betaFeedback

WorksPhysics and Astronomy

Quantum Circuits for Variable-Coefficient Vlasov Dynamics via Schrödingerisation

Abstract

High-dimensional Vlasov equations present promising use cases for quantum simulation, but their transport coefficients must be implemented coherently in order to preserve a compact quantum representation. We develop explicit Schrödingerised circuits for prescribed external field Vlasov dynamics without classical field updates. In 1D1V, velocity-dependent streaming is compiled directly from the encoded velocity register, while a non-linear cubic field is evaluated using reversible fixed-point arithmetic. For fixed polynomial degree, the field workspace scales linearly with the coordinate register width, while the field evaluation subroutine has polynomial depth. As a proof of concept, we simulate a minimal coupled 3D3V anharmonic Penning system and realise mixed electric-field terms and transverse magnetic coupling on a six-dimensional phase space in agreement with classical calculations. Resource analysis shows approximately linear growth with the number of time steps, while identifying the auxiliary register as the dominant depth bottleneck. Structured preparation and register recovery of the auxiliary state reveal an interplay between loading cost and readout. Together, our algorithm demonstrates a fully coherent quantum simulation of high-dimensional Vlasov dynamics with variable coefficients and identifies the practical resource costs that must be controlled for large-scale quantum implementation.

Introduction

The Vlasov equation provides a kinetic description of collisionless plasmas, evolving a distribution function over both configuration and velocity space. In the physically standard three-dimensional setting, the distribution f=f(x,v,t)f = f(\mathbf{x}, \mathbf{v}, t) therefore occupies a six-dimensional phase space. The resulting curse of dimensionality is a central computational challenge of grid-based Vlasov simulation [1]. For a direct tensor-product discretisation with NN points along each coordinate, the phase-space representation contains N6N^{6} values, causing memory and computational requirements to grow rapidly with resolution. Consequently, fully kinetic 4D–6D Vlasov calculations have required increasingly large high-performance computing resources, including implementations on large-scale supercomputers [2].

The high-dimensional Vlasov equation is a natural candidate for quantum simulation. A discretised function on dd coordinates with N=2nN = 2^{n} points per coordinate can be amplitude encoded using dndn qubits rather than explicitly storing the corresponding NdN^{d} grid. Sato et al. developed explicit Hamiltonian simulation circuits for finite-difference hyperbolic partial differential equations, whose resource requirements scale polynomially with the register width nn [3]. Hu et al. subsequently incorporated these circuit constructions into the Schrödingerisation framework, extending them to linear PDE discretisations with non-unitary evolution [4]. Assuming the spatial resolution scales with algorithmic error (h=O(ε)h = O(\varepsilon)), their complexity analysis identifies high-dimensional regimes where a quantum advantage is possible, including d>5d > 5 for the heat equation and d>4d > 4 for upwind advection [4].

Quantum algorithms have been proposed for several forms of Vlasov dynamics. Engel et al. formulated a Hamiltonian simulation approach for linearised Vlasov dynamics [5]. Ameri et al. considered the linear Vlasov equation in a Fourier–Hermite representation, including collisional effects [6]. Toyoizumi et al. applied quantum singular-value transformation to Hamiltonian simulation of the linearised Vlasov–Poisson equation [7]. Miyamoto et al. developed a quantum algorithm for the 6 + 1-dimensional Vlasov description of massive-neutrino dynamics in large-scale structure formation [8]. More recently, Higuchi et al. numerically evaluated Hamiltonian simulation circuits for Vlasov advection and two-stream instability within a hybrid Vlasov–Maxwell calculation [9].

These studies establish the potential of quantum algorithms for kinetic equations, but leave open the problem of constructing an explicit quantum circuit for Vlasov transport with coefficients that depend directly on the encoded phase-space coordinates. In particular, the velocity-dependent streaming term and non-linear force term require the coefficient dependence itself to be implemented coherently. This is especially relevant for hybrid quantum–classical Vlasov–Maxwell schemes [9], where repeated measurement of a high-dimensional quantum state and re-encoding of updated classical fields can introduce substantial data-transfer and readout overheads.

In the present work, we consider Vlasov dynamics under a prescribed external field which arises naturally in plasma confinement and control. Einkemmer et al. formulated suppression of Vlasov–Poisson instabilities through optimisation of an externally applied electric field [10], while subsequent work developed analytical external-field constructions for instability control [11]. External potentials can modify particle characteristics and trapping behaviour [12]. The prescribed-field setting provides a useful class of linear kinetic problems in which the field structure is known in advance but the resulting transport remains spatially nonuniform and nontrivial.

The contributions of this work are threefold. First, we develop and validate an explicit Schrödingerised circuit for variable-coefficient 1D1V Vlasov transport, including coherent evaluation of a non-linear prescribed field. Second, we extend the construction to coupled 3D3V dynamics, providing an end-to-end circuit implementation on a six-dimensional phase space. Third, we quantify the resulting logical circuit resources and examine the auxiliary pp-register, including its effect on circuit cost, recovery probability, and structured state preparation. Other than state preparation and final measurement, the quantum simulation acts entirely on quantum information and requires no intermediate field evaluation or quantum–classical feedback.

Theory

1D1V Vlasov equation with prescribed external field

We consider the one-dimensional-in-space, one-dimensional-in-velocity (1D1V) Vlasov transport equation with a prescribed external acceleration field E(x)E(x):

∂tf(x,v,t)+v∂xf(x,v,t)+E(x)∂vf(x,v,t)=0,(1)\partial_t f(x,v,t) + v \partial_x f(x,v,t) + E(x) \partial_v f(x,v,t) = 0, \tag*{(1)}

where f(x,v,t)≥0f(x,v,t) \ge0 denotes the phase-space density for time t≥0t \ge0. In this work, we focus on linear, externally forced dynamics without a self-consistent Poisson solver, which arise naturally in studies of confinement [13] and external-field control [10].

To obtain a nontrivial but structured external field, we consider the double-well potential

Vdw(x)=κ4(x2−a02)2.(2)V_{\mathrm{dw}}(x) = \frac{\kappa}{4}\left(x^2-a_0^2\right)^2. \tag*{(2)}

The corresponding prescribed acceleration field is

E(x)=−dVdwdx=−κ(x3−a02x).(3)E(x) = -\frac{\mathrm{d}V_{\mathrm{dw}}}{\mathrm{d}x} = -\kappa\left(x^3-a_0^2x\right). \tag*{(3)}

It is convenient to rewrite this in the polynomial form

E(x)=αx+βx3,α=κa02,β=−κ.(4)E(x) = \alpha x + \beta x^3, \qquad\alpha= \kappa a_0^2, \qquad\beta= -\kappa. \tag*{(4)}

The prescribed field is cubic in xx, with coefficients α\alpha and β\beta that can subsequently be represented in fixed-point form in the arithmetic subroutines.

We use uniform cell-centred grids in both spatial and velocity coordinates. The quantum registers store the indices j∈{0,…,Nx−1}j \in\{0,\ldots,N_x-1\} and k∈{0,…,Nv−1}k \in\{0,\ldots,N_v-1\}, corresponding to

xj=−Lx2+(j+12)Δx,vk=−Lv+(k+12)Δv.(5)x_j = -\frac{L_x}{2} + \left(j+\frac{1}{2}\right)\Delta x, \qquad v_k = -L_v + \left(k+\frac{1}{2}\right)\Delta v. \tag*{(5)}

For the field-evaluation circuit, the spatial index is represented by the centred integer c(j)=j−Nx/2c(j)=j-N_x/2, corresponding to x~j=Δx c(j)=xj−Δx/2\tilde{x}_j = \Delta x\,c(j) = x_j-\Delta x/2. Thus, the implemented field is evaluated on a grid shifted by half a spatial cell relative to the initial distribution.

Periodic boundary conditions are imposed on both discrete transport coordinates,

f(−Lx2,v,t)=f(Lx2,v,t),f(x,−Lv,t)=f(x,Lv,t),(6)f\left(-\frac{L_x}{2},v,t\right)=f\left(\frac{L_x}{2},v,t\right), \qquad f(x,-L_v,t)=f(x,L_v,t), \tag*{(6)}

consistent with the periodic finite-difference boundary blocks. The prescribed polynomial field E(x)E(x) is evaluated at x~j\tilde{x}_j and is not assumed to be a smooth periodic function of xx.

Expressing Eq. (1) in operator form,

∂tf=(Lx+Lv)f,Lx:=−v∂x,Lv:=−E(x)∂v,(7)\partial_t f = (\mathcal{L}_x + \mathcal{L}_v)f,\qquad\mathcal{L}_x := -v\partial_x,\qquad\mathcal{L}_v := -E(x)\partial_v, \tag*{(7)}

using second-order Strang splitting [14, 15],

fn+1≈eΔt2LxeΔtLveΔt2Lxfn,(8)f^{n+1} \approx e^{\frac{\Delta t}{2}\mathcal{L}_x}e^{\Delta t\mathcal{L}_v}e^{\frac{\Delta t}{2}\mathcal{L}_x}f^n, \tag*{(8)}

where fn(x,v)≈f(x,v,tn)f^n(x,v) \approx f(x,v,t_n) and tn+1=tn+Δtt_{n+1}=t_n+\Delta t. This conveniently separates the dynamics into two transport generators that act separately on the xx- and vv-registers.

Schrödingerisation

The Strang splitting scheme (Eq. (8)) can be implemented on explicit quantum circuits acting on registers that encode the discretised phase-space using Hamiltonian simulation [3]. In this case, however, the finite-difference scheme for the advection equation is limited to central differencing (notorious for spurious numerical oscillations), since quantum evolution must be unitary. Schrödingerisation [16] circumvents this via dimension-lifting, where one introduces an auxiliary coordinate to map a non-unitary equation to a higher-dimensional Schrödinger equation that is unitary, enabling dissipative equations such as the heat equation to be solved [4].

Consider a semi-discrete system first-order in time [16],

∂tf(t)=Af(t),(9)\partial_t\mathbf{f}(t)=A\mathbf{f}(t), \tag*{(9)}

where f(t)\mathbf{f}(t) is the vectorised grid representation of the phase-space density ff and AA is a linear evolution operator that is not necessarily skew-Hermitian. Following [4, 16], the semi-discrete evolution operator is first decomposed into Hermitian and anti-Hermitian parts as

A=A1+iA2,A1=A+A†2,A2=A−A†2i,(10)A=A_1+iA_2,\qquad A_1=\frac{A+A^\dagger}{2},\qquad A_2=\frac{A-A^\dagger}{2i}, \tag*{(10)}

with both A1A_1 and A2A_2 Hermitian. Schrödingerisation introduces an auxiliary warped phase coordinate pp, truncated to p∈[−πR,πR)p\in[-\pi R,\pi R) and discretised using Np=2npN_p=2^{n_p} points,

pl=−πR+lΔp,Δp=2πRNp,l=0,…,Np−1.(11)p_l=-\pi R+l\Delta p,\qquad\Delta p=\frac{2\pi R}{N_p},\qquad l=0,\ldots,N_p-1. \tag*{(11)}

The corresponding npn_p qubit register is denoted by PP. For p≥0p\ge0, we define

Φ(t,p)=e−pf(t).(12)\Phi(t,p)=e^{-p}f(t). \tag*{(12)}

The initial data are extended to p<0p<0 by

Φ(0,p)=e−∣p∣f(0),(13)\Phi(0,p)=e^{-|p|}f(0), \tag*{(13)}

or using the smoothed profile described in Appendix § 10. Using ∂pΦ=−Φ\partial_p\Phi=-\Phi, the lifted equation becomes

∂tΦ(t,p)=−A1∂pΦ(t,p)+iA2Φ(t,p).(14)\partial_t\Phi(t,p)=-A_1\partial_p\Phi(t,p)+iA_2\Phi(t,p). \tag*{(14)}

Fourier transforming in pp, with ∂p↦−iη\partial_p\mapsto-i\eta, gives

∂tΦ^(t,η)=i(ηA1+A2)Φ^(t,η),(15)\partial_t\widehat{\Phi}(t,\eta)=i(\eta A_1+A_2)\widehat{\Phi}(t,\eta), \tag*{(15)}

where

H(η)=ηA1+A2(16)H(\eta) = \eta A_1 + A_2 \tag*{(16)}

is Hermitian.

Let XX and VV denote the position and velocity index registers, with nxn_x and nvn_v qubits, respectively. We encode the discretised phase-space field at time tnt_n as an amplitude-encoded state,

∣ψn⟩=1∥fn∥2∑j=0Nx−1∑k=0Nv−1fj,kn∣j⟩X∣k⟩V,(17)\lvert\psi^n \rangle= \frac{1}{\lVert f^n \rVert_2} \sum_{j=0}^{N_x-1} \sum_{k=0}^{N_v-1} f^n_{j,k} \lvert j \rangle_X \lvert k \rangle_V, \tag*{(17)}

where

fj,kn:=f(xj,vk,tn),∥fn∥2=(∑j,k∣fj,kn∣2)1/2.f^n_{j,k} := f(x_j,v_k,t_n), \qquad\lVert f^n \rVert_2 = \left(\sum_{j,k} \lvert f^n_{j,k}\rvert^2\right)^{1/2}.

The Schrödingerised state is obtained by augmenting this physical X,VX,V state with the auxiliary PP-register profile (Eq. (12)). Time evolution is implemented on the lifted Schrödingerised state, which augments the physical phase-space grid by the auxiliary warped coordinate pp. We follow the explicit circuit construction of [3, 4], in which finite-difference generators are decomposed into structured unitary blocks acting on encoded grid states. This way, the resulting implementation remains fully explicit and does not rely on black-box coefficient oracles.

The Schrödingerised implementation [4] further uses an auxiliary register PP and, for the force step, an ancillary field register EE together with temporary work registers required for reversible arithmetic. The corresponding Strang-split (Eq. (8)) quantum update is

U(Δt)≈Ux(Δt2)Uv(Δt)Ux(Δt2),(18)U(\Delta t) \approx U_x\left(\frac{\Delta t}{2}\right) U_v(\Delta t) U_x\left(\frac{\Delta t}{2}\right), \tag*{(18)}

where UxU_x and UvU_v denote the Schrödingerised circuit blocks for the two split substeps.

Here, we use the original auxiliary pp-register fp(p)=e−∣p∣f_p(p)=e^{-\lvert p\rvert}, which admits a fixed central readout at p=0p=0. Alternative smooth profiles and structured loading strategies are discussed in Appendix A.

Velocity-dependent free-streaming circuit (Block 1)

The first split substep is the free-streaming update

∂tf+v ∂xf=0.(19)\partial_t f + v \,\partial_x f = 0. \tag*{(19)}

We refer to the corresponding quantum implementation of UxU_x simply as Block 1. Unlike [4], the advection coefficient is not a fixed scalar, but the encoded velocity itself. Consequently, the circuit angles depend explicitly on the computational basis state of the velocity register.

Let the velocity register encode the index k∈{0,…,Nv−1}k \in\{0,\ldots,N_v-1\} of a uniform cell-centred velocity grid,

vk=−Lv+(k+12)Δv,Δv=2LvNv.(20)v_k = -L_v + \left(k+\frac{1}{2}\right)\Delta v, \qquad\Delta v = \frac{2L_v}{N_v}. \tag*{(20)}

Because this grid is uniform, the encoded velocity depends affinely on the register index,

vk=c0+c1k,c0=−Lv+12Δv,c1=Δv.(21)v_k = c_0 + c_1 k, \qquad c_0 = -L_v + \frac{1}{2}\Delta v, \qquad c_1 = \Delta v. \tag*{(21)}

It follows that any phase angle proportional to vkv_k can be decomposed into a constant offset and a sum of bitwise contributions from the binary digits of kk.

If we let

k=∑r=0nv−12rbr(k),br(k)∈{0,1},(22)k = \sum_{r=0}^{n_v-1} 2^r b_r(k), \qquad b_r(k) \in\{0,1\}, \tag*{(22)}

where br(k)b_r(k) denotes the rr-th binary digit of kk. The corresponding block angle takes the form

θ(vk)=θoff+θb∑r=0nv−12rbr(k),(23)\theta(v_k) = \theta_{\mathrm{off}} + \theta_b \sum_{r=0}^{n_v-1} 2^r b_r(k), \tag*{(23)}

where

θoff=θxc0,θb=θxc1,θx=ΔtΔx.(24)\theta_{\mathrm{off}} = \theta_x c_0,\qquad\theta_b = \theta_x c_1,\qquad\theta_x = \frac{\Delta t}{\Delta x}. \tag*{(24)}

For Schrödingerisation, this bitwise decomposition is used with two different angle scales [4]. The A1A_1 block, coupled to the auxiliary Fourier variable, uses

θx,1=θxR,(25)\theta_{x,1} = \frac{\theta_x}{R}, \tag*{(25)}

whereas the A2A_2 block uses

θx,2=θx.(26)\theta_{x,2} = \theta_x. \tag*{(26)}

Here RR denotes the finite truncation parameter for the auxiliary Fourier variable introduced by Schrödingerisation. Thus, the coefficient dependence can be implemented explicitly using one unconditional phase rotation for the offset and a collection of controlled phase rotations from the velocity qubits.

Explicit circuit implementation

Following [3], the time evolution operator

exp⁡(iθ(eiλsj−+e−iλsj+)),(27)\exp\left(i\theta\left(e^{i\lambda s_j^-} + e^{-i\lambda s_j^+}\right)\right), \tag*{(27)}

can be implemented using

Wj(θ,λ):=Bj(λ) CRZj1,…,j−1(−2θ) Bj(λ)†,(28)W_j(\theta,\lambda) := B_j(\lambda)\,\mathrm{CRZ}_{j}^{1,\ldots,j-1}(-2\theta)\,B_j(\lambda)^\dagger, \tag*{(28)}
Bj(λ):=(∏m=1j−1CNOTmj)Pj(−λ)Hj,(29)B_j(\lambda) := \left(\prod_{m=1}^{j-1}\mathrm{CNOT}_m^j\right)P_j(-\lambda)H_j, \tag*{(29)}

where λ\lambda is the phase parameter used in the Bell-basis decomposition of the shift operators sj±s_j^\pm acting as finite difference operators (here ∂x\partial_x and ∂v\partial_v). θ\theta is the evolution angle due to the finite-difference coupling, where in Block 1, it is the velocity-dependent angle θ(vk)\theta(v_k) (Eq. (23)).

Here, an additional block UU is required to enforce periodic boundary conditions

exp⁡(iθ(eiλσ01⊗nx+e−iλσ10⊗nx)),(30)\exp\left(i\theta\left(e^{i\lambda\sigma_{01}^{\otimes n_x}} + e^{-i\lambda\sigma_{10}^{\otimes n_x}}\right)\right), \tag*{(30)}

where σ01:=∣0⟩⟨1∣\sigma_{01} := \lvert0\rangle\langle1\rvert and σ10:=∣1⟩⟨0∣\sigma_{10} := \lvert1\rangle\langle0\rvert. The explicit constructions of the WjW_j and UU blocks are shown in Figure 1. The reader is referred to [3, 4] for details.

Figure 1. Generic coefficient-dependent (a) Wj(θ,λ)W_j(\theta,\lambda) block and (b) U(θ,λ)U(\theta,\lambda) boundary block. Adapted from [3, 4].

To incorporate the encoded velocity dependence, the evolution angle θ\theta in each WjW_j block is replaced by the bitwise decomposition in Eq. (23). The resulting velocity-controlled free-streaming block, denoted vWjvW_j, is shown in Figure 2. Its phase labels θ0\theta_0, 20θb2^0\theta_b, 21θb2^1\theta_b, …\ldots represent the elementary controlled-RZR_Z rotations generated by the binary expansion of the velocity index. The left portion of the circuit implements the signed contribution proportional to vkv_k, while the separated portion applies the additional ∣vk∣|v_k|-dependent correction required by the A1A_1 part of the Schrödingerised generator in Eq. (10). The boundary unitary vUvU is modified in the same manner.

Figure 2. Velocity-dependent free-streaming block vWjvW_j. The phase labels indicate the elementary controlled-RZR_Z angles arising from the decomposition of WjW_j. The left portion implements the signed velocity contribution for the centred velocity encoding, while the portion after the gap applies the additional ∣v∣|v|-dependent correction required by the A1A_1 term in Eq. (10).

The free-streaming update is assembled from two composite operators, vV1vV_1 and vV2vV_2, corresponding respectively to the A1A_1 and A2A_2 components of the Schrödingerised generator. The block vV1vV_1 is constructed from the bulk block vWj(λ=0)vW_j(\lambda= 0) together with the boundary block vU(λ=0)vU(\lambda= 0). The companion block vV2vV_2 has the same circuit structure, but uses the phase choices λ=−π/2\lambda= -\pi/2 for the bulk terms and λ=π/2\lambda= \pi/2 for the boundary term. In both cases, the phase rotations θ0\theta_0 and θb\theta_b act on the auxiliary phase register PP and are controlled by the velocity register VV, thereby coupling the lifted auxiliary dynamics to the spatial transport.

Figure 3 displays the circuit structure of vV1vV_1. The dashed region applies the velocity-controlled phase rotations on the auxiliary PP-register, while the remaining controlled vWjvW_j and vUvU blocks realise the bulk and boundary finite-difference couplings in the spatial register. The dotted region highlights the corresponding vV2vV_2 block. Its architecture is identical, differing only in the phase selections vWj(λ=−π/2)vW_j(\lambda= -\pi/2) and vU(λ=π/2)vU(\lambda= \pi/2) required for the A2A_2 contribution.

Circuit diagram for the $vV_1$ block

Figure 3. vV1vV_1 block for A1A_1 term. The dashed region applies the phase rotations θ0\theta_0 and θb\theta_b on the pp-register, controlled by the velocity register, while the remaining controlled vWjvW_j and vUvU blocks implement the bulk and boundary finite-difference couplings. The dotted region indicates the corresponding vV2vV_2 block for the A2A_2 term, which has the same layout but uses phase selections vWj(λ=−π/2)vW_j(\lambda=-\pi/2) and vU(λ=π/2)vU(\lambda=\pi/2).

Block 1 is obtained by combining binary-controlled powers of vV1vV_1 with a single application of vV2vV_2, as shown in Figure 4. For each l=0,1,…,np−1l = 0, 1, \ldots, n_p - 1, the ll-th qubit of the auxiliary phase register PP controls the power vV12lvV_1^{2^l}, which encodes the dependence on the discretised auxiliary variable through its binary expansion. The most significant qubit controls the adjoint branch. As noted in [4], the total number of elementary vV1vV_1 applications scales as ∑l=0np−12l\sum_{l=0}^{n_p-1} 2^l, and therefore grows exponentially with the size of the auxiliary phase register.

Figure 4. Construction of Block 1 from binary-controlled powers of vV1vV_1, followed by a single vV2vV_2 block. Each PP-register qubit controls the corresponding power vV12lvV_1^{2^l}, and the most significant qubit controls the adjoint unitaries.

Although npn_p can be chosen independently of the physical phase-space resolution, its appropriate value depends on the problem, the auxiliary-domain parameter RR, and the desired accuracy. Increasing npn_p could substantially increase circuit resources without necessarily improving solution accuracy (§ 5.2 and § 5.3). The associated trade-offs in auxiliary state preparation and readout are further examined in Appendix § 10.

Field-dependent acceleration circuit (Block 2)

The second split substep is the field-driven acceleration evolution

∂tf+E(x) ∂vf=0.(31)\partial_t f + E(x)\,\partial_v f = 0. \tag*{(31)}

We refer to the quantum implementation of UvU_v as Block 2. As in Block 1, the Schrödingerised construction is based on the discretised transport generator for the split substep, but the advection coefficient is now the position-dependent field E(x)E(x).

For the cubic double-well field in Eq. (4), the required phase angles are not affine in the binary register index, so their implementation requires explicit arithmetic to evaluate the field dependence before applying the controlled rotations. Using the spatial grid defined in Eq. (5), the field value corresponding to the computational basis state ∣j⟩X\lvert j\rangle_X must first be constructed coherently on an auxiliary arithmetic register through the reversible map

∣j⟩X∣0⟩E↦∣j⟩X∣Efp(j)⟩E,Efp(j)S≈E(x~j).(32)\lvert j\rangle_X\lvert0\rangle_E \mapsto\lvert j\rangle_X\lvert E_{\mathrm{fp}}(j)\rangle_E,\qquad\frac{E_{\mathrm{fp}}(j)}{S} \approx E(\widetilde{x}_j). \tag*{(32)}

Schematically, the principal data registers for coherent field evaluation are

∣j⟩X∣0⟩c∣0⟩cabs∣0⟩xcopy∣0⟩tmp∣0⟩sq∣0⟩cube∣0⟩E.(33)\lvert j\rangle_X\lvert0\rangle_c\lvert0\rangle_{c_{\mathrm{abs}}}\lvert0\rangle_{\mathrm{xcopy}}\lvert0\rangle_{\mathrm{tmp}}\lvert0\rangle_{\mathrm{sq}}\lvert0\rangle_{\mathrm{cube}}\lvert0\rangle_E. \tag*{(33)}

Here XX is the position register, cc is the centred signed register derived from the encoded position value, cabsc_{\mathrm{abs}} stores its unsigned magnitude, xcopy, tmp, sq, and cube are work registers used in the polynomial evaluation, and EE stores the fixed-point field value Efp(j)E_{\mathrm{fp}}(j) corresponding to E(x~j)E(\widetilde{x}_j). Additional carry and sign ancillas required by the reversible arithmetic routines are omitted from Eq. (33) for clarity.

A conservative choice of the output-register width nEn_E is obtained by bounding the maximum magnitude of the fixed-point field accumulator. Let αfp\alpha_{\mathrm{fp}} and βfp\beta_{\mathrm{fp}} denote the signed fixed-point integer representations of the real coefficients α\alpha and β\beta. If the centred position register ranges over integers of magnitude at most 2nx−12^{n_x-1}, then the fixed-point field accumulator satisfies

∣Efp∣≤ME,ME:=∣αfp∣2nx−1+∣βfp∣23(nx−1).(34)\lvert E_{\mathrm{fp}}\rvert\le M_E,\qquad M_E := \lvert\alpha_{\mathrm{fp}}\rvert2^{n_x-1}+\lvert\beta_{\mathrm{fp}}\rvert2^{3(n_x-1)}. \tag*{(34)}

A sufficient two’s-complement width for the output register is therefore

nE=max⁡(3nx+1,⌈log⁡2(ME+1)⌉+1),(35)n_E = \max(3n_x+1,\lceil\log_2(M_E+1)\rceil+1), \tag*{(35)}

where the second term provides enough bits to represent any signed integer of magnitude at most MEM_E, and the first term enforces a structural lower bound compatible with the signed cubic workspace.

For the present cubic double-well field, the additional arithmetic required by Block 2 does not introduce an exponential qubit overhead in the grid size.

Lemma 1. For the cubic double-well field E(x)=αx+βx3E(x)=\alpha x+\beta x^3 with fixed-point coefficients of bounded precision, the auxiliary arithmetic workspace of Block 2 scales linearly with nxn_x, and hence as O(log⁡Nx)O(\log N_x) in the spatial grid size Nx=2nxN_x=2^{n_x}.

Proof. The coherent field-evaluation subroutine uses principal registers of widths nxn_x, nxn_x, nxn_x, nxn_x, 2nx2n_x, 4nx4n_x, and nEn_E, corresponding respectively to cc, cabsc_{\mathrm{abs}}, xcopy, tmp, sq, cube, and EE. In addition, the reversible signed-arithmetic routines require carry and sign ancillas of total width O(nx+nE)O(n_x+n_E). Hence the total auxiliary workspace is O(nx+nE)O(n_x+n_E). From Eq. (35), for bounded fixed-point precision one has nE=3nx+O(1)n_E=3n_x+O(1). Therefore the total auxiliary workspace is O(nx)O(n_x), which is O(log⁡Nx)O(\log N_x) since Nx=2nxN_x=2^{n_x}.

The linear workspace scaling is not specific to the cubic field of choice here.

Remark 1. More generally, if E(x)E(x) is a polynomial of fixed degree dd with coefficients represented at bounded fixed-point precision, then coherent evaluation of E(x~j)E(\tilde{x}_j) still requires only O(nx)O(n_x) auxiliary qubits. The number of monomial terms is constant, and the largest monomial c(j)dc(j)^d requires only O(d nx)=O(nx)O(d\,n_x)=O(n_x) bits when dd is independent of nxn_x. Hence the total arithmetic workspace remains linear in nxn_x, or equivalently logarithmic in Nx=2nxN_x=2^{n_x}.

Coherent construction of the auxiliary field register

We coherently prepare a field register whose contents encode the fixed-point field value used in the acceleration step. First, the unsigned grid index jj is converted into the centred signed integer

c(j)=j−2nx−1,(36)c(j)=j-2^{n_x-1}, \tag*{(36)}

stored in two’s-complement form.

The centred integer corresponds to the field-evaluation coordinate

x~j=Δx c(j)=xj−Δx2.(37)\tilde{x}_j=\Delta x\,c(j)=x_j-\frac{\Delta x}{2}. \tag*{(37)}

Hence, the cubic double-well field evaluated by the circuit is

E(x~j)=αΔx c(j)+β(Δx)3c(j)3.(38)E(\tilde{x}_j)=\alpha\Delta x\,c(j)+\beta(\Delta x)^3c(j)^3. \tag*{(38)}

We use a signed fixed-point representation with ss fractional bits and scale factor S:=2sS:=2^s. The real coefficients of c(j)c(j) and c(j)3c(j)^3 are encoded as fixed-point integers by

αfp=round⁡(SαΔx),βfp=round⁡(Sβ(Δx)3).(39)\alpha_{\mathrm{fp}}=\operatorname{round}(S\alpha\Delta x),\qquad\beta_{\mathrm{fp}}=\operatorname{round}(S\beta(\Delta x)^3). \tag*{(39)}

The target integer stored in the output field register is therefore

Efp(j)=αfpc(j)+βfpc(j)3.(40)E_{\mathrm{fp}}(j)=\alpha_{\mathrm{fp}}c(j)+\beta_{\mathrm{fp}}c(j)^3. \tag*{(40)}

Figure 5 summarises the auxiliary field-register construction as high-level reversible data flow into the output field register EfpE_{\mathrm{fp}}. Additional gate-level arithmetic details can be found in Appendix § 13. Here, we outline step-by-step the reversible construction of the EfpE_{\mathrm{fp}} register:

Figure 5. High-level workflow of Block 2. Here, ctr denotes the MSB flip that yields the centred signed register cc, tc denotes controlled two’s complement negation, mul denotes reversible multiplication, sign denotes sign restoration, and acc denotes weighted accumulation into EE. All CNOT gates shown denote bitwise CNOT ladders and the dotted box indicates ladders applied in reverse order for uncomputation. Additional work and ancilla registers are omitted for clarity.

Step 1: Preparation of the centred signed register. Starting from the encoded position register, the spatial index is reversibly mapped to a register storing the centred signed integer c(j)c(j):

∣j⟩X∣0⟩c↦∣j⟩X∣c(j)⟩c.(41)|j\rangle_X|0\rangle_c\mapsto|j\rangle_X|c(j)\rangle_c. \tag*{(41)}

Step 2: Magnitude extraction. The unsigned magnitude register is then prepared from the signed register cc:

∣c⟩c∣0⟩cabs↦∣c⟩c∣∣c∣⟩cabs.(42)|c\rangle_c|0\rangle_{c_{\mathrm{abs}}}\mapsto|c\rangle_c||c|\rangle_{c_{\mathrm{abs}}}. \tag*{(42)}

Step 3: Square and cubic workspaces. Using reversible multiplication, the magnitude register is promoted to quadratic and cubic work registers. First, the square register is constructed as

∣c⟩cabs∣0⟩sq↦∣∣c∣⟩cabs∣∣c∣2⟩sq.(43)\lvert c\rangle_{\mathrm{cabs}}\lvert0\rangle_{\mathrm{sq}} \mapsto\lvert\lvert c\rvert\rangle_{\mathrm{cabs}}\lvert\lvert c\rvert^{2}\rangle_{\mathrm{sq}} . \tag*{(43)}

This intermediate result is then used to form the cubic register,

∣∣c∣⟩cabs∣∣c∣2⟩sq∣0⟩cube↦∣∣c∣⟩cabs∣∣c∣2⟩sq∣∣c∣3⟩cube.(44)\lvert\lvert c\rvert\rangle_{\mathrm{cabs}}\lvert\lvert c\rvert^{2}\rangle_{\mathrm{sq}}\lvert0\rangle_{\mathrm{cube}} \mapsto\lvert\lvert c\rvert\rangle_{\mathrm{cabs}}\lvert\lvert c\rvert^{2}\rangle_{\mathrm{sq}}\lvert\lvert c\rvert^{3}\rangle_{\mathrm{cube}} . \tag*{(44)}

Once the cubic term has been obtained, the square register sq is uncomputed to ∣0⟩\lvert0\rangle.

Step 4: Sign restoration for the cubic term. The cubic register is then conditionally negated according to the sign of cc, yielding a signed representation of c3c^{3}:

∣c⟩c∣∣c∣3⟩cube↦∣c⟩c∣c3⟩cube.(45)\lvert c\rangle_{c}\lvert\lvert c\rvert^{3}\rangle_{\mathrm{cube}} \mapsto\lvert c\rangle_{c}\lvert c^{3}\rangle_{\mathrm{cube}} . \tag*{(45)}

Step 5: Weighted accumulation into the field register. Let αfp\alpha_{\mathrm{fp}} and βfp\beta_{\mathrm{fp}} denote the fixed-point integer representations of the coefficients α\alpha and β\beta. The output register EE is then assembled by reversibly accumulating the weighted linear and cubic contributions,

∣c⟩c∣c3⟩cube∣0⟩E↦∣c⟩c∣c3⟩cube∣αfpc+βfpc3⟩E.(46)\lvert c\rangle_{c}\lvert c^{3}\rangle_{\mathrm{cube}}\lvert0\rangle_{E} \mapsto\lvert c\rangle_{c}\lvert c^{3}\rangle_{\mathrm{cube}}\lvert\alpha_{\mathrm{fp}}c+\beta_{\mathrm{fp}}c^{3}\rangle_{E} . \tag*{(46)}

Thus the field register stores the fixed-point EfpE_{\mathrm{fp}} which represents the discretised field value associated with the basis state ∣j⟩X\lvert j\rangle_{X}.

To form the weighted field value Efp=αfpc+βfpc3E_{\mathrm{fp}}=\alpha_{\mathrm{fp}}c+\beta_{\mathrm{fp}}c^{3}, the circuit must apply the fixed coefficients αfp\alpha_{\mathrm{fp}} and βfp\beta_{\mathrm{fp}} to the signed source registers cc and c3c^{3}. Here, this is done through controlled constant-addition blocks, namely reversible operations that, conditioned on a single source bit, add the corresponding classical shifted constant into the output register EE while leaving the source register unchanged.

Resource estimates for polynomial field accumulation

The number of controlled constant-addition blocks required depends on the widths of the signed source registers.

Lemma 2 (Controlled-addition count for polynomial field accumulation). Suppose the field is written as

E(x)=∑m∈Samxm,(47)E(x)=\sum_{m\in S}a_mx^m, \tag*{(47)}

where S⊂NS\subset\mathbb{N} is a fixed finite set of nonzero polynomial exponents. Assume that, for each m∈Sm\in S, the signed value xmx^m is stored in a register of width mnxmn_x, and that the contribution amxma_mx^m is accumulated into the output register using the two’s-complement bits of this source register. Then the number of controlled constant-addition blocks used by this accumulation step is

nx∑m∈Sm.(48)n_x\sum_{m\in S}m. \tag*{(48)}

In particular, if SS is independent of nxn_x, then the number of controlled constant-addition blocks scales as O(nx)O(n_x).

Proof. For a given exponent m∈Sm\in S, the quantity xmx^m is stored in an mnxmn_x-bit signed source register. In a bitwise two’s complement accumulation, one controlled constant-addition block is used for each bit of the source register. The mnx−1mn_x-1 non-sign bits each contribute a positively weighted shifted constant, while the sign bit contributes the corresponding negatively weighted shifted constant. Therefore the contribution amxma_mx^m requires

(mnx−1)+1=mnx(49)(mn_x-1)+1=mn_x \tag*{(49)}

controlled constant-addition blocks. Summing over all nonzero polynomial terms gives

∑m∈Smnx=nx∑m∈Sm.(50)\sum_{m\in S}mn_x=n_x\sum_{m\in S}m. \tag*{(50)}

If SS is fixed independently of nxn_x, then ∑m∈Sm\sum_{m\in S}m is a constant, so the accumulation cost is O(nx)O(n_x). □

For the cubic double-well field E(x)=αx+βx3E(x)=\alpha x+\beta x^3, we have S={1,3}S=\{1,3\}. Hence the weighted accumulation step uses nx(1+3)=4nxn_x(1+3)=4n_x controlled constant-addition blocks.

We next estimate the depth cost of the auxiliary field-register construction used in Block 2.

Lemma 3 (Depth estimate for the auxiliary field-register construction). For the cubic auxiliary field E(x)=αx+βx3E(x)=\alpha x+\beta x^3, let Dmul(m)D_{\mathrm{mul}}(m) denote the depth of the reversible multiplier used on mm-bit inputs. Assume that each controlled constant-addition block into the nEn_E-qubit output register has depth O(nE)O(n_E). Then the depth DED_E of the coherent field-register construction satisfies

DE≤2Dmul(nx)+Dmul(2nx)+O(nxnE).(51)D_E\leq2D_{\mathrm{mul}}(n_x)+D_{\mathrm{mul}}(2n_x)+O(n_xn_E). \tag*{(51)}

In particular, since nE=3nx+O(1)n_E = 3n_x + O(1), this gives

DE≤2Dmul(nx)+Dmul(2nx)+O(nx2).(52)D_E \le2D_{\mathrm{mul}}(n_x) + D_{\mathrm{mul}}(2n_x) + O(n_x^2). \tag*{(52)}

Hence the auxiliary field-register construction has polynomial depth whenever the chosen reversible multiplier family has polynomial depth.

Proof. The preprocessing and sign-handling steps have depth O(nx)O(n_x), which is absorbed into the O(nxnE)O(n_xn_E) accumulation term. The cubic contribution is formed by computing ∣c∣2|c|^2, then computing ∣c∣3|c|^3, and finally uncomputing the intermediate square register. This gives the multiplier-depth contribution

2Dmul(nx)+Dmul(2nx).(53)2D_{\mathrm{mul}}(n_x) + D_{\mathrm{mul}}(2n_x). \tag*{(53)}

Here the second multiplication is bounded by Dmul(2nx)D_{\mathrm{mul}}(2n_x) because it multiplies the nxn_x-bit magnitude ∣c∣|c| by the 2nx2n_x-bit square ∣c∣2|c|^2.

The accumulation stage adds αfpc+βfpc3\alpha_{\mathrm{fp}}c + \beta_{\mathrm{fp}}c^3 into EE by scanning the two’s-complement bits of the signed source registers. Since the linear source register has width nxn_x and the cubic source register has width 3nx3n_x, Lemma 2 gives 4nx4n_x controlled constant-addition blocks. Each such block has depth O(nE)O(n_E), so the total accumulation depth is O(nxnE)O(n_xn_E).

Field-controlled acceleration block

We now assemble the Schrödingerised circuit for the field-driven acceleration step Eq. (31) as an analogue of the velocity-controlled free-streaming block in § 2.3, except that the finite difference now acts on the velocity register VV, and the advection coefficient is the coherently computed field value E(x~j)E(\tilde{x}_j) rather than the encoded velocity vkv_k.

For the field-controlled rotations, the coefficient has already been computed into the nEn_E-qubit field register. The field-dependent angle is

θE(j)≈θsEfp(j),θs=θvS,θv=ΔtΔv.(54)\theta_E(j) \approx\theta_s E_{\mathrm{fp}}(j), \qquad\theta_s = \frac{\theta_v}{S}, \qquad\theta_v = \frac{\Delta t}{\Delta v}. \tag*{(54)}

Substituting the two’s-complement expansion gives

θE(j)=θs(∑r=0nE−22rer(j)−2nE−1enE−1(j)),er∈{0,1},(55)\theta_E(j) = \theta_s \left( \sum_{r=0}^{n_E-2} 2^r e_r(j) - 2^{n_E-1}e_{n_E-1}(j) \right), \qquad e_r \in\{0,1\}, \tag*{(55)}

where enE−1(j)e_{n_E-1}(j) is the sign bit. Note that unlike Eq. (24) in Block 1, the field-register decomposition has no affine offset term, i.e. θoff(E)=0\theta_{\mathrm{off}}^{(E)} = 0. Similarly, the A1A_1 block, coupled to the auxiliary Fourier variable, uses

θv,1=θvR,(56)\theta_{v,1} = \frac{\theta_v}{R}, \tag*{(56)}

whereas the A2A_2 block uses

θv,2=θv.(57)\theta_{v,2} = \theta_v. \tag*{(57)}

A corresponding field-dependent acceleration block eWjeW_j is shown in Figure 6. This is the direct analogue of the velocity-dependent free-streaming block vWjvW_j in Figure 2, with the controlled phases now driven by the bits of the computed field register EfpE_{\mathrm{fp}}.

Figure 6. Field-dependent acceleration block eWjeW_j. The phase labels indicate the elementary controlled-RzR_z angles arising from the two’s-complement decomposition of the computed field register EfpE_{\mathrm{fp}}. The left portion implements the signed field contribution, while the portion after the gap applies the additional ∣E∣|E|-dependent correction required by the A1A_1 term in Eq. (10). For the A1A_1 block the scale is θs=θv/(RS)\theta_s=\theta_v/(RS); for the A2A_2 block only the signed left portion is used, with θs=θv/S\theta_s=\theta_v/S.

Similarly, the field-controlled acceleration blocks are implemented in direct analogy with the velocity-controlled blocks in Figure 3–Figure 4. After the field register has been prepared, vWjvW_j and vUvU are replaced by their field-controlled counterparts eWjeW_j and eUeU, acting on the velocity register VV with rotation controls supplied by the bits of EfpE_{\mathrm{fp}}. The A1A_1 block eV1eV_1 uses the scale θv/(RS)\theta_v/(RS), while the A2A_2 block eV2eV_2 uses the scale θv/S\theta_v/S, where θv=Δt/Δv\theta_v=\Delta t/\Delta v.

Thus the full Block 2 update has the compute–apply–uncompute form

Uv(Δt)=UE†Uv(E)(Δt)UE,(58)U_v(\Delta t)=U_E^\dagger U_v^{(E)}(\Delta t)U_E, \tag*{(58)}

where UEU_E is the coherent field-evaluation circuit and Uv(E)(Δt)U_v^{(E)}(\Delta t) is assembled from binary-controlled powers of eV1eV_1, followed by a single eV2eV_2, exactly as in Block 1.

For multiple time steps, the one-step Strang operator in Eq. (18) is applied repeatedly. Since adjacent half steps in UxU_x combine into a full step, the NtN_t-step evolution can be written as

U(NtΔt)≈Ux(Δt2)[Uv(Δt)Ux(Δt)]Nt−1Uv(Δt)Ux(Δt2),(59)U(N_t\Delta t)\approx U_x\left(\frac{\Delta t}{2}\right)\left[U_v(\Delta t)U_x(\Delta t)\right]^{N_t-1}U_v(\Delta t)U_x\left(\frac{\Delta t}{2}\right), \tag*{(59)}

where

Ux(Δt)=Ux(Δt2)Ux(Δt2).(60)U_x(\Delta t)=U_x\left(\frac{\Delta t}{2}\right)U_x\left(\frac{\Delta t}{2}\right). \tag*{(60)}

Let DxD_x and DvD_v denote the circuit depths of UxU_x and UvU_v, respectively. Since Ux(Δt/2)U_x(\Delta t/2) and Ux(Δt)U_x(\Delta t) have identical circuit depth, the depth per time step is

DNtNt=Dv+(1+1Nt)Dx.(61)\frac{D_{N_t}}{N_t}=D_v+\left(1+\frac{1}{N_t}\right)D_x. \tag*{(61)}

In particular,

DNtNt⟶Dv+DxasNt→∞.(62)\frac{D_{N_t}}{N_t}\longrightarrow D_v+D_x \qquad\text{as}\qquad N_t\to\infty. \tag*{(62)}

Thus the two half steps contribute only an O(Dx/Nt)O(D_x/N_t) overhead per time step.

Measurement and readout

After applying the Schrödingerised evolution (Eq. (59)) in the auxiliary Fourier basis, we apply the inverse of the initial Fourier-basis transformation to the auxiliary register PP to return to the pp-grid. The physical solution is recovered by measuring PP and selecting an appropriate non-negative pp-slice [4, 16]. For the original auxiliary profile used here, we select the central p=0p=0 slice.

Let the post-IQFT state at time T=NtΔtT=N_t\Delta t be

∣Ψ(T)⟩=∑j,k,lψjkl(T)∣j⟩X∣k⟩V∣l⟩P(63)\lvert\Psi(T)\rangle= \sum_{j,k,l} \psi_{jkl}(T)\lvert j\rangle_X \lvert k\rangle_V \lvert l\rangle_P \tag*{(63)}

where plp_l denotes the auxiliary grid defined in Eq. (11). The probability of measuring P=lP=l is

Pl=∑j,k∣ψjkl(T)∣2(64)P_l = \sum_{j,k} \lvert\psi_{jkl}(T)\rvert^2 \tag*{(64)}

and the corresponding normalised physical state is

∣ψl(T)⟩XV=1Pl∑j,kψjkl(T)∣j⟩X∣k⟩V.(65)\lvert\psi_l(T)\rangle_{XV} = \frac{1}{\sqrt{P_l}}\sum_{j,k}\psi_{jkl}(T)\lvert j\rangle_X\lvert k\rangle_V. \tag*{(65)}

For pl>0p_l>0, the warped representation gives

Φ(T,xj,vk,pl)=e−plf(T,xj,vk).(66)\Phi(T,x_j,v_k,p_l)=e^{-p_l f(T,x_j,v_k)}. \tag*{(66)}

Thus the selected plp_l-slice is proportional to the physical solution. Since the factor e−ple^{-p_l} is independent of jj and kk, it cancels after normalisation, giving

∣ψl(T)⟩XV≈1∥f(T)∥∑j,kf(T,xj,vk)∣j⟩X∣k⟩V.(67)\lvert\psi_l(T)\rangle_{XV} \approx\frac{1}{\lVert f(T)\rVert}\sum_{j,k}f(T,x_j,v_k)\lvert j\rangle_X\lvert k\rangle_V. \tag*{(67)}

Reconstructing every amplitude of the full NxNvN_xN_v-dimensional solution would require tomography and is not assumed to be efficient. Alternative smooth auxiliary profiles, including the Hermite-spliced and error-function profiles, are discussed in Appendix § 10.

Memory-efficient readout for classical circuit simulation

For classical circuit validation, the quantum circuits are simulated using Qiskit Aer [17] with its matrix-product-state (MPS) simulation method. The MPS representation avoids storing the full statevector explicitly which enables emulation of large circuits, although its computational cost still depends on the bond dimensions generated during the evolution. In this study, the MPS simulator is configured with a maximum bond dimension of χcap=512\chi_{\mathrm{cap}}=512 and a truncation threshold of 10−1210^{-12}, using two CPU threads. These settings are fixed for a consistent assessment of the classical emulation cost.

Directly constructing the reduced density matrix on the full XX, VV, PP subsystem requires 4nx+nv+np4^{n_x+n_v+n_p} complex entries. To reduce this memory cost, we append a clean flag qubit FF after the Schrödingerised evolution and reversibly set

F=1  ⟺  P=qrec,F=1 \;\Longleftrightarrow\; P=q_{\mathrm{rec}},

where qrec=Np/2q_{\mathrm{rec}}=N_p/2 corresponds to p=0p=0 for the original auxiliary profile.

The simulator then constructs only the reduced density matrix

ρXVF=Tr⁡P,W(∣ΨF⟩⟨ΨF∣),(68)\rho_{XVF} = \operatorname{Tr}_{P,W}\left(\lvert\Psi_F\rangle\langle\Psi_F\rvert\right), \tag*{(68)}

where WW denotes all field-arithmetic and workspace registers. The unnormalised postselected state is obtained from the flag-1 block,

(qrec)ρXV=F⟨1∣ρXVF∣1⟩F,(69){}^{(q_{\mathrm{rec}})}\rho_{XV} = {}_{F}\langle1\lvert\rho_{XVF}\rvert1\rangle_{F}, \tag*{(69)}

and its trace gives the postselection probability. After normalisation, the principal eigenvector is used as the recovered X,VX,V amplitude vector when the conditional state is numerically pure. This reduces the dense readout storage to 4nx+nv+14^{n_x+n_v+1}, a reduction factor of 4np−14^{n_p-1}. On quantum hardware, the PP-register may of course be measured directly with the X,VX,V outcomes postselected on qrecq_{\mathrm{rec}}.

High-dimensional test case: 3D3V Vlasov equation

To test whether the explicit circuit architecture remains viable beyond the 1D1V Vlasov equation, we construct a low-resolution, three-dimensional in space and three-dimensional in velocity (3D3V),

f=f(x,y,z,vx,vy,vz,t),(70)f = f(x,y,z,v_x,v_y,v_z,t), \tag*{(70)}

defined on a six-dimensional phase space. A direct tensor product grid with NN points in each coordinate contains N6N^6 phase-space values, whereas the corresponding amplitude encoding requires 6log⁡2N6\log_2 N physical qubits, in addition to the auxiliary and arithmetic registers.

Anharmonic Penning trap

A Penning trap confines charged particles by combining a static magnetic field with an electrostatic trapping potential. In the idealised configuration, a uniform magnetic field provides transverse magnetic confinement, while an electrostatic quadrupole potential provides confinement along the axial direction.

We consider the prescribed-field kinetic equation

∂tf+vx∂xf+vy∂yf+vz∂zf+(Ex+ωcvy)∂vxf+(Ey−ωcvx)∂vyf+Ez∂vzf=0,(71)\partial_t f + v_x\partial_x f + v_y\partial_y f + v_z\partial_z f + (E_x + \omega_c v_y)\partial_{v_x}f + (E_y - \omega_c v_x)\partial_{v_y}f + E_z\partial_{v_z}f = 0, \tag*{(71)}

where ωc\omega_c is a prescribed cyclotron-frequency parameter associated with a uniform magnetic field in the zz direction. The electrostatic field is obtained from the quadrupole–quartic potential

Vtrap(x,y,z)=A2(z2−x2+y22)+A4[z4−3z2(x2+y2)+38(x2+y2)2].(72)V_{\mathrm{trap}}(x,y,z) = A_2\left(z^2 - \frac{x^2+y^2}{2}\right) + A_4\left[z^4 - 3z^2(x^2+y^2) + \frac{3}{8}(x^2+y^2)^2\right]. \tag*{(72)}

The corresponding field components are

Ex=A2x+A4(6xz2−32x3−32xy2),(73)E_x = A_2x + A_4\left(6xz^2 - \frac{3}{2}x^3 - \frac{3}{2}xy^2\right), \tag*{(73)}
Ey=A2y+A4(6yz2−32y3−32x2y),(74)E_y = A_2y + A_4\left(6yz^2 - \frac{3}{2}y^3 - \frac{3}{2}x^2y\right), \tag*{(74)}
Ez=−2A2z+A4(6x2z+6y2z−4z3).(75)E_z = -2A_2z + A_4\left(6x^2z + 6y^2z - 4z^3\right). \tag*{(75)}

Note that the mixed monomials {xz2,xy2,yz2,x2y,x2z,y2z}\{xz^2,xy^2,yz^2,x^2y,x^2z,y^2z\} couple the three spatial coordinates, so the 3D3V problem cannot be factorised into three independent 1D1V problems.

Circuit compilation of the coupled electric field

In the preceding 1D1V Vlasov setup (§ 2.5), the cubic field was evaluated by first constructing signed coordinate and polynomial work registers. For the present 3D3V Vlasov proof of concept, we compile the known polynomial field directly from the binary digits of the spatial registers.

Let br(μ)∈{0,1}b_r^{(\mu)} \in\{0,1\} denote bit rr of the register representing the spatial coordinate μ∈{x,y,z}\mu\in\{x,y,z\}. On a uniform cell-centred grid, the coordinate associated with a computational-basis state is

μ=−Lμ2+Δμ2+Δμ∑r=0nμ−12rbr(μ).(76)\mu= -\frac{L_\mu}{2} + \frac{\Delta\mu}{2} + \Delta\mu\sum_{r=0}^{n_\mu-1} 2^r b_r^{(\mu)} . \tag*{(76)}

Substituting Eq. (76) into the analytic expressions for ExE_x, EyE_y, and EzE_z (Eqs. (73)–(75)) converts the spatial polynomials into polynomials of the register bits. For example, the term xz2xz^2 becomes a sum of constant, linear, quadratic, and cubic products involving bits from the XX and ZZ registers. Since the bits are in {0,1}\{0,1\}, repeated powers satisfy (br(μ))m=br(μ)(b_r^{(\mu)})^m = b_r^{(\mu)} on computational basis. Hence, a cubic spatial field produces Boolean monomials of degree at most three.

Collecting the bits of the three spatial registers into a single sequence b1,…,bnsb_1,\ldots,b_{n_s}, where ns=nx+ny+nzn_s = n_x+n_y+n_z, each field component can therefore be written as

Eμ=cμ,0+∑icμ,ibi+∑i<jcμ,ijbibj+∑i<j<kcμ,ijkbibjbk,μ∈{x,y,z}.(77)E_\mu= c_{\mu,0} + \sum_i c_{\mu,i}b_i + \sum_{i<j} c_{\mu,ij}b_i b_j + \sum_{i<j<k} c_{\mu,ijk}b_i b_j b_k,\qquad\mu\in\{x,y,z\}. \tag*{(77)}

The coefficients in Eq. (77) are calculated only once during circuit compilation and no classical field evaluation is performed during the quantum time evolution.

The coefficients cc are represented in signed fixed-point (fp) form with ss fractional bits,

cfp=round⁡(2sc).(78)c^{\mathrm{fp}} = \operatorname{round}(2^s c). \tag*{(78)}

For each field component, these fixed-point contributions are accumulated into a signed field register EE.

To implement a linear term cibic_i b_i, the spatial qubit bib_i is used directly as a control. When bi=1b_i=1, a temporary register WW is prepared in the computational basis representing the fixed-point coefficient cifpc_i^{\mathrm{fp}}. The register WW is then reversibly added into the field register EE and subsequently returned to ∣0⟩|0\rangle. For a quadratic or cubic term, the product of the required spatial bits is first computed into a single reusable Boolean ancilla. Schematically,

compute monomial⟶add fixed-point coefficient to E⟶uncompute monomial.(79)\text{compute monomial} \longrightarrow\text{add fixed-point coefficient to } E \longrightarrow\text{uncompute monomial}. \tag*{(79)}

In this way, the mixed-coordinate terms of the Penning field can be represented without storing a separate numerical register for xz2xz^2, xy2xy^2, or other coordinate products.

Finally, the field register contains the fixed-point value Eμfp(X,Y,Z)E_\mu^{\mathrm{fp}}(X,Y,Z). Then, the three electric components are applied sequentially via

UEx†Uvx(Ex)UEx,UEy†Uvy(Ey)UEy,UEz†Uvz(Ez)UEz,(80)U_{E_x}^{\dagger} U_{v_x}^{(E_x)} U_{E_x},\qquad U_{E_y}^{\dagger} U_{v_y}^{(E_y)} U_{E_y},\qquad U_{E_z}^{\dagger} U_{v_z}^{(E_z)} U_{E_z}, \tag*{(80)}

where UEμU_{E_\mu} computes the corresponding field component and Uvμ(Eμ)U_{v_\mu}^{(E_\mu)} denotes the field-controlled velocity update. The field components ExE_x, EyE_y, and EzE_z are computed, applied to the corresponding velocity update, and uncomputed sequentially, allowing the same field and workspace registers to be reused for all three components.

Low-resolution 3D3V circuit

For the proof of concept, we allocate two qubits to each physical coordinate,

nx=ny=nz=nvx=nvy=nvz=2,(81)n_x = n_y = n_z = n_{v_x} = n_{v_y} = n_{v_z} = 2, \tag*{(81)}

corresponding to four grid points per coordinate and Ngrid=46=4096N_{\mathrm{grid}} = 4^6 = 4096 points in the discretised six-dimensional phase space. The spatial domains have lengths Lx=Ly=Lz=2L_x = L_y = L_z = 2, while the velocity domains have lengths Lvx=Lvy=Lvz=4L_{v_x} = L_{v_y} = L_{v_z} = 4. The prescribed field parameters are A2=0.45A_2 = 0.45, A4=0.025A_4 = 0.025, and ωc=2\omega_c = 2.

For this low-resolution case with s=7s = 7 fixed-point precision, the binary field construction produces 20 non-constant terms for each of ExE_x, EyE_y, and EzE_z. These are linear, quadratic, or cubic products of the spatial register bits. A second 10-qubit register WW is used to prepare the fixed-point constants required by the reversible additions. Since the three field components are computed and uncomputed sequentially, this workspace is shared between ExE_x, EyE_y, and EzE_z.

For baseline calculations, we allocate np=5n_p = 5 qubits to the auxiliary register with R=6R = 6, giving 2np=322^{n_p} = 32 points in the auxiliary coordinate. Altogether, the evolution circuit contains 39 algorithmic logical qubits, of which 12 encode the physical 3D3V phase space.

The initial condition is chosen as a product of 1D Gaussian distributions in the six phase space coordinates,

f0(x,y,z,vx,vy,vz)=Z∏μ∈{x,y,z,vx,vy,vz}exp⁡[−(μ−μ0)22σμ2],(82)f_0(x,y,z,v_x,v_y,v_z) = \mathcal{Z} \prod_{\mu\in\{x,y,z,v_x,v_y,v_z\}} \exp\left[-\frac{(\mu-\mu_0)^2}{2\sigma_\mu^2}\right], \tag*{(82)}

where Z\mathcal{Z} is the discrete normalisation constant. The centres and widths are, in (x,y,z,vx,vy,vz)(x,y,z,v_x,v_y,v_z) order,

μ0=(0.20,0,0.10,0,0.40,0),σ=(0.35,0.35,0.40,0.90,0.90,0.90).(83)\boldsymbol{\mu}_0 = (0.20,0,0.10,0,0.40,0), \qquad\boldsymbol{\sigma} = (0.35,0.35,0.40,0.90,0.90,0.90). \tag*{(83)}

Because the initial distribution is a product of six 1D Gaussians, each register can be initialised separately, thus avoiding the need to amplitude-load a full 6D oracle.

The 3D3V evolution contains three types of transport. The spatial streaming terms vx∂xv_x\partial_x, vy∂yv_y\partial_y, and vz∂zv_z\partial_z are implemented using the velocity-controlled construction of § 2.3. The electric-field terms are implemented sequentially using the compute–apply–uncompute procedure of Eq. (80).

The magnetic contribution couples the two transverse velocity coordinates,

LBf=ωcvx∂vyf−ωcvy∂vxf.(84)\mathcal{L}_B f = \omega_c v_x \partial_{v_y} f - \omega_c v_y \partial_{v_x} f. \tag*{(84)}

We implement this contribution using the symmetric splitting

UB(τ)≈Uvx←vy(τ2)Uvy←vx(τ)Uvx←vy(τ2),(85)U_B(\tau) \approx U_{v_x\leftarrow v_y}\left(\frac{\tau}{2}\right) U_{v_y\leftarrow v_x}(\tau) U_{v_x\leftarrow v_y}\left(\frac{\tau}{2}\right), \tag*{(85)}

where the arrow indicates which velocity coordinate is updated and which coordinate supplies the coefficient. The outer updates are proportional to +ωcvy+\omega_c v_y, while the central update is proportional to −ωcvx-\omega_c v_x.

Accordingly, the Strang-split time step is

U3D3V(Δt)≈UK(Δt2)UB(Δt2)UE(Δt)UB(Δt2)UK(Δt2),(86)U_{\mathrm{3D3V}}(\Delta t) \approx U_K\left(\frac{\Delta t}{2}\right) U_B\left(\frac{\Delta t}{2}\right) U_E(\Delta t) U_B\left(\frac{\Delta t}{2}\right) U_K\left(\frac{\Delta t}{2}\right), \tag*{(86)}

where UKU_K, UEU_E, and UBU_B denote the configuration-space streaming, electric-field, and magnetic updates, respectively.

Results

We evaluate both logical circuit resources and the practical limits of matrix-product-state (MPS) emulation. While qubit counts, circuit depth, and gate counts characterise the quantum implementation, MPS runtime and bond dimension also depend on the entanglement generated during evolution.

1D1V prescribed-field circuit validation

We first validate the complete 1D1V Schrödingerised circuit developed in § 2.1 against classical computation. The spatial and velocity registers each contain five qubits, nx=nv=5n_x = n_v = 5, corresponding to Nx=Nv=32N_x = N_v = 32 grid points, while the auxiliary Schrödingerisation register contains np=7n_p = 7 qubits. We take Lx=Lv=6L_x = L_v = 6, R=6R = 6, and use the original auxiliary profile fp(p)=e−∣p∣f_p(p) = e^{-|p|} for which the physical solution is recovered from the central p=0p = 0 slice (§ 3). The initial distribution is sampled on the cell-centred grid, while the prescribed field is evaluated using the centred-integer coordinates defined in § 2.6. The field coefficients are represented using the same signed fixed-point convention in both the quantum circuit and the classical reference.

For this benchmark, we set a0=κ=1a_0 = \kappa= 1 and use fixed-point precision s=6s = 6, giving αfp=12\alpha_{\mathrm{fp}} = 12 and βfp=0\beta_{\mathrm{fp}} = 0. Thus, coefficient rounding suppresses the cubic contribution in this particular instance, although the circuit retains the general cubic-field arithmetic. The reduced test in Appendix § 14 additionally validates a non-zero cubic coefficient.

The initial phase-space distribution is chosen as

f0(x,v)=Z−1exp⁡[−β0Vdw(x)]exp⁡(−v22σ2),(87)f_0(x,v) = \mathcal{Z}^{-1}\exp[-\beta_0 V_{\mathrm{dw}}(x)]\exp\left(-\frac{v^2}{2\sigma^2}\right), \tag*{(87)}

where Z\mathcal{Z} denotes the discrete normalisation constant, with β0=2\beta_0 = 2 and σ=1\sigma= 1. Together with Lx=Lv=6L_x = L_v = 6, the initial distribution is well within the truncated computational domain.

Figure 7 compares the initial distribution with the quantum result after one Strang step using Nt=1N_t = 1, Δt=0.05\Delta t = 0.05, and T=0.05T = 0.05. The recovered quantum distribution retains the two-lobe phase-space structure of the initial state slightly shifted by the free-streaming and field-acceleration blocks.

Phase-space validation of the 1D1V Schrödingerised Vlasov circuit

Figure 7. Phase-space validation of the 1D1V Schrödingerised Vlasov circuit for nx=nv=5n_x = n_v = 5, np=7n_p = 7, Nt=1N_t = 1, Δt=0.05\Delta t = 0.05, and T=0.05T = 0.05. (a) Initial distribution f0(x,vx)f_0(x, v_x). (b) Recovered quantum distribution fQ(x,vx,T)f_Q(x, v_x, T). (c) Pointwise difference between the quantum and scheme-matched classical solutions, fQ−fCf_Q - f_C. The difference is of order 10−510^{-5}, compared with distribution amplitudes of order 10−210^{-2}.

Agreement with the classical reference is shown directly in Figure 7(c). The pointwise difference (fQ−fC)∼O(10−5)(f_Q - f_C) \sim O(10^{-5}) is concentrated at hotspots and less than phase-space densities of order O(10−2)O(10^{-2}). For this run, the relative L2L_2 discrepancy between the recovered quantum solution and the matching classical solution is ϵQ−C=4.9×10−3\epsilon_{Q-C} = 4.9 \times10^{-3}. The residual difference contains the accumulated effects of the Schrödingerised circuit approximation, finite auxiliary register, fixed-point coefficient representation, and classical tensor-network emulation.

Next, we project onto the spatial and velocity coordinates and define the respective discrete marginals

ρ(xj,t)=∑kf(xj,vk,t),(88)\rho(x_j,t) = \sum_k f(x_j,v_k,t), \tag*{(88)}
g(vk,t)=∑jf(xj,vk,t).(89)g(v_k,t) = \sum_j f(x_j,v_k,t). \tag*{(89)}

Figure 8 compares these quantities for the initial distribution, the classical solution, and the quantum result at T=0.05T = 0.05. The spatial marginal preserves the two maxima associated with the two localised populations, while the velocity marginal remains approximately symmetric about v=0v = 0. The near overlap of the t=0t = 0 and T=0.05T = 0.05 curves is consistent with the short simulated time and confirms that the small quantum–classical difference in Figure 7(c) is not produced by a qualitatively incorrect redistribution of the phase-space density.

Marginal distributions for the 1D1V validation case

Figure 8. Marginal distributions for the 1D1V validation case at T=0.05T = 0.05. (a) Spatial marginal ρ(x)=∑vxf(x,vx)\rho(x) = \sum_{v_x} f(x, v_x). (b) Velocity marginal g(vx)=∑xf(x,vx)g(v_x) = \sum_x f(x, v_x). The initial distribution (t=0t = 0, dashed), scheme-matched classical solution (solid), and recovered quantum solution (markers) are shown.

Together, Figure 7 and Figure 8 provide an end-to-end validation of the Schrödingerisation workflow, following amplitude preparation, auxiliary Fourier transformation, velocity-dependent free streaming, coherently field-controlled acceleration, uncomputation of the arithmetic workspace, and recovery of the physical pp-slice.

Resource scaling and temporal behaviour

Here, we estimate the circuit resources required by the explicit 1D1V construction and the behaviour of the quantum–classical discrepancy. Figure 9(a) shows the peak number of logical qubits required by the complete circuit as the spatial and velocity resolutions are increased together, nx=nv=2,…,5n_x = n_v = 2,\ldots,5, with the auxiliary register size fixed at np=7n_p = 7. The corresponding physical grids range from Nx=Nv=4N_x = N_v = 4 to Nx=Nv=32N_x = N_v = 32. The total circuit width increases from approximately 57 to 106 qubits over this range. For fixed polynomial degree and auxiliary precision, the total qubit requirement grows approximately linearly with nx+nvn_x+n_v, and hence only logarithmically with the number of phase-space grid points in each coordinate (Lemma 1 and Remark 1).

Resource scaling and temporal behaviour plots

Figure 9. Resource scaling and temporal behaviour of the 1D1V Schrödingerised Vlasov circuit. (a) Peak logical-qubit count nqn_q as the spatial and velocity grid resolutions are increased together from nx=nv=2n_x=n_v=2 to 5, with np=7n_p=7 fixed. The horizontal axis shows the corresponding combined number of physical grid points, Nx+NvN_x+N_v. (b) Decomposed CNOT count NCXN_{\mathrm{CX}} and rotation-like gate count NrotN_{\mathrm{rot}} as functions of the number of Strang steps NtN_t for (nx,nv)=(3,3)(n_x,n_v)=(3,3), (4,4)(4,4), and (5,5)(5,5), with np=7n_p=7. Both resource measures grow approximately linearly with NtN_t. (c) Relative quantum–classical L2L_2 discrepancy ϵQ−C\epsilon_{Q-C} versus total simulated time T=NtΔtT=N_t\Delta t for the (nx,nv,np)=(4,4,7)(n_x,n_v,n_p)=(4,4,7) circuit using Δt=0.005\Delta t=0.005, 0.01, and 0.05. Results obtained with different time step sizes are close at common values of TT, while the discrepancy increases with accumulated evolution time.

The gate-count depends on the number of time steps (Figure 9(b)). We report the number of CNOT gates, NCXN_{\mathrm{CX}}, together with the rotation-like gate count Nrot=Nu+Nphase+NRz+NCPN_{\mathrm{rot}} = N_u + N_{\mathrm{phase}} + N_{R_z} + N_{\mathrm{CP}} obtained after decomposing the composite circuit blocks, where uu, phase, RzR_z, and CPCP denote the corresponding parameterised Qiskit gates (NphaseN_{\mathrm{phase}} counts the number of pp gates). The quantity NrotN_{\mathrm{rot}} is intended as a proxy for non-Clifford workload in a fault-tolerant implementation and should therefore be interpreted as logical circuit resources.

For each of the (nx,nv)(n_x,n_v) cases, we see that gate counts increase approximately linearly with NtN_t. This agrees with Eq. (59), as each additional time step adds one free-streaming block (§ 2.3) and one field-driven block (§ 2.5), excluding the two half steps at the beginning and end of the evolution. Increasing the phase-space resolution only shifts the curves upward because the velocity-dependent transport and the field subroutines acting on the ancilla registers do not scale with the number of time steps.

Finally, Figure 9(c) shows the relative quantum–classical discrepancy

ϵQ−C=∥fQ∥fQ∥2−fC∥fC∥2∥2.(90)\epsilon_{Q-C} = \left\lVert\frac{f_Q}{\lVert f_Q\rVert_2} - \frac{f_C}{\lVert f_C\rVert_2} \right\rVert_2 . \tag*{(90)}

for the (nx,nv,np)=(4,4,7)(n_x,n_v,n_p) = (4,4,7) circuit as a function of the total simulated time T=NtΔtT = N_t\Delta t. The discrepancy increases monotonically with TT, from 10−410^{-4} to several 10−210^{-2} by T=0.25T = 0.25. Within the range tested, the total quantum–classical discrepancy depends more strongly on the total evolution time itself than on the particular Strang decomposition into multiple time steps.

Note that ϵQ−C\epsilon_{Q-C} measures agreement between the explicit Schrödingerised circuit and its classical counterpart, rather than the total discretisation error with respect to the continuum Vlasov equation.

3D3V proof of concept validation

In this section, we validate the low-resolution 3D3V circuit instance defined in § 4.2 against classical calculation. Since the full 3D3V solution depends on six phase-space coordinates, we visualise the result through two-dimensional pair marginals obtained by summing over the remaining four coordinates. We consider

fzvz(z,vz,t)=∑x,y,vx,vyf(x,y,z,vx,vy,vz,t),(91)f_{zv_z}(z,v_z,t) = \sum_{x,y,v_x,v_y} f(x,y,z,v_x,v_y,v_z,t), \tag*{(91)}
fvxvy(vx,vy,t)=∑x,y,z,vzf(x,y,z,vx,vy,vz,t).(92)f_{v_xv_y}(v_x,v_y,t) = \sum_{x,y,z,v_z} f(x,y,z,v_x,v_y,v_z,t). \tag*{(92)}

The first retains the axial phase-space coordinates (z,vz)(z,v_z), for which the velocity evolution is driven by the electric field, while the second retains the two transverse velocity coordinates (vx,vy)(v_x,v_y) which are related to transverse magnetic coupling.

Figure 10 shows these marginals for np=5n_p = 5, Nt=2N_t = 2, Δt=0.05\Delta t = 0.05, and T=0.1T = 0.1. In both cases, the quantum result qualitatively reproduces the classical solution, with residual differences small relative to the marginal amplitudes.

Representative pair marginals for the 3D3V proof of concept calculation

Figure 10. Representative pair marginals for the 3D3V proof of concept calculation with np=5n_p = 5, Nt=2N_t = 2, Δt=0.05\Delta t = 0.05, and T=0.1T = 0.1. (a) Quantum axial marginal fzvzQf^{Q}_{zv_z} and (b) its pointwise difference from classical result, fzvzQ−fzvzCf^{Q}_{zv_z} - f^{C}_{zv_z}. (c) Quantum transverse-velocity marginal fvxvyQf^{Q}_{v_xv_y} and (d) the corresponding difference.

Table 1 summarizes the gate-level simulation cases performed with an allocated memory of 440 GB per job and a walltime limit of 120 hours. The baseline cases (Eq. (81)) mainly use np=5n_p = 5, including one np=7n_p = 7 case to examine the effect of increasing the size of the auxiliary register on solution error ϵQ−C\epsilon_{Q-C} (Eq. (90)). The table also reports the final maximum MPS bond dimension and Aer simulation time to indicate the cost of classically verifying the complete quantum circuit.

npn_pNtN_tΔt\Delta tTTϵQ−C\epsilon_{Q-C}mass errorχmax⁡\chi_{\max}MPS time (h)
510.0050.0059.78×10−49.78 \times10^{-4}1.14×10−51.14 \times10^{-5}616.9
520.0050.0101.69×10−31.69 \times10^{-3}2.64×10−52.64 \times10^{-5}8516.8
550.0050.0252.72×10−32.72 \times10^{-3}5.76×10−55.76 \times10^{-5}11958.5
510.0500.0507.74×10−37.74 \times10^{-3}5.65×10−55.65 \times10^{-5}15922.5
520.0500.1003.88×10−23.88 \times10^{-2}2.68×10−42.68 \times10^{-4}40159.3
720.0050.0101.78×10−31.78 \times10^{-3}1.64×10−41.64 \times10^{-4}20797.0

Table 1. Simulations of low-resolution 3D3V circuits (§ 4.2). The relative mass error is evaluated with respect to the classical solution. χmax⁡\chi_{\max} is the maximum bond dimension of the final MPS representation. Reported MPS time is the Aer simulation time of classical tensor networks.

For the np=5n_p = 5, Δt=0.005\Delta t = 0.005 cases, the full six-dimensional discrepancy increases from 9.78×10−49.78 \times10^{-4} at T=0.005T = 0.005 to 1.69×10−31.69 \times10^{-3} at T=0.01T = 0.01 and 2.72×10−32.72 \times10^{-3} at T=0.025T = 0.025. For the larger step Δt=0.05\Delta t = 0.05, the discrepancy is 7.74×10−37.74 \times10^{-3} at T=0.05T = 0.05 and 3.88×10−23.88 \times10^{-2} at T=0.1T = 0.1. The relative mass error remains below 3×10−43 \times10^{-4} in all completed cases. These results show increasing quantum–classical discrepancy with evolution time, while the relative mass error remains below 3×10−43 \times10^{-4}.

For np=5n_p = 5 and Nt=1N_t = 1, increasing Δt\Delta t from 0.005 to 0.05 increases the final maximum MPS bond dimension from 61 to 159, while the Aer simulation time increases from 6.9 to 22.5 hours. Although the logical circuit topology is unchanged, the modified rotation angles lead to quantum states with different MPS representation costs. This shows that the computational difficulty of gate-level MPS emulation cannot be inferred solely from circuit depth or gate count.

We compare the auxiliary-register size at Nt=2N_t = 2, Δt=0.005\Delta t = 0.005, and T=0.01T = 0.01. Increasing npn_p from 5 to 7 changes the full six-dimensional discrepancy from 1.69×10−31.69 \times10^{-3} to 1.78×10−31.78 \times10^{-3}, so no

improvement is observed for this particular low-resolution instance. In contrast, the classical MPS cost increases substantially as the simulation time rises from approximately 16.8 h to 97.0 h and the maximum bond dimension χmax⁡\chi_{\max} increases from 85 to 207.

The probability PrecP_{\mathrm{rec}} of obtaining the selected p=0p = 0 recovery slice remains between approximately 0.759 and 0.827 for the completed np=5n_p = 5 cases. However, for the np=7n_p = 7, Nt=2N_t = 2 case, PrecP_{\mathrm{rec}} decreases to approximately 0.285. Thus, increasing the auxiliary-register size also increases the postselection overhead at readout.

As an additional verification, the conditional fraction of the field workspace recovered in ∣0⟩\lvert0\rangle exceeds 0.99999999996 for every completed calculation. This confirms that the field and arithmetic work registers are cleanly uncomputed before readout.

The accessible parameter range is constrained by the computational cost of classical MPS emulation. Several more demanding circuit instances did not complete within the imposed 120-hour walltime limit. These observations characterise the practical verification limits under the specified simulation settings and computational resources, but not quantum hardware execution times.

3D3V resource scaling

Table 2 summarises the logical resources of the low-resolution 3D3V circuit. At fixed np=5n_p = 5, the decomposed depth, CNOT count, and rotation-like gate count increase approximately linearly with NtN_t, consistent with repeated application of the split evolution blocks.

The size of the auxiliary register has a more significant effect. At fixed Nt=2N_t = 2, increasing npn_p from 5 to 7 adds only two logical qubits, but increases the decomposed depth and both reported gate counts by approximately a factor of four. This scaling behaviour is consistent with the binary-controlled powers of the Blocks and that increasing the auxiliary resolution directly increases circuit depth (Figure 4).

Discussion

We implemented an explicit Schrödingerised circuit for Vlasov transport with non-constant coefficients coherently. In the 1D1V problem, the streaming coefficient is the encoded velocity vv, while the acceleration coefficient is the non-linear field E(x)E(x). These two cases require different circuit constructions. The affine dependence of vv on the velocity-register bits can be compiled directly into controlled rotations, whereas the cubic field is evaluated coherently using reversible fixed-point arithmetic. The 3D3V construction extends this idea to an anharmonic Penning field problem that contains mixed spatial terms, while the magnetic contribution couples vxv_x and vyv_y. This ensures the resulting evolution is not an otherwise trivial tensor product of independent 1D1V problems.

npn_pNtN_tlogical qubitsdepthNCXN_{\mathrm{CX}}NrotN_{\mathrm{rot}}
51399.95×1059.95 \times10^55.77×1055.77 \times10^58.43×1058.43 \times10^5
52391.92×1061.92 \times10^61.10×1061.10 \times10^61.60×1061.60 \times10^6
55394.68×1064.68 \times10^62.66×1062.66 \times10^63.88×1063.88 \times10^6
72417.71×1067.71 \times10^64.35×1064.35 \times10^66.54×1066.54 \times10^6

Table 2. Logical circuit resources for the low-resolution 3D3V construction. The phase-space resolution is fixed at two qubits per coordinate.

A six-dimensional grid grows exponentially in the number of phase-space coordinates, while its amplitude encoding requires only the sum of the register widths. However, this does not by itself imply an efficient quantum algorithm. The transport operators must still be compiled into gates, and their cost depends on the algebraic structure of the coefficients. The present polynomial examples remain tractable because the velocity coefficient is affine and the electric fields have fixed polynomial degree. More general fields would require a correspondingly more expensive coherent representation.

The resource accounts identify the auxiliary Schrödingerisation register as a separate and more severe source of circuit cost. Increasing the number of time steps produces the expected approximately linear increase in gate count, since the split evolution blocks are repeated. In contrast, increasing npn_p presents a far more significant cost on circuit depth, because the logarithmic encoding of the phase-space grid coexists with a circuit depth that is exponential in npn_p. Hence, the practical utility of the method depends on whether the required auxiliary resolution can remain sufficiently modest for the problem and accuracy of interest.

The numerical results also show that the number of auxiliary phase qubits npn_p should not be regarded as a convergence parameter in isolation. In a representative 3D3V instance, increasing npn_p does not reduce the quantum–classical discrepancy, while the probability of post-selection decreases and the circuit cost increases substantially. This suggests that given sufficient auxiliary representation, increasing npn_p alone need not improve the final solution.

The auxiliary phase profile study in Appendix § 10 also exposes a related trade-off. The optimal error-function profile has a nearly geometric positive tail, which has the potential to be loaded much more economically than a generic amplitude vector. However, the smoother profile removes the fixed central recovery point which is otherwise available for e−∣p∣e^{-|p|}, and this creates post-selection ambiguity at recovery. This means that auxiliary state preparation and state readout cannot be treated as independent design problems.

A further limitation in the present construction is that the fields are prescribed. In a self-consistent Vlasov–Poisson or Vlasov–Maxwell system, the force field depends on the evolving distribution itself. Extending the present construction would therefore require a coherent field solve and feedback of that result into the transport circuit, or alternatively a hybrid quantum–classical update.

Finally, the practical utility of amplitude encoding depends on what is required from the solution. The full NxNvN_xN_v, or six-dimensional phase-space array is represented compactly in the quantum state, but recovering every amplitude would remove this advantage through tomography. The reality may be that only selected observables, moments, marginal probabilities, or other quantities can be estimated directly from the encoded state. The present study demonstrates the explicit gate-level realisation of Schrödingerised Vlasov transport with prescribed, variable coefficients in both 1D1V and coupled 3D3V settings. Classical MPS emulation verifies selected low-resolution instances while revealing substantial computational limitations as the evolution time and auxiliary-register size increase. Our results identify the auxiliary-register overhead as a key quantum resource challenge. Future work should focus on reducing this overhead, developing more efficient coefficient compilation, and extracting physically relevant observables without reconstructing the full phase-space distribution.

Acknowledgements

This project is supported by the National Research Foundation, Singapore through the National Quantum Office, hosted in A*STAR, under the Hybrid Quantum Classical Computing (HQCC) 1.0 Funding Initiative (S24Q7D7001) and the Advanced Quantum Algorithms and Solutions (AQAS) Funding Initiative (S25Q9DA001). The computational work for this article was partially performed on resources of the National Supercomputing Centre (NSCC), Singapore (https://www.nscc.sg).

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data Availability Statement

The source code for the 1D1V simulations, including postprocessing scripts, is available at https://github.com/Soracak/schrodingerised-vlasov.

Fourier-space initialisation and structured loading

In the Schrödingerisation framework, the auxiliary Fourier-space register encodes an initial profile in the warped phase variable pp. Its choice affects both the convergence approximation and the circuit complexity of the state preparation. Starting from the warped phase transformation Φ(t,p)=e−pf(t)\Phi(t,p)=e^{-p}f(t), the original equation is extended to p<0p<0 via initial condition Φ(0,p)=e−∣p∣f(0)\Phi(0,p)=e^{-|p|}f(0) [4, 16]. The natural choice of auxiliary variable fp(p)=e−∣p∣f_p(p)=e^{-|p|} is continuous but not differentiable at p=0p=0 so the Fourier transform exhibit only algebraic decay in pp (cf. Remark 5 [4]).

A natural idea is to replace the cusp of e−∣p∣e^{-|p|} by a smooth splice. For example, one may define [18]

fpher(p)={ep,p<−1,h(p),−1≤p<0,e−p,p≥0.(93)f_p^{\mathrm{her}}(p)= \begin{cases} e^p, & p<-1,\\ h(p), & -1\le p<0,\\ e^{-p}, & p\ge0. \end{cases} \tag*{(93)}

where hh is a Hermite interpolation polynomial fitted at p=−1p=-1 and p=0p=0. This fpher(p)f_p^{\mathrm{her}}(p) improves the regularity relative to e−∣p∣e^{-|p|}, but it remains only a near-optimal remedy.

Error-function profile for optimal scaling

To achieve optimal scaling, Jin et al. [19] introduced an error-function profile

fperf(p)=e−pϕ(p),ϕ(p)=1+erf⁡(ap)2,(94)f_p^{\mathrm{erf}}(p)=e^{-p}\phi(p), \qquad\phi(p)=\frac{1+\operatorname{erf}(ap)}{2}, \tag*{(94)}

where ϕ(p)\phi(p) interpolates smoothly between 0 on the negative side and 1 on the positive side. Unlike the original choice e−∣p∣e^{-|p|}, which has a cusp at p=0p=0, the error-function profile varies smoothly through the origin and is in fact analytic in pp.

In practice, the sharpness parameter aa is chosen from a prescribed tolerance ε\varepsilon. Writing the switching factor as ϕ(p)\phi(p), one requires ∣ϕ(p)−1∣≤ε|\phi(p)-1|\le\varepsilon for p≥p∗p\ge p^*, which leads to

p∗=log⁡(1/ε)a.(95)p^*=\frac{\sqrt{\log(1/\varepsilon)}}{a}. \tag*{(95)}

A canonical choice of a=2log⁡(1/ε)a=2\sqrt{\log(1/\varepsilon)} gives p∗=1/2p^*=1/2. On the warped grid {pj}\{p_j\}, the sampled error-function profile is then

fp,jerf=e−pj1+erf⁡(apj)2.(96)f_{p,j}^{\mathrm{erf}}=e^{-p_j}\frac{1+\operatorname{erf}(ap_j)}{2}. \tag*{(96)}

Unlike the original profile [4], this vector is strongly asymmetric. In particular,

fperf(0)=12,(97)f_p^{\mathrm{erf}}(0)=\frac{1}{2}, \tag*{(97)}

whereas fp(0)=1f_p(0)=1 for the original profile. For the error-function initialisation, the physical solution should be reconstructed from a positive pp-slice lying beyond the transition region, where ϕ(p)≈1\phi(p)\approx1 and hence fperf(p)≈e−pf_p^{\mathrm{erf}}(p)\approx e^{-p}.

Once the error-function profile Eq. (94) is adopted, the next question is how to prepare the auxiliary pp-register efficiently. A generic Möttönen state preparation operator [20], such as the initialize instruction set [17], loads the normalised sampled vector exactly, but it treats the target as an arbitrary amplitude vector. Let

Np=2np,Δp=2πRNp,pj=−πR+jΔp.(98)N_p = 2^{n_p}, \qquad\Delta p = \frac{2\pi R}{N_p}, \qquad p_j = -\pi R + j\Delta p. \tag*{(98)}

Remark 2 (Readout slice for the error-function profile. [19 Theorem 5.2]). For e−∣p∣e^{-|p|}, p=0p=0 is a natural recovery point because fp(0)=1f_p(0)=1. For fperf⁡f_p^{\operatorname{erf}}, however, p=0p=0 lies in the transition layer, where ϕ(0)=1/2\phi(0)=1/2. Recovery should instead be performed from a positive pp-slice for which the switching factor has already saturated, so that

ϕ(pj)≈1,fperf⁡(pj)≈e−pj.\phi(p_j) \approx1, \qquad f_p^{\operatorname{erf}}(p_j) \approx e^{-p_j}.

Using the grid notation defined above, we select a recovery index jrecj_{\mathrm{rec}} such that pjrecp_{j_{\mathrm{rec}}} lies away from both the switching region and the right boundary. The admissible window is

pmin⁡=p∗+Λ+T+nguardΔp,pmax⁡=pNp−1−Λ−T−nguardΔp,(99)p_{\min} = p^{*} + \Lambda_{+}T + n_{\mathrm{guard}}\Delta p, \qquad p_{\max} = p_{N_p-1} - \Lambda_{-}T - n_{\mathrm{guard}}\Delta p, \tag*{(99)}

where p∗p^{*} is the point beyond which ∣ϕ(p)−1∣≤ϵ|\phi(p)-1| \le\epsilon. nguardn_{\mathrm{guard}} is a non-negative integer specifying the number of auxiliary grid spacings reserved as a safety margin, while Λ±T\Lambda_{\pm}T denote conservative bounds on the auxiliary-direction drift over the evolution time TT.

Here, the readout slice is selected by

jrec=arg min⁡0≤j<Np∣pj−clamp⁡(π,pmin⁡,pmax⁡)∣.(100)j_{\mathrm{rec}} = \underset{0 \le j < N_p}{\operatorname{arg\,min}}\left|p_j-\operatorname{clamp}(\pi,p_{\min},p_{\max})\right|. \tag*{(100)}

Note that the target p=πp=\pi is not assumed to be universally optimal. At the selected slice, the full physical-grid vector

Φ(T,pjrec)≈e−pjrecf(T),(101)\Phi(T,p_{j_{\mathrm{rec}}}) \approx e^{-p_{j_{\mathrm{rec}}} }f(T), \tag*{(101)}

is obtained by fixing the auxiliary coordinate to pjrecp_{j_{\mathrm{rec}}}.

Remark 3 (Nearly geometric structure of the positive tail). The error-function profile Eq. (94) [19] has an additional discrete structure on the positive half of the pp-grid. Since the positive half of the grid begins at j=Np/2j=N_p/2,

pNp/2+k=kΔp,k=0,…,2np−1−1.p_{N_p/2+k} = k\Delta p, \qquad k=0,\ldots,2^{n_p-1}-1.

Substituting j=Np/2+kj=N_p/2+k into Eq. (96) gives

fp,Np/2+kerf⁡=e−kΔp1+erf⁡(akΔp)2.(102)f_{p,N_p/2+k}^{\operatorname{erf}} = e^{-k\Delta p}\frac{1+\operatorname{erf}(ak\Delta p)}{2}. \tag*{(102)}

The first factor is exactly geometric in the grid index kk,

e−kΔp=(e−Δp)k=rk,r:=e−Δp,(103)e^{-k\Delta p} = \left(e^{-\Delta p}\right)^k = r^k, \qquad r := e^{-\Delta p}, \tag*{(103)}

while the second factor is a smooth switching function, which equals 1/21/2 at p=0p=0.

For kΔp≳p∗k\Delta p \gtrsim p^{*}, however, the switching factor is close to one. Thus the positive tail satisfies

fp,Np/2+kerf⁡≈rk.(104)f_{p,N_p/2+k}^{\operatorname{erf}} \approx r^k. \tag*{(104)}

The auxiliary profile is not an arbitrary vector, but it consists of a nearly geometric positive tail, with a small non-geometric transition region near p=0p=0.

This offers possibilities for efficient state preparation, as follows.

Geometric half-support loading

To exploit this structure, let m=np−1m=n_p-1 denote the number of qubits used to index the positive half of the pp-register. Using the binary expansion

k=∑l=0m−12lbl,bl∈{0,1},(105)k=\sum_{l=0}^{m-1}2^l b_l,\qquad b_l\in\{0,1\}, \tag*{(105)}

the geometric factor reads

rk=∏l=0m−1(r2l)bl.(106)r^k=\prod_{l=0}^{m-1}\left(r^{2^l}\right)^{b_l}. \tag*{(106)}

Define the normalised geometric tail state as a product state,

∣gr⟩=⨂l=0m−1∣0⟩+r2l∣1⟩1+r2l+1.(107)\lvert g_r\rangle=\bigotimes_{l=0}^{m-1}\frac{\lvert0\rangle+r^{2^l}\lvert1\rangle}{\sqrt{1+r^{2^{l+1}}}}. \tag*{(107)}

This state can be prepared by applying one RyR_y rotation to each of the mm lower qubits,

Ry(θl),θl=2tan⁡−1(r2l),l=0,…,m−1.(108)R_y(\theta_l),\qquad\theta_l=2\tan^{-1}\left(r^{2^l}\right),\qquad l=0,\ldots,m-1. \tag*{(108)}

An additional NOT gate on the most significant pp-qubit selects the nonnegative half of the grid. Thus the geometric half-support loader prepares

∣0⟩⊗np⟼∣1⟩∣gr⟩,(109)\lvert0\rangle^{\otimes n_p}\longmapsto\lvert1\rangle\lvert g_r\rangle, \tag*{(109)}

using only O(np)O(n_p) single-qubit rotations.

Patched geometric loading

Let P=2bP=2^b be the number of positive grid points to patch, where b≤mb\le m, and

tk:=fp,Np/2+kref,k=0,…,P−1.t_k:=f^{\mathrm{ref}}_{p,N_p/2+k},\qquad k=0,\ldots,P-1.

The normalised target patch, embedded in the mm-qubit positive-half register, is

∣tP(m)⟩:=1TP∑k=0P−1tk∣k⟩m,TP:=(∑k=0P−1∣tk∣2)1/2,(110)\lvert t_P^{(m)}\rangle:=\frac{1}{\mathcal{T}_P}\sum_{k=0}^{P-1}t_k\lvert k\rangle_m,\qquad\mathcal{T}_P:=\left(\sum_{k=0}^{P-1}\lvert t_k\rvert^2\right)^{1/2}, \tag*{(110)}

where ∣k⟩m\lvert k\rangle_m denotes the mm-qubit computational basis state corresponding to the integer kk. On this selected subspace, the circuit first unprepares the local geometric state on the first bb qubits and then prepares the exact target patch state ∣tP(m)⟩\lvert t_P^{(m)}\rangle.

g~k={BPtk(∑l=0P−1∣tl∣2)1/2,0≤k<P,gk,k≥P.(111)\widetilde{g}_k= \begin{cases} B_P\dfrac{t_k}{\left(\sum_{l=0}^{P-1}\lvert t_l\rvert^2\right)^{1/2}}, & 0\le k<P,\\ g_k, & k\ge P. \end{cases} \tag*{(111)}

where the geometric probability assigned to the first PP amplitudes is

BP2=∑k=0P−1∣gk∣2.(112)B_P^2=\sum_{k=0}^{P-1}\lvert g_k\rvert^2. \tag*{(112)}

The resulting loader is a product-state geometric loader plus a local exact correction on PP amplitudes.

Circuit description Let GbG_b denote the geometric product loader restricted to the first bb positive-half qubits and let

Tb∣0⟩⊗b=1TP∑k=0P−1tk∣k⟩.(113)T_b|0\rangle^{\otimes b}=\frac{1}{T_P}\sum_{k=0}^{P-1}t_k|k\rangle. \tag*{(113)}

After the full geometric loader has prepared ∣1⟩∣gr⟩|1\rangle|g_r\rangle, the circuit applies the zero-controlled operation C0(TbGb†)C_0\left(T_bG_b^\dagger\right), where the controls are qb,…,qm−1q_b,\ldots,q_{m-1}. Thus, the local geometric state is unprepared and replaced by the exact patch state:

∣gr(b)⟩→Gb†∣0⟩⊗b→Tb1TP∑k=0P−1tk∣k⟩,(114)|g_r^{(b)}\rangle\xrightarrow{G_b^\dagger}|0\rangle^{\otimes b}\xrightarrow{T_b}\frac{1}{T_P}\sum_{k=0}^{P-1}t_k|k\rangle, \tag*{(114)}

with other positive-half blocks unchanged. Hence, the patched loader prepares

∣0⟩⊗np⟼∣1⟩∣gr⟩⟼∣1⟩∣g⟩,(115)|0\rangle^{\otimes n_p}\longmapsto|1\rangle|g_r\rangle\longmapsto|1\rangle|g\rangle, \tag*{(115)}

where the second map is the zero-controlled C0(TbGb†)C_0\left(T_bG_b^\dagger\right).

Benchmark and scaling

We quantify the loader error using the state fidelity

F(ψ,φ):=∣⟨ψ∣φ⟩∣2,(116)\mathcal{F}(\psi,\varphi):=|\langle\psi\mid\varphi\rangle|^2, \tag*{(116)}

where ∣φ⟩|\varphi\rangle is the normalised sampled error-function target and ∣ψ⟩|\psi\rangle is the state produced by the loader. Figure A1 shows a comparison of amplitude ψp(pj)\psi_p(p_j) and pointwise-error ∣Δψp(pj)∣|\Delta\psi_p(p_j)| for structured pp-register loaders for np=7n_p=7 and R=4R=4. The geometric half-support loader captures the positive exponential tail but has localised error near p=0p=0. Patched geometric loaders reduce this transition-region error by applying an exact local correction to the first PP positive-grid amplitudes.

Amplitude and pointwise-error plots for structured $p$-register loaders

Figure A1. Amplitude ψp(pj)\psi_p(p_j) and pointwise-error ∣Δψp(pj)∣|\Delta\psi_p(p_j)| plotted against p∈(−5,5)p \in(-5,5) for structured pp-register loaders with np=7n_p = 7 and R=4R = 4.

Table A1 compares two representative grids with the same R=4R=4. For np=5n_p=5, the grid spacing is Δp=π/4\Delta p=\pi/4, and a four-point patch P=4P=4 gives fidelity F≈0.9994\mathcal{F}\approx0.9994. For np=7n_p=7, the grid is finer Δp=π/16\Delta p=\pi/16, and P=4P=4 gives fidelity F≈0.9931\mathcal{F}\approx0.9931. To refine the pp-grid, one may increase PP to cover more of the switching region.

Loadernp=5n_p=5np=7n_p=7
F\mathcal{F}DepthSizeCNOTF\mathcal{F}DepthSizeCNOT
Exact tree1.00005488172901.0000407370552872
Exact half-support tree1.00004085761960.9998308648961860
Geometric half-support0.89861500.9282170
Patched geometric, P=2P=20.985085127400.976113221980
Patched geometric, P=4P=40.99942333281240.993110161525576

Table A1. Loader fidelities F\mathcal{F} and circuit metrics for R=4R=4 and ε=10−6\varepsilon=10^{-6}. Fidelity values are rounded to four decimal places. The exact tree loaders grow rapidly with npn_p. The patched loaders trade a local exact correction near p=0p=0 for improved fidelity.

A generic exact tree on Np=2npN_p=2^{n_p} amplitudes requires O(2np)O(2^{n_p}) controlled rotations, and the circuit metrics grow rapidly as npn_p increases from 5 to 7. By contrast, the geometric half-support

loader uses only m=np−1m = n_p - 1 independent one-qubit rotations together with a single NOT gate, which represents a clear scaling advantage, but at the cost of fidelity.

The patched loader supplements this geometric construction with a local exact correction acting on the first PP positive amplitudes. To ensure that the patch spans the entire switching region, one should choose PP such that PΔp≳p∗P\Delta_p \gtrsim p^{*}, or equivalently,

P≳p∗2np2πR.(117)P \gtrsim\frac{p^{*}2^{n_p}}{2\pi R}. \tag*{(117)}

The geometric part of the construction is scalable, while the patch introduces a controlled local state-preparation overhead whose cost depends on both PP and the number of control qubits. Figure A2 summarises this accuracy–cost tradeoff. The left panel plots the error 1−F1 - \mathcal{F} against the decomposed gate count NgateN_{\mathrm{gate}}, which is the total number of elementary gates after circuit decomposition into Qiskit’s standard one- and two-qubit gate set. The right panel plots the same circuit-size metric as a function of npn_p. The geometric loader remains small because it uses only one-qubit rotations, while the patched loaders improve fidelity at the price of controlled-patch overhead.

Accuracy–cost and cost-scaling comparison for the structured p-register loaders

Figure A2. Accuracy–cost and cost-scaling comparison for the structured pp-register loaders as npn_p varies from 4 to 8, with larger markers indicating larger npn_p. (Left) Error 1−F1 - \mathcal{F} versus decomposed gate count NgateN_{\mathrm{gate}}, with each curve corresponding to a fixed loader. (Right) Decomposed circuit size as a function of npn_p for the same loaders.

Reality of finite-register readout

The preceding discussion shows that the error-function profile has two attractive features in that it removes the cusp of e−∣p∣e^{-|p|}, and its positive tail is nearly geometric on the discrete pp-grid. These properties make it appealing for structured state preparation; in particular, the geometric half-support achieves fidelity up to F≈0.93\mathcal{F} \approx0.93 using O(np)\mathcal{O}(n_p) single-qubit rotations arranged in O(1)\mathcal{O}(1) circuit depth. However, its finite-register readout has additional practical considerations. For the original profile e−∣p∣e^{-|p|}, the central slice p=0p=0 is a natural readout point because fp(0)=1f_p(0)=1. By contrast, the error-function profile is one-sided, and p=0p=0 lies inside the switching layer where ϕ(0)=1/2\phi(0)=1/2. Thus the error-function profile requires readout from a positive pp-slice beyond the switching region.

Table A2 reports a reduced proof-of-concept comparison for the centred-grid double-well Vlasov test. All runs use

npn_poriginal rel. L2L^2original masserf rel. L2L^2erf mass
56.12×10−36.12 \times10^{-3}1.01544.56×10−34.56 \times10^{-3}1.0140
69.02×10−39.02 \times10^{-3}1.01421.18×10−21.18 \times10^{-2}1.0063
71.22×10−21.22 \times10^{-2}1.01221.72×10−21.72 \times10^{-2}1.0019

Table A2. Finite-register comparison of the original auxiliary profile and the error-function profile.

nx=nv=2,R=4,Nt=5,Δt=0.01.n_x=n_v=2,\qquad R=4,\qquad N_t=5,\qquad\Delta t=0.01.

The original profile is read out at p=0p=0, while the error-function profile uses the bounded-π\pi rule.

The np=5n_p=5 case shows that the error-function profile can perform well with the p=πp=\pi slice. However, increasing npn_p does not produce a monotone improvement in this reduced test. Therefore p=πp=\pi is not a universally optimal readout slice, and it is unclear whether a universally optimal slice exists.

The ambiguity of selecting the readout slice within the admissible interval (pmin⁡,pmax⁡)(p_{\min},p_{\max}) introduces an additional practical reality in the direct application of the error-function profile since the number of discrete grid points in this interval may scale as O(2np)O(2^{n_p}). Hence, the original profile e−∣p∣e^{-|p|}, or even an interpolated polynomial profile, could be more practical for PDE solution simply because its recovery slice is fixed a priori at p=0p=0.

Gate-level details for Block 2 arithmetic

This appendix expands the coherent field-register construction used in Block 2. The goal is to implement the reversible map

∣j⟩X∣0⟩E⟼∣j⟩X∣Efp(j)⟩E,Efp(j)=αfpc(j)+βfpc(j)3,(118)\lvert j\rangle_X\lvert0\rangle_E \longmapsto\lvert j\rangle_X\lvert E_{\mathrm{fp}}(j)\rangle_E,\qquad E_{\mathrm{fp}}(j)=\alpha_{\mathrm{fp}}c(j)+\beta_{\mathrm{fp}}c(j)^3, \tag*{(118)}

where c(j)c(j) is the centred signed integer associated with the encoded position index jj. All arithmetic is reversible and all temporary workspaces are uncomputed unless explicitly stated otherwise.

Step 1: Preparation of the centred signed register. The position index is first copied into an nxn_x-qubit work register cc,

∣j⟩X∣0⟩c⟼∣j⟩X∣j⟩c.(119)\lvert j\rangle_X\lvert0\rangle_c \longmapsto\lvert j\rangle_X\lvert j\rangle_c. \tag*{(119)}

Then we flip the most-significant bit of cc and interpret the bit string in two’s-complement form, giving the centred signed integer

c(j)=j−2nx−1.(120)c(j)=j-2^{n_x-1}. \tag*{(120)}

This way, the unsigned range j=0,…,2nx−1j=0,\ldots,2^{n_x}-1 is mapped to the centred signed range

−2nx−1,…,2nx−1−1.-2^{n_x-1},\ldots,2^{n_x-1}-1.

Step 2: Magnitude extraction. The bit string stored in cc is copied into an nxn_x-qubit unsigned magnitude register cabsc_{\mathrm{abs}}. Let ss denote the sign bit of cc. We map

∣c⟩c∣0⟩cabs⟼∣c⟩c∣∣c∣⟩cabs,(121)\lvert c\rangle_c\lvert0\rangle_{c_{\mathrm{abs}}} \longmapsto\lvert c\rangle_c\lvert\lvert c\rvert\rangle_{c_{\mathrm{abs}}}, \tag*{(121)}

where the target register is left unchanged when s=0s = 0, and is two’s-complement negated when s=1s = 1. Thus cabsc_{\mathrm{abs}} stores the unsigned magnitude of cc, while the signed source register cc is preserved for later use.

Step 3: Square and cubic workspaces. Using reversible multiplication, the magnitude register is promoted first to a square workspace

∣∣c∣⟩cabs∣0⟩sq⟼∣∣c∣⟩cabs∣∣c∣2⟩sq,(122)\lvert\lvert c\rvert\rangle_{c_{\mathrm{abs}}}\lvert0\rangle_{\mathrm{sq}} \longmapsto\lvert\lvert c\rvert\rangle_{c_{\mathrm{abs}}}\lvert\lvert c\rvert^2\rangle_{\mathrm{sq}}, \tag*{(122)}

and then to a cubic workspace

∣∣c∣⟩cabs∣∣c∣2⟩sq∣0⟩cube⟼∣∣c∣⟩cabs∣∣c∣2⟩sq∣∣c∣3⟩cube.(123)\lvert\lvert c\rvert\rangle_{c_{\mathrm{abs}}}\lvert\lvert c\rvert^2\rangle_{\mathrm{sq}}\lvert0\rangle_{\mathrm{cube}} \longmapsto\lvert\lvert c\rvert\rangle_{c_{\mathrm{abs}}}\lvert\lvert c\rvert^2\rangle_{\mathrm{sq}}\lvert\lvert c\rvert^3\rangle_{\mathrm{cube}}. \tag*{(123)}

The intermediate square register is then uncomputed. Although the implementation may allocate a larger register cube, only the low-order 3nx3n_x-qubit subregister is needed to store the signed cubic value. We denote this subregister by

C≔cube[0:3nx−1].(124)C \coloneqq\mathrm{cube}_{[0:3n_x-1]}. \tag*{(124)}

Step 4: Sign restoration of the cubic term. The register cube now contains the unsigned magnitude ∣c∣3\lvert c\rvert^3. To obtain the signed cubic value c3c^3, the circuit applies a controlled two’s-complement negation, controlled on the sign bit ss, such that

∣s⟩MSB(c)∣y⟩C⟼{∣0⟩MSB(c)∣y⟩C,s=0,∣1⟩MSB(c)∣−y⟩C,s=1.(125)\lvert s\rangle_{\mathrm{MSB}(c)}\lvert y\rangle_C \longmapsto \begin{cases} \lvert0\rangle_{\mathrm{MSB}(c)}\lvert y\rangle_C, & s = 0,\\ \lvert1\rangle_{\mathrm{MSB}(c)}\lvert-y\rangle_C, & s = 1. \end{cases} \tag*{(125)}

Thus,

∣c⟩c∣∣c∣3⟩C⟼∣c⟩c∣c3⟩C.(126)\lvert c\rangle_c\lvert\lvert c\rvert^3\rangle_C \longmapsto\lvert c\rangle_c\lvert c^3\rangle_C. \tag*{(126)}

At the gate level, the controlled negation is implemented by a sign-controlled layer of NOT gates on CC, followed by a sign-controlled increment by one, with the carry ancillas uncomputed at the end. The higher-order padding bits of the full cube workspace are left unchanged.

Step 5: Weighted accumulation into the field register. The output register EE is assembled by accumulating the signed linear and cubic contributions into a clean nEn_E-qubit register,

∣c⟩c∣c3⟩C∣0⟩E⟼∣c⟩c∣c3⟩C∣αfpcc+βfpcc3⟩E.(127)\lvert c\rangle_c\lvert c^3\rangle_C\lvert0\rangle_E \longmapsto\lvert c\rangle_c\lvert c^3\rangle_C\lvert\alpha f_{\mathrm{pc}}c+\beta f_{\mathrm{pc}}c^3\rangle_E. \tag*{(127)}

This accumulation is implemented bit by bit using the two’s-complement expansions of the signed source registers. For the linear term,

c=∑i=0nx−22ici−2nx−1cmsb,(128)c = \sum_{i=0}^{n_x-2} 2^i c_i - 2^{n_x-1}c_{\mathrm{msb}}, \tag*{(128)}

and so

αfpC=∑i=0nx−2ci(αfp2i)+cmsb(−αfp2nx−1).(129)\alpha_{\mathrm{fp}}C = \sum_{i=0}^{n_x-2} c_i\left(\alpha_{\mathrm{fp}}2^i\right) + c_{\mathrm{msb}}\left(-\alpha_{\mathrm{fp}}2^{n_x-1}\right). \tag*{(129)}

Thus, for each non-sign bit cic_i, the circuit conditionally adds the shifted constant αfp2i\alpha_{\mathrm{fp}}2^i into EE, while for the sign bit it conditionally adds the two’s-complement constant −αfp2nx−1-\alpha_{\mathrm{fp}}2^{n_x-1}. This uses nxn_x controlled constant-addition blocks.

The cubic term is accumulated in the same way from the signed 3nx3n_x-qubit register

C=∑i=03nx−22ici−23nx−1cmsb,(130)C = \sum_{i=0}^{3n_x-2} 2^i c_i - 2^{3n_x-1}c_{\mathrm{msb}}, \tag*{(130)}

and

βfpC=∑i=03nx−2ci(βfp2i)+cmsb(−βfp23nx−1).(131)\beta_{\mathrm{fp}}C = \sum_{i=0}^{3n_x-2} c_i\left(\beta_{\mathrm{fp}}2^i\right) + c_{\mathrm{msb}}\left(-\beta_{\mathrm{fp}}2^{3n_x-1}\right). \tag*{(131)}

Therefore, the cubic term requires 3nx3n_x controlled constant-addition blocks. In total, Step 5 uses nx+3nx=4nxn_x + 3n_x = 4n_x controlled constant-addition blocks.

Note that each controlled constant-addition block has the same structure. First, the appropriate shifted classical constant is loaded into the nEn_E-qubit workspace

W=(tmp,sq,wmsb,wpad).(132)W = (\mathrm{tmp},\mathrm{sq},w_{\mathrm{msb}},\mathrm{wpad}). \tag*{(132)}

Next, a CDKM ripple-carry adder adds this workspace register into EE, controlled by the relevant source bit. Finally, the workspace preparation and carry initialisation are uncomputed, so that only the updated output register EE remains.

Sensitivity to spatial grid

The original 1D1V implementation samples the initial phase-space distribution on a cell-centred spatial grid (Eq. (5)), whereas the fixed-point acceleration field is evaluated using centred-integer coordinates corresponding to a nodal grid. To assess the numerical sensitivity of the circuit validation to this half-cell offset, we performed an additional reduced 1D1V study.

The test uses the following setting:

nx=nv=3,np=5,R=4,Nt=1,Δt=0.05,n_x = n_v = 3,\qquad n_p = 5,\qquad R = 4,\qquad N_t = 1,\qquad\Delta t = 0.05,

with Lx=Lv=6L_x = L_v = 6, fixed-point precision s=6s = 6, and zero mean drift v0=0v_0 = 0. For this configuration, αfp=48\alpha_{\mathrm{fp}} = 48, βfp=−27\beta_{\mathrm{fp}} = -27.

Two cases were considered:

  1. Cell-centred convention, in which the initial distribution is sampled on a cell-centred spatial grid while the field is evaluated on the nodal grid.

  1. Nodal convention, in which both the initial distribution and the field use nodal spatial coordinates.

Figure C1 compares the spatial marginals for the two cases. In both conventions, the recovered quantum marginal closely follows the scheme-matched classical reference. The reported quantum–classical discrepancies are ϵQ−C=3.81×10−3\epsilon_{Q-C} = 3.81 \times10^{-3} for the original cell-centred initialization and ϵQ−C=3.42×10−3\epsilon_{Q-C} = 3.42 \times10^{-3} for the nodal initialization. The corresponding Aer MPS execution times are 9.29 and 9.69 minutes, respectively, indicating that the classical emulation cost is also broadly insensitive to the choice of initialization grid in this reduced test.

Sensitivity of the reduced 1D1V validation test to the spatial grid convention

Figure C1. Sensitivity of the reduced 1D1V validation test to the spatial grid convention for T=0.05T = 0.05. (a) Spatial marginal for the original implementation, in which the initial distribution is sampled on a cell-centred spatial grid. (b) Spatial marginal for the consistent nodal grid initialization. (c) Quantum–classical difference in the spatial marginal for the cell-centred initialization. (d) Quantum–classical difference in the spatial marginal for the nodal initialization.

The two grid conventions produce different discretised distributions, with an approximately 17% relative L2L_2 difference between the classical solutions after periodic spatial-grid alignment. A comparable difference is already present in the initial distributions, reflecting the sensitivity of this coarse grid to spatial sampling. Nevertheless, both conventions yield similar quantum–classical discrepancies and MPS execution times. Thus, the grid convention affects the discretised solution but has limited impact on circuit validation and MPS emulation cost in this reduced test.

References

  1. [1]Minna Palmroth, Urs Ganse, Yann Pfau-Kempf, Markus Battarbee, Markku Alho, Joonas Nätilä, Ivan Zaitsev, Giulia Cozzani, Konstantinos Papadakis, Leo Kotipalo, Hongyang Zhou, Lucile Turc, Sanni Hoilijoki, Maxime Grandin, Lauri Pänkäläinen, Arto Sandroos, and Sebastian von Alfthan. Vlasov methods in space physics and astrophysics. Living Reviews in Computational Astrophysics, 11(1):3, 2025.
  2. [2]Lukas Einkemmer and Alexander Moriggl. Semi-Lagrangian 4d, 5d, and 6d kinetic plasma simulation on large-scale GPU-equipped supercomputers. The International Journal of High Performance Computing Applications, 37(2):180–196, 2023.DOI
  3. [3]Yuki Sato, Ruho Kondo, Ikko Hamamura, Tamiya Onodera, and Naoki Yamamoto. Hamiltonian simulation for hyperbolic partial differential equations by scalable quantum circuits. Physical Review Research, 6(3):033246, 2024.DOI
  4. [4]Junpeng Hu, Shi Jin, Nana Liu, and Lei Zhang. Quantum Circuits for partial differential equations via Schrödingerisation. Quantum, 8:1563, 2024.DOI
  5. [5]Alexander Engel, Graeme Smith, and Scott E. Parker. Quantum algorithm for the Vlasov equation. Physical Review A, 100(6):062315, 2019.DOI
  6. [6]Abtin Ameri, Erika Ye, Paola Cappellaro, Hari Krovi, and Nuno F. Loureiro. Quantum algorithm for the linear Vlasov equation with collisions. Physical Review A, 107(6):062412, 2023.
  7. [7]Kiichiro Toyoizumi, Naoki Yamamoto, and Kazuo Hoshino. Hamiltonian simulation using the quantum singular-value transformation: Complexity analysis and application to the linearized Vlasov–Poisson equation. Physical Review A, 109(1):012430, 2024.arxiv.org/abs/2304.08937
  8. [8]Koichi Miyamoto, Soichiro Yamazaki, Fumio Uchida, Kotaro Fujisawa, and Naoki Yoshida. Quantum algorithm for the Vlasov simulation of the large-scale structure formation with massive neutrinos. Physical Review Research, 6(1):013200, 2024.DOI
  9. [9]Hayato Higuchi, Juan William Pedersen, Kiichiro Toyoizumi, Kohji Yoshikawa, Chusei Kiumi, and Akimasa Yoshikawa. Quantum calculation for two-stream instability and advection test of Vlasov–Maxwell equations: numerical evaluation of Hamiltonian simulation. Journal of Plasma Physics, 91(4):E116, 2025.DOI
  10. [10]Lukas Einkemmer, Qin Li, Li Wang, and Yunan Yang. Suppressing instability in a Vlasov–Poisson system by an external electric field through constrained optimization. Journal of Computational Physics, 498:112662, 2024.arxiv.org/abs/2305.17994
  11. [11]Lukas Einkemmer, Qin Li, Clément Mouhot, and Yukun Yue. Control of instability in a Vlasov–Poisson system through an external electric field. Journal of Computational Physics, 530:113904, 2025.arxiv.org/abs/2407.15008
  12. [12]Anibal Velozo Ruiz and Renato Velozo Ruiz. Small data solutions for the Vlasov–Poisson system with a repulsive potential. Communications in Mathematical Physics, 405(3):80, 2024.
  13. [13]Megan Griffin-Pickering and Mikaela Iacobelli. Global strong solutions in ℝ³ for ionic Vlasov–Poisson systems. Kinetic and Related Models, 14(4):571–597, 2021.
  14. [14]Gilbert Strang. On the construction and comparison of difference schemes. SIAM Journal on Numerical Analysis, 5(3):506–517, 1968.DOI
  15. [15]C. Z. Cheng and Georg Knorr. The integration of the Vlasov equation in configuration space. Journal of Computational Physics, 22(3):330–351, 1976.
  16. [16]Shi Jin, Nana Liu, and Yue Yu. Quantum simulation of partial differential equations via Schrödingerization. Physical Review Letters, 133(23):230602, 2024.DOI
  17. [17]Ali Javadi-Abhari, Matthew Treinish, Kevin Krsulich, Christopher J. Wood, Jake Lishman, Julien Gacon, Simon Martiel, Paul D. Nation, Lev S. Bishop, Andrew W. Cross, Blake R. Johnson, and Jay M. Gambetta. Quantum computing with Qiskit, 2024.
  18. [18]Shi Jin, Nana Liu, and Chuwen Ma. Schrödingerization based computationally stable algorithms for ill-posed problems in partial differential equations. SIAM Journal on Scientific Computing, 47(4):B976–B1000, 2025.DOI
  19. [19]Shi Jin, Nana Liu, Chuwen Ma, Yizhe Peng, and Yue Yu. On the Schrödingerization method for linear non-unitary dynamics with optimal dependence on matrix queries. Communications in Mathematical Sciences, 24(2):565–592, 2026.
  20. [20]Mikko Möttönen, Juha J. Vartiainen, Ville Bergholm, and Martti M. Salomaa. Transformation of quantum states using uniformly controlled rotations. Quantum Information and Computation, 5(6):467–473, 2005.arxiv.org/abs/quant-ph/0407010

Paper details

vlnc:W_ge3jg7u2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

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