Classical shadows do not reconstruct a full quantum state by default. They store a compact randomised measurement record designed so that many observables can later be predicted from the same data.
Full tomography is expensive because a generic n-qubit state has 4ⁿ−1 independent real parameters. But many experiments do not need a complete density matrix. They need expectation values of selected observables, fidelities with target states, local correlations or Hamiltonian energies.
The classical-shadows framework of Huang, Kueng and Preskill showed that randomised measurements can build a reusable classical representation whose sample complexity for predicting many chosen properties grows only logarithmically with the number of target observables, multiplied by a property-dependent shadow norm. [1]
Randomise measurement → record outcome and setting → invert the average measurement channel → store a shadow snapshot → aggregate robustly → predict chosen observables.
1. The problem classical shadows are solving
Suppose a state ρ is prepared repeatedly and we want expectation values Tr(Ojρ) for many observables O₁,…,OM. Measuring each observable independently can waste data because the same state preparations may contain information relevant to several targets.
Classical shadows use a random measurement ensemble chosen before the targets need to be evaluated. The stored classical data can then be reused to estimate many observables later, provided those observables are compatible with the measurement ensemble in the sample-complexity sense.
The method therefore changes the output of an experiment. Instead of one estimate per measurement programme, we obtain a collection of random snapshots intended for broad downstream prediction.
2. One shadow snapshot
Choose a random unitary U from a specified ensemble, apply it to ρ and measure in the computational basis. Record the basis outcome b and the unitary U.
The measured post-rotation projector is
U†|b⟩⟨b|U.
Averaging this random projector over the measurement procedure defines a linear measurement channel M acting on ρ. If M is invertible on the operator space of interest, define the classical shadow snapshot
ρ_hat=M^{-1}(U†|b⟩⟨b|U).
The snapshot need not itself be a physical density matrix. Its importance is unbiasedness: the average snapshot equals the true state under the ideal measurement model.
3. Unbiasedness is the first key property
By construction,
E[ρ_hat]=ρ.
Therefore, for any observable O in the domain of the estimator,
E[Tr(Oρ_hat)]=Tr(Oρ).
One snapshot is noisy, but averaging many snapshots gives an unbiased observable estimator in the ideal model.
This is different from reconstructing a positive density matrix. A shadow snapshot is an estimator object, not necessarily a physical state.
4. Local Pauli shadows
A particularly practical scheme chooses independently on each qubit one of the X, Y or Z measurement bases uniformly at random.
For one qubit, if the measured eigenstate projector is P, the corresponding local snapshot is
ρ_hat=3P-I.
For n qubits under independent local Pauli measurements, the full snapshot is the tensor product of these one-qubit inverse-channel factors.
This scheme is shallow and hardware-friendly, though its efficiency depends strongly on the locality and Pauli weight of the observables being predicted.
5. Worked one-qubit snapshot
Suppose the random basis is Z and the outcome is |0⟩. Then P=|0⟩⟨0|=(I+Z)/2.
The shadow snapshot is
3P-I=(I+3Z)/2.
Its Z expectation is Tr[Zρ_hat]=3. That is outside the physical expectation range [−1,1], which is allowed because the snapshot is not a physical state.
Averaging over the random choice of measurement basis and outcomes restores the correct physical expectation.
6. Estimating a Pauli observable
Consider estimating ⟨Z⟩ from local Pauli shadows of one qubit. Only snapshots measured in the Z basis carry nonzero contribution to the Z estimator. When Z is chosen, the contribution is +3 or −3 depending on outcome; when X or Y is chosen, the contribution is zero.
The basis is selected with probability 1/3, so the factor of three exactly compensates the subsampling. The estimator remains unbiased.
The price is variance. Randomising measurement settings makes the record reusable, but each specific observable uses only a fraction of the snapshots effectively.
7. Pauli weight controls local-shadow variance
For a Pauli string acting nontrivially on k qubits, a local Pauli shadow matches all required bases with probability 3−k. The inverse-channel weighting contributes a corresponding factor.
Thus high-weight observables can be expensive under local shadows, while low-weight local observables are natural targets.
This is why “sample complexity independent of system size” must be read together with the observable-dependent shadow norm. A highly nonlocal target can carry exponential cost even when n does not appear explicitly in a simplified theorem statement.
8. Global Clifford shadows
Another important ensemble applies a random global Clifford unitary before computational-basis measurement.
Global Clifford shadows can be efficient for low-rank global observables such as fidelity with a pure target state. But implementing a deep random global Clifford may be costly on noisy hardware.
The measurement ensemble should therefore match both the observable family and the hardware architecture.
9. Why one measurement ensemble cannot be best for everything
Local Pauli measurements favour local Pauli observables. Global Clifford measurements favour some low-rank global properties. Locally entangled measurement ensembles create intermediate trade-offs between circuit depth and sample complexity. [2]
Measurement design is therefore an optimisation problem. Use prior knowledge of likely targets when available rather than insisting on a completely universal scheme.
10. Shadow norm as an observable difficulty measure
The classical-shadow theorem bounds the number of snapshots needed to predict a family of observables in terms of a shadow norm associated with the measurement channel and each target observable.
Informally, the shadow norm measures how much estimator variance the inverse measurement channel creates for that observable.
A small shadow norm means the observable is naturally visible to the chosen measurement ensemble. A large shadow norm means the estimator amplifies rare measurement coincidences strongly.
11. Logarithmic dependence on the number of observables
For M fixed target observables, robust classical-shadow estimators can achieve simultaneous prediction with a number of snapshots scaling like
O(max_j ||O_j||_shadow² log(M/δ)/ε²)
up to theorem-specific constants and conventions, for additive error ε and failure probability δ.
The logarithm is the important collective advantage: the same data can support many predictions. The observable-dependent norm prevents this from being a universal free lunch.
12. Median of means
Single-shot shadow estimators can have heavy tails. A robust strategy divides snapshots into groups, averages within each group and then takes the median of those group means for each observable.
This median-of-means construction converts variance control into high-probability error control without assuming Gaussian tails.
The exact group count and block size depend on the desired confidence and theorem constants. Reporting a median-of-means estimate without the grouping rule makes the statistical guarantee impossible to reproduce.
13. Worked two-qubit local observable
Suppose the target is ⟨Z⊗Z⟩. A local Pauli shadow contributes only when both qubits were measured in Z, which happens with probability 1/9.
When that setting occurs, the observed eigenvalue product is ±1 and the inverse-channel weight contributes 3²=9, giving a single-shot estimator ±9. All other basis choices contribute zero.
The estimator is unbiased but high variance compared with a dedicated ZZ measurement. Its value is reuse: the same random dataset can simultaneously estimate XX, XY, local Z values and many other Pauli observables.
14. Dedicated measurements can beat shadows for one known observable
If only ZZ matters, measuring ZZ every time gives each shot direct ±1 information. Local shadows deliberately spend most shots in other bases.
Thus classical shadows are not automatically sample-optimal for a single known target. They become attractive when many observables share one measurement budget or when the target list is selected after data collection.
Universality has a variance price. Reuse is the resource classical shadows are designed to buy.
15. Predicting Hamiltonian energy
If H=Σ_j h_jP_j is a sum of Pauli strings, estimate each ⟨P_j⟩ from the same shadow dataset and combine
⟨H⟩=Σ_jh_j⟨P_j⟩.
Correlations between term estimators matter because they reuse the same snapshots. The variance of the total energy is not generally the sum of independent term variances.
Locally biased shadow schemes can choose X,Y,Z probabilities non-uniformly to reduce variance for a known Hamiltonian family. [3]
16. Fidelity estimation
For a pure target |ψ⟩, fidelity is F=⟨ψ|ρ|ψ⟩, which is an expectation value of the rank-one projector |ψ⟩⟨ψ|.
A shadow dataset can estimate this quantity without reconstructing a full physical density matrix. The efficiency depends on the measurement ensemble and target structure.
For stabilizer targets, Clifford-structured measurements can be particularly natural because overlaps can be processed efficiently.
17. Nonlinear properties need more care
Purity Tr(ρ²), Rényi entropies and other nonlinear functions cannot be estimated by simply plugging one unbiased shadow snapshot into the nonlinear formula and expecting an unbiased result.
Use independent snapshots in U-statistic-like combinations or specialised estimators designed for the nonlinear quantity.
Experimental work has demonstrated nonlinear-property estimation from classical-shadow data, but the estimator structure matters. [4]
18. Derandomisation
If the target observables are known in advance, measurement settings can be chosen deterministically or adaptively to cover them more efficiently than purely uniform random sampling.
Derandomised classical shadows keep the reuse philosophy while reducing wasted settings for a specific target family.
This changes the promise. The measurement record is no longer equally universal for arbitrary future observables because the design exploited prior target knowledge.
19. Calibration errors bias the inverse channel
The inverse measurement map M⁻¹ is derived from an assumed measurement ensemble. If implemented rotations or readout effects differ from that model, the estimator can become biased.
Randomised measurement does not automatically average away systematic calibration error. Characterisation or error-mitigation models may be needed.
A small statistical error bar around a biased shadow estimate is still a biased answer.
20. Drift and non-identical preparations
The method assumes repeated snapshots are generated from the same state distribution, or from a model whose time variation is explicitly handled.
If the source drifts, the shadow may describe a time average. Randomising measurement order can reduce confounding between time and basis choice, but it does not remove the underlying nonstationarity.
Track calibration and time stamps when the experiment lasts long enough for drift to matter.
21. Symmetry-aware shadows
Known symmetry can reduce sample complexity if the randomisation and estimator are designed to respect it. Recent work develops shadow protocols targeted to gauge-invariant observable families and other structured settings. [5]
There is a trade-off: richer entangled or symmetry-adapted measurements can reduce sample counts while increasing circuit depth and calibration burden.
Sample complexity and circuit complexity should be reported together.
22. Common misconception: a classical shadow is a compressed density matrix
Not in the ordinary sense. A shadow is a random classical dataset and associated estimator construction. Individual snapshots can be nonphysical and need not define a low-rank approximation to ρ.
23. Common misconception: log M means arbitrarily many observables are free
The logarithmic dependence is multiplied by the largest relevant shadow norm and by precision and confidence factors. Difficult observables can remain expensive.
24. Common misconception: randomised measurements always beat tomography
If a complete physical state model is genuinely required, tomography solves a different task. Shadows are powerful when the scientific output is a family of predictions rather than a full density matrix.
25. Worked synthesis problem
Suppose a two-qubit experiment wants to estimate 300 local and two-body Pauli observables from one dataset.
Step 1: Choose a measurement ensemble. Local random Pauli measurements are natural because the targets are Pauli strings of weight at most two.
Step 2: Record snapshots. For each shot, store the X/Y/Z basis selected on each qubit and the ±1 outcomes.
Step 3: Build estimators. A weight-one Pauli term is matched with probability 1/3 and uses factor 3. A weight-two term is matched with probability 1/9 and uses factor 9.
Step 4: Aggregate robustly. Use block means and medians if a high-probability simultaneous guarantee is required.
Step 5: Compare against dedicated measurement. If only one ZZ correlator later becomes mission-critical, dedicate additional ZZ shots rather than assuming the shadow dataset is variance-optimal for that one target.
26. Practice set
- What is the main output of a classical-shadow experiment?
- Why can a shadow snapshot be nonphysical?
- What property makes snapshot averages useful?
- What bases are used in local Pauli shadows?
- What is the one-qubit inverse-channel snapshot for measured projector P?
- Why are high-weight Pauli observables expensive under local shadows?
- What does the shadow norm measure informally?
- Why does log M matter?
- What is median of means used for?
- Why can a dedicated measurement beat shadows for one observable?
- Why do nonlinear properties require specialised estimators?
- Name two non-statistical errors that can bias a shadow estimate.
Answers
- A reusable classical record of random measurement settings/outcomes plus the inverse-channel estimator rule.
- It is an unbiased estimator object, not required to satisfy positivity on each shot.
E[ρ_hat]=ρ.- Random X, Y or Z basis independently on each qubit.
3P-I.- All required local bases must match simultaneously, with probability 3−k for weight k.
- The variance amplification of an observable under the chosen inverse measurement channel.
- Many target predictions can share one dataset with only logarithmic dependence on target count.
- Robust high-probability estimation under heavy-tailed snapshot values.
- Dedicated shots never waste settings on other observables.
- Unbiasedness is not preserved by arbitrary nonlinear functions of one snapshot.
- Calibration error, readout error, gate error in randomisation or source drift.
Sources and further study
[1] Hsin-Yuan Huang, Richard Kueng and John Preskill, Predicting many properties of a quantum system from very few measurements, Nature Physics 16, 1050–1057 (2020). The foundational classical-shadows framework.
[2] Matteo Ippoliti, Classical shadows based on locally-entangled measurements, Quantum 8, 1293 (2024). An analysis of circuit-depth versus sample-complexity trade-offs beyond single-qubit random measurements.
[3] Charles Hadfield and colleagues, Measurements of Quantum Hamiltonians with Locally-Biased Classical Shadows. This work adapts local random measurement probabilities to target observables.
[4] Christian Kokail and colleagues, Experimental Quantum State Measurement with Classical Shadows. An experimental demonstration of many-observable and nonlinear-property estimation.
[5] Jacob Bringewatt and colleagues, Classical shadows for sample-efficient measurements of gauge-invariant observables, Quantum 10, 2127 (2026). A recent example of symmetry-aware shadow design and explicit circuit/sample trade-offs.
Continue through Quantum Mathematics
Guide 21: Quantum State Tomography, Informational Completeness and Physical Reconstruction solves the full-state reconstruction problem. Guide 22: Quantum Process Tomography, Choi Matrices and Channel Verification reconstructs channels. Guide 24: Randomised Benchmarking, Gate Fidelity and Error Budgets turns random sequences into operational gate-error estimates.