Limits of Error Mitigation

Overview

Error mitigation can meaningfully improve results on noisy devices. But it has fundamental limits.

This page answers:

• “Why can’t we just mitigate our way to arbitrarily accurate quantum computation?”

We’ll discuss:

• why mitigation cost grows quickly
• how noise amplification and sampling requirements appear
• what kinds of problems mitigation struggles with
• why, at some point, stronger reliability mechanisms become necessary (without diving into error-correcting codes)

Intuition

Mitigation is an estimation strategy

Most mitigation methods work by collecting extra data and estimating what the answer would be without noise.

If noise is small, estimation works well. If noise is large, estimation becomes unstable:

• the signal you want is buried
• corrections can amplify uncertainty

Noise amplification and sampling cost

Many mitigation techniques effectively multiply the amount of experimental effort you need:

• ZNE requires multiple noise-scaled circuit runs.
• Post-selection discards data, so you need more shots.
• Some corrections amplify statistical noise.

A good rule of thumb:

• as circuits get deeper and noisier, you need disproportionately more samples to maintain accuracy

Why mitigation doesn’t scale indefinitely

As you scale up qubits and depth:

• noise processes become more complex
• correlations matter more
• calibration overhead grows
• “clean extrapolation” assumptions break down

So mitigation is powerful in a regime, but it is not a universal solution.

Formal Description

We present the limits conceptually as three bottlenecks.

Bottleneck 1: sampling overhead

Mitigation often increases variance (uncertainty) of the final estimate. To compensate, you increase shots.

If the required shots grows too large, the method becomes impractical.

Bottleneck 2: model/assumption mismatch

Mitigation methods assume something about noise:

• smooth dependence on noise level (ZNE)
• stable confusion matrix (readout mitigation)
• trustworthy constraints (post-selection)

If the assumptions don’t hold, the “correction” can be biased.

Bottleneck 3: depth and complexity

For large circuits:

• gate errors and decoherence accumulate
• output distributions can become close to random

When the device output is almost independent of the ideal answer, mitigation cannot recover information that is no longer present.

Worked Example

Imagine two regimes:

1. Mild noise, shallow circuit

   • raw answer is “close but biased”
   • mitigation can correct the bias with manageable extra cost

2. Heavy noise, deep circuit

   • raw answer is close to random
   • ZNE fits become unstable
   • post-selection discards most shots
   • corrected estimates have huge uncertainty

The key point:

• mitigation works best when the ideal signal is still meaningfully present

Turtle Tip

Common Pitfalls

Quick Check

What’s Next

Mitigation is the practical toolset for the NISQ era. To go beyond shallow circuits and noisy estimates, the field ultimately needs methods that protect quantum information during computation. We’ll transition next into the motivation for that next step, while keeping the story grounded and realistic.