Suppose a quantum computer can prepare a state \(\rho\) as often as we like, but the only way to learn about it is to measure. If we know in advance which observable we want, the best strategy is to measure that observable. Often we do not: we may want thousands of properties, or the questions may arise only after the experiment, from an analysis not yet designed. Classical shadows [1, 2] are made for this situation. Each run measures the state in a randomly chosen basis, and the record of bases and outcomes is kept. Any number of properties can then be estimated later, classically, from the same record. Ref. [3] reviews the wider family of randomized-measurement methods.

This note works through the idea. I wrote it for my own education and to share; if you have comments or would like to discuss it, please write to me. Sections 1–3 build the general framework: the measurement channel, its inverse, and how the choice of random measurements decides which questions are cheap. Section 4 specializes to fermions, Sec. 5 surveys what else the record can be asked, and Sec. 6 follows one demanding application, supplying wavefunction overlaps to quantum Monte Carlo, which is the subject of our recent paper [4].

1. The general idea

Let \(O_1,\dots,O_M\) be observables on \(n\) qubits. Measuring each directly costs a separate set of runs per observable, or per group of commuting observables, and requires the list before the experiment. The randomized-measurement protocol reverses the order. On each run \(t = 1,\dots,N\):

  1. draw a unitary \(U_t\) from a fixed ensemble \(\mathcal U\);
  2. apply it to \(\rho\) and measure every qubit in the computational basis, obtaining a bit string \(b_t\);
  3. store \((U_t, b_t)\).

The \(N\) records form the classical shadow [Fig. 1]. Nothing in the protocol refers to the \(O_i\); they enter only in the classical processing. Two questions organize the rest of the note: how a record of random measurements yields unbiased estimates (Sec. 2), and how the ensemble determines which estimates are accurate and cheap to compute (Secs. 3 and 4).

Three panels. (a) Random measurement: four qubit wires of the state rho pass through a blue gate U sub t and end in measurement meters, whose outcomes are bracketed as the bit string b sub t. (b) Classical shadow: a stack of slabs, blue in front and fading to white outlines, then an ellipsis, labeled (U sub t, b sub t), t from 1 to N. (c) Predictions: tr(O 1 rho), tr(O 2 rho), an ellipsis, and tr(O M rho). Arrows lead from a to b and from b to c.
Figure 1. The randomized-measurement protocol. (a) On each run a fresh copy of \(\rho\) is rotated by a unitary \(U_t\) drawn from the ensemble \(\mathcal U\), and every qubit is measured in the computational basis, giving the bit string \(b_t\). (b) Each run leaves one record \((U_t, b_t)\); the slabs stand for the \(N\) records. (c) Observables are chosen after the experiment and estimated from the same records: Eq. (3) for each record, combined by a median of means [2] (Sec. 2). Reference [5] explains the name with an object lit from random directions and recovered from its shadows. The analogy is loose: one record is a single random outcome in one basis, not a complete projection, and the snapshot built from it is not a state (Sec. 2).

2. The measurement channel and its inverse

Averaged over the random unitary and the outcome, the post-measurement state \(U^\dagger|b\rangle\langle b|U\) defines a linear map, the measurement channel:

\[ \mathcal M(\rho) = \mathbb E_{U\sim\mathcal U} \sum_{b} \langle b|U\rho U^\dagger|b\rangle\, U^\dagger|b\rangle\langle b|U . \tag{1} \]

The channel loses information. If it is invertible, each record defines a snapshot

\[ \hat\rho = \mathcal M^{-1}\bigl(U^\dagger|b\rangle\langle b|U\bigr), \qquad \mathbb E[\hat\rho] = \rho , \tag{2} \]

and \(\hat o = \operatorname{tr}(O\hat\rho)\) is an unbiased estimate of \(\operatorname{tr}(O\rho)\). Some ensembles see only part of operator space: matchgates, for example, never see odd fermionic operators (Sec. 4). Inverting \(\mathcal M\) only on the visible subspace, Eq. (2) becomes \(\mathbb E[\hat\rho] = \Pi(\rho)\), the projection of \(\rho\) onto that subspace, and \(\hat o\) remains unbiased for every \(O\) inside it. Since \(\mathcal M\) is self-adjoint with respect to the Hilbert–Schmidt inner product, the inverse is applied to the observable rather than to the state:

\[ \hat o = \langle b|U\,\mathcal M^{-1}(O)\,U^\dagger|b\rangle . \tag{3} \]

Only \((U, b)\) is stored; the \(2^n\times2^n\) matrix \(\hat\rho\) is never formed. A snapshot is in general not positive, so it is not a state: unbiasedness is a property of the average, and clipping single estimates to a physical range would destroy it.

How many snapshots. The variance of \(\hat o\) is bounded by the squared shadow norm \(\|O_0\|_{\mathrm{sh}}^2\) of the traceless part \(O_0 = O - \operatorname{tr}(O)\,I/2^n\), which depends on both the observable and the ensemble [2]. Splitting the snapshots into batches and taking the median of the batch means, all \(M\) estimates lie within \(\epsilon\) of their exact values with probability at least \(1-\delta\) once

\[ N = O\!\left(\frac{\log(M/\delta)}{\epsilon^2}\,\max_i \|O_{i,0}\|_{\mathrm{sh}}^2\right). \tag{4} \]

The dependence on the number of questions is only logarithmic. The price is set by the shadow norm, and that is where the choice of ensemble enters. The guarantee holds for any list of observables fixed independently of the recorded outcomes. The list may be written down after the experiment, but if each question is chosen using answers already computed from the same record, the questions are adaptive, and Eq. (4) does not apply as it stands [6].

3. Choosing the ensemble

Three questions decide whether an ensemble suits a task: which operators \(\mathcal M\) can see, how large the variance of Eq. (3) is, and how expensive Eq. (3) is to evaluate classically. For the ensembles used in practice, the first two have a simple answer. Suppose each element of an operator basis is either diagonal in the computational basis or zero on its diagonal, as Pauli strings and Majorana monomials are, and conjugation by a random \(U\) maps each element \(P\) to \(\pm P'\), with \(P'\) uniformly distributed over an orbit of basis elements. A computational-basis measurement keeps the diagonal part of an operator, so it keeps \(P'\) if \(P'\) is diagonal and annihilates it otherwise:

\[ \mathcal M(P) = f_P\, P, \qquad f_P = \Pr\bigl[\,UPU^\dagger \text{ is diagonal}\,\bigr]. \tag{5} \]

The channel is diagonal in this basis and its inverse multiplies by \(1/f_P\). An operator is visible if \(f_P > 0\). For a single basis element with eigenvalues \(\pm1\), the estimate of Eq. (3) is \(\pm 1/f_P\) with probability \(f_P\) and zero otherwise, so its second moment is exactly \(1/f_P\): rare survival means large variance.

Local Pauli. Each qubit is measured in a random Pauli basis. A Pauli string of weight \(k\) is diagonal only if all \(k\) bases match, so \(f = 3^{-k}\), and the snapshot is \(\bigotimes_j (3|s_j\rangle\langle s_j| - I)\). The circuits are trivial and so is the classical processing, but the variance grows exponentially with the support of the observable.

Global Clifford. A uniformly random \(n\)-qubit Clifford maps a non-identity Pauli string, up to sign, to each of the \(4^n-1\) non-identity strings with equal probability. Of these, \(2^n-1\) are diagonal, so \(f = 1/(2^n+1)\) for every traceless Pauli string, and the snapshot is \((2^n+1)|s\rangle\langle s| - I\) with \(|s\rangle\) a stabilizer state. Single Pauli strings are now expensive, but low-rank observables are cheap: the squared shadow norm obeys \(\|O_0\|_{\mathrm{sh}}^2 \le 3\operatorname{tr}(O_0^2)\), so the fidelity with any pure state needs a number of snapshots independent of \(n\) [2]. The catch is classical: Eq. (3) requires \(\langle s|\psi\rangle\), which is efficient for a stabilizer target but not for a generic \(|\psi\rangle\).

Matchgate. Random fermionic Gaussian unitaries (Sec. 4) are matched to fermionic observables and to Slater determinants, for which Eq. (3) reduces to Pfaffians.

Table 1. Three ensembles. The second moment is that of the estimate for one basis element, \(1/f_P\) in Eq. (5).
EnsembleVisibleSecond momentClassical evaluationSuited to
Local Pauliall\(3^k\), Pauli string of weight \(k\)product of one-qubit factorslocal observables, small-subsystem entropies
Global Cliffordall\(2^n+1\), any non-identity Pauli stringstabilizer algebra; hard for generic targetsfidelities with stabilizer or otherwise tractable states
Matchgateparity-even\(\binom{2n}{2\ell}/\binom{n}{\ell}\), Hermitian Majorana monomial of degree \(2\ell\)Pfaffians\(k\)-RDMs, Gaussian projectors, Slater-determinant overlaps

4. Fermions: matchgate shadows

For \(n\) fermionic modes, define the \(2n\) Majorana operators

\[ \gamma_{2j-1} = a_j + a_j^\dagger, \qquad \gamma_{2j} = -i\,(a_j - a_j^\dagger), \qquad \{\gamma_\mu,\gamma_\nu\} = 2\delta_{\mu\nu} . \tag{6} \]

Under the Jordan–Wigner mapping, \(\gamma_{2j-1} = Z_1\cdots Z_{j-1}X_j\), \(\gamma_{2j} = Z_1\cdots Z_{j-1}Y_j\), and \(Z_j = -i\gamma_{2j-1}\gamma_{2j}\). Products of distinct Majoranas, with a phase that makes them Hermitian, form an operator basis whose elements have eigenvalues \(\pm1\); for even degree \(|S| = 2\ell\), \(\gamma_S = (-i)^\ell\gamma_{s_1}\cdots\gamma_{s_{2\ell}}\). A one-body operator \(a_p^\dagger a_q\) has degree 2 (plus a constant when \(p = q\)), and a two-body operator degree at most 4. A matchgate, or fermionic Gaussian, unitary \(U_Q\) rotates the Majoranas linearly,

\[ U_Q^\dagger\, \gamma_\mu\, U_Q = \sum_\nu Q_{\mu\nu}\,\gamma_\nu, \qquad Q \in O(2n), \tag{7} \]

and therefore preserves the degree. The parity operator \((-i)^n\gamma_1\cdots\gamma_{2n}\) picks up a factor \(\det Q\). Rotations with \(\det Q = 1\) form \(SO(2n)\), the standard parity-preserving subgroup, generated by quadratic Hamiltonians and by the nearest-neighbour two-qubit gates \(G(A,B)\) discussed below. Those with \(\det Q = -1\), such as \(\gamma_1 = X_1\) itself, flip the parity. Following Ref. [7], we call the whole \(O(2n)\) family (generalized) matchgates. Orbital rotations are the special case that conserves particle number; a Slater determinant is a rotated computational-basis state, and hence a Gaussian state.

The channel follows from Eq. (5). Take \(Q\) uniformly from the signed permutations of the Majoranas, the matchgates that are also Clifford circuits; this gives the same channel as the Haar measure on \(O(2n)\) [4, 7]. A monomial of degree \(2\ell\) is mapped uniformly onto the \(\binom{2n}{2\ell}\) monomials of that degree, and the image is diagonal only if it is a product of \(\ell\) of the \(n\) pairs \(\gamma_{2j-1}\gamma_{2j}\). Hence [7]

\[ \mathcal M = \sum_{\ell=0}^{n} f_\ell\, \mathcal P_{2\ell}, \qquad f_\ell = \binom{n}{\ell}\Big/\binom{2n}{2\ell}, \tag{8} \]

where \(\mathcal P_{2\ell}\) projects onto degree \(2\ell\). Three consequences:

  • Odd degrees are invisible. They are never diagonal, so the snapshots reproduce only the parity-even part of \(\rho\). Any quantity to be estimated must be written with even operators.
  • Low degrees are cheap. \(f_1 = 1/(2n-1)\) and \(f_2 = 3/[(2n-1)(2n-3)]\), so elements of the one- and two-body reduced density matrices have second moments of order \(n\) and \(n^2\) [7, 8]. With local Pauli shadows the same operators can be far more expensive: through the Jordan–Wigner strings, \(a_1^\dagger a_n\) becomes a sum of Pauli strings of weight \(n\), each with second moment \(3^n\).
  • Middle degrees are amplified most. For \(n = 21\), \(1/f_\ell\) rises from 41 at \(\ell = 1\) to \(1.5\times10^6\) at \(\ell = 10\) and 11, and falls back to 1 at \(\ell = 21\). Global quantities with weight at intermediate degrees are therefore the noisiest; for Gaussian projectors the variance nevertheless grows only as \(O(\sqrt n\log n)\) [7].

Before inversion, the snapshot \(U_Q^\dagger|b\rangle\langle b|U_Q\) is a Gaussian state, fully described by its \(2n\times2n\) covariance matrix \(Q^{\mathsf T}C_bQ\), where \(C_b\) is that of \(|b\rangle\). By Wick's theorem, its contraction with a Gaussian target reduces to a Pfaffian. The inverse channel adds a reweighting of each degree by \(1/f_\ell\); Sec. 6 shows how this is done efficiently.

Why "matchgate". The name is older than the fermionic picture. Valiant [9] defined a matchgate as a small weighted graph whose boundary vertices are attached to the input and output qubits. The amplitude \(\langle y|M|x\rangle\) is a sum over the perfect matchings (sets of edges that cover every vertex exactly once) of the graph left after deleting the boundary vertices whose bit is 1, each matching weighted by the product of its edge weights [Fig. 2(a)]; the full definition adds sign conventions [10]. Two properties can be read off the picture. A perfect matching needs an even number of vertices. If the graph has an even number of vertices in total, as in Fig. 2, an even number must be deleted, so the input and output bit strings have the same parity, as for \(\det Q = 1\) in Eq. (7); a graph with an odd number of vertices flips the parity instead. Two-qubit gates of the parity-preserving kind act as \(G(A,B)\), with \(A\) on the span of \(|00\rangle, |11\rangle\) and \(B\) on the span of \(|01\rangle, |10\rangle\), and with \(\det A = \det B\) [10]. And nearest-neighbour gates glue into a planar graph, whose perfect matchings Kasteleyn's method counts with a single Pfaffian [11], so the circuit can be simulated classically. A SWAP would make two edges cross; it is not a matchgate, and adding it makes the gate set universal [10]. Knill [12] and Terhal and DiVincenzo [13] identified matchgate circuits with noninteracting fermions under the Jordan–Wigner mapping. Wick's theorem is the same combinatorics [Fig. 2(b)]: the expectation value of a product of Majoranas in a Gaussian state is a sum over the ways to pair them up, which are the perfect matchings of the points, each signed by \((-1)\) to the number of crossing arcs. The Pfaffians of Sec. 6 are matching sums of this kind.

Two rows. (a) A violet two-qubit gate M on wires 1 and 2 equals a square graph whose four corners are the boundary vertices. Next, the amplitude of 01 to 01: the two bottom vertices are hollow, their edges grey, and the top edge is drawn thick in violet, the one remaining perfect matching. Last, the amplitude of 00 to 01: only the bottom-right vertex is hollow, three vertices remain and none is highlighted, and the amplitude equals zero. (b) The expectation value of gamma 1 gamma 2 gamma 3 gamma 4 equals three diagrams of four points on a line joined in pairs by violet arcs: adjacent pairs 1-2 and 3-4, minus the crossing pairs 1-3 and 2-4, plus the nested pairs 1-4 and 2-3, each labeled by its product of two-point functions; the sum equals the Pfaffian of the matrix of two-point functions.
Figure 2. The perfect matchings behind the name. (a) A two-qubit matchgate drawn as a graph with four boundary vertices (inputs left, outputs right; qubit 1 on top); edge weights are not shown. For \(\langle 01|M|01\rangle\) the vertices of qubit 2 are deleted (open circles, grey edges) and one perfect matching remains (violet). For \(\langle 01|M|00\rangle\) only one vertex is deleted; three vertices cannot be matched, so the amplitude vanishes. (b) Wick's theorem for four Majoranas in a Gaussian state: each term is a pairing, drawn as arcs, and the pairing with one crossing enters with a minus sign. The sum is the Pfaffian of the antisymmetric matrix of two-point functions.

5. Other questions the record can answer

  • Nonlinear functions. For snapshots with \(\mathbb E[\hat\rho] = \rho\), two independent snapshots estimate a quadratic function without bias: \(\mathbb E[\operatorname{tr}(\hat\rho_t\hat\rho_{t'})] = \operatorname{tr}(\rho^2)\) for \(t \ne t'\). With the restricted inverse of Sec. 2, the same average gives \(\operatorname{tr}[\Pi(\rho)^2]\) instead. Averaged over all pairs and restricted to a subsystem \(A\), this gives an unbiased estimate of the purity \(\operatorname{tr}\rho_A^2\). The Rényi-2 entropy \(-\log\operatorname{tr}\rho_A^2\) follows by taking the logarithm, which makes it biased, although the bias vanishes as the purity estimate sharpens. Brydges et al. measured Rényi-2 entropies on a trapped-ion simulator from the same kind of randomized measurements, with an estimator built from correlations between outcomes [14]. Moments of the partial transpose detect mixed-state entanglement [15]. The number of snapshots grows exponentially with the size of \(A\), so this works for small subsystems.
  • Learning. A shadow is a classical data set, so it can train a model. From shadows of ground states across a family of Hamiltonians, classical machine learning can predict ground-state properties at new parameters and classify phases [16]. Section 6 applies the same idea to a single state.
  • Noise. Device noise changes the channel. If the noise is the same for every random circuit, the random unitaries twirl it into a channel that is still diagonal in the same basis, with modified eigenvalues. Estimating these from an easily prepared calibration state and inverting them in place of the ideal \(f_P\) corrects the noise bias, up to the statistical error of the calibration (robust shadows) [17]. Snapshots can also be combined to emulate virtual distillation for error mitigation [18].
  • When not to use shadows. If only a few observables are needed and are known in advance, measuring them directly, or in commuting groups, is cheaper. Derandomized [19] and locally biased [20] shadows interpolate between the two strategies by tuning the measurement distribution to a known list of observables.

6. Case study: overlaps for quantum Monte Carlo

In projector quantum Monte Carlo (QMC), the ground state is sampled by an ensemble of walkers, here Slater determinants \(|\phi\rangle\), propagated in imaginary time. A trial state \(|\Psi_T\rangle\) controls the sign problem, and it enters the calculation only through the overlap \(f_T(\phi) = \langle\Psi_T|\phi\rangle\) and its derivatives along one-body directions, which give the force bias and the local energy [4, 21]. In quantum–classical QMC (QC-QMC), \(|\Psi_T\rangle\) is prepared on a quantum computer [21]. The walkers are generated during the calculation, of order \(N_wN_\tau\) of them for \(N_w\) walkers and \(N_\tau\) time steps, so the questions are numerous and not known in advance: the setting of Sec. 1. They are also adaptive in the sense of Sec. 2: each walker moves according to overlaps already estimated from the same shadows, so Eq. (4) does not carry over to the QMC energy automatically.

An overlap is not an expectation value in \(|\Psi_T\rangle\), but it is one in a slightly larger state. Let \(|\chi\rangle = (|0\rangle + |\Psi_T\rangle)/\sqrt2\), with \(|0\rangle\) the vacuum. Because the vacuum is orthogonal to every state with \(\eta \ge 1\) electrons,

\[ \langle\Psi_T|\phi\rangle = 2\operatorname{tr}\bigl(|\chi\rangle\langle\chi|\; |\phi\rangle\langle0|\bigr), \qquad X_\phi = 2\operatorname{tr}\bigl(|\phi\rangle\langle0|\,\hat\rho\bigr), \tag{9} \]

and \(X_\phi\), evaluated on matchgate snapshots of \(|\chi\rangle\), is an unbiased, complex-valued estimate of the overlap [7, 21]. The vacuum supplies the phase reference that a fidelity \(|\langle\Psi_T|\phi\rangle|^2\) lacks. The operator \(|\phi\rangle\langle0|\) changes the particle number by \(\eta\), so it is parity-even, and visible, only for even \(\eta\). For odd \(\eta\), one ancilla mode is added, occupied in the trial branch and in the walker but empty in the vacuum branch: \(|\tilde\chi\rangle = (|0\rangle|0\rangle_a + |\Psi_T\rangle|1\rangle_a)/\sqrt2\) and \(|\tilde\phi\rangle = |\phi\rangle|1\rangle_a\). The overlap is unchanged, and \(|\tilde\phi\rangle\langle0|\langle0|_a\) changes the particle number by \(\eta + 1\), which is even [4, 7].

Because each walker is a product of a spin-up and a spin-down determinant, the two spin registers can be measured with independent random circuits [4]. The channel is then \(\mathcal M_\alpha\otimes\mathcal M_\beta\), and \(|\phi\rangle\langle0| = |\phi_\alpha\rangle\langle0_\alpha|\otimes|\phi_\beta\rangle\langle0_\beta|\) factorizes, so Eq. (9) is evaluated in each register separately, with that register's modes, electrons and \(f_\ell\), giving \(X_\alpha\) and \(X_\beta\). For the joint superposition \(|\chi\rangle\) above, \(X_\alpha X_\beta/2\) estimates the overlap; when the trial is a product over spins and each register is prepared in its own superposition, \(X_\alpha X_\beta\) does [4].

For the CAS(30e,20o) model of [2Fe–2S] constructed in a previous note, each spin has 15 electrons, an odd number, so each register needs an ancilla: two registers of 20 + 1 qubits. In our hardware run the trial was a product over spins and each register was prepared in its own superposition, so the estimator is \(X_\alpha X_\beta\) [4].

Evaluating Eq. (9) for one snapshot of one register uses the structure of Eq. (8), with \(n\) and \(\eta\) the modes and electrons of that register. Splitting the trace by degree,

\[ X_\phi = 2\sum_{\ell} f_\ell^{-1}\, q_\ell, \qquad q_\ell = \operatorname{tr}\bigl[\,|\phi\rangle\langle0|\;\mathcal P_{2\ell}\bigl(U_Q^\dagger|b\rangle\langle b|U_Q\bigr)\bigr] . \tag{10} \]

For even \(\eta\), Wan et al. showed that the \(q_\ell\) are the coefficients of a Pfaffian [7]: \(\sum_\ell q_\ell z^\ell = i^{\eta/2}\,2^{-r}\operatorname{pf}(B + zC)\), with \(2r = 2n - \eta\). Here \(B\) is a fixed \(2r\times2r\) skew-symmetric matrix that depends only on \(n\) and \(\eta\), and \(C\) is the reduced skew-symmetric matrix obtained from the snapshot covariance matrix through the walker's orbital rotation, a fixed complex change of basis, and restriction to \(2r\) rows and columns [4]. The power \(\ell\) of \(z\) labels Majorana degree \(2\ell\).

The inverse channel thus reweights the coefficients of a polynomial. Computing them by evaluating the Pfaffian at \(O(n)\) points and interpolating costs \(O(n^4)\). Interpolation at real nodes followed by conversion to monomial coefficients is also ill-conditioned; interpolation at roots of unity, a discrete Fourier transform, avoids this. Reducing the matrix pencil \((B, C)\) to a structured form gives all coefficients in \(O(n^3)\) [4].

QMC uses only ratios of overlaps: the phase of \(f_T(\phi')/f_T(\phi)\) in the walker update, and the force bias and local energy, which are derivatives divided by \(f_T\). A common factor therefore drops out, including the normalization of \(|\chi\rangle\) and, for noise that the random circuits twirl into a global scale, the effect of device noise [4]. Two properties work against it. A ratio of unbiased estimates is biased, and overlaps can decay exponentially with system size, so a fixed relative precision requires a growing number of snapshots.

The remaining obstacle is classical cost. Every overlap must be evaluated for every snapshot, walker and time step, \(O(n^3 N N_w N_\tau)\) for \(N\) snapshots, before derivatives. In Ref. [4] we decouple the two: overlaps are estimated from the shadows for a modest set of \(N_{\mathrm{train}} \ll N_wN_\tau\) determinants, a differentiable surrogate \(f_\theta(\phi) \approx \langle\Psi_T|\phi\rangle\) is trained on them, and QMC queries only the surrogate. This is the learning use of Sec. 5 applied to a single state. It made QC-QMC with shadows measured on Google's Willow processor affordable for up to 40 system qubits (42 with the two ancillas), including the [2Fe–2S] model above.

About this note. This note was developed in dialogue with Opus and Astra, with assistance in drafting, revision, and critical review.

References

  1. [1] S. Aaronson, in Proceedings of the 50th Annual ACM SIGACT Symposium on Theory of Computing (STOC 2018), p. 325.
  2. [2] H.-Y. Huang, R. Kueng, and J. Preskill, Nat. Phys. 16, 1050 (2020).
  3. [3] A. Elben, S. T. Flammia, H.-Y. Huang, R. Kueng, J. Preskill, B. Vermersch, and P. Zoller, Nat. Rev. Phys. 5, 9 (2023).
  4. [4] T. Jiang, A. Zhao, S. Kumar, J. Nigam, E. Rosenberg, R. Babbush, A. D. Baczewski, W. J. Huggins, and J. Lee, arXiv:2611.08151 (2026).
  5. [5] H.-Y. Hu, Hamiltonian-driven shadow tomography: a short introduction (web page); the protocol is in H.-Y. Hu and Y.-Z. You, arXiv:2102.10132.
  6. [6] J. Huang, L. Lewis, H.-Y. Huang, and J. Preskill, arXiv:2410.15501 (2024).
  7. [7] K. Wan, W. J. Huggins, J. Lee, and R. Babbush, Commun. Math. Phys. 404, 629 (2023).
  8. [8] A. Zhao, N. C. Rubin, and A. Miyake, Phys. Rev. Lett. 127, 110504 (2021).
  9. [9] L. G. Valiant, SIAM J. Comput. 31, 1229 (2002).
  10. [10] R. Jozsa and A. Miyake, Proc. R. Soc. A 464, 3089 (2008).
  11. [11] P. W. Kasteleyn, Physica 27, 1209 (1961).
  12. [12] E. Knill, arXiv:quant-ph/0108033 (2001).
  13. [13] B. M. Terhal and D. P. DiVincenzo, Phys. Rev. A 65, 032325 (2002).
  14. [14] T. Brydges, A. Elben, P. Jurcevic, B. Vermersch, C. Maier, B. P. Lanyon, P. Zoller, R. Blatt, and C. F. Roos, Science 364, 260 (2019).
  15. [15] A. Elben, R. Kueng, H.-Y. Huang, R. van Bijnen, C. Kokail, M. Dalmonte, P. Calabrese, B. Kraus, J. Preskill, P. Zoller, and B. Vermersch, Phys. Rev. Lett. 125, 200501 (2020).
  16. [16] H.-Y. Huang, R. Kueng, G. Torlai, V. V. Albert, and J. Preskill, Science 377, eabk3333 (2022).
  17. [17] S. Chen, W. Yu, P. Zeng, and S. T. Flammia, PRX Quantum 2, 030348 (2021).
  18. [18] A. Seif, Z.-P. Cian, S. Zhou, S. Chen, and L. Jiang, PRX Quantum 4, 010303 (2023).
  19. [19] H.-Y. Huang, R. Kueng, and J. Preskill, Phys. Rev. Lett. 127, 030503 (2021).
  20. [20] C. Hadfield, S. Bravyi, R. Raymond, and A. Mezzacapo, Commun. Math. Phys. 391, 951 (2022).
  21. [21] W. J. Huggins, B. A. O'Gorman, N. C. Rubin, D. R. Reichman, R. Babbush, and J. Lee, Nature 603, 416 (2022).