Virtual Distillation: Quantum Error Mitigation

• Virtual Distillation (VD) is a quantum-statistical protocol that virtually purifies a state by reconstructing its measurement statistics from multiple noisy copies, bypassing the need for physical state preparation.
• It utilizes multi-copy techniques such as cyclic shift operators and signed combinations of free operations to amplify the dominant eigenvector and reduce error impacts.
• Experimental and theoretical studies demonstrate that VD improves coherence, entanglement, and variational quantum eigensolver (VQE) performance while managing sampling overhead and circuit noise.

Virtual distillation (VD) denotes a family of quantum-statistical distillation protocols in which the target of purification is shifted from a physically prepared quantum state to the measurement statistics associated with a cleaner object. In its original error-mitigation form, VD estimates observables on the nonlinear state $\rho^M/\mathrm{Tr}(\rho^M)$ by jointly measuring $M$ noisy copies of $\rho$, thereby amplifying the dominant eigenspace without explicitly preparing the purified state (Huggins et al., 2020). In the broader resource-theoretic formulation, often called virtual resource distillation, the objective is to reproduce the measurement statistics of an ideal resource by signed combinations of free operations and classical postprocessing, rather than to output the ideal resource state itself (Yuan et al., 2023). This distinction between physical purification and statistical reconstruction underlies later extensions to coherence, entanglement, magic, quantum channels, higher-order processes, and near-term experimental implementations (Takagi et al., 2024).

1. Conceptual definition and scope

In the multi-copy error-mitigation setting, VD starts from a noisy density matrix

$\rho=\sum_i p_i |i\rangle\langle i|,$

so that

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$

As $M$ increases, the weights of subdominant eigencomponents are suppressed by powers of $p_i/p_0$, and the normalized state approaches the dominant eigenvector exponentially quickly. The method is “virtual” because it reconstructs expectation values as though the cleaner state were available, without physically preparing that state (Huggins et al., 2020).

In virtual resource distillation, the same adjective has a more explicitly operational meaning. Conventional distillation demands that a free operation physically transform a noisy resource into a target pure resource, whereas virtual distillation only demands that the resulting measurement statistics match those of the target. The relevant map may be a signed linear combination of free operations,

$$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$

so the final object need not be a physical density matrix or a CPTP map. What is reconstructed are expectation values, and those expectation values can still be used for tasks such as tomography, teleportation, or computation provided the resource is eventually measured (Yuan et al., 2023, Zhang et al., 2023).

2. Mathematical structure

For an observable $O$, the standard VD estimator is

$$\langle O\rangle_{\mathrm{VD}}=\frac{\mathrm{Tr}(O\rho^M)}{\mathrm{Tr}(\rho^M)}.$$

The central identity converts this nonlinear functional into a measurable multi-copy observable,

$M$0

where $M$1 is the cyclic shift operator on the $M$2 copies. For $M$3, the denominator is the purity $M$4, and the numerator and denominator can be implemented either with ancilla-controlled SWAP-type constructions or with circuits that diagonalize the relevant permutation operators before computational-basis measurement (Li et al., 2023, Karim et al., 2024).

The resource-theoretic framework introduces a one-shot conventional distillation rate $M$5, a virtual overhead $M$6, and a virtual distillation rate

$M$7

The overhead is the signed Monte Carlo cost of simulating $M$8 target copies through free operations and classical postprocessing. Estimating an expectation value to additive accuracy $M$9 with failure probability $\rho$0 requires

$\rho$1

samples conventionally, but

$\rho$2

samples under virtual distillation. The square in $\rho$3 is therefore the basic statistical price of virtuality (Yuan et al., 2023).

Later work generalized these quantities from states to channels and combs, using trace distance for states, diamond distance for channels, and a comb norm for higher-order processes. It also supplied computable LP/SDP characterizations through dual witness formulations, overlap quantities such as $\rho$4, and bounds expressed via resource fidelity and robustness. In theories admitting a suitable generalized twirling map, the overhead becomes exactly computable rather than merely bounded (Takagi et al., 2024).

3. Resource-theoretic generalization

A central structural result is that virtual distillation weakly dominates conventional distillation:

$\rho$5

This encodes the basic relaxation: physical state synthesis is stricter than statistical simulation. The framework therefore permits nonzero virtual rates even when conventional zero-error distillation is impossible, including cases where multi-copy coherent manipulation would be operationally prohibitive. The same logic motivated early applications to coherence, entanglement, magic distillation, and teleportation in distributed quantum-computing settings (Yuan et al., 2023).

Within the general framework, several exact formulas become available. For coherence, under MIO or DIO and for any qubit state,

$\rho$6

and for any single-qubit state and any $\rho$7,

$\rho$8

For entanglement, the framework gives exact SDP characterizations under separability-preserving and PPT-preserving operations, and it identifies a notable limitation: bound entanglement does not help virtual distillation, in the sense that the overhead can coincide with the separable-state benchmark. For magic, exact overhead formulas are obtained for noisy $\rho$9-states and for the qutrit Strange state through stabilizer twirling. The same formalism extends to dynamical resources, including quantum memory, quantum communication, and non-Markovian combs, and it yields explicit channel-level formulas such as

$\rho=\sum_i p_i |i\rangle\langle i|,$0

for the $\rho=\sum_i p_i |i\rangle\langle i|,$1-dimensional depolarizing channel and

$\rho=\sum_i p_i |i\rangle\langle i|,$2

for amplitude damping, showing that nonzero virtual rates can persist even when conventional quantum capacity vanishes (Takagi et al., 2024).

4. Implementations and circuit architectures

Near-term VD implementations have concentrated on shallow multi-copy constructions. Two standard realizations are an ancilla-controlled SWAP approach and a diagonalizing-gate approach; the latter is often preferred because controlled-SWAP bridges between copies can be too noisy on NISQ devices. In practical studies the focus is usually on $\rho=\sum_i p_i |i\rangle\langle i|,$3, because larger copy numbers further increase circuit size, routing overhead, and crosstalk sensitivity. A major hardware-oriented refinement combines VD with circuit cutting: state-preparation fragments are executed on hardware, while diagonalizing-gate fragments are simulated classically and recombined. In the formulation studied for constrained-connectivity devices, the hardware runtime scales as $\rho=\sum_i p_i |i\rangle\langle i|,$4 and the classical simulation cost as $\rho=\sum_i p_i |i\rangle\langle i|,$5, while direct deployment of the full VD circuit can suffer SWAP-induced overhead as large as $\rho=\sum_i p_i |i\rangle\langle i|,$6 (Li et al., 2023).

A separate line of work addressed the observable dependence of duplicate-circuit VD. The original B-gate constructions work cleanly for observables commuting with the swap operator $\rho=\sum_i p_i |i\rangle\langle i|,$7, effectively 1-local measurements, but chemistry Hamiltonians include multi-qubit Pauli strings such as $\rho=\sum_i p_i |i\rangle\langle i|,$8 and $\rho=\sum_i p_i |i\rangle\langle i|,$9. Deterministic low-depth decompositions based on pairwise CNOT and Hadamard projections were introduced to reconstruct the required VD statistics for arbitrary multi-qubit observables while keeping depth linear in qubit number. In VQE settings, this matters because the variational principle can be violated when the $$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$0 measurement circuit is noisier than the observable measurement circuits; the matched low-depth decomposition was designed to preserve the variational bound more reliably than B gates (Karim et al., 2024).

Circuit-noise-resilient VD (CNR-VD) targets a different failure mode: noise in the VD circuit itself. For a calibration state $$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$1, the central estimator is

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$2

Under stochastic Pauli-like circuit noise, this calibration is intended to divide out the effective gate-noise factor. A complementary architectural modification, called noise dilution, studies settings with unavoidable noisy peripherals such as delay lines and distributes the peripheral uniformly across the circuit. Under multiqubit loss and Pauli noise, the average mitigation performance improves monotonically with the number of diluted layers, and analytical and numerical evidence indicate that second-order distillation is generally sufficient for near-optimal mitigation (Xu et al., 2023, Teo et al., 2022).

VD has also been analyzed in modular and networked architectures. Three edge-case constructions were compared: cyclic rotation (CR), qubit-efficient cyclic rotation (QECR), and a constant-depth brickwork (BW) implementation. QECR minimizes qubit count at the cost of deep sequential execution; BW uses more ancillas and a GHZ resource but achieves constant depth and fewer controlled-SWAPs. In realistic networked ion-trap simulations, the constant-depth BW implementation consistently outperformed the minimal-qubit QECR implementation, and noise in local gates rather than remote entangling operations was identified as the main limiting factor (Araki et al., 25 Mar 2025).

5. Experimental demonstrations

A direct experimental realization on a photonic platform established both coherence and entanglement versions of virtual resource distillation. In the coherence experiment, a single photon was encoded as a single ququart with basis

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$3

and the protocol virtually distilled the four-dimensional maximally coherent state

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$4

from the two-dimensional maximally coherent input

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$5

a task described as impossible in conventional coherence distillation from a single two-dimensional coherent state. The distilled state achieved fidelity

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$6

whereas the input had

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$7

The relative entropy of coherence increased from

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$8

to

$$\frac{\rho^M}{\mathrm{Tr}(\rho^M)}=\frac{\sum_i p_i^M |i\rangle\langle i|}{\sum_j p_j^M}.$$9

For entanglement, the experiment used Werner states

$M$0

prepared eight values of $M$1, and reported that the fidelity of the virtually distilled state was close to $M$2 for all $M$3, while the undistilled Werner-state fidelity increased only linearly with $M$4. The negativity of the virtually distilled state was likewise close to the Bell-state value even though several input states with $M$5 were separable. In the teleportation demonstration, the virtually distilled resource produced fidelity near the ideal entanglement-assisted value, and for $M$6 it surpassed the classical threshold

$M$7

where the undistilled resource fell below it. The authors emphasized that this is only virtual or effective entanglement, paid for by sampling overhead and classical postprocessing (Zhang et al., 2023).

A second experimental direction studied VD-enhanced VQE for MaxCut with a RealAmplitudes ansatz, Qiskit v0.39.0, Qiskit Aer, and the 27-qubit device ibm_hanoi. On the real device, for the 4-qubit circuit the reported absolute errors were $M$8 with no mitigation, $M$9 with VD, $p_i/p_0$0 with VD plus zero-noise extrapolation, and $p_i/p_0$1 with VD plus circuit cutting. For the 6-qubit circuit the corresponding errors were $p_i/p_0$2, $p_i/p_0$3, $p_i/p_0$4, and $p_i/p_0$5. The same study reported that for a 10-qubit VQE instance the cut subcircuits were still too noisy to yield satisfactory pairwise distributions, showing that the improvement has a circuit-size limit (Li et al., 2023).

Low-depth duplicate-circuit VD was also demonstrated on ibm_hanoi for molecular VQE, including H$p_i/p_0$6 in a two-qubit tapered mapping, H$p_i/p_0$7 with three qubits, and H$p_i/p_0$8 with four qubits. The corrected energies were reported to move toward the ideal dissociation curves, and the matched low-depth decomposition preserved the variational principle in simulations where B-gate-based VD could violate it (Karim et al., 2024).

For circuit-noise-resilient calibration, numerical studies reported that CNR-VD can boost accuracy by up to tenfold over standard VD, remain beneficial under high noise where standard VD fails, and extend beyond VD to general Hadamard-test circuits (Xu et al., 2023).

6. Limitations, finite-shot behavior, and current directions

VD does not remove all error indiscriminately. In the original spectral analysis, its asymptotic output approaches the dominant eigenvector of the noisy state, not necessarily the ideal state, so drift of that dominant eigenvector creates a residual noise floor. In resource-theoretic form, the protocol pays a sampling overhead $p_i/p_0$9, with effective sample complexity scaling as $$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$0. In hardware realizations, extra bridging gates, routing overhead, instruction crosstalk, and readout crosstalk can erase the theoretical advantage of multi-copy purification, especially as circuit size grows (Huggins et al., 2020, Yuan et al., 2023, Li et al., 2023).

A recent finite-shot operating-window theory made this limitation explicit by treating VD as a quotient estimator with

$$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$1

The certified mean-squared-error expansion is

$$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$2

with an explicit denominator-concentration certificate and a sufficient reliability threshold

$$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$3

In the tested QAOA instances, idealized VD operating windows existed in white-box models, but realistic interferometry overhead and denominator instability often erased them, and calibrated symmetry verification was the practical winner on the studied devices and simulations (Alfaro, 13 Jun 2026).

Hybrid strategies have therefore become increasingly important. Symmetry-averaged virtual distillation (SAVD) applies symmetry averaging before VD, treating the averaging step as a state-preconditioning layer. If

$$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$4

then SAVD performs

$$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$5

so that symmetry averaging first restructures the residual error and VD subsequently amplifies the dominant eigencomponent of the averaged state. Numerical demonstrations on an isotropic Heisenberg chain reported improved accuracy in the simultaneous presence of coherent algorithmic errors and hardware noise (He et al., 20 Jun 2026).

A recurring empirical theme is that moderate copy number is often the practical regime. In noise-dilution analyses, second-order distillation was generally sufficient for near-optimal mitigation under both multiqubit loss and Pauli channels, while many near-term hardware studies concentrated on $$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$6 because larger $$\tilde\Gamma=\gamma_+\Gamma_+ - \gamma_-\Gamma_-, \qquad \gamma_+-\gamma_-=1,\ \gamma_\pm\ge 0,$$7 rapidly increases depth, routing cost, and instability (Teo et al., 2022). The resulting picture is neither that VD is universally superior nor that it is merely formal; rather, its performance is regime-dependent, architecture-dependent, and tightly controlled by the distinction between physical purification and statistical reconstruction.