Readout and Measurement Error

Every quantum computation ends the same way: delicate superpositions are converted, irreversibly, into classical bits on a wire — and that last step is itself noisy. Readout (or measurement) error is the probability that the bit the apparatus reports differs from the bit corresponding to the qubit's true logical state as measurement begins. Unlike gate error it cannot be echoed away or diluted by a deeper circuit — it sits exactly at the quantum-to-classical boundary, and on today's superconducting and trapped-ion processors it is often the single largest line in a shallow circuit's error budget, with per-qubit assignment fidelities of $97$ – $99\%$ typical rather than exceptional. Two physical mechanisms feed the number. The first is discrimination error: the detector reads a continuous analog signal — a dispersively shifted resonator quadrature, or a count of scattered fluorescence photons — whose distributions for $|0\rangle$ and $|1\rangle$ overlap, so a threshold classifier sometimes mislabels even a perfectly prepared state. The second is decay during readout: the integration window has finite duration $\tau_m$ , and while it runs an excited qubit can relax via amplitude damping ( $T_1$ ), so a true $|1\rangle$ is progressively recorded as $0$ . Together they produce an asymmetric confusion matrix that calibration measures directly. The mathematics that tames it is, surprisingly, classical and linear — the measured distribution is a stochastic matrix times the true one — and this article builds that model from the measurement postulate, shows why inverting it is dangerous, and ties it to the amplitude-damping channel and POVM formalism.

Key takeaways

Readout error is the classical misassignment of an outcome — a true $|0\rangle$ reported as $1$ , or $|1\rangle$ as $0$ — living on the classical register, not on any surviving quantum state.
One qubit is captured by a $2\times2$ column-stochastic confusion matrix $A$ with $A_{ij}=P(\text{measure }i\mid\text{prepare }j)$ ; the model is $\mathbf{p}_{\text{meas}}=A\,\mathbf{p}_{\text{true}}$ .
Physically it blends discrimination (overlap of the $|0\rangle,|1\rangle$ signal blobs in the IQ plane) with $T_1$ relaxation during the integration time $\tau_m$ , which makes $\varepsilon_1>\varepsilon_0$ on superconducting hardware.
Mitigation inverts $A$ , but $A^{-1}$ amplifies shot noise by the condition number $\kappa(A)$ and returns negative "probabilities", so constrained least-squares on the simplex is used instead.
For $n$ qubits $A$ is $2^n\times2^n$ ; uncorrelated readout factorizes it into a Kronecker product of single-qubit blocks, so calibration costs $2n$ experiments, not $2^n$ .
Key idea: model the imperfect measurement as a calibrated column-stochastic confusion matrix $A$ acting on probability vectors, $\mathbf{p}_{\text{meas}}=A\,\mathbf{p}_{\text{true}}$ , and recover expectation values by inverting or constrained-fitting this one linear relation.

The core insight

Model the imperfect measurement as a calibrated column-stochastic confusion matrix $A$ acting on probability vectors, $\mathbf{p}_{\text{meas}}=A\,\mathbf{p}_{\text{true}}$ ; recover the true distribution or its expectation values by inverting, or better constrained-fitting, this linear relation.

Where this lives today

Readout mitigation is now a default layer in every serious quantum-software stack, and the frontier has moved from whether to correct to how to correct at scale. The full $2^n$ matrix is hopeless beyond a dozen qubits, so the field uses scalable successors: the tensored model of the third derivation, and matrix-free methods such as M3 (Nation et al., 2021) that never build $A$ — solving the constrained system only in the tiny subspace of bitstrings that actually appear, bringing mitigation to hundreds of qubits. On the hardware side, teams re-optimize threshold and integration time along the tradeoff curve of the second figure, and train small neural-network or matched-filter classifiers to push discrimination error below the $T_1$ floor. The stakes are highest in error correction. Every stabilizer measurement in a surface code is an ancilla readout, and a misreported syndrome is indistinguishable, to the decoder, from a data error — so readout rates feed directly into the code threshold. Real decoders repeat syndrome extraction over many rounds and match defects in a space-time graph, and the asymmetry $\varepsilon_1>\varepsilon_0$ can be handed to a biased decoder weighting $1\to0$ faults more heavily. Mid-circuit measurement raises the bar further: the readout must be fast, accurate, and non-demolishing enough to leave neighbors undisturbed — making the humble confusion matrix a live constraint at the center of the fault-tolerance roadmap.

The physical picture

Strip the measurement to its bones and it is a classical wire that should carry one honest bit, followed by a binary asymmetric channel that occasionally flips it. A true $1$ can arrive as $0$ (mostly because the excited state decayed on the way), a true $0$ as $1$ (a rare thermal excitation or a misfired classifier crossing threshold). The confusion matrix is just the transition table of that channel — and because the two flip probabilities are unequal, it has a direction. Where does the flip come from? Not from the qubit's Hilbert space but from the analog signal the detector integrates. A dispersive superconducting readout maps $|0\rangle,|1\rangle$ onto two clouds in the in-phase/quadrature (IQ) plane; a fluorescence readout maps them onto two photon-count distributions. Either way the classifier draws a threshold between two overlapping blobs, and the overlap area is the irreducible discrimination error — pushing the threshold merely trades $0\to1$ for $1\to0$ errors.

Figure

Discrimination error is the overlap of two readout-signal distributions. With the threshold placed between the ground blob 0 and the excited blob 1, a true 0 spills above threshold with probability ε₀ ≈ 0.01 and a true 1 spills below with probability ε₁ ≈ 0.05, so the single-shot assignment fidelity is F = 1 − ½(ε₀ + ε₁) = 0.97

The asymmetry $\varepsilon_1>\varepsilon_0$ is not an accident of one platform — it is the fingerprint of energy relaxation. During the microseconds the detector integrates, an excited qubit has a real chance of dropping to ground, after which it emits a ground-state signal for the rest of the window. There is no reverse process pumping a cold qubit up, so the $1\to0$ channel carries an extra, purely temporal term the $0\to1$ channel lacks. This creates a design tension: integrating longer collects more photons and sharpens discrimination, but gives the qubit more time to relax — so an optimal integration time sits between, exactly the balance a calibration routine hunts for.

Figure

The calibration tradeoff for the 1→0 readout error. Longer integration τ_m sharpens discrimination so the amber term falls, but gives the excited state more time to relax so the green T₁-decay term 1 − e^(−τ_m/T₁) rises; their sum ε₁ = ε₁ᵈⁱˢᶜ + (1 − e^(−τ_m/T₁)) is minimized near τ_m ≈ 2.2 μs at ε₁ ≈ 0.027 for T₁ = 100 μs

Finally, mitigation is deconvolution: you observed $\mathbf{p}_{\text{meas}}$ , the sharp $\mathbf{p}_{\text{true}}$ blurred by $A$ . Like every deconvolution it amplifies noise, so the craft is not to slam $A^{-1}$ onto a finite-shot histogram — which routinely yields negative "probabilities" — but to regularize the inverse by constraining the answer to the probability simplex.

The core equations

The entire model rests on one linear relation. A perfect measurement would output the probability vector $\mathbf{p}_{\text{true}}$ over the $2^n$ bitstrings; the noisy apparatus reports $\mathbf{p}_{\text{meas}}$ , and the two are related by a fixed stochastic matrix. Because probabilities sum to one and stay nonnegative, $A$ is column-stochastic.

\mathbf{p}_{\text{meas}} = A\,\mathbf{p}_{\text{true}}, \qquad \sum_i A_{ij}=1,\ \ A_{ij}\ge 0 \qquad (1)

where: $\mathbf{p}_{\text{true}}$ — true outcome-probability vector (length $2^n$ ) of the ideal projective measurement; $\mathbf{p}_{\text{meas}}$ — observed/reported outcome-probability vector; $A_{ij}$ — probability of reporting outcome $i$ given true outcome $j$ ; the stated conditions make each column a valid distribution. Reading it: $A$ smears the distribution you should have seen into the one you did see. The matrix is not postulated but measured. Prepare each computational basis state in turn, histogram the reported bitstrings, and each normalized histogram is one column of $A$ — which is why the columns, not the rows, sum to one.

A_{ij} = P(\text{measure } i \mid \text{prepare } |j\rangle) \qquad (2)

where: $|j\rangle$ — a computational basis state prepared as a calibration input, $j\in\{0,1\}^n$ ; $i$ — the recorded bitstring outcome. Reading it: column $j$ answers "if the truth is $j$ , how are my reports distributed?" For a single qubit this collapses to a $2\times2$ table with two independent error rates that already tell the whole asymmetry story.

A = \begin{pmatrix} 1-\varepsilon_0 & \varepsilon_1 \\ \varepsilon_0 & 1-\varepsilon_1 \end{pmatrix}, \qquad \varepsilon_0=P(1\mid0),\ \ \varepsilon_1=P(0\mid1) \qquad (3)

where: $\varepsilon_0$ — probability of reading $1$ when truly $0$ (thermal population plus classifier overlap); $\varepsilon_1$ — probability of reading $0$ when truly $1$ (classifier overlap plus $T_1$ decay). Reading it: the diagonal is "reported correctly", the off-diagonal "flipped", and $\varepsilon_1\neq\varepsilon_0$ makes the channel directional. The larger rate decomposes into a static classifier piece and a dynamic decay piece — the two mechanisms of the physical picture, now quantitative, and the relation the calibration figure minimizes.

\varepsilon_1 \approx \varepsilon_1^{\text{disc}} + \big(1-e^{-\tau_m/T_1}\big) \qquad (4)

where: $\varepsilon_1^{\text{disc}}$ — discrimination (classifier-overlap) contribution, falling as the window lengthens; $\tau_m$ — effective integration time; $T_1$ — energy-relaxation time during readout. Reading it: the $0\to1$ error has no analogue of the second term, which is why $\varepsilon_1>\varepsilon_0$ . Given a calibrated $A$ , the naive undo is the exact inverse. It is unbiased and it is a trap: on finite statistics it amplifies noise and can leave the simplex.

\hat{\mathbf{p}}_{\text{true}} = A^{-1}\,\mathbf{p}_{\text{meas}} \qquad (5)

where: $A^{-1}$ — matrix inverse, exists iff $\det A\neq0$ ; $\hat{\mathbf{p}}_{\text{true}}$ — unbiased but possibly unphysical (negative-entry) estimate. Reading it: correct on average, but its variance is set by how close $A$ is to singular. The fix keeps the same linear model but enforces physicality, solving a constrained least-squares problem instead of raw inversion.

\hat{\mathbf{p}}_{\text{true}} = \arg\min_{\mathbf{q}\ge0,\ \mathbf{1}^\top\mathbf{q}=1}\ \big\lVert A\mathbf{q}-\mathbf{p}_{\text{meas}}\big\rVert_2^2 \qquad (6)

where: $\mathbf{q}$ — candidate true distribution restricted to the simplex; $\mathbf{1}$ — all-ones vector enforcing normalization; $\lVert\cdot\rVert_2$ — Euclidean norm of the model residual. Reading it: find the physical distribution whose blurred image best matches what you saw — a small bias for a large variance reduction. When the observable is a single Pauli $Z$ , the correction reduces to one scalar formula, from inverting the $2\times2$ model at the expectation-value level. (Verify the sign in Worked example 1 — an easy place to slip.)

\langle Z\rangle_{\text{mit}} = \frac{\langle Z\rangle_{\text{meas}} + (\varepsilon_0-\varepsilon_1)}{1-\varepsilon_0-\varepsilon_1} \qquad (7)

where: $\langle Z\rangle_{\text{meas}}$ — raw measured expectation of $Z=|0\rangle\langle0|-|1\rangle\langle1|$ ; $\langle Z\rangle_{\text{mit}}$ — readout-corrected estimate; $1-\varepsilon_0-\varepsilon_1$ is the readout contrast. Reading it: readout error shrinks $\langle Z\rangle$ by the contrast and offsets it by $-(\varepsilon_0-\varepsilon_1)$ ; dividing back out restores the true value, provided the contrast is nonzero. The number hardware teams actually report is the average assignment fidelity, a compact summary of both rates.

F_{\text{ro}} = 1-\tfrac12\big(\varepsilon_0+\varepsilon_1\big) \qquad (8)

where: $F_{\text{ro}}$ — average single-shot assignment fidelity, the standard figure of merit; $\tfrac12(\varepsilon_0+\varepsilon_1)$ — the mean assignment error over the two prepared inputs. Reading it: " $99\%$ readout" means $\varepsilon_0+\varepsilon_1\approx0.02$ , not that either rate alone is $1\%$ .

Derivation AReadout error as a noisy two-outcome POVM

The goal is to show that the classical confusion matrix is the exact shadow of a genuine quantum measurement — a two-outcome POVM whose effects are noisy mixtures of the ideal projectors. This pins down the article's central point: readout error is a quantum-to-classical map from a density matrix to a distribution, and therefore not a CPTP map on a surviving state, which is why it composes with amplitude damping rather than being one.

Step 1. Start from the ideal projective measurement in the computational basis, with projectors $\Pi_0=|0\rangle\langle0|$ , $\Pi_1=|1\rangle\langle1|$ , $\Pi_0+\Pi_1=\mathbb{I}$ , giving $p^{\text{true}}_j=\mathrm{Tr}(\Pi_j\rho)$ . Why: this is the noiseless reference the confusion matrix must reduce to when the noise vanishes.

Step 2. Model the imperfect detector by replacing the sharp projectors with noisy effects $E_0=(1-\varepsilon_0)\Pi_0+\varepsilon_1\Pi_1$ and $E_1=\varepsilon_0\Pi_0+(1-\varepsilon_1)\Pi_1$ . Each is positive and $E_0+E_1=\mathbb{I}$ , so $\{E_0,E_1\}$ is a valid POVM. Why: a real detector cannot resolve the projectors cleanly, so its "yes" for $0$ keeps some weight on the true $|1\rangle$ — the off-diagonal leakage the effects encode.

Step 3. Apply the Born rule for POVMs, $p^{\text{meas}}_i=\mathrm{Tr}(E_i\rho)$ , and expand: $p^{\text{meas}}_i=\sum_j A_{ij}\,\mathrm{Tr}(\Pi_j\rho)=\sum_j A_{ij}\,p^{\text{true}}_j$ , with $A_{ij}$ the weights from Step 2. Why: linearity of the trace turns the operator statement into the matrix statement $\mathbf{p}_{\text{meas}}=A\mathbf{p}_{\text{true}}$ — the confusion matrix is the coordinate form of the POVM in the projector basis.

Step 4. Note the physical status: the input is a density matrix $\rho$ , but the output lives on the classical register as a distribution over labels, with no post-measurement quantum state. Why: this separates readout error from amplitude damping (a CPTP map $\rho\mapsto\rho'$ acting before measurement) — it is the measurement's own imperfection and no quantum recovery can undo it, the information being already gone. *Verified: $E_0,E_1\ge0$ and $E_0+E_1=\mathbb{I}$ by construction, and each column of $A$ sums to one — column-stochastic as required (Nielsen & Chuang §2.2.6, Watrous Ch. 2).*

Go deeper · Derivation B · 1 minWhy inverting the confusion matrix amplifies noise

Derivation BWhy inverting the confusion matrix amplifies noise

The goal is to quantify the danger in equation (5): the exact inverse $A^{-1}$ is unbiased, yet on finite shots it can be worse than doing nothing and can return negative probabilities. Propagating the multinomial shot-noise covariance through the linear inverse shows the blow-up is set by the condition number of $A$ , which diverges as the readout loses contrast — the concrete case for the constrained fit of equation (6).

Step 1. For $N$ shots the histogram is multinomial, $\mathrm{Cov}(\mathbf{p}_{\text{meas}})_{ii'}=\tfrac{1}{N}\big(p^{\text{meas}}_i\delta_{ii'}-p^{\text{meas}}_i p^{\text{meas}}_{i'}\big)$ , with entries scaling as $1/N$ . Why: sampling a fixed distribution $N$ times gives variance $\sim1/N$ — the raw noise we push through the inverse.

Step 2. The estimator $\hat{\mathbf{p}}_{\text{true}}=A^{-1}\mathbf{p}_{\text{meas}}$ is linear, so $\mathrm{Cov}(\hat{\mathbf{p}}_{\text{true}})=A^{-1}\,\mathrm{Cov}(\mathbf{p}_{\text{meas}})\,(A^{-1})^\top$ . Why: standard linear propagation, a map $M$ sends $\Sigma$ to $M\Sigma M^\top$ , and here $M=A^{-1}$ has a potentially tiny smallest singular value.

Step 3. Bound the total variance by the condition number: $\mathrm{tr}\,\mathrm{Cov}(\hat{\mathbf{p}}_{\text{true}})\lesssim \kappa(A)^2/N$ , with $\kappa(A)=\lVert A\rVert\,\lVert A^{-1}\rVert$ . As $\varepsilon_0+\varepsilon_1\to1$ , $\det A=1-\varepsilon_0-\varepsilon_1\to0$ and $\kappa\to\infty$ . Why: the inverse stretches noise most along $A$ 's smallest singular value, so a near-singular matrix magnifies shot noise catastrophically — still "reading", but barely constraining the truth.

Step 4. For a symmetric flip $p$ , $A^{-1}=\tfrac{1}{1-2p}\begin{pmatrix}1-p&-p\\-p&1-p\end{pmatrix}$ and $\kappa\sim1/(1-2p)$ ; as $p\to\tfrac12$ mitigation is hopeless. Why: constrained least-squares sacrifices a little bias to project onto the simplex, trading the divergent variance of raw inversion for a controlled, physical answer — the production default over equation (5). *Verified: the eigenvalues of the symmetric $A$ are $1$ and $1-2p$ , so $\kappa=1/(1-2p)$ diverges as $p\to\tfrac12$ ; at $p=0.1$ , $\kappa=1.25$ , matching Worked example 2.*

Go deeper · Derivation C · 1 minTensored mitigation for n uncorrelated qubits

Derivation CTensored mitigation for n uncorrelated qubits

The goal is to defeat the exponential wall: the full confusion matrix for $n$ qubits is $2^n\times2^n$ , impossible to calibrate or store beyond a handful of qubits. If readout errors are uncorrelated across qubits — no crosstalk — the giant matrix factorizes into a Kronecker product of $2\times2$ blocks, so calibration and mitigation both scale linearly rather than exponentially in the qubit count.

Step 1. Assume independence: $P(i_1\cdots i_n\mid j_1\cdots j_n)=\prod_{k=1}^n P(i_k\mid j_k)$ . Why: crosstalk-free readout means each qubit's classifier sees only its own signal, so a whole-bitstring flip probability is the product of single-qubit flips — the content of "uncorrelated".

Step 2. The product of conditionals is a product of matrix entries, $A_{(i_1\cdots i_n),(j_1\cdots j_n)}=\prod_k A^{(k)}_{i_k j_k}$ , i.e. the Kronecker product $A=\bigotimes_{k=1}^n A^{(k)}$ . Why: the exponential matrix is fully specified by $n$ small $2\times2$ blocks — an exponential compression bought by independence.

Step 3. Each block $A^{(k)}$ needs only its two columns, from preparing $|0\rangle$ and $|1\rangle$ on qubit $k$ , so $2n$ experiments suffice instead of $2^n$ . Why: you never prepare a joint basis state; single-qubit calibrations run in parallel reconstruct every block.

Step 4. Invert factor-by-factor with $\big(\bigotimes_k M_k\big)^{-1}=\bigotimes_k M_k^{-1}$ , giving $\hat{\mathbf{p}}_{\text{true}}=\big(\bigotimes_k (A^{(k)})^{-1}\big)\mathbf{p}_{\text{meas}}$ . Why: mitigation never forms the $2^n\times2^n$ matrix — it applies $n$ small inverses, collapsing to the per-qubit $\langle Z\rangle$ correction of equation (7) when only Pauli expectations are wanted. *Verified: the Kronecker-inverse identity is standard; correlated readout violates the Step 1 factorization and forces calibration of off-diagonal crosstalk — the regime that motivates matrix-free methods like M3.*

Limiting cases

The confusion matrix interpolates between several regimes that tell you at a glance when mitigation is easy, hard, or impossible.

Case	Result	Meaning
Ideal readout	$A=\mathbb{I}_{2^n}$	No misassignment; $\mathbf{p}_{\text{meas}}=\mathbf{p}_{\text{true}}$ , mitigation is the identity.
Symmetric flip $p$	$A=\begin{pmatrix}1-p&p\\p&1-p\end{pmatrix}$	Equal $0\to1$ , $1\to0$ rates; invertible iff $p\neq\tfrac12$ , condition number $1/(1-2p)$ .
Relaxation-dominated	$A=\begin{pmatrix}1&1-e^{-\tau_m/T_1}\\0&e^{-\tau_m/T_1}\end{pmatrix}$	Pure $T_1$ decay: only $1\to0$ errors, $0$ read perfectly — the maximally asymmetric channel.
Uncorrelated $n$ qubits	$A=\bigotimes_{k=1}^n A^{(k)}$	Tensored model; $2n$ calibration states, $A^{-1}=\bigotimes_k (A^{(k)})^{-1}$ scales linearly.
Maximally confusing	$\det A=0$	Distinct true distributions map to one measured distribution; mitigation is impossible.

Common trapReadout error is just another gate error you can echo or invert away

It is not, on two counts. First, it is not a coherent or CPTP error on the qubit that a recovery unitary could reverse — it is a classical relabeling of an outcome that has already collapsed, so no pulse, echo, or quantum operation can undo it; the information is gone the moment the classifier fires. Second, "inverting" it does not mean applying $A^{-1}$ blindly: on finite shots that estimator amplifies noise by $\kappa(A)$ and returns physically meaningless negative "probabilities". The correct move is constrained least-squares (or iterative Bayesian unfolding) on the simplex. Treating readout error as a unitary defect, or trusting a raw pseudo-inverse, are the two most common mistakes — and near $\varepsilon_0+\varepsilon_1=1$ no post-processing helps, because the measurement has stopped carrying information.

Worked examples

1 · Correcting a single-qubit expectation value (fully worked)

A qubit has asymmetric readout errors $\varepsilon_0=0.01$ , $\varepsilon_1=0.05$ . Over many shots you record $\langle Z\rangle_{\text{meas}}=0.510$ . Report the assignment fidelity, recover the true $\langle Z\rangle$ , and check the sign.

Step 1. Assignment fidelity, equation (8): $F_{\text{ro}}=1-\tfrac12(0.06)=\mathbf{0.97}$ . The readout contrast is $1-\varepsilon_0-\varepsilon_1=\mathbf{0.94}$ — the factor by which raw readout shrinks every $\langle Z\rangle$ .

Step 2. Apply equation (7): $\langle Z\rangle_{\text{mit}}=\dfrac{0.510+(0.01-0.05)}{0.94}=\dfrac{0.470}{0.94}=\mathbf{0.500}$ . The raw $0.510$ was biased high, because the offset $-(\varepsilon_0-\varepsilon_1)=+0.04$ lifts the measured value when $\varepsilon_1>\varepsilon_0$ .

Step 3. Check by forward simulation: a true $\langle Z\rangle=0.5$ means $p_0=0.75,\ p_1=0.25$ , giving $p_0^{\text{meas}}=0.99(0.75)+0.05(0.25)=0.755$ and $p_1^{\text{meas}}=0.01(0.75)+0.95(0.25)=0.245$ , so $\langle Z\rangle_{\text{meas}}=0.510$ . The mitigation exactly inverts the forward map — and the $+(\varepsilon_0-\varepsilon_1)$ sign is essential: the $-$ sign one often sees would give $0.585$ , off by $17\%$ .

2 · The peril of raw inversion (negative probabilities)

A qubit has symmetric readout error $p=0.10$ . A finite run measures $\mathbf{p}_{\text{meas}}=(0.08,\,0.92)$ . Apply the exact inverse, interpret the result, and state what constrained least-squares does instead.

Step 1. Here $A=\begin{pmatrix}0.9&0.1\\0.1&0.9\end{pmatrix}$ , $\det A=0.8$ , so $A^{-1}=\begin{pmatrix}1.125&-0.125\\-0.125&1.125\end{pmatrix}$ and $\kappa(A)=1/(1-2p)=\mathbf{1.25}$ , a modest but nonzero amplification.

Step 2. Apply equation (5): $\hat p_0=1.125(0.08)-0.125(0.92)=\mathbf{-0.025}$ and $\hat p_1=-0.125(0.08)+1.125(0.92)=\mathbf{1.025}$ . The entries sum to one, but $\hat p_0=-0.025<0$ is not a probability — raw inversion left the simplex on an ordinary fluctuation.

Step 3. Equation (6) solves $\min_{\mathbf q\ge0,\ \mathbf 1^\top\mathbf q=1}\lVert A\mathbf q-\mathbf p_{\text{meas}}\rVert_2^2$ , clipping to the nearest physical distribution, here $\hat{\mathbf p}_{\text{true}}=(0,\,1)$ . A small bias buys a physical, bounded-variance estimate — why production stacks never ship a bare pseudo-inverse.

3 · Budgeting the readout error against T₁

A transmon has $T_1=100\ \mu\mathrm{s}$ and a discrimination error $\varepsilon_1^{\text{disc}}=0.40\,e^{-\tau_m/0.5\,\mu\mathrm{s}}$ . Find the $1\to0$ error at $\tau_m=1\ \mu\mathrm{s}$ and the window that minimizes it.

Step 1. At $\tau_m=1\ \mu\mathrm{s}$ : decay term $1-e^{-0.01}=\mathbf{0.00995}$ , discrimination term $0.40\,e^{-2}=\mathbf{0.0541}$ , so by equation (4) $\varepsilon_1\approx\mathbf{0.064}$ — still discrimination-limited, so lengthening the window helps.

Step 2. Minimizing $\varepsilon_1(\tau_m)=0.40\,e^{-\tau_m/0.5}+(1-e^{-\tau_m/100})$ sets $\tfrac{0.40}{0.5}e^{-\tau_m/0.5}=\tfrac{1}{100}e^{-\tau_m/100}$ , giving $\tau_m\approx\mathbf{2.2\ \mu s}$ and $\varepsilon_1\approx 0.0049+0.0218=\mathbf{0.027}$ . Beyond this, discrimination is already tiny and further integration only feeds $T_1$ decay — the optimum is decay-limited, exactly the minimum in the second figure.

Physics of the portrait

Readout error is a modern engineering concern, but the questions it forces — what a measurement is, what it leaves behind, how imperfect it can be while still being a measurement — were settled in their essentials nearly a century ago by John von Neumann, whose formalism draws the very line between "quantum evolution" and "classical outcome" that readout error lives on.

Origins · 1 minJohn von Neumann and how this was discovered

John von Neumann

1903–1957 · the mathematical foundations of quantum measurement

In Mathematische Grundlagen der Quantenmechanik (1932) von Neumann gave quantum theory its operator-theoretic backbone: states as density operators, observables as self-adjoint operators, and — crucially for us — two utterly different dynamics. His "process 2" is smooth, reversible, unitary Schrödinger evolution; his "process 1" is measurement, the abrupt, irreversible, non-unitary projection onto an eigenstate yielding a classical outcome with Born-rule probability. Readout error is a controlled failure of an idealized process 1: sharp projectors become fuzzy effects, and the crisp outcome a stochastically relabeled one. He also introduced the density matrix and the entropy $S(\rho)=-\mathrm{Tr}\,\rho\log\rho$ that anchor quantum information, and his measurement "cut" between quantum system and classical apparatus is exactly where the confusion matrix sits. The generalization from projective measurements to the POVMs describing real detectors came later — Naimark's dilation theorem and Kraus's formalism of effects and operations — but all of it is written in the language he built. Calibrate a two-outcome POVM with imperfect effects and read its confusion matrix off the columns, and you are working inside his framework three-quarters of a century on.

History

1932

— von Neumann's Mathematische Grundlagen formalizes projective measurement, the density operator, and the quantum entropy

1940

— Naimark's dilation theorem shows every POVM is a projective measurement on a larger space

1970

— Davies and Lewis axiomatize generalized measurements as instruments and POVMs

1983

— Kraus's States, Effects, and Operations codifies effects and completely positive operations

2017

— matrix-inversion readout mitigation becomes standard practice on cloud superconducting processors

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.2.3–2.2.6 (measurements, POVMs) · Preskill, Lecture Notes on Quantum Computation, Ch. 3 · Watrous, The Theory of Quantum Information, Ch. 2 · Wilde, Quantum Information Theory, Ch. 4 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, §B · Mermin, Quantum Computer Science, Ch. 1.

Provenance. All equations are standard and recomputable. A verifier should: (1) confirm $\sum_i A_{ij}=1$ and $E_0+E_1=\mathbb{I}$ (Nielsen & Chuang §2.2.6, Watrous Ch. 2); (2) reproduce the corrected sign $\langle Z\rangle_{\text{mit}}=(\langle Z\rangle_{\text{meas}}+(\varepsilon_0-\varepsilon_1))/(1-\varepsilon_0-\varepsilon_1)$ , checked in Worked example 1 against forward simulation ( $\varepsilon_0=0.01,\varepsilon_1=0.05$ , $\langle Z\rangle_{\text{meas}}=0.510\Rightarrow0.500$ ); (3) verify $A^{-1}=\tfrac{1}{1-2p}(\cdots)$ and $\kappa=1/(1-2p)$ , with $p=0.1$ giving $\hat p_0=-0.025$ ; (4) check the survival probability $e^{-\tau_m/T_1}$ and the $\tau_m\approx2.2\ \mu\mathrm{s}$ minimum for $T_1=100\ \mu\mathrm{s}$ . The $97$ – $99\%$ ranges are typical, not exact. The $T_1$ term links to aq-qci-amplitude-damping-channel and aq-qci-relaxation-dephasing-t1-t2; the POVM/CPTP framing to aq-qci-cptp-maps-quantum-operations and aq-qci-open-quantum-systems-reduced-dynamics.

Prerequisites. CPTP Maps and Quantum Operations · Amplitude Damping Channel · Relaxation and Dephasing: T₁ and T₂ · Open Quantum Systems and Reduced Dynamics. Leads to. Pauli and Bit/Phase-Flip Channels · Decoherence · Kraus Representation of Quantum Channels.

Ask the tutor: Want to see how the tensored confusion matrix and the constrained-least-squares fit actually run on a two-qubit histogram, step by step?

Is readout error the same thing as decoherence or amplitude damping?

No, though one feeds the other. Amplitude damping and decoherence are CPTP maps on $\rho$ during the computation — quantum states to quantum states, partly counteracted by error correction on the live qubit. Readout error is the measurement's own imperfection: it takes $\rho$ to a classical distribution and stochastically relabels it. The link is that the $T_1$ relaxation driving amplitude damping also acts during the measurement window, the exact source of the asymmetric $1\to0$ term. So amplitude damping contributes to readout error, but readout error is a broader, classical-register phenomenon, not a channel on any surviving state.

Why is the confusion matrix column-stochastic rather than row-stochastic?

Because each column answers a conditional with a fixed premise. Column $j$ is built by preparing $|j\rangle$ and histogramming what the detector reports; since it must report some outcome every shot, those conditionals sum to one, $\sum_i A_{ij}=1$ . Rows instead collect "all the ways to report outcome $i$ " across different true states, with no normalization forcing a sum. This is why the model is $\mathbf{p}_{\text{meas}}=A\,\mathbf{p}_{\text{true}}$ with the truth on the right, and column-stochasticity is exactly the statement that $A$ preserves total probability.

If mitigation just inverts a known matrix, why not always use A inverse directly?

Because the exact inverse is unbiased but statistically reckless. On finite shots your histogram is a noisy estimate of $\mathbf{p}_{\text{meas}}$ , and $A^{-1}$ amplifies that noise by roughly $\kappa(A)$ , which diverges as the contrast $1-\varepsilon_0-\varepsilon_1$ approaches zero. Worse, the inverse does not respect the simplex, so it returns negative "probabilities" on ordinary fluctuations, as Worked example 2 shows. The standard fix is constrained least-squares (or iterative Bayesian unfolding): find the physical distribution whose blurred image best matches the data, paying a small controlled bias for a bounded-variance, physical estimate.

How does readout error scale to many qubits without an exponential blow-up?

The full confusion matrix is $2^n\times2^n$ , intractable beyond about a dozen qubits. Two ideas rescue it. First, if readout errors are uncorrelated — no crosstalk — the matrix factorizes into a Kronecker product of single-qubit blocks, so you calibrate only $2n$ states and apply $n$ tiny inverses. Second, matrix-free methods (such as M3) exploit that only a handful of the $2^n$ bitstrings ever appear in a real dataset, solving the constrained problem within that sparse subspace by an iterative solver. Together these push mitigation to hundreds of qubits; when crosstalk is present, one adds the leading off-diagonal correlations as small correction blocks.

Does readout error have a preferred direction, and can that be exploited?

Yes on both counts. On superconducting hardware $\varepsilon_1$ is typically several times larger than $\varepsilon_0$ , because energy relaxation during the window only pushes population downhill — there is no comparable uphill process. This directionality is real physics, not a calibration artifact, and worth exploiting: biased decoders weight $1\to0$ syndrome faults more heavily to match the true noise, and expectation estimators use the asymmetric formula (7), whose $(\varepsilon_0-\varepsilon_1)$ offset encodes the bias. Assuming a symmetric channel on asymmetric hardware leaves accuracy on the table.

Where exactly does the "collapse" happen, and does readout error change the post-measurement state?

In the model used here, readout error acts entirely after the quantum information has become a classical outcome — a relabeling on the classical register, with no surviving quantum state. Real measurements are not always perfectly projective: a non-demolition readout can leave the qubit in an approximate eigenstate, and imperfections there (measurement-induced transitions, leakage, heating) are separate effects on top of the confusion matrix. For the standard end-of-circuit readout the statement is clean: the confusion matrix distorts the statistics you report, not any state you keep — which is why no quantum recovery can undo it and the correction is classical post-processing.