betaFeedback

WorksPhysics and Astronomy

Residual Carleman lift for nonlinear lattice Boltzmann simulations

Abstract

Low-order Carleman embeddings replace nonlinear dynamics by a larger linear system, but truncation errors can accumulate even when the underlying nonlinear model is accurately approximated by a quadratic equation. We construct a residual Carleman lift (RC2) of lattice Boltzmann dynamics in which an independently evolved spectral background supplies the coefficients and forcing of a second-order residual lift. In a double-shear-layer test, conventional fixed-rest Carleman reaches 87.3%87.3\% final error, whereas prepared RC2 achieves 0.136%0.136\% through vortex roll-up, without restarting the lift or using fine-reference feedback. The results demonstrate control of the finite-time accuracy of a fixed-order linear embedding through nonlinear preparation.

Nonlinear interactions govern the formation of structures and the transfer of energy across scales in fluids, plasmas, and many other physical systems. Their evolution nevertheless admits linear descriptions in enlarged spaces: observables, phase-space distributions, or products of the dynamical variables. Such embeddings preserve nonlinearity through the representation of the state, while transferring its evolution to a linear operator. Their renewed interest is partly motivated by quantum computation, whose linear state evolution makes linear embeddings natural candidates for simulating nonlinear classical physics [1, 2].

The central difficulty is to turn an infinite linear representation into a useful finite computation. Carleman-based quantum algorithms have established conditional results for dissipative nonlinear differential equations [3], with subsequent improvements through rescaling, higher-order numerical methods, and tighter error analysis [4]. However, finite-dimensional projection can introduce accumulating errors even when the underlying nonlinear dynamics remains well behaved [5]. The accuracy of the truncated embedding is therefore a physical and numerical question that precedes its quantum realization.

Carleman embedding evolves products of the original variables in a linear hierarchy. Lattice Boltzmann (LB) models provide a concrete setting: streaming is linear, while the local equilibrium carries the nonlinearity. Carleman–LB studies have demonstrated useful accuracy in selected regimes [6, 7]. Together with general truncation estimates [8], they motivate examining how the dynamics presented to a finite lift controls its accuracy horizon.

A quadratic nonlinear model generates cubic and quartic products in the discrete evolution of its second sector. Removing those products can destroy accuracy even when the quadratic model itself remains accurate. Increasing the embedding order enlarges the state. Here we investigate a different control: which part of the dynamics should be supplied by an independent nonlinear background before the residual is lifted?

We introduce a residual Carleman lift truncated at degree two (RC2). A time-dependent background supplies its coefficients and forcing, receiving no information from the residual or the fine nonlinear reference. The pair sector is initialized once and propagated independently; it is never reconstructed from the evolving first sector. This differs operationally from methods that reconstruct a Carleman embedding when switching local charts [9]. Our objective is fidelity to nonlinear LB dynamics on the same lattice. Quantum realization and computational advantage are outside the present study.

Conventional Carleman lift.—Let fn∈R9N2f_n \in\mathbb{R}^{9N^2} collect all postcollision populations on a periodic two-dimensional N×NN \times N lattice; nn denotes the time step. Its value at one node is fn(x)∈R9f_n(\boldsymbol{x}) \in\mathbb{R}^9. With weight vector w∈R9\boldsymbol{w} \in\mathbb{R}^9, the uniform equilibrium at fixed density ρ0\rho_0 and zero velocity is

f(0)=ρ0w⊗1N2.(1)f^{(0)} = \rho_0 \boldsymbol{w} \otimes\boldsymbol{1}_{N^2}. \tag*{(1)}

It is independent of position and time. The full lattice update is fn+1=F(fn)f_{n+1} = \mathcal{F}(f_n), so F(f(0))=f(0)\mathcal{F}(f^{(0)}) = f^{(0)}; its increment vanishes at equilibrium.

We use regularized D2Q9 dynamics [10] because its postcollision state has an exact six-field representation: density ρ\rho, two momentum components j\boldsymbol{j}, and three stress moments. It provides a compact nonlinear demonstrator. The residual-lift construction does not depend on this collision model, and accuracy is assessed within the stable operating regime of the underlying discrete dynamics. The population update and its six-field implementation are specified in the Supplemental Material [11].

The conventional construction expands the local equilibrium about (ρ0,j=0)(\rho_0,\boldsymbol{j}=0). Its quadratic term contains j⊗j/ρ0\boldsymbol{j}\otimes\boldsymbol{j}/\rho_0; density variations in the denominator first contribute at cubic order. Combining this expansion with collision and linear streaming gives, for xn=fn−f(0)\boldsymbol{x}_n=f_n-f^{(0)},

xn+1=A∗xn+B∗[xn⊗xn]+R≥3,∗(xn).(2)\boldsymbol{x}_{n+1} = A_* \boldsymbol{x}_n + B_*[\boldsymbol{x}_n \otimes\boldsymbol{x}_n] + \mathcal{R}_{\geq3,*}(\boldsymbol{x}_n). \tag*{(2)}

Here A∗=DF(f(0))A_* = D\mathcal{F}(f^{(0)}) is the Jacobian at the uniform equilibrium (1), B∗B_* contracts a global pair tensor and R≥3,∗\mathcal{R}_{\geq3,*} is the remainder of order three and higher. Streaming translations are included in these operators. Promoting products to independent sectors and truncating at degree two yields conventional second-order Carleman lift C2:

xn+1=A∗xn+B∗[Zn],Zn+1=(A∗⊗A∗)Zn.(3)\begin{aligned} x_{n+1} &= A_{\ast}x_n+B_{\ast}[Z_n],\\ Z_{n+1} &= (A_{\ast}\otimes A_{\ast})Z_n. \tag*{(3)} \end{aligned}

The initial pair is Z0=x0⊗x0Z_0=x_0\otimes x_0. It includes populations at different nodes and subsequently evolves independently of xn⊗xnx_n\otimes x_n. This conventional lift is constructed about the static background (1). As the flow evolves away from that background, low-order truncation can accumulate substantial error, as in Figure 1.

Figure 1. Population errors relative to direct LB on 642642 sites in the double-shear layer simulation at Re⁡=105\operatorname{Re}=10^{5}, with the common normalization in Eq. (13). Curves show conventional fixed-rest C2, the prepared background at K=28K=28, RC2 using that background, and the nonlinear quadratic residual solution at K=28K=28. The green curve is obtained by evolving Eq. (12) nonlinearly, without Carleman lifting; its error reflects the accumulated omission of cubic and higher residual terms.

Preparing the nonlinear background and lifting the residual.— To address the limitations of the uniform background (1) for the Carleman lift, we aim at constructing a nonlinear background cnc_n and lift the residual rn=fn−cnr_n=f_n-c_n instead of lifting xn=fn−f(0)x_n=f_n-f^{(0)}. To that end, fix the flow setup, Reynolds number, a convective-time horizon 0≤t∗≤T∗0\leq t^{\ast}\leq T^{\ast}, and a lattice pair (Nc,Nt)(N_c,N_t), with Nc<NtN_c<N_t. The coarse lattice NcN_c supplies the initial trajectory for preparation; the prepared background and residual lift both evolve on the target lattice NtN_t. In other words, our goal is to prepare a background that carries enough nonlinear evolution for a low-order Carleman residual lift to remain accurate throughout the target interval T∗T^{\ast}. The coarse trajectory on the NcN_c grid alone need not provide such a background.

Let the coarse-grid populations gm∈VNcg_m\in\mathcal{V}_{N_c} evolve by gm+1=Fc(gm)g_{m+1}=\mathcal{F}_c(g_m) on Nc2N_c^2 coarse sites; Ft\mathcal{F}_t denotes the LB map on Nt2N_t^2 target sites. For common lattice velocity UU, coarse and target times are t∗=Um/Nct^{\ast}=Um/N_c and Un/NtUn/N_t, respectively. The required step counts are Mc=T∗Nc/UM_c=T^{\ast}N_c/U and Mt=T∗Nt/UM_t=T^{\ast}N_t/U. Spatial prolongation Pc→t\mathcal{P}_{c\to t} and interpolation Ic\mathcal{I}_c of the coarse time series define the transferred coarse trajectory bnb_n on the target grid NtN_t:

bn=Pc→tIc[{gm}](Un/Nt)∈VNt.(4)b_n=\mathcal{P}_{c\to t}\mathcal{I}_c[\{g_m\}](Un/N_t)\in\mathcal{V}_{N_t}. \tag*{(4)}

Thus bnb_n represents the coarse solution at target nodes and times; it is not evolved by Ft\mathcal{F}_t.

To prepare the background cn∈VNtc_n\in\mathcal{V}_{N_t}, we evolve selected modes nonlinearly under the target map and prescribe their complement from bnb_n:

c0=PKf0+(I−PK)b0,cn+1=PKFt(cn)+(I−PK)bn+1.(5)\begin{aligned} c_0 &= P_Kf_0+(I-P_K)b_0,\\ c_{n+1} &= P_K\mathcal{F}_t(c_n)+(I-P_K)b_{n+1}. \tag*{(5)} \end{aligned}

Here, the spectral projection PKP_K retains integer Fourier wavevectors ∣k∣≤K\lvert\boldsymbol{k}\rvert\leq K in each target population field,

PKh^(k)={h^(k),kx2+ky2≤K2,0,otherwise.(6)\widehat{P_Kh}(\boldsymbol{k})= \begin{cases} \widehat{h}(\boldsymbol{k}), & k_x^2+k_y^2\leq K^2,\\ 0, & \text{otherwise}. \end{cases} \tag*{(6)}

The projector retains the zero mode, so PKf(0)=f(0)P_Kf^{(0)}=f^{(0)}. It acts only in background preparation. The spectral cut-off KK will be specified below. With this, the nonlinear background cnc_n is constructed. While the background is intended to capture some essential nonlinearity of the flow, its remaining error can still exceed the desired accuracy, as in Figure 1.

We now introduce the residual rnr_n relative to the background by writing fn=cn+rnf_n=c_n+r_n. Both the residual and its unrestricted pair vector evolve on the target lattice NtN_t. Importantly, by construction, the background cnc_n receives no feedback from the residual dynamics. The prescribed forcing is the one-step defect of the background:

dn=Ft(cn)−cn+1=(I−PK)[Ft(cn)−bn+1].(7)d_n=\mathcal{F}_t(c_n)-c_{n+1}=(I-P_K)[\mathcal{F}_t(c_n)-b_{n+1}]. \tag*{(7)}

The defect (7) is an input to the lift, not an auxiliary diagnostic: even a zero residual evolves to dnd_n at the next step. Indeed, expanding the exact identity, rn+1=Ft(cn+rn)−cn+1r_{n+1}=\mathcal{F}_t(c_n+r_n)-c_{n+1}, gives a closed equation for the residual,

rn+1=dn+Anrn+Bn[rn⊗rn]+R≥3,n(rn),(8)r_{n+1}=d_n+A_nr_n+B_n[r_n\otimes r_n]+\mathcal{R}_{\geq3,n}(r_n), \tag*{(8)}

where An=DFt(cn)A_n=D\mathcal{F}_t(c_n) and Bn[h⊗k]=12D2Ft(cn)[h,k]B_n[h\otimes k]=\frac{1}{2}D^2\mathcal{F}_t(c_n)[h,k]. Similar to the conventional map (2), the rational density dependence of the LB equilibrium gives an explicit higher-order remainder R≥3,n\mathcal{R}_{\geq3,n} [11]. As in C2, RC2 evolves an independent symmetric pair tensor, now initialized from the residual: r0=f0−c0r_0=f_0-c_0 and Z0=r0⊗r0Z_0=r_0\otimes r_0. Introduce linear combinations

sn=Anrn+qn,qn=Bn[Zn].(9)\begin{aligned} s_n &= A_nr_n+q_n,\\ q_n &= B_n[Z_n]. \tag*{(9)} \end{aligned}

The RC2 system is

rn+1=dn+sn,(10)r_{n+1}=d_n+s_n, \tag*{(10)}
Zn+1=(An⊗An)Zn+dn⊗sn+sn⊗dn+dn⊗dn.(11)Z_{n+1}=(A_n\otimes A_n)Z_n+d_n\otimes s_n+s_n\otimes d_n+d_n\otimes d_n. \tag*{(11)}

All terms of residual degree at most two are retained, treating the prescribed background forcing dnd_n (7) as degree zero, including its couplings to the lifted quadratic sector qnq_n (9). With the background prescribed, the RC2 equations (10) and (11) are affine linear in (r,Z)(r,Z), or homogeneous linear after adjoining a constant. No products of the unknown first-sector state are used to close the pair sector. We propagate the RC2 lift classically through an exact factor representation, retaining all spatial pairs without rank truncation or reinitialization [11]. For the uniform-background special case, replace the initialization in Eq. (5) by c0=f(0)c_0=f^{(0)} and set bn=0b_n=0. Keeping the same projector and update gives cn+1=PKFt(f(0))=f(0)c_{n+1}=P_K\mathcal{F}_t(f^{(0)})=f^{(0)} in (5) and dn=0d_n=0 in (7), and RC2 (10) and (11) becomes the conventional C2 (3).

For diagnostic purposes only, we separately evolve the nonlinear quadratic equation obtained by neglecting the remainder in the exact residual equation (8),

r~n+1=dn+Anr~n+Bn[r~n⊗r~n].(12)\widetilde{r}_{n+1}=d_n+A_n\widetilde{r}_n+B_n[\widetilde{r}_n\otimes\widetilde{r}_n]. \tag*{(12)}

With r~0=r0\widetilde{r}_0=r_0, this auxiliary evolution uses the same background and actual residual products. It is not required to execute RC2 and supplies no feedback to it. The comparison separates errors from the quadratic approximation of the residual dynamics from those introduced by truncating its Carleman lift; an exact identity for the latter’s pair-consistency defect is derived in the Supplemental Material.

A coarse pilot and transfer of preparation.— Finally, it remains to specify the spectral cutoff KK. The latter is evaluated using a pilot simulation construction. The pair of coarse-target lattices (Nc,Nt)(N_c,N_t) above is modeled by a surrogate pair (Nc′,Nt′)(N'_c,N'_t), with Nt′=NcN'_t=N_c, Nc′<Nt′N'_c<N'_t. For the surrogate pair (Nc′,Nt′)(N'_c,N'_t), we first compute the nonlinear LB trajectories for the same flow and horizon, then screen predefined cutoffs to select K′K', which is transferred to the (Nc,Nt)(N_c,N_t) pair. In this work, we test a double shear layer in the periodic unit square at Re⁡=105\operatorname{Re}=10^5, with lattice velocity U=0.08U=0.08, layer width 0.040.04, and transverse perturbation 0.05Usin⁡(2πx)0.05U\sin(2\pi x). The grid-dependent relaxation time τN=1/2+3UN/Re⁡\tau_N=1/2+3UN/\operatorname{Re} preserves UU and Reynolds number. We prescribe T∗=3T^*=3: 600, 1200, and 2400 steps on Nc′=16N'_c=16, Nt′=Nc=32N'_t=N_c=32 and Nt=64N_t=64 lattices, respectively. These tests measure fidelity to the same-lattice nonlinear LB dynamics, rather than resolved continuum turbulence. All population errors use

ϵ(n)=100∥fnapp−fnLB∥w∥f0−f(0)∥w.(13)\epsilon(n)=100\frac{\lVert f_n^{\mathrm{app}}-f_n^{\mathrm{LB}}\rVert_w}{\lVert f_0-f^{(0)}\rVert_w}. \tag*{(13)}

Here app specifies the method and NN is the lattice size used for that comparison. The weighted norm is ∥g∥w2=gTW−1g\lVert g\rVert_w^2=g^{\mathsf{T}}W^{-1}g, where W=diag⁡(w)⊗IN2W=\operatorname{diag}(w)\otimes I_{N^2}. It equals the Euclidean norm in the six orthonormal moment coordinates.

The pilot pair (Nc′,Nt′)=(16,32)(N'_c,N'_t)=(16,32) uses shared nonlinear LB trajectories on 16216^2 and 32232^2 sites. For each integer K′=0,…,14K'=0,\ldots,14, the candidate background and its linear residual correction both evolve on the pilot target lattice, 32232^2, with rn+1lin=dn+Anrnlinr^{\mathrm{lin}}_{n+1}=d_n+A_n r^{\mathrm{lin}}_n. The score is the maximum of their reconstruction error against the shared 32232^2 reference over the entire prescribed interval. The smallest cutoff meeting a 1% tolerance is K′=14K'=14 (Figure 2). This screening step assesses preparation; it is not a bound on RC2 error.

Figure 2. Cutoff selection on the (Nc′,Nt′)=(16,32)(N'_c,N'_t)=(16,32) pilot at Re⁡=105\operatorname{Re}=10^5. Each point is the peak population error through t∗=3t^*=3 for a candidate prepared background with its linear residual correction. The dashed line is the prescribed 1% tolerance; the open circle marks the smallest passing cutoff, K′=14K'=14.

We transfer this preparation to (Nc,Nt)=(32,64)(N_c,N_t)=(32,64) using K/Nt=K′/Nt′K/N_t=K'/N'_t, preserving approximately the fraction of Fourier modes evolved nonlinearly. Its target accuracy is then tested over the prescribed interval. Both lattice sizes and the cutoff are doubled, giving K=28K=28, with the fine-to-coarse ratio remaining two. The same flow and horizon T∗T^* are used; the choice is specific to both. The Nt=64N_t=64 nonlinear reference is used only for validation, never for preparation or lifted evolution.

Accuracy through nonlinear roll-up.— Figure 1 compares the conventional C2 lift (3), the prepared background (5), the residual RC2 lift (10) and (11), and the nonlinear quadratic residual solution (12). For this Nt=64N_t=64, Re⁡=105\operatorname{Re}=10^5 test, the fixed-rest linearization (C1, not shown) first exceeds the 1% population error after one lattice step, t∗=0.00125t^*=0.00125; its error subsequently oscillates. Conventional C2 first crosses 1% at t∗=0.15375t^*=0.15375, after 123 lattice steps, and ends at 87.34%. At K=28K=28, the prepared background has 1.539% final error, RC2 has 0.1358%, and the nonlinear quadratic residual solution has 2.48×10−52.48 \times10^{-5}%. The green curve thus measures the error accumulated by omitting cubic and higher terms of the exact residual equation, evolved nonlinearly; the blue curve is obtained by Carleman-lifting that quadratic equation and truncating its hierarchy at degree two. It is noted that, while the nonlinear background by itself is more accurate than the standard C2, the RC2 is a genuine improvement of the background. All four curves use the same LB reference and normalization.

The final RC2 velocity and vorticity errors, normalized by the corresponding current reference norms, are 0.1309% and 0.9045%. Figure 3 shows the formation and deformation of the vortices over the full interval. Additional cutoff sweeps and the diagnostic return of omitted cubic sources to the population sector are reported in the supplement. For the selected cutoff K=28K=28, the prepared background retains 59.9% of the target Fourier modes. RC2 reduces its final population error from 1.54% to 0.136%, demonstrating that the remaining deviation can be corrected by a second-order linear lift over the prescribed interval.

Figure 3. Double-shear-layer roll-up at Re=105\mathrm{Re}=10^{5} on a 64264^{2} lattice. Direct nonlinear LB (top) and RC2 with K=28K=28 preparation (bottom), at three convective times. All panels use the same reference-based vorticity scale and display unsmoothed lattice fields. The final relative vorticity error is 0.9045%. The RC2 lift is propagated for 2400 steps without reinitialization.

Residual Carleman lifting is a hybrid construction: a small nonlinear evolution supplies the coefficients and forcing of a larger linear lifted evolution through one-way coupling. The coarse pilot serves as a numerical sandbox, rehearsing the flow over the prescribed interval of convective time and selecting the nonlinear spectral fraction to retain. That preparation is then transferred to the larger lattice. For the tested shear layer, it markedly reduces error accumulation through nonlinear roll-up at unchanged lifting order and without restarting the lift. The construction applies beyond LB; its usefulness depends on preparing a residual that a finite lift can evolve accurately over the required time interval.

Acknowledgement of AI assistance. During the development and preparation of this manuscript, the author used ChatGPT (OpenAI, GPT-5.6 Sol) as an interactive research and writing assistant. All results and final text were reviewed and accepted by the author, who assumes full responsibility for the content of the manuscript.

References

References

  1. [1]I. Joseph, Koopman–von Neumann approach to quantum simulation of nonlinear classical dynamics, Phys. Rev. Research 2, 043102 (2020), doi:10.1103/PhysRevResearch.2.043102.DOI
  2. [2]A. Engel, G. Smith, and S. E. Parker, Linear embedding of nonlinear dynamical systems and prospects for efficient quantum algorithms, Phys. Plasmas 28, 062305 (2021), doi:10.1063/5.0040313.DOI
  3. [3]J.-P. Liu, H. Ø. Kolden, H. K. Krovi, N. F. Loureiro, K. Trivisa, and A. M. Childs, Efficient quantum algorithm for dissipative nonlinear differential equations, Proc. Natl. Acad. Sci. U.S.A. 118, e2026805118 (2021), doi:10.1073/pnas.2026805118; corrected version, arXiv:2103.01835v4 (2026).
  4. [4]P. C. S. Costa, P. Schleich, M. E. S. Morales, and D. W. Berry, Further improving quantum algorithms for nonlinear differential equations via higher-order methods and rescaling, npj Quantum Inf. 11, 141 (2025), doi:10.1038/s41534-025-01084-z.
  5. [5]Y. T. Lin, R. B. Lowrie, D. Aslangil, Y. Subaşı, and A. T. Sornborger, Challenges for quantum computation of nonlinear dynamical systems using linear representations, arXiv:2202.02188v3 (2024).arxiv.org/abs/2202.02188
  6. [6]W. Itani and S. Succi, Analysis of Carleman linearization of lattice Boltzmann, Fluids 7, 24 (2022), doi:10.3390/fluids7010024.DOI
  7. [7]C. Sanavio, W. A. Simon, A. Ralli, P. Love, and S. Succi, Carleman-lattice-Boltzmann quantum circuit with matrix access oracles, Phys. Fluids 37, 037123 (2025), doi:10.1063/5.0254588.DOI
  8. [8]M. Forets and A. Pouly, Explicit error bounds for Carleman linearization, arXiv:1711.02552 (2017).arxiv.org/abs/1711.02552
  9. [9]I. Novikau and I. Joseph, Globalizing the Carleman linear embedding method for nonlinear dynamics, arXiv:2510.15715 (2025).arxiv.org/abs/2510.15715
  10. [10]J. Latt and B. Chopard, Lattice Boltzmann method with regularized pre-collision distribution functions, Math. Comput. Simul. 72, 165–168 (2006), doi:10.1016/j.matcom.2006.05.017.DOI
  11. [11]See Supplemental Material for the population map, the discrete lift and its factor realization, the defect estimate, the shear-layer setup, and numerical checks.

Paper details

vlnc:W_zymmt52k
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents