Dequantized Particle Algorithm

• Dequantized Particle Algorithms are classical methods that substitute quantum evolution with deterministic representations using configurations, trajectories, and tensor network states.
• They employ cluster expansion, Chebyshev filtering, and tensor-network reformulations to approximate outcomes in ground state energy estimation and guided local Hamiltonian problems.
• Their applications bridge complexity theory, many-body approximation, and kinetic plasma theory, providing insights into the quantum-classical boundary and simulation efficiency.

Searching arXiv for the specified papers and closely related dequantization work on guided local Hamiltonians and dequantized particle methods. The expression dequantized particle algorithm has been used in arXiv literature for several closely related classical surrogates of quantum procedures. In one usage, it denotes a classical algorithm for the Guided Local Hamiltonian (GLH) problem that replaces quantum imaginary-time filtering by combinatorial evaluation of partition functions over product-state components and connected interaction clusters (Zhang et al., 2024). In a second usage, it denotes a tensor-network reformulation of dequantized ground-state energy estimation in which Monte Carlo sampling over local-term “paths” is replaced by deterministic propagation of Chebyshev vectors represented as MPS or more general tensor networks (Manabe et al., 15 Dec 2025). In a third, explicit usage, Qin, May, and Molina use the name for a finite-dimensional dequantization of the many-body quantum theory underlying the Vlasov–Poisson system, yielding a structure-preserving discretization of the Schrödinger–Poisson equations in configuration space (Qin et al., 7 Jul 2025). An earlier swarm-based formulation of quantum dynamics by classical point samples, symplexes, and corteges is conceptually adjacent to this terminology, although it arises from a different physical program (Ozhigov, 2010).

1. Terminological scope and major formulations

Across these works, the common feature is not a single canonical algorithmic template but a recurring operation: a quantum evolution, filter, or many-body Hamiltonian is replaced by a classical representation whose elementary objects are configurations, trajectories, tensor-network states, or mode amplitudes. The phrase therefore identifies a family resemblance rather than a unique formalism.

Context	Classical objects	Quantum object being replaced
GLH / GSEE dequantization	Product configurations and connected interaction clusters	Randomized Quantum Imaginary-Time Evolution
Tensor-network GSEE dequantization	Chebyshev vectors as MPS / tensor networks	QSVT / QEVT polynomial filtering
Vlasov–Poisson dequantization	Fourier amplitudes $a_l$ as finite-dimensional canonical variables	Second-quantized bosonic Hamiltonian
Dynamic diffusion	Swarm samples, symplexes, corteges, bonds	Schrödinger evolution in Hilbert space

In the GLH setting, the term is partly interpretive: the algorithm is expressed as a deterministic cluster expansion, but it can be viewed as a particle-like classical simulation of imaginary-time evolution, where basis configurations play the role of particles and connected clusters encode their interaction histories (Zhang et al., 2024). In the tensor-network setting, the analogy is explicit but contrastive: the “overhead arising from sampling is replaced by the growth of the bond dimension required to represent Chebyshev vectors as tensor network states,” so a stochastic particle picture is traded for deterministic compression by entanglement structure (Manabe et al., 15 Dec 2025). In the Vlasov–Poisson setting, the phrase is literal and names the algorithm introduced in the paper title itself (Qin et al., 7 Jul 2025).

2. Guided local Hamiltonians and the imaginary-time filtering problem

The most technically developed use of the term in recent complexity-theoretic work arises from the guided Ground State Energy Estimation (GSEE) problem. The starting point is an $n$-qubit $k$-local Hamiltonian

$H=\sum_{X\in S}\lambda_X h_X,$

where each $h_X$ acts on at most $k=O(1)$ qubits, $\|h_X\|=1$, $|\lambda_X|\le 1$, and the interaction graph has maximum degree $\mathfrak d$. With eigen-decomposition

$$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$

guided GSEE asks for an estimate $n$0 satisfying $n$1, given a gap lower bound $n$2 and a guiding state $n$3 with overlap $n$4 (Zhang et al., 2024).

This is the GLH problem studied by Cade–Lucia and Gharibian et al. The problem is known to be BQP-complete: any BQP computation can be encoded into a clock Hamiltonian plus a guiding state, and near-optimal quantum algorithms solve GLH efficiently when such overlap is available (Zhang et al., 2024). The dequantized algorithm of interest targets a specific imaginary-time quantum algorithm called Randomized Quantum Imaginary-Time Evolution (RQITE).

RQITE is built from the partition function

$n$5

and from the residual

$n$6

The algorithm evaluates $n$7 on a grid in an interval $n$8 containing $n$9, and stops when $k$0 falls below a threshold $k$1, outputting that $k$2 as $k$3. A key condition is

$k$4

under which one chooses

$k$5

The imaginary-time operator is represented by the Fourier identity

$k$6

so the quantum algorithm reduces partition-function estimation to real-time evolutions and Hadamard tests (Zhang et al., 2024).

The same paper proves that the quantum RQITE scheme achieves additive error $k$7 with maximal evolution time $k$8 and total evolution time $k$9. It is not query-optimal, but it is designed to admit dequantization (Zhang et al., 2024).

3. Cluster expansion, semi-classical guiding states, and the particle interpretation

In the GLH dequantization, dequantization means constructing a classical algorithm that, given classical access to the guiding state and the local structure of $H=\sum_{X\in S}\lambda_X h_X,$0, simulates the quantum RQITE outcomes sufficiently well to solve GSEE under specified parameter regimes. The paper emphasizes that this is neither tensor networks nor naive brute-force simulation; it is a classical cluster expansion algorithm for expectations of $H=\sum_{X\in S}\lambda_X h_X,$1, combined with analytic continuation when necessary (Zhang et al., 2024).

The required guiding state is semi-classical: $H=\sum_{X\in S}\lambda_X h_X,$2 where $H=\sum_{X\in S}\lambda_X h_X,$3 and each $H=\sum_{X\in S}\lambda_X h_X,$4 is a product state. Then

$H=\sum_{X\in S}\lambda_X h_X,$5

with matrix-element partition functions

$H=\sum_{X\in S}\lambda_X h_X,$6

For sufficiently small $H=\sum_{X\in S}\lambda_X h_X,$7, the algorithm approximates $H=\sum_{X\in S}\lambda_X h_X,$8 by a cluster expansion over connected clusters of local interaction terms. The expansion is organized by connected clusters $H=\sum_{X\in S}\lambda_X h_X,$9, connected partitions $h_X$0, coupling monomials $h_X$1, and local matrix elements

$h_X$2

The central bound is

$h_X$3

Consequently, if

$h_X$4

the truncated cluster expansion has exponentially small tail, and $h_X$5, hence $h_X$6, can be approximated in polynomial time (Zhang et al., 2024).

The “particle” reading arises from how this expansion can be reinterpreted. Each product configuration $h_X$7 in $h_X$8 acts like a classical particle or walker with initial weight $h_X$9. Imaginary-time evolution propagates these configurations through local terms, and each connected cluster $k=O(1)$0 is an interaction history by which $k=O(1)$1 scatters into $k=O(1)$2. The full matrix element $k=O(1)$3 is then a sum over connected histories, analogous in structure to a sum over trajectories or branching histories. The authors explicitly note that they do not formulate the method as a Monte Carlo particle filter; they use deterministic enumeration rather than stochastic branching. A common misconception is therefore to equate the method with a sampling-based particle simulation. In the paper’s actual construction, the particle language is structural rather than literal (Zhang et al., 2024).

4. Accuracy regimes, analytic continuation, and the classical–quantum boundary

A central contribution of the GLH dequantization is the removal of two earlier restrictions in dequantized GLH algorithms: constant accuracy and the unrealistic uniform norm bound $k=O(1)$4. Earlier QSVT-based dequantizations became inefficient when $k=O(1)$5, because the polynomial degree scales as $k=O(1)$6. The new approach works directly with local terms $k=O(1)$7 and interaction degree $k=O(1)$8, so the cluster-expansion bounds depend on $k=O(1)$9 rather than directly on $\|h_X\|=1$0 (Zhang et al., 2024).

In the limited-accuracy regime, when $\|h_X\|=1$1, the classical algorithm solves GSEE in runtime

$\|h_X\|=1$2

The paper characterizes this as polynomial in system parameters $\|h_X\|=1$3, $\|h_X\|=1$4, and $\|h_X\|=1$5, and super-polynomial but still sub-exponential in $\|h_X\|=1$6. A normalization corollary states that if $\|h_X\|=1$7, then the same algorithm is efficient down to accuracy $\|h_X\|=1$8, so for $\|h_X\|=1$9, accuracy $|\lambda_X|\le 1$0 remains allowed (Zhang et al., 2024).

For arbitrary constant accuracy, real-$|\lambda_X|\le 1$1 cluster expansion ceases to converge, so the paper imposes a stronger overlap condition, $|\lambda_X|\le 1$2, and proves a zero-free region. If

$|\lambda_X|\le 1$3

then

$|\lambda_X|\le 1$4

is nonzero for all $|\lambda_X|\le 1$5 with $|\lambda_X|\le 1$6. This makes $|\lambda_X|\le 1$7 analytic in the right half-plane and permits analytic continuation using the protocol of Wild. The resulting runtime is not efficient in general; the paper states that it is doubly exponential in $|\lambda_X|\le 1$8, hence exponential in $|\lambda_X|\le 1$9 and super-exponential in $\mathfrak d$0. Even so, the result refines the placement of classical and quantum hardness: GLH with large overlap and constant accuracy is classically solvable with an explicit, although huge, complexity bound (Zhang et al., 2024).

The broader complexity implication is sharply stated. The work does not alter worst-case BQP-completeness. Rather, it indicates that quantum advantage for GLH must come from at least one of the following: a genuinely quantum guiding state rather than a semi-classical one, inverse-polynomial rather than constant accuracy, small overlaps, or Hamiltonian structure outside the reach of cluster expansion. Conversely, for sparse physical Hamiltonians with constant-degree interaction graphs, semi-classical guiding states, and constant accuracy, no exponential quantum speedup is expected (Zhang et al., 2024).

5. Tensor-network reformulation as a deterministic analogue

The tensor-network formulation of dequantized GSEE provides a second major interpretation of dequantized particle methods. It begins from the same QSVT/QEVT eigenvalue filtering framework but removes Monte Carlo sampling entirely. Instead of sampling products of local terms $\mathfrak d$1, it represents the Chebyshev vectors

$\mathfrak d$2

as tensor-network states and computes

$\mathfrak d$3

deterministically by contraction (Manabe et al., 15 Dec 2025).

The Chebyshev recursion is

$\mathfrak d$4

Moments satisfy

$\mathfrak d$5

so all moments up to order $\mathfrak d$6 are generated from Chebyshev vectors up to degree $\mathfrak d$7. The choice of Chebyshev basis rather than monomials is crucial because $\mathfrak d$8 for $\mathfrak d$9, and the coefficients remain $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$0, unlike monomial coefficients, which can grow like $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$1 (Manabe et al., 15 Dec 2025).

This paper explicitly frames tensor-network dequantization as the deterministic, compressed-representation counterpart of dequantized particle approaches. In sampling-based dequantization, the expectation value is approximated by Monte Carlo over an exponentially large path space; in tensor-network dequantization, the same expectation is evaluated by representing the entire propagated state and allowing complexity to be controlled by bond-dimension growth. Under exact contraction assumptions, the paper proves a classical runtime

$$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$2

where $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$3 is the bond-dimension growth factor for one application of $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$4. For MPS guiding states and Pauli Hamiltonians with $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$5 Pauli strings, the cost becomes $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$6 (Manabe et al., 15 Dec 2025).

The approximate algorithm truncates MPS bond dimension to $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$7, introducing local truncation error $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$8, cosine error $$H\ket{\psi_j}=E_j\ket{\psi_j},\qquad E_0<E_1\le \cdots\le E_{N-1},$$9, and global vector error $n$00. If the cumulative error satisfies

$n$01

then the filter expectation is accurate enough to run the GSEE binary search. If bounded-bond approximants exist with bond dimension $n$02, the cost becomes $n$03, which is linear rather than exponential in $n$04 (Manabe et al., 15 Dec 2025).

Numerically, the tensor-network method constructs high-degree polynomials up to $n$05 for Hamiltonians with up to $n$06 qubits. For the 1D transverse-field Ising model, up to $n$07 sites, the method yields accurate energy estimates and shows that for this setting ground-state energy estimation can be classically dequantized via MPS Chebyshev filtering. For the 2D transverse-field Ising model, up to $n$08, the required entanglement quickly defeats the MPS representation, and the paper interprets this breakdown as evidence of a crossover into a genuinely quantum-advantaged regime (Manabe et al., 15 Dec 2025).

6. Configuration-space particle dequantization for kinetic equations and earlier swarm formulations

In kinetic plasma theory, Qin, May, and Molina introduce the “Dequantized particle algorithm” as a specific method for the nonlinear Vlasov–Poisson system. They begin from the $n$09-body bosonic Hamiltonian

$n$10

pass to a second-quantized plane-wave representation, truncate to a finite mode set $n$11, and then dequantize by replacing $n$12 with complex c-numbers $n$13 and commutators with Poisson brackets (Qin et al., 7 Jul 2025).

The resulting finite-dimensional Hamiltonian is

$n$14

with

$n$15

and equations of motion

$n$16

This system preserves total energy, the particle number $n$17, and the total momentum $n$18. It is shown to be a structure-preserving discretization of the Schrödinger–Poisson system, and through the Wigner or Husimi transforms it provides an approximation of the Vlasov–Poisson system when quantum effects are negligible. Unlike conventional structure-preserving algorithms formulated in 6D phase space, it operates in 3D configuration space (Qin et al., 7 Jul 2025).

The paper’s numerical example is the nonlinear two-stream instability. In normalized units with

$n$19

the simulation uses parameters

$n$20

a mode set $n$21, and therefore $n$22 dequantized particles. The measured linear growth rate is $n$23, compared with the theoretical value $n$24, while energy, momentum, and particle number are conserved to high precision (Qin et al., 7 Jul 2025).

A more speculative but historically relevant precursor is the dynamic diffusion model of quantum behavior. There, the quantum state is represented by a swarm of classical point samples; many-particle states are represented by corteges, and samples aggregate into symplexes connected by rigid bonds. The swarm density approximates $n$25, while bond creation and annihilation generate a two-fraction dynamics of slow clustered “liquid” samples and fast “gas” samples. The paper presents this as an approximation to unitary quantum dynamics that avoids differentiating the density, but it also stresses fixed spatial grain $n$26, strong $n$27-dependence of effective coefficients, and an intrinsic decoherence mechanism caused by limited sample number (Ozhigov, 2010). This suggests a broader historical meaning of “dequantized particle algorithm”: a class of schemes in which quantum evolution is approximated by classical-like populations endowed with interaction rules, at the cost of structural assumptions, coarse graining, or both.

Taken together, these usages identify a technically coherent research direction. In complexity theory, dequantized particle algorithms delineate when semi-classical guiding states and bounded local structure suffice for classical simulation of GLH. In tensor-network theory, they become deterministic spectral filters whose cost is governed by entanglement rather than variance. In kinetic theory, they provide a structure-preserving route from second-quantized many-body dynamics to configuration-space approximations of Vlasov–Poisson. A plausible implication is that the term will continue to function as a bridge concept between quantum-inspired classical algorithms, many-body approximation theory, and the practical verification of quantum advantage.