Can you eliminate mid-circuit measurements entirely in variational algorithms, or do some remain fundamental?

Can You Eliminate Mid-Circuit Measurements in Variational Algorithms?

No—some measurements appear fundamental, but you can substantially reduce them.

### What Can Be Eliminated

The optimistic finding is that *some* mid-circuit measurements can be replaced entirely. [AG-2024.05-2315] presents a static circuit optimization that substitutes certain mid-circuit measurements with randomized gate applications, using classical precomputation of measurement outcome probabilities. The runtime scales polynomially, making this practical. This works because not every measurement outcome genuinely needs to feed back into the circuit; sometimes you can "burn" that information by randomizing gates instead.

Similarly, [AG-2024.01-1477] reduces measurement *costs* in adaptive variational quantum algorithms (like ADAPT-VQE) by recycling second-derivative information (the Hessian) across iterations—fewer fresh measurements are needed if you carry over optimization data. And [AG-2024.06-1943] introduces a novel measurement strategy that avoids the Hadamard test entirely, instead using linear combinations of ansatz circuits, thereby cutting reliance on two-qubit gates and the measurements that accompany them.

### Why Some Measurements Stay Fundamental

However, mid-circuit measurements themselves appear to serve a deeper computational role in certain regimes. [AG-2024.01-1336] shows that algorithms using mid-circuit measurements and classical feedback can naturally resist decoherence—even infinitely deep noisy circuits maintain useful quantum correlations. This suggests that mid-circuit measurements aren't just overhead; they can be a *computational tool* that stabilizes the quantum state, not just an error-correction patch.

The broader picture: [AG-2024.04-1475] compares Quantum Non-Demolition Measurement (QNDM)—a way to estimate gradients efficiently—to standard approaches, finding QNDM more resource-efficient. This hints that how you *measure* matters as much as whether you measure.

### The Bottom Line

You can eliminate *redundant* mid-circuit measurements through circuit optimization and careful algorithm design, and you can reduce measurement *cost* by reusing classical information across iterations. But measurements that feed forward quantum information appear to remain valuable—sometimes essential—for maintaining signal in deep noisy circuits.