Introduction
In two spatial dimensions, quantum mechanics allows for unconventional particles known as anyons, that exhibit exchange statistics beyond the familiar bosonic and fermionic cases [1 , 2 , 3 ] . In particular, a counterclockwise exchange of two Abelian anyons adiabatically multiplies their quantum wavefunction by e i π α e^{i\pi\alpha} e iπ α , where 0 < α < 1 0 < \alpha< 1 0 < α < 1 . In condensed matter systems, such Abelian anyons naturally emerge as quasiparticle excitations of topologically ordered phases. Prominent examples include charge-neutral semions (α = 1 / 2 \alpha= 1/2 α = 1/2 ) in chiral spin liquid phases of frustrated magnets [4 , 5 , 6 , 7 ] and fractionally charged e / 3 e/3 e /3 Laughlin anyons (α = 1 / 3 \alpha= 1/3 α = 1/3 ) in fractional quantum Hall phases [8 , 9 , 10 , 11 ] . To connect the exchange statistics of anyons to their collective behavior at finite density, it is crucial to develop a microscopic description of interacting, mobile anyons [12 , 13 , 14 , 15 ] .
Recent experimental developments in two-dimensional Van der Waals heterostructures and quantum simulators have accelerated the need to describe quantum states at finite anyon density. The discovery of fractional Chern insulators in moiré transition metal dichalcogenides (TMDs) [16 , 17 , 18 , 19 ] and multilayer graphene [20 , 21 ] , where anyons are not confined by a strong external magnetic field and can hence disperse [22 , 23 , 24 ] , opens up the possibility to study mobile anyons in quantum materials. In parallel, the rapid development of quantum simulation platforms with carefully engineered Hamiltonians allow anyons to emerge as low-energy quasiparticle excitations in correlated bosonic systems [25 , 26 , 27 ] . These discoveries have stimulated widespread theoretical efforts to characterize the many-body physics of anyons at finite density [28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 ] .
The canonical starting point to study many-body phases of interacting bosons or fermions is to formulate their continuum non-interacting ground states with periodic boundary conditions, topologically equivalent to a torus in two dimensions. However, for anyons, even this non-interacting problem is complicated for two reasons. First, even in the absence of any interaction potential V ( r ) V(r) V ( r ) between two anyons, their braiding statistics makes its way into the Hamiltonian via a statistical interaction mediated by Chern-Simons gauge fields [11 , 42 ] . This couples motion of anyons even at large separation, leading to a non-trivial spectral problem. Second, the torus introduces an additional global constraint: transporting anyons around its two non-contractible loops do not commute. This global transport algebra cannot be represented by an ordinary scalar-periodic wavefunction, so the wavefunction must carry additional structure that encode the braid algebra [43 , 44 ] . The inherent complexity of the toroidal many-anyon Hilbert space thus restricts direct access to many-anyon states on the torus.
Owing to its fundamental importance, several alternative approaches have been developed to theoretically tackle the problem of anyons at finite density. Analytical field-theoretic descriptions often model anyons as charge-flux composites [45 ] , and approximate the fluctuating statistical gauge field by an average effective flux generated by the other anyons [13 , 14 ] . Such approaches permit controlled perturbative limits near bosonic (α → 0 + \alpha\to0^+ α → 0 + ) or fermionic (α → 1 − \alpha\to1^- α → 1 − ) statistics [46 , 47 , 48 , 49 ] , self-consistent magnetic states such as nonlinear Landau levels [48 , 50 ] or specific large-N N N limits [51 ] . Their applicability to generic statistics and parameter regimes, however, is limited. Alternately, numerical exact diagonalization (ED) studies have focused on few-anyon spectral problems [15 ] , often confining the anyonic wave-functions using potential wells [52 , 53 ] . Extending these calculations to delocalized anyonic states on a torus, and scalability with anyon number remain challenging. Similar scalability limitations arise when the anyons are allowed to emerge as excitations by solving for excited states of correlated bosonic or fermionic systems via ED or density matrix renormalization group (DMRG) [35 , 54 , 55 ] . These challenges exemplify the importance of developing complementary techniques to study the fate of anyons at finite density on a torus.
Motivated by the success of neural networks [56 ] for continuum electronic systems [57 , 58 , 59 , 60 , 61 , 62 , 63 , 64 , 65 ] , here we introduce AnyFormer: an attention-based variational neural network ansatz for anyons on a torus. As a minimal setting beyond the reach of perturbation theory in the small deviation in statistics from bosons/fermions, we specialize to semions with α = 1 / 2 \alpha= 1/2 α = 1/2 on a torus, without an external field or additional interaction potential, isolating correlations generated by statistics. Our AnyFormer ansatz combines a smooth two-component orbital frame that encodes the required torus transport structure with periodic neural-network coefficients that capture many-body correlations. The orbital geometry and neural coefficients are learned jointly in the training process, while an analytic pair factor enforces the appropriate short-distance behavior. We formulate a variational Monte Carlo (VMC) optimization protocol applicable across particle numbers, and validate it against an exact two-semion solution and against continuum ED for four and six semions. We then apply AnyFormer to systems of up to N = 16 N = 16 N = 16 semions, well beyond the reach of ED benchmarks. Across these system sizes, we characterize the pair correlations, density fluctuations, and currents across these sizes, finding a pronounced correlation hole, suppressed long-wavelength fluctuations, and, for the first time, a reversal in the sign of the pair-circulation profile at a nearly semion-number independent distance measured in units of the mean semion spacing. At N = 12 , 16 N = 12, 16 N = 12 , 16 , the circulation changes sign again at larger separation, revealing a damped oscillatory radial profile. Collectively, our results establish a route toward microscopic variational studies of bulk anyonic matter, that incorporate both exchange statistics and global constraints imposed by toroidal geometry.
The rest of the paper is organized as follows. In Sec. II, we specify the continuum semion Hamiltonian and its torus boundary conditions, which inform our choice of variational wavefunctions in Sec. III. Sec. III constructs the learned orbital-frame ansatz, and Sec. IV describes its VMC optimization and evaluation. We validate the method against exact solutions and continuum ED in § 1.4 , then present correlations and currents beyond ED in Sec. VI. Section VII concludes with a summary of our key findings and a discussion of future directions.
Continuum semion gas on a torus
We begin with an introduction to the many-anyon Hamiltonian and Hilbert space on a torus. Specifically, we consider N N N identical point semions of mass m m m on a rectangular torus with sides L x , L y L_x,L_y L x , L y and area V = L x L y \mathcal{V}=L_xL_y V = L x L y . We study their kinetic energy without any external field or pair interaction potential V ( r ) V(r) V ( r ) . The statistical connection, global transport, and collision domain specify the key facets of the model used to develop our variational neural state.
Let us denote the coordinate of the i th i^{\text{th}} i th anyon by r i = ( x i , y i ) \mathbf{r}_i=(x_i,y_i) r i = ( x i , y i ) , z i = x i + i y i z_i=x_i+\mathrm{i}y_i z i = x i + i y i , and define R = ( r 1 , … , r N ) \mathbf{R}=(\mathbf{r}_1,\ldots,\mathbf{r}_N) R = ( r 1 , … , r N ) . In the anyon gauge, the wavefunction Ψ A ( R ) \Psi_{\mathrm{A}}(\mathbf{R}) Ψ A ( R ) acquires a phase e i α π e^{\mathrm{i}\alpha\pi} e i α π under a counterclockwise exchange of any two anyons, with α = 1 / 2 \alpha=1/2 α = 1/2 for semions. This leads to a multi-valued wavefunction for a many-semion state. The kinetic Hamiltonian acting on such multi-valued wavefunctions is H A = − κ ∑ i ∇ i 2 H_{\mathrm{A}}=-\kappa\sum_i\nabla_i^2 H A = − κ ∑ i ∇ i 2 , where κ = ℏ 2 / ( 2 m ) \kappa=\hbar^2/(2m) κ = ℏ 2 / ( 2 m ) .
Statistical transmutation factors this phase out of the wavefunction:
Ψ A ( R ) = e i χ ( R ) Ψ ( R ) , A i = ∇ i χ . (1) \Psi_{\mathrm{A}}(\mathbf{R})=e^{\mathrm{i}\chi(\mathbf{R})}\Psi(\mathbf{R}),\qquad\mathcal{A}_i=\nabla_i\chi. \tag*{(1)} Ψ A ( R ) = e i χ ( R ) Ψ ( R ) , A i = ∇ i χ . (1)
Choosing χ \chi χ to increase by π / 2 \pi/2 π /2 under a counterclockwise exchange leaves Ψ \Psi Ψ permutation symmetric. Differentiation gives − i ∇ i Ψ A = e i χ D i Ψ -\mathrm{i}\nabla_i\Psi_{\mathrm{A}}=e^{\mathrm{i}\chi}D_i\Psi − i ∇ i Ψ A = e i χ D i Ψ , so conjugating the anyon-gauge operator, H = e − i χ H A e i χ H=e^{-\mathrm{i}\chi}H_{\mathrm{A}}e^{\mathrm{i}\chi} H = e − i χ H A e i χ , gives the bosonic-gauge Hamiltonian
H = κ ∑ i = 1 N D i ⋅ D i , D i = − i ∇ i + A i . (2) H=\kappa\sum_{i=1}^{N}D_i\cdot D_i,\qquad D_i=-\mathrm{i}\nabla_i+\mathcal{A}_i. \tag*{(2)} H = κ i = 1 ∑ N D i ⋅ D i , D i = − i ∇ i + A i . (2)
This is the torus anyon Hamiltonian [44 ] .
On the plane, χ = 1 2 ∑ i < j arg ( z i − z j ) \chi=\frac{1}{2}\sum_{i<j}\arg(z_i-z_j) χ = 2 1 ∑ i < j arg ( z i − z j ) . On the torus, we replace the relative angle by the phase of the Jacobi theta function
f ( z ) = ϑ 1 ( π z L x | τ ) , τ = i L y / L x , (3) f(z)=\vartheta_1\left(\frac{\pi z}{L_x}\middle|\tau\right),\qquad\tau=\mathrm{i}L_y/L_x, \tag*{(3)} f ( z ) = ϑ 1 ( L x π z τ ) , τ = i L y / L x , (3)
which vanishes linearly at z = 0 z=0 z = 0 and at all of its periodic images. (The first argument of ϑ 1 \vartheta_1 ϑ 1 has periods π \pi π and π τ \pi\tau π τ .) We take χ = 1 2 ∑ i < j arg f ( z i − z j ) \chi=\frac{1}{2}\sum_{i<j}\arg f(z_i-z_j) χ = 2 1 ∑ i < j arg f ( z i − z j ) on a collision-free simply connected region, choosing each argument continuously there. Differentiating this phase, each pair z = z i − z j z=z_i-z_j z = z i − z j contributes to particle i i i the connection
a ( z ) = 1 2 ( Im [ f ′ ( z ) / f ( z ) ] Re [ f ′ ( z ) / f ( z ) ] ) , A i = ∑ j ≠ i a ( z i − z j ) . (4) \mathbf{a}(z)=\frac{1}{2}\begin{pmatrix}\operatorname{Im}[f'(z)/f(z)]\\ \operatorname{Re}[f'(z)/f(z)]\end{pmatrix},\qquad\mathcal{A}_i=\sum_{j\ne i}\mathbf{a}(z_i-z_j). \tag*{(4)} a ( z ) = 2 1 ( Im [ f ′ ( z ) / f ( z )] Re [ f ′ ( z ) / f ( z )] ) , A i = j = i ∑ a ( z i − z j ) . (4)
The prime denotes the derivative with respect to the dimensional coordinate z z z . Near a coincidence, a ≃ ( − y , x ) / ( 2 r 2 ) \mathbf{a}\simeq(-y,x)/(2r^2) a ≃ ( − y , x ) / ( 2 r 2 ) , with counterclockwise circulation π \pi π : half a flux quantum and a phase − 1 -1 − 1 for a full pair loop. Since A i = ∇ i χ \mathcal{A}_i=\nabla_i\chi A i = ∇ i χ , the connection is locally pure gauge away from collisions, and the statistical flux is entirely concentrated on the particles. Because f f f is odd, a ( − z ) = − a ( z ) \mathbf{a}(-z)=-\mathbf{a}(z) a ( − z ) = − a ( z ) . The connection is dimensionally an inverse length, acts as the identity on the components introduced below, and carries no adjustable coupling. Supplemental Material, Sec. S1 verifies the sign of D i D_i D i against the stated exchange phase.
The noncontractible torus cycles require additional transport data. Transport around the y y y cycle gives a position-dependent Aharonov–Bohm phase. Translating another particle around the x x x cycle changes this phase by − 1 -1 − 1 , so the two cycle transports anticommute. This algebra requires a multicomponent wavefunction [43 ] . For semions its minimal representation has two components,
Ψ ( R ) = ( Ψ 1 ( R ) Ψ 2 ( R ) ) , Ψ ( P R ) = Ψ ( R ) , (5) \Psi(\boldsymbol{R}) =
\begin{pmatrix}
\Psi_1(\boldsymbol{R})\\
\Psi_2(\boldsymbol{R})
\end{pmatrix},
\qquad
\Psi(P\boldsymbol{R}) = \Psi(\boldsymbol{R}), \tag*{(5)} Ψ ( R ) = ( Ψ 1 ( R ) Ψ 2 ( R ) ) , Ψ ( P R ) = Ψ ( R ) , (5)
where P P P is any permutation of the particles. At a fixed configuration, Ψ 1 \Psi_1 Ψ 1 and Ψ 2 \Psi_2 Ψ 2 are amplitudes for two global winding states of the same N N N particles. They retain information about transport around the torus handles that the instantaneous particle positions alone do not specify. In the basis chosen below, carrying any one particle around the x x x cycle interchanges these amplitudes; the y y y cycle gives them opposite signs, in addition to a common position-dependent phase. Thus there is one two-component amplitude for the whole configuration, rather than an extra two-state label for each particle. The probability density for finding the semions in configuration R \boldsymbol{R} R is Ψ † ( R ) Ψ ( R ) = ∣ Ψ 1 ( R ) ∣ 2 + ∣ Ψ 2 ( R ) ∣ 2 \Psi^\dagger(\boldsymbol{R})\Psi(\boldsymbol{R}) = |\Psi_1(\boldsymbol{R})|^2 + |\Psi_2(\boldsymbol{R})|^2 Ψ † ( R ) Ψ ( R ) = ∣ Ψ 1 ( R ) ∣ 2 + ∣ Ψ 2 ( R ) ∣ 2 .
The sign from anti-commutation when transported around the two cycles of the torus is carried by the Pauli-like matrices
X = ( 0 1 1 0 ) , Z = ( 1 0 0 − 1 ) , X Z = − Z X . (6) X =
\begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix},
\qquad
Z =
\begin{pmatrix}
1 & 0\\
0 & -1
\end{pmatrix},
\qquad
XZ = -ZX. \tag*{(6)} X = ( 0 1 1 0 ) , Z = ( 1 0 0 − 1 ) , X Z = − Z X . (6)
We work with unwrapped coordinates on the plane covering the torus. Moving particle i i i once around either cycle gives
Ψ ( R + L x e ^ i x ) = e i β x X Ψ ( R ) , (7) \Psi(\boldsymbol{R} + L_x\hat{e}_{ix}) = e^{i\beta_x}X\Psi(\boldsymbol{R}), \tag*{(7)} Ψ ( R + L x e ^ i x ) = e i β x X Ψ ( R ) , (7)
Ψ ( R + L y e ^ i y ) = e i β y + i π s i / L x Z Ψ ( R ) , (8) \Psi(\boldsymbol{R} + L_y\hat{e}_{iy}) = e^{i\beta_y+i\pi s_i/L_x}Z\Psi(\boldsymbol{R}), \tag*{(8)} Ψ ( R + L y e ^ i y ) = e i β y + iπ s i / L x Z Ψ ( R ) , (8)
with s i = ∑ j ≠ i ( x i − x j ) = N x i − X s_i = \sum_{j\ne i}(x_i-x_j) = Nx_i-X s i = ∑ j = i ( x i − x j ) = N x i − X and X = ∑ j x j X = \sum_j x_j X = ∑ j x j . Here e ^ i μ \hat{e}_{i\mu} e ^ i μ shifts one coordinate of one particle. The twists β x , β y \beta_x,\beta_y β x , β y specify additional Aharonov–Bohm phases around the handles. A phase redefinition can move these twists from the boundary rules into a constant vector potential, but cannot in general remove them from both. Supplemental Material, Sec. S1, gives the explicit transformation. Figure 1 summarizes these rules.
Figure 1. The three requirements on a semion wavefunction in the bosonic gauge. Crossing an x x x edge swaps the components. Crossing a y y y edge flips the sign of the second component and adds the Aharonov–Bohm (AB) phase of the flux carried by the other particles. Exchanging two particles leaves Ψ \Psi Ψ unchanged, since the exchange phase sits in A \mathcal{A} A . A close pair vanishes as r 1 / 2 r^{1/2} r 1/2 (Sec. IIIB).
Crossing an x x x edge exchanges the two components. Crossing a y y y edge flips the sign of the second component and adds a position-dependent phase. As a function of x i x_i x i , that phase is exactly the Landau-gauge boundary phase of a charged particle on a torus pierced by ( N − 1 ) / 2 (N-1)/2 ( N − 1 ) /2 flux quanta, the half quanta carried by the other N − 1 N-1 N − 1 particles. Its dependence on the other x j x_j x j records where those fluxes sit. The matrices also ensure consistency when a single particle crosses both periods. Moving particle i i i first by L x L_x L x and then by L y L_y L y must give the same amplitude at the lifted configuration as performing these coordinate shifts in the opposite order. Interchanging X X X and Z Z Z gives a minus sign, while evaluating the y y y -cycle phase after the x x x shift gives the factor e i π ( N − 1 ) e^{i\pi(N-1)} e iπ ( N − 1 ) . These signs cancel precisely for even N N N . The different-particle shifts obey the same consistency requirement (Supplemental Material, Sec. S1).
The singular statistical connection makes the behavior at a coincidence essential to the kinetic energy. We choose the regular, Friedrichs condition, with no contact parameter or finite core radius. Consider an isolated pair with relative separation r i − r j = r ( cos φ , sin φ ) \boldsymbol{r}_i-\boldsymbol{r}_j=r(\cos\varphi,\sin\varphi) r i − r j = r ( cos φ , sin φ ) , while the other particles remain separated. The connection approaches a = e ^ φ / ( 2 r ) \boldsymbol{a}=\hat{e}_{\varphi}/(2r) a = e ^ φ / ( 2 r ) , shifting the relative angular derivative from − i ∂ φ -i\partial_\varphi − i ∂ φ to − i ∂ φ + 1 / 2 -i\partial_\varphi+1/2 − i ∂ φ + 1/2 . With reduced mass m / 2 m/2 m /2 , the leading relative kinetic operator is
H r e l ≃ 2 κ [ − ∂ r 2 − 1 r ∂ r + ( − i ∂ φ + 1 / 2 ) 2 r 2 ] . (9) H_{\mathrm{rel}} \simeq2\kappa\left[-\partial_r^2-\frac{1}{r}\partial_r+\frac{\left(-i\partial_\varphi+1/2\right)^2}{r^2}\right]. \tag*{(9)} H rel ≃ 2 κ [ − ∂ r 2 − r 1 ∂ r + r 2 ( − i ∂ φ + 1/2 ) 2 ] . (9)
Exchanging the particles sends φ → φ + π \varphi\to\varphi+\pi φ → φ + π . Since the bosonic-gauge amplitude is unchanged, an angular harmonic e i ℓ φ e^{i\ell\varphi} e i ℓ φ must satisfy e i ℓ π = 1 e^{i\ell\pi}=1 e i ℓ π = 1 , so only even integers ℓ \ell ℓ occur. In each harmonic, the shifted angular term is the centrifugal barrier ( ℓ + 1 / 2 ) 2 / r 2 (\ell+1/2)^2/r^2 ( ℓ + 1/2 ) 2 / r 2 . Balancing this term against the radial derivatives gives the regular behavior
r ν ℓ e i ℓ φ , ν ℓ = ∣ ℓ + 1 / 2 ∣ , ℓ ∈ 2 Z . (10) r^{\nu_\ell}e^{i\ell\varphi},
\qquad
\nu_\ell= |\ell+1/2|,
\qquad
\ell\in2\mathbb{Z}. \tag*{(10)} r ν ℓ e i ℓ φ , ν ℓ = ∣ ℓ + 1/2∣ , ℓ ∈ 2 Z . (10)
The leading channel, ℓ = 0 \ell=0 ℓ = 0 , vanishes as r 1 / 2 r^{1/2} r 1/2 ; the next, ℓ = − 2 \ell=-2 ℓ = − 2 , starts as r 3 / 2 e − 2 i φ r^{3/2}e^{-2i\varphi} r 3/2 e − 2 i φ . These are the regular solutions selected by the Friedrichs condition. A nonzero contact limit would instead give a kinetic density proportional to 1 / r 2 1/r^2 1/ r 2 , with a logarithmically divergent integral ∫ 0 r d r / r 2 \int_0^r dr/r^2 ∫ 0 r d r / r 2 .
The learned orbital-frame ansatz
Transport, smoothness, and completeness
Our starting idea is to let known two-component functions carry the boundary transport, while a neural network supplies ordinary periodic scalar amplitudes. We denote the neural variational parameters by θ \theta θ . If a given function s ( R ) \boldsymbol{s}(\boldsymbol{R}) s ( R ) obeys Eqs. (7), then
f θ ( R ) = s ( R ) c θ ( R ) (11) f_\theta(\boldsymbol{R}) = \boldsymbol{s}(\boldsymbol{R})c_\theta(\boldsymbol{R}) \tag*{(11)} f θ ( R ) = s ( R ) c θ ( R ) (11)
obeys the same rules for any periodic scalar c θ c_\theta c θ . Choosing both factors to be permutation symmetric also preserves bosonic exchange symmetry. The prescribed transport is then enforced analytically, and the network learns the remaining configuration dependence.
Mathematically, s s s is a section of the rank-two vector bundle specified by the boundary rules [66 ] : it assigns a two-component vector to each configuration, with vectors at identified boundaries related by the prescribed transport matrices. A single product fixes f θ , 2 / f θ , 1 = s 2 / s 1 f_{\theta,2}/f_{\theta,1}=s_2/s_1 f θ , 2 / f θ , 1 = s 2 / s 1 wherever that ratio is defined, and inherits every zero of s s s . To vary both components independently, we use several known sections with independent scalar coefficients,
f θ ( R ) = ∑ k = 1 K N s k ( R ) c θ , k ( R ) = S 0 ( R ) c θ ( R ) . (12) f_\theta(\mathbf{R})=\sum_{k=1}^{K_N}s_k(\mathbf{R})c_{\theta,k}(\mathbf{R})=S_0(\mathbf{R})c_\theta(\mathbf{R}). \tag*{(12)} f θ ( R ) = k = 1 ∑ K N s k ( R ) c θ , k ( R ) = S 0 ( R ) c θ ( R ) . (12)
The sections are stored as the columns of the 2 × K N 2\times K_N 2 × K N matrix S 0 S_0 S 0 , and every coefficient is complex, periodic, and permutation symmetric. A collection spanning the two-component space at every configuration is called a frame ; redundant columns are allowed [66 ] .
Correct transport alone does not determine a useful choice of frame. The kinetic energy differentiates the given functions as well as their neural coefficients. Applied to Eq. (12) , the product rule gives D i f θ = ( D i S 0 ) c θ − i S 0 ∇ i c θ D_i f_\theta=(D_iS_0)c_\theta-\mathrm{i}S_0\nabla_i c_\theta D i f θ = ( D i S 0 ) c θ − i S 0 ∇ i c θ . A rapidly varying frame therefore asks the coefficients to compensate for its variation when representing a low-energy state. We seek a frame that is smooth across the torus boundaries and distributes the required winding over the cell, with controlled spatial variation.
We also require uniform conditioning of its two physical components,
S 0 ( R ) S 0 ( R ) † = I 2 . (13) S_0(\mathbf{R})S_0(\mathbf{R})^\dagger=\mathbb{I}_2. \tag*{(13)} S 0 ( R ) S 0 ( R ) † = I 2 . (13)
This identity, called tightness [67 ] , says that the sections resolve both components at every configuration. Individual sections may vanish, but the collection always spans both components. Any smooth, symmetric target f ∗ f_* f ∗ with the prescribed boundary transport can then be reconstructed using a coefficient function
c ∗ ( R ) = S 0 ( R ) † f ∗ ( R ) . (14) c_*(\mathbf{R})=S_0(\mathbf{R})^\dagger f_*(\mathbf{R}). \tag*{(14)} c ∗ ( R ) = S 0 ( R ) † f ∗ ( R ) . (14)
These coefficients are periodic because the transport of S 0 S_0 S 0 and f ∗ f_* f ∗ cancels, and they are permutation-symmetric because both factors are symmetric. Since the coefficients are functions of the full configuration, a finite number of sections in a frame does not restrict the representable space of smooth targets. In practice, however, a finite neural network must approximate these coefficients, and the choice of frame can strongly affect the complexity of that approximation. We now construct a smooth, tight starting frame from single-particle orbitals and a center-of-mass frame.
The analytic frame and pair factor
Our goal in this section is to construct an initial variational ansatz for the many-semion state. To this end, we specify a smooth initial frame S 0 ( R ) S_0(\mathbf{R}) S 0 ( R ) compatible with boundary transport, and supplement it with an analytical pair-factor J η ( R ) J_\eta(\mathbf{R}) J η ( R ) that enforces the leading semion cusp at coincidence or collision.
To construct S 0 ( R ) S_0(\mathbf{R}) S 0 ( R ) , we begin by confactoring the boundary phase into single-particle and collective parts.
e i π ( N x i − X ) / L x = e i π N x i / L x e − i π X / L x . (15) e^{\mathrm{i}\pi(Nx_i-X)/L_x}=e^{\mathrm{i}\pi Nx_i/L_x}e^{-\mathrm{i}\pi X/L_x}. \tag*{(15)} e i π ( N x i − X ) / L x = e i π N x i / L x e − i π X / L x . (15)
Accordingly, we seek a factorized frame
S 0 ( R ) = C ( X , Y ) ⊗ p ( R ) , (16) S_0(\mathbf{R})=C(X,Y)\otimes p(\mathbf{R}), \tag*{(16)} S 0 ( R ) = C ( X , Y ) ⊗ p ( R ) , (16)
with X = ∑ j x j X=\sum_jx_j X = ∑ j x j and Y = ∑ j y j Y=\sum_jy_j Y = ∑ j y j . Here ⊗ \otimes ⊗ denotes the Kronecker product with p p p treated as a row vector: ( S 0 ) α , ( a , k ) = C α a p k (S_0)_{\alpha,(a,k)}=C_{\alpha a}p_k ( S 0 ) α , ( a , k ) = C α a p k , where α = 1 , 2 \alpha=1,2 α = 1 , 2 labels the transport component, a a a a column of C C C , and k k k a scalar function. The matrix C C C carries the collective, two-component transport, while the scalar functions p k p_k p k carry the single-particle magnetic phases.
The first phase factor involves particle i i i alone. It is the boundary phase of one particle in a uniform field of n ϕ = N / 2 n_\phi=N/2 n ϕ = N /2 flux quanta. This is the familiar average-field picture of anyons [14 ] , in which each particle sees the half quanta of the others smeared over the torus. The second factor depends only on the center of mass. It carries the remaining − 1 / 2 -1/2 − 1/2 flux quantum, together with both components and the anticommuting matrices. The center-of-mass matrix must therefore satisfy
C ( X + L x , Y ) = X C ( X , Y ) , (17) C(X+L_x,Y)=XC(X,Y), \tag*{(17)} C ( X + L x , Y ) = X C ( X , Y ) , (17)
C ( X , Y + L y ) = e − i π X / L x Z C ( X , Y ) , (18) C(X,Y+L_y)=e^{-\mathrm{i}\pi X/L_x}ZC(X,Y), \tag*{(18)} C ( X , Y + L y ) = e − i π X / L x Z C ( X , Y ) , (18)
and the scalar functions must obey, for each particle,
p k ( R + L x e ^ i ) = e i β x p k ( R ) , (19) p_k(\mathbf{R}+L_x\hat{\mathbf{e}}_i)=e^{\mathrm{i}\beta_x}p_k(\mathbf{R}), \tag*{(19)} p k ( R + L x e ^ i ) = e i β x p k ( R ) , (19)
p k ( R + L y e ^ i ) = e i β y + i π N x i / L x p k ( R ) . (20) p_k(\mathbf{R}+L_y\hat{\mathbf{e}}_i)=e^{\mathrm{i}\beta_y+\mathrm{i}\pi Nx_i/L_x}p_k(\mathbf{R}). \tag*{(20)} p k ( R + L y e ^ i ) = e i β y + i π N x i / L x p k ( R ) . (20)
The product then has exactly the transport (7).
These are precisely the boundary conditions of a symmetric many-body wavefunction of particles in the same uniform auxiliary field B ∗ = 2 π n ϕ / ν = π N / ν B_* = 2\pi n_\phi/\nu=\pi N/\nu B ∗ = 2 π n ϕ / ν = π N / ν , with twists β x , β y \beta_x,\beta_y β x , β y . For one particle, the y y y -cycle phase is e i β y + i B ∗ L x y e^{\mathrm{i}\beta_y+\mathrm{i}B_*L_xy} e i β y + i B ∗ L x y ; in the many-body function, moving particle i i i gives that same phase with x x x replaced by x i x_i x i . This holds for correlated states as well as product states. Symmetric products of single-particle magnetic orbitals therefore supply building blocks for the many-body scalars p k p_k p k . Landau-level wavefunctions provide a natural starting point for the orbital texture in Sec. III B, while the center-of-mass frame supplies the matrix-valued transport.
The following construction defines S 0 S_0 S 0 at the physical configuration R \mathbf{R} R . We subsequently allow the network to adapt this starting geometry in Sec. III C.
In this factorization the two transport components are carried by the center-of-mass frame. This parallels the fractional quantum Hall effect on a torus, where the topological degeneracy is a property of the center-of-mass motion [68 ] . We therefore need a two-component function of ( X , Y ) (X,Y) ( X , Y ) alone that obeys Eqs. (16) and never vanishes.
The magnetic phase in Eq. (16) corresponds to a uniform auxiliary field B C = − π / V B_C=-\pi/\mathcal{V} B C = − π / V , or − 1 / 2 -1/2 − 1/2 flux quantum through the L x × L y L_x\times L_y L x × L y cell. The two components accommodate this half flux: the extra sign from the magnetic phase around the cell is canceled by X Z = − Z X XZ=-ZX X Z = − Z X . Since an X X X lap swaps the components and X 2 = I 2 X^2=\mathbb{I}_2 X 2 = I 2 , two laps restore each component. We can therefore unfold the two-component problem into a scalar one on the doubled cell 2 L x × L y 2L_x\times L_y 2 L x × L y , carrying one flux quantum in magnitude.
As a smooth starting choice, take a lowest-Landau-level orbital on this doubled cell. Its magnetic length fixes the Gaussian width, w C 2 = 1 / ∣ B C ∣ = V / π w_C^2=1/|B_C|=\mathcal{V}/\pi w C 2 = 1/∣ B C ∣ = V / π . In Landau gauge, each orbital mode is a plane wave in X X X times a Gaussian in Y Y Y . A single Gaussian is not compatible with identifying Y Y Y and Y + L y Y+L_y Y + L y . We add its translates by every j L y jL_y j L y , with the corresponding magnetic phases, so crossing the boundary simply relabels the terms in the sum. This gives
F ( X , Y ) = ∑ j ∈ Z e − ( Y − j L y ) 2 / ( 2 w C 2 ) e − i π j X / L x . (21) F(X,Y)=\sum_{j\in\mathbb{Z}}e^{-(Y-jL_y)^2/(2w_C^2)}e^{-i\pi jX/L_x}. \tag*{(21)} F ( X , Y ) = j ∈ Z ∑ e − ( Y − j L y ) 2 / ( 2 w C 2 ) e − iπ j X / L x . (21)
Equivalently, this is a Jacobi theta function [69 ] times a nonzero Gaussian:
F ( X , Y ) = e − Y 2 / ( 2 w C 2 ) ϑ 3 ( π ( X + i Y ) 2 L x | τ C ) , (22) F(X,Y)=e^{-Y^2/(2w_C^2)}\vartheta_3\left(\frac{\pi(X+iY)}{2L_x}\,\middle|\,\tau_C\right), \tag*{(22)} F ( X , Y ) = e − Y 2 / ( 2 w C 2 ) ϑ 3 ( 2 L x π ( X + iY ) τ C ) , (22)
where τ C = i L y / ( 2 L x ) \tau_C=iL_y/(2L_x) τ C = i L y / ( 2 L x ) and ϑ 3 ( z ∣ τ ) = ∑ j ∈ Z e i π τ j 2 + 2 i j z \vartheta_3(z|\tau)=\sum_{j\in\mathbb{Z}}e^{i\pi\tau j^2+2ijz} ϑ 3 ( z ∣ τ ) = ∑ j ∈ Z e iπ τ j 2 + 2 ij z . Each term is 2 L x 2L_x 2 L x periodic in X X X . Reindexing the sum under Y → Y + L y Y\to Y+L_y Y → Y + L y gives F ( X , Y + L y ) = e − i π X / L x F ( X , Y ) F(X,Y+L_y)=e^{-i\pi X/L_x}F(X,Y) F ( X , Y + L y ) = e − iπ X / L x F ( X , Y ) . We fold this scalar orbital back into two components on the original cell,
Φ ( X , Y ) = 1 2 ( F ( X , Y ) F ( X + L x , Y ) ) . (23) \Phi(X,Y)=\frac{1}{\sqrt{2}}\begin{pmatrix}F(X,Y)\\F(X+L_x,Y)\end{pmatrix}. \tag*{(23)} Φ ( X , Y ) = 2 1 ( F ( X , Y ) F ( X + L x , Y ) ) . (23)
An X X X lap exchanges the entries, giving X X X . A Y Y Y lap multiplies the first entry by e − i π X / L x e^{-i\pi X/L_x} e − iπ X / L x and the second by its negative, giving Z Z Z and the common phase. Thus the folded orbital obeys Eqs. (16) exactly.
The zeros of ϑ 3 \vartheta_3 ϑ 3 occur at z = π ( m + 1 2 ) + π τ C ( n + 1 2 ) z=\pi(m+\frac{1}{2})+\pi\tau_C(n+\frac{1}{2}) z = π ( m + 2 1 ) + π τ C ( n + 2 1 ) for integers m , n m,n m , n [69 ] . Thus F F F has exactly one zero on the doubled cell, at ( X , Y ) = ( L x , L y / 2 ) (X,Y)=(L_x,L_y/2) ( X , Y ) = ( L x , L y /2 ) modulo ( 2 L x , L y ) (2L_x,L_y) ( 2 L x , L y ) . The two entries of Φ \Phi Φ sample points separated by L x L_x L x , so they cannot vanish together. Its normalized version φ = Φ / ∥ Φ ∥ \varphi=\Phi/\|\Phi\| φ = Φ/∥Φ∥ is therefore smooth everywhere.
One section still locks the ratio of components. Two “boosts”, which multiply by a unit-modulus phase and a constant matrix, map sections to sections:
C 0 = [ φ , B x φ , B y φ ] , C_0=[\varphi,B_x\varphi,B_y\varphi], C 0 = [ φ , B x φ , B y φ ] ,
B x = e i π X / L x Z , B y = e i π Y / L y X . (24) B_x=e^{i\pi X/L_x}Z,\qquad B_y=e^{i\pi Y/L_y}X. \tag*{(24)} B x = e iπ X / L x Z , B y = e iπ Y / L y X . (24)
For any unit spinor, C 0 C 0 † C_0C_0^\dagger C 0 C 0 † has eigenvalues exactly 1 1 1 and 2 2 2 (Supplemental Material, Sec. S3). Pointwise normalization makes the frame tight, C = ( C 0 C 0 † ) − 1 / 2 C 0 C=(C_0C_0^\dagger)^{-1/2}C_0 C = ( C 0 C 0 † ) − 1/2 C 0 , with C C † = I 2 CC^\dagger=\mathbb{I}_2 C C † = I 2 at every point.
Having constructed the center-of-mass matrix C C C , we now move on to construct the scalar functions p k ( R ) p_k(R) p k ( R ) . Guided by the magnetic boundary conditions in Eq. (17), we use two lowest-Landau-level (LLL) functions for the scalar factor, ϕ A \phi_A ϕ A and ϕ B \phi_B ϕ B , each carrying the full n ϕ n_\phi n ϕ flux quanta. We set β x = β y = 0 \beta_x=\beta_y=0 β x = β y = 0 for the calculations reported here, as specified in Sec. IV. Each has n ϕ n_\phi n ϕ vortices, points where its amplitude vanishes and its phase winds. We arrange the vortices of ϕ A \phi_A ϕ A in a near-regular array and obtain ϕ B \phi_B ϕ B by a translation with a phase that preserves the LLL and its torus boundary rules. The two vortex arrays interleave. Where one function vanishes the other remains nonzero, so the pair has no common zero. Supplemental Material, Sec. S5 gives the construction. Two functions here do not mean two flux quanta: as N N N increases, each function carries more vortices.
The locally normalized orbital vector is
v ( r ) = ( ϕ A ( r ) , ϕ B ( r ) ) ρ ( r ) , ρ = ∣ ϕ A ∣ 2 + ∣ ϕ B ∣ 2 . (25) \boldsymbol{v}(\boldsymbol{r})=\frac{(\phi_A(\boldsymbol{r}),\phi_B(\boldsymbol{r}))}{\sqrt{\rho(\boldsymbol{r})}},\qquad\rho=|\phi_A|^2+|\phi_B|^2. \tag*{(25)} v ( r ) = ρ ( r ) ( ϕ A ( r ) , ϕ B ( r )) , ρ = ∣ ϕ A ∣ 2 + ∣ ϕ B ∣ 2 . (25)
It has two entries at every even N ≥ 4 N\geq4 N ≥ 4 , and unit norm everywhere. Normalization uses the combined density, never division by a single orbital. The relative weights and phase define an orbital pseudospin direction n = v † σ v \boldsymbol{n}=\boldsymbol{v}^\dagger\boldsymbol{\sigma}\boldsymbol{v} n = v † σ v , where σ \boldsymbol{\sigma} σ denotes the three Pauli matrices. This is an auxiliary way of describing the two orbital amplitudes, rather than a physical particle spin. The direction lies on a unit sphere: a zero of ϕ A \phi_A ϕ A is its south pole and a zero of ϕ B \phi_B ϕ B its north pole. Between the vortices it changes smoothly. The sphere is covered n ϕ n_\phi n ϕ times as one traverses the torus; this winding is a skyrmion texture. Figure 2 shows the six-particle starting texture.
Figure 2. Analytic starting orbital texture for N = 6 N = 6 N = 6 . Color shows the orbital pseudospin component n z n_z n z , and arrows show ( n x , n y ) (n_x,n_y) ( n x , n y ) . Crosses and circles mark the three vortices of each orbital function. The disjoint zero sets give a smooth normalized orbital vector.
Three features motivate this compact choice. First, it carries exactly the scalar transport of Eq. (17). Second, the interleaved zeros make it smooth everywhere. Third, remaining within the LLL gives the smallest possible average twisting for that flux. The flux requires the orbital direction to wind a fixed number of times. The first-order relation between the x x x and y y y derivatives of LLL functions makes their gradient cost attain the lower bound set by this winding. More precisely, the phase-independent derivative cost is
g = ∑ μ = x , y ( ∥ ∂ μ v ∥ 2 − ∣ v † ∂ μ v ∣ 2 ) = 1 4 ∣ ∇ n ∣ 2 . (26) g=\sum_{\mu=x,y}\left(\lVert\partial_\mu\boldsymbol{v}\rVert^2-\lvert\boldsymbol{v}^\dagger\partial_\mu\boldsymbol{v}\rvert^2\right)=\frac{1}{4}\lvert\nabla\boldsymbol{n}\rvert^2. \tag*{(26)} g = μ = x , y ∑ ( ∥ ∂ μ v ∥ 2 − ∣ v † ∂ μ v ∣ 2 ) = 4 1 ∣ ∇ n ∣ 2 . (26)
This measures how quickly the orbital direction changes, after removing changes of its common phase. Its torus average is B ∗ = 2 π n ϕ / V B_* = 2\pi n_\phi/\mathcal{V} B ∗ = 2 π n ϕ / V , the auxiliary field strength, and its integral attains the flux lower bound of Belavin and Polyakov (Supplemental Material, Sec. S5) [70 ] . This property does not depend on the number of orbital functions. The natural length is the magnetic length ℓ ∗ = ( V / π N ) 1 / 2 \ell_*=(\mathcal{V}/\pi N)^{1/2} ℓ ∗ = ( V / π N ) 1/2 , proportional to the mean particle spacing. Interleaving the vortices also aims to spread this variation over the cell, rather than concentrating it into sharp regions.
The two orbital entries are distinct from the physical transport components, which come from C \mathbf{C} C . For N = 2 N = 2 N = 2 , the auxiliary flux is a single quantum and the LLL is one dimensional: every LLL function is proportional to the same orbital and vanishes at its zero. Two such functions would therefore leave ρ = 0 \rho= 0 ρ = 0 there and the normalized orbital vector undefined. We instead use a first excited Landau-level partner that is nonzero at the LLL zero, keeping ρ > 0 ( x r e f ‘ S u p p l e m e n t a l M a t e r i a l , S e c . S 5 ‘ ) . E v e r y t h i n g e l s e i s u n c h a n g e d , i n c l u d i n g t h e c o u n t o f $ 3 ( N + 1 ) = 9 \rho> 0 ({xref}`Supplemental Material, Sec. S5`). Everything else is unchanged, including the count of \$3(N + 1) = 9 ρ > 0 ( x r e f ‘ S u ppl e m e n t a l M a t er ia l , S ec . S 5‘ ) . E v er y t hin g e l se i s u n c han g e d , in c l u d in g t h eco u n t o f $3 ( N + 1 ) = 9 sections, so the same ansatz construction applies at every even N ≥ 2 N \ge2 N ≥ 2 .
Each particle samples a unit orbital vector v i = v ( r i ) \mathbf{v}_i = \mathbf{v}(\mathbf{r}_i) v i = v ( r i ) . To form an exchange-symmetric product of these vectors, we can place k = 0 , … , N k = 0,\ldots,N k = 0 , … , N particles in the second orbital v 1 \mathbf{v}_1 v 1 while keeping the rest of the particles in the first orbital v 0 \mathbf{v}_0 v 0 , leading to N + 1 N + 1 N + 1 occupation amplitudes. They can be obtained by multiplying a polynomial in a bookkeeping variable t t t ,
∏ i = 1 N [ v 0 ( r i ) + v 1 ( r i ) t ] = ∑ k = 0 N a k ( R ) t k , (27) \prod_{i=1}^{N} [v_0(r_i) + v_1(r_i)t] = \sum_{k=0}^{N} a_k(\mathbf{R})t^k, \tag*{(27)} i = 1 ∏ N [ v 0 ( r i ) + v 1 ( r i ) t ] = k = 0 ∑ N a k ( R ) t k , (27)
and applying the usual occupation-state normalization,
p ~ k = a k ( N k ) , N 2 = ∑ k = 0 N ∣ p ~ k ∣ 2 , p k = p ~ k N . (28) \widetilde{p}_k = \frac{a_k}{\sqrt{\binom{N}{k}}}, \qquad\mathcal{N}^2 = \sum_{k=0}^{N} \lvert\widetilde{p}_k\rvert^2, \qquad p_k = \frac{\widetilde{p}_k}{\mathcal{N}}. \tag*{(28)} p k = ( k N ) a k , N 2 = k = 0 ∑ N ∣ p k ∣ 2 , p k = N p k . (28)
Every term carries the same boundary phase, so each p k p_k p k is symmetric and obeys the scalar transport. The count is M N = N + 1 M_N = N + 1 M N = N + 1 . In pseudospin language, symmetrizing N N N spin-1 2 \frac{1}{2} 2 1 orbital vectors gives a collective spin N / 2 N/2 N /2 with these N + 1 N+1 N + 1 components. The individual orbital directions are its Majorana stars [71 ] .
The normalization N 2 \mathcal{N}^2 N 2 is the weight of the product in the symmetric subspace. It equals one for parallel orbital directions and falls as they spread over the sphere. It never vanishes: each factor in (27) is a nonzero polynomial, so their product is nonzero. A quantitative bound is ( N + 1 ) e − N ≤ N 2 ≤ 1 (N + 1)e^{-N} \le\mathcal{N}^2 \le1 ( N + 1 ) e − N ≤ N 2 ≤ 1 (Supplemental Material, Sec. S6). This proves regularity at each fixed N N N and gives an explicit size-dependent lower bound.
Combining the scalar amplitudes with the three COM sections gives
( S 0 ) a k ( R ) = C a ( X , Y ) p k ( R ) . (29) (S_0)_{ak}(\mathbf{R}) = C_a(X,Y)p_k(\mathbf{R}). \tag*{(29)} ( S 0 ) ak ( R ) = C a ( X , Y ) p k ( R ) . (29)
Here a = 1 , 2 , 3 a = 1,2,3 a = 1 , 2 , 3 and k = 0 , … , N k = 0,\ldots,N k = 0 , … , N , so K N = 3 ( N + 1 ) K_N = 3(N + 1) K N = 3 ( N + 1 ) physical sections suffice for the starting frame. Because ∑ k ∣ p k ∣ 2 = 1 \sum_k \lvert p_k\rvert^2 = 1 ∑ k ∣ p k ∣ 2 = 1 and C C † = I 2 CC^\dagger= \mathbb{I}_2 C C † = I 2 , S 0 S 0 † = I 2 S_0S_0^\dagger= \mathbb{I}_2 S 0 S 0 † = I 2 . The frame is smooth, tight, exchange symmetric, and has exact transport. The reconstruction in Eq. (14) therefore applies with this compact frame: configuration-dependent coefficients provide the remaining many-body freedom.
The analytic frame and smooth coefficient functions need not vanish when particles meet. We enforce the leading semion cusp with the pair factor
J η ( R ) = ∏ i < j ( d i j 2 d i j 2 + η 2 ) 1 / 4 , η > 0. (30) J_\eta(\mathbf{R}) = \prod_{i<j}\left(\frac{d_{ij}^2}{d_{ij}^2+\eta^2}\right)^{1/4}, \qquad\eta> 0. \tag*{(30)} J η ( R ) = i < j ∏ ( d ij 2 + η 2 d ij 2 ) 1/4 , η > 0. (30)
Here d i j d_{ij} d ij is the chord distance
d i j 2 = ∑ μ = x , y [ L μ π sin ( π ( r i μ − r j μ ) L μ ) ] 2 , (31) d_{ij}^2 = \sum_{\mu=x,y}\left[\frac{L_\mu}{\pi}\sin\left(\frac{\pi(r_{i\mu}-r_{j\mu})}{L_\mu}\right)\right]^2, \tag*{(31)} d ij 2 = μ = x , y ∑ [ π L μ sin ( L μ π ( r i μ − r j μ ) ) ] 2 , (31)
whose square root approaches the separation r i j r_{ij} r ij at short distance but, unlike the minimum-image distance, is smooth and periodic. Each pair factor behaves as ( r i j / η ) 1 / 2 (r_{ij}/\eta)^{1/2} ( r ij / η ) 1/2 when the pair is close and tends to a constant far apart. The exponent is fixed by the statistics. The range η \eta η only sets how fast the factor recovers. It is not an excluded radius and does not change H H H , and it may be fixed or optimized. Multiplying a smooth bounded frame expansion by this factor gives a state in the Friedrichs form domain. Supplemental Material, Sec. S2 gives the domain argument and discusses subdividing pair channels and their representation by additional analytic factors.
Learned geometry and neural amplitudes
The analytic frame and pair factor supply the required transport and collision behavior. We now let the network adapt the orbital geometry and learn the coefficients that combine the frame sections. Replacing S 0 S_0 S 0 by a learned frame G θ G_\theta G θ preserves the same transport and tightness. Adapting the geometry gives a finite network another way to express correlations, alongside learning the coefficients. The network assigns a real two-dimensional auxiliary displacement u θ , i ( R ) \mathbf{u}_{\theta,i}(R) u θ , i ( R ) to each particle and forms
q i = r i + u θ , i ( R ) , Q θ = ( q 1 , … , q N ) . (32) \mathbf{q}_i=\mathbf{r}_i+\mathbf{u}_{\theta,i}(R),\qquad Q_\theta=(\mathbf{q}_1,\ldots,\mathbf{q}_N). \tag*{(32)} q i = r i + u θ , i ( R ) , Q θ = ( q 1 , … , q N ) . (32)
These auxiliary displacements are smooth and periodic in every physical coordinate. They are also permutation equivariant : exchanging two particle labels exchanges their displacements. Thus particles in different local environments can sample different parts of the orbital texture without introducing distinguishable particle labels. The auxiliary coordinates play a role similar to backflow coordinates [72 ] : a particle’s orbital value responds to its neighbors. All sections use the same displacement map.
Simply using S 0 ( Q θ ) S_0(Q_\theta) S 0 ( Q θ ) would give the wrong y y y -cycle phase, because that phase depends on the x x x coordinates. The required correction has an explicit form. With y ˉ = N − 1 ∑ i y i \bar{y}=N^{-1}\sum_i y_i y ˉ = N − 1 ∑ i y i , define
φ θ ( R ) = − π N ν ∑ i ( y i − y ˉ ) u θ , i , x ( R ) , (33) \varphi_\theta(R)=-\frac{\pi N}{\nu}\sum_i(y_i-\bar{y})u_{\theta,i,x}(R), \tag*{(33)} φ θ ( R ) = − ν π N i ∑ ( y i − y ˉ ) u θ , i , x ( R ) , (33)
G θ ( R ) = e i φ θ ( R ) S 0 ( Q θ ) . (34) G_\theta(R)=e^{i\varphi_\theta(R)}S_0(Q_\theta). \tag*{(34)} G θ ( R ) = e i φ θ ( R ) S 0 ( Q θ ) . (34)
This phase realigns the orbital winding with the physical boundary rule. When particle i i i crosses the y y y boundary, the internally evaluated frame gains an extra phase π [ N u θ , i , x − ∑ j u θ , j , x ] / L x \pi[Nu_{\theta,i,x}-\sum_j u_{\theta,j,x}]/L_x π [ N u θ , i , x − ∑ j u θ , j , x ] / L x . The change in φ θ \varphi_\theta φ θ is exactly its negative. The x x x boundary rule needs no correction. Exchange symmetry is preserved because the sum in Eq. (33) is unchanged by relabeling particles. Supplemental Material, § 1.12 gives the short check.
The distinction between physical and auxiliary positions is essential. The pair factor J η ( R ) J_\eta(R) J η ( R ) remains evaluated at the physical positions, as does the statistical connection A i ( R ) A_i(R) A i ( R ) . Sampling and expectation-value integrals use the physical coordinates R R R with measure d R = ∏ i d 2 r i dR=\prod_i d^2r_i d R = ∏ i d 2 r i . Only the smooth frame is evaluated at Q θ Q_\theta Q θ . The internal map need not be invertible, and two particles may have the same auxiliary position. The analytic frame is smooth even there. Such an auxiliary coincidence neither imposes a physical collision nor forces a node. There is no change of integration coordinates or volume factor.
The learned frame remains tight since
G θ G θ † = S 0 ( Q θ ) S 0 ( Q θ ) † = I 2 . (35) G_\theta G_\theta^\dagger=S_0(Q_\theta)S_0(Q_\theta)^\dagger=I_2. \tag*{(35)} G θ G θ † = S 0 ( Q θ ) S 0 ( Q θ ) † = I 2 . (35)
The learned frame evaluates the same N + 1 N+1 N + 1 scalar amplitudes at Q θ Q_\theta Q θ , so its section count stays 3 ( N + 1 ) 3(N+1) 3 ( N + 1 ) throughout training. Zero displacement recovers the starting frame exactly. Learning therefore begins with the analytic orbital construction and can adapt it continuously while preserving transport and the ability to resolve both components. The frame derivative in the product rule now acts on G θ G_\theta G θ . It includes both the change of the internal coordinates and the explicit phase correction, with the physical connection A i ( R ) A_i(R) A i ( R ) . All are differentiated along with the coefficients. The starting orbitals have minimal twisting, while a learned deformation can increase or decrease their derivative cost. The variational energy, which includes every product-rule term, determines which deformations are useful. Supplemental Material, § 1.12 gives the explicit chain rule.
Combining the pair factor with the learned frame and neural coefficients completes the trial wavefunction,
Ψ θ ( R ) = J η ( R ) G θ ( R ) c θ ( R ) . (36) \Psi_\theta(R)=J_\eta(R)G_\theta(R)c_\theta(R). \tag*{(36)} Ψ θ ( R ) = J η ( R ) G θ ( R ) c θ ( R ) . (36)
The frame supplies smooth two-component transport, the coefficients learn amplitudes, phases, and nodes, and the pair factor supplies the leading collision behavior. Including this last factor in the product rule gives
D i Ψ θ = − i ( ∇ i J η ) G θ c θ + J η ( D i G θ ) c θ − i J η G θ ∇ i c θ . (37) D_i\Psi_\theta=-i(\nabla_iJ_\eta)G_\theta c_\theta+J_\eta(D_iG_\theta)c_\theta
-iJ_\eta G_\theta\nabla_i c_\theta. \tag*{(37)} D i Ψ θ = − i ( ∇ i J η ) G θ c θ + J η ( D i G θ ) c θ − i J η G θ ∇ i c θ . (37)
These choices enforce statistics, transport, and finite collision energy for every parameter value. The frame and coefficients share one neural encoder, as described below. Figure 3 shows the complete structure. Equation (36) is the building block for the momentum-projected state in Sec. III D.
Figure 3. Computational graph of the learned orbital-frame ansatz. (a) Periodic position and pair features pass through one shared attention encoder to give a hidden vector for each particle. (b) A particlewise readout produces displacements; their internal coordinates sample the analytic frame, and a common phase restores physical transport. The same hidden vectors are mean-pooled and augmented by the orbital weights ∣ p k ( Q θ ) ∣ 2 |p_k(Q_\theta)|^2 ∣ p k ( Q θ ) ∣ 2 before the complex coefficient readout. The pair factor uses the physical positions. Contracting the frame with its coefficients and multiplying by J η J_\eta J η gives the two-component amplitude. The final momentum projection coherently averages evaluations of this entire graph at collectively shifted positions, with shared parameters [Eq. (42)]. Blue blocks contain trainable maps; green blocks are prescribed operations.
For a target Ψ ∗ \Psi_* Ψ ∗ whose remainder f ∗ = Ψ ∗ / J η f_*=\Psi_*/J_\eta f ∗ = Ψ ∗ / J η is smooth, the reconstruction in Eq. (14) applies equally to the learned frame, with c ∗ = G θ † f ∗ c_* = G_\theta^\dagger f_* c ∗ = G θ † f ∗ . Its smoothness at contact depends on the target’s collision channels (Supplemental Material, Sec. S2).
The displacement and coefficients come from one permutation-equivariant attention network [73 ] . For each direction μ = x , y \mu=x,y μ = x , y and harmonic h = 1 , … , H f h=1,\ldots,H_f h = 1 , … , H f , collect sin ( 2 π h r μ / L μ ) \sin(2\pi h r_\mu/L_\mu) sin ( 2 π h r μ / L μ ) and cos ( 2 π h r μ / L μ ) \cos(2\pi h r_\mu/L_\mu) cos ( 2 π h r μ / L μ ) into γ ( r ) \gamma(r) γ ( r ) . Particle i i i receives features of its own position and its separations from the others:
f i i n = [ γ ( r i ) , 1 N − 1 ∑ j ≠ i γ ( r i − r j ) ] , (38) \mathbf{f}_i^{\mathrm{in}}=\left[\gamma(\mathbf{r}_i),\frac{1}{N-1}\sum_{j\ne i}\gamma(\mathbf{r}_i-\mathbf{r}_j)\right], \tag*{(38)} f i in = γ ( r i ) , N − 1 1 j = i ∑ γ ( r i − r j ) , (38)
h i ( 0 ) = tanh ( W i n f i i n + b i n ) . \mathbf{h}_i^{(0)}=\tanh\left(W_{\mathrm{in}}\mathbf{f}_i^{\mathrm{in}}+\mathbf{b}_{\mathrm{in}}\right). h i ( 0 ) = tanh ( W in f i in + b in ) .
Square brackets denote concatenation, and tanh \tanh tanh acts entrywise. All matrices W W W and biases b \mathbf{b} b are learned. The Fourier inputs make every output periodic; the hidden vectors h i ( 0 ) \mathbf{h}_i^{(0)} h i ( 0 ) have width w w w and describe each particle’s environment.
Each attention layer ℓ = 1 , … , L \ell=1,\ldots,L ℓ = 1 , … , L refines these descriptions. Head m = 1 , … , H a m=1,\ldots,H_a m = 1 , … , H a forms query, key, and value vectors q i ( ℓ , m ) \mathbf{q}_i^{(\ell,m)} q i ( ℓ , m ) , k i ( ℓ , m ) \mathbf{k}_i^{(\ell,m)} k i ( ℓ , m ) , and v i ( ℓ , m ) \mathbf{v}_i^{(\ell,m)} v i ( ℓ , m ) by affine maps of h i ( ℓ − 1 ) \mathbf{h}_i^{(\ell-1)} h i ( ℓ − 1 ) . These are auxiliary features, distinct from the internal particle positions q i \mathbf{q}_i q i . With head width d h = w / H a d_h=w/H_a d h = w / H a , the update is
A i j ( ℓ , m ) = exp [ q i ( ℓ , m ) ⋅ k j ( ℓ , m ) / d h ] ∑ j ′ = 1 N exp [ q i ( ℓ , m ) ⋅ k j ′ ( ℓ , m ) / d h ] , o i ( ℓ , m ) = ∑ j = 1 N A i j ( ℓ , m ) v j ( ℓ , m ) , h i ( ℓ ) = tanh [ h i ( ℓ − 1 ) + W O ( ℓ ) o i ( ℓ ) + b O ( ℓ ) ] . (39) \begin{aligned}
A_{ij}^{(\ell,m)}&=\frac{\exp\left[q_i^{(\ell,m)}\cdot k_j^{(\ell,m)}/\sqrt{d_h}\right]}{\sum_{j'=1}^{N}\exp\left[q_i^{(\ell,m)}\cdot k_{j'}^{(\ell,m)}/\sqrt{d_h}\right]},\\
o_i^{(\ell,m)}&=\sum_{j=1}^{N}A_{ij}^{(\ell,m)}v_j^{(\ell,m)},\\
h_i^{(\ell)}&=\tanh\left[h_i^{(\ell-1)}+W_O^{(\ell)}o_i^{(\ell)}+b_O^{(\ell)}\right]. \tag*{(39)}
\end{aligned} A ij ( ℓ , m ) o i ( ℓ , m ) h i ( ℓ ) = ∑ j ′ = 1 N exp [ q i ( ℓ , m ) ⋅ k j ′ ( ℓ , m ) / d h ] exp [ q i ( ℓ , m ) ⋅ k j ( ℓ , m ) / d h ] , = j = 1 ∑ N A ij ( ℓ , m ) v j ( ℓ , m ) , = tanh [ h i ( ℓ − 1 ) + W O ( ℓ ) o i ( ℓ ) + b O ( ℓ ) ] . (39)
The weights sum to one over j j j , so each head returns a learned weighted average over particles. Here o i ( ℓ ) o_i^{(\ell)} o i ( ℓ ) concatenates all head outputs. The last line adds their contribution to the previous hidden vector. Parameters are shared across particles, with no particle-index encoding; relabeling particles therefore only relabels the hidden vectors.
A shared readout converts each final hidden vector into a displacement,
u θ , i , μ = s L μ tanh [ ( W u h i ( L ) + b u ) μ ] . (40) u_{\theta,i,\mu}=sL_\mu\tanh\left[(W_u h_i^{(L)}+b_u)_\mu\right]. \tag*{(40)} u θ , i , μ = s L μ tanh [ ( W u h i ( L ) + b u ) μ ] . (40)
The dimensionless scale s s s bounds each component by s L μ sL_\mu s L μ . We then evaluate the analytic frame at Q θ Q_\theta Q θ and apply the phase in Eq. (33) . Initializing W u W_u W u and b u b_u b u to zero recovers the analytic starting frame exactly.
The coefficient readout uses the whole configuration. Mean pooling gives a permutation-invariant hidden vector, augmented by the deformed frame’s orbital weights:
h ˉ = 1 N ∑ i h i ( L ) , \bar{h}=\frac{1}{N}\sum_i h_i^{(L)}, h ˉ = N 1 i ∑ h i ( L ) ,
g θ = tanh { W g [ h ˉ , { ∣ p k ( Q θ ) ∣ 2 } k = 0 N ] + b g } , c θ = ( W R g θ + b R ) + i ( W I g θ + b I ) . (41) \begin{aligned}
g_\theta&=\tanh\left\{W_g\left[\bar{h},\{|p_k(Q_\theta)|^2\}_{k=0}^{N}\right]+b_g\right\},\\
c_\theta&=(W_Rg_\theta+b_R)+\mathrm{i}(W_Ig_\theta+b_I). \tag*{(41)}
\end{aligned} g θ c θ = tanh { W g [ h ˉ , { ∣ p k ( Q θ ) ∣ 2 } k = 0 N ] + b g } , = ( W R g θ + b R ) + i ( W I g θ + b I ) . (41)
Here brackets denote vector concatenation. The pooled hidden layer has learned real weights W g ∈ R w × ( N + 1 ) W_g\in\mathbb{R}^{w\times(N+1)} W g ∈ R w × ( N + 1 ) and bias b g ∈ R w b_g\in\mathbb{R}^{w} b g ∈ R w . The learned real matrices were W R , W I ∈ R K N × w W_R,W_I\in\mathbb{R}^{K_N\times w} W R , W I ∈ R K N × w and biases b R , b I ∈ R K N b_R,b_I\in\mathbb{R}^{K_N} b R , b I ∈ R K N produce the real and imaginary parts, respectively, of the K N = 3 ( N + 1 ) K_N=3(N+1) K N = 3 ( N + 1 ) complex coefficients. The orbital weights are symmetric and periodic because their common boundary phase cancels. They tell the network which orbital occupations dominate at the internal configuration. The complex coefficients allow interference and movable nodes. Absolute positions are retained, so translation invariance is not imposed; exact transport comes from the learned frame for every parameter value. The calculation is sequential: encode the physical positions, move the frame, then mix its sections. Supplemental Material, Sec. S7 gives the full layer equations, dimensions, and a pedagogical account of the shared encoder and its two readouts.
Momentum selection and computational cost
Each pair connection is odd, so ∑ i A i = 0 \sum_i \mathbf{A}_i=0 ∑ i A i = 0 . Consequently, the total momentum P = − i ℏ ∑ i ∇ i \mathbf{P}=-i\hbar\sum_i\nabla_i P = − i ℏ ∑ i ∇ i is conserved. The internal coordinates adapt the orbital geometry, but do not by themselves give a definite total momentum. A learned state can still depend on where the whole configuration sits on the torus. Such dependence mixes momentum sectors, called momentum leakage here. Its energy contains a nonnegative center-of-mass part,
E C O M [ Ψ ] = κ N ∫ ∥ ∑ i D i Ψ ∥ 2 d R ∫ ∥ Ψ ∥ 2 d R , (42) E_{\mathrm{COM}}[\Psi]=\frac{\kappa}{N}\frac{\int\left\|\sum_iD_i\Psi\right\|^2d\mathbf{R}}{\int\left\|\Psi\right\|^2d\mathbf{R}}, \tag*{(42)} E COM [ Ψ ] = N κ ∫ ∥ Ψ ∥ 2 d R ∫ ∥ ∑ i D i Ψ ∥ 2 d R , (42)
which equals κ ∣ K ∣ 2 / N \kappa\lvert\mathbf{K}\rvert^2/N κ ∣ K ∣ 2 / N in a state of momentum ℏ K \hbar\mathbf{K} ℏ K (Supplemental Material, Sec. S3). The remainder is the relative energy. Reporting both separates improved correlations from residual collective motion. We suppress unwanted momentum components by averaging translated copies of the complete learned wavefunction with the phases appropriate to the target momentum.
Set β x = β y = 0 \beta_x=\beta_y=0 β x = β y = 0 , as in our calculations. Write R + a = ( r 1 + a , … , r N + a ) \mathbf{R}+\mathbf{a}=(\mathbf{r}_1+\mathbf{a},\ldots,\mathbf{r}_N+\mathbf{a}) R + a = ( r 1 + a , … , r N + a ) for a collective translation. A momentum eigenstate transforms by Ψ K ( R + a ) = e i K ⋅ a Ψ K ( R ) \Psi_{\mathbf{K}}(\mathbf{R}+\mathbf{a})=e^{i\mathbf{K}\cdot\mathbf{a}}\Psi_{\mathbf{K}}(\mathbf{R}) Ψ K ( R + a ) = e i K ⋅ a Ψ K ( R ) . The continuous projection onto K = 0 \mathbf{K}=0 K = 0 is thus
( P 0 Ψ θ ) ( R ) = 1 V ∫ T 2 Ψ θ ( R + a ) d 2 a . (43) (\mathcal{P}_0\Psi_\theta)(\mathbf{R})=\frac{1}{\mathcal{V}}\int_{\mathbb{T}^2}\Psi_\theta(\mathbf{R}+\mathbf{a})\,d^2\mathbf{a}. \tag*{(43)} ( P 0 Ψ θ ) ( R ) = V 1 ∫ T 2 Ψ θ ( R + a ) d 2 a . (43)
All copies add coherently in the zero-momentum sector; the oscillating phases of other sectors average to zero. A nonzero target K ∗ \mathbf{K}_* K ∗ is selected by inserting e − i K ∗ ⋅ a e^{-i\mathbf{K}_*\cdot\mathbf{a}} e − i K ∗ ⋅ a in the average. The zero-momentum choice removes the center-of-mass energy in Eq. (42) . The relative-motion energies can also differ between sectors, so the lowest sector is identified by comparing their variational energies, rather than from the center-of-mass term alone.
In practice we use an M × M M\times M M × M grid of collective shifts,
a a b = ( a L x M , b L y M ) , a , b = 0 , … , M − 1 , (44) \mathbf{a}_{ab}=\left(\frac{aL_x}{M},\frac{bL_y}{M}\right),\qquad a,b=0,\ldots,M-1, \tag*{(44)} a ab = ( M a L x , M b L y ) , a , b = 0 , … , M − 1 , (44)
and train the projected ansatz
Ψ θ ( M ) ( R ) = J η ( R ) M 2 ∑ a , b = 0 M − 1 G θ ( R + a a b ) c θ ( R + a a b ) . (45) \Psi_\theta^{(M)}(\mathbf{R})=\frac{J_\eta(\mathbf{R})}{M^2}\sum_{a,b=0}^{M-1}G_\theta(\mathbf{R}+\mathbf{a}_{ab})c_\theta(\mathbf{R}+\mathbf{a}_{ab}). \tag*{(45)} Ψ θ ( M ) ( R ) = M 2 J η ( R ) a , b = 0 ∑ M − 1 G θ ( R + a ab ) c θ ( R + a ab ) . (45)
The pair factor comes outside because it depends only on separations. Every shifted copy reevaluates the encoder, displacement, orbital frame, phase correction, and
coefficients. We average amplitudes, so interference removes unwanted momenta before computing the density or energy. Sampling and differentiation use this projected state throughout training.
The finite grid removes every sector except those with both integer momentum labels divisible by M M M , where K μ L μ = 2 π n μ K_\mu L_\mu=2\pi n_\mu K μ L μ = 2 π n μ . It therefore suppresses leakage while retaining possible higher-momentum copies. To suppress the dominant residual momentum sectors, we choose the smallest M ≥ 2 M\geq2 M ≥ 2 that shares no factor with N / 2 N/2 N /2 : M = 2 , 3 , 2 , 3 , 5 , 5 M=2,3,2,3,5,5 M = 2 , 3 , 2 , 3 , 5 , 5 for N = 2 , 4 , 6 , 8 , 10 , 12 , 16 N=2,4,6,8,10,12,16 N = 2 , 4 , 6 , 8 , 10 , 12 , 16 . The residual center-of-mass energy, and refinement of the shift grid, check the sector selection. All collective translates have the same particle boundary rules: separations and N x i − X N x_i-X N x i − X are unchanged by a collective shift. Their sum preserves the exact transport, exchange symmetry, and physical collision behavior. The tightness identity applies to each learned frame inside the sum; projection is performed on the whole state, so it needs no separately normalized projected frame. Projected-state derivatives are obtained by differentiating the coherent shift sum itself, including the physical pair factor.
The center-of-mass energy provides a direct bound on residual momentum weight. If the evaluation grid is Q × Q Q\times Q Q × Q , every surviving nonzero momentum on the square has magnitude at least 2 π Q / L 2\pi Q/L 2 π Q / L . For normalized weights w n w_n w n in the orthogonal momentum sectors, E C O M = ( κ / N ) ∑ n w n ∣ K n ∣ 2 E_{\mathrm{COM}}=(\kappa/N)\sum_n w_n\lvert\mathbf{K}_n\rvert^2 E COM = ( κ / N ) ∑ n w n ∣ K n ∣ 2 . Each nonzero sector therefore contributes at least κ ( 2 π Q / L ) 2 / N \kappa(2\pi Q/L)^2/N κ ( 2 π Q / L ) 2 / N per unit weight, giving the bound on their total norm fraction,
w l e a k ≤ N E C O M κ ( 2 π Q / L ) 2 . (46) w_{\mathrm{leak}}\leq\frac{NE_{\mathrm{COM}}}{\kappa(2\pi Q/L)^2}. \tag*{(46)} w leak ≤ κ ( 2 π Q / L ) 2 N E COM . (46)
We use Q = 2 M Q=2M Q = 2 M for the reported energies and observables. This diagnostic complements the change in energy when the evaluation grid is refined.
The adopted pseudospin-texture frame has M N = N + 1 M_N=N+1 M N = N + 1 scalar amplitudes and K N = 3 ( N + 1 ) K_N=3(N+1) K N = 3 ( N + 1 ) physical sections at all even N ≥ 4 N\geq4 N ≥ 4 . The two-orbital symmetric products are built by polynomial multiplication with a number of operations proportional to N 2 N^2 N 2 . At fixed series accuracy and aspect ratio, the orbital construction also has quadratic forward cost. At fixed neural width and depth, attention has quadratic particle-number dependence because it couples every pair of particles. The movement head adds two real outputs per particle and 2 ( w + 1 ) 2(w+1) 2 ( w + 1 ) parameters at hidden width w w w ; it requires no extra encoder. For example, the N = 6 N=6 N = 6 two-orbital model has twenty-one complex coefficient outputs and 41 836 real parameters at width 64, including the 130 movement parameters.
An M × M M\times M M × M projection evaluates M 2 M^2 M 2 shifted copies with shared parameters. This increases evaluation work, not the parameter count.
Variational Monte Carlo
For the numerical results we use the unit square, L x = L y = 1 L_x = L_y = 1 L x = L y = 1 , β x = β y = 0 \beta_x = \beta_y = 0 β x = β y = 0 , and κ = 1 \kappa= 1 κ = 1 . Energies are quoted in units of κ / ( L x L y ) \kappa/(L_xL_y) κ / ( L x L y ) . The Hamiltonian has no intrinsic length scale, so at fixed N N N and aspect ratio its energies scale as κ / ( L x L y ) \kappa/(L_xL_y) κ / ( L x L y ) ; the intensive length scale is set by the density n = N / V n = N/\mathcal{V} n = N / V .
We minimize the normalized kinetic-energy expectation of the full complex two-component wavefunction,
E [ Ψ ] = ⟨ Ψ ∣ H ∣ Ψ ⟩ ⟨ Ψ ∣ Ψ ⟩ , E 0 = inf Ψ ≠ 0 E [ Ψ ] , (47) E[\Psi] = \frac{\langle\Psi|H|\Psi\rangle}{\langle\Psi|\Psi\rangle}, \qquad E_0 = \inf_{\Psi\ne0} E[\Psi], \tag*{(47)} E [ Ψ ] = ⟨ Ψ∣Ψ ⟩ ⟨ Ψ∣ H ∣Ψ ⟩ , E 0 = Ψ = 0 inf E [ Ψ ] , (47)
where the infimum runs over states satisfying the specified transport, exchange symmetry, and collision condition. Metropolis sampling [74 ] uses the Born density ∥ Ψ M ∥ 2 \lVert\Psi^{\mathcal{M}}\rVert^2 ∥ Ψ M ∥ 2 , with all learned factors reevaluated inside the translation sum. Writing Ψ = J η f \Psi= J_\eta f Ψ = J η f and u = ln J η u = \ln J_\eta u = ln J η , integration by parts gives the first-derivative estimator
E [ Ψ ] = ⟨ e W ⟩ ∥ Ψ ∥ 2 , E[\Psi] = \left\langle e_W \right\rangle_{\lVert\Psi\rVert^2}, E [ Ψ ] = ⟨ e W ⟩ ∥ Ψ ∥ 2 ,
e W = κ ∑ i [ ∥ D i f ∥ 2 ∥ f ∥ 2 − ∇ i 2 u − ∣ ∇ i u ∣ 2 ] . (48) e_W = \kappa\sum_i \left[\frac{\lVert D_i f\rVert^2}{\lVert f\rVert^2} - \nabla_i^2 u - \lvert\nabla_i u\rvert^2\right]. \tag*{(48)} e W = κ i ∑ [ ∥ f ∥ 2 ∥ D i f ∥ 2 − ∇ i 2 u − ∣ ∇ i u ∣ 2 ] . (48)
For a close pair at separation r r r , the cusp J η ∝ r 1 / 2 J_\eta\propto r^{1/2} J η ∝ r 1/2 gives u ≃ 1 2 ln r u \simeq\frac{1}{2}\ln r u ≃ 2 1 ln r . The pair contributions to ∣ A i ∣ 2 \lvert\mathcal{A}_i\rvert^2 ∣ A i ∣ 2 and ∣ ∇ i u ∣ 2 \lvert\nabla_i u\rvert^2 ∣ ∇ i u ∣ 2 are then both 1 / ( 4 r 2 ) 1/(4r^2) 1/ ( 4 r 2 ) . They enter Eq. (48) with opposite signs, so the leading 1 / r 2 1/r^2 1/ r 2 terms cancel. We combine these terms analytically before numerical evaluation to avoid subtracting large, nearly equal quantities. The energy gradient differentiates both numerator and normalization, including the displacement and complex coefficient parameters. Updates use Kronecker-factored approximate curvature (K-FAC) [57 , 75 ] .
The training protocol shares the attention architecture and optimizer, with lengths set by the mean particle spacing and the projection grid chosen as in Sec. III D. Energies and observables are measured in separate VMC runs with frozen parameters and independent random streams. We estimate statistical errors by blocking [76 ] and examine the energy tails. Supplemental Material, Sec. S9, gives the sampling moves, gradient formulation, hyperparameters, and error analysis. All reported production states are trained through P M \mathcal{P}_M P M and evaluated through the finer P 2 M \mathcal{P}_{2M} P 2 M filter, using the same frozen checkpoint. The width and walker count of each state are specified in § 1.4 and VI.
Figure Figure 4 shows the training energies of the selected states. From the random start, most of the energy reduction occurs within the first 10 3 10^3 1 0 3 updates; a slower decline follows and is still visible at the end of training for the larger systems. The finer evaluation filter lowers the energies further, most noticeably at the larger sizes.
Figure 4. Training energies of the selected states, per particle in units of the average-field value π n κ \pi n\kappa π nκ (Sec. VI). Lines show the walker average of e W e_W e W for the state projected through the training filter P M \mathcal{P}_M P M , averaged over logarithmically spaced update intervals; the N = 16 N = 16 N = 16 state is trained for 6 × 10 4 6 \times10^4 6 × 1 0 4 updates. Open circles are the frozen P 2 M \mathcal{P}_{2M} P 2 M energies of Tables I and II.
Validation against exact solutions and diagonalization
We validate the representation at N = 2 , 4 , 6 N = 2, 4, 6 N = 2 , 4 , 6 before studying the larger systems. For two semions, center-of-mass and relative motion separate, and an exact solution in a convergent Bessel series gives the ground energy (Supplemental Material, Sec. S8). For N = 4 , 6 N = 4, 6 N = 4 , 6 we use continuum exact diagonalization (ED) in the zero-momentum sector. Symmetric Fourier configurations with total squared integer wavenumber at most Λ \Lambda Λ multiply fixed Gaussian-periodized two-component sections and a regular collision factor. A collective translation average selects total momentum zero. The expansion coefficients follow from a generalized eigenvalue problem for the continuum kinetic and overlap matrices. Supplemental Material, § 1.18 , gives the basis, the additional pair channel used at N = 4 N = 4 N = 4 , and the independent integration checks.
With exact matrix integrals, finite-basis ED gives variational upper bounds. Numerical integration introduces a separate error, assessed by evaluating frozen eigenvectors on independent quadratures. We therefore compare the neural energies with the exact two-particle result and the best available finite-cutoff estimates at N = 4 , 6 N = 4, 6 N = 4 , 6 , retaining their different levels of numerical precision.
Table Table I shows close agreement with the exact two-semion energy and neural energies below the available finite-cutoff ED estimates at N = 4 , 6 N = 4, 6 N = 4 , 6 . The exact solution also tests the pair-correlation and current estimators (Supplemental Material, Sec. S8).
Doubling the momentum-filter grid leaves the N = 4 N = 4 N = 4 energy unchanged within statistical errors. At N = 6 N = 6 N = 6 , it lowers both the total and center-of-mass energies. The total reduction is larger, showing that the removed momentum sectors also had higher relative-motion energy.
Correlations and currents beyond exact diagonalization
We now apply the representation at N = 8 , 10 , 12 , 16 N=8,10,12,16 N = 8 , 10 , 12 , 16 , beyond the reach of our continuum diagonalization. All states use the learned two-orbital frame, a random start, and 2 × 10 4 2\times10^{4} 2 × 1 0 4 updates through P M \mathcal{P}_M P M (6 × 10 4 6\times10^{4} 6 × 1 0 4 at N = 16 N=16 N = 16 ), followed by frozen evaluation through P 2 M \mathcal{P}_{2M} P 2 M . We use width 64 at N = 8 , 10 , 12 N=8,10,12 N = 8 , 10 , 12 and width 256 at N = 16 N=16 N = 16 . Table II specifies the states used for all observables in this section. Smaller systems enter the comparisons of size dependence.
The average-field estimates provide an energy scale. Smearing the statistical flux gives flux density π n \pi n π n . Starting from bosons gives E / N = π n κ E/N=\pi n\kappa E / N = π nκ ; starting from fermions and filling two Landau levels gives E / N = 2 π n κ E/N=2\pi n\kappa E / N = 2 π nκ [14 ] . The dimensionless ratio E / ( N π n κ ) E/(N\pi n\kappa) E / ( N π nκ ) lies between these values at every size Figure 5 (a).
Figure 5. Particle-number dependence of the frozen variational states, with mean spacing a = ( V / N ) 1 / 2 a=(\mathcal{V}/N)^{1/2} a = ( V / N ) 1/2 . All states are evaluated through P 2 M \mathcal{P}_{2M} P 2 M ; N = 16 N=16 N = 16 uses width 256. (a) Energy in average-field units, with and without the center-of-mass part; point labels give N N N and error bars show two statistical standard errors. (b) Radial pair correlation at N = 4 N=4 N = 4 –16 16 16 ; the logarithmic inset shows the contact law for N = 8 , 10 , 12 , 16 N=8,10,12,16 N = 8 , 10 , 12 , 16 . (c) Pair relative angular momentum, with two-standard-error bands and the exact two-semion result. (d) Shell-averaged static structure factor at N = 4 N=4 N = 4 –16 16 16 . The dashed line fits the N ≥ 8 N\geq8 N ≥ 8 points with ∣ k ∣ a < 3.6 |k|a<3.6 ∣ k ∣ a < 3.6 .
Finer projection lowers the energies at the larger sizes. Increasing the network width and extending training further improve the N = 16 N=16 N = 16 state. The residual center-of-mass energies in Table II bound the remaining momentum weight through Eq. (46) .
Pair correlations and chiral motion
We normalize the pair correlation to one for independent uniform positions,
g 2 ( d ) = V N ( N − 1 ) ⟨ ∑ i ≠ j δ T 2 ( d − r i + r j ) ⟩ . (49) g_2(d)=\frac{\mathcal{V}}{N(N-1)}\left\langle\sum_{i\ne j}\delta_{\mathrm{T}^2}(d-r_i+r_j)\right\rangle. \tag*{(49)} g 2 ( d ) = N ( N − 1 ) V ⟨ i = j ∑ δ T 2 ( d − r i + r j ) ⟩ . (49)
Here δ T 2 \delta_{\mathrm{T}^2} δ T 2 is the periodic delta function and the expectation sums both physical components. Its integral over the separation cell is V \mathcal{V} V . Each beyond-ED state has a
Table 1. Validation on the unit square at β x = β y = 0 \beta_x=\beta_y=0 β x = β y = 0 , in units of κ / ( L x L y ) \kappa/(L_xL_y) κ / ( L x L y ) . The neural states are trained through P M \mathcal{P}_M P M for 2 × 10 4 2\times10^{4} 2 × 1 0 4 updates and evaluated through P 2 M \mathcal{P}_{2M} P 2 M , using width 64 and 2048 walkers. Parentheses give one blocked standard error from 4000 frozen-parameter measurements. The N = 4 N=4 N = 4 ED entry is a fitted finite-basis Ritz value; the N = 6 N=6 N = 6 entry is the mean of two independent positive-form evaluations of a frozen ED state. ED convergence and integration checks are given in Supplemental Material, Sec. S10.
pronounced correlation hole despite the absence of a repulsive potential Figure 6 . At contact, the ℓ = 0 \ell=0 ℓ = 0 semion channel gives g 2 ∝ r 2 α = r g_2\propto r^{2\alpha}=r g 2 ∝ r 2 α = r , as imposed by the cusp factor Figure 5 (b), inset. The optimized profiles determine the crossover out of the contact regime and the extent of the correlation hole.
Figure 6. Pair correlations and currents beyond ED. Color maps show g 2 ( d ) g_2(\mathbf{d}) g 2 ( d ) for the selected N = 8 , 10 , 12 , 16 N = 8, 10, 12, 16 N = 8 , 10 , 12 , 16 states, on their full minimum-image separation cells in spacing units. Arrows show the measured joint pair current; every second map bin is displayed. Arrow lengths are proportional to ln ( 1 + j / 0.05 ) \ln(1 + j/0.05) ln ( 1 + j /0.05 ) , where j = ∣ J 2 ∣ / [ ℏ / ( m a ) ] j = |\mathcal{J}_2|/[\hbar/(ma)] j = ∣ J 2 ∣/ [ ℏ/ ( ma )] , using the same scale in all panels. This logarithmic scale makes weak currents visible while preserving their directions. Dashed circles mark the first zero of the radially averaged L ( r ) L(r) L ( r ) . No rotation or reflection averaging is applied.
In spacing units the short-distance profiles are similar. At intermediate distances the size dependence is clearer Figure 5 (b). The first maximum is resolved inside r < L / 2 r<L/2 r < L /2 from N = 10 N=10 N = 10 onward, near 1.3 1.3 1.3 –1.4 a 1.4a 1.4 a , and becomes weaker as particle number increases. The modest maxima and suppressed contact probability describe weak spatial order.
The covariant current measures the motion accompanying these correlations. At a configuration R \mathbf{R} R , the velocity of particle i i i is
v i ( R ) = ℏ m Re [ Ψ † D i Ψ ] ∣ Ψ ∣ 2 . (50) v_i(\mathbf{R})=\frac{\hbar}{m}\frac{\operatorname{Re}[\Psi^\dagger D_i\Psi]}{|\Psi|^2}. \tag*{(50)} v i ( R ) = m ℏ ∣Ψ ∣ 2 Re [ Ψ † D i Ψ ] . (50)
The derivative uses the physical positions and the statistical connection, and the numerator sums both components. Averaging v i − v j v_i-v_j v i − v j at a fixed pair separation removes common drift. Its azimuthal part defines the pair relative angular momentum,
L ( r ) = m 2 r ⟨ ( v i − v j ) ⋅ e ^ φ ⟩ r . (51) L(r)=\frac{m}{2}r\left\langle(\mathbf{v}_i-\mathbf{v}_j)\cdot\hat{\mathbf{e}}_\varphi\right\rangle_r . \tag*{(51)} L ( r ) = 2 m r ⟨ ( v i − v j ) ⋅ e ^ φ ⟩ r . (51)
where e ^ φ \hat{\mathbf{e}}_\varphi e ^ φ points counterclockwise and m / 2 m/2 m /2 is the reduced mass. The joint pair current is J 2 ( d ) = g 2 ( d ) v r e l ( d ) \boldsymbol{\mathcal{J}}_2(d)=g_2(d)\mathbf{v}_{\mathrm{rel}}(d) J 2 ( d ) = g 2 ( d ) v rel ( d ) . For an isolated close pair, L → ( ℓ + α ) ℏ = ℏ / 2 L\to(\ell+\alpha)\hbar=\hbar/2 L → ( ℓ + α ) ℏ = ℏ/2 . The sign records the chosen semion chirality: time reversal conjugates the exchange phase + i +i + i to − i -i − i and reverses the current.
All four beyond-ED states counter-rotate at larger separation Figure 5 (c). The first zero lies near 0.74 0.74 0.74 –0.76 a 0.76a 0.76 a , followed by a negative minimum near r = a r=a r = a . The reversal also occurs at N = 4 , 6 N=4,6 N = 4 , 6 , whereas the exact two-semion current remains positive. As particle number increases, the reversal remains near the mean spacing while moving inward relative to the box size. This supports a density-scale interpretation of the screening of the contact circulation.
Table 2. States beyond ED used for the energy and observable results. Training uses P M \mathcal{P}_M P M and frozen evaluation uses P 2 M \mathcal{P}_{2M} P 2 M throughout. w w w is the hidden width; the walker count is 2048 at N = 8 , 10 N=8,10 N = 8 , 10 and 1024 at N = 12 , 16 N=12,16 N = 12 , 16 . Energies are in units of κ / ( L x L y ) \kappa/(L_xL_y) κ / ( L x L y ) ; parentheses are one blocked standard error. The last column uses the density n = N / L 2 n=N/L^2 n = N / L 2 .
Density fluctuations
For ρ k = ∑ i e i k ⋅ r i \rho_k=\sum_i e^{i\mathbf{k}\cdot\mathbf{r}_i} ρ k = ∑ i e i k ⋅ r i , with k μ = 2 π n μ / L μ k_\mu=2\pi n_\mu/L_\mu k μ = 2 π n μ / L μ and integer n μ n_\mu n μ , define
S ( k ) = ⟨ ∣ ρ k ∣ 2 ⟩ N , S c ( k ) = S ( k ) − ∣ ⟨ ρ k ⟩ ∣ 2 N . (52) S(k)=\frac{\langle|\rho_k|^2\rangle}{N},\qquad S_{\mathrm{c}}(k)=S(k)-\frac{|\langle\rho_k\rangle|^2}{N}. \tag*{(52)} S ( k ) = N ⟨ ∣ ρ k ∣ 2 ⟩ , S c ( k ) = S ( k ) − N ∣ ⟨ ρ k ⟩ ∣ 2 . (52)
Independent uniform positions give S = 1 S=1 S = 1 for nonzero k k k , while fixed particle number gives S ( 0 ) = N S(0)=N S ( 0 ) = N and S c ( 0 ) = 0 S_{\mathrm{c}}(0)=0 S c ( 0 ) = 0 . The smallest nonzero shell shows increasingly suppressed density fluctuations as longer wavelengths become accessible in spacing units. The shell averages approximately collapse against ∣ k ∣ a |k|a ∣ k ∣ a Figure 5 (d). A fit through the origin to the N ≥ 8 N\geq8 N ≥ 8 points with ∣ k ∣ a < 3.6 |k|a<3.6 ∣ k ∣ a < 3.6 gives S ≃ 0.19 ∣ k ∣ a S\simeq0.19|k|a S ≃ 0.19∣ k ∣ a . The smallest available ∣ k ∣ a |k|a ∣ k ∣ a is π / 2 \pi/2 π /2 at N = 16 N=16 N = 16 ; larger systems would test the continuation of this trend toward zero wavevector.
The one-particle density is largely uniform at all four sizes.
Discussion and conclusion
We have constructed a neural variational representation of continuum semions on a torus that preserves the statistical connection, global transport, and collision domain. Its compact orbital frame contains 3 ( N + 1 ) 3(N+1) 3 ( N + 1 ) generating sections. Configuration-dependent neural coefficients retain the smooth many-body freedom, while learned orbital coordinates adapt the frame without changing the physical boundary conditions. A common phase correction and coherent translation average allow this geometry to be optimized in the target momentum sector.
Validation at two, four, and six semions anchors the many-body calculation. The two-particle energy closely matches the exact solution, and the neural energies at four and six particles lie below our best available finite-cutoff ED estimates (Table I). The same construction reaches up to 16 semions, beyond the reach of our continuum diagonalization. Their observables show a correlation hole, suppressed long-wavelength density fluctuations, and a first reversal of pair circulation near 0.74–0.76 mean spacings. At N = 12 , 16 N = 12, 16 N = 12 , 16 , the circulation returns to positive values at larger separation, giving a damped oscillatory radial profile. Comparisons with smaller systems support a length for the first reversal set by the density. The contact law and limiting ℏ / 2 \hbar/2 ℏ/2 pair circulation follow from the imposed statistical structure; their finite-distance profiles are learned.
Our work represents a first attempt at constructing neural variational states for anyons on a torus. Further methodological improvements can focus on the joint choice of orbital frame and neural architecture. Can one identify an optimal starting frame, in both its geometry and number of generating sections, that reaches lower variational energies at a fixed computational cost? How does accuracy depend on network depth, and which combinations of width and depth are most effective for a given frame size and particle number? Exploring these choices systematically could yield more efficient variational states and extend calculations of energy densities and correlations toward the thermodynamic limit.
Importantly, our work sets the stage to study several open problems via anyonic neural quantum states. Within our framework, both short-range interaction potentials and long-range Coulomb repulsion for charged anyons can be incorporated immediately, allowing us to study whether anyonic crystals [51 ] or superfluids [13 , 14 ] can be stabilized. Our work can also be extended to characterize correlations in multi-component anyon systems with mutual anyonic statistics, e.g., Z 2 \mathbb{Z}_2 Z 2 spin liquids where the spinon and vison excitations are self-bosons but mutual semions. Finally, a long term goal would be to generalize our learned frame ansatz to non-Abelian anyons on a torus [77 ] , thereby opening the door to study the effect of non-Abelian statistics on anyonic correlations at finite density.
Acknowledgments
We thank Eslam Khalaf for insightful conversations and ongoing collaboration on related topics, and T. Senthil for inspiring discussions. The computations in this work were done using Bridges-2 at PSC and SDSC Expanse cluster through allocation PHY250361 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program [78 ] , which is supported by U.S. National Science Foundation grants No. 2138259, No. 2138286, No. 2138307, No. 2137603, and No. 2138296.
Note added. While our manuscript was in preparation, we became aware of related work by Perez Fadon et al. [79 ] . They develop a neural-network based variational ansatz for continuum anyons in harmonic and Coulomb traps. Here we pursue a complementary direction: we address the continuum anyon problem on a torus with a neural ansatz that exactly enforces global transport and provides microscopic access to bulk anyon physics.
Supplemental Material
Consistency of the boundary rules and of the connection
a. Corner algebra and even N N N . Translating x j x_j x j by L x L_x L x changes the phase in the y i y_i y i rule of Eq. (7) by exp [ i π ( N δ i j − 1 ) ] \exp[i\pi(N\delta_{ij}-1)] exp [ iπ ( N δ ij − 1 )] . For i ≠ j i\ne j i = j this is − 1 -1 − 1 for every N N N . For i = j i=j i = j it is ( − 1 ) N − 1 (-1)^{N-1} ( − 1 ) N − 1 , which equals − 1 -1 − 1 only for even N N N . The anticommutation X Z = − Z X XZ=-ZX X Z = − Z X compensates a factor − 1 -1 − 1 for every pair i , j i,j i , j . Both orders of the coordinate translations therefore agree for even N N N , while for odd N N N the same-particle corner fails.
b. Boundary twists and gauge choice. Let b = ( β x / L x , β y / L y ) \mathbf b=(\beta_x/L_x,\beta_y/L_y) b = ( β x / L x , β y / L y ) and χ β ( R ) = b ⋅ ∑ i r i \chi_\beta(\mathbf R)=\mathbf b\cdot\sum_i\mathbf r_i χ β ( R ) = b ⋅ ∑ i r i . The rephasing
Ψ ~ = e − i χ β Ψ , A ~ i = A i + b (53) \widetilde{\Psi}=e^{-i\chi_\beta}\Psi,\qquad\widetilde{\mathcal A}_i=\mathcal A_i+\mathbf b \tag*{(53)} Ψ = e − i χ β Ψ , A i = A i + b (53)
removes the scalar factors e i β μ e^{i\beta_\mu} e i β μ from both boundary rules. With D ~ i = − i ∇ i + A ~ i \widetilde D_i=-i\nabla_i+\widetilde{\mathcal A}_i D i = − i ∇ i + A i , however,
D ~ i Ψ ~ = e − i χ β D i Ψ . (54) \widetilde D_i\widetilde{\Psi}=e^{-i\chi_\beta}D_i\Psi. \tag*{(54)} D i Ψ = e − i χ β D i Ψ. (54)
Thus the same Hamiltonian expectation is represented by untwisted boundaries and an additional constant vector potential. Its curl is zero, but its phases around the handles retain the twist information. For generic β μ \beta_\mu β μ , e − i χ β e^{-i\chi_\beta} e − i χ β is not a single-valued periodic scalar gauge transformation on the torus. Setting both the boundary twists and this constant potential to zero specifies a physical boundary choice; it is not a gauge fixing of arbitrary twists.
c. Sign of the covariant derivative. The inverse of the statistical transmutation in Eq. (1) recovers the locally free anyon-gauge derivative. On a simply connected collision-free chart,
Ψ A = e i χ Ψ ⟹ − i ∇ i Ψ A = e i χ D i Ψ . (55) \Psi_A=e^{i\chi}\Psi\quad\Longrightarrow\quad-i\nabla_i\Psi_A=e^{i\chi}D_i\Psi. \tag*{(55)} Ψ A = e i χ Ψ ⟹ − i ∇ i Ψ A = e i χ D i Ψ. (55)
Near a pair, f ( z ) ∝ z f(z)\propto z f ( z ) ∝ z , so a counterclockwise exchange raises χ \chi χ by π / 2 \pi/2 π /2 . The anyon-gauge wavefunction Ψ A \Psi_A Ψ A thus acquires + i +i + i , and a full pair loop gives − 1 -1 − 1 . Reference 44 uses a fermionic reference wavefunction. Its exchange factor e − i π / 2 e^{-i\pi/2} e − iπ /2 describes the same physical semion phase. Our bosonic representation and boundary convention are fixed by Eqs. (7) and (55) .
d. Quasiperiodicity of the connection. The theta-function identities give
a ( z + L x ) = a ( z ) , a ( z + i L y ) = a ( z ) − π L x x ^ . (56) a(z+L_x)=a(z),\qquad a(z+iL_y)=a(z)-\frac{\pi}{L_x}\hat{\mathbf x}. \tag*{(56)} a ( z + L x ) = a ( z ) , a ( z + i L y ) = a ( z ) − L x π x ^ . (56)
Translating y i y_i y i therefore changes the connection acting on particle k k k by
A k ( R + L y e ^ i y ) = A k ( R ) − π L x ( N δ i k − 1 ) x ^ . (57) \mathcal A_k(\mathbf R+L_y\hat{\mathbf e}_{iy})=\mathcal A_k(\mathbf R)-\frac{\pi}{L_x}(N\delta_{ik}-1)\hat{\mathbf x}. \tag*{(57)} A k ( R + L y e ^ i y ) = A k ( R ) − L x π ( N δ ik − 1 ) x ^ . (57)
This is exactly minus the gradient of the phase in Eq. (7). Hence D k Ψ D_k\Psi D k Ψ obeys the same unitary transport as Ψ \Psi Ψ , and the kinetic density is single valued. The connection is locally a gradient away from coincidences. Its singular circulation represents point statistical flux, not a smooth external field.
Collision domain and pair channels
The coincidence set Δ \Delta Δ contains all configurations with r i = r j \mathbf r_i=\mathbf r_j r i = r j modulo a torus period for at least one pair. Positions are continuous, and the configuration space is F N ( T 2 ) = ( T 2 ) N ∖ Δ F_N(T^2)=(T^2)^N\setminus\Delta F N ( T 2 ) = ( T 2 ) N ∖ Δ , with exchange symmetry imposed on labeled coordinates. The kinetic expectation in Eq. (47) is defined through the nonnegative form
q [ Ψ ] = κ ∑ i ∫ ∥ D i Ψ ( R ) ∥ 2 d R , E [ Ψ ] = q [ Ψ ] ⟨ Ψ ∣ Ψ ⟩ . (58) q[\Psi]=\kappa\sum_i\int\lVert D_i\Psi(\mathbf R)\rVert^2\,dR,\qquad E[\Psi]=\frac{q[\Psi]}{\langle\Psi\mid\Psi\rangle}. \tag*{(58)} q [ Ψ ] = κ i ∑ ∫ ∥ D i Ψ ( R ) ∥ 2 d R , E [ Ψ ] = ⟨ Ψ ∣ Ψ ⟩ q [ Ψ ] . (58)
We define it first on smooth physical sections supported away from Δ \Delta Δ and then close it. The constant factor associated with indistinguishability cancels in the energy ratio. This domain choice is part of our model; it is not inferred from the differential expression of Ref. 44 alone.
For an isolated pair the singular angular operator is ( − i ∂ φ + 1 / 2 ) 2 / r 2 (-i\partial_\varphi+1/2)^2/r^2 ( − i ∂ φ + 1/2 ) 2 / r 2 . The regular radial solutions give the powers ν ℓ = ∣ ℓ + 1 / 2 ∣ \nu_\ell=|\ell+1/2| ν ℓ = ∣ ℓ + 1/2∣ of Eq. (10) , with leading power ν 0 = 1 / 2 \nu_0=1/2 ν 0 = 1/2 . A wavefunction with nonzero value at coincidence has a kinetic density ∝ ∣ Ψ ∣ 2 / ( 4 r 2 ) \propto|\Psi|^2/(4r^2) ∝ ∣Ψ ∣ 2 / ( 4 r 2 ) near the pair and a logarithmically divergent form energy.
With smooth bounded frame and coefficient derivatives, both D i Ψ θ D_i\Psi_\theta D i Ψ θ and the collision cutoff approximants of Ψ θ \Psi_\theta Ψ θ have finite form norm, so the pair factor places the ansatz in the Friedrichs form domain. It does not prescribe all eigenfunction asymptotics. The ℓ = − 2 \ell=-2 ℓ = − 2 channel starts as r 3 / 2 e − 2 i φ r^{3/2}e^{-2i\varphi} r 3/2 e − 2 i φ , whereas a smooth even Cartesian remainder multiplied by r 1 / 2 r^{1/2} r 1/2 starts that angular channel at r 5 / 2 r^{5/2} r 5/2 . This affects the resolution required of the coefficient functions near a collision. In local relative coordinates z = r e i φ z=re^{i\varphi} z = r e i φ , the missing remainder is proportional to z ˉ 2 / r \bar{z}^{2}/r z ˉ 2 / r . Smooth, exchange-even approximants are
b ϵ ( z ) = z ˉ 2 r 2 + ϵ 2 , ϵ > 0. (59) b_\epsilon(z)=\frac{\bar{z}^{2}}{\sqrt{r^{2}+\epsilon^{2}}},\qquad\epsilon>0. \tag*{(59)} b ϵ ( z ) = r 2 + ϵ 2 z ˉ 2 , ϵ > 0. (59)
After multiplication by J ∼ r 1 / 2 J\sim r^{1/2} J ∼ r 1/2 and a smooth local cutoff, these converge to r 3 / 2 e − 2 i φ r^{3/2}e^{-2i\varphi} r 3/2 e − 2 i φ in the local kinetic-form norm as ϵ → 0 \epsilon\to0 ϵ → 0 . Thus increasing coefficient resolution can approximate this channel in energy. Supplying z ˉ 2 / r \bar{z}^{2}/r z ˉ 2 / r explicitly in an additional analytic factor represents its leading power directly, reducing the near-collision resolution required of the network. For a simultaneous collision, let k k k particles approach a common point with all pair separations proportional to a size ϱ \varrho ϱ , keeping their relative shape fixed and the other particles separated. Their k ( k − 1 ) / 2 k(k-1)/2 k ( k − 1 ) /2 internal pair factors each contribute ϱ 1 / 2 \varrho^{1/2} ϱ 1/2 , so J ∼ ϱ k ( k − 1 ) / 4 J\sim\varrho^{k(k-1)/4} J ∼ ϱ k ( k − 1 ) /4 ; for three particles, J ∼ ϱ 3 / 2 J\sim\varrho^{3/2} J ∼ ϱ 3/2 . A bounded smooth remainder makes the full ansatz vanish at least this fast. The physical collective exponent follows from the k k k -particle relative Schrödinger problem, rather than the isolated-pair cusp alone [79 ] . Representing a channel that vanishes more slowly requires coefficients Ψ / J \Psi/J Ψ/ J that grow as the cluster shrinks. Approximating this growth can increase the resolution required of the network; an adapted collective factor can supply it analytically.
Center of mass
a. Momentum and COM energy. Since ∑ i A i = 0 \sum_i\mathcal{A}_i=0 ∑ i A i = 0 , collective translations commute with H H H . A state with P Ψ = ℏ K Ψ \mathbf{P}\Psi=\hbar\mathbf{K}\Psi P Ψ = ℏ K Ψ obeys
K μ L μ = N β μ + 2 π n μ , n μ ∈ Z . (60) K_\mu L_\mu=N\beta_\mu+2\pi n_\mu,\qquad n_\mu\in\mathbb{Z}. \tag*{(60)} K μ L μ = N β μ + 2 π n μ , n μ ∈ Z . (60)
The N N N particle derivatives decompose orthogonally into their sum and relative parts. The sum contributes Eq. (42) , which equals κ ∣ K ∣ 2 / N \kappa|\mathbf{K}|^2/N κ ∣ K ∣ 2 / N for a definite momentum.
b. The COM section. From Eq. (21) , F ( X + 2 L x , Y ) = F ( X , Y ) F(X+2L_x,Y)=F(X,Y) F ( X + 2 L x , Y ) = F ( X , Y ) and F ( X , Y + L y ) = e − i π X / L x F ( X , Y ) F(X,Y+L_y)=e^{-i\pi X/L_x}F(X,Y) F ( X , Y + L y ) = e − iπ X / L x F ( X , Y ) . Consequently Φ ( X + L x ) = X Φ \Phi(X+L_x)=X\Phi Φ ( X + L x ) = X Φ and Φ ( Y + L y ) = e − i π X / L x Z Φ \Phi(Y+L_y)=e^{-i\pi X/L_x}Z\Phi Φ ( Y + L y ) = e − iπ X / L x Z Φ , which are Eqs. (16). The boosts B x B_x B x and B y B_y B y preserve these laws.
c. Gram eigenvalues. For a unit spinor v \boldsymbol{v} v with Bloch vector m \boldsymbol{m} m ,
v v † + Z v v † Z + X v v † X = 1 2 [ 3 + ( m x , − m y , m z ) ⋅ σ ] . (61) \boldsymbol{v}\boldsymbol{v}^{\dagger}+Z\boldsymbol{v}\boldsymbol{v}^{\dagger}Z+X\boldsymbol{v}\boldsymbol{v}^{\dagger}X=\frac{1}{2}\left[3+(m_x,-m_y,m_z)\cdot\boldsymbol{\sigma}\right]. \tag*{(61)} v v † + Z v v † Z + X v v † X = 2 1 [ 3 + ( m x , − m y , m z ) ⋅ σ ] . (61)
with eigenvalues 1 1 1 and 2 2 2 . The boost phases have unit modulus and do not change this. The inverse square root of G = C 0 C 0 † G=C_0C_0^\dagger G = C 0 C 0 † is therefore the fixed polynomial
G − 1 / 2 = ( 2 − 1 2 ) I 2 + ( 1 2 − 1 ) G , (62) G^{-1/2}=\left(2-\frac{1}{\sqrt{2}}\right)\mathbb{I}_2+\left(\frac{1}{\sqrt{2}}-1\right)G, \tag*{(62)} G − 1/2 = ( 2 − 2 1 ) I 2 + ( 2 1 − 1 ) G , (62)
which is regular everywhere and requires no eigenvector convention.
Transport and derivatives of the learned frame
Write U x = ∑ j u θ , j , x U_x=\sum_j u_{\theta,j,x} U x = ∑ j u θ , j , x and Q x = ∑ j q j , x Q_x=\sum_j q_{j,x} Q x = ∑ j q j , x . Under y i → y i + L y y_i\to y_i+L_y y i → y i + L y , periodicity of the displacement means that only q i , y q_{i,y} q i , y changes, by the same period. The analytic frame therefore transforms as
S 0 ( Q θ + L y e ^ y ) = e i β y + i π ( N q i , x − Q x ) / L x Z S 0 ( Q θ ) . (63) S_0(Q_\theta+L_y\hat{e}_y)=e^{i\beta_y+i\pi(Nq_{i,x}-Q_x)/L_x}ZS_0(Q_\theta). \tag*{(63)} S 0 ( Q θ + L y e ^ y ) = e i β y + iπ ( N q i , x − Q x ) / L x Z S 0 ( Q θ ) . (63)
Meanwhile Eq. (33) gives
Δ φ θ = − π L x ( N u θ , i , x − U x ) . (64) \Delta\varphi_\theta=-\frac{\pi}{L_x}(Nu_{\theta,i,x}-U_x). \tag*{(64)} Δ φ θ = − L x π ( N u θ , i , x − U x ) . (64)
Since N q i , x − Q x = N x i − X + N u θ , i , x − U x Nq_{i,x}-Q_x=Nx_i-X+Nu_{\theta,i,x}-U_x N q i , x − Q x = N x i − X + N u θ , i , x − U x , these extra phases cancel exactly. Under an x i x_i x i period, φ θ \varphi_\theta φ θ is unchanged and the analytic frame supplies e i β x X e^{i\beta_x}X e i β x X . Permutation equivariance of the displacement makes both S 0 ( Q θ ) S_0(Q_\theta) S 0 ( Q θ ) and φ θ \varphi_\theta φ θ invariant under particle relabeling. These arguments use coordinate values continued across the boundary, retaining the analytic frame’s quasiperiodic phase.
The ordinary chain rule, for a Cartesian direction μ = x , y \mu= x,y μ = x , y , is
∂ i μ G θ = e i φ θ [ i ( ∂ i μ φ θ ) S 0 ( Q θ ) + ∑ j , ν ( δ i j δ μ ν + ∂ i μ u θ , j , ν ) ∂ q j , ν S 0 ( Q θ ) ] (65) \partial_{i\mu}G_\theta=e^{i\varphi_\theta}\left[i(\partial_{i\mu}\varphi_\theta)S_0(Q_\theta)+\sum_{j,\nu}(\delta_{ij}\delta_{\mu\nu}+\partial_{i\mu}u_{\theta,j,\nu})\partial_{q_{j,\nu}}S_0(Q_\theta)\right] \tag*{(65)} ∂ i μ G θ = e i φ θ [ i ( ∂ i μ φ θ ) S 0 ( Q θ ) + j , ν ∑ ( δ ij δ μν + ∂ i μ u θ , j , ν ) ∂ q j , ν S 0 ( Q θ ) ] (65)
Here ∂ i μ \partial_{i\mu} ∂ i μ differentiates with respect to the physical coordinate r i μ r_{i\mu} r i μ and the Kronecker delta equals one for matching indices and zero otherwise. The matrix of coordinate derivatives in parentheses can lose rank without affecting the tightness of the frame. No inverse coordinate map appears.
For clarity, setting b = S 0 ( Q θ ) c θ b=S_0(Q_\theta)c_\theta b = S 0 ( Q θ ) c θ gives
D i ( e i φ θ b ) = e i φ θ [ − i ∇ i b + ( A i ( R ) + ∇ i φ θ ) b ] . (66) D_i(e^{i\varphi_\theta}b)=e^{i\varphi_\theta}\left[-i\nabla_i b+(\mathcal{A}_i(R)+\nabla_i\varphi_\theta)b\right]. \tag*{(66)} D i ( e i φ θ b ) = e i φ θ [ − i ∇ i b + ( A i ( R ) + ∇ i φ θ ) b ] . (66)
Thus the phase contributes to kinetic and parameter derivatives even though it cancels from the unprojected density. Between translated copies it also affects interference. Keeping J η ( R ) J_\eta(R) J η ( R ) and differentiating the full smooth amplitude allows the same integration-by-parts kinetic form and isolated-pair pole cancellation as for the analytic starting frame. The smoothness and collision-domain argument in Supplemental Material, Sec. S2 applies to both the learned amplitude and its finite projection sum.
Local orbital frame
a. Orbitals. Let B ∗ = 2 π n φ / V B_* = 2\pi n_\varphi/V B ∗ = 2 π n φ / V and ℓ ∗ = B ∗ − 1 / 2 \ell_* = B_*^{-1/2} ℓ ∗ = B ∗ − 1/2 . A basis of the auxiliary LLL is
ϕ k ( x , y ) = ∑ j ∈ Z exp [ − B ∗ 2 ( y − L y λ j k ) 2 ] × exp [ 2 π i n φ L x λ j k x + i β y y j ] (67) \phi_k(x,y)=\sum_{j\in\mathbb{Z}}\exp\left[-\frac{B_*}{2}(y-L_y\lambda_{jk})^2\right]\times\exp\left[\frac{2\pi i n_\varphi}{L_x}\lambda_{jk}x+i\beta_y y_j\right] \tag*{(67)} ϕ k ( x , y ) = j ∈ Z ∑ exp [ − 2 B ∗ ( y − L y λ j k ) 2 ] × exp [ L x 2 π i n φ λ j k x + i β y y j ] (67)
λ j k = j + k n φ + β x 2 π n φ , k = 0 , … , n φ − 1. \lambda_{jk}=j+\frac{k}{n_\varphi}+\frac{\beta_x}{2\pi n_\varphi},\qquad k=0,\ldots,n_\varphi-1. λ j k = j + n φ k + 2 π n φ β x , k = 0 , … , n φ − 1.
Each orbital has the single-particle boundary law of Eq. (17).
b. Two interleaved orbital functions. Put m = n φ m=n_\varphi m = n φ in this paragraph. Choose a near-regular index-m m m point lattice on the torus; “index m m m ” means that it has m m m sites per cell. In the implementation its sites are r j = ( j L x / m , ( s j m o d m ) L y / m ) r_j=(jL_x/m,(sj\bmod m)L_y/m) r j = ( j L x / m , ( s j mod m ) L y / m ) for j = 0 , … , m − 1 j=0,\ldots,m-1 j = 0 , … , m − 1 , with the integer s s s chosen to maximize their minimum separation. Shift all sites so that their sum equals, modulo torus periods,
( m L x 2 − β y L x 2 π , m L y 2 + β x L y 2 π ) . (68) \left(\frac{mL_x}{2}-\frac{\beta_yL_x}{2\pi},\frac{mL_y}{2}+\frac{\beta_xL_y}{2\pi}\right). \tag*{(68)} ( 2 m L x − 2 π β y L x , 2 m L y + 2 π β x L y ) . (68)
This is the zero-sum condition for an LLL function with these boundary twists. The coefficients in ϕ A = ∑ l = 0 m − 1 c l ϕ l \phi_A=\sum_{l=0}^{m-1}c_l\phi_l ϕ A = ∑ l = 0 m − 1 c l ϕ l are found by requiring zeros at these sites and normalizing the coefficient vector. This selects a single orbital function, rather than retaining the full LLL as local features.
Its partner is the magnetic translate
ϕ B ( x , y ) = e i B ∗ d y x ϕ A ( x − d x , y − d y ) , d μ ∈ ( L μ / m ) Z . (69) \phi_B(x,y)=e^{iB_*d_yx}\phi_A(x-d_x,y-d_y),\qquad d_\mu\in(L_\mu/m)\mathbb{Z}. \tag*{(69)} ϕ B ( x , y ) = e i B ∗ d y x ϕ A ( x − d x , y − d y ) , d μ ∈ ( L μ / m ) Z . (69)
The phase in this translation keeps the boundary twists and the LLL condition. The shift is chosen to maximize the minimum separation between the two vortex sets. As two distinct translates of the same index-m m m lattice, their zero sets are disjoint, ensuring ρ > 0 \rho>0 ρ > 0 . For N = 6 N=6 N = 6 and zero twists, the three zeros of ϕ A \phi_A ϕ A are ( 1 / 6 , 1 / 6 ) (1/6,1/6) ( 1/6 , 1/6 ) , ( 1 / 2 , 1 / 2 ) (1/2,1/2) ( 1/2 , 1/2 ) , and ( 5 / 6 , 5 / 6 ) (5/6,5/6) ( 5/6 , 5/6 ) on the unit square, with d = ( 0 , 1 / 3 ) d=(0,1/3) d = ( 0 , 1/3 ) for its partner. These are the constants used in Figure 2 and the trained compact model. For N = 2 N=2 N = 2 , m = 1 m=1 m = 1 allows only one independent LLL function. The first-excited partner ϕ 1 e x = − 2 / B ∗ ∂ y ϕ 0 \phi^{\mathrm{ex}}_1=-\sqrt{2/B_*}\,\partial_y\phi_0 ϕ 1 ex = − 2/ B ∗ ∂ y ϕ 0 is nonzero at the simple zero of ϕ 0 \phi_0 ϕ 0 , so the sum of their squared magnitudes remains positive there.
c. Normalization density and metric. For the two selected LLL functions, Eq. (26) satisfies g = B ∗ + 1 2 ∇ 2 log ρ g = B_* + \frac{1}{2}\nabla^2 \log\rho g = B ∗ + 2 1 ∇ 2 log ρ , and hence ⟨ g ⟩ T 2 = B ∗ \langle g\rangle_{T^2} = B_* ⟨ g ⟩ T 2 = B ∗ . Here ρ \rho ρ is the sum of squared magnitudes of these two functions. Disjoint zeros ensure a regular denominator; interleaving aims to keep the derivative cost spread over the cell. The geometry checks measure this distribution as well as its fixed mean.
d. Why this is minimal. With the projected derivatives D μ v = ∂ μ v − v ( v † ∂ μ v ) D_\mu v = \partial_\mu v - v(v^\dagger\partial_\mu v) D μ v = ∂ μ v − v ( v † ∂ μ v ) , g = ∥ D x v ∥ 2 + ∥ D y v ∥ 2 g = \lVert D_xv\rVert^2 + \lVert D_yv\rVert^2 g = ∥ D x v ∥ 2 + ∥ D y v ∥ 2 , and the Berry curvature of v v v is F = 2 Im ⟨ D x v , D y v ⟩ F = 2\operatorname{Im}\langle D_xv,D_yv\rangle F = 2 Im ⟨ D x v , D y v ⟩ . Pointwise, g ≥ 2 ∥ D x v ∥ ∥ D y v ∥ ≥ ∣ F ∣ g \ge2\lVert D_xv\rVert\lVert D_yv\rVert\ge|F| g ≥ 2 ∥ D x v ∥ ∥ D y v ∥ ≥ ∣ F ∣ . Because v v v carries the boundary phases of n ϕ n_\phi n ϕ flux quanta, ∫ T 2 F d 2 r = ± 2 π n ϕ \int_{T^2} F\,\mathrm{d}^2r = \pm2\pi n_\phi ∫ T 2 F d 2 r = ± 2 π n ϕ . Hence ∫ g d 2 r ≥ 2 π n ϕ \int g\,\mathrm{d}^2r \ge2\pi n_\phi ∫ g d 2 r ≥ 2 π n ϕ , with equality for the LLL textures used here. For the two-component orbital vector, g = 1 4 ∣ ∇ n ∣ 2 g = \frac{1}{4}|\nabla n|^2 g = 4 1 ∣∇ n ∣ 2 and F F F is the skyrmion density, so this is the Belavin–Polyakov bound [70 ] .
These properties control the local orbital geometry. The complete physical derivatives also act on the COM frame, the product normalization, the collision factor and the learned coefficients.
Normalization of the symmetric products
The amplitudes p ~ k \tilde{p}_k p ~ k of Eq. (28) are the components of the symmetric projection of v 1 ⊗ ⋯ ⊗ v N v_1\otimes\cdots\otimes v_N v 1 ⊗ ⋯ ⊗ v N , in the normalized two-orbital occupation basis. This projection removes the parts that change under particle permutations, so its squared norm is at most one. To obtain a lower bound, average over unit two-component vectors z z z , with uniform probability measure d μ ( z ) \mathrm{d}\mu(z) d μ ( z ) on their pseudospin sphere:
N 2 = ( N + 1 ) ∫ ∏ i ∣ z † v i ∣ 2 d μ ( z ) . (70) \mathcal{N}^2=(N+1)\int\prod_i\lvert z^\dagger v_i\rvert^2\,\mathrm{d}\mu(z). \tag*{(70)} N 2 = ( N + 1 ) ∫ i ∏ ∣ z † v i ∣ 2 d μ ( z ) . (70)
For each unit v i v_i v i , the average of log ∣ z † v i ∣ 2 \log\lvert z^\dagger v_i\rvert^2 log ∣ z † v i ∣ 2 is − 1 -1 − 1 . The inequality that the average of an exponential is at least the exponential of its average then gives N 2 ≥ ( N + 1 ) e − N \mathcal{N}^2\ge(N+1)e^{-N} N 2 ≥ ( N + 1 ) e − N , proving the stated lower bound.
Network equations
The network constructs the two learned ingredients of Eq. (36) : the displacement u θ ( R ) u_\theta(R) u θ ( R ) that changes where the analytic frame is evaluated, and the complex coefficients c θ ( R ) c_\theta(R) c θ ( R ) that combine its columns. In terms of the three two-component COM columns C a C_a C a and the N + 1 N+1 N + 1 scalar occupation amplitudes p k p_k p k , the resulting wavefunction is
Ψ θ ( R ) = J η ( R ) e i φ θ ( R ) ∑ a = 1 3 ∑ k = 0 N C a ( X Q , Y Q ) p k ( Q θ ) c θ , a k ( R ) , (71) \Psi_\theta(R)=J_\eta(R)e^{i\varphi_\theta(R)}\sum_{a=1}^{3}\sum_{k=0}^{N}C_a(X_Q,Y_Q)\,p_k(Q_\theta)\,c_{\theta,ak}(R), \tag*{(71)} Ψ θ ( R ) = J η ( R ) e i φ θ ( R ) a = 1 ∑ 3 k = 0 ∑ N C a ( X Q , Y Q ) p k ( Q θ ) c θ , ak ( R ) , (71)
where Q θ = R + u θ ( R ) Q_\theta=R+u_\theta(R) Q θ = R + u θ ( R ) , X Q = ∑ i q i , x X_Q=\sum_i q_{i,x} X Q = ∑ i q i , x , and Y Q = ∑ i q i , y Y_Q=\sum_i q_{i,y} Y Q = ∑ i q i , y . The pair factor, COM columns, scalar amplitudes, and compensating phase (30) have the analytic forms given in the main text. The following layers specify how the network produces u θ u_\theta u θ and c θ c_\theta c θ from the physical positions R R R ; they introduce no additional wavefunction factor.
a. Periodic inputs and the first layer. For harmonics h = 1 , … , H f h=1,\ldots,H_f h = 1 , … , H f , collect sin ( 2 π h r μ / L μ ) \sin(2\pi h r_\mu/L_\mu) sin ( 2 π h r μ / L μ ) and cos ( 2 π h r μ / L μ ) \cos(2\pi h r_\mu/L_\mu) cos ( 2 π h r μ / L μ ) into a vector γ ( r ) \gamma(r) γ ( r ) with 4 H f 4H_f 4 H f entries. These smooth periodic functions describe positions consistently across the edges of the torus. Particle i i i receives the 8 H f 8H_f 8 H f -component vector
f i i n = [ γ ( r i ) , 1 N − 1 ∑ j ≠ i γ ( r i − r j ) ] , (72) f_i^{\mathrm{in}}=\left[\gamma(r_i),\ \frac{1}{N-1}\sum_{j\ne i}\gamma(r_i-r_j)\right], \tag*{(72)} f i in = γ ( r i ) , N − 1 1 j = i ∑ γ ( r i − r j ) , (72)
where square brackets denote concatenation. The first part records its position, and the second describes its environment through an average over the other particles. The input layer is explicitly
h i ( 0 ) = tanh ( W i n f i i n + b i n ) , W i n ∈ R w × 8 H f , b i n ∈ R w . (73) h_i^{(0)}=\tanh\left(W_{\mathrm{in}}f_i^{\mathrm{in}}+b_{\mathrm{in}}\right),\qquad W_{\mathrm{in}}\in\mathbb{R}^{w\times8H_f},\quad b_{\mathrm{in}}\in\mathbb{R}^{w}. \tag*{(73)} h i ( 0 ) = tanh ( W in f i in + b in ) , W in ∈ R w × 8 H f , b in ∈ R w . (73)
An affine map multiplies its input by a learned matrix and adds a learned bias; tanh \tanh tanh then acts on each entry, giving a smooth nonlinear transformation. The resulting hidden vector h i ( 0 ) ∈ R w h_i^{(0)}\in\mathbb{R}^{w} h i ( 0 ) ∈ R w is a learned description of particle i i i in its environment, rather than a physical wavefunction component. Its width w w w is the number of internal features. The same matrix and bias act on every particle.
b. Successive attention layers. The encoder refines these descriptions through h i ( 0 ) → h i ( 1 ) → ⋯ → h i ( L ) h_i^{(0)} \to h_i^{(1)} \to\cdots\to h_i^{(L)} h i ( 0 ) → h i ( 1 ) → ⋯ → h i ( L ) . The superscript counts successive transformations, not particles or time steps. At layer ℓ = 1 , … , L \ell= 1,\ldots,L ℓ = 1 , … , L , each of H a H_a H a attention heads m = 1 , … , H a m = 1,\ldots,H_a m = 1 , … , H a maps the previous hidden vector into three vectors of width d h = w / H a d_h = w/H_a d h = w / H a :
q i ( ℓ , m ) = W Q ( ℓ , m ) h i ( ℓ − 1 ) + b Q ( ℓ , m ) , k i ( ℓ , m ) = W K ( ℓ , m ) h i ( ℓ − 1 ) + b K ( ℓ , m ) , v i ( ℓ , m ) = W V ( ℓ , m ) h i ( ℓ − 1 ) + b V ( ℓ , m ) . (74) \begin{aligned}
q_i^{(\ell,m)} &= W_Q^{(\ell,m)}h_i^{(\ell-1)} + b_Q^{(\ell,m)},\\
k_i^{(\ell,m)} &= W_K^{(\ell,m)}h_i^{(\ell-1)} + b_K^{(\ell,m)},\\
v_i^{(\ell,m)} &= W_V^{(\ell,m)}h_i^{(\ell-1)} + b_V^{(\ell,m)}. \tag*{(74)}
\end{aligned} q i ( ℓ , m ) k i ( ℓ , m ) v i ( ℓ , m ) = W Q ( ℓ , m ) h i ( ℓ − 1 ) + b Q ( ℓ , m ) , = W K ( ℓ , m ) h i ( ℓ − 1 ) + b K ( ℓ , m ) , = W V ( ℓ , m ) h i ( ℓ − 1 ) + b V ( ℓ , m ) . (74)
These are called query, key, and value vectors. The query of particle i i i is compared with the key of particle j j j to determine how much of j j j ’s value vector is passed to i i i :
A i j ( ℓ , m ) = exp [ q i ( ℓ , m ) ⋅ k j ( ℓ , m ) / d h ] ∑ j ′ = 1 N exp [ q i ( ℓ , m ) ⋅ k j ′ ( ℓ , m ) / d h ] , o i ( ℓ , m ) = ∑ j = 1 N A i j ( ℓ , m ) v j ( ℓ , m ) . (75) A_{ij}^{(\ell,m)} = \frac{\exp\left[q_i^{(\ell,m)} \cdot k_j^{(\ell,m)} / \sqrt{d_h}\right]}{\sum_{j'=1}^{N}\exp\left[q_i^{(\ell,m)} \cdot k_{j'}^{(\ell,m)} / \sqrt{d_h}\right]}, \qquad o_i^{(\ell,m)} = \sum_{j=1}^{N} A_{ij}^{(\ell,m)}v_j^{(\ell,m)}. \tag*{(75)} A ij ( ℓ , m ) = ∑ j ′ = 1 N exp [ q i ( ℓ , m ) ⋅ k j ′ ( ℓ , m ) / d h ] exp [ q i ( ℓ , m ) ⋅ k j ( ℓ , m ) / d h ] , o i ( ℓ , m ) = j = 1 ∑ N A ij ( ℓ , m ) v j ( ℓ , m ) . (75)
The weights are positive and sum to one over j j j , so each head returns a configuration-dependent weighted average. All particles, including i i i , enter this sum. Different heads can learn different ways of combining the information. These attention queries are unrelated to the internal coordinates q i \mathbf{q}_i q i , and A i j ( ℓ , m ) A_{ij}^{(\ell,m)} A ij ( ℓ , m ) is unrelated to the statistical connection A i \mathcal{A}_i A i .
The head outputs are concatenated back into a vector of width w w w , mixed by another affine map, and added to the previous hidden vector:
o i ( ℓ ) = [ o i ( ℓ , 1 ) , … , o i ( ℓ , H a ) ] , h i ( ℓ ) = tanh [ h i ( ℓ − 1 ) + W O ( ℓ ) o i ( ℓ ) + b O ( ℓ ) ] , W O ( ℓ ) ∈ R w × w . (76) \begin{aligned}
o_i^{(\ell)} &= [o_i^{(\ell,1)},\ldots,o_i^{(\ell,H_a)}],\\
h_i^{(\ell)} &= \tanh\left[h_i^{(\ell-1)} + W_O^{(\ell)}o_i^{(\ell)} + b_O^{(\ell)}\right], \qquad W_O^{(\ell)} \in\mathbb{R}^{w \times w}. \tag*{(76)}
\end{aligned} o i ( ℓ ) h i ( ℓ ) = [ o i ( ℓ , 1 ) , … , o i ( ℓ , H a ) ] , = tanh [ h i ( ℓ − 1 ) + W O ( ℓ ) o i ( ℓ ) + b O ( ℓ ) ] , W O ( ℓ ) ∈ R w × w . (76)
This addition, often called a residual connection, lets each layer modify the existing description using information from the other particles. The parameters differ between layers and heads, but are shared across particles; no particle-index labels are supplied. Relabeling particles therefore only relabels the hidden vectors. Repeating Eq. (76) L L L times gives the common encoder output { h i ( L ) } \{h_i^{(L)}\} { h i ( L ) } used by both readouts below.
c. From hidden vectors to the learned frame. A readout is a map from hidden features to an output needed by the ansatz. The displacement readout acts separately on each particle:
u θ , i , μ = s L μ tanh [ ( W u h i ( L ) + b u ) μ ] , (77) u_{\theta,i,\mu} = sL_\mu\tanh\left[(W_u h_i^{(L)} + b_u)_\mu\right], \tag*{(77)} u θ , i , μ = s L μ tanh [ ( W u h i ( L ) + b u ) μ ] , (77)
where W u ∈ R 2 × w W_u \in\mathbb{R}^{2 \times w} W u ∈ R 2 × w , b u ∈ R 2 b_u \in\mathbb{R}^2 b u ∈ R 2 , and s = 0.1 s = 0.1 s = 0.1 is the dimensionless displacement scale. Thus ∣ u θ , i , μ ∣ < s L μ \lvert u_{\theta,i,\mu}\rvert< sL_\mu ∣ u θ , i , μ ∣ < s L μ , and a particle’s internal displacement can depend on the entire physical configuration through its hidden vector. The matrix and bias start at zero, giving Q θ = R Q_\theta= R Q θ = R initially. The bound controls scale; exact transport needs smooth, periodic, permutation-equivariant displacements.
Once these displacements are known, the analytic construction gives
Q θ = R + u θ ( R ) , G θ ( R ) = e i φ θ ( R ) S 0 ( Q θ ) . (78) Q_\theta= R + u_\theta(R), \qquad G_\theta(R) = e^{i\varphi_\theta(R)}S_0(Q_\theta). \tag*{(78)} Q θ = R + u θ ( R ) , G θ ( R ) = e i φ θ ( R ) S 0 ( Q θ ) . (78)
In particular, φ θ \varphi_\theta φ θ is calculated from u θ u_\theta u θ by Eq. (33) ; it is not a separate neural output. Only the frame uses the internal coordinates. The pair factor and physical statistical connection remain evaluated at R R R .
d. From hidden vectors to complex coefficients. The coefficient vector describes the whole configuration, so its input must be invariant under particle relabeling. Mean pooling forms h ˉ = N − 1 ∑ i h i ( L ) \bar{h} = N^{-1}\sum_i h_i^{(L)} h ˉ = N − 1 ∑ i h i ( L ) . We concatenate this vector with the N + 1 N+1 N + 1 orbital weights { ∣ p k ( Q θ ) ∣ 2 } k = 0 N \{\lvert p_k(Q_\theta)\rvert^2\}_{k=0}^{N} {∣ p k ( Q θ ) ∣ 2 } k = 0 N to give the coefficient readout information about the frame it will combine:
g θ = tanh { W g [ h ˉ , { ∣ p k ( Q θ ) ∣ 2 } k = 0 N ] + b g } , W g ∈ R w × ( w + N + 1 ) . (79) g_\theta= \tanh\left\{W_g\left[\bar{h},\{\lvert p_k(Q_\theta)\rvert^2\}_{k=0}^{N}\right] + b_g\right\}, \qquad W_g \in\mathbb{R}^{w \times(w+N+1)}. \tag*{(79)} g θ = tanh { W g [ h ˉ , {∣ p k ( Q θ ) ∣ 2 } k = 0 N ] + b g } , W g ∈ R w × ( w + N + 1 ) . (79)
The orbital weights are symmetric and periodic because the common boundary phase of the p k p_k p k cancels in their absolute squares. The vector g θ ∈ R w g_\theta\in\mathbb{R}^{w} g θ ∈ R w is another hidden representation. Two linear readouts produce the real and imaginary parts of the K N = 3 ( N + 1 ) K_N = 3(N+1) K N = 3 ( N + 1 ) complex coefficients:
c θ = ( W R g θ + b R ) + i ( W I g θ + b I ) . (80) c_\theta= (W_Rg_\theta+ b_R) + i(W_Ig_\theta+ b_I). \tag*{(80)} c θ = ( W R g θ + b R ) + i ( W I g θ + b I ) . (80)
Here W R , W I ∈ R K N × w W_R,W_I \in\mathbb{R}^{K_N \times w} W R , W I ∈ R K N × w and b R , b I ∈ R K N b_R,b_I \in\mathbb{R}^{K_N} b R , b I ∈ R K N . Each coefficient labels a COM column and an occupation amplitude, not an individual particle. Both readouts use the same encoder. The calculation is sequential: encode R R R , obtain $u_\theta, evaluate p k p_k p k and the frame at Q θ Q_\theta Q θ , and then obtain c θ \mathbf{c}_\theta c θ . The hidden vectors are computed once from R \mathbf{R} R ; the appended orbital weights account for the dependence of the coefficient readout on the learned coordinates.
For the width-64 architecture, H f = 2 H_f=2 H f = 2 , w = 64 w=64 w = 64 , L = 2 L=2 L = 2 , and H a = 4 H_a=4 H a = 4 . Thus the input layer maps 16 features to 64 hidden entries per particle; two attention updates preserve this width, with 16 entries per head. The displacement readout maps each final hidden vector to two real numbers, while pooling and the complex readout give one vector of 3 ( N + 1 ) 3(N+1) 3 ( N + 1 ) complex coefficients for the whole configuration. For N = 6 N=6 N = 6 these are 21 coefficients multiplying a 2 × 21 2 \times21 2 × 21 frame, whose contraction gives the two physical components in Eq. (71) . The layer count L L L refers to the attention updates; the input layer and readouts are specified separately. All learned matrices and biases above belong to θ \theta θ and are optimized together by the variational energy. Momentum projection reevaluates this entire construction at each shifted configuration with the same parameters and combines the resulting amplitudes.
Exact ground state of two semions
For N = 2 N=2 N = 2 the problem separates, and its ground state can be written in closed form up to one quantization condition.
a. Reduction. With w = r 1 − r 2 \mathbf{w}=\mathbf{r}_1-\mathbf{r}_2 w = r 1 − r 2 , R c = ( r 1 + r 2 ) / 2 \mathbf{R}_c=(\mathbf{r}_1+\mathbf{r}_2)/2 R c = ( r 1 + r 2 ) /2 and A 1 = − A 2 = a ( w ) \mathbf{A}_1=-\mathbf{A}_2=\mathbf{a}(\mathbf{w}) A 1 = − A 2 = a ( w ) ,
H = 1 2 ( − i ∇ R c ) 2 + 2 ( − i ∇ w + a ) 2 . (81) H=\frac{1}{2}(-i\nabla_{\mathbf{R}_c})^2+2(-i\nabla_{\mathbf{w}}+\mathbf{a})^2. \tag*{(81)} H = 2 1 ( − i ∇ R c ) 2 + 2 ( − i ∇ w + a ) 2 . (81)
The boundary rules give Ψ = e i K ⋅ R c Φ ( w ) \Psi=e^{i\mathbf{K}\cdot\mathbf{R}_c}\Phi(\mathbf{w}) Ψ = e i K ⋅ R c Φ ( w ) with K = 2 π ( n x , n y ) \mathbf{K}=2\pi(n_x,n_y) K = 2 π ( n x , n y ) , and the relative function obeys
Φ ( w + 1 ) = s x X Φ ( w ) , Φ ( w + i ) = s y e i π Re w Z Φ ( w ) , Φ ( − w ) = Φ ( w ) , (82) \Phi(w+1)=s_xX\Phi(w),\qquad\Phi(w+i)=s_y e^{i\pi\operatorname{Re}w}Z\Phi(w),\qquad\Phi(-w)=\Phi(w), \tag*{(82)} Φ ( w + 1 ) = s x X Φ ( w ) , Φ ( w + i ) = s y e iπ Re w Z Φ ( w ) , Φ ( − w ) = Φ ( w ) , (82)
with s μ = ( − 1 ) n μ s_\mu=(-1)^{n_\mu} s μ = ( − 1 ) n μ . The four sign sectors are unitarily equivalent, since Z Φ Z\Phi Z Φ flips s x s_x s x and X Φ X\Phi X Φ flips s y s_y s y . Hence E = ∣ K ∣ 2 / 2 + 2 λ E=|\mathbf{K}|^2/2+2\lambda E = ∣ K ∣ 2 /2 + 2 λ , with λ \lambda λ an eigenvalue of one relative problem ( − i ∇ w + a ) 2 Φ = λ Φ (-i\nabla_{\mathbf{w}}+\mathbf{a})^2\Phi=\lambda\Phi ( − i ∇ w + a ) 2 Φ = λ Φ . The ground state has K = 0 \mathbf{K}=0 K = 0 . Its center-of-mass boosts lie 2 π 2 2\pi^2 2 π 2 higher, and the first internal excitation is at E = 44.6066 E=44.6066 E = 44.6066 .
b. Local solutions. On the cell C = { ∣ Re w ∣ , ∣ Im w ∣ ≤ 1 / 2 } C=\{|\operatorname{Re}w|,|\operatorname{Im}w|\le1/2\} C = { ∣ Re w ∣ , ∣ Im w ∣ ≤ 1/2 } around the collision, write w = r e i φ w=re^{i\varphi} w = r e i φ and χ = φ / 2 + η ( w ) \chi=\varphi/2+\eta(w) χ = φ /2 + η ( w ) , with η = 1 2 arg [ f ( w ) / w ] \eta=\frac{1}{2}\arg[f(w)/w] η = 2 1 arg [ f ( w ) / w ] smooth and single valued on C C C [f f f of Eq. (3) ]. In the anyon gauge each component solves the free Helmholtz equation and changes sign around w = 0 w=0 w = 0 . The solutions that are regular in the Friedrichs sense and even under exchange are J n + 1 / 2 ( k r ) e i ( m n + 1 / 2 ) φ J_{n+1/2}(kr)e^{i(m_n+1/2)\varphi} J n + 1/2 ( k r ) e i ( m n + 1/2 ) φ , with m n = n m_n=n m n = n for even n n n and m n = − ( n + 1 ) m_n=-(n+1) m n = − ( n + 1 ) for odd n n n , n = 0 , 1 , 2 , … n=0,1,2,\ldots n = 0 , 1 , 2 , … . The local powers ν = n + 1 / 2 \nu=n+1/2 ν = n + 1/2 are those of Eq. (10) . Half-integer Bessel functions are elementary, for example J 1 / 2 ( x ) = 2 / ( π x ) sin x J_{1/2}(x)=\sqrt{2/(\pi x)}\sin x J 1/2 ( x ) = 2/ ( π x ) sin x . In the bosonic gauge the ground state is, for w ∈ C w\in C w ∈ C ,
Φ c ( w ) = e − i η ( w ) ∑ n ≥ 0 a c , n J n + 1 / 2 ( k r ) J n + 1 / 2 ( k / 2 ) e i m n φ , c = 1 , 2 , (83) \Phi_c(w)=e^{-i\eta(w)}\sum_{n\ge0}a_{c,n}\frac{J_{n+1/2}(kr)}{J_{n+1/2}(k/2)}e^{im_n\varphi},\qquad c=1,2, \tag*{(83)} Φ c ( w ) = e − i η ( w ) n ≥ 0 ∑ a c , n J n + 1/2 ( k /2 ) J n + 1/2 ( k r ) e i m n φ , c = 1 , 2 , (83)
continued to the whole plane by the relative boundary rules. The series converges geometrically on C C C : its radius of convergence is the distance 1 to the nearest image collision, while the cell corners lie at 0.707.
c. Quantization and result. The wavenumber k k k is fixed by requiring that the value and the normal derivative of Φ \Phi Φ match across opposite edges of C C C under the boundary rules. We solve this with the method of particular solutions, locating the minimum of the Betcke–Trefethen subspace angle with interior normalization [80 ] . With M = 40 M=40 M = 40 terms per component,
k = 2.614 701 921 8 , E 0 = 2 k 2 = 13.673 332 280 ± 10 − 9 . (84) k=2.614\,701\,921\,8,\qquad E_0=2k^2=13.673\,332\,280\pm10^{-9}. \tag*{(84)} k = 2.614 701 921 8 , E 0 = 2 k 2 = 13.673 332 280 ± 1 0 − 9 . (84)
Between 32 and 56 terms per component, k k k changes by less than 10 − 10 10^{-10} 1 0 − 10 . The ground state is non-degenerate in each sector: at k k k the second subspace angle is 0.73, while the first is about 10 − 8 10^{-8} 1 0 − 8 . The coefficients are real. The leading ones are
Table 1 (PDF p. 21).
They are normalized so that ∫ C ∥ Φ ∥ 2 = 1 \int_{\mathcal{C}}\|\Phi\|^2=1 ∫ C ∥Φ ∥ 2 = 1 , which makes g 2 ( d ) = ∥ Φ ( d ) ∥ 2 g_2(d)=\|\Phi(d)\|^2 g 2 ( d ) = ∥Φ ( d ) ∥ 2 in the convention of Eq. (49) . Two relations hold numerically to 3 × 10 − 8 3\times10^{-8} 3 × 1 0 − 8 without being imposed: c 2 , 2 j = − c 1 , 2 j + 1 c_{2,2j}=-c_{1,2j+1} c 2 , 2 j = − c 1 , 2 j + 1 and c 2 , 2 j + 1 = c 1 , 2 j c_{2,2j+1}=c_{1,2j} c 2 , 2 j + 1 = c 1 , 2 j for the unscaled coefficients c c , n = a c , n / J n + 1 / 2 ( k / 2 ) c_{c,n}=a_{c,n}/J_{n+1/2}(k/2) c c , n = a c , n / J n + 1/2 ( k /2 ) . The contact spinor points along ( cos π 8 , sin π 8 ) (\cos\frac{\pi}{8},\sin\frac{\pi}{8}) ( cos 8 π , sin 8 π ) , the + 1 +1 + 1 eigenvector of ( X + Z ) / 2 (X+Z)/\sqrt{2} ( X + Z ) / 2 . Near contact ∥ Φ ∥ = 1.86324 r 1 / 2 [ 1 + O ( r ) ] \|\Phi\|=1.86324\,r^{1/2}[1+O(r)] ∥Φ∥ = 1.86324 r 1/2 [ 1 + O ( r )] , and the next channel is r 3 / 2 e − 2 i φ r^{3/2}e^{-2i\varphi} r 3/2 e − 2 i φ .
d. Exact observables. Figure S1 shows the exact pair correlation and the pair relative angular momentum L ( r ) L(r) L ( r ) of Eq. (51) . For N = 2 N=2 N = 2 with K = 0 K=0 K = 0 , v 1 − v 2 = ( 2 ℏ / m ) Re [ Φ † ( − i ∇ w + a ) Φ ] / ∥ Φ ∥ 2 v_1-v_2=(2\hbar/m)\operatorname{Re}[\Phi^\dagger(-i\nabla_w+a)\Phi]/\|\Phi\|^2 v 1 − v 2 = ( 2ℏ/ m ) Re [ Φ † ( − i ∇ w + a ) Φ ] /∥Φ ∥ 2 . The pair correlation rises as 1.8632 2 r / L 1.8632^2r/L 1.863 2 2 r / L and exceeds one beyond r ≃ 0.35 L r\simeq0.35L r ≃ 0.35 L . L ( r ) L(r) L ( r ) starts at ℏ / 2 \hbar/2 ℏ/2 and decreases monotonically to 0.12 ℏ 0.12\hbar 0.12ℏ at r = L / 2 r=L/2 r = L /2 . The radial part vanishes to 10 − 17 10^{-17} 1 0 − 17 . These curves are exact references for the neural N = 2 N=2 N = 2 state.
Figure S1. Exact two-semion ground state on the unit torus. (a) Relative density ∥ Φ ( d ) ∥ 2 = g 2 ( d ) \|\Phi(d)\|^2=g_2(d) ∥Φ ( d ) ∥ 2 = g 2 ( d ) . (b) Angular average with the contact law. (c) Pair relative angular momentum; the dotted line is the isolated-pair value α ℏ = ℏ / 2 \alpha\hbar=\hbar/2 α ℏ = ℏ/2 .
VMC estimators, optimization, and evaluation
The ansatz is optimized and evaluated by variational Monte Carlo (VMC). For training through the projection, the same projected state, Eq. (45) , is used for sampling, the energy, and its gradient. Measurements use the frozen state specified for each result.
a. Sampling. Configurations are drawn from the Born density ∥ Ψ θ ( M ) ( R ) ∥ 2 = ∣ Ψ 1 ∣ 2 + ∣ Ψ 2 ∣ 2 \|\Psi_\theta^{(M)}(R)\|^2=|\Psi_1|^2+|\Psi_2|^2 ∥ Ψ θ ( M ) ( R ) ∥ 2 = ∣ Ψ 1 ∣ 2 + ∣ Ψ 2 ∣ 2 with the Metropolis algorithm [74 ] . One sweep combines a Gaussian move of all particles with a move of one randomly chosen particle to a uniformly drawn position. During training, the Gaussian width is adjusted every 100 updates from the measured acceptance, following [57 ] , and is capped at the shortest box length. Ten sweeps separate consecutive optimization steps. Positions are wrapped into the cell only for sampling. The density is periodic, while the amplitude itself obeys Eqs. (7).
b. Energy estimator. Write Ψ = J η f \Psi=J_\eta f Ψ = J η f , where f f f is the smooth, momentum-projected amplitude, and u = ln J η u=\ln J_\eta u = ln J η . Integrating the cross terms between J η J_\eta J η and f f f by parts turns the form (S6) into an average of first derivatives only, with e W e_W e W as given in Eq. (48) . The boundary terms at coincidences vanish for the r 1 / 2 r^{1/2} r 1/2 cusp. Near a collision, both ∥ D i f ∥ 2 / ∥ f ∥ 2 \|D_i f\|^2/\|f\|^2 ∥ D i f ∥ 2 /∥ f ∥ 2 and ∣ ∇ i u ∣ 2 |\nabla_i u|^2 ∣ ∇ i u ∣ 2 contain a 1 / ( 4 r 2 ) 1/(4r^2) 1/ ( 4 r 2 ) pole. The statistical connection and the rotated gradient of u u u share the same singular part, so we subtract it analytically before summing. The local estimator e W e_W e W has no leading isolated-pair 1 / r 2 1/r^2 1/ r 2 divergence. It has the same expectation as the kinetic form, but unlike the operator estimator Re [ Ψ † H Ψ ] / ∥ Ψ ∥ 2 \operatorname{Re}[\Psi^\dagger H\Psi]/\|\Psi\|^2 Re [ Ψ † H Ψ ] /∥Ψ ∥ 2 it does not have zero variance at an eigenstate. It needs only first derivatives, avoiding spatial second derivatives of the neural amplitude.
c. Optimization. The gradient differentiates the sampled energy expectation, numerator and norm together, with the Born measure held fixed. It is therefore the exact gradient of the variational energy, for the full complex two-component state. Updates use Kronecker-factored approximate curvature (K-FAC) [57 , 75 ] in its kfac-jax implementation, with the learning rate ϵ t = 0.05 / ( 1 + t / 10 4 ) \epsilon_t=0.05/(1+t/10^4) ϵ t = 0.05/ ( 1 + t /1 0 4 ) , damping 10 − 3 10^{-3} 1 0 − 3 and a norm constraint 10 − 3 10^{-3} 1 0 − 3 . Every coefficient, attention, embedding and displacement parameter is registered with the curvature estimator. Training starts from random initialization with a fixed seed. The displacement head starts at zero, so the first state uses the analytic starting frame.
d. Production training protocol. The pair range is fixed at η = 0.4 a \eta=0.4a η = 0.4 a , and the initial Gaussian width parameter is 0.32 a 2 0.32a^2 0.32 a 2 , with α = ( V / N ) 1 / 2 \alpha=(\mathcal{V}/N)^{1/2} α = ( V / N ) 1/2 and ν = 1 \nu=1 ν = 1 . All production states use two attention layers, four heads, two Fourier harmonics, displacement scale s = 0.1 s=0.1 s = 0.1 , the same optimizer and schedule, and 2 × 10 4 2\times10^4 2 × 1 0 4 updates from random initialization, continued to 6 × 10 4 6\times10^4 6 × 1 0 4 for the selected N = 16 N=16 N = 16 state. The training grid is the smallest M ≥ 2 M\ge2 M ≥ 2 coprime to N / 2 N/2 N /2 : M = 2 , 3 , 2 , 3 , 2 , 5 , 3 M = 2, 3, 2, 3, 2, 5, 3 M = 2 , 3 , 2 , 3 , 2 , 5 , 3 for N = 2 , 4 , 6 , 8 , 10 , 12 , 16 N = 2, 4, 6, 8, 10, 12, 16 N = 2 , 4 , 6 , 8 , 10 , 12 , 16 . The hidden width is 64 through N = 12 N = 12 N = 12 and 256 for the selected N = 16 N = 16 N = 16 state. We use 2048 walkers through N = 10 N = 10 N = 10 and 1024 at N = 12 , 16 N = 12, 16 N = 12 , 16 . Walkers are split over the available GPUs; the parameters and optimizer arrays are replicated. The N = 16 N = 16 N = 16 width controls use the same seed, training grid, walker count, and 2 × 10 4 2 \times10^{4} 2 × 1 0 4 updates. At N = 6 , 8 N = 6, 8 N = 6 , 8 , width-128 controls lower the P 2 M P_{2M} P 2 M energies by 0.27 and 1.05, respectively. Independent seed-53 runs through N = 8 N = 8 N = 8 , evaluated on their training grids, reproduce the corresponding seed-52 energies within 0.15; the larger production states use one training seed.
e. Evaluation. Reported energies and observables come from a separate VMC run with frozen parameters, a new random stream, and the sampling width of the final checkpoint. All selected states are evaluated through P 2 M P_{2M} P 2 M on 4000 measurements, each following ten sweeps, with evaluation seed 2 and the walker counts stated above. Statistical errors come from blocking the measurement series [76 ] . We quote the largest blocked error among block sizes that leave at least 32 blocks. Spatial maps and profiles use errors from 32 contiguous groups of blocks. The same run measures the center-of-mass energy (42) , total momentum, and the observables of Sec. VI.
Continuum exact diagonalization
The ED calculation is a linear Rayleigh–Ritz expansion in explicitly known continuum functions. It uses the Hamiltonian, two global components, and boundary conditions of Sec. II, with the Friedrichs collision condition described in Sec. III B. All energies below are for L x = L y = 1 L_x = L_y = 1 L x = L y = 1 , β x = β y = 0 \beta_x = \beta_y = 0 β x = β y = 0 , and κ = 1 \kappa= 1 κ = 1 . The basis is constructed independently of the neural wavefunction: its sections and collision factors are fixed, and only linear coefficients are determined by diagonalization.
Boundary sections, collision factor, and Fourier cutoff
An unordered configuration μ \mu μ consists of N N N integer vectors k i = ( k i x , k i y ) \mathbf{k}_i = (k_{ix}, k_{iy}) k i = ( k i x , k i y ) . Let O μ \mathcal{O}_{\mu} O μ be its distinct assignments to the labeled particles. The symmetric Fourier polynomial is
F μ ( R ) = 1 ∣ O μ ∣ ∑ k ∈ O μ exp [ 2 π i ∑ i = 1 N ( k i x x i L x + k i y y i L y ) ] , λ μ = ∑ i ( k i x 2 + k i y 2 ) ≤ Λ . (85) F_{\mu}(\mathbf{R}) = \frac{1}{\sqrt{|\mathcal{O}_{\mu}|}} \sum_{\mathbf{k}\in\mathcal{O}_{\mu}} \exp\left[2\pi i \sum_{i=1}^{N}\left(\frac{k_{ix}x_i}{L_x}+\frac{k_{iy}y_i}{L_y}\right)\right], \qquad\lambda_{\mu} = \sum_i \left(k_{ix}^{2}+k_{iy}^{2}\right) \leq\Lambda. \tag*{(85)} F μ ( R ) = ∣ O μ ∣ 1 k ∈ O μ ∑ exp [ 2 π i i = 1 ∑ N ( L x k i x x i + L y k i y y i ) ] , λ μ = i ∑ ( k i x 2 + k i y 2 ) ≤ Λ. (85)
Every configuration satisfying this inequality is included. Thus Λ \Lambda Λ truncates coefficient functions. The fixed factors below also carry spatial structure, so Λ \Lambda Λ is not a sharp Fourier cutoff on the full physical wavefunction.
To supply the boundary transport, periodize a product Gaussian with the matrices X X X and Z Z Z of Eq. (6) . With s i = N x i − ∑ j x j s_i = Nx_i - \sum_j x_j s i = N x i − ∑ j x j , define the 2 × 2 2 \times2 2 × 2 matrix
U σ ( R ) = Y σ ( R ) X σ ( R ) , U_{\sigma}(\mathbf{R}) = Y_{\sigma}(\mathbf{R})X_{\sigma}(\mathbf{R}), U σ ( R ) = Y σ ( R ) X σ ( R ) ,
X σ ( R ) = ∏ i = 1 N [ ∑ n ∈ Z e − ( x i − n L x ) 2 / ( 2 σ 2 L x 2 ) X n ] , X_{\sigma}(\mathbf{R}) = \prod_{i=1}^{N}\left[\sum_{n\in\mathbb{Z}} e^{-(x_i-nL_x)^2/(2\sigma^2L_x^2)}X^n\right], X σ ( R ) = i = 1 ∏ N [ n ∈ Z ∑ e − ( x i − n L x ) 2 / ( 2 σ 2 L x 2 ) X n ] ,
Y σ ( R ) = ∏ i = 1 N [ ∑ m ∈ Z e − ( y i − m L y ) 2 / ( 2 σ 2 L y 2 ) e i π m s i / L x Z m ] . (86) Y_{\sigma}(\mathbf{R}) = \prod_{i=1}^{N}\left[\sum_{m\in\mathbb{Z}} e^{-(y_i-mL_y)^2/(2\sigma^2L_y^2)}e^{i\pi ms_i/L_x}Z^m\right]. \tag*{(86)} Y σ ( R ) = i = 1 ∏ N [ m ∈ Z ∑ e − ( y i − m L y ) 2 / ( 2 σ 2 L y 2 ) e iπ m s i / L x Z m ] . (86)
Each column is permutation symmetric and obeys the required two-component transport. Relabeling the Gaussian images proves the y y y rule directly; for the x x x rule, X Z = − Z X XZ = -ZX X Z = − Z X cancels the change in the y y y -image phases. The product form evaluates the many-particle image sum through one-dimensional sums and provides analytic first derivatives. The reported bases use fixed width σ = 0.3 \sigma= 0.3 σ = 0.3 , images ∣ m ∣ , ∣ n ∣ ≤ 6 |m|, |n| \leq6 ∣ m ∣ , ∣ n ∣ ≤ 6 , and 16 trapezoidal nodes per direction for the projection below.
The ED collision factor is the range-free function
J E D ( R ) = ∏ i < j d i j 1 / 2 , ϕ μ b ( R ) = J E D ( R ) P K [ F μ ( R ) U σ ( R ) e b ] , b = 1 , 2 , (87) J_{\mathrm{ED}}(\mathbf{R}) = \prod_{i<j} d_{ij}^{1/2}, \qquad\phi_{\mu b}(\mathbf{R}) = J_{\mathrm{ED}}(\mathbf{R})P_K[F_{\mu}(\mathbf{R})U_{\sigma}(\mathbf{R})\mathbf{e}_b], \qquad b = 1, 2, \tag*{(87)} J ED ( R ) = i < j ∏ d ij 1/2 , ϕ μ b ( R ) = J ED ( R ) P K [ F μ ( R ) U σ ( R ) e b ] , b = 1 , 2 , (87)
where d i j d_{ij} d ij is the chord distance of Sec. III B and e b \mathbf{e}_b e b is a component unit vector. This factor has the same isolated-pair r 1 / 2 r^{1/2} r 1/2 cusp as the neural factor J η J_{\eta} J η , but has no range parameter. It places the basis functions in the same Friedrichs form domain. The collective translation projection is
P K f ( R ) = 1 V ∫ T 2 d 2 t e − i K ⋅ t f ( R + t ) , K a = 2 π n a L a . (88) P_K f(\mathbf{R}) = \frac{1}{\mathcal{V}}\int_{T^2} d^2t\,e^{-i\mathbf{K}\cdot\mathbf{t}}f(\mathbf{R}+\mathbf{t}), \qquad K_a = \frac{2\pi n_a}{L_a}. \tag*{(88)} P K f ( R ) = V 1 ∫ T 2 d 2 t e − i K ⋅ t f ( R + t ) , K a = L a 2 π n a . (88)
The comparisons use K = 0 K = 0 K = 0 . Since F μ F_\mu F μ has known total Fourier momentum and s i s_i s i is invariant under collective translation, this projection separates into two one-dimensional integrals over the Gaussian blocks. The translation nodes approximate these smooth integrals. Increasing their number from 16 to 24 and the image cutoff from 6 to 7 changes tested N = 4 N = 4 N = 4 columns and first derivatives through Λ = 8 \Lambda= 8 Λ = 8 by less than 2 × 10 − 11 2 \times10^{-11} 2 × 1 0 − 11 in relative norm.
There are two nominal columns per Fourier configuration. At N = 4 N = 4 N = 4 , the Λ = 4 , 5 , 6 \Lambda= 4, 5, 6 Λ = 4 , 5 , 6 spaces contain 288, 576, and 992 columns. At N = 6 N = 6 N = 6 , Λ = 6 , 7 , 8 , 9 \Lambda= 6, 7, 8, 9 Λ = 6 , 7 , 8 , 9 gives 1552, 3072, 5526, and 9558 columns. These counts precede the removal of overlap dependencies.
The N = 4 N = 4 N = 4 entry marked + T +T + T in Table I also includes the regular ℓ = − 2 \ell= -2 ℓ = − 2 channel of Sec. S2. For a minimum-image pair displacement z i j = δ x i j + i δ y i j z_{ij} = \delta x_{ij} + i\delta y_{ij} z ij = δ x ij + i δ y ij , let r i j = ∣ z i j ∣ r_{ij} = \lvert z_{ij}\rvert r ij = ∣ z ij ∣ and
t i j = χ ( r i j ) z ‾ i j 2 r 0 r i j , T ( R ) = ∑ i < j t i j , r 0 = 0.4 L . (89) t_{ij} = \chi(r_{ij})\frac{\overline{z}_{ij}^{2}}{r_{0}r_{ij}},\qquad T(\mathbf{R}) = \sum_{i<j}t_{ij},\qquad r_{0} = 0.4L. \tag*{(89)} t ij = χ ( r ij ) r 0 r ij z ij 2 , T ( R ) = i < j ∑ t ij , r 0 = 0.4 L . (89)
Here χ ( r ) = 1 \chi(r) = 1 χ ( r ) = 1 for r ≤ r 0 / 2 r \le r_{0}/2 r ≤ r 0 /2 , χ ( r ) = 0 \chi(r) = 0 χ ( r ) = 0 for r ≥ r 0 r \ge r_{0} r ≥ r 0 , and χ ( r ) = 1 − h ( 2 r / r 0 − 1 ) \chi(r) = 1 - h(2r/r_{0} - 1) χ ( r ) = 1 − h ( 2 r / r 0 − 1 ) in between, with h ( v ) = e − 1 / v / [ e − 1 / v + e − 1 / ( 1 − v ) ] h(v) = e^{-1/v}/[e^{-1/v} + e^{-1/(1-v)}] h ( v ) = e − 1/ v / [ e − 1/ v + e − 1/ ( 1 − v ) ] for 0 < v < 1 0 < v < 1 0 < v < 1 . The support lies inside the minimum-image cell, making T T T periodic; it is also symmetric and invariant under collective translation. The added columns are T ϕ μ b T\phi_{\mu b} T ϕ μ b with their own coefficient cutoff λ μ ≤ Λ T \lambda_\mu\le\Lambda_T λ μ ≤ Λ T . Their pair contribution has the leading form r 3 / 2 e − 2 i φ r^{3/2}e^{-2i\varphi} r 3/2 e − 2 i φ , giving this channel an independent amplitude. The bounded first derivatives of t i j t_{ij} t ij keep the columns in the form domain. The quoted N = 4 N = 4 N = 4 result uses Λ = 6 \Lambda= 6 Λ = 6 and Λ T = 5 \Lambda_T = 5 Λ T = 5 , for 992 + 576 = 1568 992 + 576 = 1568 992 + 576 = 1568 nominal columns. The reported N = 6 N = 6 N = 6 calculations use the base family of Eq. (87) , without T T T .
Continuum matrices and numerical validation
Writing Ψ = ∑ a c a ϕ a \Psi= \sum_a c_a\phi_a Ψ = ∑ a c a ϕ a , assemble
S a b = ∫ d R ϕ a † ϕ b , H a b = κ ∫ d R ∑ i ( D i ϕ a ) † ⋅ ( D i ϕ b ) , H c = E S c . (90) S_{ab} = \int dR\,\phi_a^\dagger\phi_b,\qquad H_{ab} = \kappa\int dR\sum_i(D_i\phi_a)^\dagger\cdot(D_i\phi_b),\qquad Hc = ESc. \tag*{(90)} S ab = ∫ d R ϕ a † ϕ b , H ab = κ ∫ d R i ∑ ( D i ϕ a ) † ⋅ ( D i ϕ b ) , H c = E S c . (90)
All component contractions use the full two-component state. Derivatives of the Gaussian sums, Fourier polynomials, and collision factors are analytic. At fixed total momentum the integrands are invariant under collective translation. Fixing r N = 0 r_N = 0 r N = 0 therefore removes one position integral; the remaining 2 ( N − 1 ) 2(N-1) 2 ( N − 1 ) coordinates are integrated continuously, with the common area factor restored.
For ϕ a = J E D f a \phi_a = J_{\mathrm{ED}}f_a ϕ a = J ED f a , integration by parts gives
H a b κ = ∫ d R J E D 2 [ ∑ i ( D i f a ) † ⋅ ( D i f b ) − Δ J E D J E D f a † f b ] , Δ = ∑ i ∇ i 2 . (91) \frac{H_{ab}}{\kappa} = \int dR\,J_{\mathrm{ED}}^2\left[\sum_i(D_if_a)^\dagger\cdot(D_if_b)-\frac{\Delta J_{\mathrm{ED}}}{J_{\mathrm{ED}}}f_a^\dagger f_b\right],\qquad\Delta= \sum_i\nabla_i^2. \tag*{(91)} κ H ab = ∫ d R J ED 2 [ i ∑ ( D i f a ) † ⋅ ( D i f b ) − J ED Δ J ED f a † f b ] , Δ = i ∑ ∇ i 2 . (91)
Torus-face terms cancel by boundary covariance, and the collision terms vanish with the regular pair cusp. This is the original kinetic form, with no additional scalar interaction. With g i = ∇ i ln J E D \mathbf{g}_i = \nabla_i\ln J_{\mathrm{ED}} g i = ∇ i ln J ED , the scalar term in the expanded kernel is ∑ i ( ∣ A i ∣ 2 − ∣ g i ∣ 2 ) − Δ ln J E D \sum_i(\lvert\mathbf{A}_i\rvert^2-\lvert\mathbf{g}_i\rvert^2)-\Delta\ln J_{\mathrm{ED}} ∑ i (∣ A i ∣ 2 − ∣ g i ∣ 2 ) − Δ ln J ED . The equal leading 1 / ( 4 r 2 ) 1/(4r^2) 1/ ( 4 r 2 ) pair poles of ∣ A i ∣ 2 \lvert\mathbf{A}_i\rvert^2 ∣ A i ∣ 2 and ∣ g i ∣ 2 \lvert\mathbf{g}_i\rvert^2 ∣ g i ∣ 2 cancel before quadrature. This reduces integration scatter substantially compared with directly integrating the positive form.
The reduced matrices use uniform scrambled Sobol points, assembled independently for each scramble and averaged before diagonalization. Independent scrambles and increasing point counts test the matrix integrals. We also freeze a vector fitted without a given scramble and evaluate its Rayleigh quotient with the excluded matrices. A separate check freezes the averaged-matrix vector and evaluates the original positive form on new points, without another eigensolve. For this latter check, half of the integration measure is uniform and half samples close pairs. The pair component has uniform angle and radial density proportional to ( 1 − r / R ) 2 (1-r/R)^2 ( 1 − r / R ) 2 on 0 < r < R = min ( L x , L y ) / 2 0<r<R=\min(L_x,L_y)/2 0 < r < R = min ( L x , L y ) /2 ; exact mixture weights restore the continuum integral. Permutation symmetry allows separate integration of the uniform term and one representative pair term.
The basis is nonorthogonal and has exact and near dependencies. First scale with B a a = S a a − 1 / 2 B_{aa}=S_{aa}^{-1/2} B aa = S aa − 1/2 and diagonalize S ~ = B S B \widetilde{S}=BSB S = B S B . Keep overlap eigenvalues s j > ϵ s max s_j>\epsilon s_{\max} s j > ϵ s m a x and canonically orthogonalize with W = B V k e e p s k e e p − 1 / 2 W=BV_{\mathrm{keep}}s_{\mathrm{keep}}^{-1/2} W = B V keep s keep − 1/2 . Diagonalizing W † H W W^\dagger HW W † H W gives the retained Ritz spectrum and c = W v c=Wv c = W v . Both the retained rank and the sensitivity to ϵ \epsilon ϵ are recorded.
Table S1 separates fixed-basis integration refinement from the larger change on increasing Λ \Lambda Λ . Changing ϵ \epsilon ϵ from 10 − 9 10^{-9} 1 0 − 9 to 10 − 12 10^{-12} 1 0 − 12 changes the fine energies from 73.87399 to 73.83746 at Λ = 4 \Lambda= 4 Λ = 4 and from 73.17856 to 73.17819 at Λ = 5 \Lambda= 5 Λ = 5 . Thresholds 10 − 12 10^{-12} 1 0 − 12 and 10 − 13 10^{-13} 1 0 − 13 retain identical spaces.
Table S1. Refined N = 4 N = 4 N = 4 , K = 0 K = 0 K = 0 base-basis results. Each point count uses four independent reduced-kernel scrambles; the energies diagonalize their averaged matrices with ϵ = 10 − 12 \epsilon= 10^{-12} ϵ = 1 0 − 12 . The last column evaluates vectors selected using only the other three fine scrambles. Its range describes integration variation, rather than a statistical confidence interval.
Original-positive-form checks on independently frozen 10 − 9 10^{-9} 1 0 − 9 vectors agree at the few-hundredths level. Deterministic two-particle Gauss quadrature also verifies agreement between the original and reduced kernels.
The N = 4 N = 4 N = 4 , Λ = 6 + T ( 5 ) \Lambda= 6 + T(5) Λ = 6 + T ( 5 ) calculation uses two 2 16 2^{16} 2 16 -point scrambles and ϵ = 10 − 12 \epsilon= 10^{-12} ϵ = 1 0 − 12 , retaining 819 of 1568 columns. The averaged-matrix value is 72.4283; vectors fitted to one scramble and evaluated on the other give 72.4807 and 72.5017. These checks place the integration variation on the scale of a few hundredths. The remaining Fourier-cutoff dependence is larger: at the same integration settings, the Λ = 5 + T ( 5 ) \Lambda= 5 + T(5) Λ = 5 + T ( 5 ) value is 72.7924.
For N = 6 N = 6 N = 6 , Table S2 reports the completed Λ = 8 , 9 \Lambda= 8, 9 Λ = 8 , 9 runs. Their matrix scrambles use seeds 9101 and 9102. The mean of the two fine Λ = 9 \Lambda= 9 Λ = 9 positive-form evaluations is 176.2543, quoted as about 176.3 in Table I. Averaging the independently evaluated Rayleigh quotients here gives a practical finite-state reference. The roughly 1.35 difference from its fitted value 174.9042 demonstrates why matrix fitting and state evaluation must be distinguished. The coarser run gives a positive-form mean 176.6382, so integration refinement also changes the selected state on the scale of a few tenths. The two validation values alone do not determine a confidence interval.
Table S2. N = 6 N = 6 N = 6 , K = 0 K = 0 K = 0 base-basis calculations. The first two columns identify the Fourier cutoff and the point count per reduced-matrix scramble. Two independent matrix scrambles are used. Excluded-scramble checks evaluate each single-scramble vector with the other matrices. The last column gives two independent original-positive-form evaluations of the frozen averaged-matrix vector; no coefficients are refitted for these checks.
The Fourier cutoff, numerical integration, overlap threshold, Gaussian width, and momentum sector are separate convergence controls. In particular, the present N = 6 N = 6 N = 6 sequence still changes with both cutoff and integration. A sharper ED reference requires their joint refinement, followed by independent positive-form and overlap-threshold checks at the desired energy resolution. Establishing the global ground sector additionally requires comparison with the inequivalent nonzero-momentum sectors. For exact integrals the conforming trial spaces give variational upper bounds; the numerical values above are quadrature estimates of those finite-basis energies.