What if you used error mitigation techniques instead of reducing depth—could that preserve expressivity while maintaining noise resilience?

This is an excellent question that cuts to the heart of near-term quantum computing strategy. The short answer: yes, error mitigation can preserve expressivity better than depth reduction alone—but with important trade-offs.

The expressivity-resilience tension

When you reduce circuit depth to fight noise, you're making a hard choice: lose computational power to gain stability. Error mitigation techniques, by contrast, try to *estimate what a noiseless circuit would have produced* without throwing away the deep circuit entirely [AG-2024.03-1985]. This is conceptually more appealing because you keep the circuit's expressiveness intact.

How error mitigation works

The main strategies fall into a few families:

Zero-noise extrapolation (ZNE) runs your circuit at different noise levels (by adding extra gates), then extrapolates backward to what zero-noise would give [AG-2024.03-1638]. The key insight is that you can *amplify* noise deliberately to map the trend, then fit a curve back to the noiseless limit.

Probabilistic error cancellation (PEC) characterizes your noisy gates, then adds random Pauli corrections to "undo" the noise in post-processing [AG-2024.03-2192]. This works beautifully for Clifford gates but breaks down for non-Clifford operations—until recently. The field has now generalized PEC to weakly-entangling non-Clifford gates (like small-angle interactions), opening up simulation of quantum dynamics without sacrificing noise-resilience [AG-2024.03-2192].

Tensor-network error mitigation (TEM) uses structure from classical simulation to reconstruct the error-free answer and turns out to be *optimal*—it saturates the universal lower bound on sampling overhead needed for any error mitigation strategy [AG-2024.03-1985].

The catch: sampling overhead

Here's the tension: error mitigation doesn't break the laws of physics. To recover a noiseless answer from noisy measurements, you need to run the circuit many more times. The sampling cost grows exponentially with noise level in the worst case. So you're trading *shots* (measurement repetitions) for *depth*. If your hardware can tolerate thousands of shots but struggles with deep circuits, this is a win. If shots are scarce, it's not [AG-2024.03-1985].

Keeping expressivity alive

What does the archive show about preserving expressivity *while* staying noise-resilient?

A practical angle: post-selection with smart abort criteria. Rather than running every instance to the end, you can reject measurements that look "too noisy" during error correction [AG-2024.05-1757]. Remarkably, below a noise threshold, the exclusion rate *decays* with code distance, meaning larger systems actually need to throw away fewer results—a path to scalable quantum computing that doesn't sacrifice depth.

Another approach: leverage "junk" information. Measurements that violate symmetry constraints are normally discarded as garbage, but they actually encode error signatures [AG-2024.02-1735]. A new method called the Self-Trained Quantum Noise Filter (SQNF) uses these junk events to learn the noise structure without extra experiments, improving accuracy without extra cost.

For algorithms with a single correct output (like many practical applications), a simple qubit-wise majority vote recovers the right answer even when it never appeared in raw measurements [AG-2024.02-1801]. This requires fewer shots than existing techniques.

The hard constraint: noise instability

One subtle catch: error mitigation relies on *knowing* your noise model. If your device's noise drifts (as happens in superconducting qubits interacting with stray defects), your mitigation strategy becomes unreliable [AG-2024.07-1608]. Physical stabilization of the device—tuning out unwanted interactions—is sometimes necessary before mitigation even works.

The final verdict on expressivity

Deep circuits under realistic noise are not automatically more expressive: uncorrected noise tends to truncate circuits to logarithmic effective depth anyway [AG-2024.03-2009]. But error mitigation *can* change this story by recovering the deep circuit's output *without* actually needing it to stay quantum-coherent. You're outsourcing the coherence requirement to classical postprocessing and extra measurements.

In summary: Error mitigation preserves expressivity better than naive depth reduction, especially with techniques like TEM (optimal sampling cost) and SQNF (free improvements via junk information). The real cost is measurement shots and device noise stability, not gate depth. For near-term quantum advantage, this is likely the right trade [AG-2024.03-1985].