Mead avoid derivatives entirely and tolerate moderate noise, though they scale poorly to many parameters.

Shot budgeting also matters: spend few shots early when you only need the rough downhill direction, and ramp up shots near convergence when you need to resolve small energy differences.

Error mitigation

Distinct from full error correction (which needs many physical qubits per logical qubit), error mitigation post-processes noisy results to estimate the noiseless value cheaply:

Zero-noise extrapolation (ZNE). Deliberately amplify the noise — e.g. by stretching gate durations or inserting gate–inverse pairs — measure the energy at several noise levels, then extrapolate back to the zero-noise limit. Temme et al. showed a low-order polynomial fit removes much of the leading bias.
Readout-error mitigation. Characterize the measurement confusion matrix on basis states and invert it to correct the output distribution.
Probabilistic error cancellation. Represent the inverse of the noise channel as a quasi- probability mixture of implementable operations; it removes bias exactly in expectation, at the cost of increased variance (more shots).

Putting it together

A practical noise-aware VQE loop looks like: prepare $|\psi(\boldsymbol{\theta})\rangle$ with a shallow, symmetry-respecting ansatz; measure grouped commuting Pauli terms with readout mitigation; estimate the energy with zero-noise extrapolation; drive the optimization with SPSA on a growing shot budget. Each ingredient targets a specific failure mode — depth, measurement count, readout bias, gate bias, and noisy gradients respectively.

Key takeaways

Noise splits into shrinkable variance ( $1/\sqrt{N_{\text{shots}}}$ ) and stubborn systematic bias.
Robust optimizers (SPSA, gradient-free) handle noisy cost evaluations; shot budgets should grow toward convergence.
Error mitigation — ZNE, readout correction, probabilistic error cancellation — removes bias without the qubit overhead of error correction.
VQE tolerates some coherent error by retuning parameters, but incoherent noise must be mitigated.