Source-linked AI summary
Fermionic tomography and learning
Bryan O'Gorman
TL;DR
Estimating properties of quantum states is a central subroutine in quantum algorithms, and classical shadows offer a randomized-measurement framework for doing so. This paper analyzes the affine-matchgate ensemble to bound estimator variance and derive estimation protocols. It reports rigorous fidelity-estimation and Slater-learning results, numerical evidence for an X-type observable, and a smaller ensemble with the same channel.
Problem
Quantum algorithms need subroutines to estimate state properties, and classical shadows provide a randomized-measurement approach for estimating their expectations.
Source: §1 · p. 1Method
The paper analyzes the affine-matchgate channel to bound shadow norms and estimator variance, then uses those results to develop estimation protocols.
Source: §1 · p. 1Results
The paper gives efficient fidelity-estimation protocols for pure fermionic Gaussian states, numerical evidence for an X-type observable, and a rigorous sample-complexity bound for learning Slater determinants.
Source: §1 · pp. 1, 4Takeaways & Limitations
A smaller ensemble of measurement bases can yield the exact same channel as the full permutation group.
Source: §1 · p. 5The occupation-basis bitstrings need not coincide with the usual qubit basis under every fermion-to-qubit encoding, though the paper's results do not depend on the encoding.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Shadow tomography via classical shadows is a state-of-the-art approach for estimating properties of a quantum state. We present a simplified, combinatorial analysis of a recently proposed instantiation of this approach based on the ensemble of unitaries that are both fermionic Gaussian and Clifford. Using this analysis, we derive a corrected expression for the variance of the estimator. We then show how this leads to efficient estimation protocols for the fidelity with a pure fermionic Gaussian state (provably) and for an $X$-like operator of the form ($|\mathbf 0\rangle\langle\psi|$ + h.c.) (via numerical evidence). We also construct much smaller ensembles of measurement bases that yield the exact same quantum channel, which may help with compilation. We use these tools to show that an $n$-electron, $m$-mode Slater determinant can be learned to within $\epsilon$ fidelity given $O(n^2 m^7 \log(m / \delta) / \epsilon^2)$ samples of the Slater determinant.
1 Introduction
The introduction frames classical shadows as randomized measurements for estimating state properties, then develops an analysis of the affine-matchgate ensemble. It presents results on estimator variance, efficient estimation and Slater-determinant learning, alongside a smaller ensemble with the same channel.
1 Introduction
- Classical shadows use randomized measurements to build a shared record from which expectations of state properties can be estimated.The measurements are independent of the specific properties being estimated.Source: §1 · p. 1
- The analysis derives a general-operator second-moment expression for the affine-matchgate channel, bounding shadow norms and estimator variance.The affine-matchgate ensemble consists of unitaries that are both affine (Clifford) and fermionic Gaussian (matchgate).Source: §1 · pp. 1, 3
- The paper rigorously bounds the sample complexity of learning a Slater determinant by estimating its 1-RDMs and relating their error to fidelity.A Slater determinant is uniquely defined by its 1-RDMs.Source: §1 · p. 1
- The paper presents efficient estimation protocols for fidelity with pure fermionic Gaussian states and for an X-type operator, with the latter supported by numerical evidence.The X-type form is |0⟩⟨ψ|; the stated numerical evidence concerns its shadow norm.Source: §1 · p. 4
- A smaller set of measurement bases yields the exact same channel as the full permutation group, with potential to reduce circuit depth.The proposed circuit-depth reduction from approximately m to approximately m/2 is conjectured, not established.Source: §1 · p. 5
Learning a Slater determinant
Computational-basis measurements alone cannot distinguish all Slater determinants, but the paper gives a sample bound for learning one with high fidelity using affine-Gaussian measurements and classical processing.
Learning a Slater determinant
- Orthogonal n-electron Slater determinants can have identical computational-basis measurement distributions, so those measurements alone are insufficient for learning them.The cited example uses two n-electron, 2n-mode Slater determinants.Source: Learning a Slater determinant · p. 5
- The analysis notes that a Slater determinant is uniquely specified by its 2-RDMs and derives a rigorous, quantitative upper bound on the samples needed to learn a Slater determinant.Source: Learning a Slater determinant · p. 5
- Theorem 6 gives a quantitative sample upper bound for learning an m-mode, n-electron Slater determinant to fidelity at least 1 − ϵfid with probability at least 1 − δ.The bound's full sample expression is not present in the supplied passage.Source: Learning a Slater determinant · p. 5
- The algorithm outputs a classical description of the learned Slater determinant, using independent measurements in random affine-Gaussian bases followed by classical processing.The passage describes the quantum component as measurements of copies of the input state.Source: Learning a Slater determinant · p. 5
A.1 Additional notation
This notation maps diagonal Majorana index sequences to bitstrings and identifies a related subset whose elements can also be represented by bitstrings.
A.1 Additional notation
- For a diagonal Majorana operator Γµ, bin(µ) gives its corresponding m-bit string, and seq(x) is the inverse mapping.The notation is defined for µ in D2m.Source: §A.1 · p. 6
- For µ in D*2m, there is a unique µ′ in D2m with Γµ = ±Γµ′, and sgn(µ) records that sign.The supplied passage introduces the sign as an element of {±1}.Source: §A.1 · p. 6
- Elements of B2m,j are identified with j-bit strings, with binx and seqx mapping between the two representations.The passage gives these as analogues of bin and seq.Source: §A.1 · p. 6
A.2 Useful facts
The section records a range of admissible values for a parameter a given k and k′.
A.2 Useful facts
- Fact 4 places a in the range max{0, k + k′ − m} ≤ a ≤ min{k, k′}, for integers k and k′ from 0 to m.The passage introduces this as one of a set of useful facts.Source: §A.2 · pp. 6–7
A.3 Channel eigenvalues
The section gives a simplified combinatorial proof that the channel is diagonal in the Majorana basis and explains the associated eigenvalue expression through random permutations.
A.3 Channel eigenvalues
- The combinatorial proof simplifies an earlier finite-frames proof of channel diagonality [ZRM21; HLS00; Wal18].The passage attributes the earlier result to [ZRM21] and its method to finite frames [HLS00; Wal18].Source: §A.3 · p. 7
- The channel is diagonal in the Majorana basis, so each Γµ is an eigenoperator with eigenvalue λm,µ.The proof reduces the claim by linearity to showing the eigenvalue relation for every µ in C2m.Source: §A.3 · p. 7
- The eigenvalue λm,µ is the probability that a uniformly random permutation maps an index sequence of a given length to a diagonal one.The proof connects this probability interpretation to the sequence and diagonal-operator mappings.Source: §A.3 · p. 8
A.4 Second moment of the estimator for general state and observable
Theorem 3 gives a second-moment result for any state and even-degree Hermitian observable, with a combinatorial interpretation of its coefficients in terms of random permutations.
A.4 Second moment of the estimator for general state and observable
- Theorem 3 applies to any state and even-degree Hermitian observable.The supplied theorem statement is truncated before the bound is visible.Source: §A.4 · p. 8
- The coefficient λ_m,μ,μ′ = λ_m,k,k′,a is the probability that a uniformly random permutation maps four disjoint index sequences to diagonal sequences.The sequence sizes are 2k, 2k′, 2a, and 2(m − k − k′ + a).Source: §A.4 · p. 8
A.5 Computation of the estimator
The section gives polynomial-time classical procedures for computing estimators involving fermionic Gaussian states, with an additional result for fixed-even-particle-number Slater determinants.
A.5 Computation of the estimator
- For pure fermionic Gaussian states, the stated inverse-channel trace can be calculated exactly in poly(m) classical time.The trace involves the inverse measurement channel applied to |ψ⟩⟨ψ| and the projector |φ⟩⟨φ|.Source: §A.5 · p. 9
- For fixed, even-particle-number Slater determinants, the corresponding inverse-channel trace for |0⟩⟨ψ| can also be calculated exactly in poly(m) classical time.The restriction to fixed even particle number is explicit.Source: §A.5 · p. 9
- Although the inverse-channel image of |0⟩⟨0| is not a matchgate, it can be expressed as a sum of m + 1 matchgate tensor networks.The passage notes that the proof is deferred.Source: §A.5 · p. 10
- An even-n Slater determinant can be prepared from |1⟩^⊗n|0⟩^⊗(m−n) by a number-preserving fermionic Gaussian unitary.This representation is used for the computation argument.Source: §A.5 · p. 10
A.6 Shadow norm of projector onto zero state
Lemma 2 establishes an upper bound on the shadow norm of the projector onto the zero state.
A.6 Shadow norm of projector onto zero state
- Lemma 2 upper-bounds the shadow norm of the projector onto |0⟩.The supplied statement does not include the bound's expression.Source: §A.6 · p. 12
A.7 Numerical evaluation of the shadow norm
The numerical evaluation examines the shadow-norm bound at two endpoint particle numbers, while the following construction decomposes permutations into structured factors for smaller ensembles.
A.7 Numerical evaluation of the shadow norm
- The plotted endpoint cases are n = 0 and n = m, with conjectured upper bounds F0(m) and F1(m), respectively.Figure 1 presents the numerical evaluation of f(m, n).Source: Figure 1 · p. 13, §A.7 · p. 13
- For m ≤ 50, f(m, n) decreases monotonically as n increases.The numerical evaluation plots the cases n = 0 and n = m alongside conjectured upper bounds F0(m) and F1(m).Source: §A.7 · p. 13, Figure 1 · p. 13
B Smaller ensembles yielding the same channel
- The permutation set is decomposed as Perm(2m) = F · S · P, with F and S described through pairwise swaps.F consists of permutations that swap only within each pair; S consists of permutations that swap adjacent pairs together.Source: B Smaller ensembles yielding the same… · pp. 13–14
C Learning a Slater determinant
The learning procedure estimates the Slater determinant’s 1-RDMs, then extracts a classical description of a determinant whose fidelity with the target is at least 1 − ε_fid with probability at least 1 − δ.
C Learning a Slater determinant
- Because a Slater determinant is uniquely specified by its 1-RDMs, the proof estimates those matrices and converts their approximation error into fidelity error.The extraction step constructs a determinant from the approximated 1-RDMs.Source: C Learning a Slater determinant · p. 14
- With probability at least 1 − δ, the procedure outputs a classical description of a Slater determinant with fidelity at least 1 − ε_fid to the target.The theorem concerns an m-mode, n-electron Slater determinant and fidelity tolerance ε_fid ∈ (0, n/m].Source: C Learning a Slater determinant · p. 14
- An additive ε_shdw approximation to all degree-2 Majorana operators gives an additive ε_shdw approximation to every entry of the matrix R.The passage identifies these degree-2 Majorana expectations with the 1-RDMs.Source: C Learning a Slater determinant · p. 15
- The spectral argument separates the estimated matrix’s eigenvalues into groups near 1 and 0, identifying the subspace used to recover the determinant.The proof invokes the Gershgorin circle theorem to bound how much estimation error the two subspaces can tolerate.Source: C Learning a Slater determinant · pp. 16–17