Measurement Calibration & Readout Correction

Overview

Measurement calibration answers a practical question:

• “When the device reports 0 or 1, how often is it wrong?”

Because readout errors can dominate observed results, you often need to:

• estimate the readout confusion probabilities
• interpret measured counts in light of those probabilities

This page explains how confusion matrices are built conceptually and how simple readout correction works as a post-processing step.

Intuition

Measurement as classification

Readout typically converts a noisy analog signal into a bit. That conversion is a classifier. Classifiers can mislabel.

So it’s natural to model readout as:

• true bitstring → reported bitstring

Confusion matrices

A confusion matrix summarizes misclassification. For one qubit, it captures:

• how often true 0 is reported as 1
• how often true 1 is reported as 0

For multiple qubits, the same idea applies to bitstrings.

Why calibration must be repeated

Readout behavior can drift because:

• electronics drift
• temperature changes
• device state changes between runs

So a confusion matrix measured yesterday may not be accurate today. Calibration is not “one and done.”

Formal Description

We define a calibration experiment and the correction idea without heavy math.

Calibration experiments (conceptual)

For one qubit:

1. Prepare $|0\rangle$ many times and measure. Estimate $P(\text{report }0 \mid \text{true }0)$ and $P(\text{report }1 \mid \text{true }0)$.

2. Prepare $|1\rangle$ many times and measure. Estimate $P(\text{report }0 \mid \text{true }1)$ and $P(\text{report }1 \mid \text{true }1)$.

These four numbers form the 2×2 confusion matrix.

For multiple qubits, you prepare basis states like $|00\cdots 0\rangle$, $|00\cdots 1\rangle$, etc., and estimate how often the device reports each outcome.

Basic readout correction (idea)

Suppose your observed counts are distorted by readout confusion. If you have an estimated confusion matrix, you can attempt to “undo” the distortion in post-processing:

• you treat the observed distribution as the true distribution passed through a noisy measurement channel
• you mathematically invert (or approximately invert) that channel to estimate the underlying distribution

Key interpretation:

• this does not make the hardware noiseless
• it is an estimation step to better interpret measurements

Why correction is imperfect

Correction can amplify statistical noise if:

• the confusion matrix is inaccurate
• the measurement error rates are high
• you have limited shots

So readout correction is useful, but it must be applied carefully.

Worked Example

Assume a 1-qubit device has:

• $P(1\mid 0)=0.03$
• $P(0\mid 1)=0.07$

You run an experiment and observe:

• 600 zeros
• 400 ones

Without calibration, you might conclude the true probability of 1 is 0.40.

But if the device flips 0→1 about 3% of the time and 1→0 about 7% of the time, then:

• some of the observed ones are actually misread zeros
• some of the observed zeros are actually misread ones

Using the confusion matrix, you can estimate a corrected underlying distribution. Even conceptually, the point is clear: calibration lets you separate “state physics” from “measurement mistakes.”

Turtle Tip

Common Pitfalls

Quick Check

What’s Next

Now you have tools to interpret state closeness (fidelity) and measurement behavior (calibration). Next we’ll connect the whole toolbox: how these metrics fit together, what you can realistically know about a device, and why no single number captures “noise.”