betaFeedback

WorksPhysics and Astronomy

Rotation Collision based Quantum Lattice Boltzmann Methods

Abstract

In the existing quantum algorithms for the lattice Boltzmann method, resolving the unitarity problem of the BGK collision operator from first principles remains a fundamental challenge, and a systematically constructed unitary alternative that reproduces the BGK relaxation structure to leading order has not yet been established in the literature. In this work, we propose a novel quantum lattice Boltzmann method (QLBM) based on the rotation collision operator, in which the non-unitary BGK relaxation is replaced by a unitary rotation in the amplitude space of the distribution function, and the corresponding gate-level quantum circuits are also constructed. Specifically, the rotation collision operator is developed for both the D2Q5 model of the convection-diffusion equation and the D2Q9 model of the incompressible Navier–Stokes equations. It is worth noting that, for the D2Q5 model, the equilibrium-state preparation circuit can be precompiled and reused, and the resulting unitary blocks can be sequentially composed once the state-dependent collision parameters are specified. Finally, some numerical experiments are performed to validate the developed QLBM, demonstrating its accuracy and effectiveness.

Introduction

Over the past three decades, the lattice Boltzmann method (LBM) has emerged as a competitive numerical tool in computational fluid dynamics, and has been widely applied to a broad range of fluid flow and heat/mass transfer problems [4, 22]. As a mesoscopic approach, the LBM not only can recover the macroscopic governing equations in the hydrodynamic limit, but also retains a collision–streaming structure that is well suited for parallel computing [11].

Recent advances in the theory and hardware of quantum computing [1, 17] have motivated researchers to explore quantum algorithms for scientific computing, and computational fluid dynamics (CFD) has emerged as one of the most promising application areas [2, 10]. In particular, several quantum CFD methods have been proposed for the simulation of fluid flows governed by the Euler or Navier–Stokes equations [8], exploiting the potential quantum speedup in solving large-scale linear systems [10]. Among these approaches, the LBM provides a natural bridge to quantum

Funding: This work was financially supported by the National Natural Science Foundation of China (Grant Nos. 12501599 and 123B2018), the Postdoctoral Fellowship Program of CPSF (Grant No. GZB20250714), China Postdoctoral Science Foundation (Grant No. 2025M773077), the Open Research Fund of State Key Laboratory of Mesoscience and Process Engineering (Grant No. MESO-25-D04) and the Interdisciplinary Research Program of Hust (Grant No. 2024JCYJ001). computing, because its collision-streaming structure maps to quantum circuits naturally [14, 31]. For this reason, several quantum lattice Boltzmann methods (QLBMs) have been developed for the convection-diffusion equation and the Navier–Stokes equations [13, 25]. However, in such implementations, there are two fundamental problems that need to be addressed. The first one is the unitarity issue, i.e., the quantum evolution is required to be unitary, whereas the BGK collision operator in LBM is dissipative and non-unitary. The second is the nonlinearity issue, i.e., the equilibrium distribution function in the D2Q9 model contains some quadratic terms of the fluid velocity, which cannot be directly encoded in quantum circuits without violating the no-cloning theorem [28]. To overcome these two problems, several strategies have been developed in some previous works. For instance, Mezzacapo et al. [14] first decomposed the non-unitary collision operator into a linear combination of unitary operators and then implemented it probabilistically with the help of an ancilla qubit, while the success probability of this approach decreases exponentially with the system size. Unlike this method, Todorova and Steijl [23] avoided the collision term by adopting a collisionless model, which, however, fundamentally restricts the physical scope of the method. Later, Zeng [32, 33] proposed a linearized collision operator within the QLBM framework, but the linearization inevitably introduces approximation error that deviates from the original physical model. Xiao et al. [29, 30] further extended QLBMs to incompressible flows on nonuniform meshes and in domains with curved boundaries, and improved numerical stability at high Reynolds numbers via fractional-step formulations. Kumar and Frankel [12] represented the collision operator as a unitary matrix via singular value decomposition, achieving an exact unitary encoding. However, the resulting matrix depends on the current quantum state and must be recomputed at each time step, with gate complexity growing exponentially in the number of qubits. In addition, Wang et al. [24] proposed a novel ensemble method based on the lattice gas automaton, which provides an alternative perspective for the quantum implementation of the LBM. Moreover, it is worth mentioning that such a linearized operator no longer preserves the genuine BGK structure at the level of quantum encoding. Nevertheless, to the best of our knowledge, none of these approaches that can be used to solve the unitarity problem of the collision operator from the first principles.

In this work, we develop a quantum lattice Boltzmann framework based on a rotation collision operator used to address the unitarity issue of the BGK relaxation. The proposed framework provides a unitary formulation of the collision process in the amplitude space and corresponding quantum circuit implementations for both the D2Q5 model of the convection–diffusion equation and the D2Q9 model of the incompressible Navier–Stokes equations.

The rest of the paper is organized as follows. In Section 2, the rotation collision operator is introduced, followed by its quantum circuit implementations for the D2Q5 and D2Q9 models. Numerical results are presented in Section 4, and some conclusions are summarized in Section 5. Finally, the gate-level circuit constructions are provided in the Appendices.

Rotation Collision Operator

Lattice Boltzmann Method

In the lattice Boltzmann method with the BGK collision operator, also called lattice BGK method, the distribution function {fi}i=0Q−1\{f_i\}_{i=0}^{Q-1} at a single lattice site is updated according to the lattice Boltzmann equation (LBE) [4, 18],

fi(x+eiδt,t+δt)=fi(x,t)−1τ(fi(x,t)−fieq(x,t))(1)f_i(\boldsymbol{x} + \boldsymbol{e}_i\delta t, t + \delta t) = f_i(\boldsymbol{x},t) - \frac{1}{\tau}\left(f_i(\boldsymbol{x},t) - f_i^{eq}(\boldsymbol{x},t)\right) \tag*{(1)}

where QQ is the number of discrete velocity directions, eie_i is the ii-th discrete velocity, δt\delta t is the time step, τ\tau is the relaxation time, and fieqf_i^{eq} is the equilibrium distribution function. The macroscopic density and velocity are determined by the moments of the distribution function,

ρ=∑ifi,ρu=∑ifiei.(2)\rho= \sum_i f_i,\qquad\rho\boldsymbol{u} = \sum_i f_i\boldsymbol{e}_i. \tag*{(2)}

The equilibrium distribution function satisfies the following conservation conditions [11],

∑ifieq=ρ,∑ifieqei=ρu.(3)\sum_i f_i^{eq} = \rho,\qquad\sum_i f_i^{eq}\boldsymbol{e}_i = \rho\boldsymbol{u}. \tag*{(3)}

Based on Eq. (1), the collision step in Eq. (1) can be written as

fi∗=fi−1τ(fi−fieq)=(1−1τ)fi+1τfieq,(4)f_i^* = f_i - \frac{1}{\tau}\left(f_i - f_i^{eq}\right) = \left(1-\frac{1}{\tau}\right)f_i + \frac{1}{\tau}f_i^{eq}, \tag*{(4)}

which is a linear interpolation between fif_i and fieqf_i^{eq}. Thus, the deviation from equilibrium state is reduced by a factor of (1−1/τ)(1-1/\tau) at each collision step,

fi∗−fieq=(1−1τ)(fi−fieq).(5)f_i^* - f_i^{eq} = \left(1-\frac{1}{\tau}\right)\left(f_i - f_i^{eq}\right). \tag*{(5)}

Non-Unitarity of the BGK Collision Operator

In quantum computing, each admissible operation must be implemented as a unitary operator [15]. The BGK collision operator, however, presents two fundamental obstacles that prevent its direct implementation as a quantum gate.

The first one is the nonlinearity in amplitude space. The natural encoding of the distribution function into a quantum state is given by the amplitudes ψi=fi\psi_i = \sqrt{f_i} [14], that is, ∣ψ⟩=∑ifi∣i⟩|\psi\rangle= \sum_i \sqrt{f_i}|i\rangle. Under this encoding, the BGK collision operator takes the form

fi∗=(1−1τ)fi+1τfieq,(6)\sqrt{f_i^*} = \sqrt{\left(1-\frac{1}{\tau}\right)f_i+\frac{1}{\tau}f_i^{eq}}, \tag*{(6)}

and due to the fact a+b≠a+b\sqrt{a+b} \ne\sqrt{a}+\sqrt{b} in general, this is not equal to (1−1/τ)fi+(1/τ)fieq\left(1-1/\tau\right)\sqrt{f_i}+(1/\tau)\sqrt{f_i^{eq}}. This means that the BGK collision operator is nonlinear in amplitude space and cannot be represented as a linear operator on quantum state.

The second obstacle is the non-unitarity of linear interpolation. Even we consider the corresponding linear interpolation directly in amplitude space, such an operation is not unitary. Specifically, for any ∣ψ⟩≠∣ψeq⟩|\psi\rangle\ne|\psi_{eq}\rangle, we have

∥(1−1τ)∣ψ⟩+1τ∣ψeq⟩∥≠∥∣ψ⟩∥,(7)\left\|\left(1-\frac{1}{\tau}\right)|\psi\rangle+\frac{1}{\tau}|\psi_{eq}\rangle\right\| \ne\||\psi\rangle\|, \tag*{(7)}

which indicates that the linear interpolation does not preserve the norm of the state vector. This is in conflict with the requirement that any quantum operation must satisfy ∥U∣ψ⟩∥=∥∣ψ⟩∥\|U|\psi\rangle\|=\||\psi\rangle\|. From the above analysis, it can be seen that the BGK collision operator cannot be directly realized as a quantum gate, and a unitary alternative is therefore required.

Amplitude Encoding

Given a single-site distribution function {fi}i=0Q−1\{f_i\}_{i=0}^{Q-1}, the corresponding normalized quantum state [19] is defined by

∣ψ⟩=1ρ∑i=0Q−1fi∣i⟩,⟨ψ∣ψ⟩=1,(8)|\psi\rangle=\frac{1}{\sqrt{\rho}}\sum_{i=0}^{Q-1}\sqrt{f_i}|i\rangle,\qquad\langle\psi|\psi\rangle=1, \tag*{(8)}

from which the distribution function can be recovered by fi=ρ∣⟨i∣ψ⟩∣2f_i=\rho|\langle i|\psi\rangle|^2. Similarly, the equilibrium distribution function is encoded as

∣ψeq⟩=1ρ∑i=0Q−1fieq∣i⟩,(9)|\psi_{eq}\rangle=\frac{1}{\sqrt{\rho}}\sum_{i=0}^{Q-1}\sqrt{f_i^{eq}}|i\rangle, \tag*{(9)}

and it can be verified that ∣ψeq⟩|\psi_{eq}\rangle is also a normalized state by the conservation condition Eq. (3).

Construction of the Rotation Collision Operator

Since the amplitudes fi\sqrt{f_i} and fieq\sqrt{f_i^{eq}} are real and non-negative, the inner product ⟨ψ∣ψeq⟩\langle\psi|\psi_{eq}\rangle is a real number in [0,1][0,1], and ∣ψ⟩|\psi\rangle and ∣ψeq⟩|\psi_{eq}\rangle are both unit vectors in the Hilbert space. We define the angle between them as

ϕ=arccos⁡⟨ψ∣ψeq⟩∈[0,π/2].(10)\phi=\arccos\langle\psi|\psi_{eq}\rangle\in[0,\pi/2]. \tag*{(10)}

Based on Eq. (5), the BGK collision reduces the distance between ff and feqf^{eq} by a factor of (1−1/τ)(1-1/\tau) at each step. In the Hilbert space representation, this deviation is characterized by the angle ϕ\phi between ∣ψ⟩|\psi\rangle and ∣ψeq⟩|\psi_{eq}\rangle, the unitary analogue of the BGK collision is therefore expected to contract the angle ϕ\phi by the same factor corresponding to a rotation of ∣ψ⟩|\psi\rangle toward ∣ψeq⟩|\psi_{eq}\rangle such that the angle decreases from ϕ\phi to

ϕ(1−1τ)=ϕ−θ,θ=ϕτ.(11)\phi\left(1-\frac{1}{\tau}\right)=\phi-\theta,\qquad\theta=\frac{\phi}{\tau}. \tag*{(11)}

If ϕ=0\phi=0, the state ∣ψ⟩|\psi\rangle already coincides with ∣ψeq⟩|\psi_{eq}\rangle, θ=0\theta=0, and Ucol=IU_{\mathrm{col}}=I trivially. In what follows assume ϕ≠0\phi\ne0, and accordingly define the unit vector in the plane spanned by ∣ψ⟩|\psi\rangle and ∣ψeq⟩|\psi_{eq}\rangle that is orthogonal to ∣ψ⟩|\psi\rangle and in the direction of ∣ψeq⟩|\psi_{eq}\rangle as

∣ψ⊥⟩=∣ψeq⟩−cos⁡ϕ∣ψ⟩sin⁡ϕ.(12)|\psi_\perp\rangle=\frac{|\psi_{eq}\rangle-\cos\phi|\psi\rangle}{\sin\phi}. \tag*{(12)}

Substituting Eq. (10) gives ⟨ψ∣ψ⊥⟩=(⟨ψ∣ψeq⟩−cos⁡ϕ)/sin⁡ϕ=0\langle\psi|\psi_\perp\rangle=(\langle\psi|\psi_{eq}\rangle-\cos\phi)/\sin\phi=0 and ⟨ψ⊥∣ψ⊥⟩=1\langle\psi_\perp|\psi_\perp\rangle=1, thus {∣ψ⟩,∣ψ⊥⟩}\{|\psi\rangle,|\psi_\perp\rangle\} forms an orthonormal basis for the two-dimensional subspace V=span⁡{∣ψ⟩,∣ψeq⟩}\mathcal{V}=\operatorname{span}\{|\psi\rangle,|\psi_{eq}\rangle\}. The rotation collision operator is then defined as the rotation by angle θ\theta within V\mathcal{V} and the identity on its orthogonal complement,

Ucol=I+(cos⁡θ−1)(∣ψ⟩⟨ψ∣+∣ψ⊥⟩⟨ψ⊥∣)+sin⁡θ(∣ψ⊥⟩⟨ψ∣−∣ψ⟩⟨ψ⊥∣).(13)U_{\mathrm{col}}=I+(\cos\theta-1)(|\psi\rangle\langle\psi|+|\psi_\perp\rangle\langle\psi_\perp|)+\sin\theta(|\psi_\perp\rangle\langle\psi|-|\psi\rangle\langle\psi_\perp|). \tag*{(13)}

and the post-collision quantum state is given by

∣ψ′⟩=Ucol∣ψ⟩=cos⁡θ∣ψ⟩+sin⁡θ∣ψ⊥⟩,(14)|\psi'\rangle= U_{\mathrm{col}}|\psi\rangle= \cos\theta|\psi\rangle+ \sin\theta|\psi_{\perp}\rangle, \tag*{(14)}

from which the post-collision distribution function is recovered by fi∗=ρ∣⟨i∣ψ′⟩∣2f_i^* = \rho|\langle i|\psi'\rangle|^2.

It is clear that Eq. (13) defines UcolU_{\mathrm{col}} in terms of the explicit vector ∣ψ⊥⟩|\psi_{\perp}\rangle, which does not need to be prepared as a separate quantum state at the circuit level: an equivalent realization of UcolU_{\mathrm{col}} can be assembled from the state-preparation circuits of ∣ψ⟩|\psi\rangle and ∣ψeq⟩|\psi_{\mathrm{eq}}\rangle alone, using a sequence of generalized reflections; the details of this circuit construction are given in Appendix B.

The rotation collision operator UcolU_{\mathrm{col}} defined in Eq. (13) is unitary, i.e., Ucol†Ucol=IU_{\mathrm{col}}^\dagger U_{\mathrm{col}} = I. This follows from the fact that the restriction of UcolU_{\mathrm{col}} to the subspace VV is the rotation matrix R=(cos⁡θ−sin⁡θsin⁡θcos⁡θ)R = \begin{pmatrix} \cos\theta& -\sin\theta\\ \sin\theta& \cos\theta\end{pmatrix}, which satisfies RTR=IR^{\mathsf{T}}R = I, while UcolU_{\mathrm{col}} acts as the identity on the orthogonal complement of VV.

As a direct consequence of the unitarity of UcolU_{\mathrm{col}}, the rotation collision operator conserves mass exactly. Indeed, one can obtain ∑ifi∗=ρ∑i∣⟨i∣ψ′⟩∣2=ρ\sum_i f_i^* = \rho\sum_i |\langle i|\psi'\rangle|^2 = \rho since ∥∣ψ′⟩∥=∥∣ψ⟩∥=1\lVert|\psi'\rangle\rVert= \lVert|\psi\rangle\rVert= 1.

In addition, we can show that the inner product of the post-collision state ∣ψ′⟩|\psi'\rangle with ∣ψeq⟩|\psi_{\mathrm{eq}}\rangle also satisfies the following relation,

⟨ψ′∣ψeq⟩=cos⁡θcos⁡ϕ+sin⁡θsin⁡ϕ=cos⁡(ϕ−θ),(15)\langle\psi'|\psi_{\mathrm{eq}}\rangle= \cos\theta\cos\phi+ \sin\theta\sin\phi= \cos(\phi-\theta), \tag*{(15)}

so the post-collision angle is ϕ−θ=ϕ(1−1/τ)\phi-\theta= \phi(1-1/\tau), which agrees exactly with the relaxation rate (1−1/τ)(1-1/\tau) of the BGK collision operator in Eq. (5).

The rotation collision operator UcolU_{\mathrm{col}} is an exact unitary transformation in the Hilbert space. The difference from the classical BGK collision arises when the amplitude-space evolution is mapped back to the distribution-function space through

firot=ρ∣⟨i∣ψ′⟩∣2.(16)f_i^{\mathrm{rot}} = \rho|\langle i|\psi'\rangle|^2. \tag*{(16)}

We show below that, in the continuum limit δt→0\delta t \to0, the rotation collision agrees with the BGK collision up to first order in δt\delta t.

Consider the evolution equation Eq. (1), and introduce Di=∂t+ei⋅∇D_i = \partial_t + \mathbf{e}_i \cdot\nabla, we can obtain the following equation with the help of Taylor expansion

δtDifi+δt22Di2fi+O(δt3)=−1τ(fi−fieq).(17)\delta t D_i f_i + \frac{\delta t^2}{2}D_i^2 f_i + O(\delta t^3) = -\frac{1}{\tau}(f_i-f_i^{\mathrm{eq}}). \tag*{(17)}

From above equation one can derive [3]

fi−fieq=O(δt).(18)f_i-f_i^{\mathrm{eq}} = O(\delta t). \tag*{(18)}

Under the condition of non-negative equilibrium/distribution functions, and let ai=fi/ρa_i = \sqrt{f_i/\rho} and aieq=fieq/ρa_i^{\mathrm{eq}} = \sqrt{f_i^{\mathrm{eq}}/\rho} denote the normalized amplitudes of the distribution and equilibrium states, respectively, the following equation can be obtained from Eq. (18)

ai−aieq=O(δt).(19)a_i-a_i^{\mathrm{eq}} = O(\delta t). \tag*{(19)}

Since the two quantum states are normalized,

cos⁡ϕ=⟨ψ∣ψeq⟩=1−12∑i(ai−aieq)2,(20)\cos\phi= \langle\psi|\psi_{\mathrm{eq}}\rangle= 1-\frac{1}{2}\sum_i(a_i-a_i^{\mathrm{eq}})^2, \tag*{(20)}

we have

ϕ=O(δt).(21)\phi= O(\delta t). \tag*{(21)}

Substituting ∣ψ⊥⟩|\psi_{\perp}\rangle into the post-collision state yields

∣ψ′⟩=(cos⁡θ−sin⁡θcos⁡ϕsin⁡ϕ)∣ψ⟩+sin⁡θsin⁡ϕ∣ψeq⟩.(22)|\psi'\rangle= \left(\cos\theta-\frac{\sin\theta\cos\phi}{\sin\phi}\right)|\psi\rangle+\frac{\sin\theta}{\sin\phi}|\psi_{\mathrm{eq}}\rangle. \tag*{(22)}

Consider θ=ϕ/τ\theta= \phi/\tau and ϕ=O(δt)\phi= O(\delta t), we have

sin⁡θsin⁡ϕ=1τ+O(δt2),cos⁡θ−sin⁡θcos⁡ϕsin⁡ϕ=1−1τ+O(δt2),(23)\frac{\sin\theta}{\sin\phi} = \frac{1}{\tau}+O(\delta t^2), \qquad\cos\theta-\frac{\sin\theta\cos\phi}{\sin\phi}=1-\frac{1}{\tau}+O(\delta t^2), \tag*{(23)}

which yields

∣ψ′⟩=(1−1τ)∣ψ⟩+1τ∣ψeq⟩+O(δt2).(24)|\psi'\rangle= \left(1-\frac{1}{\tau}\right)|\psi\rangle+\frac{1}{\tau}|\psi_{\mathrm{eq}}\rangle+O(\delta t^2). \tag*{(24)}

Taking the iith component and recovering the distribution function, together with

fifieq=fi+fieq2+O(δt2),(25)\sqrt{f_i f_i^{\mathrm{eq}}} = \frac{f_i + f_i^{\mathrm{eq}}}{2} + O(\delta t^2), \tag*{(25)}

gives

firot=fi+fieq−fiτ+O(δt2).(26)f_i^{\mathrm{rot}} = f_i + \frac{f_i^{\mathrm{eq}} - f_i}{\tau} + O(\delta t^2). \tag*{(26)}

On the other hand, the post-collision distribution function is given by

fiBGK=fi+fieq−fiτ,(27)f_i^{\mathrm{BGK}} = f_i + \frac{f_i^{\mathrm{eq}} - f_i}{\tau}, \tag*{(27)}

thus we have

firot−fiBGK=O(δt2).(28)f_i^{\mathrm{rot}} - f_i^{\mathrm{BGK}} = O(\delta t^2). \tag*{(28)}

From above discussion, one can find that compared to the BGK collision, the rotation collision only introduces a discrepancy with a second order of time step, which is consistent with truncation error of the standard LBM.

In particular, if we consider the case of τ=1\tau= 1, one can obtain θ=ϕ\theta= \phi and ∣ψ′⟩=∣ψeq⟩|\psi'\rangle= |\psi_{\mathrm{eq}}\rangle exactly. Consequently,

firot=fieq=fiBGK(29)f_i^{\mathrm{rot}} = f_i^{\mathrm{eq}} = f_i^{\mathrm{BGK}} \tag*{(29)}

Quantum Circuit Implementation

For the D2Q9 model, the equilibrium state ∣ψeq⟩|\psi_{\mathrm{eq}}\rangle depends on the density and velocity moments of the current state ∣ψ⟩|\psi\rangle. As a result, the collision operator UcolU_{\mathrm{col}} is a 16×1616 \times16 state-dependent unitary that must be updated at each time step according to the macroscopic moments and applied to the four-qubit register encoding ∣ψ⟩|\psi\rangle.

In contrast, the equilibrium state ∣ψeq⟩|\psi_{\mathrm{eq}}\rangle in the D2Q5 model reduces to a fixed vector independent of the current state (see § 3.4). Consequently, UcolU_{\mathrm{col}} becomes a fixed unitary acting on the three-qubit register, and can be precomputed once at the circuit construction stage.

In this case, UcolU_{\mathrm{col}} admits a direct gate-level synthesis using the Shende–Bullock–Markov state-preparation scheme [21] together with multi-controlled phase gates [15], without requiring dynamic recompilation. The resulting gate-level circuits for both models are provided in Appendices B and C.

Quantum Lattice Boltzmann Method and Gate-Concatenability

Full Quantum LBM Evolution Structure

At each time step, the classical LBM updates the distribution function through two successive steps: a collision step followed by a streaming step. In the quantum framework, each step corresponds to a unitary operator acting on the joint state

∣Ψ⟩∈Hdir⊗Hgrid.(30)|\Psi\rangle\in\mathcal{H}_{\mathrm{dir}} \otimes\mathcal{H}_{\mathrm{grid}}. \tag*{(30)}

The Hilbert space can be decomposed as H=Hdir⊗Hgrid\mathcal{H} = \mathcal{H}_{\mathrm{dir}} \otimes\mathcal{H}_{\mathrm{grid}}, where Hdir\mathcal{H}_{\mathrm{dir}} encodes the QQ discrete velocity directions at each lattice site, as defined in Eq. (8), and Hgrid\mathcal{H}_{\mathrm{grid}} encodes the lattice coordinates. The single-step quantum evolution operator is then given by

Ustep=Ustream⋅Ucol,(31)U_{\mathrm{step}} = U_{\mathrm{stream}} \cdot U_{\mathrm{col}}, \tag*{(31)}

where UcolU_{\mathrm{col}} is the rotation collision operator acting on Hdir\mathcal{H}_{\mathrm{dir}} at each site, and UstreamU_{\mathrm{stream}} denotes the streaming operator defined below, which acts jointly on Hdir⊗Hgrid\mathcal{H}_{\mathrm{dir}} \otimes\mathcal{H}_{\mathrm{grid}}. The operator ordering in Eq. (31) reflects that UcolU_{\mathrm{col}} is applied prior to UstreamU_{\mathrm{stream}}, consistent with the classical collide-then-stream update. The evolution over TT steps is then given by

∣Ψ(T)⟩=UstepT∣Ψ(0)⟩.(32)|\Psi(T)\rangle= U_{\mathrm{step}}^T |\Psi(0)\rangle. \tag*{(32)}

Quantum Implementation of the Streaming Step

The streaming step shifts the distribution function at one lattice spacing along each discrete velocity direction by

fi(x+eiδt,t+δt)=fi∗(x,t).(33)f_i(\boldsymbol{x} + \boldsymbol{e}_i \delta t, t + \delta t) = f_i^*(\boldsymbol{x}, t). \tag*{(33)}

In the quantum framework, the streaming step corresponds to a permutation operator UstreamU_{\mathrm{stream}} acting jointly on the direction and lattice-coordinate registers. We describe two implementations that realize this permutation by different circuit constructions.

Scheme 1: Quantum walk [5, 16]. The streaming step is represented as a quantum walk with the discrete velocity direction serving as the coin degree of freedom. Actually, the streaming operator can be decomposed as

Ustream=S⋅(C⊗Igrid),(34)U_{\mathrm{stream}} = S \cdot\left(C \otimes I_{\mathrm{grid}}\right), \tag*{(34)}

where CC acts on Hdir\mathcal{H}_{\mathrm{dir}} as the coin operator, and SS is the conditional shift operator that translates the lattice register by one site along each velocity direction. Note that CC is distinct from the collision operator UcolU_{\mathrm{col}} in Eq. (31) and serves only to route the shift SS, not to relax the distribution function toward its equilibrium state. Under periodic boundary conditions, the conditional shift reduces to a modular addition, which can be implemented efficiently with quantum adder circuits.

Scheme 2: Quantum Fourier transform [6, 20]. Under periodic boundary conditions, the spatial shift operator is diagonal in the Fourier basis. The quantum Fourier transform F\mathcal{F} diagonalizes the shift TeiT_{\mathbf{e}_i} as

F⋅Tei⋅F†=diag⁡(e2πikei/N)k=0N−1,(35)\mathcal{F} \cdot T_{\mathbf{e}_i} \cdot\mathcal{F}^{\dagger} = \operatorname{diag}\left(e^{2\pi i k e_i/N}\right)_{k=0}^{N-1}, \tag*{(35)}

where i=−1i = \sqrt{-1} denotes the imaginary unit, written in upright type to distinguish it from the direction index ii used elsewhere. The streaming step is thus implemented in three stages: the QFT, diagonal phase gates, and inverse QFT. The QFT circuit requires O(n2)O(n^2) gates with n=log⁡2Nn = \log_2 N, compared with O(Nlog⁡N)=O(n2n)O(N \log N) = O(n 2^n) for the classical fast Fourier transform, giving rise to an exponential reduction in gate count.

Both schemes can realize the same permutation operator UstreamU_{\mathrm{stream}} through different circuit constructions, and both are applicable to the D2Q5 and D2Q9 models.

Sequential Concatenation of QLBM Steps

The implementation of QLBM consists of a collision operation followed by a streaming operation. Since the rotation collision operator is constructed from the instantaneous state, we introduce a time-step index and write the one-step evolution operator as

Ustep,t=UstreamUcol,t,(36)U_{\mathrm{step},t} = U_{\mathrm{stream}} U_{\mathrm{col},t}, \tag*{(36)}

where Ucol,tU_{\mathrm{col},t} is the rotation collision operator at time tt, UstreamU_{\mathrm{stream}} denotes the streaming operator.

Accordingly, the evolution over TT time steps can be represented by the ordered composition

∣Ψ(T)⟩=Ustep,T−1Ustep,T−2⋯Ustep,1Ustep,0∣Ψ(0)⟩,(37)|\Psi(T)\rangle= U_{\mathrm{step},T-1}U_{\mathrm{step},T-2} \cdots U_{\mathrm{step},1}U_{\mathrm{step},0}|\Psi(0)\rangle, \tag*{(37)}

or equivalently,

∣Ψ(T)⟩=[∏t=T−10UstreamUcol,t]∣Ψ(0)⟩.(38)|\Psi(T)\rangle= \left[ \prod_{t=T-1}^{0} U_{\mathrm{stream}}U_{\mathrm{col},t} \right]|\Psi(0)\rangle. \tag*{(38)}

We refer to this ordered composition of successive unitary collision–streaming blocks as sequential circuit concatenation. Unlike repeated application of a single time-independent operator, this construction does not require the collision operators at different time steps to be identical. It should be noted that when the complete one-step operator is independent of both the instantaneous state and the time step, the following special form can be recovered

∣Ψ(T)⟩=UstepT∣Ψ(0)⟩(39)|\Psi(T)\rangle= U_{\mathrm{step}}^{T}|\Psi(0)\rangle \tag*{(39)}

Once the parameters defining the individual collision operators are specified, the corresponding unitary blocks can be connected sequentially without requiring a projective measurement solely for the composition of adjacent quantum operations. The determination of state-dependent collision parameters from an intermediate quantum state is a separate issue from the unitary composition itself.

Linearity of the D2Q5 Equilibrium Distribution for convection-diffusion equation

In this work, the D2Q5 model is used to solve the convection–diffusion equation

∂tϕ+∇⋅(vϕ)=κ∇2ϕ,(40)\partial_t \phi+ \nabla\cdot(\mathbf{v}\phi) = \kappa\nabla^2 \phi, \tag*{(40)}

where v\mathbf{v} is an externally prescribed velocity field and κ=cs2(τ−1/2)δt\kappa= c_s^2(\tau- 1/2)\delta t is the diffusion coefficient. The equilibrium distribution function and macroscopic variable ϕ\phi are given by [27]

fieq=λiϕ,λi=wi(1+ei⋅vcs2),ϕ=∑jfj.(41)f_i^{\mathrm{eq}} = \lambda_i\phi,\qquad \lambda_i = w_i\left(1 + \frac{\mathbf{e}_i \cdot\mathbf{v}}{c_s^2}\right),\qquad \phi= \sum_j f_j. \tag*{(41)}

The coefficients λi\lambda_i depend only on the prescribed parameters {wi,ei,v}\{w_i,e_i,v\} and are independent of the instantaneous distribution {fj}\{f_j\}. Consequently, fieqf_i^{\mathrm{eq}} is a linear function of the distribution function. This structure leads to an important simplification in the quantum representation: the normalized equilibrium direction is independent of the instantaneous distribution.

Indeed, the normalized equilibrium state can be written as

∣e^eq⟩=∑iλi∣i⟩(∑iλi)1/2.(42)\lvert\hat{e}_{\mathrm{eq}}\rangle= \frac{\sum_i \sqrt{\lambda_i}\lvert i\rangle}{\left(\sum_i \lambda_i\right)^{1/2}}. \tag*{(42)}

From Eq. (41) we have fieq=λiϕ\sqrt{f_i^{\mathrm{eq}}}=\sqrt{\lambda_i}\sqrt{\phi}. Since the common factor ϕ\sqrt{\phi} cancels upon normalization, the amplitude-encoded equilibrium state satisfies

∣ψeq⟩=∣e^eq⟩,(43)\lvert\psi_{\mathrm{eq}}\rangle= \lvert\hat{e}_{\mathrm{eq}}\rangle, \tag*{(43)}

and is therefore independent of the instantaneous distribution.

Rotation Collision and Sequential Evolution for the D2Q5 Model

As shown in the § 3.4, the normalized equilibrium state of the D2Q5 model is the fixed direction ∣e^eq⟩\lvert\hat{e}_{\mathrm{eq}}\rangle, independent of the instantaneous distribution function. This property simplifies the implementation of the rotation collision, although the complete collision operator generally remains state dependent.

Let ∣ψt⟩\lvert\psi_t\rangle denote the amplitude-encoded distribution state at time tt. The angle between the instantaneous state and the fixed equilibrium direction is

ϕt=arccos⁡⟨ψt∣e^eq⟩.(44)\phi_t = \arccos\langle\psi_t \lvert\hat{e}_{\mathrm{eq}}\rangle. \tag*{(44)}

Following the rotation collision construction introduced in § 2, the corresponding rotation angle is

θt=ϕtτ.(45)\theta_t = \frac{\phi_t}{\tau}. \tag*{(45)}

If ϕt≠0\phi_t \ne0, the unit vector orthogonal to ∣ψt⟩\lvert\psi_t\rangle within the two-dimensional rotation subspace span⁡{∣ψt⟩,∣e^eq⟩}\operatorname{span}\{\lvert\psi_t\rangle,\lvert\hat{e}_{\mathrm{eq}}\rangle\} is given by

∣ψ⊥,t⟩=∣e^eq⟩−cos⁡ϕt∣ψt⟩sin⁡ϕt.(46)\lvert\psi_{\perp,t}\rangle= \frac{\lvert\hat{e}_{\mathrm{eq}}\rangle-\cos\phi_t\lvert\psi_t\rangle}{\sin\phi_t}. \tag*{(46)}

Accordingly, the rotation collision operator at time tt takes the form

Ucol,t=I+(cos⁡θt−1)(∣ψt⟩⟨ψt∣+∣ψ⊥,t⟩⟨ψ⊥,t∣)+sin⁡θt(∣ψ⊥,t⟩⟨ψt∣−∣ψt⟩⟨ψ⊥,t∣),(47)U_{\mathrm{col},t}=I+(\cos\theta_t-1)\left(\lvert\psi_t\rangle\langle\psi_t\rvert+\lvert\psi_{\perp,t}\rangle\langle\psi_{\perp,t}\rvert\right)+\sin\theta_t\left(\lvert\psi_{\perp,t}\rangle\langle\psi_t\rvert-\lvert\psi_t\rangle\langle\psi_{\perp,t}\rvert\right), \tag*{(47)}

and its action on the instantaneous state is

Ucol,t∣ψt⟩=cos⁡θt∣ψt⟩+sin⁡θt∣ψ⊥,t⟩.(48)U_{\mathrm{col},t}\lvert\psi_t\rangle=\cos\theta_t\lvert\psi_t\rangle+\sin\theta_t\lvert\psi_{\perp,t}\rangle. \tag*{(48)}

Therefore, the angular distance from the equilibrium direction changes from ϕt\phi_t to

ϕt−θt=(1−1τ)ϕt,(49)\phi_t-\theta_t=\left(1-\frac{1}{\tau}\right)\phi_t, \tag*{(49)}

which is consistent with the relaxation mechanism of the rotation collision introduced in § 2.

The complete evolution of D2Q5 at time tt consists of the rotation collision followed by the streaming operation,

Ustep,t=UstreamUcol,t.(50)U_{\mathrm{step},t}=U_{\mathrm{stream}}U_{\mathrm{col},t}. \tag*{(50)}

Following the sequential composition introduced in § 3.3, the evolution over TT time steps can be written as

∣Ψ(T)⟩=Ustep,T−1Ustep,T−2⋯Ustep,1Ustep,0∣Ψ(0)⟩,(51)\lvert\Psi(T)\rangle=U_{\mathrm{step},T-1}U_{\mathrm{step},T-2}\cdots U_{\mathrm{step},1}U_{\mathrm{step},0}\lvert\Psi(0)\rangle, \tag*{(51)}

or equivalently,

∣Ψ(T)⟩=[∏t=T−10UstreamUcol,t]∣Ψ(0)⟩.(52)\lvert\Psi(T)\rangle=\left[\prod_{t=T-1}^{0}U_{\mathrm{stream}}U_{\mathrm{col},t}\right]\lvert\Psi(0)\rangle. \tag*{(52)}

It should be emphasized that the fixed equilibrium direction does not imply that the complete rotation collision operator is time-independent. In general, the instantaneous state ∣ψt⟩\lvert\psi_t\rangle, the angle ϕt\phi_t, the rotation angle θt\theta_t, and the associated rotation subspace vary during the evolution. Thus, Ucol,tU_{\mathrm{col},t} may be different at successive time steps.

Nevertheless, the D2Q5 model provides an important circuit-level simplification. Since the equilibrium direction ∣e^eq⟩\lvert\hat{e}_{\mathrm{eq}}\rangle is fixed, the state-preparation circuit associated with the equilibrium state and its inverse can be compiled once and reused in every collision block. Furthermore, for a fixed lattice and streaming implementation, UstreamU_{\mathrm{stream}} is also reused throughout the evolution. Thus, only the state-dependent part of the rotation collision needs to be updated as the distribution evolves.

Once the state-dependent parameters of the individual collision blocks are specified, the resulting collision–streaming blocks of D2Q5 model can be connected sequentially into a single unitary circuit, and no projective measurement is required solely for the composition of adjacent unitary blocks. The coherent determination of the state-dependent collision parameters from an intermediate quantum state is, however, a separate algorithmic problem, and is not addressed in the present construction.

State-Dependent Rotation Collision for the D2Q9 Model

For the D2Q9 model, the equilibrium distribution function is given by [18]

fieq=ρwi(1+ei⋅ucs2+(ei⋅u)22cs4−∣u∣22cs2)(53)f_i^{\mathrm{eq}}=\rho w_i\left(1+\frac{e_i\cdot u}{c_s^2}+\frac{(e_i\cdot u)^2}{2c_s^4}-\frac{\lvert u\rvert^2}{2c_s^2}\right) \tag*{(53)}

where u=∑jejfj/ρu=\sum_j e_jf_j/\rho is determined from the first-order moment of distribution function. Unlike the D2Q5 model, the equilibrium distribution function depends nonlinearly on the current state, and the normalized equilibrium state ∣ψeq,t⟩\lvert\psi_{\mathrm{eq},t}\rangle varies during the evolution. Consequently, the rotation angle and the collision operator Ucol,tU_{\mathrm{col},t} must be updated at each time step.

Although the evolution of D2Q9 model can still be written as a sequential composition of collision–streaming blocks, the equilibrium-state preparation circuit cannot be precompiled and reused as that in D2Q5 model. In the present implementation, the distribution function is recovered by measurement at each time step, after which uu, fieqf_i^{\mathrm{eq}}, and the corresponding rotation collision operator are updated before the next step.

Structure of the Sequential Collision Circuit

According to § 3.3, the TT-step evolution of the D2Q5 model is represented by the ordered sequence

Ustep,T−1⋯Ustep,1Ustep,0,Ustep,t=UstreamUcol,t(54)U_{\mathrm{step},T-1}\cdots U_{\mathrm{step},1}U_{\mathrm{step},0},\qquad U_{\mathrm{step},t}=U_{\mathrm{stream}}U_{\mathrm{col},t} \tag*{(54)}

where the streaming operator is fixed, while the rotation collision operator generally depends on the instantaneous state.

The gate-level implementation of the collision operator is based on the Shende–Bullock–Markov state preparation scheme (SBM) [21] and phase reflection operators, with the detailed SBM construction provided in Appendix § 7. A phase reflection about a state ∣ϕ⟩\lvert\phi\rangle is defined as

Sϕ(α)=I+(eiα−1)∣ϕ⟩⟨ϕ∣(55)S_{\phi}(\alpha)=I+(e^{i\alpha}-1)\lvert\phi\rangle\langle\phi\rvert \tag*{(55)}

which is unitary for any real α\alpha. The collision operator at time tt can be decomposed as

Ucol,t=Sψt(γt)Seq(βt)(56)U_{\mathrm{col},t}=S_{\psi_t}(\gamma_t)S_{\mathrm{eq}}(\beta_t) \tag*{(56)}

where the phase parameters are determined from the instantaneous angles ϕt\phi_t and θt=ϕt/τ\theta_t=\phi_t/\tau. In particular,

sin⁡(βt/2)=sin⁡θtsin⁡2ϕt,γt=−arg⁡[eiα0,t(1+cos⁡2ϕt(eiβt−1))](57)\sin(\beta_t/2)=\frac{\sin\theta_t}{\sin2\phi_t},\qquad \gamma_t=-\arg\left[e^{i\alpha_{0,t}}\left(1+\cos^2\phi_t(e^{i\beta_t}-1)\right)\right] \tag*{(57)}

where α0,t=−π/2−βt/2\alpha_{0,t}=-\pi/2-\beta_t/2. Thus, no numerical optimization is required once ϕt\phi_t is specified.

If SBMϕ\mathrm{SBM}_{\phi} prepares ∣ϕ⟩\lvert\phi\rangle from ∣0⟩\lvert0\rangle, the reflection can be implemented as

Sϕ(α)=SBMϕP0(α)SBMϕ†(58)S_{\phi}(\alpha)=\mathrm{SBM}_{\phi}P_0(\alpha)\mathrm{SBM}_{\phi}^{\dagger} \tag*{(58)}

where

P0(α)=I+(eiα−1)∣0⟩⟨0∣(59)P_0(\alpha)=I+(e^{i\alpha}-1)\lvert0\rangle\langle0\rvert \tag*{(59)}

is a multi-controlled phase operation. The collision circuit is therefore implemented as

∣0⟩→SBMψt→SBMeq†→P0(βt)→SBMeq→SBMψt†→P0(γt)→SBMψt→∣ψ′⟩.(60)\lvert0\rangle\to\mathrm{SBM}_{\psi_t} \to\mathrm{SBM}_{\mathrm{eq}}^\dagger\to P_0(\beta_t) \to\mathrm{SBM}_{\mathrm{eq}} \to\mathrm{SBM}_{\psi_t}^\dagger\to P_0(\gamma_t) \to\mathrm{SBM}_{\psi_t} \to\lvert\psi'\rangle. \tag*{(60)}

For the D2Q5 model, the normalized equilibrium direction ∣e^eq⟩\lvert\hat{e}_{\mathrm{eq}}\rangle is independent of the instantaneous distribution function. Consequently, SBMeq\mathrm{SBM}_{\mathrm{eq}} and SBMeq†\mathrm{SBM}_{\mathrm{eq}}^\dagger can be compiled once and reused in every collision block. By contrast, SBMψt\mathrm{SBM}_{\psi_t}, the phase parameters βt\beta_t and γt\gamma_t generally vary with the instantaneous state.

Once the state-dependent parameters of the individual collision blocks are specified, these unitary collision–streaming blocks can be connected sequentially without projective measurement solely for their composition. The coherent determination of these state-dependent parameters during evolution is a separate algorithmic problem.

Numerical results

D2Q9 Model: Taylor–Green Vortex

The first benchmark we consider is the two-dimensional Taylor–Green vortex, which admits an exact solution of the incompressible Navier–Stokes equations defined on the periodic domain [0,1]2[0,1]^2,

ux(x,y,t)=−U0cos⁡(2πx)sin⁡(2πy)e−2ν(2π)2t,uy(x,y,t)=U0sin⁡(2πx)cos⁡(2πy)e−2ν(2π)2t.(61)\begin{aligned} u_x(x,y,t) &= -U_0 \cos(2\pi x)\sin(2\pi y)e^{-2\nu(2\pi)^2t},\\ u_y(x,y,t) &= U_0 \sin(2\pi x)\cos(2\pi y)e^{-2\nu(2\pi)^2t}. \tag*{(61)} \end{aligned}

The exponential decay is governed by viscous dissipation and involves fully nonlinear Navier–Stokes dynamics, which cannot be represented within the reduced D2Q5 velocity set, and the D2Q9 model should be adopted. The simulation parameters are N=32N = 32, Re=100Re = 100, U0=0.1U_0 = 0.1, and τ=0.596\tau= 0.596. The initial condition of distribution function is constructed by its equilibrium part with the exact velocity field Eq. (61) at t=0t = 0, and the periodic boundary conditions are imposed in both spatial directions. To evaluate the difference between the rotation collision operator and the classical BGK operator, we consider the vorticity ω=∂xuy−∂yux\omega= \partial_x u_y - \partial_y u_x, approximated using a second-order central difference scheme. The comparison is performed at the half-life time t1/2t_{1/2}, defined as the time at which the amplitude of the velocity field decays to half of its initial value, i.e., U(t1/2)/U0=1/2U(t_{1/2})/U_0 = 1/2.

Figure 1 shows the vorticity fields obtained from both collision operators. From this figure, one can observe that the two vorticity distributions exhibit nearly identical spatial patterns, including the locations and signs of the vortices, indicating that the rotation collision operator preserves the essential dissipative dynamics of the classical BGK model. We also note that there are some minor differences in the vorticity magnitude, which arise from the nonlinear mapping between the amplitude evolution in the quantum representation and the reconstruction of distribution function. These results demonstrate that the proposed unitary collision operator provides a consistent approximation to the classical BGK relaxation for the Taylor–Green vortex.

Vorticity contours for the Taylor–Green vortex at $t=1/2$, comparing BGK and QLBM

Figure 1. Vorticity contours for the Taylor–Green vortex at Re=100Re = 100, N=32N = 32, U0=0.1U_0 = 0.1, τ=0.596\tau= 0.596, evaluated at the half-life time t1/2t_{1/2} of the velocity decay. Left: vorticity obtained by the classical BGK collision operator. Right: vorticity obtained by the rotation collision operator

D2Q9 Model: Poiseuille Flow

The second benchmark is the two-dimensional plane Poiseuille flow driven by a uniform body force, and the cases with three different Reynolds numbers, i.e., Re=10Re = 10, 4040, and 100100, are used to test the D2Q9 model. In the present QLBM, the collision step is implemented as a unitary rotation operator on the quantum state, as described in § 2. The streaming step and the boundary conditions can be handled within a fully quantum framework via the singular value decomposition approach [12], which embeds non-unitary operators into unitary dilations. In the present work, however, both the streaming step and the half-way bounce-back scheme are treated classically, and the forcing scheme developed by Guo et al. [9] is adopted to incorporate the effect of body force.

The driving force is fixed for all three cases, and the kinematic viscosity is varied to achieve the different Reynolds numbers. Figure 2 presents the velocity profiles ux(y)u_x(y) at Re=10Re = 10, 4040, and 100100, and the numerical results are in good agreement with the analytical solution

Velocity profiles of three cases with $Re = 10$, $40$, and $100$

Figure 2. Velocity profiles of three cases with Re=10Re = 10, 4040, 100100. The results of lattice BGK method and QLBM agree with the analytical solution.

ux(y)=G2νy(H−y),(62)u_x(y) = \frac{G}{2\nu}y(H-y), \tag*{(62)}

and also, the classical lattice BGK method and QLBM yield almost the same results.

Error Scaling with Time Step and Relaxation Time

To quantify the discrepancy between the rotation collision and the classical BGK collision, we perform two parameter sweeps using the two-dimensional Taylor–Green vortex on a uniform grid N×NN \times N with N=32N = 32. In the first test,

time step δt\delta t is varied with a fixed τ=0.7\tau= 0.7, while in the second test, τ\tau is varied at a fixed U0=0.1U_0 = 0.1. The lattice BGK method and QLBM use same initial conditions, streaming procedures, and boundary conditions.

To evaluate the discrepancy, the following relative velocity errors are adopted

EBGK,exu=∥uBGK−uex∥2,h∥uex∥2,h,(63)E^{u}_{\mathrm{BGK,ex}} = \frac{\lVert u^{\mathrm{BGK}} - u^{\mathrm{ex}}\rVert_{2,h}}{\lVert u^{\mathrm{ex}}\rVert_{2,h}}, \tag*{(63)}
EQ,exu=∥uQ−uex∥2,h∥uex∥2,h,(64)E^{u}_{\mathrm{Q,ex}} = \frac{\lVert u^{\mathrm{Q}} - u^{\mathrm{ex}}\rVert_{2,h}}{\lVert u^{\mathrm{ex}}\rVert_{2,h}}, \tag*{(64)}
EQ,BGKu=∥uQ−uBGK∥2,h∥uBGK∥2,h,(65)E^{u}_{\mathrm{Q,BGK}} = \frac{\lVert u^{\mathrm{Q}} - u^{\mathrm{BGK}}\rVert_{2,h}}{\lVert u^{\mathrm{BGK}}\rVert_{2,h}}, \tag*{(65)}

where ∥⋅∥2,h\lVert\cdot\rVert_{2,h} denotes the discrete L2L_2 norm.

Figure 3(a) shows the relative errors at different time steps, and the slope of dotted line is 2. From this figure, one can see that the errors of the lattice BGK method is almost the same as those of QLBM, and the discrepancy EQ,BGKuE^{u}_{\mathrm{Q,BGK}} between the rotation collision and the classical BGK collision follows the scaling

Figure 3. Relative errors of velocity based on the lattice BGK method and QLBM. (a) Effect of time step (τ=0.7\tau= 0.7), the dotted line denotes the reference scaling αδt2\alpha\delta t^2. (b) Effect of relaxation time (N=32N = 32 and U0=0.1U_0 = 0.1); the two collision operators coincide at τ=1\tau= 1 and discrepancy between them is up to floating-point round-off.

EQ,BGKu=O(δt2),(66)E^{u}_{\mathrm{Q,BGK}} = O(\delta t^2), \tag*{(66)}

which is consistent with the theoretical analysis in § 2.

Figure 3(b) presents the effect of the relaxation time on the numerical results. It is also found that these two methods almost give the same results, while EQ,BGKuE^{u}_{\mathrm{Q,BGK}} reaches the floating-point round-off level at τ=1\tau= 1. As stated previously, when τ=1\tau= 1, θ=ϕ\theta= \phi so that both collision operators map the distribution function exactly to its equilibrium part. However, if τ≠1\tau\ne1, the discrepancy increases due to the higher-order difference between the two collision formulations.

Overall, the rotation collision preserves the accuracy of lattice BGK method, and exhibits the predicted second-order discrepancy with respect to δt\delta t.

Discussion

Circuit reuse and multi-step composition For the D2Q5 model, the linear form of the equilibrium distribution function can render the normalized equilibrium state independent of the instantaneous distribution, and in this case the equilibrium-state preparation circuit can be compiled once and reused throughout the evolution. The complete collision operator, however, generally depends on the instantaneous quantum state through the corresponding rotation parameters. Once the state-dependent parameters of the individual collision blocks are specified, the unitary collision–streaming blocks can be composed sequentially without projective measurement solely for their composition, thereby preserving quantum coherence during the corresponding unitary evolution.

State-dependent equilibrium distribution function of the D2Q9 model. In contrast, the equilibrium distribution function of D2Q9 model depends nonlinearly on the local macroscopic velocity, which causes both the equilibrium state and collision operator to vary with the current state. In the present hybrid implementation, the macroscopic variables are reconstructed from the current distribution function and used classically to update the equilibrium state and collision parameters at each time step. A fully coherent determination of these state-dependent quantities would require additional quantum subroutines.

Comparison with existing approaches and limitations. Compared with existing quantum collision schemes, the present rotation-based collision operator requires no ancilla qubits or post-selection [14], and does not need to remove the collision step [23], and additionally, the approximation of the non-equilibrium distribution function based on the Chapman-Enskog expansion [32] can also be avoided. Moreover, the difference between the rotation collision and BGK collision is of the order O(δt2)O(\delta t^{2}).

We would also like to point out that the present QLBM still has several limitations. First, the complexity analysis does not include the cost of state preparation and terminal measurement, which scales as O(N)O(N) and may partially offset the quantum advantage. Second, the streaming step and boundary conditions are treated classically. How to realize the quantum implementation of them remains an open problem, which may be potentially addressable via the singular value decomposition framework [12]. Finally, all results are obtained via classical simulations, and the impact of hardware noise on real quantum devices also needs to be investigated.

Conclusion

In this work, we developed a quantum lattice Boltzmann framework based on a rotation collision operator, providing a unitary formulation of the classical BGK relaxation step. The operator is implemented as a rotation in the two-dimensional subspace spanned by the current state and its equilibrium counterpart, and is unitary by construction. It is also shown that the difference between the rotation collision and the classical BGK operator is of the order O(δt2)O(\delta t^{2}).

A key theoretical result of this work is the reusable equilibrium-state structure of the D2Q5 model. Owing to the linear dependence of its equilibrium distribution on the conservative scalar ϕ\phi, the normalized equilibrium state is independent of the distribution function, allowing the corresponding equilibrium-state preparation circuit to be compiled once and reused throughout the evolution. The complete collision operator, however, generally depends on the instantaneous state. Accordingly, the TT-step evolution is represented by the ordered composition ∣Ψ(T)⟩=Ustep,T−1⋯Ustep,0∣Ψ(0)⟩|\Psi(T)\rangle=U_{\mathrm{step},T-1}\cdots U_{\mathrm{step},0}|\Psi(0)\rangle with Ustep,t=UstreamUcol,tU_{\mathrm{step},t}=U_{\mathrm{stream}}U_{\mathrm{col},t}. In contrast, the equilibrium distribution function of D2Q9 depends on the macroscopic velocity, so that both the equilibrium state preparation and collision parameters generally vary during the evolution.

Numerical results of the Taylor–Green vortex and Poiseuille flow demonstrate that there is a good agreement between the proposed QLBM and the classical lattice BGK method, and the difference between them is of the order O(δt2)O(\delta t^{2}). The circuit implementation of D2Q5 model, constructed via the Shende–Bullock–Markov state preparation scheme and phase reflection operator, is provided in the appendix together with explicit gate-count comparison.

Overall, the present work establishes a unitary and gate-level realizable formulation of lattice Boltzmann method, and identifies the reusable equilibrium-state circuit structure of the D2Q5 model as an advantageous feature for quantum implementation.

Shende–Bullock–Markov State Preparation Circuit

The Shende–Bullock–Markov (SBM) scheme [21] provides a systematic construction for quantum-state preparation. In the present work, the target amplitudes are real and non-negative because they are defined by ai=fi/ρa_i=\sqrt{f_i/\rho}, thus only the real-amplitude part of the SBM construction is required. The resulting state-preparation circuit provides the basic primitive used to implement the phase reflections in the D2Q5 collision circuit.

Problem statement

Given a normalized real state

∣φ⟩=∑k=02n−1φk∣k⟩,φk≥0,∑kφk2=1,(67)|\varphi\rangle= \sum_{k=0}^{2^n-1} \varphi_k |k\rangle,\qquad\varphi_k \ge0,\qquad\sum_k \varphi_k^2 = 1, \tag*{(67)}

the objective is to construct a unitary SBMφ\mathrm{SBM}_{\varphi} satisfying

SBMφ∣0⟩⊗n=∣φ⟩.(68)\mathrm{SBM}_{\varphi}|0\rangle^{\otimes n} = |\varphi\rangle. \tag*{(68)}

Recursive construction

The circuit is constructed recursively from the most significant qubit (level ℓ=0\ell= 0) to the least significant qubit (level ℓ=n−1\ell= n - 1). At the level ℓ\ell, the amplitudes are divided into 2ℓ2^\ell blocks, each of them is further separated into left and right halves, φL(g)\varphi_L^{(g)} and φR(g)\varphi_R^{(g)}. The corresponding rotation angle is

ϑℓ,g=2arctan⁡(∥φR(g)∥∥φL(g)∥),(69)\vartheta_{\ell,g} = 2\arctan\left(\frac{\lVert\varphi_R^{(g)}\rVert}{\lVert\varphi_L^{(g)}\rVert}\right), \tag*{(69)}

where the zero-norm branches are omitted. The gate at level ℓ\ell is an Ry(ϑℓ,g)R_y(\vartheta_{\ell,g}) rotation on qubit q=n−1−ℓq = n - 1 - \ell, controlled by the preceding ℓ\ell qubits according to the binary pattern associated with block gg. Thus, ℓ=0\ell= 0 corresponds to an uncontrolled RyR_y rotation, whereas higher levels involve controlled rotations.

Three-qubit case (D2Q5 model)

For the D2Q5 model, three qubits are used to embed the five discrete velocity populations into the eight-dimensional state

∣φ⟩=∑i=04φi∣i⟩,(70)|\varphi\rangle= \sum_{i=0}^{4} \varphi_i |i\rangle, \tag*{(70)}

where the remaining amplitudes are set to be zero.

At the level ℓ=0\ell= 0, one RyR_y rotation is applied to q2q_2. At the level ℓ=1\ell= 1, two controlled RyR_y rotations act on q1q_1, conditioned on q2=0q_2 = 0 and q2=1q_2 = 1, respectively. At the level ℓ=2\ell= 2, the non-zero amplitude pairs (φ0,φ1)(\varphi_0,\varphi_1) and (φ2,φ3)(\varphi_2,\varphi_3) give two nontrivial doubly controlled rotations on q0q_0. The remaining padded branches produce zero rotation angles and are omitted.

For the equilibrium state in D2Q5 model,

∣e^eq⟩∝∑iλi∣i⟩,(71)|\hat{e}_{\mathrm{eq}}\rangle\propto\sum_i \sqrt{\lambda_i}|i\rangle, \tag*{(71)}

the amplitudes are independent of the instantaneous distribution functions. Consequently, SBMeq\mathrm{SBM}_{\mathrm{eq}} and its inverse can be compiled once and reused in every collision block.

By contrast, SBMψt\mathrm{SBM}_{\psi_t} depends on the instantaneous state. Once the amplitudes defining ∣ψt⟩|\psi_t\rangle are specified, the rotation angles can be obtained directly from Eq. (A.3). The determination of these amplitudes during a fully coherent multistep evolution is separate from the SBM state-preparation procedure.

Three-qubit SBM circuit for the amplitude encoding of D2Q5 model

Figure 4. Three-qubit SBM circuit for the amplitude encoding of D2Q5 model. Qubits are ordered from q2q_2 (MSB) to q0q_0 (LSB). Open and filled controls denote conditions on ∣0⟩|0\rangle and ∣1⟩|1\rangle, respectively. Gates associated with zero-padded branches are omitted.

Collision Circuit Construction of the D2Q5 Model

The collision circuit of D2Q5 model is implemented by using generalized phase reflections. At time tt, the prescribed rotation angle is θt=ϕt/τ\theta_t=\phi_t/\tau, where ϕt=arccos⁡⟨ψt∣e^eq⟩\phi_t=\arccos\langle\psi_t|\hat{e}_{\mathrm{eq}}\rangle.

Phase-reflection realization

For a normalized state ∣ϕ⟩|\phi\rangle, the generalized phase reflection is defined as

Sϕ(α)=I+(eiα−1)∣ϕ⟩⟨ϕ∣.(72)S_{\phi}(\alpha)=I+(e^{i\alpha}-1)|\phi\rangle\langle\phi|. \tag*{(72)}

Let SBMϕ\mathrm{SBM}_{\phi} satisfy SBMϕ∣0⟩⊗n=∣ϕ⟩\mathrm{SBM}_{\phi}|0\rangle^{\otimes n}=|\phi\rangle, then we have

Sϕ(α)=SBMϕP0(α)SBMϕ†.(73)S_{\phi}(\alpha)=\mathrm{SBM}_{\phi}P_0(\alpha)\mathrm{SBM}_{\phi}^{\dagger}. \tag*{(73)}

where

P0(α)=I+(eiα−1)∣0⟩⟨0∣.(74)P_0(\alpha)=I+(e^{i\alpha}-1)|0\rangle\langle0|. \tag*{(74)}

where the phase eiαe^{i\alpha} has been applied to ∣0⟩⊗n|0\rangle^{\otimes n}.

The collision operator can be decomposed as

Ucol,t=Sψt(γt)Seq(βt).(75)U_{\mathrm{col},t}=S_{\psi_t}(\gamma_t)S_{\mathrm{eq}}(\beta_t). \tag*{(75)}

The phase parameters are chosen so that the two reflections generate the desired rotation in span⁡{∣ψt⟩,∣e^eq⟩}\operatorname{span}\{|\psi_t\rangle,|\hat{e}_{\mathrm{eq}}\rangle\}:

sin⁡(βt2)=sin⁡θtsin⁡2ϕt,γt=−arg⁡[eiα0,t(1+cos⁡2ϕt(eiβt−1))],(76)\sin\left(\frac{\beta_t}{2}\right)=\frac{\sin\theta_t}{\sin2\phi_t},\qquad\gamma_t=-\arg\left[e^{i\alpha_{0,t}}\left(1+\cos^2\phi_t\left(e^{i\beta_t}-1\right)\right)\right], \tag*{(76)}

where

α0,t=−π2−βt2.(77)\alpha_{0,t}=-\frac{\pi}{2}-\frac{\beta_t}{2}. \tag*{(77)}

Thus, once ϕt\phi_t is specified, βt\beta_t and γt\gamma_t can be obtained directly without numerical optimization.

With the help of Eq. (B.2), the single-step collision circuit takes the form

∣ψt⟩→SBMeq†→P0(βt)→SBMeq→SBMψt†→P0(γt)→SBMψt→∣ψt′⟩.(78)|\psi_t\rangle\rightarrow\mathrm{SBM}_{\mathrm{eq}}^{\dagger}\rightarrow P_0(\beta_t)\rightarrow\mathrm{SBM}_{\mathrm{eq}}\rightarrow\mathrm{SBM}_{\psi_t}^{\dagger}\rightarrow P_0(\gamma_t)\rightarrow\mathrm{SBM}_{\psi_t}\rightarrow|\psi'_t\rangle. \tag*{(78)}

For the D2Q5 model, ∣e^eq⟩|\hat{e}_{\mathrm{eq}}\rangle is fixed, thus SBMeq\mathrm{SBM}_{\mathrm{eq}} and its inverse can be compiled once and reused. By contrast, SBMψt\mathrm{SBM}_{\psi_t}, βt\beta_t, and γt\gamma_t generally depend on the instantaneous state.

Sequential multi-step circuit

A complete time step is

Ustep,t=UstreamUcol,t,(79)U_{\mathrm{step},t}=U_{\mathrm{stream}}U_{\mathrm{col},t}, \tag*{(79)}

and the TT-step evolution is

∣Ψ(T)⟩=Ustep,T−1⋯Ustep,1Ustep,0∣Ψ(0)⟩.(80)|\Psi(T)\rangle=U_{\mathrm{step},T-1}\cdots U_{\mathrm{step},1}U_{\mathrm{step},0}|\Psi(0)\rangle. \tag*{(80)}

The collision blocks are generally state-dependent, while the equilibrium-state preparation circuit and the streaming operator are reused throughout the evolution. Once the parameters of the individual collision blocks are specified, these unitary blocks can be connected sequentially without projective measurement solely for their composition.

Sequential T-step circuit of D2Q5 model

Figure 5. Sequential TT-step circuit of D2Q5 model. Each step consists of a state-dependent collision Ucol,tU_{\mathrm{col},t} followed by the fixed streaming operator.

The Collision Circuit Resource of D2Q9 model

Two quantum implementations of the rotation collision operator in D2Q9 model are compared. Method A: a UnitaryGate to apply the full 16×1616 \times16 unitary matrix UcolU_{\mathrm{col}} directly on 4 qubits; Method B: UcolU_{\mathrm{col}} [26], encode the state in the eigenbasis, and apply the eigenphases.

Both circuits were transpiled to the {U3,CX}\{U_3, CX\} basis at optimization level 0 by using Qiskit [7]. Table 1 reports the gate counts and circuit depth for the collision subcircuit only, with state preparation excluded.

MetricUnitaryGateDiagonalGate
Qubits45
CXCX gates95196
U3U_3 gates180286
Total gates275482
Circuit depth189328

Table 1. The collision circuit resources after transpilation to the {U3,CX}\{U_3, CX\} basis with optimization level 0 (state preparation excluded).

The UnitaryGate approach requires approximately half number of CXCX gates compared with the DiagonalGate approach (ratio 196/95≈2.06196/95 \approx2.06), while using one fewer qubit. Both methods require a projective measurement at each time step to extract the macroscopic velocity u(t)\mathbf{u}^{(t)} and reconstruct Ucol(t+1)U_{\mathrm{col}}^{(t+1)}, consistent with the non-concatenability of the D2Q9 model in § 3.6.

References

  1. [1]F. Arute, K. Arya, R. Babుషh, D. Bacon, J. C. Bardin, R. Barends, R. Biswas, S. Boixo, F. G. Brandao, D. A. Buell, et al., Quantum supremacy using a programmable superconducting processor, Nature, 574 (2019), pp. 505–510.DOI
  2. [2]Y. Cao, A. Papageorgiou, I. Petras, J. Traub, and S. Kais, Quantum algorithm and circuit design solving the Poisson equation, New Journal of Physics, 15 (2013), p. 013021.DOI
  3. [3]Z. Chai and B. Shi, Multiple-relaxation-time lattice boltzmann method for the navier-stokes and nonlinear convection-diffusion equations: Modeling, analysis, and elements, Physical Review E, 102 (2020), p. 023306.DOI
  4. [4]S. Chen and G. D. Doolen, Lattice Boltzmann method for fluid flows, Annual Review of Fluid Mechanics, 30 (1998), pp. 329–364.
  5. [5]A. M. Childs, On the relationship between continuous- and discrete-time quantum walk, Communications in Mathematical Physics, 294 (2010), pp. 581–603.DOI
  6. [6]D. Coppersmith, An approximate Fourier transform useful in quantum factoring, arXiv preprint quant-ph/0201067, (2002).DOI
  7. [7]M. Fingerhuth, T. Babej, and P. Wittek, Open source software in quantum computing, PLOS ONE, 13 (2018), p. e0208561.DOI
  8. [8]F. Gaitan, Finding flows of a Navier–Stokes fluid through quantum computing, npj Quantum Information, 6 (2020), p. 61.DOI
  9. [9]Z. Guo, C. Zheng, and B. Shi, Discrete lattice effects on the forcing term in the lattice Boltzmann method, Physical review E, 65 (2002), p. 046308.pubmed.ncbi.nlm.nih.gov/12006014
  10. [10]A. W. Harrow, A. Hassidim, and S. Lloyd, Quantum algorithm for linear systems of equations, Physical Review Letters, 103 (2009), p. 150502.pubmed.ncbi.nlm.nih.gov/19905613
  11. [11]T. Krüger, H. Kusumaatmaja, A. Kuzmin, O. Shardt, G. Silva, and E. M. Viggen, The lattice Boltzmann method, vol. 10, Springer, 2017.
  12. [12]E. D. Kumar and S. H. Frankel, Quantum unitary matrix representation of the lattice Boltzmann model for low reynolds fluid flow simulation, AVS Quantum Science, 7 (2025), p. 013802.
  13. [13]B. Ljubomir, Quantum algorithm for the Navier–Stokes equations by using the streamfunction-vorticity formulation and the lattice Boltzmann method, International Journal of Quantum Information, 20 (2022), p. 2150039.DOI
  14. [14]A. Mezzacapo, M. Sanz, L. Lamata, I. Egusquiza, S. Succi, and E. Solano, Quantum simulator for transport phenomena in fluid flows, Scientific Reports, 5 (2015), p. 13153.DOI
  15. [15]M. A. Nielsen and I. L. Chuang, Quantum computation and quantum information, Cambridge University Press, 2010.
  16. [16]R. Portugal, Quantum walks and search algorithms, vol. 19, Springer, 2013.DOI
  17. [17]J. Preskill, Quantum Computing in the NISQ era and beyond, Quantum, 2 (2018), p. 79, https://doi.org/10.22331/q-2018-08-06-79, https://doi.org/10.22331/q-2018-08-06-79.
  18. [18]Y.-H. Qian, D. D’Humières, and P. Lallemand, Lattice BGK models for Navier-Stokes equation, EPL (Europhysics Letters), 17 (1992), pp. 479–484.DOI
  19. [19]M. Schuld and N. Killoran, Quantum machine learning in feature Hilbert spaces, Physical Review Letters, 122 (2019), p. 040504.DOI
  20. [20]A. Shakeel, Efficient and scalable quantum walk algorithms via the quantum Fourier transform, Quantum Information Processing, 19 (2020), p. 323.DOI
  21. [21]V. V. Shende, S. S. Bullock, and I. L. Markov, Synthesis of quantum logic circuits, in Proceedings of the 2005 Asia and South Pacific Design Automation Conference, 2005, pp. 272–275.
  22. [22]S. Succi, The lattice Boltzmann equation: for fluid dynamics and beyond, Oxford university press, 2001.
  23. [23]B. N. Todorova and R. Steijl, Quantum algorithm for the collisionless Boltzmann equation, Journal of Computational Physics, 409 (2020), p. 109347.DOI
  24. [24]B. Wang, Z. Meng, Y. Zhao, and Y. Yang, Quantum lattice Boltzmann method for simulating nonlinear fluid dynamics, npj Quantum Information, 11 (2025), p. 196, https://doi.org/10.1038/s41534-025-01142-6.
  25. [25]D. Wawrzyniak, J. Winter, S. Schmidt, T. Indinger, C. F. Janssen, U. Schramm, and N. A. Adams, A quantum algorithm for the lattice-Boltzmann method advection-diffusion equation, Computer Physics Communications, 306 (2025), p. 109373.DOI
  26. [26]J. Welch, D. Greenbaum, S. Mostame, and A. Aspuru-Guzik, Efficient quantum circuits for diagonal unitaries without ancillas, New Journal of Physics, 16 (2014), p. 033040.
  27. [27]D. A. Wolf-Gladrow, Lattice-gas cellular automata and lattice Boltzmann models: an introduction, Springer, 2004.
  28. [28]W. K. Wootters and W. H. Zurek, A single quantum cannot be cloned, Nature, 299 (1982), pp. 802–803.DOI
  29. [29]Y. Xiao, L. Yang, H. Dong, and Y. Du, Interpolation-supplemented quantum lattice Boltzmann method for incompressible flows with nonuniform mesh and curved boundary, Physical Review E, 113 (2026), p. 045306.DOI
  30. [30]Y. Xiao, L. Yang, C. Shu, and Y. Du, A stable and general quantum fractional-step lattice Boltzmann method for incompressible flows, arXiv preprint arXiv:2603.00558, (2026).arxiv.org/abs/2603.00558
  31. [31]J. Yepez, Quantum lattice-gas model for the diffusion equation, International Journal of Modern Physics C, 12 (2001), pp. 1285–1303.DOI
  32. [32]K.-Y. Zeng, X.-D. Niu, A. Khan, D.-C. Li, and H. Yamaguchi, A quantum computing-based lattice Boltzmann method with a linearized non-equilibrium collision operator and modular circuit for practical flow simulation, Physics of Fluids, 37 (2025), p. 081701.DOI
  33. [33]K.-Y. Zeng, Z.-D. Wei, X.-D. Niu, A. Khan, D.-C. Li, and H. Yamaguchi, A quantum lattice Boltzmann method for solving the navier-stokes systems with a linearized non-equilibrium collision operator and modular circuit, Acta Mechanica Sinica, 42 (2026), p. 725741.DOI

Paper details

vlnc:W_hknek2v4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents