Introduction
Strong interactions can reorganize a lattice many-body Hilbert space long before they produce conventional long-range order. A simple example is a repulsively bound pair, or doublon: when the interaction energy exceeds the available kinetic bandwidth, two particles occupying the same site form a long-lived composite excitation whose motion proceeds only through higher-order tunneling. Repulsively bound pairs were observed experimentally in an optical lattice [1 , 2 ] and have since become a standard example of interaction-induced composite dynamics in Hubbard-type systems.
In the two-dimensional electronic Hubbard problem, Souza and Claro [3 ] showed that the repulsive pair forms a distinct “U-band” and behaves dynamically as a charge-2e composite object, the doubled Bloch-oscillation frequency of their simulations being its dynamical signature. Claro and Robles [4 ] solved the one-dimensional bound pair exactly and showed semiclassically, at leading order in λ / U \lambda/U λ / U , that its Bloch-oscillation frequency is doubled, compatible with charge 2e. More recent work has extended this picture to tunable doublon–doublon interactions [5 ] , multidoublon Bloch oscillations [6 ] , doublon-plus-particle dynamics [7 ] , doublon/holon attraction from strong-coupling perturbation theory [8 ] , QED-analogous pair creation [9 ] , non-Hermitian bosonic doublon dynamics [10 ] , and tunable extended-range bound pairs in a Rydberg-dressed Bose-Hubbard lattice [11 ] . Effective doublon Hamiltonians, with parameters derived perturbatively from the underlying repulsion, are therefore well established as a description of strong-coupling dynamics.
There is a complementary, geometric way of viewing this problem. A wavefunction of n n n distinguishable particles on a 2D lattice is a single-particle wavefunction on a 2 n 2n 2 n -dimensional configuration lattice; such mappings of correlated few-particle problems onto higher-dimensional single-particle tight-binding problems have precedent [12 , 13 ] . Binding and projection restrict this high-dimensional lattice to lower-dimensional subspaces: a doublon is the simplest instance, translating with two coordinates where two independent particles need four. This coordinate-counting device is not itself the novelty of the present work; it is the starting point for the question we address: once strong repulsion restricts n n n fermions to a coordinate skeleton, what determines the density of states on it (through its moments), and at what order does fermionic exchange become visible? For the N d = 0 N_d = 0 N d = 0 band, the subject of §§5–7, the skeleton is the full 2 n 2n 2 n -dimensional configuration lattice with the doubly occupied points removed; the reduced doublon sectors enter through §§2, 4 and 8. A second, independent line of work treats a many-body problem as a hopping problem on a high-dimensional Fock-space or configuration graph, most influentially for quasiparticle lifetimes and many-body localization [14 , 15 ] ; we use the same language below, likewise established rather than novel.
The precedent most conceptually adjacent to §6 is the strong-coupling single-hole problem: Brinkman and Rice [16 ] obtained hole-band spectral moments in the U → ∞ U \to\infty U → ∞ Hubbard model by counting hopping paths that return the carrier to its origin while restoring or altering the surrounding spin background, and ten Haaf et al. [17 ] made this explicit in closed-walk/permutation language: a closed walk induces a permutation of the background spins, decomposable into disjoint cycles, with the background restored precisely when every spin within a nontrivial cycle shares the same orientation (see also [18 ] ). That closed-walk→permutation-cycle connection is established, not new here. What differs is the spectral object: their moments concern a single carrier hopping through a thermodynamic-limit spin background, evaluated probabilistically through the background magnetization, whereas we study the eigenvalue distribution of the entire fixed-particle-number, no-double-occupancy kinetic operator at finite N N N and arbitrary ( N ↑ , N ↓ ) (N_{\uparrow}, N_{\downarrow}) ( N ↑ , N ↓ ) . Turning that path-counting idea into exact finite-N N N composition laws, namely that four projected hops admit only the identity, transpositions, and three-cycles, forcing μ 4 = A ( N , n ) + α ( N , n ) P s a m e \mu_4 = A(N,n) + \alpha(N,n)P_{\mathrm{same}} μ 4 = A ( N , n ) + α ( N , n ) P same exactly, with α ( N , n ) \alpha(N,n) α ( N , n ) derived in closed form, is the content of § 6 .
Spectral-moment methods have a separate long history in Hubbard-model theory, from Harris and Lange’s moment expansion of the single-particle spectral function [19 ] through exact strong-coupling spectral-moment constraints in modern work [20 ] ; both also use a strong-coupling no-double-occupancy projection, so projection itself should likewise not be read as novel here. That prior work characterizes the one-particle addition/removal spectral weight; the traces considered here instead characterize the eigenvalue distribution of the projected many-body kinetic operator at fixed particle number:
M ℓ s p ( k ) = ∫ d ω ω ℓ A ( k , ω ) v s . μ ℓ ( n ) = 1 D 0 Tr H n ( 0 ) [ ( P 0 H t P 0 ) ℓ ] , (1) M_{\ell}^{\mathrm{sp}}(\mathbf{k}) = \int d\omega\,\omega^{\ell}\mathcal{A}(\mathbf{k},\omega) \quad\mathrm{vs.} \quad\mu_{\ell}^{(n)} = \frac{1}{D_0}\operatorname{Tr}_{\mathcal{H}_{n}^{(0)}}\left[(P_0 H_t P_0)^{\ell}\right], \tag*{(1)} M ℓ sp ( k ) = ∫ d ω ω ℓ A ( k , ω ) vs. μ ℓ ( n ) = D 0 1 Tr H n ( 0 ) [ ( P 0 H t P 0 ) ℓ ] , (1)
with A \mathcal{A} A the single-particle spectral function, H n ( 0 ) \mathcal{H}_{n}^{(0)} H n ( 0 ) the n n n -electron no-double-occupancy space of one spin composition ( N ↑ , N ↓ ) (N_{\uparrow},N_{\downarrow}) ( N ↑ , N ↓ ) , and D 0 D_0 D 0 its dimension. M ℓ s p M_{\ell}^{\mathrm{sp}} M ℓ sp probes the frequency distribution of adding or removing a particle from a many-body state; μ ℓ ( n ) \mu_{\ell}^{(n)} μ ℓ ( n ) is a moment of the eigenvalue distribution of the projected kinetic operator itself, restricted to fixed n n n and doublon occupation, so that individual closed walks can be classified directly by real-space occupancy and exchange. Both involve hopping-path combinatorics, but they concern inequivalent spectral measures.
Principal results. After establishing the two-particle doublon problem as a benchmark (§ 2 ), we address two questions, explicitly for n = 2 , 3 , 4 n = 2, 3, 4 n = 2 , 3 , 4 fermions and for every n n n where the arguments allow (how many independent translational coordinates remain after strong-coupling projection, and how does fermionic exchange modify the spectral moments on the resulting configuration graph) and then extend the static analysis to a gauge field. Item 1 below organizes the analysis; items 2–5 are the four results of the abstract:
A cluster-coordinate counting rule (§ 3 ): n n n fermions in C C C independently-mobile clusters occupy a d c l = 2 C d_{\mathrm{cl}} = 2C d cl = 2 C coordinate skeleton, specializing to d c l = 2 ( n − N d ) d_{\mathrm{cl}} = 2(n-N_d) d cl = 2 ( n − N d ) for doublons plus unpaired electrons. The sector structure it implies (which N d N_d N d bands exist, and their centres near N d ε d N_d\varepsilon_d N d ε d ) is verified by exact diagonalization for n = 3 , 4 n = 3, 4 n = 3 , 4 (§ 4 ).
An exact second-moment theorem (§ 5 ), valid for any spin-independent Hermitian hopping amplitudes (t j i = t i j ∗ t_{ji}=t_{ij}^{*} t j i = t ij ∗ , no on-site terms) on N ≥ 2 N \ge2 N ≥ 2 sites, μ 2 ( U → ∞ ) = n ( N − n ) N ( N − 1 ) ∑ i ≠ j ∣ t i j ∣ 2 \mu_2(U\to\infty)=\frac{n(N-n)}{N(N-1)}\sum_{i\ne j}|t_{ij}|^2 μ 2 ( U → ∞ ) = N ( N − 1 ) n ( N − n ) ∑ i = j ∣ t ij ∣ 2 , independent of spin composition; on a z z z -regular lattice, μ 2 = λ 2 n z ( N − n ) / ( N − 1 ) \mu_2=\lambda^2nz(N-n)/(N-1) μ 2 = λ 2 n z ( N − n ) / ( N − 1 ) .
Where exchange enters the fourth moment (§ 6 ): μ 4 \mu_4 μ 4 depends on spin composition only through an exactly linear dependence on the same-spin pair count P s a m e P_{\mathrm{same}} P same , on any finite graph, because four projected hops admit only the identity, a transposition, or a 3-cycle (never a four-particle permutation, shown combinatorially unreachable in four hops). For uniform
hopping the slope is exact in closed form at every n n n on any finite simple graph with N ≥ 4 N \ge4 N ≥ 4 , α ( N , n ) = − 8 C 4 ( N − n ) ( 8 N − 11 n − 2 ) / [ N ( N − 1 ) ( N − 2 ) ( N − 3 ) ] \alpha(N,n)=-8C_4(N-n)(8N-11n-2)/[N(N-1)(N-2)(N-3)] α ( N , n ) = − 8 C 4 ( N − n ) ( 8 N − 11 n − 2 ) / [ N ( N − 1 ) ( N − 2 ) ( N − 3 )] with C 4 C_4 C 4 the number of 4-cycles: transpositions around a 4-cycle (on the square lattice, two geometrically distinct ring-exchange channels) plus 3-cycles of three particles around one. On the L × L L \times L L × L torus with L = 3 L=3 L = 3 or L ≥ 5 L \ge5 L ≥ 5 , α ( N , 2 ) = − 64 / ( N − 1 ) \alpha(N,2)=-64/(N-1) α ( N , 2 ) = − 64/ ( N − 1 ) , and on hypercubic tori whose sides are all 3 or ≥ 5 \ge5 ≥ 5 , α d ( N , 2 ) = − 16 ( d − 1 ) z / ( N − 1 ) \alpha_d(N,2)=-16(d-1)z/(N-1) α d ( N , 2 ) = − 16 ( d − 1 ) z / ( N − 1 ) .
A Schrieffer–Wolff expansion at finite U U U (§ 7 ): the 1 / U 1/U 1/ U contribution to μ 2 \mu_2 μ 2 vanishes identically on triangle-free graphs, including the torus with L ≥ 4 L \ge4 L ≥ 4 (a proven cancellation), and the U − 2 U^{-2} U − 2 coefficient, which includes a third-order term, is an exact trace confirmed by exact diagonalization; numerically, at low filling it is negative (the band narrows) and close to proportional to the opposite-spin pair count N ↑ N ↓ N_{\uparrow}N_{\downarrow} N ↑ N ↓ (at fixed n n n , equivalently close to affine in P s a m e P_{\mathrm{same}} P same with its exact zero at a single spin species), and near half filling it turns positive.
A gauge-field hierarchy (§ 8 ): the flux dependence of the exchange slope is exact at n = 2 n=2 n = 2 (L × L L \times L L × L tori with L = 3 L=3 L = 3 or L ≥ 5 L \ge5 L ≥ 5 ), α ( N , 2 , Φ ) = − 64 N − 1 cos Φ □ \alpha(N,2,\Phi)=-\frac{64}{N-1}\cos\Phi_{\square} α ( N , 2 , Φ ) = − N − 1 64 cos Φ □ , sharing one ordinary charge-e e e harmonic between orbital and exchange loops (exchange changes interference parity, not charge); the Schrieffer–Wolff-projected doublon hopping itself carries a doubled phase, t d e 2 i A i j t_{\mathrm{d}}e^{2iA_{ij}} t d e 2 i A ij , so that the second-order effective single-doublon Hamiltonian is exactly π \pi π -periodic (h / 2 e h/2e h /2 e ) under a boundary twist, a periodicity the exact two-electron U-band follows up to corrections of order ( 2 λ ) L / U L − 1 (2\lambda)^L/U^{L-1} ( 2 λ ) L / U L − 1 .
A closely related question, whether two doublons bind into a larger composite, is a self-contained problem with its own subtleties, which we leave to separate work, in the spirit of [8 ] . The single-doublon charge-2 e 2e 2 e result above is the static half of a longer program. The dynamical extension (driven pair-transport and current-response observables, and ultimately many-pair phase coherence) remains future work (§ 9 ).
Machine-checked results. The second-moment theorem, the four-hop permutation bound and the three-cycle identity behind the linearity of μ 4 \mu_4 μ 4 (the linearity itself, which also needs a sign lemma, and the closed form of α ( N , n ) \alpha(N,n) α ( N , n ) are not formalized), the vanishing cross term, the doubled doublon phase, and the finite-population identities are also formalized and proved in the Lean 4 proof assistant [21 ] with its mathematical library Mathlib [22 ] , with the projected operator H ∞ H_{\infty} H ∞ built there from fermion operators satisfying the canonical anticommutation relations. Appendix § 16 lists every machine-checked statement and states precisely what it does not cover; none of the numerical results depends on it. Derivations too long for the paper are collected in the Supplemental Material, twelve notes cited as S1–S12: S1 configuration space and densities of states, S2 the doublon effective Hamiltonian, S3 the second-moment theorem, S4 finite-population corrections, S5 the exchange topology of μ 4 \mu_4 μ 4 , S6 lattice specificity and finite size, S7 the Schrieffer–Wolff expansion, S8 the fermion sign rule, S9–S11 the gauge-field extension, and S12 the Lean statements and definitions.
Model and the two-particle benchmark
Two electrons occupy a periodic L × L L \times L L × L square lattice of N = L 2 N=L^2 N = L 2 sites with nearest-neighbor hopping λ \lambda λ and on-site Hubbard repulsion U U U , the sum running over nearest-neighbor bonds ⟨ i j ⟩ \langle ij\rangle ⟨ ij ⟩ and both spin species σ ∈ { ↑ , ↓ } \sigma\in\{\uparrow,\downarrow\} σ ∈ { ↑ , ↓ }
H = − λ ∑ ⟨ i j ⟩ , σ ( c i σ † c j σ + h . c . ) + U ∑ i n i ↑ n i ↓ . (2) H=-\lambda\sum_{\langle ij\rangle,\sigma}(c_{i\sigma}^{\dagger}c_{j\sigma}+\mathrm{h.c.})+U\sum_i n_{i\uparrow}n_{i\downarrow}. \tag*{(2)} H = − λ ⟨ ij ⟩ , σ ∑ ( c iσ † c j σ + h.c. ) + U i ∑ n i ↑ n i ↓ . (2)
Two free (U = 0 U = 0 U = 0 ) electrons in 2D are exactly a single particle hopping on a 4-dimensional hypercubic lattice, since Ψ ( x 1 , y 1 , x 2 , y 2 ) \Psi(x_1, y_1, x_2, y_2) Ψ ( x 1 , y 1 , x 2 , y 2 ) depends on four independent coordinates with additive dispersion
E = − 2 λ ∑ α = 1 4 cos k α . E = -2\lambda\sum_{\alpha=1}^{4} \cos k_\alpha. E = − 2 λ α = 1 ∑ 4 cos k α .
This gives an exact, parameter-free prediction for the U = 0 U = 0 U = 0 density of states via the Bessel-integral representation of Appendix § 11 , confirmed by dense exact diagonalization at L = 7 L = 7 L = 7 (the lattice of Souza and Claro’s original density-of-states figure [3 ] ) and, at L = 9 L = 9 L = 9 , sharpened to a near-exact match by twisted-boundary-condition averaging (Figure 1 ).
Figure 1. Two-electron density of states at U = 0 U = 0 U = 0 (one electron of each spin) for an L = 9 L = 9 L = 9 periodic square lattice. The fixed-boundary spectrum contains 6561 states; averaging over 300 independent boundary twists gives 1,968,300 eigenvalues. The twist-averaged result approaches the exact continuum ρ 4 D ( E ) \rho_{4D}(E) ρ 4 D ( E ) , the free 4D configuration space made explicit.
The doublon effective Hamiltonian. At fixed total momentum and every U U U , the pair’s reduced (relative-coordinate) equation is exactly the textbook Koster–Slater bound-state-at-a-repulsive-defect problem [23 ] . Expanding for large U U U and converting to the physical pair momentum ( K x , K y ) (K_x, K_y) ( K x , K y ) gives, to leading order in λ / U \lambda/U λ / U , an effective 2D tight-binding band,
E p a i r ( K x , K y ) ≈ ε d + 2 t d ( cos K x + cos K y ) , ε d = U + 8 λ 2 U , t d = 2 λ 2 U , (3) E_{\mathrm{pair}}(K_x,K_y) \approx\varepsilon_d + 2t_d(\cos K_x + \cos K_y), \qquad\varepsilon_d = U + \frac{8\lambda^2}{U}, \qquad t_d = \frac{2\lambda^2}{U}, \tag*{(3)} E pair ( K x , K y ) ≈ ε d + 2 t d ( cos K x + cos K y ) , ε d = U + U 8 λ 2 , t d = U 2 λ 2 , (3)
i.e. a composite boson hopping on the same L × L L \times L L × L lattice with its own 2D dispersion, inverted relative to a bare particle’s − 2 λ ( cos K x + cos K y ) -2\lambda(\cos K_x + \cos K_y) − 2 λ ( cos K x + cos K y ) sign convention: the doublon band has its maximum, not its minimum, at K = 0 K = 0 K = 0 . This sign is fixed, not a free convention. Direct second-order degenerate perturbation theory (the N N N -fold-degenerate manifold { ∣ r , r ⟩ } \{|r,r\rangle\} { ∣ r , r ⟩} at energy U U U (zeroth order), coupled through the continuum of states with no doubly occupied site, N d = 0 N_d = 0 N d = 0 , at energy 0 by two sequential single-electron hops) has exactly two contributing virtual paths between adjacent doublon sites r , s r,s r , s : ↑ \uparrow ↑ hops r → s r \to s r → s first, then ↓ \downarrow ↓ ; or the reverse order. Each contributes ( − λ ) ( − λ ) / ( U − 0 ) = + λ 2 / U (-\lambda)(-\lambda)/(U - 0) = +\lambda^2/U ( − λ ) ( − λ ) / ( U − 0 ) = + λ 2 / U to the effective hopping matrix element, with no additional fermion sign (a single particle per spin species carries none, under the sorted-tuple convention of Appendix E), giving a net + 2 λ 2 / U +2\lambda^2/U + 2 λ 2 / U , the same sign as λ 2 \lambda^2 λ 2 , not its negative. This positive matrix element Fourier-transforms to + 2 t d ( cos K x + cos K y ) +2t_d(\cos K_x + \cos K_y) + 2 t d ( cos K x + cos K y ) , opposite in sign to the bare kinetic term, whose convention is H t = − λ ∑ ( ⋯ ) H_t = -\lambda\sum(\cdots) H t = − λ ∑ ( ⋯ ) . The sign is corroborated independently by gauge covariance: the doublon operator d r = c r ↓ c r ↑ d_r = c_{r\downarrow}c_{r\uparrow} d r = c r ↓ c r ↑ transforms as d r → e 2 i θ r d r d_r \to e^{2i\theta_r}d_r d r → e 2 i θ r d r under a local phase redefinition c r σ → e i θ r c r σ c_{r\sigma} \to e^{i\theta_r}c_{r\sigma} c r σ → e i θ r c r σ , twice the phase of a single fermion, so on a bipartite lattice (even L L L ) the staggered transformation (θ = 0 , π \theta= 0,\pi θ = 0 , π on the two sublattices) that reverses the bare hopping’s sign leaves the doublon’s derived hopping sign unchanged (e 2 i π = + 1 e^{2i\pi} = +1 e 2 iπ = + 1 ), ruling out the naive expectation that the two signs are gauge-equivalent. This inverted dispersion is the standard, if underappreciated, signature of a pair bound by repulsion : the doublon descends from the upper Hubbard band. It is derived in full, together with both consistency checks above, in S2. The positive sign is also consistent with the exact one-dimensional repulsively bound-pair dispersion [4 ] , whose strong-coupling expansion gives E p a i r = ε d + 2 t d cos K C M + O ( λ 4 / U 3 ) E_{\mathrm{pair}} = \varepsilon_d + 2t_d \cos K_{\mathrm{CM}} + O(\lambda^4/U^3) E pair = ε d + 2 t d cos K CM + O ( λ 4 / U 3 ) with t d = 2 λ 2 / U t_d = 2\lambda^2/U t d = 2 λ 2 / U (and ε d = U + 4 λ 2 / U \varepsilon_d = U + 4\lambda^2/U ε d = U + 4 λ 2 / U in one dimension), an independent, exact two-body corroboration of the sign correction above, not merely a consistency check on our own perturbative bookkeeping. For the same one-dimensional problem, a semiclassical treatment at leading order in λ / U \lambda/U λ / U gives a doubled Bloch-oscillation frequency under an applied field, which that work reads as compatible with the pair’s composite charge 2 e 2e 2 e [4 ] ; we return to this, and to its relation to the gauge-covariant result of § 8 , below. This particular perturbative form is by now a standard type of result; comparable effective doublon Hamiltonians have recently been derived for an extended, mass-imbalanced Hubbard model [6 ] and for a periodically driven one [5 ] , and our contribution here is not the functional form itself but the zero-fitted-parameter connection between it and the exact two-electron U-band on this specific lattice: the predicted band centre matches the observed mean to within 0.06% (at most 0.03 t d 0.03t_d 0.03 t d ), and the predicted band edge matches the observed maximum to within 0.5% (at most 0.26 t d 0.26t_d 0.26 t d ) (U = 10 , 15 , 20 λ U = 10,15,20\lambda U = 10 , 15 , 20 λ , L = 7 L = 7 L = 7 , the top N N N levels of the spectrum), and the full U-band collapses onto the predicted 2D dispersion once plotted in the doublon’s own reduced variables (Figure 2 , from a relative-coordinate momentum scan, Appendix E). The band-centre test and the collapse are insensitive to the sign of t d t_d t d on any lattice: the grid average of cos K \cos K cos K vanishes, and the two signs give mirror-image densities of states about ε d \varepsilon_d ε d , which deviate equally from the symmetric 2D density. The band edge is not: on the odd L = 7 L = 7 L = 7 lattice ( π , π ) (\pi,\pi) ( π , π ) is not an allowed momentum, so the opposite sign would place the maximum at ε d + 4 t d cos ( π / 7 ) \varepsilon_d + 4t_d\cos(\pi/7) ε d + 4 t d cos ( π /7 ) rather than ε d + 4 t d \varepsilon_d + 4t_d ε d + 4 t d .1 The observed maximum favours the derived sign, deviating from its prediction by 0.45%, 0.11% and 0.04% (0.26, 0.13 and 0.08 t d 0.08t_d 0.08 t d ) at U = 10 , 15 , 20 λ U = 10,15,20\lambda U = 10 , 15 , 20 λ , against 1.14%, 0.44% and 0.23% (0.66, 0.53 and 0.47 t d 0.47t_d 0.47 t d ) for the opposite sign. This benchmark is the starting point for the rest of the paper.
Figure 2. The two-electron U-band at U = 10 , 15 , 20 , 50 λ U = 10,15,20,50\lambda U = 10 , 15 , 20 , 50 λ , from a relative-coordinate scan (relative lattice L r e l = 21 L_{\mathrm{rel}} = 21 L rel = 21 , 48 × 48 48 \times48 48 × 48 pair momenta K K K ; Appendix E), histogrammed (50 bins, unit area) in the reduced variable x = ( E − ε d ) / ( 2 t d ) x = (E-\varepsilon_d)/(2t_d) x = ( E − ε d ) / ( 2 t d ) : ρ ~ U ( x ) \widetilde{\rho}_U(x) ρ U ( x ) is the U-band density of states in that variable. It collapses onto the 2D tight-binding density of states of Eq. (3) in the same variable, with t d t_d t d and ε d \varepsilon_d ε d predicted from U U U , λ \lambda λ and no fitted parameters: the n = 2 n = 2 n = 2 benchmark for everything that follows.
Cluster-coordinate counting
A coordinate count, not a spectral claim
Ignoring interactions and exchange entirely, n n n particles in 2D, Ψ ( x 1 , y 1 , … , x n , y n ) \Psi(x_1,y_1,\ldots,x_n,y_n) Ψ ( x 1 , y 1 , … , x n , y n ) , live on a configuration space that is exactly a single particle hopping on a 2 n 2n 2 n -dimensional hypercubic lattice, ρ n ( 0 ) ( E ) = ρ 2 n D ( E ) \rho_n^{(0)}(E) = \rho_{2nD}(E) ρ n ( 0 ) ( E ) = ρ 2 n D ( E ) (Appendix A). This is pure kinematics, a relabeling of coordinates: if C C C independently translating clusters move in two spatial dimensions, then by definition they carry 2 C 2C 2 C centre-of-mass coordinates. We state this plainly because a natural objection is that nothing has yet been proved: coordinate counting alone is not a dynamical result. We use d c l d_{\mathrm{cl}} d cl exclusively as a coordinate-counting dimension : the number of independent real-space coordinates needed to translate C C C composite or unpaired objects in two spatial dimensions. It is not the dimension of the many-body Hilbert space, the spectral dimension of the configuration graph, or a scaling dimension, and should not be confused with any of these. The nontrivial physical content begins only once we ask which sectors the actual Hubbard spectrum organizes into, and what spectral measure lives on the resulting skeleton: the subjects of §4 onward.
From coordinate count to cluster rule
Each pairwise on-site interaction imposes x i = x j x_i = x_j x i = x j , y i = y j y_i = y_j y i = y j , two constraints, reducing the manifold’s dimension by 2, consistent with 4 D → 2 D 4\mathrm{D} \to2\mathrm{D} 4 D → 2 D at n = 2 n = 2 n = 2 . More generally, if n n n particles organize into C C C independently-mobile clusters (a partition n = n 1 + ⋯ + n C n = n_1 + \cdots+ n_C n = n 1 + ⋯ + n C ), each cluster carries exactly one 2D centre-of-mass coordinate, so
d c l = 2 C ( codimension 2 ( n − C ) ) . (4) d_{\mathrm{cl}} = 2C \qquad(\text{codimension }2(n-C)). \tag*{(4)} d cl = 2 C ( codimension 2 ( n − C )) . (4)
For n = 4 n = 4 n = 4 : four free electrons (C = 4 C = 4 C = 4 ) occupy 8D; two doublons (C = 2 C = 2 C = 2 ) occupy 4D; an (off-site) bound tetramer (C = 1 C = 1 C = 1 , if one exists) would occupy 2D.
Single-band Hubbard physics forbids on-site clusters larger than 2 (only one ↑ \uparrow ↑ and one ↓ \downarrow ↓ electron fit per site), so composites built purely from doublons plus unpaired electrons obey a testable rule in terms of the doublon occupation N d N_d N d alone:
d c l = 2 ( n − N d ) . (5) d_{\mathrm{cl}} = 2(n - N_d). \tag*{(5)} d cl = 2 ( n − N d ) . (5)
Sector structure for n = 3 , 4 n = 3, 4 n = 3 , 4
Below, a composition is a fixed ( N ↑ , N ↓ ) (N_{\uparrow}, N_{\downarrow}) ( N ↑ , N ↓ ) ; its states with N d = k N_d = k N d = k doubly occupied sites form the N d = k N_d = k N d = k sector, whose energies form the N d = k N_d = k N d = k band at large U U U ; and a momentum block is a fixed total momentum K K K .
Figure 3 tests the sector structure implied by the counting rule for n = 3 , 4 n = 3, 4 n = 3 , 4 : which N d N_d N d bands exist, and where they are centred. Each eigenstate of the full exact spectrum is assigned to the band N d = k N_d = k N d = k with ∣ ⟨ N d ⟩ − k ∣ < 1 / 2 \lvert\langle N_d\rangle- k\rvert< 1/2 ∣⟨ N d ⟩ − k ∣ < 1/2 . In every case the band centres E ˉ N d \bar{E}_{N_d} E ˉ N d are confirmed against N d ε d N_d\varepsilon_d N d ε d , the leading isolated-cluster prediction obtained by simply adding N d N_d N d copies of the single-doublon self-energy: to ≤ 0.10 % \leq0.10\% ≤ 0.10% at U = 40 λ U = 40\lambda U = 40 λ (0.02 % 0.02\% 0.02% for n = 3 n = 3 n = 3 ; 0.10 % 0.10\% 0.10% and 0.03 % 0.03\% 0.03% for N d = 1 , 2 N_d = 1, 2 N d = 1 , 2 at n = 4 n = 4 n = 4 ), and to ≤ 1.1 % \leq1.1\% ≤ 1.1% already at U = 10 λ U = 10\lambda U = 10 λ . This residual is not numerical noise or a higher-order 1 / U 2 1/U^2 1/ U 2 tail: exact diagonalization at U = 10 , 15 , 20 , 30 , 40 , 60 λ U = 10, 15, 20, 30, 40, 60\lambda U = 10 , 15 , 20 , 30 , 40 , 60 λ shows U ⋅ ( E ˉ N d − N d ε d ) U \cdot(\bar{E}_{N_d} - N_d\varepsilon_d) U ⋅ ( E ˉ N d − N d ε d ) converging to a nonzero constant as U U U grows (not U 2 ⋅ ( … ) U^2 \cdot(\ldots) U 2 ⋅ ( … ) , which instead diverges), confirming a genuine O ( λ 2 / U ) O(\lambda^2/U) O ( λ 2 / U ) many-body correction: with more than one cluster present, other particles can block or modify some of the virtual hopping processes that generate the isolated-pair shift 8 λ 2 / U 8\lambda^2/U 8 λ 2 / U , and this finite-density correction is not yet captured in closed form. For N d = 1 N_d = 1 N d = 1 in the n = 3 n = 3 n = 3 (L = 5 L = 5 L = 5 ) and n = 4 n = 4 n = 4 (L = 4 L = 4 L = 4 ) sectors this constant is ≈ − 0.33 λ 2 \approx-0.33\lambda^2 ≈ − 0.33 λ 2 and ≈ − 1.60 λ 2 \approx-1.60\lambda^2 ≈ − 1.60 λ 2 respectively (measured at U = 60 λ U = 60\lambda U = 60 λ ); for N d = 2 N_d = 2 N d = 2 in the n = 4 n = 4 n = 4 sector it is ≈ − 1.06 λ 2 \approx-1.06\lambda^2 ≈ − 1.06 λ 2 , still visibly converging at that U U U . Being finite-density corrections, these constants depend on L L L .
Figure 3. Sector structure implied by the counting rule, tested by full exact diagonalization of every interacting composition. All panels share the energy axis of (a), E E E in units of λ \lambda λ . In (a,b), color encodes the doublon expectation value ⟨ N d ⟩ \langle N_d\rangle ⟨ N d ⟩ on a continuous white-to-green scale (the manuscript’s virtual-doublon color; darker = more doublons), never a proxy for U U U . In (a,b) the horizontal axis is the true, unjittered ⟨ N d ⟩ \langle N_d\rangle ⟨ N d ⟩ of each eigenstate (small pale points: the full spectrum, a random subset of 4000 levels where it is larger; marker shape distinguishes U = 10 , 20 , 40 λ U = 10,20,40\lambda U = 10 , 20 , 40 λ ); bold markers with black edges mark each band’s exact mean energy over all its states (a state belongs to band N d = k N_d = k N d = k when ∣ ⟨ N d ⟩ − k ∣ < 1 / 2 |\langle N_d\rangle- k| < 1/2 ∣ ⟨ N d ⟩ − k ∣ < 1/2 ), and the short dotted horizontal ticks mark the independently predicted sector center (0 0 0 , ε d \varepsilon_d ε d , or 2 ε d 2\varepsilon_d 2 ε d ): coincidence of marker and tick is the closure test, not a fitted line. (a) n = 3 n = 3 n = 3 (2 ↑ + 1 ↓ 2 \uparrow+1 \downarrow 2 ↑ + 1 ↓ , L = 5 L = 5 L = 5 ): the ⟨ N d ⟩ ≈ 1 \langle N_d\rangle\approx1 ⟨ N d ⟩ ≈ 1 sector separates cleanly, converging to ε d \varepsilon_d ε d . (b) n = 4 n = 4 n = 4 (2 ↑ + 2 ↓ 2 \uparrow+2 \downarrow 2 ↑ + 2 ↓ , L = 4 L = 4 L = 4 ): the same test, now with three sectors (N d = 0 , 1 , 2 N_d = 0,1,2 N d = 0 , 1 , 2 ). (c) n = 4 n = 4 n = 4 , three spin compositions (N ↑ , N ↓ N_{\uparrow},N_{\downarrow} N ↑ , N ↓ ) = ( 4 , 0 ) , ( 3 , 1 ) , ( 2 , 2 ) (4,0),(3,1),(2,2) ( 4 , 0 ) , ( 3 , 1 ) , ( 2 , 2 ) at fixed U = 40 λ U = 40\lambda U = 40 λ ; the ( 4 , 0 ) (4,0) ( 4 , 0 ) column needs no diagonalization, its levels being sums of four distinct single-particle energies, of which a random sample is shown (horizontal position is the categorical composition, not N d N_d N d ; color is the band each state is assigned to): spin composition alone controls which sectors are reachable (( 4 , 0 ) (4,0) ( 4 , 0 ) : only N d = 0 N_d = 0 N d = 0 ; ( 3 , 1 ) (3,1) ( 3 , 1 ) : N d ∈ { 0 , 1 } N_d \in\{0,1\} N d ∈ { 0 , 1 } ; ( 2 , 2 ) (2,2) ( 2 , 2 ) : N d ∈ { 0 , 1 , 2 } N_d \in\{0,1,2\} N d ∈ { 0 , 1 , 2 } ).
n = 3 n = 3 n = 3 : one doublon plus one electron
Taking N ↑ = 2 N_{\uparrow} = 2 N ↑ = 2 , N ↓ = 1 N_{\downarrow} = 1 N ↓ = 1 (L = 5 L = 5 L = 5 ), the energy spectrum separates cleanly by N d N_d N d (Figure 3 a): even at U = 10 λ U = 10\lambda U = 10 λ the N d ≈ 1 N_d \approx1 N d ≈ 1 cloud is visually separated from the N d ≈ 0 N_d \approx0 N d ≈ 0 continuum, centred almost exactly at ε d \varepsilon_d ε d ; by U = 40 λ U = 40\lambda U = 40 λ the upper-band mean is within 0.02 % 0.02\% 0.02% of ε d = U + 8 λ 2 / U \varepsilon_d = U + 8\lambda^2/U ε d = U + 8 λ 2 / U . The relative coordinate is unbound, not a genuine three-body composite. Total momentum K K K is conserved, and in every momentum block K K K every N d ≈ 1 N_d \approx1 N d ≈ 1 level lies inside the doublon–electron continuum ε d + 2 t d [ cos ( K x − q x ) + cos ( K y − q y ) ] − 2 λ ( cos q x + cos q y ) \varepsilon_d + 2t_d[\cos(K_x - q_x) + \cos(K_y - q_y)] - 2\lambda(\cos q_x + \cos q_y) ε d + 2 t d [ cos ( K x − q x ) + cos ( K y − q y )] − 2 λ ( cos q x + cos q y ) (over all q q q ): in exact diagonalization at U = 30 λ U = 30\lambda U = 30 λ for L ≤ 9 L \leq9 L ≤ 9 , and, with t d → 0 t_d \to0 t d → 0 , up to L = 32 L = 32 L = 32 in the first-order (U → ∞ U \to\infty U → ∞ ) model of a static doublon and a hopping electron, which reproduces the exact band to O ( λ 2 / U ) O(\lambda^2/U) O ( λ 2 / U ) . The only short-range process at this order, the doublon’s ↓ \downarrow ↓ electron hopping onto an adjacent electron’s site (which exchanges the two), produces a resonance inside the band near E − ε d ≈ + 1.2 λ E - \varepsilon_d \approx+1.2\lambda E − ε d ≈ + 1.2 λ at K = 0 K = 0 K = 0 (and − 1.2 λ -1.2\lambda − 1.2 λ at K = ( π , π ) K = (\pi,\pi) K = ( π , π ) , by the exact symmetry of the first-order model on even L L L that maps K → K + ( π , π ) K \to K + (\pi,\pi) K → K + ( π , π ) and E − ε d → − ( E − ε d ) E - \varepsilon_d \to-(E - \varepsilon_d) E − ε d → − ( E − ε d ) , a sublattice sign change), not a bound state: its weight on the adjacent configurations spreads over more and more eigenstates as L L L grows (in the first-order model, computed for this purpose up to L = 64 L = 64 L = 64 , the largest single-state weight falls from 0.36 0.36 0.36 at L = 8 L = 8 L = 8 to 0.034 0.034 0.034 at L = 64 L = 64 L = 64 ). Such doublon-plus-particle configurations have recently been studied dynamically on quantum hardware, including the same N ↑ = 2 , N ↓ = 1 N_{\uparrow} = 2, N_{\downarrow} = 1 N ↑ = 2 , N ↓ = 1 sector [7 ] ; here we instead use them as a controlled static test of the counting rule.
n = 4 n = 4 n = 4 : spin composition as a sector-control knob
Taking N ↑ = N ↓ = 2 N_{\uparrow} = N_{\downarrow} = 2 N ↑ = N ↓ = 2 (L = 4 L = 4 L = 4 ), the predicted three-sector hierarchy sharpens with increasing U U U : three bands at E ≈ 0 , U , 2 U E \approx0, U, 2U E ≈ 0 , U , 2 U (Figure 3 b), with the N d ≈ 1 , 2 N_d \approx1, 2 N d ≈ 1 , 2 means converging to ε d \varepsilon_d ε d and 2 ε d = 2 U + 16 λ 2 / U 2\varepsilon_d = 2U + 16\lambda^2/U 2 ε d = 2 U + 16 λ 2 / U (within 0.10% and 0.03% at U = 40 λ U = 40\lambda U = 40 λ , where the bare values U U U and 2 U 2U 2 U are still 0.40% and 0.47% away); N d = 0 , 1 , 2 N_d = 0,1,2 N d = 0 , 1 , 2 correspond to C = 4 , 3 , 2 C = 4,3,2 C = 4 , 3 , 2 clusters, the predicted 8D → \to → 6D → \to → 4D sequence, and all three bands are present in a single spectrum. Repeating at three spin compositions, same U = 40 λ U = 40\lambda U = 40 λ (Figure 3 c), shows spin composition directly controls which sectors exist: 4 ↑ 4 \uparrow 4 ↑ is a single band at N d = 0 N_d = 0 N d = 0 (no interaction is even possible with one spin species); 3 + 1 3+1 3 + 1 shows exactly two bands (N d ∈ { 0 , 1 } N_d \in\{0,1\} N d ∈ { 0 , 1 } ); 2 + 2 2+2 2 + 2 shows exactly three (N d ∈ { 0 , 1 , 2 } N_d \in\{0,1,2\} N d ∈ { 0 , 1 , 2 } ).
The detailed spectral shape within a sector is a separate matter, addressed from § 5 onward: coordinate counting alone does not determine the many-body density of states, since fermionic projection changes the spectral measure on the same coordinate skeleton. For one electron of each spin at L = 5 L = 5 L = 5 , for example, the N d = 0 N_d = 0 N d = 0 band has variance 23 λ 2 / 3 23\lambda^2/3 23 λ 2 /3 against 8 λ 2 8\lambda^2 8 λ 2 for the free four-dimensional lattice with the same coordinate count, and excess kurtosis μ 4 / μ 2 2 − 3 = − 204 / 529 ≈ − 0.386 \mu_4/\mu_2^2 - 3 = -204/529 \approx-0.386 μ 4 / μ 2 2 − 3 = − 204/529 ≈ − 0.386 against − 3 / 8 -3/8 − 3/8 (Appendix § 11 ); the variance is fixed for every n n n by the theorem of § 5.2 (Figure 5 a).
Figure 5. μ 2 \mu_2 μ 2 of the N d = 0 N_d = 0 N d = 0 sector, n = 2 , 3 , 4 n = 2, 3, 4 n = 2 , 3 , 4 at L = 4 L = 4 L = 4 , against three increasingly complete references, all judged against the same exact U → ∞ U \to\infty U → ∞ value: (a) independent-particle; (b) same-spin Pauli/FPC; (c) full U = ∞ U = \infty U = ∞ projection. Single-species compositions (filled markers) collapse in panel (b); mixed-spin compositions (open markers) remain off the diagonal there by design, not by failure: the same-spin FPC correction accounts only for momentum-space exclusion within each spin species, so a mixed-spin residual is expected wherever opposite-spin real-space exclusion (the U = ∞ U = \infty U = ∞ Gutzwiller constraint) also matters. That residual measures exactly that missing exclusion, and vanishes only once panel (c) restores it.
We do not extend this sector-by-sector escalation to n = 5 n=5 n = 5 : the results of § 5 –§ 7 below are stated for every n n n (the μ 2 \mu_2 μ 2 theorem, the exact linearity of μ 4 \mu_4 μ 4 and its coefficient α ( N , n ) \alpha(N,n) α ( N , n ) , the vanishing 1 / U 1/U 1/ U term), and larger n n n enter only as checks of them.
Projected configuration graph at U → ∞ U \to\infty U → ∞
Setup
Let H t = − λ ∑ ⟨ i j ⟩ , σ ( c i σ † c j σ + h . c . ) H_t=-\lambda\sum_{\langle ij\rangle,\sigma}(c^\dagger_{i\sigma}c_{j\sigma}+\mathrm{h.c.}) H t = − λ ∑ ⟨ ij ⟩ , σ ( c iσ † c j σ + h.c. ) and let P 0 P_0 P 0 project onto configurations with no double occupancy of any kind. Define H ∞ = P 0 H t P 0 H_\infty=P_0H_tP_0 H ∞ = P 0 H t P 0 , acting on the N d = 0 N_d=0 N d = 0 sector for n = N ↑ + N ↓ n=N_\uparrow+N_\downarrow n = N ↑ + N ↓ electrons on N N N sites of a graph G G G with edge set E E E . H ∞ H_\infty H ∞ is the adjacency operator of a configuration graph : its vertices are n n n -electron no-double-occupancy configurations, its edges single allowed hops, with matrix elements of modulus λ \lambda λ whose signs are fixed by the fermionic ordering of the orbitals (Appendix § 15 ).
Exact second-moment theorem
Every basis state in the N d = 0 N_d=0 N d = 0 sector has zero diagonal energy, since a hop always moves a particle to a distinct, previously empty site, so E ‾ = Tr ( H ∞ ) / D 0 = 0 \overline{E}=\operatorname{Tr}(H_\infty)/D_0=0 E = Tr ( H ∞ ) / D 0 = 0 and μ 2 = Tr ( H ∞ 2 ) / D 0 \mu_2=\operatorname{Tr}(H_\infty^2)/D_0 μ 2 = Tr ( H ∞ 2 ) / D 0 . For a configuration a a a (occupied site set S S S , ∣ S ∣ = n |S|=n ∣ S ∣ = n ), ( H ∞ 2 ) a a = λ 2 M a (H_\infty^2)_{aa}=\lambda^2M_a ( H ∞ 2 ) aa = λ 2 M a , where M a M_a M a counts the allowed single-particle hops out of a a a : the number of directed (occupied site, empty neighbor) pairs. So μ 2 = λ 2 ⟨ M ⟩ \mu_2=\lambda^2\langle M\rangle μ 2 = λ 2 ⟨ M ⟩ , averaged uniformly over the D 0 = ( N n ) ( n N ↑ ) D_0=\binom{N}{n}\binom{n}{N_\uparrow} D 0 = ( n N ) ( N ↑ n ) configurations of the sector. Each n n n -site subset S S S occurs in exactly ( n N ↑ ) \binom{n}{N_\uparrow} ( N ↑ n ) of them (one per spin labelling), and M a M_a M a depends only on S S S , so the average is over a uniformly random n n n -element subset S S S of the N N N sites. This reduces to elementary combinatorics: for any two distinct sites i ≠ j i\ne j i = j , Pr ( i ∈ S , j ∉ S ) = n ( N − n ) / [ N ( N − 1 ) ] \Pr(i\in S,j\notin S)=n(N-n)/[N(N-1)] Pr ( i ∈ S , j ∈ / S ) = n ( N − n ) / [ N ( N − 1 )] , independent of whether i , j i,j i , j happen to be adjacent. Summing over the directed edges of G G G , we obtain:
Theorem (exact second moment; S3). For any finite simple graph G = ( V , E ) G=(V,E) G = ( V , E ) with ∣ V ∣ = N ≥ 2 |V|=N\ge2 ∣ V ∣ = N ≥ 2 , hopping amplitude λ \lambda λ , and 0 ≤ n ≤ N 0\le n\le N 0 ≤ n ≤ N , the U → ∞ U\to\infty U → ∞ projected hopping operator in any spin-composition sector with N ↑ + N ↓ = n N_\uparrow+N_\downarrow=n N ↑ + N ↓ = n satisfies
μ 2 ( U → ∞ ) = λ 2 ⋅ 2 ∣ E ∣ ⋅ n ( N − n ) N ( N − 1 ) (6) \mu_2(U\to\infty)=\frac{\lambda^2\cdot2|E|\cdot n(N-n)}{N(N-1)} \tag*{(6)} μ 2 ( U → ∞ ) = N ( N − 1 ) λ 2 ⋅ 2∣ E ∣ ⋅ n ( N − n ) (6) independent of spin composition.
Here 2 ∣ E ∣ = ∑ v deg ( v ) 2|E|=\sum_v\deg(v) 2∣ E ∣ = ∑ v deg ( v ) is the total number of directed edges. This holds on any finite simple graph , regular or not, since the derivation never used translational invariance, dimensionality, or a fixed coordination number, only that the N d = 0 N_d=0 N d = 0 basis consists of all n n n -site subsets. For a z z z -regular graph, 2 ∣ E ∣ = N z 2|E|=Nz 2∣ E ∣ = N z , giving the form used throughout the rest of this paper:
μ 2 ( U → ∞ ) = λ 2 n z N − n N − 1 . (7) \mu_2(U\to\infty)=\lambda^2nz\frac{N-n}{N-1}. \tag*{(7)} μ 2 ( U → ∞ ) = λ 2 n z N − 1 N − n . (7)
Crucially, N ↑ N_\uparrow N ↑ and N ↓ N_\downarrow N ↓ never enter: M a M_a M a depends only on which sites are occupied, not on how the occupied sites are split between spins, so every spin composition with the same total n n n shares the identical μ 2 ( ∞ ) \mu_2(\infty) μ 2 ( ∞ ) . (L = 4 L=4 L = 4 square lattice, N = 16 N=16 N = 16 , z = 4 z=4 z = 4 , n = 4 n=4 n = 4 : μ 2 = ( 64 / 5 ) λ 2 \mu_2=(64/5)\lambda^2 μ 2 = ( 64/5 ) λ 2 , one numerical example, not the content of the result.)
Generality. We verified the graph-general form directly by exhaustive enumeration on three markedly non-regular graphs (an open chain of degrees 1 and 2, a star graph with one hub of degree N − 1 N-1 N − 1 , and an irregular graph with mixed degrees 1–4) for n = 2 , 3 , 4 n=2,3,4 n = 2 , 3 , 4 ; every case matches λ 2 ⋅ 2 ∣ E ∣ ⋅ n ( N − n ) / [ N ( N − 1 ) ] \lambda^{2}\cdot2|E|\cdot n(N-n)/[N(N-1)] λ 2 ⋅ 2∣ E ∣ ⋅ n ( N − n ) / [ N ( N − 1 )] to machine precision. The square-lattice regular-graph form is therefore a special case of a general graph-combinatorial statement, not a two-dimensional or lattice-specific coincidence, and the enumeration below (Table 1) exercises the regular-graph specialization that the rest of the paper uses. The boxed theorem, and the weighted form below, are also machine-checked, for H ∞ H_{\infty} H ∞ built from fermion operators with any spin-independent Hermitian amplitudes (in particular any Peierls phases) and any ordering of the orbitals, which fixes the fermionic signs (Appendix F).
The same argument extends immediately to a general (possibly asymmetric-looking but Hermitian) weighted hopping matrix H t = − ∑ i ≠ j , σ t i j c i σ † c j σ H_t=-\sum_{i\ne j,\sigma}t_{ij}c^\dagger_{i\sigma}c_{j\sigma} H t = − ∑ i = j , σ t ij c iσ † c j σ (t j i = t i j ∗ t_{ji}=t_{ij}^{*} t j i = t ij ∗ , the same for both spins), since nothing in the derivation used equal hopping amplitudes either: ( H ∞ 2 ) a a = ∑ i ∈ S , j ∉ S ∣ t i j ∣ 2 (H_{\infty}^{2})_{aa}=\sum_{i\in S,j\notin S}|t_{ij}|^{2} ( H ∞ 2 ) aa = ∑ i ∈ S , j ∈ / S ∣ t ij ∣ 2 , and averaging over the uniformly random n n n -subset S S S gives
μ 2 ( U → ∞ ) = n ( N − n ) N ( N − 1 ) ∑ i ≠ j ∣ t i j ∣ 2 . (8) \mu_{2}(U\to\infty)=\frac{n(N-n)}{N(N-1)}\sum_{i\ne j}|t_{ij}|^{2}. \tag*{(8)} μ 2 ( U → ∞ ) = N ( N − 1 ) n ( N − n ) i = j ∑ ∣ t ij ∣ 2 . (8)
Defining the mean coordination z ˉ ≡ 1 N ∑ i ≠ j ∣ t i j ∣ 2 / λ 2 \bar{z}\equiv\frac{1}{N}\sum_{i\ne j}|t_{ij}|^{2}/\lambda^{2} z ˉ ≡ N 1 ∑ i = j ∣ t ij ∣ 2 / λ 2 for some reference scale λ \lambda λ recovers μ 2 = λ 2 z ˉ n ( N − n ) / ( N − 1 ) \mu_{2}=\lambda^{2}\bar{z}n(N-n)/(N-1) μ 2 = λ 2 z ˉ n ( N − n ) / ( N − 1 ) for any weighted graph, regular or not: the theorem is properly a statement about the squared Frobenius norm of the off-diagonal part of the one-body hopping matrix T = ( t i j ) T=(t_{ij}) T = ( t ij ) , ∑ i ≠ j ∣ t i j ∣ 2 \sum_{i\ne j}|t_{ij}|^{2} ∑ i = j ∣ t ij ∣ 2 , factored against the purely combinatorial blocking probability n ( N − n ) / [ N ( N − 1 ) ] n(N-n)/[N(N-1)] n ( N − n ) / [ N ( N − 1 )] ; the regular, equal-amplitude square lattice used throughout this paper is simply the case that makes the factorization z ˉ λ 2 \bar{z}\lambda^{2} z ˉ λ 2 transparent.
Table 1. Exact μ 2 ( U → ∞ ) / λ 2 \mu_{2}(U\to\infty)/\lambda^{2} μ 2 ( U → ∞ ) / λ 2 : the analytic column is Eq. (8); the sparse column is obtained by explicit construction of H ∞ H_{\infty} H ∞ (here in the single-species sector; by the theorem every composition gives the same value) and evaluation of Tr H ∞ 2 / D 0 \operatorname{Tr}H_{\infty}^{2}/D_{0} Tr H ∞ 2 / D 0 , without diagonalization. The two are independent computations, not a formula checked against itself.
Same-spin exclusion at U = 0 U=0 U = 0 versus projection at U = ∞ U=\infty U = ∞
Does Pauli exclusion among same-spin electrons alone account for the reduced variance at U = ∞ U=\infty U = ∞ ? The non-interacting problem answers this, because its variance is also exact in closed form. Non-interacting identical fermions are exactly Slater determinants: for a single spin species, their many-body spectrum is exactly every way of choosing k k k of the N N N single-particle levels (distinct labels, possibly degenerate energies), without replacement (population mean 0, since the hopping matrix is traceless, and variance σ 1 2 = z λ 2 \sigma_{1}^{2}=z\lambda^{2} σ 1 2 = z λ 2 ), giving the classical finite-population-correction (FPC) identity Var = k σ 1 2 ( N − k ) / ( N − 1 ) \operatorname{Var}=k\sigma_{1}^{2}(N-k)/(N-1) Var = k σ 1 2 ( N − k ) / ( N − 1 ) (S4), identical in form to the U = ∞ U=\infty U = ∞ theorem with k → n k\to n k → n , but not in mechanism: at U = 0 U=0 U = 0 exclusion is momentum-space (same-spin antisymmetry, so FPC applies only within one spin species); at U = ∞ U = \infty U = ∞ it is real-space (a spin-blind Gutzwiller constraint on all electrons). Both reduce to the same “one site blocks one neighbor” combinatorics, hence the shared form. An analogous coincidence-pattern classification of the fourth moment (Appendix B) gives the exact μ 4 \mu_4 μ 4 counterpart. Both are verified against exact diagonalization (Table 2 ), and the variance also against full enumeration across L = 3 L = 3 L = 3 –10 10 10 (Figure 4 ): forming all ( N k ) \binom{N}{k} ( k N ) many-body energies of a single species (k = 2 , 3 , 4 k = 2, 3, 4 k = 2 , 3 , 4 , up to 3.9 × 10 6 3.9 \times10^6 3.9 × 1 0 6 of them at L = 10 L = 10 L = 10 ) and taking their variance reproduces k σ 1 2 ( N − k ) / ( N − 1 ) k\sigma_1^2(N-k)/(N-1) k σ 1 2 ( N − k ) / ( N − 1 ) to 10 − 15 10^{-15} 1 0 − 15 (relative) at every L L L . For n = 4 n = 4 n = 4 at L = 4 L = 4 L = 4 this is an exact 20% narrowing relative to the naive prediction, a finite-volume statistical effect (vanishing as L → ∞ L \to\infty L → ∞ ), not a finite-U U U or numerical artifact, since a single spin species has no Hubbard interaction at all. At U = 0 U = 0 U = 0 the n = 2 n = 2 n = 2 opposite-spin pair (no mutual exclusion) shows zero such narrowing: precisely zero Pauli narrowing, not zero finite-size deviation of any kind.
Table 2. Exact finite-population-correction formulas for μ 2 , μ 4 \mu_2,\mu_4 μ 2 , μ 4 at U = 0 U = 0 U = 0 ; “ED error” is the relative residual against independent exact diagonalization, confirming the closed form rather than defining it.
Figure 4. Many-fermion variance of a single spin species at U = 0 U=0 U = 0 , in units of k σ 1 2 k\sigma_1^2 k σ 1 2 (σ 1 2 = z λ 2 \sigma_1^2=z\lambda^2 σ 1 2 = z λ 2 , the variance of the single-particle band), by full enumeration of all ( N k ) \binom{N}{k} ( k N ) many-body energies (markers), against the parameter-free finite-population prediction ( N − k ) / ( N − 1 ) (N-k)/(N-1) ( N − k ) / ( N − 1 ) (line), L = 3 L=3 L = 3 –10 10 10 .
For example, at k = 4 k = 4 k = 4 , N = 16 N = 16 N = 16 , the FPC formula gives μ 2 / λ 2 = k z ( N − k ) / ( N − 1 ) = 4 × 4 × 12 / 15 = 64 / 5 \mu_2/\lambda^2 = kz(N-k)/(N-1) = 4 \times4 \times12/15 = 64/5 μ 2 / λ 2 = k z ( N − k ) / ( N − 1 ) = 4 × 4 × 12/15 = 64/5 directly from the sampling argument above, with no diagonalization required.
Does momentum-space exchange alone repair the naive sector prediction? Figure 5 tests this at every N d = 0 N_d = 0 N d = 0 composition for n = 2 , 3 , 4 n = 2, 3, 4 n = 2 , 3 , 4 (L = 4 L = 4 L = 4 ), judging three progressively more complete predictions against the same exact U → ∞ U \to\infty U → ∞ value of μ 2 \mu_2 μ 2 : (a) the naive coordinate-counting prediction, the variance 2 ( 2 n ) λ 2 = 4 n λ 2 2(2n)\lambda^2 = 4n\lambda^2 2 ( 2 n ) λ 2 = 4 n λ 2 of the free 2 n 2n 2 n -dimensional hypercubic lattice, is uniformly too large; (b) the exact U = 0 U = 0 U = 0 variance, ∑ σ N σ z λ 2 ( N − N σ ) / ( N − 1 ) \sum_\sigma N_\sigma z\lambda^2(N-N_\sigma)/(N-1) ∑ σ N σ z λ 2 ( N − N σ ) / ( N − 1 ) (the FPC identity applied within each spin species), brings single-species compositions exactly onto the diagonal (expected, since FPC is the exact result for one species), but leaves mixed-spin compositions with a residual growing with the degree of spin mixing, the additional real-space exclusion between opposite-spin electrons that only exists at strong coupling; (c) the full U = ∞ U = \infty U = ∞ formula of § 5.2 brings every composition onto the diagonal, as it must, being exact on any graph; panels (a) and (b) show that neither the coordinate count nor same-spin exclusion alone reproduces it. Doublon-containing (N d ≥ 1 N_d \ge1 N d ≥ 1 ) sectors are a separate question, since they only exist at finite U U U ; their band centres agree with N d ε d N_d\varepsilon_d N d ε d to ≤ 0.10 % \le0.10\% ≤ 0.10% at U = 40 λ U = 40\lambda U = 40 λ for every composition tested, ( 2 , 1 ) (2,1) ( 2 , 1 ) , ( 3 , 1 ) (3,1) ( 3 , 1 ) and ( 2 , 2 ) (2,2) ( 2 , 2 ) (§ 4.1 –§ 4.2 ).
Exchange topology in the fourth moment
Writing μ 4 = Tr ( H 4 ) / D 0 = ∑ a , b ( H a b 2 ) 2 / D 0 \mu_4 = \operatorname{Tr}(H^4)/D_0 = \sum_{a,b}(H^2_{ab})^2/D_0 μ 4 = Tr ( H 4 ) / D 0 = ∑ a , b ( H ab 2 ) 2 / D 0 makes the physical content explicit: μ 4 \mu_4 μ 4 measures the squared multiplicity of two-hop connections between configurations, summed over every pair ( a , b ) (a,b) ( a , b ) : the diagonal terms ( a = b ) (a=b) ( a = b ) are backtracking/local connectivity, and the off-diagonal terms measure interference between distinct two-hop paths connecting the same two configurations. Exchange is not one closed-walk type among several but an interference effect: it arises whenever two different particles’ two-hop paths connect the same pair of configurations, since whether the resulting amplitude constructively or destructively interferes with the direct path depends on the fermionic sign of the intervening permutation. The closed-walk classification below organizes the same sum
differently; the two bookkeepings assign the exchange contribution to different terms but give the same same-spin minus opposite-spin difference (Appendix § 13 ).
A length-4 closed walk in the N d = 0 N_d=0 N d = 0 configuration graph returns the occupied site set to itself and induces a permutation of “which original particle ends up where”; we classify walks by the cycle type of that permutation. Hard-core blocking in this sector does not depend on spin, so which walks exist from a configuration depends only on its occupied site set; spin enters only through the requirement that the permutation map every particle onto one of the same spin, so that the walk returns to the identical occupation-number state. Identity walks (a single particle’s self-returning excursions, or several particles’ excursions that each return) are therefore spin-blind. Ring exchange: two occupied sites i , j i,j i , j admit a 4-hop sequence, through two empty sites, that returns the occupied site set to { i , j } \{i,j\} { i , j } but exchanges which particle sits where; it contributes if and only if i , j i,j i , j hold the same spin . This is the first, and for n = 2 n=2 n = 2 the only, spin-dependent closed-walk mechanism; for n ≥ 3 n\ge3 n ≥ 3 , four-hop three-cycles provide a second. The complete second-quantized classification, including every fermionic sign, is given in Appendix § 13 . Throughout, we write μ 4 ( ∞ ) = A ( N , n ) + α ( N , n ) P s a m e \mu_4(\infty)=A(N,n)+\alpha(N,n)P_{\mathrm{same}} μ 4 ( ∞ ) = A ( N , n ) + α ( N , n ) P same in units of λ 4 \lambda^4 λ 4 , with P s a m e = ( N ↑ 2 ) + ( N ↓ 2 ) P_{\mathrm{same}}=\binom{N_{\uparrow}}{2}+\binom{N_{\downarrow}}{2} P same = ( 2 N ↑ ) + ( 2 N ↓ ) the same-spin pair count; A A A and α \alpha α depend on the graph as well as on N N N and n n n .
Two geometrically distinct exchange channels. On the L × L L\times L L × L torus with L = 3 L=3 L = 3 or L ≥ 5 L\ge5 L ≥ 5 , ring exchange occurs through two channels of equal weight: a diagonal channel (next-nearest-neighbor pairs sharing two common neighbors, an elementary plaquette’s opposite corners) and an edge channel (nearest-neighbor pairs, via a 3-hop+1-hop traversal of an adjacent plaquette). For each eligible pair of sites, the diagonal element ( H ∞ 4 ) a a (H^4_{\infty})_{aa} ( H ∞ 4 ) aa of a configuration a a a occupying that pair is 8 λ 4 8\lambda^4 8 λ 4 lower when the two spins are equal than when they are opposite (derived in Appendix § 13 , and confirmed by exact sparse-trace evaluation at L = 5 , 6 , 7 L=5,6,7 L = 5 , 6 , 7 ); this per-pair value is a local plaquette-incidence count, not a function of the coordination number alone (see the hypercubic comparison in Appendix § 13 ). With 2 N 2N 2 N pairs in each channel, the n = 2 n=2 n = 2 coefficient follows in closed form, not fitted :
α ( N , 2 ) = − 8 ⋅ 4 N ( N 2 ) = − 64 N − 1 , (9) \alpha(N,2)=\frac{-8\cdot4N}{\binom{N}{2}}=-\frac{64}{N-1}, \tag*{(9)} α ( N , 2 ) = ( 2 N ) − 8 ⋅ 4 N = − N − 1 64 , (9)
matching every lattice of these sizes tested exactly. At L = 4 L=4 L = 4 the torus has extra 4-cycles (its axis rings): wraparound adds 16 diagonal-type pairs and raises the edge channel to − 12 -12 − 12 per pair, giving α ( 16 , 2 ) = − 32 / 5 \alpha(16,2)=-32/5 α ( 16 , 2 ) = − 32/5 , which the graph-general form below reproduces (Appendix § 13 ). The fermion-sign path-by-path derivation of both channels is given in Appendix § 13 .
Exact linearity on any graph. That μ 4 \mu_4 μ 4 is a symmetric polynomial of degree at most four in the spin labels does not by itself exclude a genuine P s a m e 2 P_{\mathrm{same}}^2 P same 2 term: that would require a four-particle cyclic permutation or a double transposition, both of which force a 1 + 1 + 1 + 1 1+1+1+1 1 + 1 + 1 + 1 hop distribution among four occupied sites. No such schedule has a legal first move, since each candidate first hop needs its target site vacated by another hop of the same schedule, a closed dependency loop with no entry point (machine-checked on any graph without self-loops, Appendix § 16 ). Four-hop closed walks therefore induce only the identity, transpositions, and 3-cycles. Now average over spin labels: the trace over a sector ( N ↑ , N ↓ ) (N_{\uparrow},N_{\downarrow}) ( N ↑ , N ↓ ) runs over site sets S S S (∣ S ∣ = n |S|=n ∣ S ∣ = n ) and, for each, over the ( n N ↑ ) \binom{n}{N_{\uparrow}} ( N ↑ n ) spin labellings. A walk inducing π \pi π contributes λ 4 sgn ( π ) \lambda^4\operatorname{sgn}(\pi) λ 4 sgn ( π ) whatever the spectators’ spins (the closed-walk amplitude lemma of Appendix § 13 ), but only in the labellings that π \pi π preserves. Two given particles share a spin in a fraction P s a m e / ( n 2 ) P_{\mathrm{same}}/\binom{n}{2} P same / ( 2 n ) of the labellings, and three given particles all share a spin in a fraction P 3 / ( n 3 ) P_3/\binom{n}{3} P 3 / ( 3 n ) , with P 3 = ( N ↑ 3 ) + ( N ↓ 3 ) P_3=\binom{N_{\uparrow}}{3}+\binom{N_{\downarrow}}{3} P 3 = ( 3 N ↑ ) + ( 3 N ↓ ) . Hence
μ 4 ( ∞ ) = I − T 2 P s a m e ( n 2 ) + T 3 P 3 ( n 3 ) , (10) \mu_4(\infty)=I-T_2\frac{P_{\mathrm{same}}}{\binom{n}{2}}+T_3\frac{P_3}{\binom{n}{3}}, \tag*{(10)} μ 4 ( ∞ ) = I − T 2 ( 2 n ) P same + T 3 ( 3 n ) P 3 , (10)
where I I I , T 2 T_2 T 2 , T 3 T_3 T 3 are the average numbers of identity, transposition and 3-cycle walks per configuration, averaged over uniformly random n n n -site sets: they depend on the graph and ( N , n ) (N,n) ( N , n ) , not on spin. (They are numbers, not to be confused with the hopping operators T 0 T_0 T 0 , T 1 T_1 T 1 of §7.) Finally P 3 P_3 P 3 is an exact affine function of P s a m e P_{\mathrm{same}} P same at fixed n n n , P 3 = ( n 3 ) − n − 2 2 ( ( n 2 ) − P s a m e ) P_3=\binom{n}{3}-\frac{n-2}{2}\left(\binom{n}{2}-P_{\mathrm{same}}\right) P 3 = ( 3 n ) − 2 n − 2 ( ( 2 n ) − P same ) for every n n n (a combinatorial identity, machine-checked in Appendix F), despite looking cubic. So μ 4 ( ∞ ) \mu_4(\infty) μ 4 ( ∞ ) is exactly affine in P s a m e P_{\mathrm{same}} P same on any finite graph, with
α ( N , n ) = − T 2 ( n 2 ) + ( n − 2 ) T 3 2 ( n 3 ) . (11) \alpha(N,n)=-\frac{T_2}{\binom{n}{2}}+\frac{(n-2)T_3}{2\binom{n}{3}}. \tag*{(11)} α ( N , n ) = − ( 2 n ) T 2 + 2 ( 3 n ) ( n − 2 ) T 3 . (11)
This is a first-principles proof of exact linearity, not an empirical observation, confirmed independently by an over-determined n = 6 n=6 n = 6 test (Table 3 : four distinct spin compositions at fixed ( L , n ) = ( 5 , 6 ) (L,n)=(5,6) ( L , n ) = ( 5 , 6 ) , over-determining a linear fit by two degrees of freedom): in exact rational arithmetic the four points lie exactly on one line, μ 4 = 259863 / 253 − ( 38 / 23 ) P s a m e \mu_4=259863/253-(38/23)P_{\mathrm{same}} μ 4 = 259863/253 − ( 38/23 ) P same .
Table 3. Over-determined n = 6 n=6 n = 6 linearity test (L = 5 L=5 L = 5 ): four distinct spin compositions, all sharing the same spin-blind μ 2 = 19 \mu_2=19 μ 2 = 19 , against two free parameters ( A , α ) (A,\alpha) ( A , α ) in μ 4 = A + α P s a m e \mu_4=A+\alpha P_{\mathrm{same}} μ 4 = A + α P same . The μ 4 \mu_4 μ 4 are exact rationals ( 253593 / 253 , 255683 / 253 , 256937 / 253 , 257355 / 253 ) (253593/253, 255683/253, 256937/253, 257355/253) ( 253593/253 , 255683/253 , 256937/253 , 257355/253 ) , and the four points lie exactly on the line μ 4 = 259863 / 253 − ( 38 / 23 ) P s a m e \mu_4=259863/253-(38/23)P_{\mathrm{same}} μ 4 = 259863/253 − ( 38/23 ) P same .
The exchange coefficient at every n n n . On any finite simple graph, T 2 T_2 T 2 and T 3 T_3 T 3 can be counted exactly, because every spin-sensitive four-hop walk lives on a 4-cycle of the graph. A transposition walk exchanges two particles on a 4-cycle (opposite corners, through the other two; or adjacent corners, three hops around the cycle and one back) and needs exactly its pair occupied and the other two corners empty; there are 32 C 4 32C_4 32 C 4 such walks in all at n = 2 n=2 n = 2 (per 4-cycle, eight for each of its two diagonal pairs and four for each of its four edges; (S5, Lemma 3)), C 4 C_4 C 4 being the number of 4-cycles of the graph (cycles of length four counted as subgraphs, chords allowed). A 3-cycle walk is the empty corner of a 4-cycle going once around it, in either direction, with the other three corners occupied: 8 C 4 8C_4 8 C 4 walks in all (four empty corners, two directions). Averaging the occupation pattern over uniformly random n n n -site sets (hypergeometric probabilities, exact) gives
T 2 ( n 2 ) = 64 C 4 ( N − n ) ( N − n − 1 ) N ( N − 1 ) ( N − 2 ) ( N − 3 ) , ( n − 2 ) T 3 2 ( n 3 ) = 24 C 4 ( n − 2 ) ( N − n ) N ( N − 1 ) ( N − 2 ) ( N − 3 ) , (12) \frac{T_2}{\binom{n}{2}}=\frac{64C_4(N-n)(N-n-1)}{N(N-1)(N-2)(N-3)},\qquad
\frac{(n-2)T_3}{2\binom{n}{3}}=\frac{24C_4(n-2)(N-n)}{N(N-1)(N-2)(N-3)}, \tag*{(12)} ( 2 n ) T 2 = N ( N − 1 ) ( N − 2 ) ( N − 3 ) 64 C 4 ( N − n ) ( N − n − 1 ) , 2 ( 3 n ) ( n − 2 ) T 3 = N ( N − 1 ) ( N − 2 ) ( N − 3 ) 24 C 4 ( n − 2 ) ( N − n ) , (12)
so that, with Eq. (11) ,
α ( N , n ) = − 8 C 4 ( N − n ) ( 8 N − 11 n − 2 ) N ( N − 1 ) ( N − 2 ) ( N − 3 ) ( N ≥ 4 ) , (13) \alpha(N,n)=-\frac{8C_4(N-n)(8N-11n-2)}{N(N-1)(N-2)(N-3)}\qquad(N\ge4), \tag*{(13)} α ( N , n ) = − N ( N − 1 ) ( N − 2 ) ( N − 3 ) 8 C 4 ( N − n ) ( 8 N − 11 n − 2 ) ( N ≥ 4 ) , (13)
which reduces to − 64 C 4 / [ N ( N − 1 ) ] -64C_4/[N(N-1)] − 64 C 4 / [ N ( N − 1 )] at n = 2 n=2 n = 2 . On the L × L L \times L L × L torus C 4 = N C_4=N C 4 = N for L = 3 L=3 L = 3 and L ≥ 5 L\geq5 L ≥ 5 (the plaquettes; at L = 3 L=3 L = 3 the axis rings are triangles), giving α ( N , 2 ) = − 64 / ( N − 1 ) \alpha(N,2)=-64/(N-1) α ( N , 2 ) = − 64/ ( N − 1 ) , while at L = 4 L=4 L = 4 the eight axis rings of length four make C 4 = 24 C_4=24 C 4 = 24 . In exact rational arithmetic Eq. (13) reproduces every slope we computed: on the square torus at L = 3 L=3 L = 3 (n = 3 , 4 , 5 n=3,4,5 n = 3 , 4 , 5 ), L = 4 L=4 L = 4 (n = 2 , 3 , 4 n=2,3,4 n = 2 , 3 , 4 ), L = 5 L=5 L = 5 (n = 3 , 4 , 6 n=3,4,6 n = 3 , 4 , 6 , including the − 38 / 23 -38/23 − 38/23 of Table 3 ), L = 6 L=6 L = 6 (n = 3 , 4 , 5 n=3,4,5 n = 3 , 4 , 5 ) and L = 7 L=7 L = 7 (n = 3 n=3 n = 3 ), and on graphs unlike the square lattice: K 6 K_6 K 6 , the Petersen graph (C 4 = 0 C_4=0 C 4 = 0 , α = 0 \alpha=0 α = 0 ), a triangular torus, a random graph and the kagome lattice (C 4 = 0 C_4=0 C 4 = 0 ). The transposition term alone, the independent blocking of a ring exchange by the n − 2 n-2 n − 2 other particles, is therefore not an approximation of α \alpha α but its exact transposition part; the 3-cycle term is exactly the remainder that the hop-class analysis of Appendix C finds in the 2+1+1 class. The walk counts 32 C 4 32C_4 32 C 4 and 8 C 4 8C_4 8 C 4 are derived in S5; the n = 2 n=2 n = 2 channels are worked out in Appendix C.
The n = 4 n=4 n = 4 data of Table 4 give the same conclusion at L = 4 L=4 L = 4 : the two independently-determined slopes ( 4 , 0 ) → ( 3 , 1 ) (4,0)\to(3,1) ( 4 , 0 ) → ( 3 , 1 ) and ( 3 , 1 ) → ( 2 , 2 ) (3,1)\to(2,2) ( 3 , 1 ) → ( 2 , 2 ) agree exactly ( − 1968 / 455 ) (-1968/455) ( − 1968/455 ) .
Table 4. Exact sparse-trace μ 2 , μ 4 \mu_2,\mu_4 μ 2 , μ 4 at U = ∞ U=\infty U = ∞ (in units of λ 2 , λ 4 \lambda^2,\lambda^4 λ 2 , λ 4 ), given as exact rationals (μ 4 = Tr ( H ∞ 4 ) / D 0 \mu_4=\operatorname{Tr}(H_\infty^4)/D_0 μ 4 = Tr ( H ∞ 4 ) / D 0 with D 0 = ( 16 − N ↑ N ↓ ) D_0=\binom{16-N_{\uparrow}}{N_{\downarrow}} D 0 = ( N ↓ 16 − N ↑ ) an integer, so μ 4 \mu_4 μ 4 is exactly rational, no rounding): μ 2 \mu_2 μ 2 is spin-blind (identical within each n n n -group, separated by horizontal rules), while μ 4 \mu_4 μ 4 depends on P s a m e P_{\mathrm{same}} P same .
Proposition (exact fourth moment). On any finite simple graph with N ≥ 4 N\geq4 N ≥ 4 sites, C 4 C_4 C 4 four-cycles and uniform hopping λ \lambda λ , for n ≥ 2 n\geq2 n ≥ 2 projected fermions in units of λ 4 \lambda^4 λ 4 ,
μ 4 ( ∞ ) = A ( N , n ) + α ( N , n ) P s a m e , α ( N , n ) = − 8 C 4 ( N − n ) ( 8 N − 11 n − 2 ) N ( N − 1 ) ( N − 2 ) ( N − 3 ) (14) \mu_4(\infty)=A(N,n)+\alpha(N,n)P_{\mathrm{same}},\qquad\alpha(N,n)=-\frac{8C_4(N-n)(8N-11n-2)}{N(N-1)(N-2)(N-3)} \tag*{(14)} μ 4 ( ∞ ) = A ( N , n ) + α ( N , n ) P same , α ( N , n ) = − N ( N − 1 ) ( N − 2 ) ( N − 3 ) 8 C 4 ( N − n ) ( 8 N − 11 n − 2 ) (14) with A ( N , n ) A(N,n) A ( N , n ) independent of spin composition. On the square-lattice torus with L = 3 L=3 L = 3 or L ≥ 5 L\geq5 L ≥ 5 , C 4 = N C_4=N C 4 = N and α ( N , 2 ) = − 64 / ( N − 1 ) \alpha(N,2)=-64/(N-1) α ( N , 2 ) = − 64/ ( N − 1 ) .
The resulting picture: μ 2 \mu_2 μ 2 resolves the purely local connectivity of the configuration graph, while on triangle-free graphs, where no closed three-hop walk exists and so μ 3 = 0 \mu_3=0 μ 3 = 0 , μ 4 \mu_4 μ 4 is the first moment to resolve closed-loop geometry : every area-enclosing four-hop walk, of which fermionic exchange is one instance and ordinary single-particle orbital motion (§ 8 ) is another. Under a magnetic flux the two are distinguished only by the sign with which their loops interfere (their interference parity), not by any difference in the underlying harmonic (§ 8 ). On graphs with triangles (the L = 3 L=3 L = 3 torus, K 6 K_6 K 6 , the triangular torus), μ 3 \mu_3 μ 3 is in general nonzero and already contains three-hop loops, including a three-hop exchange of two particles around a triangle.
Finite U U U : virtual doublons
This section derives how μ 2 \mu_2 μ 2 of the N d = 0 N_d=0 N d = 0 band deforms as U U U is lowered from infinity, through order U − 2 U^{-2} U − 2 , and checks the result against full exact diagonalization of the Hubbard model.
Effective Hamiltonian through third order
Write H = H t + U N ^ d H = H_t + U\hat{N}_d H = H t + U N ^ d (N ^ d \hat{N}_d N ^ d the number of doubly occupied sites). From the N d = 0 N_d=0 N d = 0 sector P 0 P_0 P 0 , a single hop reaches only the N d = 1 N_d=1 N d = 1 sector Q 1 Q_1 Q 1 , and a path P 0 → Q 1 → Q 1 → P 0 P_0 \to Q_1 \to Q_1 \to P_0 P 0 → Q 1 → Q 1 → P 0 cannot pass through N d ≥ 2 N_d \ge2 N d ≥ 2 , so through third order Q 1 Q_1 Q 1 is the only virtual sector. The Hermitian Schrieffer–Wolff expansion [24 , 25 ] then gives
H e f f = T 0 − 1 U V + 1 U 2 H ( 3 ) + O ( 1 / U 3 ) , H ( 3 ) = T 1 † T 0 Q T 1 − 1 2 { V , T 0 } , (15) H_{\mathrm{eff}} = T_0 - \frac{1}{U}V + \frac{1}{U^2}H^{(3)} + O(1/U^3), \qquad H^{(3)} = T_1^\dagger T_0^Q T_1 - \frac{1}{2}\{V,T_0\}, \tag*{(15)} H eff = T 0 − U 1 V + U 2 1 H ( 3 ) + O ( 1/ U 3 ) , H ( 3 ) = T 1 † T 0 Q T 1 − 2 1 { V , T 0 } , (15)
where T 0 = P 0 H t P 0 T_0 = P_0H_tP_0 T 0 = P 0 H t P 0 is the operator H ∞ H_\infty H ∞ of § 5 (a hop into an empty site; T 0 T_0 T 0 is the usual name in the strong-coupling expansion), T 1 = Q 1 H t P 0 T_1 = Q_1H_tP_0 T 1 = Q 1 H t P 0 (hop onto an opposite-spin occupied site, creating one doublon), V ≡ T 1 † T 1 V \equiv T_1^\dagger T_1 V ≡ T 1 † T 1 , and T 0 Q = Q 1 H t Q 1 T_0^Q = Q_1H_tQ_1 T 0 Q = Q 1 H t Q 1 (a hop that keeps exactly one doublon: a single electron moving into an empty site, or the doublon moving by an exchange hop); the construction is given in Appendix § 14 . Writing ⟨ X ⟩ ≡ Tr ( X ) / D 0 \langle X\rangle\equiv\operatorname{Tr}(X)/D_0 ⟨ X ⟩ ≡ Tr ( X ) / D 0 and ⟨ X , Y ⟩ H S ≡ Tr ( X † Y ) / D 0 \langle X,Y\rangle_{\mathrm{HS}} \equiv\operatorname{Tr}(X^\dagger Y)/D_0 ⟨ X , Y ⟩ HS ≡ Tr ( X † Y ) / D 0 for the Hilbert–Schmidt inner product, with ∥ X ∥ H S 2 = ⟨ X , X ⟩ H S \|X\|_{\mathrm{HS}}^2 = \langle X,X\rangle_{\mathrm{HS}} ∥ X ∥ HS 2 = ⟨ X , X ⟩ HS , we must be explicit about raw vs. central moments here, unlike at U = ∞ U=\infty U = ∞ : V V V has a nonzero diagonal, so the band’s mean energy is E ‾ ( U ) = − ⟨ V ⟩ / U + O ( 1 / U 2 ) \overline{E}(U) = -\langle V\rangle/U + O(1/U^2) E ( U ) = − ⟨ V ⟩ / U + O ( 1/ U 2 ) . The physically meaningful quantity is the central second moment μ 2 , c ( U ) ≡ ⟨ ( H e f f − E ‾ ( U ) ) 2 ⟩ \mu_{2,c}(U) \equiv\langle(H_{\mathrm{eff}}-\overline{E}(U))^2\rangle μ 2 , c ( U ) ≡ ⟨( H eff − E ( U ) ) 2 ⟩ . Since Tr T 0 = 0 \operatorname{Tr}T_0=0 Tr T 0 = 0 ,
μ 2 , c ( U ) = μ 2 ( ∞ ) − 2 U ⟨ T 0 , V ⟩ + c 2 U 2 + O ( 1 / U 3 ) , c 2 = Var ( V ) + 2 ⟨ T 0 H ( 3 ) ⟩ , (16) \mu_{2,c}(U) = \mu_2(\infty) - \frac{2}{U}\langle T_0,V\rangle+ \frac{c_2}{U^2} + O(1/U^3), \qquad c_2 = \operatorname{Var}(V) + 2\langle T_0H^{(3)}\rangle, \tag*{(16)} μ 2 , c ( U ) = μ 2 ( ∞ ) − U 2 ⟨ T 0 , V ⟩ + U 2 c 2 + O ( 1/ U 3 ) , c 2 = Var ( V ) + 2 ⟨ T 0 H ( 3 ) ⟩ , (16)
with Var ( V ) ≡ ⟨ V 2 ⟩ − ⟨ V ⟩ 2 \operatorname{Var}(V) \equiv\langle V^2\rangle- \langle V\rangle^2 Var ( V ) ≡ ⟨ V 2 ⟩ − ⟨ V ⟩ 2 . Every term is an exact sparse trace (an exact rational at λ = 1 \lambda=1 λ = 1 ). Two operators contribute at order U − 2 U^{-2} U − 2 : the square of the second-order correction, Var ( V ) = ∥ V − ⟨ V ⟩ I ∥ H S 2 ≥ 0 \operatorname{Var}(V) = \|V-\langle V\rangle I\|_{\mathrm{HS}}^2 \ge0 Var ( V ) = ∥ V − ⟨ V ⟩ I ∥ HS 2 ≥ 0 , and the overlap of the third-order term with the bare hopping. At the low fillings of Table 5 the second is larger and of opposite sign, so c 2 < 0 c_2 < 0 c 2 < 0 : at large finite U U U the N d = 0 N_d=0 N d = 0 band is narrower than at U = ∞ U=\infty U = ∞ . This sign is a computed fact, not a theorem, and it reverses near half filling, where the bare hopping has little room to act (at half filling T 0 = 0 T_0=0 T 0 = 0 and c 2 = Var ( V ) > 0 c_2=\operatorname{Var}(V)>0 c 2 = Var ( V ) > 0 ; at L = 4 L=4 L = 4 , c 2 > 0 c_2>0 c 2 > 0 already for ( 13 , 1 ) (13,1) ( 13 , 1 ) and ( 12 , 2 ) (12,2) ( 12 , 2 ) ).
Table 5. The exact U − 2 U^{-2} U − 2 coefficient c 2 c_2 c 2 of the N d = 0 N_{d=0} N d = 0 band variance at L = 4 L = 4 L = 4 (units of λ 4 \lambda^4 λ 4 ), Eq. (16) , and its second-order part Var ( V ) \operatorname{Var}(V) Var ( V ) alone (to six decimals). For ( 1 , 1 ) (1,1) ( 1 , 1 ) and ( 2 , 1 ) (2,1) ( 2 , 1 ) , c 2 c_2 c 2 agrees with full exact diagonalization to better than 10 − 3 10^{-3} 1 0 − 3 . These are exact numbers, not a closed form: β ( L , n ) \beta(L,n) β ( L , n ) is not derived (§7).
The 1 / U 1/U 1/ U term vanishes on triangle-free graphs
The 1 / U 1/U 1/ U coefficient is − 2 ⟨ T 0 , V ⟩ H S -2\langle T_0,V\rangle_{\mathrm{HS}} − 2 ⟨ T 0 , V ⟩ HS . Direct computation gives ⟨ T 0 , V ⟩ H S = 0 \langle T_0,V\rangle_{\mathrm{HS}}=0 ⟨ T 0 , V ⟩ HS = 0 exactly (an integer trace) for L = 4 , 5 , 6 L=4,5,6 L = 4 , 5 , 6 , every composition tested. The reason is graph locality: T 0 T_0 T 0 ’s edges are single nearest-neighbor hops, while V V V ’s off-diagonal elements connect configurations related either by an exchange of the spins on a bond (the superexchange piece) or by a hop between two sites sharing a common mediating neighbor (the correlated three-site-hopping piece, with its spin-flip part), the standard decomposition of the strong-coupling expansion [24 , 25 ] (Appendix § 14 ). Neither shares an off-diagonal matrix element with the bare projected hopping, unless a mediated pair of sites is itself a bond, which requires a triangle. On any triangle-free graph Tr ( T 0 V ) = 0 \operatorname{Tr}(T_0V)=0 Tr ( T 0 V ) = 0 , and the L × L L \times L L × L torus (L ≥ 3 L\ge3 L ≥ 3 ) is triangle-free exactly when L ≠ 3 L\ne3 L = 3 ; both statements are machine-checked (Appendix § 16 ). The second-order correction operator V V V is therefore exactly Hilbert–Schmidt orthogonal to the projected kinetic operator for L ≥ 4 L\ge4 L ≥ 4 , and μ 2 , c ( U ) \mu_{2,c}(U) μ 2 , c ( U ) has no 1 / U 1/U 1/ U term:
μ 2 , c ( U ) − μ 2 ( ∞ ) = c 2 U 2 + O ( 1 / U 3 ) ( L ≥ 4 ) . (17) \mu_{2,c}(U)-\mu_2(\infty) = \frac{c_2}{U^2} + O(1/U^3) \qquad(L\ge4). \tag*{(17)} μ 2 , c ( U ) − μ 2 ( ∞ ) = U 2 c 2 + O ( 1/ U 3 ) ( L ≥ 4 ) . (17)
This is a proven cancellation, not a fit. (The third-order operator is not orthogonal to T 0 T_0 T 0 ; its overlap is what enters c 2 c_2 c 2 .) At L = 3 L = 3 L = 3 the overlap is nonzero (a small-torus wraparound effect where the mediated site coincides with a true nearest neighbor), and full exact diagonalization confirms the resulting 1 / U 1/U 1/ U term: its coefficient matches − 2 ⟨ T 0 , V ⟩ H S -2\langle T_0,V\rangle_{\mathrm{HS}} − 2 ⟨ T 0 , V ⟩ HS to better than 10 − 3 10^{-3} 1 0 − 3 (relative) for ( 1 , 1 ) (1,1) ( 1 , 1 ) and ( 2 , 1 ) (2,1) ( 2 , 1 ) .
The U − 2 U^{-2} U − 2 coefficient and its N ↑ N ↓ N_{\uparrow}N_{\downarrow} N ↑ N ↓ dependence
Exact sparse traces give c 2 c_2 c 2 for every composition (Table 5 at L = 4 L = 4 L = 4 ), and full exact diagonalization of the Hubbard model reproduces it: for ( 1 , 1 ) (1,1) ( 1 , 1 ) and ( 2 , 1 ) (2,1) ( 2 , 1 ) at L = 4 L = 4 L = 4 , the U − 2 U^{-2} U − 2 coefficient of the exact band variance agrees with c 2 c_2 c 2 to better than 10 − 3 10^{-3} 1 0 − 3 (relative). Here the band is the lowest D 0 D_0 D 0 levels of the exact spectrum, and the coefficient is the Richardson extrapolation of U 2 [ μ 2 , c ( U ) − μ 2 ( ∞ ) ] U^2[\mu_{2,c}(U)-\mu_2(\infty)] U 2 [ μ 2 , c ( U ) − μ 2 ( ∞ )] from U = 400 U = 400 U = 400 and 800 λ 800\lambda 800 λ . Var ( V ) \operatorname{Var}(V) Var ( V ) alone has the wrong sign; for ( 2 , 1 ) (2,1) ( 2 , 1 ) , Var ( V ) = 9.643175 \operatorname{Var}(V) = 9.643175 Var ( V ) = 9.643175 while 2 ⟨ T 0 H ( 3 ) ⟩ = − 816 / 35 2\langle T_0H^{(3)}\rangle= -816/35 2 ⟨ T 0 H ( 3 ) ⟩ = − 816/35 . Single-species sectors have c 2 = 0 c_2 = 0 c 2 = 0 : no doublon can form. For two particles the computed values follow a simple form, c 2 ( 1 , 1 ) = − 96 / ( N − 1 ) − 64 / ( N − 1 ) 2 c_2(1,1) = -96/(N-1)-64/(N-1)^2 c 2 ( 1 , 1 ) = − 96/ ( N − 1 ) − 64/ ( N − 1 ) 2 exactly at L = 5 L = 5 L = 5 –8 8 8 , which we have not derived.
At low filling, c 2 / ( N ↑ N ↓ ) c_2/(N_{\uparrow}N_{\downarrow}) c 2 / ( N ↑ N ↓ ) is close to composition-independent at fixed ( L , n ) (L,n) ( L , n ) , and its relative spread between two compositions (∣ difference ∣ |\text{difference}| ∣ difference ∣ divided by the mean) shrinks as L L L grows: 0.14 % 0.14\% 0.14% , 0.06 % 0.06\% 0.06% , 0.03 % 0.03\% 0.03% between ( 3 , 1 ) (3,1) ( 3 , 1 ) and ( 2 , 2 ) (2,2) ( 2 , 2 ) at L = 4 , 5 , 6 L = 4,5,6 L = 4 , 5 , 6 , and 0.34 % 0.34\% 0.34% , 0.13 % 0.13\% 0.13% between ( 4 , 1 ) (4,1) ( 4 , 1 ) and ( 3 , 2 ) (3,2) ( 3 , 2 ) at L = 4 , 5 L = 4,5 L = 4 , 5 . The near-independence degrades as the filling grows (at L = 4 L = 4 L = 4 , n = 6 n = 6 n = 6 , the spread is 0.79 % 0.79\% 0.79% between ( 5 , 1 ) (5,1) ( 5 , 1 ) and ( 3 , 3 ) (3,3) ( 3 , 3 ) ), and near half filling c 2 c_2 c 2 changes sign. At low filling this is numerical evidence that the leading finite-U U U correction is governed by the opposite-spin pair count, c 2 ≈ β ( L , n ) N ↑ N ↓ c_2 \approx\beta(L,n)N_{\uparrow}N_{\downarrow} c 2 ≈ β ( L , n ) N ↑ N ↓ with a composition-independent β ( L , n ) < 0 \beta(L,n) < 0 β ( L , n ) < 0 . At fixed n n n this is the same statement as c 2 c_2 c 2 being close to affine in P s a m e P_{\mathrm{same}} P same , since N ↑ N ↓ = ( n 2 ) − P s a m e N_{\uparrow}N_{\downarrow} = \binom{n}{2} - P_{\mathrm{same}} N ↑ N ↓ = ( 2 n ) − P same , with the exact zero c 2 = 0 c_2 = 0 c 2 = 0 of a single species fixing the intercept; with two or three compositions per ( L , n ) (L,n) ( L , n ) , each test is one ratio. We do not derive β ( L , n ) \beta(L,n) β ( L , n ) ; at fixed n n n , ∣ β ∣ |\beta| ∣ β ∣ is at most of order 1 / N 1/N 1/ N , by a bound ∣ c 2 ∣ ≤ C ( n , z ) λ 4 N ↑ N ↓ / ( N − 1 ) |c_2| \le C(n,z)\lambda^4N_{\uparrow}N_{\downarrow}/(N-1) ∣ c 2 ∣ ≤ C ( n , z ) λ 4 N ↑ N ↓ / ( N − 1 ) , with C ( n , z ) C(n,z) C ( n , z ) depending only on n n n and the coordination z z z , proved in S7. The N ↑ N ↓ N_{\uparrow}N_{\downarrow} N ↑ N ↓ law is a numerical trend, epistemically distinct from the exact μ 2 ( ∞ ) \mu_2(\infty) μ 2 ( ∞ ) theorem and from the exact vanishing of the 1 / U 1/U 1/ U term.
Both spin-resolved results are therefore statements about the dependence on P s a m e P_{\mathrm{same}} P same at fixed n n n : μ 4 ( ∞ ) \mu_4(\infty) μ 4 ( ∞ ) is exactly affine in it, with the exact coefficient of §6; c 2 c_2 c 2 , the leading finite-U U U correction to μ 2 \mu_2 μ 2 , is close to affine in it, with its zero at a single species, as a numerical trend. We keep this asymmetry in confidence explicit rather than presenting both as theorems.
Gauge response: the flux-resolved fourth moment and the doublon
Promoting the hopping amplitudes to gauge-dependent variables (S9), t i j → t i j e i A i j t_{ij} \to t_{ij}e^{iA_{ij}} t ij → t ij e i A ij (A j i = − A i j A_{ji}=-A_{ij} A j i = − A ij ), with Φ □ = ∑ □ A i j \Phi_{\square}=\sum_{\square}A_{ij} Φ □ = ∑ □ A ij the flux through an elementary plaquette (for a uniform flux on the L × L L\times L L × L torus, Φ □ = 2 π p / L 2 \Phi_{\square}=2\pi p/L^2 Φ □ = 2 π p / L 2 with integer p p p ), is especially natural in the present formulation because spectral moments are sums over closed configuration-space paths: at second order the contributing paths backtrack (occupied site to empty neighbor and back), and are therefore gauge-blind, μ 2 ( Φ ) = μ 2 ( 0 ) \mu_2(\Phi)=\mu_2(0) μ 2 ( Φ ) = μ 2 ( 0 ) identically, since a backtrack’s phase is e i A i j e − i A i j = 1 e^{iA_{ij}}e^{-iA_{ij}}=1 e i A ij e − i A ij = 1 for any A i j A_{ij} A ij ; at fourth order and higher, closed walks enclosing nonzero area can acquire nontrivial gauge phases. We report the n = 2 n=2 n = 2 calculation carried through exactly, together with a machine-precision numerical cross-check against an explicit Hofstadter-type Peierls Hamiltonian (Figure 10 , Appendix § 15 ). Full derivations are in S10 and S11; the one item not fully closed, the flux-independent baseline, is flagged below.
Figure 10. Exchange contribution to the fourth moment: the same-spin-minus-opposite-spin difference μ 4 ( 2 , 0 ) ( Φ ) − μ 4 ( 1 , 1 ) ( Φ ) \mu_4^{(2,0)}(\Phi)-\mu_4^{(1,1)}(\Phi) μ 4 ( 2 , 0 ) ( Φ ) − μ 4 ( 1 , 1 ) ( Φ ) against the closed-form prediction − 64 / ( N − 1 ) cos Φ □ -64/(N-1)\cos\Phi_{\square} − 64/ ( N − 1 ) cos Φ □ (88), n = 2 n=2 n = 2 , N = 36 N=36 N = 36 (L = 6 L=6 L = 6 ; the horizontal axis Φ \Phi Φ is the plaquette flux Φ □ \Phi_{\square} Φ □ ), exact sparse-trace evaluation with a genuine Peierls Hamiltonian, no fitted parameters. Taking this spin-sector difference, rather than the raw fourth moment, cancels the spin-blind orbital plaquette-loop contribution exactly, isolating the exchange mechanism alone.
The organizing principle. Since H ∞ ( Φ ) H_{\infty}(\Phi) H ∞ ( Φ ) is Hermitian for any Φ \Phi Φ , μ 4 ( Φ ) = ∑ a , b ∣ ( H ∞ 2 ( Φ ) ) a b ∣ 2 / D 0 \mu_4(\Phi)=\sum_{a,b}\lvert(H_{\infty}^2(\Phi))_{ab}\rvert^2/D_0 μ 4 ( Φ ) = ∑ a , b ∣( H ∞ 2 ( Φ ) ) ab ∣ 2 / D 0 term by term. A closed-walk contribution reached by a single path is therefore exactly flux-independent (∣ λ 2 s e i θ ∣ 2 = λ 4 \lvert\lambda^2se^{i\theta}\rvert^2=\lambda^4 ∣ λ 2 s e i θ ∣ 2 = λ 4 , with s = ± 1 s=\pm1 s = ± 1 a fermion sign, regardless of θ \theta θ ); flux dependence requires interference between at least two distinct paths reaching the same intermediate configuration, with their relative phase equal to the gauge-invariant Wilson-loop phase Θ C = ∑ ( i j ) ∈ C A i j \Theta_C=\sum_{(ij)\in C}A_{ij} Θ C = ∑ ( ij ) ∈ C A ij of the closed loop they form together (an integer multiple of Φ □ \Phi_{\square} Φ □ for the contractible loops of the square lattice; on the 4 × 4 4\times4 4 × 4 torus some four-hop loops wind around it and pick up the boundary holonomies instead). Multiple interfering paths are necessary but not sufficient for flux sensitivity, since two paths could still carry a vanishing relative phase, but this organizes every closed-walk type classified above into “single-path, hence flux-blind” or “multi-path, hence potentially flux-sensitive.”
Flux-sensitive four-hop mechanisms. In the classification of Appendix § 13 , three closed-walk types are multi-path: the single-particle plaquette loop (Type I, spin-blind), the two ring-exchange channels (Type II(b), diagonal and edge), and the 3-cycle (Type III, needed only for n ≥ 3 n\ge3 n ≥ 3 ). All of them give the identical functional dependence on Φ □ \Phi_{\square} Φ □ , differing only in interference parity : the plaquette loop and the 3-cycle are constructive at zero flux (their two contributing routes share a fermion sign, exhaustively verified for the 3-cycle in Appendix § 13 ), giving ∝ ( 1 + cos Φ □ ) \propto(1+\cos\Phi_{\square}) ∝ ( 1 + cos Φ □ ) ; both ring-exchange channels are destructive at zero flux (their two contributing assignments carry opposite signs, the exact cancellation derived in Appendix § 13 ), giving ∝ ( 1 − cos Φ □ ) \propto(1-\cos\Phi_{\square}) ∝ ( 1 − cos Φ □ ) for the same-spin case that survives, with the opposite-spin case exactly flux-blind. For n = 2 n=2 n = 2 on the L × L L\times L L × L torus with L = 3 L=3 L = 3 or L ≥ 5 L\ge5 L ≥ 5 this assembles into an exact closed form for the exchange coefficient,
α ( N , 2 , Φ ) = − 64 N − 1 cos Φ □ , (18) \alpha(N,2,\Phi)=-\frac{64}{N-1}\cos\Phi_{\square}, \tag*{(18)} α ( N , 2 , Φ ) = − N − 1 64 cos Φ □ , (18)
confirmed to machine precision (deviation below 10 − 12 10^{-12} 1 0 − 12 ) against exact sparse-trace evaluation with Peierls phases at every quantized flux for L = 3 , 5 , 6 L=3,5,6 L = 3 , 5 , 6 and at every third one for L = 7 L=7 L = 7 . At L = 4 L=4 L = 4 it fails already at zero flux (α ( 16 , 2 ) = − 32 / 5 \alpha(16,2)=-32/5 α ( 16 , 2 ) = − 32/5 , §6), and under flux the L = 4 L=4 L = 4 value also depends on the boundary twists. A companion check at L = 4 , 5 L=4,5 L = 4 , 5 finds no cos 2 Φ □ \cos2\Phi_{\square} cos 2 Φ □ component in the 3-cycle’s contribution to the precision of the same sparse-trace method, as the constructive-interference derivation (S10, Sec. 8) requires; at L = 4 L=4 L = 4 , loops winding around the torus add further harmonics (cos 4 Φ □ \cos4\Phi_{\square} cos 4 Φ □ and cos 8 Φ □ \cos8\Phi_{\square} cos 8 Φ □ , observed numerically), in line with the L = 4 L=4 L = 4 caveat. The flux-independent “retrace” baseline (Type I’s other sub-types together with Type II(a)) is only partially assembled into closed form for general N N N : the part in which both two-hop paths return through the reference configuration is in closed form, the rest is pinned down numerically (S10). What is resolved exactly, on L × L L \times L L × L tori with L = 3 L = 3 L = 3 or L ≥ 5 L \ge5 L ≥ 5 , is the flux factor of each mechanism, ( 1 + cos Φ □ ) (1+\cos\Phi_{\square}) ( 1 + cos Φ □ ) for the plaquette loop and the 3-cycle and ( 1 − cos Φ □ ) (1-\cos\Phi_{\square}) ( 1 − cos Φ □ ) for same-spin ring exchange; the closed form of the exchange coefficient under flux is given here for n = 2 n = 2 n = 2 only.
What this does, and does not, establish. On L × L L \times L L × L tori with L = 3 L = 3 L = 3 or L ≥ 5 L \ge5 L ≥ 5 , every flux-sensitive mechanism in μ 4 ( Φ ) \mu_4(\Phi) μ 4 ( Φ ) carries the same harmonic, cos Φ □ \cos\Phi_{\square} cos Φ □ , the ordinary single-particle Aharonov–Bohm periodicity. At fourth order there is no cos 2 Φ □ \cos2\Phi_{\square} cos 2 Φ □ term (derived in S10, and confirmed numerically for the 3-cycle at L = 4 , 5 L = 4,5 L = 4 , 5 ): the bare, projected fourth moment shows no charge-2 e 2e 2 e signature. This answers the question that motivates the section, whether the bare moments carry a charge-2 e 2e 2 e signature: fermionic exchange changes the interference parity of a closed walk (constructive vs. destructive), not the effective electromagnetic charge of the object completing it. A clean consequence is that the two mechanisms trade off maximally at Φ □ = π \Phi_{\square} = \pi Φ □ = π : the orbital and 3-cycle contributions vanish there ( 1 + cos π = 0 ) (1+\cos\pi= 0) ( 1 + cos π = 0 ) while the same-spin exchange contribution is maximal ( 1 − cos π = 2 ) (1-\cos\pi= 2) ( 1 − cos π = 2 ) . For n = 2 n = 2 n = 2 the difference μ 4 ( 2 , 0 ) ( Φ ) − μ 4 ( 1 , 1 ) ( Φ ) \mu_4^{(2,0)}(\Phi)-\mu_4^{(1,1)}(\Phi) μ 4 ( 2 , 0 ) ( Φ ) − μ 4 ( 1 , 1 ) ( Φ ) cancels the spin-blind orbital background exactly , isolating pure exchange topology as a function of flux: a concrete diagnostic of exchange under flux, not merely a demonstration that μ 4 ( Φ ) ≠ μ 4 ( 0 ) \mu_4(\Phi) \ne\mu_4(0) μ 4 ( Φ ) = μ 4 ( 0 ) .
The doublon operator d i = c i ↓ c i ↑ d_i = c_{i\downarrow}c_{i\uparrow} d i = c i ↓ c i ↑ transforms under a local gauge transformation as d i → e 2 i θ i d i d_i \to e^{2i\theta_i}d_i d i → e 2 i θ i d i , twice the microscopic phase of a single fermion, so a genuine charge-2 e 2e 2 e periodicity, if present at all, must appear in a projected , doublon-sector observable rather than in the bare N d = 0 N_d = 0 N d = 0 fourth moment computed above. This question has an exact answer. Applying the same Peierls substitution to the microscopic H t H_t H t and carrying out the second-order Schrieffer–Wolff projection onto the doublon (one doubly occupied site) sector, the counterpart for N d = 1 N_d = 1 N d = 1 of the projection onto N d = 0 N_d = 0 N d = 0 in §7, every virtual path contributing to the effective doublon hopping (the two hop orderings of §2) traverses the same bond twice, once per spin species, so, unlike the ring-exchange channels of §6, no path goes around a loop and there is no second route with a different phase to interfere against and no possibility of cancellation: the two-electron process simply accumulates the bond phase twice. In the two-electron (single-doublon) sector, with energies measured from U U U , this gives
H d ( 2 ) ( A ) = ε d ( 2 ) ∑ i n i d + t d ∑ ⟨ i j ⟩ ( e 2 i A i j d j † d i + h . c . ) , t d = 2 λ 2 U , (19) H_d^{(2)}(A)=\varepsilon_d^{(2)}\sum_i n_i^d+t_d\sum_{\langle ij\rangle}\left(e^{2iA_{ij}}d_j^\dagger d_i+\mathrm{h.c.}\right),\qquad t_d=\frac{2\lambda^2}{U}, \tag*{(19)} H d ( 2 ) ( A ) = ε d ( 2 ) i ∑ n i d + t d ⟨ ij ⟩ ∑ ( e 2 i A ij d j † d i + h.c. ) , t d = U 2 λ 2 , (19)
with A i j = − A j i A_{ij}=-A_{ji} A ij = − A j i and the on-site energy shift ε d ( 2 ) = 8 λ 2 / U \varepsilon_d^{(2)}=8\lambda^2/U ε d ( 2 ) = 8 λ 2 / U (so that ε d = U + ε d ( 2 ) \varepsilon_d=U+\varepsilon_d^{(2)} ε d = U + ε d ( 2 ) ) exactly flux-blind at this order (a backtracking argument, identical to μ 2 ( Φ ) = μ 2 ( 0 ) \mu_2(\Phi)=\mu_2(0) μ 2 ( Φ ) = μ 2 ( 0 ) ), while the inter-site hopping inherits twice the bare link phase, as local gauge covariance requires: d j † d i → e 2 i ( θ i − θ j ) d j † d i d_j^\dagger d_i\to e^{2i(\theta_i-\theta_j)}d_j^\dagger d_i d j † d i → e 2 i ( θ i − θ j ) d j † d i under a local gauge transformation must be matched by the transformation of the coefficient, which depends only on its own bond’s phase. The doubled amplitude t d e 2 i A i j t_d e^{2iA_{ij}} t d e 2 i A ij , the absence of longer-range doublon hopping at this order, and the flux-blind on-site term ( 2 z λ 2 / U (2z\lambda^2/U ( 2 z λ 2 / U on a site of coordination z ) z) z ) are machine-checked on an arbitrary finite graph for the explicit second-order expression of the two-electron problem; the Schrieffer–Wolff expansion that leads to it is not formalized (Appendix F). With several doublons, further second-order terms (doublon–doublon interactions and correlated hoppings) appear, which we do not derive. This gives two logically distinct, and equally exact, consequences. Around one elementary plaquette, the enclosed Wilson phase is doubled,
∏ □ e 2 i A i j = e 2 i Φ □ , (20) \prod_{\square} e^{2iA_{ij}}=e^{2i\Phi_{\square}}, \tag*{(20)} □ ∏ e 2 i A ij = e 2 i Φ □ , (20)
a local statement, the direct analog at doubled phase of the orbital plaquette-loop mechanism above. Independently, threading a global boundary twist θ \theta θ along x x x , distributed as a phase θ / L \theta/L θ / L on every x x x -bond (gauge-equivalent to θ \theta θ on one boundary bond; a bare electron then sees cos ( k x + θ / L ) \cos(k_x+\theta/L) cos ( k x + θ / L ) ), shifts the doublon’s own dispersion as E d ( K x , K y ; θ ) = ε d + 2 t d [ cos ( K x + 2 θ / L ) + cos K y ] E_d(K_x,K_y;\theta)=\varepsilon_d+2t_d[\cos(K_x+2\theta/L)+\cos K_y] E d ( K x , K y ; θ ) = ε d + 2 t d [ cos ( K x + 2 θ / L ) + cos K y ] , whose full spectrum is invariant under θ → θ + π \theta\to\theta+\pi θ → θ + π , half the period of the bare electron’s own 2 π 2\pi 2 π -periodic twisted-boundary spectrum,
Spec H d ( θ + π ) = Spec H d ( θ ) , (21) \operatorname{Spec} H_d(\theta+\pi)=\operatorname{Spec} H_d(\theta), \tag*{(21)} Spec H d ( θ + π ) = Spec H d ( θ ) , (21)
an exact, genuinely global statement of h / ( 2 e ) h/(2e) h / ( 2 e ) periodicity for the effective doublon Hamiltonian, kept explicitly distinct from the local Wilson-loop result above (derivation in S11, Sec. 5). The exact two-electron problem is not π \pi π -periodic as a whole: its N d = 0 N_{d=0} N d = 0 continuum keeps the bare 2 π 2\pi 2 π period. Its U-band follows the doublon prediction: on the 6 × 6 6\times6 6 × 6 torus it is π \pi π -periodic in θ \theta θ up to deviations of about 10 − 4 λ 10^{-4}\lambda 1 0 − 4 λ at U = 20 λ U=20\lambda U = 20 λ and 3 × 10 − 6 λ 3\times10^{-6}\lambda 3 × 1 0 − 6 λ at U = 40 λ U=40\lambda U = 40 λ , falling as ( 2 λ ) L / U L − 1 (2\lambda)^L/U^{L-1} ( 2 λ ) L / U L − 1 (breaking the π \pi π -periodicity requires the pair to dissociate virtually around the torus), and it matches the doubled-twist dispersion above to 8 × 10 − 3 λ 8\times10^{-3}\lambda 8 × 1 0 − 3 λ and 10 − 3 λ 10^{-3}\lambda 1 0 − 3 λ , i.e. to O ( λ 4 / U 3 ) O(\lambda^4/U^3) O ( λ 4 / U 3 ) . The local, plaquette-flux statement was tested against the exact flux-threaded two-electron Hubbard problem with no fitted parameters: on the 6 × 6 6\times6 6 × 6 torus, over all 36 quantized uniform fluxes Φ □ = 2 π p / 36 \Phi_{\square}=2\pi p/36 Φ □ = 2 π p /36 , the doubled-phase prediction for the U-band outperforms an undoubled control (the same model with phase A i j A_{ij} A ij instead of 2 A i j 2A_{ij} 2 A ij ) in RMS residual by a factor of 12.5 at U = 20 λ U=20\lambda U = 20 λ and 48 at U = 40 λ U=40\lambda U = 40 λ . The residual at one flux is the RMS difference between the N N N sorted U-band levels of the exact spectrum and the N N N sorted eigenvalues of the effective model; it is averaged over the fluxes. The ratio of the doubled to the undoubled residual falls steadily with U U U (0.28, 0.080, 0.021, 0.0052 at U = 10 U=10 U = 10 , 20, 40, 80 λ 80\lambda 80 λ ), as expected for a leading-order match.
This single-pair result was anticipated dynamically, if not at the Hamiltonian level, by Souza and Claro’s two-dimensional simulations [3 ] , in which the pair’s Bloch-oscillation frequency is doubled, and by the one-dimensional treatment of Claro and Robles [4 ] , which finds the same doubling semiclassically at leading order in λ / U \lambda/U λ / U , compatible with charge 2 e 2e 2 e . A doubled Peierls phase is what one expects of a charge-2 e 2e 2 e pair, and we do not claim it as new. What this paper adds is the explicit derivation: the same effective charge emerges from gauge-covariant Schrieffer–Wolff projection of the Peierls-coupled microscopic Hubbard model on any graph, with its second-order terms machine-checked for the two-electron problem, giving both the local Wilson-loop doubling and the global π \pi π -periodicity above as consequences of a single derived operator, Eq. (19) , tested without fitted parameters against the exact problem.
The resulting contrast with the flux-sensitive four-hop mechanisms established earlier in this section is worth stating as one outcome: the bare projected operator P 0 H t P 0 P_0H_tP_0 P 0 H t P 0 ’s closed loops carry ordinary charge-e e e interference at every order computed here (cos Φ □ \cos\Phi_{\square} cos Φ □ , with exchange changing interference parity rather than harmonic ), while Schrieffer–Wolff-projected doublon dynamics carries a distinct, phase-preserving charge-2 e 2e 2 e degree of freedom (e 2 i A i j e^{2iA_{ij}} e 2 i A ij , cos 2 Φ □ \cos2\Phi_{\square} cos 2 Φ □ , θ → θ + π \theta\to\theta+\pi θ → θ + π ). The two are not competing descriptions of the same physics but successive levels of one hierarchy: exchange-sensitive interference of charge-e e e microscopic paths at one order, unitary propagation of a single charge-2 e 2e 2 e composite degree of freedom at the next: not yet the collective coherence of many such composites, which is the open question of §9.
Discussion and outlook
These results form a hierarchy of configuration-space constraints, spectral signatures, and gauge response (Table 6 ). Figure 6 makes the microscopic origin of that hierarchy explicit, at the level of individual sites and hops rather than abstract equations: single hops out of one occupied site fix the spin-blind connectivity moment μ 2 \mu_2 μ 2 ; two geometrically distinct four-hop closed paths (diagonal and edge, of equal weight at n = 2 n=2 n = 2 ) first resolve same-spin exchange in μ 4 \mu_4 μ 4 ; and the virtual doubloncreating hop T 1 T_1 T 1 , together with its return T 1 † / U T_1^\dagger/U T 1 † / U (and, at third order, a hop within the one-doublon sector in between), generates the leading finite-U U U correction to μ 2 \mu_2 μ 2 .
Table 6. The spectral-moment hierarchy established in this paper: each successive row resolves a more nonlocal structure of the projected configuration graph, and (final row) a distinct gauge-coupled degree of freedom.
Figure 6. Microscopic origin of the spectral-moment hierarchy, colored by the three concepts used consistently throughout this paper’s figures: connectivity (μ 2 \mu_2 μ 2 ), exchange (μ 4 \mu_4 μ 4 ), and virtual doublon (finite U U U ). (a) μ 2 \mu_2 μ 2 counts single allowed hops out of an occupied site i i i to an empty neighbor j j j , a local, spin-blind connectivity count. (b) μ 4 \mu_4 μ 4 first becomes exchange-sensitive through two geometrically distinct four-hop closed paths that return the occupied site set { i , j } \{i,j\} { i , j } to itself: a diagonal (2+2 hop) channel through shared empty neighbors k , l k,l k , l (drawn as the closed sequence of both particles’ hops, each particle moving twice), and an edge (3+1 hop) channel through an adjacent plaquette’s other corners m , p m,p m , p . Both return to the identical occupation-number state ∣ α ⟩ |\alpha\rangle ∣ α ⟩ (the initial configuration) only when i , j i,j i , j carry the same spin, making them the exchange-sensitive closed paths of § 6 . (c) At finite U U U , the virtual hop T 1 T_1 T 1 onto an opposite-spin occupied site creates a doublon at energy cost U U U ; the return hop T 1 † / U T_1^\dagger/U T 1 † / U (light green, dashed) generates the second-order term V V V of the effective Hamiltonian, which contributes to μ 2 \mu_2 μ 2 only at order U − 2 U^{-2} U − 2 , through Var ( V ) \operatorname{Var}(V) Var ( V ) ; with the third-order term H ( 3 ) H^{(3)} H ( 3 ) (a hop within the one-doublon sector between T 1 T_1 T 1 and T 1 † T_1^\dagger T 1 † , and the anticommutator − 1 2 { V , T 0 } -\frac{1}{2}\{V,T_0\} − 2 1 { V , T 0 } ), it fixes the U − 2 U^{-2} U − 2 coefficient discussed in §7.
Generality and open problems. Several quantities are exact traces for which we have no derived closed form: β ( L , n ) \beta(L,n) β ( L , n ) , the prefactor of the opposite-spin pair count in the U − 2 U^{-2} U − 2 coefficient of μ 2 \mu_2 μ 2 (§ 7 ), whose N ↑ N ↓ N_{\uparrow}N_{\downarrow} N ↑ N ↓ law is established only numerically and only at low filling (exact traces at L = 4 , 5 , 6 L=4,5,6 L = 4 , 5 , 6 for n = 4 n=4 n = 4 , L = 4 , 5 , 6 L=4,5,6 L = 4 , 5 , 6 for n = 5 n=5 n = 5 and L = 4 L=4 L = 4 for n = 6 n=6 n = 6 ; for n = 2 n=2 n = 2 , c 2 c_2 c 2 follows a computed but underived form); the intercept A ( N , n ) A(N,n) A ( N , n ) of μ 4 \mu_4 μ 4 ; the flux-independent baseline of μ 4 ( Φ ) \mu_4(\Phi) μ 4 ( Φ ) (§ 8 ); and the many-cluster O ( λ 2 / U ) O(\lambda^2/U) O ( λ 2 / U ) shift of the sector centres (§ 4 ). The exchange coefficient α ( N , n ) \alpha(N,n) α ( N , n ) for uniform hopping, by contrast, is exact at every n n n on any finite simple graph, through its number of 4-cycles (Eq. (13) ). Further directions include higher moments (μ 6 \mu_6 μ 6 and beyond) of the same projected operators, and extending the cluster-coordinate rule to non-doublon clusters (e.g. a bound tetramer, C = 1 C=1 C = 1 , d c l = 2 d_{\mathrm{cl}}=2 d cl = 2 ), where the two-doublon binding problem mentioned in § 1 reconnects with this framework.
This program should be pursued in a specific order, because each step establishes a qualitatively stronger claim than the last, and only the last begins to justify superconductivity language. (i) Flux sensitivity of exchange loops : established exactly in § 8 , at h / e h/e h / e periodicity. This demonstrates orbital/gauge sensitivity of the exchange topology already identified in this paper, and nothing more: a completely normal, non-superconducting system can have flux-sensitive closed paths. (ii) A charge-2e signature at the single-pair level : established in § 8 for the spectrum (a static property; transport is not computed here), Eq. (19) ; the projected pair sector does carry the doubled Peierls phase, confirmed against the exact microscopic problem with no fitted parameters. (iii) Collective stiffness : whether any such doubling survives , or grows , with particle number and system size , rather than remaining the dynamics of one isolated doublon, e.g. via the finite-system charge stiffness D s ∼ ∂ 2 E 0 ( θ ) / ∂ θ 2 ∣ θ = 0 D_s \sim\partial^2 E_0(\theta)/\partial\theta^2|_{\theta=0} D s ∼ ∂ 2 E 0 ( θ ) / ∂ θ 2 ∣ θ = 0 compared across sectors with unpaired fermions, one doublon, and multiple pairs, and whether any collective response is genuinely cooperative rather than merely additive across independently-mobile doublons. Only this last step addresses coherence rather than local bound-pair motion, and remains entirely future work, together with the doublon–doublon interactions and correlated hoppings that a multi-doublon Schrieffer–Wolff projection would need to derive first, so that a stiffness calculation tests the repulsive Hubbard model itself rather than an assumed hard-core-boson model.
The present paper’s two spin-resolved results suggest a specific question to ask of this program. The fourth moment found here is controlled by the same-spin pair count P s a m e P_{\mathrm{same}} P same (§ 6 ), while the lead-
ing finite-U U U correction to μ 2 \mu_2 μ 2 is, numerically, close to proportional to the opposite-spin pair count N ↑ N ↓ N_{\uparrow}N_{\downarrow} N ↑ N ↓ § 7 . At fixed n n n the two pair counts are affinely related, so the distinction is not between two variables but between two mechanisms: same-spin exchange, exact at U = ∞ U = \infty U = ∞ , and opposite-spin virtual doublon formation, a numerical trend at finite U U U . Under flux, only the first has been examined so far: it appears as an interference parity of the ordinary h / e h/e h / e harmonic Table 6 . We phrase the resulting question deliberately short of superconductivity: can virtual opposite-spin pairing and same-spin exchange topology combine to produce anomalously collective (non-additive) charge-2 e 2e 2 e transport in the repulsive Hubbard model? This paper establishes an explicit ladder toward that question without yet answering it: repulsive binding into a mobile composite § 2 , [3 , 4 ] , a microscopic charge-2 e 2e 2 e gauge coupling and single-pair h / ( 2 e ) h/(2e) h / ( 2 e ) flux periodicity at leading order in 1 / U 1/U 1/ U § 8 , but not yet many-pair phase stiffness, off-diagonal long-range order, or condensation. Only these collective properties, step (iii) above and beyond, would justify superconductivity language, and they are left, deliberately, entirely to follow-on work.
Numerical statements. The following results are computed, not derived, and are presented as such: the two-electron benchmark numbers, the U U U -band collapse and the odd-L L L bandwidth convergence § 2 ; the variance and kurtosis comparison with the free lattice, the agreement of the sector centres with N d ε d N d\varepsilon_d N d ε d , the many-cluster constants, and the n = 3 n = 3 n = 3 resonance inside the doublon–electron continuum § 4 ; the sign of c 2 c_2 c 2 , its closeness to β ( L , n ) N ↑ N ↓ \beta(L,n)N_{\uparrow}N_{\downarrow} β ( L , n ) N ↑ N ↓ at low filling, and the form of c 2 ( 1 , 1 ) c_2(1,1) c 2 ( 1 , 1 ) at L = 5 L = 5 L = 5 –8 8 8 § 7 ; and, in § 8 , the twist dependence and higher harmonics at L = 4 L = 4 L = 4 under flux, the π \pi π -periodicity of the exact U U U -band up to corrections of order ( 2 λ ) L / U L − 1 (2\lambda)^L/U^{L-1} ( 2 λ ) L / U L − 1 , its agreement with the doubled-twist doublon dispersion to O ( λ 4 / U 3 ) O(\lambda^4/U^3) O ( λ 4 / U 3 ) , and the RMS agreement of the effective doublon model with the exact flux-threaded U U U -band. Every other result is derived, and every exact closed form is also checked against exact evaluation; the local values quoted in Appendix C as cross-checks are exact enumerations at the sizes stated.
What is not claimed. We collect here the limits stated throughout. (i) β ( L , n ) \beta(L,n) β ( L , n ) is not derived: the sign of the U − 2 U^{-2} U − 2 coefficient and its opposite-spin pair-count law at low filling are numerical findings § 7 ; nor are A ( N , n ) A(N,n) A ( N , n ) , the flux-independent baseline of μ 4 ( Φ ) \mu_4(\Phi) μ 4 ( Φ ) , or the many-cluster shift of the sector centres given in closed form. (ii) For uniform hopping, the exchange coefficient depends on the graph only through its number of 4-cycles (Eq. (14), finite simple graphs, N ≥ 4 N \ge4 N ≥ 4 ); with bond-dependent amplitudes (spin-independent, Hermitian, no on-site terms) μ 4 \mu_4 μ 4 is still linear in P s a m e P_{\mathrm{same}} P same (the argument of § 6 goes through unchanged, § 5 ) but the slope differs. The square-lattice value − 64 / ( N − 1 ) -64/(N-1) − 64/ ( N − 1 ) , − 16 z / ( N − 1 ) -16z/(N-1) − 16 z / ( N − 1 ) , is not a law in the coordination number z z z (Appendix C), and the hypercubic form α d ( N , 2 ) \alpha_d(N,2) α d ( N , 2 ) holds only on tori whose sides are all 3 or ≥ 5 \ge5 ≥ 5 . (iii) The doublon Hamiltonian, Eq. (20), is derived for a single doublon; its π \pi π -periodicity under a twist is exact for that effective model, the exact two-electron U U U -band follows it only up to corrections of order ( 2 λ ) L / U L − 1 (2\lambda)^L/U^{L-1} ( 2 λ ) L / U L − 1 , and the full two-electron spectrum is not π \pi π -periodic. (iv) Nothing here concerns many-pair phase stiffness, off-diagonal long-range order, condensation, collective transport, or superconductivity; the single-pair flux response is a standard consequence of the effective doublon Hamiltonian. (v) Neither the no-double-occupancy projection, nor the moment method, nor the functional form of the doublon Hamiltonian is claimed as new § 1 , § 2 . (vi) What is and is not machine-checked is listed in Appendix F.
Acknowledgments
The formal proofs build on Mathlib, and we thank its contributors. This work received no external funding. Generative AI tools were used in this work. Claude (Anthropic) was used under the author’s direction to draft and revise text, code and Lean proofs, and as independent reviewers of the statements, proofs and prose, following the review procedure documented in the accompanying repository; GPT models (OpenAI) also served as independent reviewers of the statements. Qwen models (Alibaba Cloud) were used to test and review the scientific code, DeepSeek models in writing the numerical solvers, and the open-source coding agent opencode in the development of the code. The author checked the results and takes full responsibility for the content.
Exact hypercubic-lattice density of states
For a hypercubic lattice of dimension d d d , hopping λ \lambda λ , dispersion E ( k ) = − 2 λ ∑ i cos k i E(\mathbf{k}) = -2\lambda\sum_i \cos k_i E ( k ) = − 2 λ ∑ i cos k i , the density of states has the exact integral representation [26 ] (S1)
ρ d ( E ) = 1 π ∫ 0 ∞ cos ( E t ) [ J 0 ( 2 λ t ) ] d d t , (22) \rho_d(E) = \frac{1}{\pi} \int_0^\infty\cos(Et) [J_0(2\lambda t)]^d \,\mathrm{d}t, \tag*{(22)} ρ d ( E ) = π 1 ∫ 0 ∞ cos ( E t ) [ J 0 ( 2 λ t ) ] d d t , (22)
following from E E E being a sum of d d d independent “arcsine” random variables. For d = 1 d = 1 d = 1 the density of states is elementary (inverse-square-root band edges); for d = 2 d = 2 d = 2 it is a complete elliptic integral, logarithmically divergent at the band centre [27 ] ; for d ≥ 3 d \ge3 d ≥ 3 we evaluate the integral numerically.
Rather than display d = 1 d = 1 d = 1 –8 8 8 side by side, where the high-d d d curves become visually redundant once the van Hove structure has washed out, Figure 7 makes the approach to the central-limit Gaussian quantitative. Panel (a) shows d = 1 d = 1 d = 1 –4 4 4 on their native energy scale, where the van Hove structure (inverse-square-root divergences at d = 1 d = 1 d = 1 ; a logarithmic divergence at d = 2 d = 2 d = 2 ; square-root cusps, of finite value but divergent slope, at d = 3 d = 3 d = 3 ; a smooth maximum at d = 4 d = 4 d = 4 ) is visibly disappearing. Panel (b) variance-normalizes d = 4 , 5 , 6 , 8 d = 4, 5, 6, 8 d = 4 , 5 , 6 , 8 , x = E / μ 2 x = E/\sqrt{\mu_2} x = E / μ 2 , ρ ~ d ( x ) = μ 2 ρ d ( x μ 2 ) \tilde{\rho}_d(x) = \sqrt{\mu_2}\rho_d(x\sqrt{\mu_2}) ρ ~ d ( x ) = μ 2 ρ d ( x μ 2 ) , against the unit Gaussian g ( x ) = e − x 2 / 2 / 2 π g(x) = e^{-x^2/2}/\sqrt{2\pi} g ( x ) = e − x 2 /2 / 2 π : on this scale the curves are nearly indistinguishable by eye. Panel (c) plots the residual ρ ~ d ( x ) − g ( x ) \tilde{\rho}_d(x) - g(x) ρ ~ d ( x ) − g ( x ) , which remains clearly resolved and shrinks systematically with d d d .
Figure 7. Dimensional evolution of the hypercubic tight-binding density of states. (a) Low-dimensional DOS, showing the progressive smoothing of van Hove structure. (b) Variance-normalized DOS for d = 4 , 5 , 6 , 8 d = 4, 5, 6, 8 d = 4 , 5 , 6 , 8 , compared with the unit Gaussian approached as d → ∞ d \to\infty d → ∞ . (c) Difference from the Gaussian reference. Although the high-d d d curves are nearly indistinguishable on the scale of the DOS itself, their non-Gaussianity remains exactly resolved by the excess kurtosis γ 2 = − 3 / ( 2 d ) \gamma_2 = -3/(2d) γ 2 = − 3/ ( 2 d ) .
This is quantified exactly by the excess kurtosis. For isotropic hopping (c i = λ c_i = \lambda c i = λ , i = 1 , … , d i = 1,\ldots,d i = 1 , … , d ), μ 2 = 2 d λ 2 \mu_2 = 2d\lambda^2 μ 2 = 2 d λ 2 and μ 4 = 12 d 2 λ 4 − 6 d λ 4 \mu_4 = 12d^2\lambda^4 - 6d\lambda^4 μ 4 = 12 d 2 λ 4 − 6 d λ 4 (from the additive moment formulas below), so
γ 2 ≡ μ 4 μ 2 2 − 3 = − 3 2 d . (23) \gamma_2 \equiv\frac{\mu_4}{\mu_2^2} - 3 = -\frac{3}{2d}. \tag*{(23)} γ 2 ≡ μ 2 2 μ 4 − 3 = − 2 d 3 . (23)
The non-Gaussianity of ρ d \rho_d ρ d is therefore always present and exactly known at any d d d , even though it becomes visually imperceptible in the DOS itself for d ≳ 5 d \gtrsim5 d ≳ 5 : γ 2 → 0 \gamma_2 \to0 γ 2 → 0 as 1 / d 1/d 1/ d , not as a step function.
The representation generalizes immediately to anisotropic hopping amplitudes c i c_i c i (needed for mixed doublon/electron clusters, e.g. § 4.1 ), since the characteristic function of a sum of independent cosines factorizes regardless of amplitude:
ρ ( E ) = 1 π ∫ 0 ∞ cos ( E t ) ∏ i J 0 ( 2 c i t ) d t . (24) \rho(E) = \frac{1}{\pi} \int_0^\infty\cos(Et) \prod_i J_0(2c_i t)\,\mathrm{d}t. \tag*{(24)} ρ ( E ) = π 1 ∫ 0 ∞ cos ( E t ) i ∏ J 0 ( 2 c i t ) d t . (24)
For a sum of independent cosine terms E = − 2 ∑ i c i cos k i E = -2\sum_i c_i\cos k_i E = − 2 ∑ i c i cos k i , the exact moments follow without numerical integration: per-term variance 2 c i 2 2c_i^2 2 c i 2 and per-term 4th cumulant − 6 c i 4 -6c_i^4 − 6 c i 4 , both additive, giving Var = 2 ∑ i c i 2 \operatorname{Var} = 2\sum_i c_i^2 Var = 2 ∑ i c i 2 , μ 4 = − 6 ∑ i c i 4 + 3 Var 2 \mu_4 = -6\sum_i c_i^4 + 3\operatorname{Var}^2 μ 4 = − 6 ∑ i c i 4 + 3 Var 2 , used for the naive reference variance of §5.3 and the excess-kurtosis comparison of § 4 .
For k k k distinct labels (single-particle states, possibly degenerate) sampled without replacement from a population { ϵ 1 , … , ϵ N } \{\epsilon_1,\ldots,\epsilon_N\} { ϵ 1 , … , ϵ N } of zero mean (∑ i ϵ i = 0 \sum_i \epsilon_i = 0 ∑ i ϵ i = 0 , as for any lattice band without on-site energy), classifying the 4-index expansion of S 4 S^4 S 4 (S S S = sum of sampled values) by coincidence pattern (all equal; 3+1; 2+2; 2+1+1; all distinct) and writing p m = k ( k − 1 ) ⋯ ( k − m + 1 ) / [ N ( N − 1 ) ⋯ ( N − m + 1 ) ] p_m = k(k-1)\cdots(k-m+1)/[N(N-1)\cdots(N-m+1)] p m = k ( k − 1 ) ⋯ ( k − m + 1 ) / [ N ( N − 1 ) ⋯ ( N − m + 1 )] for the probability that m m m given distinct labels all belonging to the sample gives
μ 4 = P 4 ( p 1 − 7 p 2 + 12 p 3 − 6 p 4 ) + 3 P 2 2 ( p 2 − 2 p 3 + p 4 ) , (25) \mu_4 = P_4(p_1 - 7p_2 + 12p_3 - 6p_4) + 3P_2^2(p_2 - 2p_3 + p_4), \tag*{(25)} μ 4 = P 4 ( p 1 − 7 p 2 + 12 p 3 − 6 p 4 ) + 3 P 2 2 ( p 2 − 2 p 3 + p 4 ) , (25)
with the power sums P 2 = ∑ ε i 2 P_2 = \sum\varepsilon_i^2 P 2 = ∑ ε i 2 , P 4 = ∑ ε i 4 P_4 = \sum\varepsilon_i^4 P 4 = ∑ ε i 4 (not to be confused with the pair counts P s a m e P_{\mathrm{same}} P same , P 3 P_3 P 3 of § 6 ). Validated against brute-force enumeration on a random centred toy population (exactly, in rational arithmetic, k = 2 , 3 , 4 k = 2, 3, 4 k = 2 , 3 , 4 ) and against the real L = 4 L = 4 L = 4 single-particle spectrum (Table 2 ); for a U = 0 U=0 U = 0 spectrum with both spin species, the independent pieces are combined via the sum-of-independent-variables identity μ 4 ( X + Y ) = μ 4 ( X ) + μ 4 ( Y ) + 6 Var ( X ) Var ( Y ) \mu_4(X+Y) = \mu_4(X) + \mu_4(Y) + 6\operatorname{Var}(X)\operatorname{Var}(Y) μ 4 ( X + Y ) = μ 4 ( X ) + μ 4 ( Y ) + 6 Var ( X ) Var ( Y ) . The formula, valid for every k ≤ N k \le N k ≤ N with p m = 0 p_m = 0 p m = 0 for m > k m > k m > k , the finite-population variance, the additivity identity, and the six exact values of Table 2 are machine-checked (Appendix § 16 ).
The μ 4 ( U → ∞ ) \mu_4(U \to\infty) μ 4 ( U → ∞ ) closed-walk classification
Four-step closed walks in the N d = 0 N_d = 0 N d = 0 configuration graph classify by how many original particles participate:
Type I — single particle, self-returning. Four sub-types: same-neighbor out-and-back twice; two different neighbors, sequentially; “there, one step further, and back”; and a genuine 4-cycle/plaquette loop. Each depends only on joint “these m m m additional sites are simultaneously empty, given the source occupied” probabilities: a symmetric hypergeometric factor in ( n , N ) (n, N) ( n , N ) only, with no dependence on how n n n splits into ( N ↑ , N ↓ ) (N_{\uparrow}, N_{\downarrow}) ( N ↑ , N ↓ ) , by the identical argument as μ 2 \mu_2 μ 2 (§ 5.2 ).
Type II(a) — two different particles, each returning. Two particles each make a two-hop excursion and return (the excursions may share an intermediate site); the walk induces the identity and depends only on which sites are empty, so it is spin-blind.
Type II(b) — ring exchange (transposition). Two occupied sites i , j i, j i , j on a common 4-cycle admit a 4-hop sequence, through the cycle’s two other (empty) corners, that returns the occupied site set to { i , j } \{i,j\} { i , j } but permutes which original particle occupies which site.
Type III — 3-cycle. Three occupied sites i , j , k i, j, k i , j , k and one auxiliary empty site m m m (adjacent to two of the three, forming an L-shaped or plaquette-corner arrangement) admit the 2+1+1-hop sequence i → m i \to m i → m , j → i j \to i j → i , k → j k \to j k → j , m → k m \to k m → k : the particle originally at i i i ends at k k k (via m m m , two hops), the particle at j j j ends at i i i (one hop), the particle at k k k ends at j j j (one hop): a genuine 3-cycle ( i k j ) (i\,k\,j) ( i k j ) using exactly 4 hops. Explicit check: on the square lattice, taking i = ( 0 , 0 ) i = (0, 0) i = ( 0 , 0 ) , j = ( 1 , 0 ) j = (1, 0) j = ( 1 , 0 ) , k = ( 1 , 1 ) k = (1, 1) k = ( 1 , 1 ) , m = ( 0 , 1 ) m = (0, 1) m = ( 0 , 1 ) realizes this, since m m m is a common neighbor of both i i i and k k k .
Figure 8 draws each type, with the interference that sets its response to a magnetic flux (§ 8 ). For Type II(b), the correct statement is in the occupation-number basis, not in labelled-particle language: writing the initial state as ∣ α ⟩ = c i σ † c j σ ′ † ∣ ⋯ ⟩ |\alpha\rangle= c^\dagger_{i\sigma}c^\dagger_{j\sigma'}|\cdots\rangle ∣ α ⟩ = c iσ † c j σ ′ † ∣ ⋯ ⟩ and the representative diagonal contribution to H ∞ 4 H_\infty^4 H ∞ 4 as the normal-ordered hop sequence around the plaquette i , k , j , l i,k,j,l i , k , j , l ,
Figure 8. (a) Type I(i). out and back twice. One route: flux-blind. Any spins.
(b) Types I(ii), I(iii). two neighbors in turn, or one step further and back. One route: flux-blind. Any spins.
(c) Type I(iv). plaquette loop: two routes to the far corner, constructive, ∝ 1 + cos Φ □ \propto1 + \cos\Phi_{\square} ∝ 1 + cos Φ □ . Any spins.
(d) Type II(a). two particles, each out and back. One route each: flux-blind. Any spins.
(e) Type II(b), diagonal. ring exchange of i i i , j j j through k k k , l l l : four routes to { k , l } \{k,l\} { k , l } . Same spins: they cancel at Φ □ = 0 \Phi_{\square}=0 Φ □ = 0 ; ∝ 1 − cos Φ □ \propto1-\cos\Phi_{\square} ∝ 1 − cos Φ □ . Opposite spins: flux-blind.
(f) Type II(b), edge. i i i goes three hops around, j j j one hop along the bond: two routes per intermediate, over two plaquettes. Same spins: ∝ 1 − cos Φ □ \propto1-\cos\Phi_{\square} ∝ 1 − cos Φ □ . Opposite spins: flux-blind.
(g) Type III, 3-cycle. the empty corner goes once around: two routes (i → m i \to m i → m or k → m k \to m k → m first), constructive, ∝ 1 + cos Φ □ \propto1+\cos\Phi_{\square} ∝ 1 + cos Φ □ . All three spins equal.
occupied site empty site hop, spin-blind type hop, spin-sensitive type dashed: second route or second particle numbers: hop order
Figure 8: The four-hop closed-walk types of μ 4 ( U → ∞ ) \mu_4(U \to\infty) μ 4 ( U → ∞ ) , drawn on one plaquette. The number of routes that reach the same two-hop intermediate decides the response to a plaquette flux Φ □ \Phi_{\square} Φ □ § 8 : a single route is flux-blind, while two or more interfere with relative Wilson phase Φ □ \Phi_{\square} Φ □ . The spin-sensitive types (vermillion) return to the initial configuration only when the particles involved have equal spins.
⟨ α ∣ c j σ ′ † c k σ † c i σ † c l σ ′ † c l σ ′ c j σ ′ c k σ c i σ ∣ α ⟩ (26) \langle\alpha|c^\dagger_{j\sigma'}c^\dagger_{k\sigma}c^\dagger_{i\sigma}c^\dagger_{l\sigma'}c_{l\sigma'}c_{j\sigma'}c_{k\sigma}c_{i\sigma}|\alpha\rangle \tag*{(26)} ⟨ α ∣ c j σ ′ † c k σ † c iσ † c l σ ′ † c l σ ′ c j σ ′ c k σ c iσ ∣ α ⟩ (26)
the operator string returns every mode’s occupation number to its initial value (hence contributes to the diagonal) only when σ = σ ′ \sigma= \sigma' σ = σ ′ . For σ ≠ σ ′ \sigma\ne\sigma' σ = σ ′ it vanishes identically: its third hop (the second factor as written), c i σ † c l σ c^\dagger_{i\sigma}c_{l\sigma} c iσ † c l σ , must remove a σ \sigma σ electron from l l l , which holds σ ′ \sigma' σ ′ . Written instead with the spin labels following the particles (σ ′ \sigma' σ ′ from l l l to i i i , σ \sigma σ from k k k to $j), the string returns the occupied sites but leaves σ ′ \sigma' σ ′ at i i i and σ \sigma σ at j j j , the opposite of ∣ α ⟩ |\alpha\rangle ∣ α ⟩ , so ⟨ α ∣ ⋯ ∣ α ⟩ = 0 \langle\alpha|\cdots|\alpha\rangle= 0 ⟨ α ∣ ⋯ ∣ α ⟩ = 0 again, independent of any sign. For σ = σ ′ \sigma= \sigma' σ = σ ′ the occupation numbers do return to ∣ α ⟩ |\alpha\rangle ∣ α ⟩ and the matrix element is nonzero; its
sign and magnitude are fixed by the fermionic operator ordering above (the closed-walk amplitude lemma below).
The identical argument applies to the Type III 3-cycle: the final occupation-number state returns to ∣ α ⟩ |\alpha\rangle ∣ α ⟩ if and only if the spin at each site after the cycle matches the spin before it, which for the cycle ( i k j ) (i\ k\ j) ( i k j ) requires σ i = σ j = σ k \sigma_i=\sigma_j=\sigma_k σ i = σ j = σ k (all three participating particles the same spin), contributing a term proportional to P 3 = ( N ↑ 3 ) + ( N ↓ 3 ) P_3=\binom{N_{\uparrow}}{3}+\binom{N_{\downarrow}}{3} P 3 = ( 3 N ↑ ) + ( 3 N ↓ ) when summed over eligible triples.
Lemma (amplitude of a closed walk). Let w w w be a site walk of four hops from an occupied site set S S S that returns the occupied set to S S S , inducing the permutation π w \pi_w π w of S S S , and let σ \sigma σ be a spin labelling of S S S . Applying the four hops of w w w to the basis state ∣ S , σ ⟩ |S,\sigma\rangle ∣ S , σ ⟩ , each as its term − λ c y τ † c x τ -\lambda c^\dagger_{y\tau}c_{x\tau} − λ c y τ † c xτ of H ∞ H_\infty H ∞ (τ \tau τ the spin of the moving particle), gives a basis state; it is ∣ S , σ ⟩ |S,\sigma\rangle ∣ S , σ ⟩ itself if and only if σ \sigma σ is constant on the cycles of π w \pi_w π w , and then the product of the four matrix elements is ( − λ ) 4 sgn ( π w ) = λ 4 sgn ( π w ) (-\lambda)^4\operatorname{sgn}(\pi_w)=\lambda^4\operatorname{sgn}(\pi_w) ( − λ ) 4 sgn ( π w ) = λ 4 sgn ( π w ) , whatever the spins of the other particles and whatever ordering of the 2 N 2N 2 N modes defines the basis.
Proof. Write ∣ S , σ ⟩ = ϵ c s 1 σ ( s 1 ) † ⋯ c s n σ ( s n ) † ∣ 0 ⟩ |S,\sigma\rangle=\epsilon c^\dagger_{s_1\sigma(s_1)}\cdots c^\dagger_{s_n\sigma(s_n)}|0\rangle ∣ S , σ ⟩ = ϵ c s 1 σ ( s 1 ) † ⋯ c s n σ ( s n ) † ∣0 ⟩ , with s 1 , … , s n s_1,\ldots,s_n s 1 , … , s n the sites of S S S in any fixed order and ϵ = ± 1 \epsilon=\pm1 ϵ = ± 1 the sign relating this product to the basis vector of the chosen mode ordering. Let a hop c y τ † c x τ c^\dagger_{y\tau}c_{x\tau} c y τ † c xτ , with y y y empty, act on such a product in which c x τ † c^\dagger_{x\tau} c xτ † stands in slot k k k . Moving c x τ c_{x\tau} c xτ to slot k k k gives ( − 1 ) k − 1 (-1)^{k-1} ( − 1 ) k − 1 ; there c x τ c x τ † c_{x\tau}c^\dagger_{x\tau} c xτ c xτ † acts as the identity, since no other factor creates the mode ( x , τ ) (x,\tau) ( x , τ ) ; and moving c y τ † c^\dagger_{y\tau} c y τ † into slot k k k gives ( − 1 ) k − 1 (-1)^{k-1} ( − 1 ) k − 1 again. So a hop replaces c x τ † c^\dagger_{x\tau} c xτ † by c y τ † c^\dagger_{y\tau} c y τ † in its slot, with no sign. After the four hops, slot k k k holds c π w ( s k ) σ ( s k ) † c^\dagger_{\pi_w(s_k)\sigma(s_k)} c π w ( s k ) σ ( s k ) † and the prefactor is ϵ ( − λ ) 4 \epsilon(-\lambda)^4 ϵ ( − λ ) 4 : site π w ( s k ) \pi_w(s_k) π w ( s k ) carries spin σ ( s k ) \sigma(s_k) σ ( s k ) . This is ∣ S , σ ⟩ |S,\sigma\rangle ∣ S , σ ⟩ if and only if σ ( π w ( s ) ) = σ ( s ) \sigma(\pi_w(s))=\sigma(s) σ ( π w ( s )) = σ ( s ) for all s s s , and otherwise an orthogonal basis state. In the first case, sorting the slots back into the order s 1 , … , s n s_1,\ldots,s_n s 1 , … , s n permutes n n n anticommuting factors by π w \pi_w π w , giving sgn ( π w ) \operatorname{sgn}(\pi_w) sgn ( π w ) , and ϵ 2 = 1 \epsilon^2=1 ϵ 2 = 1 . □
The individual hop signs depend on the mode ordering and on where the spectators sit; only the product around a closed walk is the convention-free sgn ( π w ) \operatorname{sgn}(\pi_w) sgn ( π w ) . Counting the labellings that a given π \pi π preserves gives the fractions used in §6: for a transposition ( a b ) (ab) ( ab ) , [ ( N ↑ − 2 N ↑ − 2 ) + ( N ↓ − 2 N ↓ − 2 ) ] / ( N 2 ) = P s a m e / ( N 2 ) \left[\binom{N_{\uparrow}-2}{N_{\uparrow}-2}+\binom{N_{\downarrow}-2}{N_{\downarrow}-2}\right]\big/\binom{N}{2}=P_{\mathrm{same}}/\binom{N}{2} [ ( N ↑ − 2 N ↑ − 2 ) + ( N ↓ − 2 N ↓ − 2 ) ] / ( 2 N ) = P same / ( 2 N ) ; for a 3-cycle, [ ( N ↑ − 3 N ↑ − 3 ) + ( N ↓ − 3 N ↓ − 3 ) ] / ( N 3 ) = P 3 / ( N 3 ) \left[\binom{N_{\uparrow}-3}{N_{\uparrow}-3}+\binom{N_{\downarrow}-3}{N_{\downarrow}-3}\right]\big/\binom{N}{3}=P_3/\binom{N}{3} [ ( N ↑ − 3 N ↑ − 3 ) + ( N ↓ − 3 N ↓ − 3 ) ] / ( 3 N ) = P 3 / ( 3 N ) . With bond-dependent amplitudes t x y t_{xy} t x y (the same for both spins, Hermitian, no on-site terms) the same proof gives the product of the walk’s four amplitudes in place of λ 4 \lambda^4 λ 4 , a factor that depends on the walk but not on the spins, so (10) holds with I 1 , I 2 , I 3 I_1,I_2,I_3 I 1 , I 2 , I 3 replaced by amplitude-weighted walk sums, and μ 4 \mu_4 μ 4 stays affine in P s a m e P_{\mathrm{same}} P same (S5).
The 3-cycle’s local path multiplicity, exhaustively. The construction above exhibits one explicit 4-hop realization of the 3-cycle; exhaustively enumerating ( H ∞ 2 ) a b = ∑ e ( H ∞ ) a e ( H ∞ ) e b (H_\infty^2)_{ab}=\sum_e(H_\infty)_{ae}(H_\infty)_{eb} ( H ∞ 2 ) ab = ∑ e ( H ∞ ) a e ( H ∞ ) e b over every locally allowed intermediate e e e , for a = { i , j , k } a=\{i,j,k\} a = { i , j , k } and b = { i , k , m } b=\{i,k,m\} b = { i , k , m } , shows this is incomplete. Restricted to the four local sites, only two first hops are ever legal from a a a (i → m i\to m i → m or k → m k\to m k → m ; every other local pair is occupied-occupied), and each completes to b b b uniquely (j → i j\to i j → i or j → k j\to k j → k respectively): an exhaustive, not illustrative, count. Both routes carry the same fermion sign under the sorted-tuple rule (direct computation), so they interfere constructively: ( H ∞ 2 ) a b = − 2 λ 2 (H_\infty^2)_{ab}=-2\lambda^2 ( H ∞ 2 ) ab = − 2 λ 2 , twice the magnitude ( ∣ ( H ∞ 2 ) a b ∣ 2 = 4 λ 4 ) (|(H_\infty^2)_{ab}|^2=4\lambda^4) ( ∣ ( H ∞ 2 ) ab ∣ 2 = 4 λ 4 ) of a single route. The exact sparse-trace numbers of this paper sum over every second-quantized path and so include both routes.
A four-particle cyclic permutation’s diagonal spin condition is σ i = σ j = σ k = σ l \sigma_i=\sigma_j=\sigma_k=\sigma_l σ i = σ j = σ k = σ l (all four the same spin), contributing a term ∝ P 4 s a m e = ( N ↑ 4 ) + ( N ↓ 4 ) \propto P_4^{\mathrm{same}}=\binom{N_{\uparrow}}{4}+\binom{N_{\downarrow}}{4} ∝ P 4 same = ( 4 N ↑ ) + ( 4 N ↓ ) , which is not affine in P s a m e P_{\mathrm{same}} P same (unlike P 3 P_3 P 3 ). A double transposition ( i j ) ( k l ) (ij)(kl) ( ij ) ( k l ) has a different and more dangerous condition: only σ i = σ j \sigma_i=\sigma_j σ i = σ j and independently σ k = σ l \sigma_k=\sigma_l σ k = σ l , with no requirement relating the two pairs to each other: an ↑ ↑ ↓ ↓ \uparrow\uparrow\downarrow\downarrow ↑↑↓↓ configuration survives it as well as an all-↑ \uparrow ↑ one. Its spin trace is therefore a product of two pair-equality indicators, exactly the kind of structure capable of generating a genuine P s a m e 2 P_{\mathrm{same}}^2 P same 2 term. Both cycle types must therefore be excluded for exact linearity to hold.
No four-particle permutation is reachable in four hops. A permutation with no fixed points among four sites { i , j , k , l } \{i,j,k,l\} { i , j , k , l } (either cycle type above) requires all four participating particles to move at least once; with exactly four hops available, this forces a 1 + 1 + 1 + 1 1+1+1+1 1 + 1 + 1 + 1 distribution, since any particle receiving zero hops would remain fixed. But a 1 + 1 + 1 + 1 1+1+1+1 1 + 1 + 1 + 1 realization requires each particle’s single hop to land directly on its final site among the four, and every one of those four sites is occupied at the outset (the four sites are exactly the support of the permutation). The hop into a site i i i can proceed only once i i i ’s occupant has left, but that occupant’s hop is itself one of the four hops being scheduled, and by the same argument it must wait for the occupant of its own target site to leave first, and so on around the cycle back to i i i : a closed dependency loop with no legal first move. (A double transposition faces the identical obstruction independently on each of its two 2 2 2 -cycles.) Four-particle permutations are therefore impossible within a four-hop closed walk on any graph without self-loops; only e e e , transpositions, and 3 3 3 -cycles are accessible, matching the closed-walk types enumerated above. The same argument proves more: on any graph, a closed walk of h h h hard-core hops moves either no particle or fewer than h h h of them (if h h h particles moved, each would move once, onto a site whose occupant had to leave first, and the hop times would decrease around every cycle). This general statement, its h = 4 h=4 h = 4 case, and the realization of both a transposition and a 3 3 3 -cycle in four hops on every L × L L \times L L × L torus with L ≥ 3 L \ge3 L ≥ 3 are machine-checked (Appendix § 16 ).
We cross-checked this directly on a fully occupied plaquette at n = 4 n=4 n = 4 : a diagonal pair ( i , j ) (i,j) ( i , j ) with both mediators k , l k,l k , l also occupied (k , l k,l k , l are themselves a diagonal pair, with i , j i,j i , j as their mediators). With all four corners occupied, the diagonal exchanges of this plaquette are blocked; what remains spin-dependent is the edge channel, in which a same-spin nearest-neighbor pair of the plaquette exchanges by three hops around its outer neighboring plaquette and one back, four walks of sign − 1 -1 − 1 . Writing b s b_s b s for the number of same-spin bonds of the plaquette, the measured values follow exactly ( H ∞ 4 ) a a = 208 − 4 b s (H^4_{\infty})_{aa}=208-4b_s ( H ∞ 4 ) aa = 208 − 4 b s : 208 when i , j i,j i , j and k , l k,l k , l carry opposite spins (no same-spin bond), 200 when one of k , l k,l k , l shares the spin of i , j i,j i , j (two), and 192 when all four spins are equal (four), identical at L = 5 L=5 L = 5 and L = 6 L=6 L = 6 . No four-particle process contributes.
Two geometrically distinct channels, and their exact n = 2 n=2 n = 2 coefficient
Evaluating the analogous strings for every distinct pair type on the square lattice (verified against exact sparse traces at n = 2 n=2 n = 2 , where Type II(b) is isolated exactly since no other particles can occupy the intermediate sites) reveals two , not one, contributing geometries:
Diagonal pairs ( i , j (i,j ( i , j next-nearest neighbors, sharing two common neighbors k , l k,l k , l directly, an elementary plaquette’s opposite corners): the cyclic sequence i → k → j → l → i i \to k \to j \to l \to i i → k → j → l → i (either orientation) contributes.
Edge pairs ( i , j (i,j ( i , j nearest neighbors): a 3-hop+1-hop traversal of an adjacent plaquette also contributes, e.g. one particle traverses i → m → p → j i \to m \to p \to j i → m → p → j (three sides of the plaquette, m , p m,p m , p its other two corners) while the other simply crosses the shared edge j → i j \to i j → i once the direct site is vacated. This channel has no counterpart in a count of opposite plaquette corners alone.
At n = 2 n=2 n = 2 , where the two occupied sites are the entire system and the required intermediate sites are guaranteed empty, exact evaluation gives identical magnitude for both channels : − 8 λ 4 -8\lambda^4 − 8 λ 4 per eligible pair, verified at L = 5 , 6 , 7 L=5,6,7 L = 5 , 6 , 7 (at L = 3 L=3 L = 3 as well; L = 4 L=4 L = 4 differs, see below). Each channel has 2 N 2N 2 N eligible pairs, which gives α ( N , 2 ) = ( − 8 ) ⋅ 2 ⋅ 2 N / ( N 2 ) = − 64 / ( N − 1 ) \alpha(N,2)=(-8)\cdot2\cdot2N/\binom{N}{2}=-64/(N-1) α ( N , 2 ) = ( − 8 ) ⋅ 2 ⋅ 2 N / ( 2 N ) = − 64/ ( N − 1 ) (§6), matching the exact evaluation at every L ≠ 4 L\ne4 L = 4 tested. The sign mechanism, derived completely. Writing μ 4 = ∑ a , b ( H ∞ 2 ) a b 2 / D 0 \mu_4 = \sum_{a,b}(H^2_{\infty})_{ab}^2/D_0 μ 4 = ∑ a , b ( H ∞ 2 ) ab 2 / D 0 , the square-lattice − 8 λ 4 -8\lambda^4 − 8 λ 4 per-pair magnitude common to both channels follows from a direct, complete fermion-sign count of every second-quantized hopping path connecting the reference state a a a to each relevant two-hop intermediate b b b , not merely a qualitative interference argument.
Diagonal channel. Take a same-spin pair i , j i,j i , j with common neighbors k , l k,l k , l . The single relevant intermediate is the “fully swapped” configuration c = { k , l } c=\{k,l\} c = { k , l } , reached in exactly two hops by four distinct paths: two choices of which particle moves to which mediator (i → k , j → l i\to k,j\to l i → k , j → l or i → l , j → k i\to l,j\to k i → l , j → k ), each with two hop-orderings. Direct evaluation of the fermion sign (using the sorted-tuple rule of App. E) for a concrete instance (i , j i,j i , j diagonal, k , l k,l k , l the shared neighbors) gives the four path signs − 1 , + 1 , + 1 , − 1 -1,+1,+1,-1 − 1 , + 1 , + 1 , − 1 , which sum to zero: ( H ∞ 2 ) a , c = 0 (H^2_{\infty})_{a,c}=0 ( H ∞ 2 ) a , c = 0 exactly for same spin, an exact destructive cancellation between the two particle-assignment choices (which differ by an odd permutation of indistinguishable fermions). For an opposite-spin pair, the two assignments produce two distinguishable target states (↑ \uparrow ↑ at k , ↓ k,\downarrow k , ↓ at l l l , or vice versa: genuinely different basis vectors, since the spin label now marks which particle sits where); each is reached by the two hop-orderings only, which share a sign and add constructively, giving ( H ∞ 2 ) = 2 λ 2 (H^2_{\infty})=2\lambda^2 ( H ∞ 2 ) = 2 λ 2 for each of the two targets. Summing ( H ∞ 2 ) 2 (H^2_{\infty})^2 ( H ∞ 2 ) 2 over both: same spin contributes 0 0 0 , opposite spin contributes 2 × ( 2 λ 2 ) 2 = 8 λ 4 2\times(2\lambda^2)^2=8\lambda^4 2 × ( 2 λ 2 ) 2 = 8 λ 4 . The difference, 0 − 8 λ 4 = − 8 λ 4 = − 2 z λ 4 0-8\lambda^4=-8\lambda^4=-2z\lambda^4 0 − 8 λ 4 = − 8 λ 4 = − 2 z λ 4 at z = 4 z=4 z = 4 , is the diagonal channel’s coefficient, obtained with no fitted or imported number. In the signed closed-walk bookkeeping of Eq. (10) the same difference appears instead as transposition walks of sign − 1 -1 − 1 that return to the reference state only for the same-spin pair: the two bookkeepings assign the − 8 λ 4 -8\lambda^4 − 8 λ 4 to different terms but agree on the same-spin minus opposite-spin difference.
Edge channel. Take an adjacent pair i , j i,j i , j and the two other corners m , p m,p m , p of one plaquette containing edge i i i -j j j . Two two-hop intermediates are relevant, c 1 = { p , j } c_1=\{p,j\} c 1 = { p , j } and c 2 = { m , i } c_2=\{m,i\} c 2 = { m , i } (the two ways of interleaving the 3-hop+1-hop process), each reached by exactly two paths: a genuine two-particle path (both original particles move once) and a single-particle path (one particle hops twice through the plaquette). For same-spin, direct sign evaluation gives, for both c 1 c_1 c 1 and c 2 c_2 c 2 , one path of each sign (− 1 , + 1 -1,+1 − 1 , + 1 ), so both cancel exactly: ( H ∞ 2 ) a , c 1 = ( H ∞ 2 ) a , c 2 = 0 (H^2_{\infty})_{a,c_1}=(H^2_{\infty})_{a,c_2}=0 ( H ∞ 2 ) a , c 1 = ( H ∞ 2 ) a , c 2 = 0 . For opposite spin, the two paths to each target are no longer the same target : they resolve into two distinguishable, singly-reached configurations per intermediate (↑ ↓ \uparrow\downarrow ↑↓ interchanged), each with ( H ∞ 2 ) = λ 2 (H^2_{\infty})=\lambda^2 ( H ∞ 2 ) = λ 2 (a single contributing path, sign + 1 +1 + 1 ), contributing λ 4 \lambda^4 λ 4 each: four distinguishable targets (two from c 1 c_1 c 1 , two from c 2 c_2 c 2 ) give 4 λ 4 4\lambda^4 4 λ 4 from this one plaquette. Because an edge belongs to two plaquettes on the square lattice, the mirror plaquette (corners m ′ , p ′ m',p' m ′ , p ′ on the other side of edge i i i -j j j ) contributes an independent, identical 4 λ 4 4\lambda^4 4 λ 4 , same-spin again canceling to zero by the identical sign argument. Summing both plaquettes: same spin 0 0 0 , opposite spin 8 λ 4 8\lambda^4 8 λ 4 , difference − 8 λ 4 = − 2 z λ 4 -8\lambda^4=-2z\lambda^4 − 8 λ 4 = − 2 z λ 4 at z = 4 z=4 z = 4 , identical to the diagonal channel, but reaching the same total through a structurally different route (one plaquette with two constructively-interfering paths per target, versus two plaquettes each contributing single-path targets).
Both channels’ − 8 λ 4 -8\lambda^4 − 8 λ 4 coefficients are therefore derived by a complete accounting of fermion signs on every contributing path, and confirmed by exact evaluation.
Does the square-lattice value generalize? Writing the square-lattice result as α ( N , 2 ) = − 16 z / ( N − 1 ) \alpha(N,2)=-16z/(N-1) α ( N , 2 ) = − 16 z / ( N − 1 ) (that is, − 64 / ( N − 1 ) -64/(N-1) − 64/ ( N − 1 ) with z = 4 z=4 z = 4 restored) packages the per-pair magnitude and the pair count together, and unlike the second-moment theorem this packaging is not graph-general: “16 z 16z 16 z ” is a square-lattice mnemonic for one specific number, not a law in z z z . Repeating the exact sparse-trace measurement on a 3-dimensional simple-cubic lattice with periodic boundaries (z = 6 z=6 z = 6 ) gives α ( 27 , 2 ) = − 7.385 = − 96 / 13 \alpha(27,2)=-7.385=-96/13 α ( 27 , 2 ) = − 7.385 = − 96/13 , twice the substitution − 16 z / ( N − 1 ) = − 48 / 13 ≈ − 3.692 -16z/(N-1)=-48/13\approx-3.692 − 16 z / ( N − 1 ) = − 48/13 ≈ − 3.692 . This factor of two is not accounted for by the edge channel alone: direct enumeration shows both channels independently double in three dimensions, by two unrelated geometric mechanisms that happen to share the same d − 1 d-1 d − 1 scaling. The edge channel doubles because each edge of the simple-cubic lattice belongs to 4 elementary plaquettes rather than 2 (one pair for each of the two directions transverse to the edge). The diagonal channel doubles independently , for a different reason: a site has 2 d ( d − 1 ) = 12 2d(d-1)=12 2 d ( d − 1 ) = 12 face-diagonal (next-nearest, two-common-neighbor) neighbors in three dimensions, not the z = 6 z=6 z = 6 a naive substitution would suggest: the square lattice’s diagonal-type neighbor count happening to equal its coordination number z z z is a d = 2 d=2 d = 2 coincidence, not a general fact. Each channel supplies an equal half of the doubled total (− 1296 λ 4 -1296\lambda^4 − 1296 λ 4 of − 2592 λ 4 -2592\lambda^4 − 2592 λ 4 at N = 27 N=27 N = 27 ). Both mechanisms are 4-cycle counts, and the general statement is Eq. (13) at n = 2 n=2 n = 2 , α ( N , 2 ) = − 64 C 4 / [ N ( N − 1 ) ] \alpha(N,2)=-64C_4/[N(N-1)] α ( N , 2 ) = − 64 C 4 / [ N ( N − 1 )] : the exchange coefficient counts 4-cycles, not neighbors (S6). On the d d d -dimensional hypercubic torus with every side 3 or ≥ 5 \ge5 ≥ 5 , the 4-cycles are the plaquettes, C 4 = N d ( d − 1 ) / 2 C_4=Nd(d-1)/2 C 4 = N d ( d − 1 ) /2 , so
α d ( N , 2 ) = − 32 d ( d − 1 ) N − 1 = − 16 ( d − 1 ) z N − 1 ( z = 2 d ) , (27) \alpha_d(N,2)=-\frac{32d(d-1)}{N-1}=-\frac{16(d-1)z}{N-1}\qquad(z=2d), \tag*{(27)} α d ( N , 2 ) = − N − 1 32 d ( d − 1 ) = − N − 1 16 ( d − 1 ) z ( z = 2 d ) , (27)
reproducing α ( 27 , 2 ) = − 96 / 13 \alpha(27,2)=-96/13 α ( 27 , 2 ) = − 96/13 at d = 3 d=3 d = 3 (and − 48 / 31 -48/31 − 48/31 on the 5 × 5 × 5 5\times5\times5 5 × 5 × 5 lattice). Each side of length 4 adds axis rings of length four as further 4-cycles; with m m m such sides, α ( N , 2 ) = − 16 [ 2 d ( d − 1 ) + m ] / ( N − 1 ) \alpha(N,2)=-16[2d(d-1)+m]/(N-1) α ( N , 2 ) = − 16 [ 2 d ( d − 1 ) + m ] / ( N − 1 ) (4 × 4 × 4 4\times4\times4 4 × 4 × 4 : − 80 / 21 -80/21 − 80/21 ; 3 × 4 3\times4 3 × 4 : − 80 / 11 -80/11 − 80/11 ; 4 × 5 4\times5 4 × 5 : − 80 / 19 -80/19 − 80/19 ). A lattice with the same d d d and z z z but a different 4-cycle structure behaves differently: the kagome lattice, with d = 2 d=2 d = 2 and z = 4 z=4 z = 4 like the square lattice, has no 4-cycles and α ( N , 2 ) = 0 \alpha(N,2)=0 α ( N , 2 ) = 0 exactly.
The L = 4 L=4 L = 4 torus, exactly. At L = 4 L=4 L = 4 , α ( 16 , 2 ) = − 32 / 5 \alpha(16,2)=-32/5 α ( 16 , 2 ) = − 32/5 rather than the − 64 / 15 -64/15 − 64/15 of larger tori. In terms of Eq. (13) , the eight axis rings of length four are 4-cycles in addition to the sixteen plaquettes, C 4 = 24 C_4=24 C 4 = 24 , and − 64 ⋅ 24 / ( 16 ⋅ 15 ) = − 32 / 5 -64\cdot24/(16\cdot15)=-32/5 − 64 ⋅ 24/ ( 16 ⋅ 15 ) = − 32/5 . Pair by pair, this decomposes exactly into two independent, equal contributions of − 16 / 15 -16/15 − 16/15 each: (i) 16 extra diagonal-type pairs beyond the generic 32, since sites displaced by ( 2 , 0 ) (2,0) ( 2 , 0 ) or ( 0 , 2 ) (0,2) ( 0 , 2 ) acquire a spurious second common neighbor via periodic wraparound on this specific 4 × 4 4\times4 4 × 4 torus (generically they share only one, insufficient for the ring-exchange mechanism); (ii) the edge channel’s per-pair magnitude is itself enhanced from − 8 -8 − 8 to − 12 -12 − 12 by an additional wraparound-enabled route unavailable at larger L L L . Both are confirmed by direct enumeration; together, − 64 / 15 − 16 / 15 − 16 / 15 = − 96 / 15 = − 32 / 5 -64/15-16/15-16/15=-96/15=-32/5 − 64/15 − 16/15 − 16/15 = − 96/15 = − 32/5 exactly.
Extension to n > 2 n>2 n > 2 . Independent pairwise blocking (Eq. (12) in the main text) is exactly the transposition part of α ( N , n ) \alpha(N,n) α ( N , n ) ; on its own it misses the slope by 1.8% (n = 3 n=3 n = 3 ) and 3.9% (n = 4 n=4 n = 4 ) at L = 5 L=5 L = 5 . Classifying closed walks by hop partition ( 4 , 3 + 1 , 2 + 2 , 2 + 1 + 1 , 1 + 1 + 1 + 1 ) (4,\ 3+1,\ 2+2,\ 2+1+1,\ 1+1+1+1) ( 4 , 3 + 1 , 2 + 2 , 2 + 1 + 1 , 1 + 1 + 1 + 1 ) places the four-hop 3-cycles (three particles on a 4-cycle, the empty corner circulating once) in the 2 + 1 + 1 2+1+1 2 + 1 + 1 class, which is where the missing weight lies. The local values show it directly (n = 3 n=3 n = 3 ): a third particle sitting on a same-spin pair’s mediator site gives a clean, spin-dependent value of ( H ∞ 4 ) a a (H_\infty^4)_{aa} ( H ∞ 4 ) aa (180 for a same-spin and 186 for an opposite-spin third particle on a mediator of a diagonal pair; 180 and 182 on a plaquette corner next to an edge pair, in units of λ 4 \lambda^4 λ 4 ), distinct from either the blocked or fully-open case, and N N N -independent (identical at L = 5 L=5 L = 5 and L = 6 L=6 L = 6 ). Summed over configurations with the spin-averaging weights of Eq. (10) , the 3-cycles give the second term of Eq. (13) , which completes α ( N , n ) \alpha(N,n) α ( N , n ) .
Schrieffer–Wolff operator construction
Write H = H t + U N ^ d H=H_t+U\hat{N}_d H = H t + U N ^ d . From the N d = 0 N_d=0 N d = 0 sector P 0 P_0 P 0 , define T 0 = P 0 H t P 0 T_0=P_0H_tP_0 T 0 = P 0 H t P 0 and T 1 = Q 1 H t P 0 T_1=Q_1H_tP_0 T 1 = Q 1 H t P 0 , where Q 1 Q_1 Q 1 projects onto the exactly-one-doublon sector. T 1 T_1 T 1 is built by, for each P 0 P_0 P 0 basis state ( u , d ) (u,d) ( u , d ) (upand down-tuples of occupied sites), enumerating hops of a spin-σ \sigma σ particle at site s s s onto a neighbor n b nb nb already occupied by the opposite spin: the resulting state has s s s vacated and n b nb nb doubly-occupied, landing in the Q 1 Q_1 Q 1 basis. Because up and down sectors are combined as a plain tensor product with no cross Jordan–Wigner string (the standard Hubbard-model convention), the fermion sign of such a hop is exactly the same sorted-tuple sign rule used for T 0 T_0 T 0 (Appendix E), applied to the hopping species’ own tuple, treating the other species as a spectator. T 0 Q = Q 1 H t Q 1 T_0^Q = Q_1 H_t Q_1 T 0 Q = Q 1 H t Q 1 is built the same way from each Q 1 Q_1 Q 1 basis state, keeping the hops whose result again has exactly one doubly occupied site (a singly occupied site’s electron moving into an empty site, or a doublon electron hopping onto an opposite-spin singly occupied site, which moves the doublon).
The Hermitian Schrieffer–Wolff expansion [24 , 25 ] with every virtual energy denominator equal to ( 0 − U ) − 1 = − 1 / U (0-U)^{-1}=-1/U ( 0 − U ) − 1 = − 1/ U (through third order Q 1 Q_1 Q 1 is the only virtual sector, since P 0 H t Q 2 = 0 P_0H_tQ_2=0 P 0 H t Q 2 = 0 and a path P 0 → Q 1 → Q 2 → P 0 P_0 \to Q_1 \to Q_2 \to P_0 P 0 → Q 1 → Q 2 → P 0 needs two hops back) gives Eq. (15) , with V ≡ T 1 † T 1 V \equiv T_1^\dagger T_1 V ≡ T 1 † T 1 and H ( 3 ) = T 1 † T 0 Q T 1 − 1 2 { V , T 0 } H^{(3)} = T_1^\dagger T_0^Q T_1 - \frac{1}{2}\{V,T_0\} H ( 3 ) = T 1 † T 0 Q T 1 − 2 1 { V , T 0 } . Expanding ⟨ ( T 0 + X ) 2 ⟩ − ⟨ T 0 + X ⟩ 2 = Var ( T 0 ) + 2 ⟨ T 0 , X ⟩ H S + Var ( X ) \langle(T_0+X)^2\rangle-\langle T_0+X\rangle^2=\operatorname{Var}(T_0)+2\langle T_0,X\rangle_{\mathrm{HS}}+\operatorname{Var}(X) ⟨( T 0 + X ) 2 ⟩ − ⟨ T 0 + X ⟩ 2 = Var ( T 0 ) + 2 ⟨ T 0 , X ⟩ HS + Var ( X ) (using ⟨ T 0 ⟩ = 0 \langle T_0\rangle=0 ⟨ T 0 ⟩ = 0 ) for X = − V / U + H ( 3 ) / U 2 X=-V/U+H^{(3)}/U^2 X = − V / U + H ( 3 ) / U 2 gives Eq. (16) , with Tr ( T 0 H ( 3 ) ) = Tr ( T 0 T 1 † T 0 Q T 1 ) − Tr ( T 0 2 V ) \operatorname{Tr}(T_0H^{(3)})=\operatorname{Tr}(T_0T_1^\dagger T_0^Q T_1)-\operatorname{Tr}(T_0^2V) Tr ( T 0 H ( 3 ) ) = Tr ( T 0 T 1 † T 0 Q T 1 ) − Tr ( T 0 2 V ) ; all traces are evaluated as sparse matrix products, in integers at λ = 1 \lambda=1 λ = 1 .
Connection to the standard t t t -J J J /three-site expansion. V = T 1 † T 1 V=T_1^\dagger T_1 V = T 1 † T 1 is exactly the standard second-order strong-coupling operator familiar from the canonical Hubbard-to-t t t -J J J transformation [24 , 25 ] : up to the overall 1 / U 1/U 1/ U and normalization, − V -V − V decomposes into (i) the Heisenberg superexchange term J ∑ ⟨ i j ⟩ ( S i ⋅ S j − 1 4 n i n j ) J\sum_{\langle ij\rangle}(\mathbf{S}_i\cdot\mathbf{S}_j-\frac{1}{4}n_i n_j) J ∑ ⟨ ij ⟩ ( S i ⋅ S j − 4 1 n i n j ) with J = 4 λ 2 / U J=4\lambda^2/U J = 4 λ 2 / U , itself splitting into a diagonal S i z S j z − 1 4 n i n j S_i^zS_j^z-\frac{1}{4}n_i n_j S i z S j z − 4 1 n i n j piece and an off-diagonal spin-flip piece that swaps spin labels between two already-occupied sites without changing which sites are occupied; and (ii) the correlated three-site hopping term, which moves one particle between two sites both adjacent to a third (mediating, necessarily doubly-visited) site, either keeping its spin or, in the spin-flip form, exchanging spin labels with the electron on the mediating site: in both forms a second-neighbor-type move on the site-occupation graph. Both pieces are standard ingredients of the t / U t/U t / U expansion; what is specific to the present calculation is what they imply for ⟨ T 0 , V ⟩ H S \langle T_0,V\rangle_{\mathrm{HS}} ⟨ T 0 , V ⟩ HS .
Locality argument for the vanishing cross term. T 0 T_0 T 0 ’s edges are single nearest-neighbor hops, changing the site-occupation pattern at exactly one site pair. The diagonal spin-exchange piece above returns to the same configuration, contributing nothing when multiplied by the purely off-diagonal T 0 T_0 T 0 and traced; the spin-flip piece changes no site occupation pattern at all (an edge type disjoint from T 0 T_0 T 0 ’s, since T 0 T_0 T 0 always changes the site pattern); and the three-site hop is a second-neighbor-type move, again a different edge type from T 0 T_0 T 0 ’s nearest-neighbor hops on a large-enough torus. All three pieces therefore contribute zero to ⟨ T 0 , V ⟩ H S \langle T_0,V\rangle_{\mathrm{HS}} ⟨ T 0 , V ⟩ HS unless two sites with a common mediating neighbor are themselves adjacent, that is, unless the graph has a triangle. The square-lattice torus has none for L ≥ 4 L\ge4 L ≥ 4 ; at L = 3 L=3 L = 3 periodic wraparound creates triangles, the origin of the nonzero cross term reported in §7. The machine-checked statement (Appendix F) is the vanishing of the trace on every triangle-free graph; this decomposition is not used there.
Numerical methods and validation
Fermion sign rule (S8). A k k k -fermion configuration is represented as a sorted tuple of occupied sites; moving the fermion at its rank-m m m position (0-indexed) to a new site a a a picks up sign ( − 1 ) m × ( − 1 ) p (-1)^m \times(-1)^p ( − 1 ) m × ( − 1 ) p , where p p p counts how many of the remaining occupied sites are less than a a a : the standard sorted-Slater-determinant sign rule for any k k k . Verified Hermitian for k = 2 k=2 k = 2 (direct anticommutation derivation); cross-validated against an independent Jordan–Wigner bit-string implementation on 4000 random moves, for k = 1 k = 1 k = 1 –8 8 8 on 12 sites (zero mismatches); and, for a single species with k = 4 k = 4 k = 4 at L = 4 L = 4 L = 4 , the moments μ 2 \mu_2 μ 2 , μ 4 \mu_4 μ 4 of the operator built with this rule agree exactly, in rational arithmetic, with the free-fermion values obtained by choosing 4 distinct single-particle energies and summing. In the Lean formalization (Appendix § 16 ) the rule is derived from creation and annihilation operators satisfying the canonical anticommutation relations, for an arbitrary ordering of the orbitals.
Sparse-trace evaluation. All U = ∞ U = \infty U = ∞ and Schrieffer–Wolff results (§ 5 –§ 7 ) are computed via exact sparse matrix products, never diagonalization: Tr ( H 2 ) \operatorname{Tr}(H^2) Tr ( H 2 ) from the diagonal of H@H, and Tr ( H 4 ) = ∑ a b ∣ ( H 2 ) a b ∣ 2 \operatorname{Tr}(H^4) = \sum_{ab} |(H^2)_{ab}|^2 Tr ( H 4 ) = ∑ ab ∣ ( H 2 ) ab ∣ 2 for Hermitian H H H . At zero flux the traces are integers at λ = 1 \lambda= 1 λ = 1 , so the moments are exact rationals; under flux they are evaluated in floating point, to about 10 − 12 10^{-12} 1 0 − 12 . The method is tractable for every sector reported here (up to D 0 ≈ 4 × 10 6 D_0 \approx4 \times10^6 D 0 ≈ 4 × 1 0 6 ).
Relative-coordinate scan (Figs. 2 and 9). At fixed pair momentum K K K , the two-electron problem reduces exactly to a single particle on the lattice of relative coordinates, with K K K -dependent hopping amplitudes along the two axes and an on-site defect U U U at the origin (the Koster–Slater problem of § 2 ). We take a relative lattice of L r e l = 21 L_{\mathrm{rel}} = 21 L rel = 21 sites per side and scan K K K over a 48 × 48 48 \times48 48 × 48 grid, which acts as a fine set of twisted boundary conditions for the centre-of-mass motion; at each K K K the U-band level is the eigenvalue split off above the continuum by the repulsive defect. Each relative-coordinate problem has only L r e l 2 L_{\mathrm{rel}}^2 L rel 2 states, so the scan is far cheaper than diagonalizing the full two-electron Hamiltonian at many twists.
Additional numerical validation figures. Three further figures close specific analytic claims made in the main text: the two-particle U-band benchmark (Fig. 9), the exchange contribution to μ 4 \mu_4 μ 4 under flux (Fig. 10), and the doublon’s doubled Peierls phase (Fig. 11), in addition to the density-of-states validation of Fig. 7 above.
Machine-checked results
The exact statements listed in Table 7 are formalized and proved in Lean 4 [21 ] with Mathlib [22 ] (Lean v4.35.0, Mathlib release v4.35.0). Each statement is written as a proposition in one file that uses only Mathlib and the definitions of the formalization (sector, torus, configuration, fermion operators, H t H_t H t , H ∞ H_\infty H ∞ , moments), so that what is proved can be read without reading the proofs. Every headline theorem depends only on Lean’s three standard axioms (propositional extensionality, choice, quotient soundness): no unproved step (sorry) and no trusted compiled evaluation enters. Negative controls (a wrong sector dimension, a wrong multiplicity in the finite-population μ 4 \mu_4 μ 4 , the single-phase doublon amplitude, and sign-free annihilation operators) were checked by hand to be rejected or refuted (recorded with the formalization). Independently, Lean’s comparator checks that every headline theorem proves exactly its stated proposition with only the three permitted axioms, re-checking the proofs with six kernels (Lean’s default and paranoid checkers, lean4lean, nanoda, con-leche and con-ron), and it rejects four automated negative controls (a weakened statement, an added axiom, a changed definition, an unproved step); the record of this run is kept with the formalization.
Table 7. Machine-checked statements and the Lean theorems that prove them.
Scope. The following are not formalized, and the corresponding statements of the paper rest on the derivations given in the text and the Supplemental Material: the Schrieffer–Wolff expansion itself, at second and third order (the formalization evaluates the second-order terms and the operators T 0 T_0 T 0 , V V V it produces), and hence the U − 2 U^{-2} U − 2 coefficient c 2 c_2 c 2 ; the identification of the U = 0 U = 0 U = 0 many-body
Figure 9. Two-particle benchmark: the exact two-electron U-band, computed by a direct relative-coordinate K-grid scan (L r e l = 21 L_{\mathrm{rel}} = 21 L rel = 21 , 48 × 48 48 \times48 48 × 48 K-points), against the parameter-free second-order prediction ρ 2 D ( E − ε d ; t d ) \rho_{2D}(E-\varepsilon_d;t_d) ρ 2 D ( E − ε d ; t d ) of Eq. (3) , at U = 10 , 15 , 20 , 50 λ U = 10, 15, 20, 50\lambda U = 10 , 15 , 20 , 50 λ : moderate coupling through the asymptotic doublon regime, no fitted parameters. This is the same benchmark underlying the collapse of Figure 2 , shown here at the level of the individual U U U values rather than the rescaled universal curve.
Figure 11. Doublon flux response (§8; energies in units of λ \lambda λ ): the exact flux-threaded two-electron U-band’s top eigenvalue against the effective doublon model with doubled Peierls phase t d e e 2 i A i j t_{\mathrm{de}}e^{2iA_{ij}} t de e 2 i A ij , compared with an undoubled control t d e e i A i j t_{\mathrm{de}}e^{iA_{ij}} t de e i A ij , N = 36 N = 36 N = 36 (L = 6 L = 6 L = 6 ), U = 40 λ U = 40\lambda U = 40 λ , no fitted parameters. This is a standard consequence of the effective doublon Hamiltonian’s charge-2e gauge coupling rather than a distinctive result of this paper on its own; it is included because it makes the bridge to the gauge-response discussion of §8 concrete at the two-particle level, not as evidence of collective or superconducting behavior.
energies with sums of the energies of distinct single-particle states (Slater determinants) and of the tight-binding band with plane waves; every fourth-moment coefficient, including α ( N , n ) \alpha(N,n) α ( N , n ) , α d ( N , 2 ) \alpha_d(N,2) α d ( N , 2 ) , the spin-averaging step and hence the exact linearity of μ 4 ( ∞ ) \mu_4(\infty) μ 4 ( ∞ ) in P s a m e P_{\mathrm{same}} P same , which require the spin conditions and fermion signs of four-hop walks; and the physical identification of the formalized fermion operators: they satisfy the canonical anticommutation relations for an arbitrary ordering of the orbitals, and any other convention differs from them by phases of the basis states, a standard fact not proved there. All spectral, density-of-states and finite-U U U numerics are independent of the formalization. The Lean text of every statement, with the definitions it uses, is reproduced verbatim in S12; the proofs are in the accompanying repository.
Reproduction. From the lean/ directory of the accompanying repository: lake exe cache get (prebuilt Mathlib), lake build (must finish without errors or warnings), and lake env lean AxiomAudit.lean (prints the axioms of every headline theorem); on Linux, lake comparator --paranoid and lake check run the independent checks above.
Code and data availability. All exact-diagonalization and sparse-trace codes used to produce the numerical results in this paper, covering the n = 3 , 4 n = 3, 4 n = 3 , 4 exact diagonalization, the general k k k -fermion sector construction, the U = ∞ U = \infty U = ∞ sparse-trace evaluation, the Schrieffer–Wolff operator construction and traces, and the flux-threaded Peierls-Hamiltonian validation of § 8 , are public, together with the scripts that regenerate every figure and table, the records (command, environment and output) of the long runs, and the Lean 4 formalization of Appendix § 16 (its lean/ directory), at https://github.com/zhineos/hubbard-spectral-moments, release v1.0.0 (commit 34715e1), archived at Zenodo as doi:10.5281/zenodo.24244237. Notes S1–S12 form the Supplemental Material.