Skip to content

QubitWire / Self-directed portfolio project

Mitiq pathway: test whether extrapolation earns its extra samples

Compare extrapolation with an equally funded baseline and keep uncertainty visible.

Reviewed

Starting skills
Python, expectation values, elementary regression and sampling uncertainty
Portfolio output
Equal-budget mitigation audit with repeated complete fits
A useful completion check
All scaled-circuit samples counted; improvement assessed across repetitions
Hardware needed
No
Data status
Explicitly synthetic

Mitiq’s official documentation introduces zero-noise extrapolation as a way to estimate an observable at a lower effective noise level using results obtained at several amplified noise levels. This pathway asks a narrower, practical question: when does the extra sampling and model fitting improve the estimate enough to justify its cost? The project is an original statistical audit built around a small simulated circuit. It does not demonstrate error-corrected computation, a measured commercial device or an open position with a software employer.

Choose an observable with an independently calculable reference value. Define a noise model and a set of scale factors before looking at the results. For each scale factor, preserve the circuit transformation, sample budget and measured estimate. The first comparison should show the raw noisy result beside the extrapolated result, with uncertainty for both. An extrapolated number that lies closer to the reference in one trial can still have a larger spread across repetitions. That distinction is the reason to repeat the complete experiment, including the fit.

The second part is a fair resource comparison. Give the unmitigated baseline the same total sampling budget used across all scaled circuits. If circuit depth or execution cost differs, report that separately instead of hiding it inside a nominal shot count. Then compare error distributions at the same total budget. This proposed experiment may identify regimes in which extrapolation helps and regimes in which it does not. Neither outcome should be suppressed. A useful portfolio can explain a failure precisely without presenting it as a universal verdict on mitigation.

Add a model-choice review. Fit two justified extrapolation forms and examine their sensitivity to the chosen scale range, a held-out point or one noisy estimate. Explain why extending a fitted curve beyond the observed region requires assumptions. Record estimates that leave the observable’s allowed range, and distinguish an optional clipping rule from the original estimator. Clipping can change statistical properties; it should be reported as an additional procedure rather than silently making the result appear physical. Keep all synthetic noise parameters clearly labelled and avoid attributing them to an actual processor.

Questions to take with you

What is the fair baseline?

An unmitigated estimate using the same total sampling budget, with circuit execution cost also disclosed. Comparing against one underfunded raw estimate can overstate the value of mitigation.

Why preserve an estimate outside the allowed range?

It reveals behaviour of the estimator. A separate clipping rule can be evaluated, but silently clipping would hide part of the method and alter the comparison.

Self-directed project; no vacancy, completed execution, certification or employer endorsement is implied. Pin the software environment before presenting runnable code.

Sources and what they show (2)
Explore other portfolio projects