Gate Set Tomography

Every characterization of a quantum gate seems to demand a paradox. To learn what a gate does you feed it known states and read it out with known measurements — but on real hardware the states you prepare and the measurements you make are themselves flawed, and flawed by roughly as much as the gate you are trying to pin down. Calibrating the preparations needs trustworthy gates; calibrating those gates needs trustworthy preparations; and round the circle goes. Standard quantum process tomography cuts the knot by assuming perfect state preparation and measurement, which quietly stamps the very errors it ignores onto the gate it reports. Gate set tomography, GST, refuses that assumption: it treats the preparation, the whole family of gates, and the measurement as one jointly unknown object and demands nothing of the data except internal self-consistency. This is the deep-tier synthesis of the characterization thread, taking the CPTP structure of aq-qci-cptp-maps-quantum-operations, the operator-sum language of aq-qci-kraus-representation-q9, and the noise vocabulary of aq-qci-pauli-bit-phase-flip-channels and aq-qci-decoherence, and assembling from them the most rigorous self-calibrating protocol in the toolbox. Three ideas organize everything below. First, self-consistency has a price: the reconstruction is fixed only up to a gauge, an invisible similarity transformation that reshuffles the matrices without touching any measurable probability, so only gauge-invariant figures of merit are physical. Second, the estimate lives in the Pauli transfer matrix picture, where states are real vectors, measurements real covectors, and gates real matrices that multiply. Third, by repeating short "germ" circuits deep, GST amplifies coherent errors linearly in depth and reaches a precision that scales like 1/L — the hallmark of Heisenberg-limited estimation.

Key takeaways

Gate set tomography characterizes a complete set of operations — state preparation, gates, and measurement (SPAM) — self-consistently, treating every element as unknown rather than trusting a perfect reference frame.
The obstruction it must live with is the gauge freedom: a similarity transformation on all gate matrices, the state, and the effect leaves every observable probability invariant, so absolute process matrices carry no meaning.
Circuit outcome probabilities are bilinear, $p=\langle\langle E|R_{G_L}\cdots R_{G_1}|\rho\rangle\rangle$ , in the Pauli transfer matrix representation of CPTP maps.
Long-sequence GST repeats germ circuits to depth $L$ , amplifying a coherent per-gate error $\theta$ into a signal $\propto L\theta$ and reaching precision $\Delta\theta\sim 1/L$ , far past the $1/\sqrt{N}$ of naive tomography.
It reports gauge-invariant quantities only — fidelity, diamond norm, process eigenvalues — cleanly separating coherent from stochastic errors, and because it overfits a Markovian model its goodness-of-fit statistic doubles as a detector of non-Markovian, drifting, or context-dependent errors.
Key idea: treat preparation, gates, and measurement as jointly unknown, fit them to circuit frequencies up to an unobservable gauge similarity, report only gauge-invariant figures of merit, and use deep germ-power circuits to reach 1/L precision on coherent errors.

The core insight

Treat preparation, gates, and measurement as jointly unknown; fit them to circuit outcome frequencies up to an unobservable gauge similarity, then report only gauge-invariant figures of merit, using deep germ-power circuits to amplify coherent errors and reach $1/L$ estimation precision.

Where this lives today

Gate set tomography is the most rigorous single-tool answer to "how good, and how, is this gate?", and it runs on real hardware through the Sandia pyGSTi package, which trapped-ion, superconducting, and spin-qubit teams use to turn raw circuit counts into gauge-invariant error reports. Its distinctive output — a physically constrained gate set plus a diamond-norm bound and a coherent-versus-stochastic breakdown — lets a calibration engineer decide whether a bad gate needs a pulse-phase tweak (coherent, correctable) or a better qubit (stochastic, not). It is prized for what it adds to cheaper tools: randomized benchmarking returns one error rate fast but averages away structure, and the depolarizing rate GST extracts should match the RB decay constant — when they disagree, GST exposes the gate-dependent or non-Markovian error RB could not see. That model-violation test is a first-class result: a large goodness-of-fit statistic is a principled alarm for drift, crosstalk, leakage, or bath memory, tying device characterization to the study of open and non-Markovian systems. The main frontier is scale — circuit count and fit cost grow steeply with qubit number — so full GST remains mostly a one- and two-qubit instrument, with active work on scalable and crosstalk-aware variants for the many-qubit devices fault tolerance will demand.

The physical picture

Imagine you must measure the lengths of several rulers with no trusted reference — no metre stick, only the ability to lay one ruler against another. You can determine every relative length perfectly: this ruler is 1.07 times that one, that one 0.94 times a third. What you can never fix is a global rescaling of the whole set, because multiplying every ruler by the same factor changes no measured ratio. Gate set tomography is exactly this predicament. The unmeasurable rescaling — really a similarity transformation of the operational frame — is the gauge, and only relationships among gates, states, and effects are physical. A raw process matrix out of the fit is like a single ruler's "absolute" length: no operational content until you fix the frame. Now watch a small error grow. A coherent over-rotation is like a constant clock drift: a tiny per-gate angle $\theta$ is invisible in one shot, but applied $L$ times it accumulates to $L\theta$ , a large, easily read rotation. GST is built to ride that amplification, layering short germ circuits — sequences whose repetition amplifies one kind of error — between preparation and measurement fiducial subcircuits, and repeating each germ many times. The first figure shows the anatomy of one such circuit and the amplification it buys: a 0.02 rad over-rotation, hopeless to see directly, becomes a 1.28 rad swing after only 64 repetitions.

Figure

A gate set tomography circuit sandwiches a germ circuit repeated L times between preparation and measurement fiducials, so a per-gate over-rotation θ = 0.02 rad accumulates coherently to Lθ = 1.28 rad ≈ 73° after L = 64 repetitions — the depth L multiplies the circuit's sensitivity to the error

Not every error amplifies the same way. A coherent error — an over-rotation, an axis tilt, a miscalibrated phase — rotates the signal, so its imprint grows linearly with depth and can be driven to arbitrarily small uncertainty by going deeper. A stochastic error — depolarization, dephasing, energy loss — instead damps the signal as $e^{-\gamma L}$ , capping how deep it is worth going. GST is built to tell these apart, because one is a calibration mistake you can fix in software and the other is decoherence you cannot. The second figure is the payoff: the precision on a coherent angle falls like 1/L along the germ-power staircase, plunging below the flat shot-noise floor that depth-blind tomography can never leave.

The core equations

The whole machine runs in one representation, so start there. The Pauli transfer matrix (PTM) picture expands every state, effect, and channel in the normalized Pauli basis, turning density matrices into real vectors and channels into real matrices. This is what makes "compose two gates" become "multiply two matrices."

|\rho\rangle\rangle_n=\mathrm{Tr}(P_n\,\rho),\qquad \langle\langle E|_n=\mathrm{Tr}(P_n\,E),\qquad (R_G)_{mn}=\tfrac{1}{d}\,\mathrm{Tr}\!\left[P_m\,\mathcal{G}(P_n)\right] \qquad (1)

where: $P_n$ — normalized Pauli basis, $n=0,\dots,d^2-1$ , with $P_0=I$ ; $d$ — Hilbert-space dimension ( $d=2^q$ for $q$ qubits); $|\rho\rangle\rangle$ — real Hilbert–Schmidt vector encoding the prepared state; $\langle\langle E|$ — covector encoding a POVM effect; $R_G$ — real $d^2\times d^2$ transfer matrix of the channel $\mathcal{G}$ . Reading it: the Born rule $\mathrm{Tr}(AB)$ becomes an inner product $\langle\langle A|B\rangle\rangle$ , and channel composition becomes ordinary matrix multiplication. Every GST experiment is a circuit predicting one number, an outcome probability. In the PTM picture that number is a bilinear form — covector, product of matrices, vector — the structural fact from which everything else, including the gauge, follows.

p_{i,k,j}=\langle\langle E_i|\,R_{G_{j_L}}\cdots R_{G_{j_2}}R_{G_{j_1}}\,|\rho_k\rangle\rangle \qquad (2)

where: $p_{i,k,j}$ — probability of effect $i$ given preparation $k$ and gate sequence $j=(j_1,\dots,j_L)$ ; $R_{G_{j_\ell}}$ — transfer matrix of the $\ell$ -th gate; $L$ — circuit depth. Because the unknowns ( $|\rho\rangle\rangle$ , the $R_{G}$ , and $\langle\langle E|$ ) appear as a product, the estimation problem is fundamentally nonlinear, and the freedom to insert an invertible factor between any two of them is the gauge. That freedom cannot be removed: any invertible $B$ acting on all elements at once produces an indistinguishable gate set. This is the defining feature of GST and why it never reports a raw process matrix.

R_{G_j}\;\mapsto\;B\,R_{G_j}\,B^{-1},\qquad |\rho\rangle\rangle\mapsto B|\rho\rangle\rangle,\qquad \langle\langle E|\mapsto\langle\langle E|B^{-1} \qquad (3)

where: $B$ — any invertible $d^2\times d^2$ matrix (the gauge), constrained by $B|I\rangle\rangle=|I\rangle\rangle$ so trace is preserved; $R_{G_j}$ — estimated gate transfer matrix. The transformation is a similarity on the gates flanked by matched $B$ on the state and $B^{-1}$ on the effect — precisely the pattern that will telescope away inside equation (2). To turn data into an estimate, the earliest and simplest method is linear inversion. Choose informationally complete fiducials, collect the empty-circuit overlaps and the single-gate-sandwiched probabilities, and read the gate off by a matrix inverse.

g_{ik}=\langle\langle E_i|\rho_k\rangle\rangle,\qquad (M_j)_{ik}=\langle\langle E_i|R_{G_j}|\rho_k\rangle\rangle,\qquad \tilde{R}_{G_j}=g^{-1}\,M_j \qquad (4)

where: $g$ — Gram matrix of fiducial overlaps (preparation $k$ read by effect $i$ ); $M_j$ — single-gate probabilities; $\tilde{R}_{G_j}$ — linear-inversion estimate of gate $j$ in the fiducials' natural gauge. The empty circuit gives $g$ , depth-one circuits give $M_j$ , and $g^{-1}M_j$ cancels the unknown fiducial content — up to gauge, only approximately CPTP, so it seeds a nonlinear physical fit. The gauge-fixed estimate is then compared to the ideal target through gauge-invariant figures of merit. The workhorse is the average gate fidelity, tied to the process (entanglement) fidelity — a normalized transfer-matrix overlap.

F_{\mathrm{avg}}(\mathcal{G},\mathcal{U})=\frac{d\,F_{\mathrm{pro}}+1}{d+1},\qquad F_{\mathrm{pro}}=\frac{\mathrm{Tr}\!\left[R_{\mathcal{U}}^{\top}R_{\mathcal{G}}\right]}{d^2} \qquad (5)

where: $\mathcal{U}$ — ideal target unitary channel; $F_{\mathrm{pro}}$ — process fidelity, a normalized overlap of transfer matrices (with $R_{\mathcal{U}}$ real orthogonal, $R_{\mathcal{U}}^{\top}=R_{\mathcal{U}}^{-1}$ ); $F_{\mathrm{avg}}$ — average gate fidelity, the uniform-input average of the state fidelity. The infidelity is $r=1-F_{\mathrm{avg}}=\tfrac{d}{d+1}(1-F_{\mathrm{pro}})$ . Fidelity is an average, and averages hide worst cases. The diamond norm supplies the worst-case operational distance — the quantity that actually enters fault-tolerance threshold theorems — and the eigenvalues of the error transfer matrix split that distance into its coherent and stochastic parts.

\tfrac{1}{2}\|\mathcal{G}-\mathcal{U}\|_{\diamond}=\max_{\rho}\tfrac{1}{2}\bigl\|(\mathcal{G}\otimes\mathcal{I})\rho-(\mathcal{U}\otimes\mathcal{I})\rho\bigr\|_1,\qquad \lambda_k(R_{\mathcal{E}})=|\lambda_k|\,e^{i\theta_k} \qquad (6)

where: $\|\cdot\|_{\diamond}$ — diamond (completely bounded trace) norm, maximized over states of the system plus an equal-dimension ancilla; $\mathcal{I}$ — identity channel on the ancilla; $R_{\mathcal{E}}=R_{\mathcal{G}}R_{\mathcal{U}}^{-1}$ — error transfer matrix; $\theta_k$ — eigenphases (coherent over-rotations), $|\lambda_k|<1$ — eigenvalue magnitudes below one (stochastic decay). For a purely coherent single-qubit over-rotation by $\theta$ , $\tfrac12\|\cdot\|_\diamond=\sin(\theta/2)\approx\theta/2$ , while the infidelity is only $r\approx\theta^2/6$ — the two differ by a factor $\sim 3/\theta$ .

Derivation AWhy the gauge is invisible and only invariants are physical

The goal is to prove the single most important structural fact about gate set tomography: the gauge transformation of equation (3) leaves every measurable circuit probability exactly unchanged, so no experiment can ever fix the gauge and any figure of merit must be gauge-invariant or computed only after a gauge is explicitly imposed. We take a generic depth- $L$ circuit, apply the gauge to each factor, and watch the inserted matrices annihilate by telescoping — the cancellation that makes "convert to inches, convert back to centimetres" leave a length untouched.

Step 1. In the probability of equation (2), replace every element by its gauge image from (3): the effect becomes $\langle\langle E_i|B^{-1}$ , each gate $B R_{G_{j_\ell}}B^{-1}$ , the state $B|\rho_k\rangle\rangle$ . Why: the gauge acts on all objects at once, so this is not a new experiment but the same one in a rescaled frame.

Step 2. Concatenate: $\langle\langle E_i|B^{-1}\,(B R_{G_{j_L}}B^{-1})\cdots(B R_{G_{j_1}}B^{-1})\,B|\rho_k\rangle\rangle$ . Between every adjacent pair sits $B^{-1}B=\mathbb{1}$ , and the ends cancel too. Why: the inserted factors pair off like nested parentheses and every internal pair collapses to the identity — the telescoping.

Step 3. What survives is $\langle\langle E_i|R_{G_{j_L}}\cdots R_{G_{j_1}}|\rho_k\rangle\rangle=p_{i,k,j}$ , for every invertible $B$ with $B|I\rangle\rangle=|I\rangle\rangle$ . Why: the data depend only on gauge-invariant combinations, so a fit determines the gate set no finer than the orbit of $B$ — the absolute matrices are unobservable in principle.

Step 4. Consequence: because $F_{\mathrm{avg}}$ and $\|\cdot\|_\diamond$ compare the estimate to a fixed target they change under $B$ , so one first chooses the $B$ bringing the estimate closest to the target under CPTP constraints. Why: skipping this gauge optimization is the classic GST blunder — an adversarial $B$ can make the same gate set look perfect or broken. *Verified: pure similarity algebra; $B|I\rangle\rangle=|I\rangle\rangle$ keeps states normalized, and the telescoping holds at any depth.*

Go deeper · Derivation B · 1 minHeisenberg-limited estimation of a coherent over-rotation

Derivation BHeisenberg-limited estimation of a coherent over-rotation

This derivation is the metrological heart of long-sequence GST: why repeating a germ to depth $L$ makes the uncertainty on a coherent error scale as $1/L$ rather than the $1/\sqrt{N}$ of ordinary tomography, and why decoherence caps how deep it is worth going. We model a gate meant to rotate by $\phi$ about $z$ but actually rotating by $\phi+\theta$ , run it $L$ times between fiducials that turn the accumulated phase into an oscillating probability, and propagate binomial shot noise through the measured slope.

Step 1. Repeat the germ $L$ times; each application adds $\theta$ , so the sequence implements a rotation by $L\phi+L\theta$ . Why: coherent errors add in amplitude, so $L$ copies multiply the signal by $L$ — a systematic drift is easiest to catch after many ticks.

Step 2. With suitable fiducials the outcome probability is $p_{\uparrow}(L)=\tfrac12[1+\cos(L\phi+L\theta+\varphi_0)]$ , oscillating in $\theta$ at frequency $L$ ( $\varphi_0$ absorbs fiducial phases). Why: the fiducials turn an inaccessible angle into an accessible population, and depth $L$ becomes the fringe frequency in $\theta$ .

Step 3. Differentiate: $\partial p_{\uparrow}/\partial\theta=-\tfrac{L}{2}\sin(L\phi+L\theta+\varphi_0)$ , whose magnitude grows linearly in $L$ . Why: a steeper response carries more information per shot — the slope is the amplification made quantitative.

Step 4. Propagate binomial noise: $\Delta\hat\theta=\sqrt{p_{\uparrow}(1-p_{\uparrow})/N}\big/|\partial p_{\uparrow}/\partial\theta|\sim 1/(L\sqrt{N})$ , while decoherence damps the fringe as $e^{-\gamma L}$ , saturating the useful depth near $L\sim1/\gamma$ . Why: a fixed shot-noise numerator over an $O(L)$ slope gives $1/L$ scaling until contrast loss at depth $1/\gamma$ stops the gain — the trade-off in the second figure. *Verified: elementary algebra; at $N=1000$ , $\Delta\hat\theta=0.032$ at $L=1$ and $4.9\times10^{-4}$ at $L=64$ , matching the plotted line.*

Go deeper · Derivation C · 1 minReading coherent and stochastic error out of the process eigenvalues

Derivation CReading coherent and stochastic error out of the process eigenvalues

The aim is to show that the eigenvalues of the estimated gate's transfer matrix cleanly separate the two physically distinct kinds of error — coherent over-rotations, which turn eigenvalues in phase, and stochastic decoherence, which shrinks their magnitude — and to derive the leading-order infidelity each contributes. We factor the estimate into the ideal target composed with an error channel, diagonalize the error, and expand the fidelity to second order. The coefficients matter: a widely-copied version overcounts them by a factor of $d$ .

Step 1. Factor out the target. Since $R_{\mathcal{U}}$ is real orthogonal, $R_{\mathcal{E}}=R_{\mathcal{G}}R_{\mathcal{U}}^{-1}$ isolates the error, and $F_{\mathrm{pro}}=\mathrm{Tr}[R_{\mathcal{U}}^{\top}R_{\mathcal{G}}]/d^2=\mathrm{Tr}\,R_{\mathcal{E}}/d^2$ . Why: comparing to the ideal is the same as looking at the residual channel, whose deviation from the identity is the whole error.

Step 2. Diagonalize, $\lambda_k(R_{\mathcal{E}})=|\lambda_k|e^{i\theta_k}$ : eigenphases $\theta_k\neq0$ are coherent rotations, magnitudes $|\lambda_k|<1$ are stochastic decay. Why: a unitary error keeps eigenvalues on the unit circle and only turns them; an incoherent error pulls them inward — phase versus radius is the split.

Step 3. Expand $F_{\mathrm{pro}}=\tfrac{1}{d^2}\sum_k|\lambda_k|\cos\theta_k$ , so to second order $1-F_{\mathrm{pro}}\approx\tfrac{1}{2d^2}\sum_k\theta_k^2+\tfrac{1}{d^2}\sum_k(1-|\lambda_k|)$ . Why: $\cos\theta_k\approx1-\theta_k^2/2$ makes the coherent piece quadratic while $1-|\lambda_k|$ enters linearly — why over-rotations are quiet in infidelity yet loud in the worst case.

Step 4. With $r=\tfrac{d}{d+1}(1-F_{\mathrm{pro}})$ : $r_{\mathrm{coh}}\approx\tfrac{1}{2d(d+1)}\sum_k\theta_k^2$ and $r_{\mathrm{sto}}\approx\tfrac{1}{d(d+1)}\sum_k(1-|\lambda_k|)$ . Why: at $d=2$ a $\theta$ over-rotation gives $r_{\mathrm{coh}}\approx\theta^2/6$ and depolarizing strength $p$ gives $r_{\mathrm{sto}}=p/2$ — classic results reproduced only by the $1/d(d+1)$ normalization, not $1/(d+1)$ . *Verified: the $d=2$ over-rotation gives $r=\theta^2/6$ ( $\theta=0.02\Rightarrow6.7\times10^{-5}$ ) and depolarizing gives $r=p/2$ , both matching the boxed coefficients.*

Limiting cases

Case	Result	Meaning
Single gate, perfect SPAM	GST $\to$ standard process tomography	with trusted $\|\rho\rangle\rangle,\langle\langle E\|$ the self-consistency problem vanishes and GST reduces to ordinary QPT of one channel.
Short circuits only ( $L=1$ )	$\Delta\hat\theta\sim 1/\sqrt{N}$	linear GST is only shot-noise limited; the metrological advantage needs long germ-power sequences.
Long-sequence regime	$\Delta\hat\theta\sim 1/L$	germ repetition amplifies coherent errors, so precision improves linearly with maximum depth (Heisenberg-like).
Purely coherent error	$\tfrac12\\|\cdot\\|_\diamond\approx\theta/2\gg r\approx\theta^2/6$	worst-case (diamond) error dwarfs the average infidelity; over-rotations hide from fidelity.
Non-Markovian / drifting gates	$\chi^2_{\mathrm{obs}}\gg\nu$	model-violation tests reject the time-stationary CPTP model; the reconstruction becomes untrustworthy.
Gate-dependent SPAM	$\langle\langle E_i\|\to\langle\langle E_i^{(G)}\|$	GST assumes fixed preparation and measurement; SPAM correlated with the applied gate biases the fit and must be modeled.

Common trapJust gauge-transform the estimate until it matches, and read off the process matrix

Two traps hide in that sentence. First, you may not compare a raw estimate to an ideal gate without a constrained gauge optimization: the gauge is a genuine similarity orbit, and an adversarial $B$ can make a broken gate look perfect or a perfect one broken. The right procedure minimizes distance to the target subject to the estimate staying CPTP, once for the whole gate set. Second, there is no such thing as "the" process matrix of a single gate in GST — the absolute matrix is unobservable, so every number you quote must be gauge-invariant (eigenvalues, unitarity, entanglement fidelity after gauge-fixing) or defined relative to a fixed target. Treating a gauge-dependent matrix element as physical is the most common way to misread a GST report.

Worked examples

1 · A coherent over-rotation: infidelity versus worst case (fully worked)

GST reconstructs a single-qubit gate and, after gauge optimization, finds the error channel is a pure coherent over-rotation of $\theta=0.02$ rad ( $\approx 1.15°$ ) about a fixed axis, with no measurable stochastic decay. Report the process fidelity, the average gate infidelity, and the diamond-norm error, and say which one a fault-tolerance budget should use.

Step 1. Process fidelity. A unitary over-rotation by $\theta$ has $F_{\mathrm{pro}}=\cos^2(\theta/2)$ , so $F_{\mathrm{pro}}=\cos^2(0.01)=\mathbf{0.99990}$ and $1-F_{\mathrm{pro}}=1.00\times10^{-4}$ . The error is entirely in the eigenphases — the two nontrivial eigenvalues sit at $e^{\pm i\theta}$ on the unit circle, none pulled inward, confirming a purely coherent error.

Step 2. Average gate infidelity. With $r=\tfrac{d}{d+1}(1-F_{\mathrm{pro}})$ and $d=2$ : $F_{\mathrm{avg}}=(2\cdot0.99990+1)/3=\mathbf{0.999933}$ , so $r=\mathbf{6.67\times10^{-5}}$ , matching the small-angle rule $r\approx\theta^2/6=6.67\times10^{-5}$ . This is the number RB would report — reassuringly tiny.

Step 3. Diamond-norm error. For a coherent rotation the worst-case distance is $\tfrac12\|\mathcal{G}-\mathcal{U}\|_\diamond=\sin(\theta/2)=\sin(0.01)=\mathbf{1.00\times10^{-2}}$ . That is $\mathbf{150}$ times the average infidelity — exactly the factor $3/\theta$ from comparing $\theta/2$ with $\theta^2/6$ .

Step 4. Which to trust. Threshold theorems are stated in the diamond norm, so a fault-tolerance budget must use $\tfrac12\|\cdot\|_\diamond=10^{-2}$ , not the 150-times-smaller infidelity. A gate that RB calls " $10^{-4}$ good" can carry a worst-case error of $10^{-2}$ — right at a typical surface-code threshold — precisely because coherent errors accumulate as $\sqrt{r}$ , and separating them out is what GST is for.

2 · How deep must GST go to resolve that over-rotation?

Take the same $\theta=0.02$ rad over-rotation and $N=1000$ shots per circuit. Compare the precision of depth-one tomography with long-sequence GST at germ depth $L=64$ , and check the estimate is significant.

Step 1. Depth-one precision. At $L=1$ the fringe slope is at most $1/2$ , so $\Delta\hat\theta\sim 1/\sqrt{N}=1/\sqrt{1000}=\mathbf{0.032}$ rad. Larger than the $0.02$ rad signal — depth-one tomography cannot even confirm the error exists.

Step 2. Long-sequence precision. Depth $L=64$ multiplies the slope by $L$ , so $\Delta\hat\theta\sim 1/(L\sqrt{N})=1/(64\sqrt{1000})=\mathbf{4.9\times10^{-4}}$ rad. A single depth increase buys a $64\times$ precision gain at the same shots per circuit.

Step 3. Significance and the decoherence ceiling. The estimate now stands at $\theta/\Delta\hat\theta\approx\mathbf{41}$ standard deviations — an unambiguous detection. At stochastic rate $\gamma\approx10^{-3}$ the fringe contrast $e^{-\gamma L}$ is still $0.94$ at $L=64$ , safely below the ceiling $L\sim1/\gamma\approx10^3$ . The payoff is real but bounded: precision falls as $1/L$ until decoherence flattens it near $L\approx1/\gamma$ .

Physics of the portrait

Gate set tomography is less the work of one person than the resolution of a paradox that undermined quantum characterization for two decades — calibrating a ruler with no reference ruler — so its portrait is best drawn as the idea itself, with the people who forced it into rigorous form.

Origins · 1 mingate set tomography and how this was discovered

gate set tomography

self-consistent estimation of a full gate set up to gauge · quantum device characterization, verification, and validation

Gate set tomography began as an answer to an embarrassment. Quantum process tomography, formalized by Chuang and Nielsen and independently by Poyatos, Cirac, and Zoller in 1997, promised to reconstruct any quantum operation from data — but it did so by trusting the very preparations and measurements it should have distrusted, silently attributing SPAM errors to the gate. For years the field knew its answers were only as good as the assumed reference frame, with no principled fix. The turn came in 2013. Seth Merkel and collaborators at IBM introduced "self-consistent quantum process tomography," dropping the perfect-SPAM assumption and fitting states, gates, and effects together; the same year, Robin Blume-Kohout and colleagues at Sandia introduced long-sequence gate set tomography, adding the two ideas that made it rigorous — the recognition that the reconstruction is defined only up to a gauge, and the use of deep germ-power circuits to amplify coherent errors to Heisenberg-limited precision. Their 2017 demonstration certified single-qubit trapped-ion gates below a rigorous fault-tolerance threshold, the first such claim made honestly because it reported gauge-invariant diamond-norm bounds, not fidelities alone. The framework was consolidated by Erik Nielsen, John King Gamble, Kenneth Rudinger, Kevin Young, and the Sandia group in the open-source pyGSTi package and a comprehensive 2021 review. Its lasting lesson is subtle: when nothing is trusted, only relationships are knowable, and the gauge is the precise statement of that humility.

History

1997

— Chuang and Nielsen, and independently Poyatos, Cirac, and Zoller, formulate quantum process tomography

2013

— Merkel and collaborators introduce self-consistent process tomography, dropping the perfect-SPAM assumption

2013

— Blume-Kohout and colleagues introduce long-sequence gate set tomography and the gauge picture

2017

— gate set tomography certifies trapped-ion gates below a rigorous fault-tolerance threshold

2021

— the pyGSTi framework and a comprehensive review consolidate the method and its model-violation tests

Going further

Texts. Nielsen and Chuang, Quantum Computation and Quantum Information, Ch. 8 (quantum operations) and §8.4.2 (process tomography) · Preskill, Lecture Notes on Quantum Computation, Ch. 3 (channels, fidelity, and distance measures) · Watrous, The Theory of Quantum Information, Ch. 2 and 3 (superoperators and the diamond norm) · Wilde, Quantum Information Theory, Ch. 4 and 9 (channels and distinguishability) · Kitaev, Shen, and Vyalyi, Classical and Quantum Computation (the superoperator formalism and the diamond norm).

Provenance. A verifier should reproduce: (1) the telescoping proof of gauge invariance directly from the bilinear probability of equation (2); (2) the $1/L$ precision from binomial error propagation on $p_{\uparrow}(L)=\tfrac12[1+\cos(L\phi+L\theta)]$ , giving $\Delta\hat\theta=0.032$ at $L=1$ and $4.9\times10^{-4}$ at $L=64$ for $N=1000$ ; (3) the fidelity relation $F_{\mathrm{avg}}=(dF_{\mathrm{pro}}+1)/(d+1)$ and, for a coherent single-qubit over-rotation of $\theta=0.02$ , $r=\theta^2/6=6.67\times10^{-5}$ against the diamond error $\sin(\theta/2)=10^{-2}$ ; (4) the corrected eigenvalue coefficients $r_{\mathrm{coh}}\approx\tfrac{1}{2d(d+1)}\sum\theta_k^2$ and $r_{\mathrm{sto}}\approx\tfrac{1}{d(d+1)}\sum(1-|\lambda_k|)$ , checked against the depolarizing result $r=p/2$ at $d=2$ . All quantities are dimension-general textbook identities; no device-specific numbers are asserted as measured.

Prerequisites. Quantum Operations and CPTP Maps · Kraus Representation of Quantum Channels · Pauli and Bit/Phase-Flip Channels · Decoherence. Leads to. Diamond Norm and Channel Distinguishability · Markovian versus Non-Markovian Dynamics · Depolarizing Channel · Lindblad / GKSL Master Equation.

Ask the tutor: Ask me to telescope the gauge out of a specific circuit, to turn a measured over-rotation into its infidelity and diamond error, or to explain how germs and fiducials are chosen for a given gate set.

What exactly is the gauge, and why can't I just measure it away?

The gauge is an invertible similarity $B$ acting on all elements at once — $R_G\mapsto BR_GB^{-1}$ , $|\rho\rangle\rangle\mapsto B|\rho\rangle\rangle$ , $\langle\langle E|\mapsto\langle\langle E|B^{-1}$ — and it leaves every circuit probability unchanged because the $B$ and $B^{-1}$ telescope out. That invariance is a mathematical identity, not a measurement limit you could beat with more data, so no experiment can distinguish two gate sets related by a gauge. The consequence is that raw matrix elements are meaningless: you must quote gauge-invariant quantities (eigenvalues, unitarity, entanglement fidelity to the target after gauge-fixing) or first pin the gauge by a CPTP-constrained optimization toward the target. The rulers-with-no-reference picture is exact — relative facts are knowable, the absolute frame is not.

How is GST different from randomized benchmarking?

They answer different questions at different costs. Randomized benchmarking runs random Clifford sequences and fits one exponential decay to a single gauge-independent error rate — cheap, robust, blind to structure. GST reconstructs the entire gate set: process matrices (up to gauge), a diamond-norm worst-case bound, and a coherent-versus-stochastic split, at much greater cost. They are complementary: the depolarizing rate GST extracts should match the RB decay constant, and a mismatch flags gate-dependent or non-Markovian errors RB averages away. Use RB to track a number over time; use GST when you need to know what kind of error you have and how bad the worst case is.

Why does long-sequence GST beat the standard 1/√N precision?

Because it converts a small static error into a large one before measuring it. A per-gate over-rotation $\theta$ accumulates to $L\theta$ after $L$ repetitions, and with the right fiducials the outcome probability oscillates in $\theta$ at frequency $L$ . Its derivative in $\theta$ grows as $L$ , so dividing the fixed binomial shot noise $\sqrt{p(1-p)/N}$ by an $O(L)$ slope gives $\Delta\hat\theta\sim1/(L\sqrt{N})$ — depth 64 sharpens the estimate 64-fold at the same shots per circuit. The catch is decoherence: stochastic errors damp the fringe contrast as $e^{-\gamma L}$ , so the $1/L$ gain flattens once $L$ nears $1/\gamma$ , and the best experiment balances amplification against signal loss.

Why report the diamond norm when the fidelity is already so small?

Because fidelity is an average and fault tolerance cares about the worst case. For a coherent over-rotation $\theta$ the average infidelity is $r\approx\theta^2/6$ but the diamond error is $\tfrac12\|\cdot\|_\diamond\approx\theta/2$ , larger by $\sim3/\theta$ — a hundredfold for $\theta$ of a few milliradians. Threshold theorems are stated in the diamond norm precisely because coherent errors add in amplitude, accumulating quadratically worse than their infidelity suggests. A gate a benchmark calls " $10^{-4}$ good" may carry a worst-case error of $10^{-2}$ , right at a typical surface-code threshold. Separating the coherent part and bounding the diamond norm is what makes GST's error budgets honest.

What are germs and fiducials, and why are they built the way they are?

Fiducials are short subcircuits that dress the native preparation and measurement so that together they prepare an informationally complete set of input states and read out an informationally complete set of effects — enough to span the $d^2$ -dimensional space and make the inverse in equation (4) well defined. Germs are the circuits placed between the fiducials and repeated; each is chosen so its repetition amplifies a particular combination of gate errors to first order in the number of repetitions. A good germ set is amplificationally complete: every gauge-invariant parameter is amplified by at least one germ, so nothing important stays stuck at $1/\sqrt{N}$ precision. pyGSTi ships standard fiducial and germ sets for common gate sets.

When does GST fail, and how would I know?

GST assumes the gates are Markovian, time-stationary, and described by a fixed SPAM — no memory, no drift, no dependence on context. When those assumptions break the reconstruction is not just noisy but wrong. Its saving grace is that it overdetermines its model, so it detects its own failure: the goodness-of-fit statistic (a log-likelihood ratio, asymptotically $\chi^2$ ) is compared to its expected value, and a significant excess is a model violation saying the Markovian CPTP picture cannot fit the data. Common culprits are slow calibration drift over the hours-long run, crosstalk from neighbours, leakage out of the computational subspace, and non-Markovian bath memory. A large model violation is a verdict to trust more than any fidelity the fit still prints — and it is why full GST remains largely a one- and two-qubit instrument, since both the circuit count and the fit cost over $4^q\times4^q$ matrices grow steeply with qubit number.