Error Mitigation in VQE

Overview

VQE is designed for NISQ hardware, but it is still sensitive to noise. Noise affects VQE in two ways:

• it biases energy estimates
• it increases variance (uncertainty), making optimization less stable

This page explains why VQE often uses mitigation techniques and what tradeoffs they introduce.

Intuition

Why VQE needs mitigation

VQE relies on comparing energies across parameter settings. If noise shifts energies unpredictably (or consistently biases them), the optimizer can:

• move in the wrong direction
• get stuck
• “converge” to a result that is mostly noise-driven

Mitigation aims to make the energy estimates more faithful to the ideal circuit.

The core tradeoff

Mitigation typically trades:

• more circuits and more shots

for:

• a better estimate (less bias or better interpretability)

So mitigation is not free. It increases experimental cost.

Formal Description

We keep this method-focused and hardware-agnostic.

Readout mitigation in VQE

VQE estimates expectation values from measured bitstrings. If measurement misclassifies bits, expectation values are distorted.

Readout mitigation workflow:

1. calibrate measurement confusion probabilities
2. correct observed counts (conceptually: invert the confusion effect)
3. compute expectation values from corrected frequencies

This can noticeably improve results in shallow circuits. But it does not fix gate errors or decoherence.

Zero-noise extrapolation (ZNE) in VQE

ZNE idea:

• run the same logical energy-estimation circuit at multiple noise levels
• fit how the energy estimate changes with noise
• extrapolate to a zero-noise estimate

In VQE, ZNE can be applied to:

• individual Hamiltonian term estimates
• grouped measurement settings
• or the total energy estimate

Tradeoff:

• each noise scale requires extra runs, multiplying cost

Cost vs accuracy

Mitigation can improve accuracy, but it increases:

• number of circuit executions
• total shot budget
• sensitivity to calibration drift

So VQE often uses selective mitigation:

• mitigate the dominant error source first (often readout)
• use heavier mitigation only when it measurably helps

Why mitigation does not scale indefinitely

As circuits and systems scale:

• measurement cost increases
• mitigation overhead multiplies that cost
• assumptions behind mitigation (smoothness, stationarity) become harder to maintain

So mitigation can extend the useful regime, but it cannot replace fault-tolerant protection for very deep computations.

Worked Example

Suppose a VQE iteration estimates energy as 0.72. You apply readout mitigation and the estimate shifts to 0.75.

Interpretation:

• the raw result was biased by measurement misclassification
• mitigation adjusted the estimate toward what an ideal measurement would likely have produced

Now suppose you also run ZNE with three noise scales. You get energy estimates:

• 0.72 (base)
• 0.68 (more noise)
• 0.65 (even more noise)

Extrapolation might estimate a zero-noise energy closer to 0.78.

This looks like progress, but note the costs:

• you ran multiple versions of the circuit
• uncertainty may increase due to extrapolation

Turtle Tip

Common Pitfalls

Quick Check

What’s Next

With VQE in place, we can move to another major variational algorithm family: QAOA. QAOA uses a different kind of structured ansatz and is often framed around combinatorial optimization.