Weak Simulation of Quantum Resources

• Weak simulation is a sampling task that reproduces quantum output statistics using approximate algorithms without computing full quantum states.
• It employs methods like quasiprobability decomposition, rejection sampling, and state compression to efficiently approximate quantum distributions.
• The approach balances trade-offs in computational resources such as memory, communication, and energy, revealing threshold behaviors in simulating quantum processes.

Weak simulation of quantum resources denotes a family of sampling tasks in which a classical or restricted quantum procedure reproduces, exactly or approximately, the output distribution of a target quantum process without computing the full state or all amplitudes explicitly. In the circuit setting, the task is to output bit strings distributed according to the Born rule of a quantum computation; in resource-theoretic settings, it is to sample from the statistics of a target resource state or operation using only a restricted class of free states, free operations, or noisy surrogates. The notion is operationally distinct from strong simulation, which computes probabilities or amplitudes, and from expectation-value estimation, which need not produce samples at all (Hillmich et al., 2020, Onishi et al., 26 Mar 2026, Naik et al., 27 Jan 2025).

1. Definitions and operational scope

For an $n$-qubit circuit $C$ with output amplitudes $\alpha_x$, weak simulation asks for a randomized classical algorithm that outputs $x\in\{0,1\}^n$ with probability $P_C(x)=|\alpha_x|^2$, or approximately so up to small total-variation error. A standard approximate formulation requires $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$, where $S$ is the simulator’s output distribution (Hillmich et al., 2020).

In resource theories, the target is often a state $\rho\notin F$, where $F$ is a convex set of free states preparable by restricted operations $O$. The objective is to sample bit strings $C$0 according to $C$1 using only preparations from $C$2, operations in $C$3, and computational-basis measurements. A procedure is an $C$4-weak simulator of $C$5 if, with probability at least $C$6, it outputs samples from a distribution $C$7 satisfying $C$8, and its sampling cost is the total number of free-state preparations (Onishi et al., 26 Mar 2026).

For channels, the literature distinguishes weak or local simulation from strong or generic simulation. In the weak scenario, Alice communicates finite classical information to Bob so that Bob can reproduce the outcome statistics of any POVM on the channel output alone. In the strong or generic scenario, Bob may additionally hold part of an unknown ancillary system, possibly entangled with a larger environment, and must reproduce the statistics of arbitrary joint POVMs on the output together with that ancillary system. This distinction is operationally decisive: reproducing local marginals is strictly weaker than reproducing all bipartite Born-rule probabilities, especially when entangled effects are allowed (Naik et al., 27 Jan 2025).

2. Quasiprobability methods and weak distillation

A central resource-theoretic construction begins from a quasiprobability decomposition

$C$9

with $\alpha_x$0 and negativity $\alpha_x$1. In virtual resource distillation and related protocols, importance sampling over the $\alpha_x$2 enables estimation of expectation values, but this only yields quantities such as $\alpha_x$3, not samples from the Born distribution of $\alpha_x$4 itself (Onishi et al., 26 Mar 2026).

Weak distillation lifts this expectation-value paradigm to a sampling protocol. The construction first estimates an acceptance-ratio table $\alpha_x$5, where

$\alpha_x$6

by sampling preparable states according to $\alpha_x$7, measuring in the computational basis, and forming empirical counts $\alpha_x$8 and $\alpha_x$9 for positive and negative coefficients. One then sets

$x\in\{0,1\}^n$0

draws $x\in\{0,1\}^n$1, and accepts $x\in\{0,1\}^n$2 with probability $x\in\{0,1\}^n$3; rejected draws are repeated until acceptance. The accepted distribution $x\in\{0,1\}^n$4 is within $x\in\{0,1\}^n$5 of $x\in\{0,1\}^n$6 in total-variation distance with high probability if the ratio table is estimated accurately and sufficiently many support samples are used (Onishi et al., 26 Mar 2026).

The complexity gain over naive estimation can be substantial. For an optimal two-term decomposition $x\in\{0,1\}^n$7, the rejection-based weak simulator has sampling cost $x\in\{0,1\}^n$8 when $x\in\{0,1\}^n$9, whereas naive estimation scales as $P_C(x)=|\alpha_x|^2$0 even as the negativity approaches zero. In the limit $P_C(x)=|\alpha_x|^2$1, the rejection-based cost approaches $P_C(x)=|\alpha_x|^2$2, recovering direct single-sample behavior on free states. Numerical examples on $P_C(x)=|\alpha_x|^2$3–$P_C(x)=|\alpha_x|^2$4 qubits show that error mitigation under local depolarizing noise $P_C(x)=|\alpha_x|^2$5, entanglement distillation of isotropic states with noise $P_C(x)=|\alpha_x|^2$6, and magic-state injection for a 5-qubit IQP circuit with dephasing $P_C(x)=|\alpha_x|^2$7 can all reach target TVD $P_C(x)=|\alpha_x|^2$8 with $P_C(x)=|\alpha_x|^2$9 samples, whereas estimation-based methods require $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$0–$\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$1 samples for comparable accuracy (Onishi et al., 26 Mar 2026).

This framework introduces a notion of distillation that remains entirely virtual: it does not physically produce the target resource state, uses only single-copy preparations and no joint operations, and removes the restriction that only expectation values are accessible. A plausible implication is that quasiprobability negativity functions simultaneously as a resource monotone and as a direct predictor of weak-simulation overhead.

3. Approximate weak simulation of noisy and compressible circuits

One major line of work studies weak simulation by compressing many-body states that arise in noisy or imperfect devices. For random circuits on $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$2 qubits and two-qubit gate depth $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$3, matrix product state compression represents the wavefunction with bond dimension $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$4 and induces an effective per-gate error $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$5 defined through the fidelity relation

$\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$6

Single-qubit gates update the MPS exactly at $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$7 cost, while neighboring two-qubit gates require an SVD truncated to the top $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$8 singular values and cost $\|\mathrm{Dist}(S)-P_C\|_1\le \epsilon$9. The total simulation time is $S$0 and the memory footprint is $S$1 (Zhou et al., 2020).

The compression error is controlled by entanglement. Across a bipartition, the entanglement entropy satisfies $S$2, so exact representation of target entropy $S$3 requires $S$4. After each two-qubit gate, the retained singular values determine a local fidelity

$S$5

and empirically the local error $S$6 decreases toward a nonzero floor $S$7 as $S$8. For $S$9, one finds $\rho\notin F$0 with $\rho\notin F$1–$\rho\notin F$2, so the time cost remains polynomial in $\rho\notin F$3. By contrast, pushing below $\rho\notin F$4 requires exponentially growing $\rho\notin F$5 and hence exponential time. In 1D random circuits with nearest-neighbor $\rho\notin F$6 gates, $\rho\notin F$7, and $\rho\notin F$8, the observed floor is $\rho\notin F$9–$F$0 with $F$1. For a 2D $F$2-qubit layout of depth $F$3, grouping strategies such as $F$4 reach $F$5 in a few hours on one core with $F$6 GB memory; more complex iSWAP+$F$7 gates require a split-and-merge scheme and reach roughly $F$8 per-gate fidelity in a few hours (Zhou et al., 2020).

A different compression route uses edge-weighted quantum decision diagrams. A state on $F$9 qubits is stored as a rooted DAG whose paths encode amplitudes via products of complex edge weights. Weak simulation then proceeds by computing downstream probabilities

$O$0

for all nodes in $O$1 time and space, where $O$2 is the number of DD nodes, normalizing local choice probabilities at each node, and descending from the root by sampling one bit per qubit. The precomputation cost is $O$3, each sample costs $O$4, and in practical instances $O$5. By comparison, the state-vector plus prefix-sum approach requires $O$6 time and space before sampling, though each sample still costs $O$7. Benchmarks reported for generating $O$8 samples include $O$9, where the vector-based method is out of memory while the DD size is approximately $C$00, and $C$01, where vector-based sampling takes $C$02 seconds versus $C$03 seconds for DD-based sampling (Hillmich et al., 2020).

Together, these results show that weak simulators can exploit structural compressibility even when strong simulation by explicit amplitude storage is prohibitive. This suggests a recurrent principle: practical weak simulation is governed less by Hilbert-space dimension itself than by whether the relevant probability distribution can be accessed through low-entanglement or graph-compressed state representations.

4. Complexity frontiers for structured and noisy-resource models

For matchgate computations, weak simulability depends sharply on supplementary resources. Exact polynomial-time weak simulation holds for $C$04, for $C$05, and for $C$06. When only a few adaptive measurements are allowed and the input is a tensor product of blocks of at most $C$07 qubits, the runtime is

$C$08

so $C$09 gives polynomial time and $C$10 gives $C$11. When the input is otherwise product except for one $C$12-qubit entangled block, the runtime is

$C$13

which is again polynomial for $C$14 or $C$15. With unrestricted polynomial supplies of magic states plus arbitrary adaptive measurements, or magic states plus many-line final measurement, the model becomes QC-hard or $C$16-hard in the strong sense, and there is strong evidence that $C$17 is not even weakly simulable (Hebenstreit et al., 2020).

For universal Clifford+$C$18 circuits, weak simulation to additive error $C$19 is $C$20-complete and is expected to scale exponentially in the number $C$21 of $C$22 gates. Approximate stabilizer decomposition of $C$23 yields stabilizer extent $C$24, and the standard SPARSIFY approach gives cost $C$25 when $C$26. Correlated $C$27 sampling replaces i.i.d. sampling by blockwise correlated samples constructed from highly separated bit strings, reducing the worst-case bound to

$C$28

when $C$29. The leading exponential behavior remains, but the finite-$C$30 correction shows that the approximate stabilizer decomposition is not multiplicative at finite $C$31, despite multiplicativity of stabilizer extent (Kocia, 2021).

A related transition appears when only the injected magic resources are noisy. In the resource-centric noise model, a free backbone $C$32 is augmented by $C$33 noisy magic inputs $C$34, and the simulator targets TVD $C$35. The algorithm samples an ensemble decomposition of $C$36, truncates trajectories with too many genuinely magic draws using a Chernoff/KL bound, and approximates the remaining state by a superposition of at most

$C$37

free basis states, where $C$38 is the number of magic draws and $C$39. Runtime is $C$40. In the qubit dephasing case there is an exact stabilizer-only regime at

$C$41

where $C$42 and $C$43. More generally, polynomial time holds when the effective magic contribution is $C$44. For fermion-loss and fermion-dephasing channels, the paper gives corresponding polynomial-time regions in terms of $C$45 and $C$46, with thresholds $C$47 and $C$48, respectively. Numerical estimates show sharp drops in the worst-case free rank $C$49 near these thresholds, with qubit $C$50 instances falling from $C$51 to $C$52 as the dephasing strength increases, and $C$53 exactly beyond the stabilizer threshold (Heo et al., 15 Jan 2026).

These results identify several distinct mechanisms by which weak simulation becomes tractable: bounded adaptivity, limited entangled support, correlated sampling over stabilizer decompositions, and noise-induced suppression of non-free trajectories. A plausible implication is that “hard” quantum resources are often best understood not as binary features but as quantities with threshold behavior under operational restrictions.

5. Channels, instruments, and the boundary between local and generic simulation

For quantum channels, the gap between weak and generic simulation is explicit. A perfect qubit channel can be weakly simulated in the local sense, but any finite-bit classical protocol fails for generic simulation if Bob must reproduce the statistics of all joint bipartite POVMs on the output together with an unknown ancillary qubit. The obstruction already appears for the singlet-projection measurement $C$54, whose Born probability on a product input $C$55 is

$C$56

The no-go theorem shows that exact generic simulation of the identity qubit channel requires infinite one-shot communication complexity $C$57. By contrast, if joint effects are restricted to product or separable measurements, finite-bit simulation is possible. Noisy depolarizing channels $C$58 are also generically simulable with finite communication, and the exact strong-simulation cost satisfies

$C$59

as $C$60, diverging only in the perfect-channel limit (Naik et al., 27 Jan 2025).

Quantum instruments generalize measurements by specifying both the classical outcome and the post-measurement state. Projective simulation of instruments asks whether an instrument $C$61 can be realized by sampling a classical label $C$62, performing a projective measurement $C$63, and then applying a trace-preserving channel $C$64 to the post-measurement state: $C$65 In the Choi representation, simulability is linked to Schmidt-rank constraints. The resulting criterion gives a computationally efficient necessary condition for generic instruments and, for qubits, a complete characterization: qubit-instrument simulability is equivalent to the existence of a decomposition into rank-vector sectors $C$66 and $C$67, with the $C$68 sector satisfying Schmidt number at most $C$69, which for qubits is equivalent to PPT and hence SDP-testable (Khandelwal et al., 2 Mar 2025).

The operational consequences are concrete. Standard qubit unsharp Lüders instruments are non-projective except for $C$70, even though the induced binary qubit POVM is always projective-simulable. Closed-form critical visibilities are obtained for dephasing noise and worst-case noise, while $C$71 is characterized as the largest real root of an explicit octic polynomial in $C$72. Projective-simulable qubit instruments obey the linear information-disturbance bound

$C$73

whereas the optimal Lüders instrument achieves the strictly better curve

$C$74

In sequential CHSH scenarios, projective-simulable instruments yield a Pareto frontier for $C$75, with numerically accessible violations up to approximately $C$76 on one leg (Khandelwal et al., 2 Mar 2025).

Taken together, these results show that weak simulation becomes substantially more restrictive once one includes post-measurement states, unknown ancillas, or entangled measurement effects. This suggests that “sampling the right distribution” is not a unique notion but a hierarchy whose strength depends on which operational interfaces of the quantum process must be preserved.

6. Resource measures beyond runtime: non-locality, memory, and energy

Weak simulation by quasiprobability sampling can also be phrased at the level of channels. For a non-local channel $C$77, the channel robustness of non-locality is defined as

$C$78

where $C$79 is the set of local channels built from single-qubit unitaries, projective measurements, and their tensor products. A one-shot simulator applies $C$80 with probability $C$81 and records the phase of $C$82, so that $C$83. The robustness is submultiplicative under composition. For arbitrary two-qubit unitaries, an explicit Cartan-based decomposition gives a fully general analytic upper bound, and numerical exploration of the Cartan tetrahedron yields worst-case overhead around $C$84, far below the naive concatenated estimate $C$85. Estimating an observable to precision $C$86 with confidence $C$87 requires

$C$88

so the sampling overhead is quadratic in $C$89 (Mitarai et al., 2020).

Classical simulation cost can also be cast as memory. In the Peres–Mermin contextuality scenario, a classical simulator is modeled as a finite-state Mealy machine whose internal state stores the memory required to reproduce sequential measurement statistics. For the six Peres–Mermin contexts alone, no automaton with $C$90 or $C$91 states can obey all quantum-certain constraints, while a $C$92 automaton exists, giving memory cost $C$93 bits. Adding compatibility constraints raises the optimum to $C$94, i.e. $C$95 bits. For the 15-context extended Peres–Mermin square, one needs $C$96, so the memory cost exceeds $C$97 bits, surpassing the Holevo bound for two qubits (Kleinmann et al., 2010).

In continuous-variable and hybrid models, weak simulation induces a trade-off between physical size and energy. An $C$98-qubit circuit of polynomial size can be approximately weakly simulated by a unitary circuit in a hybrid qubit-oscillator architecture that encodes logical qubits into $C$99 bosonic modes using approximate GKP codewords. The simulator outputs a distribution $\alpha_x$00 satisfying $\alpha_x$01, and the asymptotic mode-energy regimes are

$\alpha_x$02

With $\alpha_x$03, polynomial energy suffices; with sublinear $\alpha_x$04, subexponential energy is required; with constant $\alpha_x$05, the energy becomes exponential. The analysis is based on truncated-comb approximate GKP encodings, preparation error $\alpha_x$06, logical-gate error $\alpha_x$07, and explicit energy bounds for the preparation and compiled simulation circuits (Brenner et al., 23 Sep 2025).

These examples show that weak simulation of quantum resources is not characterized by a single universal cost parameter. Depending on the operational model, the critical resource may be quasiprobability negativity, bond dimension, stabilizer extent, communication, memory, Schmidt number, or bosonic energy. This suggests that the subject is best understood as a collection of closely related sampling problems whose common feature is operational fidelity to quantum output statistics, but whose limiting classical resource depends on which quantum capability is being emulated.