Tomography & Gate Set Tomography
Self-consistent characterization of noisy operations.
State tomography trusts its measurements. Process tomography trusts its state preparations and measurements. But on a real device every preparation and measurement is built from the same imperfect gates you are trying to characterize. Gate set tomography breaks this circle by fitting everything at once, and pays for it with a fundamental ambiguity: the gauge freedom.
What you’ll learn
- Represent channels by Pauli transfer matrices, count their parameters, and describe quantum process tomography.
- Explain why SPAM errors make process tomography self-inconsistent, and define the gate-set model.
- Derive linear gate set tomography, exhibit the gauge freedom, and distinguish gauge-invariant from gauge-variant quantities.
- Explain how long germ sequences give precision, and how model violation tests Markovianity.
Process tomography
A channel on qubits () is fixed by its action on a basis of operators. Use the Pauli strings , for which :
How many real parameters does a general CPTP map on two qubits have?
Show the answer
240 parameters
A Hermiticity-preserving map on matrices has real parameters (its PTM is a real matrix), and trace preservation fixes the first row, removing . With : . For one qubit it is 12. Complete positivity is an inequality and removes no parameters.
Quantum process tomography (Chuang & Nielsen 1997) prepares an informationally complete set of inputs (for example all tensor products of , which is states), applies , performs state tomography on each output in Pauli bases, and inverts linearly or by maximum likelihood. That is configurations, each with its own shot noise. State tomography itself is treated in the Measurement topic's Rydberg State lesson on quantum state tomography.
The SPAM problem and the gate-set model
Process tomography assumes that the input states and measurement effects are known exactly. On real hardware they are made with gates: is " followed by a pulse", and an -basis measurement is "a pulse followed by a readout". Errors in those pulses masquerade as errors of the process under study, and the reconstructed map can even fail to be completely positive. The fix is to characterize preparations, measurements and gates together.
For a qubit, the trace-preserving gauge group has parameters. A gate set with one preparation (3 parameters), one two-outcome measurement (4) and three trace-preserving gates (12 each) has 43 parameters, of which only are physically meaningful.
Which of these quantities can, in principle, be determined by experiment?
- The process fidelity of the gate with its target
- The eigenvalues of the superoperator
- The Bloch vector of the prepared state
- The off-diagonal entries of the Pauli transfer matrix of
Show the answer
B. The eigenvalues of the superoperator
Eigenvalues are invariant under similarity transformations , so they are gauge-invariant and observable, as are all circuit probabilities and RB decay rates. The other three change under a gauge transformation, so their reported values depend on a convention, usually fixed by gauge optimization: choosing the that brings the estimate closest to the target.
Linear GST: the gauge made concrete
Choose fiducial sequences : short gate strings such that the effective states and effective effects are informationally complete. Then measure the probabilities of the circuits "prepare, , optional gate , , measure".
Stack the effective measurement covectors as the rows of and the effective prepared states as the columns of . Both are and invertible by informational completeness, but unknown.
The measured "Gram matrix" (no gate in the middle) and the data for gate .
The data determine every gate up to the same unknown similarity transformation : this is the gauge freedom.
The state and the effect come out in the same gauge, so is a complete, self-consistent estimate made without any assumption about SPAM.
Long sequences, precision and model violation
Long-sequence GST measures circuits of the form fiducial, germ repeated, fiducial. Each germ is a short gate string, repeated until the total length reaches (typically ). Germs are chosen so that every error parameter of the gate set is amplified by some germ.
A small over-rotation . Four repetitions ideally give the identity (up to sign), so the error accumulates coherently.
Prepare , apply once to reach the equator (where sensitivity is greatest), then , and measure.
With shots, near . The precision improves as , not only as .
A length- circuit takes time , so at fixed the cost of reaching precision scales like (Heisenberg-like), as long as the error stays coherent and stable over the whole sequence.
Robust phase estimation (Kimmel, Low & Yoder 2015) uses the same idea to calibrate rotation angles with Heisenberg-like scaling. The catch is that amplification only works if the errors are Markovian and stable during the long sequences.
| Method | Trusts | Learns | Cost |
|---|---|---|---|
| State tomography | the measurements | one state | exponential in |
| Process tomography | preparations and measurements | one channel ( parameters) | configurations |
| Gate set tomography | only the Markovian model | all preparations, measurements and gates, up to gauge | thousands of circuits for one qubit; rarely used beyond two |
| Randomized benchmarking | (nearly) gate-independent noise | one average error rate per gate set | polynomial |
| Pauli-noise learning, cycle benchmarking | Pauli twirling | Pauli error rates of whole layers | polynomial for sparse models |
Key takeaways
- Channels are Pauli transfer matrices with free parameters; process tomography estimates them by assuming perfectly known preparations and measurements.
- Gate set tomography models preparations, measurements and gates together; any gauge transformation , with and transformed accordingly, leaves all data unchanged.
- Linear GST recovers from the Gram matrix of fiducial circuits, the gauge freedom made explicit. Only gauge-invariant quantities (probabilities, spectra, RB decays) are truly observable.
- Long germ sequences amplify coherent errors for precision, and the goodness of fit () tests whether the noise is Markovian at all.
Check yourself
In the interactive lesson these are scored with your confidence, and what you miss comes back for review.
In the Pauli transfer matrix of a qubit channel, what does a non-zero entry in the first column below the top, such as , signal?
- The channel is not trace-preserving
- The channel is not completely positive
- The channel is not unital: it maps to something other than , as relaxation toward does
- The channel is entangling
Show the answer
C. The channel is not unital: it maps to something other than , as relaxation toward does
The first column is the image of . Trace preservation constrains the first row; a non-zero entry lower in the first column means , as for amplitude damping.
In linear GST the data give . What does this imply?
- is recovered exactly, because cancels
- Each gate is recovered up to a common similarity transformation, the gauge, which no amount of data can remove
- The method fails unless the fiducials are perfect
- can be determined by measuring the fiducials separately
Show the answer
B. Each gate is recovered up to a common similarity transformation, the gauge, which no amount of data can remove
, the matrix of effective prepared states, is unknown and enters as a similarity transformation shared by all gates, the state and the effect. Every gauge-equivalent gate set predicts the same data.
Why do GST and robust phase estimation use long, repeated gate sequences?
- To beat the shot-noise scaling by taking more shots
- To make the noise Markovian
- To twirl the noise into a Pauli channel
- Coherent errors accumulate linearly with repetition, so small error parameters can be estimated with precision
Show the answer
D. Coherent errors accumulate linearly with repetition, so small error parameters can be estimated with precision
An over-rotation repeated times becomes , so a measurement with fixed statistical error resolves to about .
A GST fit returns . What is the most reasonable conclusion?
- The data are inconsistent with any Markovian gate set of this form: look for drift, leakage or non-Markovian noise
- The gates are very bad
- The fit has converged to the target gate set
- More gauge optimization is needed
Show the answer
A. The data are inconsistent with any Markovian gate set of this form: look for drift, leakage or non-Markovian noise
compares the fit's log-likelihood with its expected distribution. A value in the hundreds means the model cannot explain the data, whatever the quality of the gates.
Further reading
- I. L. Chuang & M. A. Nielsen, "Prescription for experimental determination of the dynamics of a quantum black box", Journal of Modern Optics 44, 2455 (1997).
- S. T. Merkel et al., "Self-consistent quantum process tomography", Physical Review A 87, 062119 (2013).
- R. Blume-Kohout et al., "Demonstration of qubit operations below a rigorous fault tolerance threshold with gate set tomography", Nature Communications 8, 14485 (2017).
- E. Nielsen, J. K. Gamble, K. Rudinger, T. Scholten, K. Young & R. Blume-Kohout, "Gate set tomography", Quantum 5, 557 (2021).
- D. Greenbaum, "Introduction to quantum gate set tomography", arXiv:1509.02921 (2015).