WorksPhysics and Astronomy
Multi-reference variational states resolve stripe selection and local -wave pairing in the Hubbard model
Abstract
Whether stripes compete with -wave superconductivity or promote pairing through fluctuations remains unsettled in the two-dimensional Hubbard model. Using multi-reference neural variational auxiliary-field quantum Monte Carlo (MR-nVAFQMC) at , we superpose unpaired Slater determinants across stripe periods and domain-wall positions, reaching variational energies competitive with leading benchmarks on periodic lattices up to sites. For cuprate next-nearest-neighbor hopping (), diagonal hopping preserves electron-doped antiferromagnetism while kinetically frustrating hole motion into Yamada stripes with a dome of short-range connected -wave pairing. At hole doping, hole delocalization across antiphase spin boundaries outweighs on-site repulsion, splitting the filled stripes of the pure model into half-filled walls with strong local pairing. Coherently delocalizing shifted walls strongly suppresses static charge modulation yet leaves short-range pairing unchanged, showing that local pairing is intrinsic to half-filled walls rather than driven by stripe fluctuations. Across intervening antiferromagnetic domains, however, pairing correlations become unresolved at one stripe period and beyond. Both parameter transfer to larger lattices without reoptimization and deep benchmark checkpoints retain stripe spin order without resolving long-range pairing, within uncertainties allowing weak superconductivity. Half-filled antiphase walls thus generate local -wave pairs, whereas suppressing static stripe order alone yields no resolved coherence across stripes in the states studied.
Introduction
How unidirectional spin and charge stripes interact with -wave superconductivity has been a central question in the physics of high- cuprates for more than three decades [1, 2, 3]. In LaSrCuO, neutron scattering reveals incommensurate magnetic peaks displaced from by , which grows approximately linearly with hole concentration (, the Yamada relation) up to before saturating at higher doping [4]. The suppression of bulk superconductivity near doping in LaBaCuO is the longstanding anomaly [5]. In LaBaCuO, static period-four charge order () and period-eight antiphase spin order () coexist with a strongly depressed bulk three-dimensional (3D) transition at K, whereas transport and magnetization measurements indicate two-dimensional (2D) superconducting fluctuations below about 40 K and a Berezinskii–Kosterlitz–Thouless transition near 16 K [6]. A proposed explanation for the suppressed interlayer coherence is an intra-plane pair-density wave (PDW), whose pairing phase alternates across adjacent stripes and whose stripe direction rotates by 90° between layers, frustrating interlayer Josephson coupling [2, 7]. Because unidirectional antiphase stripes of charge period and linear hole filling carry magnetic incommensurability , the observed slope points directly to half-filled domain walls () rather than full walls (, ). Whether a single-layer 2D Hubbard model at develops phase coherence across half-filled stripes—as uniform -wave or antiphase PDW order—or supports only short-range intra-wall pairing remains a central microscopic question [2, 8, 9, 10].
Whether the single-band 2D repulsive Hubbard model captures these intertwined phenomena has been investigated intensively with modern many-body methods [11, 12, 13, 14, 15, 16]. For the nearest-neighbor model () at and , combined density-matrix renormalization group (DMRG), auxiliary-field quantum Monte Carlo (AFQMC), density-matrix embedding, and tensor-network calculations established an insulating filled-stripe ground state (, ) without long-range superconductivity [17, 18, 19]. Negative next-nearest-neighbor hopping , relevant to cuprate Fermi surfaces and electron–hole asymmetry [20, 21, 22], favors shorter stripe wavelengths; at , infinite projected entangled-pair states find period-four stripes with suppressed superconductivity [23]. Pairing also depends sensitively on cylinder width: four-leg calculations find algebraically decaying pairing at finite [24], whereas eight-leg calculations find extremely short-ranged superconducting correlations for [25]. In a distinct setting, periodically modulating nearest-neighbor hopping enhances superconducting correlations on four- and six-leg cylinders [26]. Finite-temperature simulations reveal fluctuating stripes and intertwined spin, charge, and pair correlations [27, 28]. Complementary cold-atom experiments observe precursors of individual stripes in mixed-dimensional settings [29] and a pseudogap in a two-dimensional Fermi–Hubbard simulator [30].
Determining the 2D zero-field behavior at with has nonetheless proved delicate because distinct uniform, superconducting-stripe, and non-superconducting-stripe states compete on the milli- () energy scale per site and respond sensitively to boundary conditions, applied fields, and variational bias. Combining DMRG with constrained-path AFQMC under twist-averaged boundary conditions and applied pairing fields—taking the thermodynamic limit at fixed pairing field before extrapolating the field to zero—Xu et al. reported -wave superconductivity coexisting with partially filled stripes (–) on the hole-doped side and Néel antiferromagnetism on the electron-doped side [31]. Working directly at zero external field on periodic 2D tori with neural quantum states [32, 33, 34], Rende et al. reported finite extrapolated -wave pairing up to using symmetry-preserving BCS backflow [35]. Extending Pfaffian backflow [36, 37] and neural-network fermionic projected entangled pair states (NN-fPEPS) up to tori, Chen et al. reached lower variational energies and revealed a near-degeneracy between a nearly uniform -wave state and lower-energy striped states carrying very weak or statistically unresolved long-range superconductivity [10]. On , Viteritti et al. found that near-degenerate determinant and Pfaffian backflow states can exhibit different pair correlations, but that symmetry restoration and variance reduction bring the three wavefunction families toward a common picture of coexisting stripe and pairing correlations on this finite lattice [38]. Related neural backflow calculations show how symmetry projection and multi-determinant improvements change both energies and correlation functions [39].
Two methodological considerations motivate probing this competing regime with multi-reference unpaired auxiliary-field wavefunctions. First, a single-reference variational state—even after space-group projection—inherits a single reference stripe period and rigid inter-wall spacing, preventing inequivalent stripe periods and shifted domain-wall arrangements from competing or fluctuating coherently within one quantum state. Second, variational families differ in how pairing enters their reference states. Canonical number-projected BCS and Pfaffian wavefunctions conserve total particle number just as Slater determinants do, but embed explicit pairing fields into their reference geminals; by contrast, each individual unprojected particle-conserving Slater determinant carries no explicit pairing field and has vanishing connected two-body pairing (), whereas space-group projection, multi-determinant superposition, and the shared imaginary-time auxiliary-field projector generate non-Wick four-fermion correlations without geminal pairing initialization. Although a variational upper bound on total energy implies neither upper nor lower bounds on order parameters, reaching competitive energies from particle-conserving Slater pools provides a complementary test of which features of stripe order govern local pair formation and which limit long-distance coherence.
Here we develop multi-reference neural variational auxiliary-field quantum Monte Carlo (MR-nVAFQMC) to study the 2D square-lattice –– Hubbard Hamiltonian (Figure 1A) on periodic tori ( sites),
Figure 1. Square-lattice –– Hubbard model, Yamada relation, and real-space stripe correlations across doping. (A) Schematic of the 2D square –– Hubbard model (; top) and competing filled (, ) versus half-filled (, ) antiphase stripe orders at (bottom). (B) Spin incommensurability (top) and nearest-neighbor connected -wave pairing vertex (bottom) versus signed doping (: ; : ) on (circles), (squares), and at (diamonds: blue, ; vermilion, , coinciding with in the top panel). Top: open symbols denote spin peaks below ; the grey band guides the Yamada trend ( up to followed by an plateau across ), and the dashed blue line marks filled stripes (). (C–E) Intrinsic 1D carrier-density profiles [or in (C); top] and five-row 2D staggered spin-correlation maps around a reference site (open circle; bottom) for (C) electron-doped at (, commensurate AFM), (D) at , (filled stripe, , ), and (E) at , (half-filled stripe, , ). Error bars denote 1 s.e.
where and denote nearest- and next-nearest-neighbor pairs, , and we set with . Hole doping is studied in the spin-balanced canonical sector for both and cuprate hopping ; physical electron doping ( at ) maps via the bipartite particle–hole transformation to filling at reversed diagonal hopping , displayed at signed doping .
Results
Doping asymmetry, the Yamada relation, and kinetic stabilization of half-filled stripes
Figure 1B–E traces how magnetic order and local -wave pairing evolve across electron and hole doping in the absence of external pinning fields. On the electron-doped branch (, ), doped carriers move coherently on the same magnetic sublattice without nucleating antiphase domain walls: the staggered spin correlation remains uniformly positive across the torus at (Figure 1C), keeping the magnetic Bragg peak locked at the commensurate Néel wavevector (; Figure 1B, top), while no local connected -wave pairing dome develops (Figure 1B, bottom). On the hole-doped branch (, ), by contrast, the magnetic peak splits into incommensurate pairs whose shift follows the half-filled Yamada relation across commensurate lattices at (), (), and (). Above on , the magnetic wavevector saturates at the plateau at —raising the linear wall filling from to while the stripe spin peak weakens nearly threefold—before departing from the plateau to a weaker
peak at heavy overdoping (), where static charge modulation falls below . Across the entire hole-doped range , the nearest-neighbor connected -wave pairing vertex (Section III) forms a broad positive dome, reaching its maximum around where period-eight spin stripe order begins to soften (Figure 1B, bottom).
To identify how negative selects half-filled Yamada stripes over the filled stripes of the pure Hubbard model, we compare candidate charge periods at in Figure 2. On the four-leg torus, whose length accommodates both and at deep projection (), the filled stripe () is favored at whereas introducing reverses the energetic ordering (Figure 2A): the half-filled manifold () becomes the clear variational minimum among all single-period sectors, lying per site below and well below , , and the uniform () state. When all candidate periods compete simultaneously inside unbiased and multi-period pools (Figure 2A, shaded column), the optimized wavefunctions spontaneously select the () sector and match or slightly improve upon the independently optimized energy.
Figure 2. Stripe wavelength switch and kinetic stabilization at . (A) Ground-state energy per site on across charge periods for (blue circles) and (vermilion squares, ), alongside unbiased RR-CV and multi-period pools (shaded column). (B) Longitudinal spin structure factor on at , shifting from (, ) to (, ). (C) Kinetic (), interaction (), and total () energy changes on (, ) from to single-sector control (left), to multi-sector (middle), and pinned to multi-sector at fixed (right). Error bars denote 1 s.e.
The same period-halving transition occurs on the isotropic 256-site torus at : the Bragg peak of shifts from (, ) at to (, ) at (Figure 2B), replacing two broad filled walls with four half-filled antiphase walls in real space (Figure 1D,E) and increasing the intra-wall nearest-neighbor pairing vertex more than fivefold. Decomposing the total energy into hopping and on-site repulsion, with double occupancy (Figure 2C), reveals why half-filled stripes win: splitting filled walls into twice as many half-filled walls increases local double occupancy and raises the Hubbard interaction energy (), yet this repulsion penalty is outweighed more than twofold by the gain in transverse kinetic energy () as holes hop across closer antiphase domain boundaries without leaving extended strings of overturned spins. Sublattice kinetic frustration thus unifies both sides of the phase diagram: positive effective on the electron-doped side assists intrasublattice motion within a uniform Néel background, whereas negative frustrates intrasublattice hopping on the hole-doped side and stabilizes half-filled antiphase walls.
Controlled stripe superpositions: pinned and delocalized domain walls pair alike
Testing whether transverse stripe fluctuations amplify -wave pairing requires separating three situations that are easily conflated: (i) a pinned stripe, whose domain walls sit at fixed lattice columns with static hole-density modulation ; (ii) an incoherent translation mixture of shifted copies of that pinned stripe, which sets the one-body modulation to and transfers its Bragg weight into the connected density structure factor , while leaving every linear translation-averaged observable—total energy , full , bare , and bare -wave Bloch eigenvalues—identically unchanged [whereas Wick-subtracted shifts slightly only because is translation-averaged before squaring]; and (iii) a coherent quantum superposition of shifted stripe sectors, which differs from the classical mixture solely through off-diagonal quantum interference between distinct domain-wall positions.
Figure 3 carries out this controlled comparison on at fixed Hamiltonian and projection depth (, , ), contrasting a pinned single-reference stripe (), a single-sector control localized in one wall sector, and delocalized multi-sector and superpositions distributed across translated wall positions. Relative to the matched single-sector control, coherent multi-sector mixing suppresses the intrinsic-frame period-four modulation amplitude roughly threefold while shifting weight into the connected charge peak at conserved total (Figure 3A), and re-optimizing the projector under coherent shifted-stripe superposition lowers the variational energy by per site below the pinned state.
Figure 3. Multi-reference stripe unpinning and local -wave pairing at , . (A) Intrinsic 2D hole-density maps comparing the single-sector control and fluctuating multi-sector state on (; top), together with 1D hole profiles and period-four amplitudes across pinned, control, fluctuating, and fluctuating states (bottom). (B) Nearest-neighbor connected -wave vertex along the walls [], across the walls [], and scaled cumulative short-range sum across the unpinning series, showing surviving local -wave pairing without finite-range enhancement. Error bars denote 1 s.e.
Across the entire unpinning series, short-range connected -wave pairing remains robust and essentially unchanged (Figure 3B). Subtracting the normal Wick contraction from yields a positive nearest-neighbor connected vertex along the walls in all four states; by contrast, matched-depth uniform paramagnetic and filled reference controls at identical Hamiltonian and projector architecture collapse both variational energy and local toward zero, confirming that the half-filled antiphase wall geometry itself generates the local pairing vertex. While expanding from to restores local bond isotropy between and , this restoration is already present in the pinned single-sector control; comparing the delocalized superposition directly against its matched single-sector control reveals no statistically resolved change in either the cumulative short-range vertex or its strictly non-overlapping bond subset where three-site contact terms vanish identically. At inter-stripe separations () and across distant bond pairs, connected pair correlations remain unresolved within uncertainty, and split-block cross-validated Rayleigh quotients of the full bond-singlet pair matrix on 64-site ( and ) tori remain non-extensive (–) with comparable connected weight in uniform and period-eight modes. Local -wave pairing is therefore already an intrinsic property of an individual half-filled domain wall, whereas unpinning the charge profile within our multi-reference subspace neither amplifies local pairing nor induces resolved inter-stripe coherence.
Ground-state energy benchmarks on large two-dimensional square tori
Figure 4 and Table I benchmark the ground-state energies of our multi-reference MR-nVAFQMC wavefunctions at and across periodic (256-site) and (1,024-site) square tori. Throughout, we separate fixed-checkpoint variational upper bounds from empirical Sorella projection-depth extrapolations [19] (marked “not variational” in Table I and shown as open stars in Figure 4), alongside both variational and extrapolated literature benchmarks. For the periodic VAFQMC benchmark, is the Sorella–Tocchio result reported in the V-score dataset of Wu et al. [34] and reproduced in Table S4 of Chen et al. [10]; Ref. [19] describes the underlying method and its twist-averaged TDL extrapolation.
Figure 4. Ground-state energy benchmarks across large periodic 2D square tori at (). Filled or solid-outlined symbols denote strictly variational energies; open symbols denote empirical or extrapolations. Because Monte Carlo standard errors span two orders of magnitude across sign-free VMC and finite-sign VAFQMC, error bars are omitted from the plot and listed in Table I. (A) Energy per site on 16 × 16 and 32 × 32 tori for the nearest-neighbor Hubbard model (), comparing HFPS [36], Transformer-BF [40], SCALE and ACE-BF [41], 16 × 16 PBC variational VAFQMC [10, 19, 34] and twist-averaged thermodynamic-limit (TDL) Sorella VAFQMC [19], variational () and extrapolated fPEPS, NN-fPEPS (), BCS backflow, and Pfaffian backflow [10] with our multi-reference () variational curve and same-manifold extrapolations ( on 16 × 16, on 32 × 32; open vermilion stars). (B) Energy per site for the cuprate Hubbard model () across 16 × 16, 24 × 24, and 32 × 32, comparing HFPS [36], SCALE and ACE-BF [41], symmetry-preserving BCS backflow [35], and NN-fPEPS () and Pfaffian backflow [10] with our multi-reference variational benchmarks ( on 16 × 16, on 32 × 32) and same-manifold extrapolations ( on 16 × 16, on 32 × 32; open vermilion stars).
Across both and tori, MR-nVAFQMC reaches variational energies competitive with the most accurate neural- and tensor-network simulations reported to date (Figure 4 and Table I). In the nearest-neighbor model (; Figure 4A), our states ( on , on ) match or lie slightly below BCS backflow, Sorella VAFQMC, Pfaffian backflow, ACE backflow, and NN-FPEPS within statistical uncertainty while lying well below unprojected SCALE and variational fPEPS [10, 19, 34, 41]; extrapolating in on () further agrees closely with both extrapolated fPEPS () [10] and Sorella’s twist-averaged 2D thermodynamic-limit estimate () [19]. In the frustrated cuprate model (; Figure 4B), our () checkpoint on 16 × 16 and () checkpoint on 32 × 32 likewise match or lie slightly below Pfaffian backflow, symmetry-preserving BCS backflow, and NN-fPEPS within statistical error [10, 35], demonstrating that particle-conserving multi-determinant neural auxiliary-field states access the same milli- low-energy regime without geminal pairing initialization.
Directional pairing decay and exploratory transferred-parameter scaling
We next examine how connected -wave pairing correlations decay spatially on 16 × 16 and how spin and pairing observables behave in an exploratory fixed-depth series up to 32 × 32 (Figure 5). Along the stripe direction on 16 × 16 (; Figure 5A, top), the bare two-bond correlator at short distance is dominated by the disconnected one-body Wick contribution , underscoring the need to subtract . The resulting connected intra-stripe vertex is clearly positive at nearest-neighbor separation () across , but drops sharply already by —where , 4 lie near zero within and fluctuates slightly negative within uncertainty—and remains statistically indistinguishable from zero over –8 (Figure 5A, inset). Perpendicular to the stripes (Figure 5A, bottom), decays even more rapidly, becoming indistinguishable from zero within a single charge-stripe period () and beyond. Deepening projection on 16 × 16 up to –1.52 closes the energy offset relative to Pfaffian backflow while leaving both stripe spin order and distant-window pairing unchanged within , and doping scans into and on 12 × 12 similarly enhance quasi-1D along-stripe while transverse and distant-window pairing remain unresolved.
Figure 5. Directional decay of connected -wave pairing on and exploratory transferred-parameter scaling up to at . (A) Directional 1D two-bond cuts on (). Top: along-stripe bare (open diamonds), Wick background (grey squares) for (), and connected vertex for (inset: –). Bottom: across-stripe ( shaded). (B,C) Exploratory fixed-depth transferred-parameter scaling () at (B) (, ; ) and (C) (, ; ), comparing full-SU(2) stripe spin order (top) and the exact four-bond site-centered singlet distant-window pairing Eq. (5), matching Ref. [10]; bottom of This work (transf.) (vermilion circles and dashed guides; parameters transferred from without per-size re-optimization) together with unified same-checkpoint Table 1 SOTA evaluations on (, ) and (, ) (purple diamonds in C) against NN-fPEPS, Pfaffian backflow, and BCS backflow [10, 35] as well as pinned-field CP-AFQMC [31] (open star). Error bars denote 1 s.e. [omitted for Ref. [10] points whose uncertainties lie below the quoted plot-unit threshold; the Ref. [31] star shows the asymmetric interval propagated from ].
To compare distant-window pairing directly with Chen et al. [10] across commensurate square tori up to 32 × 32, Figure 5B,C evaluates the full-SU(2) four-peak stripe spin order together with the exact four-bond site-centered singlet pairing order parameter matching (2) of Ref. [10] (Section III). Two caveats govern the interpretation of the common-depth multi-size series and motivate labeling Figure 5B,C as exploratory. First, the fixed-depth series ( at ; at ) uses shallower wavefunctions at higher variational energies than the deeper – energy-benchmark checkpoints of Table I [for instance, on at , the evaluation gives compared with −0.7476(37) for the checkpoint in Table I]. Second, the and wavefunctions in the common-depth series are constructed by tiling and transferring parameters optimized on without independent variational re-optimization on the larger tori (steps = 0 on ), and under () these large-torus states remain frame-pinned (). To remove checkpoint mismatch directly at and , we additionally performed unified same-checkpoint joint evaluations of energy, , and four-bond on the exact Table I SOTA wavefunctions (, , ; , , ; purple diamonds in Figure 5C).
Within both the common-depth transferred series and the unified Table I SOTA checkpoints, retains pronounced period-16 (; Figure 5B, top) and period-8 (; Figure 5C, top) stripe spin order up to , comparable in magnitude to NN-fPEPS and Pfaffian backflow [10], whereas -symmetric BCS backflow suppresses stripe magnetism toward a nearly uniform state [10, 35]. In the four-bond pairing channel (Figure 5B, bottom), is compatible with zero at every finite size in the transferred series [giving a three-point linear intercept of at ] and remains statistically indistinguishable from zero on the exact Table I SOTA checkpoints on and (purple diamonds); likewise, the maximum site-centered Bloch eigenvalue density across all momenta —including uniform and PDW , —scales downward as from to . By comparison, while -symmetric BCS backflow extrapolates to a sizable uniform order parameter [10, 35], the lower-energy striped wavefunctions of Chen et al. [10] retain positive finite-size pairing on – that extrapolates to very weak or fit-window-sensitive superconductivity in the thermodynamic limit [ for NN-fPEPS, and from versus a bound from for Pfaffian backflow]. Because corresponds to , comparable to our statistical uncertainty, our results neither resolve long-range pairing statistically nor exclude weak superconductivity at the NN-fPEPS scale; establishing fully converged ground-state scaling across all sizes will still require independent per-size variational re-optimization at matched deep across .
Method
Multi-reference neural VAFQMC (MR-nVAFQMC) wavefunction
Building on the general strategy of superposing nonorthogonal, symmetry-projected determinants [42], we construct the multi-reference, symmetry-restored neural VAFQMC (MR-nVAFQMC) wavefunction
where is a pool of inequivalent canonical particle-conserving Slater determinants spanning candidate stripe periods and shifted domain-wall configurations, projects onto a discrete lattice and spin-flip symmetry group of order , and is a shared parameterized auxiliary-field correlation projector of imaginary-time depth , extending single-reference VAFQMC [19] to coherent multi-determinant superpositions. Applying symmetric Suzuki–Trotter factorization and discrete Hubbard–Stratonovich decoupling across slices writes , where denotes the auxiliary Ising fields and is a product of spin-factored one-body exponentials. Beyond tunable Trotter slice weights, we incorporate a neural network in the auxiliary-field sector that takes the auxiliary-field configuration as input and modulates the auxiliary-field path weights and effective one-body slice couplings shared across all reference determinants, increasing expressive flexibility while keeping every field configuration an exact one-body Slater-determinant propagation. On 64-site (, ) and tori we also employ stripe or full space-group projections (), whereas for the (, ) and multi-reference benchmarks in Table 1 we use the eight-element group () generated by lattice reflections and global spin flip , because summing many off-diagonal translations at large increases computational cost and sign fluctuations while the sub-extensive projection gain becomes small on 256–1,024-site tori where shifted stripe sectors are represented through the multi-reference pool and translation-averaged observables. Whereas each individual unprojected Slater determinant carries no explicit BCS or Pfaffian pairing field and has zero connected two-body pairing (), symmetry restoration , multi-determinant interference (), and the neural auxiliary-field projector together generate non-Gaussian spin, charge, and pairing correlations without geminal pairing initialization. Superposition weights (), reference orbitals, and projector parameters are optimized jointly by damped stochastic reconfiguration, followed by frozen-checkpoint sign-reweighted Markov-chain evaluation [19] with statistical uncertainties estimated by coarse-grained block jackknife over independent Markov chains after thermalization burn-in.
| Lattice, | Method | () |
| 16 × 16, 0 | MR-nVAFQMC (this work) (, , extrap., not variational) | −0.7604(32) |
| fPEPS, extrap. (Chen 2026, not variational) | −0.7602(14) | |
| MR-nVAFQMC (this work) (, , ) | −0.7589(19) | |
| BCS backflow (Chen 2026) | −0.75878(2) | |
| VAFQMC (Sorella 2023; via Wu 2024, Chen 2026) | −0.75865(3) | |
| NN-fPEPS, (Chen 2026) | −0.75851(2) | |
| 16 × 16, −0.2 | nVAFQMC (this work) (, , extrap., not variational) | −0.7482(19) |
| MR-nVAFQMC (this work) (, , ) | −0.7466(38) | |
| Pfaffian backflow (Chen 2026) | −0.74533(1) | |
| NN-fPEPS, (Chen 2026) | −0.74467(2) | |
| SBP [35] | −0.74411(1) | |
| 32 × 32, 0 | nVAFQMC (this work) (, , extrap., not variational) | −0.7581(17) |
| MR-nVAFQMC (this work) (, , ) | −0.7577(26) | |
| NN-fPEPS, (Chen 2026) | −0.75677(1) | |
| BCS backflow (Chen 2026) | −0.75617(1) | |
| SCALE det. backflow [40] [41] | −0.7501 | |
| 32 × 32, −0.2 | MR-nVAFQMC (this work) (, , extrap., not variational) | −0.7494(35) |
| MR-nVAFQMC (this work) (, , ) | −0.7476(37) | |
| Pfaffian backflow (Chen 2026) | −0.74569(1) | |
| NN-fPEPS, (Chen 2026) | −0.74257(2) |
Table 1. TABLE I. Best MR-nVAFQMC ground-state energies per site (this work) on large 2D square tori at () compared with leading published benchmarks. For each periodic torus and , our best multi-reference strictly variational state and same-manifold extrapolation (marked “not variational”) plotted in Figure 4 are listed alongside the lowest published periodic-torus benchmarks [10, 19, 34, 35, 41] [for reference, Sorella’s twist-averaged 2D thermodynamic-limit extrapolation at is −0.7605(1) [19]]. Numbers in parentheses denote 1 s.e. statistical uncertainty in the last displayed digit(s) [e.g., −0.7589(19) denotes −0.7589 ± 0.0019]; denotes projection onto lattice reflections (, ) and global spin flip (), while additionally include commensurate stripe translations (Section III).
Spin, charge, and connected pairing observables
From the equal-time spin and density correlators and (), longitudinal spin structure factors and the full-SU(2) four-peak stripe magnetic order parameter (matching Ref. [10]) are obtained from
where and with spin incommensurability . Static 1D hole-density modulation at commensurate stripe period (wavevector , e.g., at ) and translation-invariant connected charge fluctuations are quantified by
satisfying identically in the canonical ensemble.
With normalized nearest-neighbor singlet creation operator and bond signs , , we distinguish two standard -wave pair operators: the two-bond link-origin operator used for short-range vertices and 1D directional cuts (Figure 1B, Figure 3B, and Figure 5A), and the four-bond site-centered operator matching Eq. (2) of Ref. [10] used for distant-window comparisons in Figure 5B,C,
From the bare pair correlator built from bond-singlet amplitudes [, ], subtracting the disconnected normal Wick contraction built from the full state’s symmetrized equal-time one-body Green’s function ,
isolates genuine connected two-body pairing in either convention [10, 43] [used throughout Figure 3 and Figure 5; the broad doping scan in Figure 1B uses a uniform archive transition- subtraction across all points that agrees within statistical uncertainty with Eq. (6), e.g., versus on at ]. From the two-bond vertex we form the cumulative short-range sum , while from the four-bond vertex we evaluate the 2D distant-window average matching Ref. [10],
over the , 355, and 623 periodic minimum-image displacements outside on , 24, and 32. On 64-site tori, we also construct the full bond-singlet matrix [44] and suppress positive noisy-matrix diagonalization bias via the 50/50 split-block cross-validated Rayleigh quotient , which provides a conservative (downward-biased in expectation relative to , up to finite-sample ratio bias) estimate of the leading pair-matrix weight.
Discussion and Conclusion
Using MR-nVAFQMC wavefunctions built from particle-conserving Slater determinants without explicit pairing fields—at variational energies matching or improving upon leading neural- and tensor-network benchmarks up to sites—our single-layer calculations establish a clear separation between intra-stripe local pair formation and inter-stripe phase coherence in the 2D –– Hubbard model. Negative next-nearest-neighbor hopping switches the state from filled () to half-filled (, ) antiphase stripes obeying the Yamada relation , stabilized by transverse kinetic delocalization against an opposing Hubbard repulsion penalty. Crucially, each half-filled domain wall already hosts robust short-range -wave singlet pairing: the connected intra-wall vertex is enhanced more than fivefold over the filled stripes at , absent on the electron-doped Néel branch, peaking on the overdoped side near where period-eight spin stripe order softens, and unchanged within uncertainty at whether the walls are pinned or coherently delocalized across translated stripe sectors so that static charge modulation drops threefold.
Comparing these strictly single-layer results with the experimental anomaly in LaBaCuO highlights both an encouraging microscopic agreement and an instructive open tension [2, 5, 6, 7]. Experimentally, bulk 3D superconductivity at occurs at only K [6], whereas each individual CuO layer exhibits 2D superconducting fluctuations setting in near the spin-stripe ordering scale ( K) and reaches a 2D zero-resistance state near 16 K; the suppression of 3D arises from interlayer Josephson decoupling, widely attributed to an intra-plane period-eight PDW that alternates sign between adjacent stripes and rotates by from layer to layer [2, 6, 7]. On the local scale, our finding that half-filled domain walls intrinsically form short-range -wave pairs—with uniform and period-eight PDW modes nearly degenerate in 64-site pair-matrix spectra because inter-wall Josephson coupling across the antiphase spin domain is very weak—captures the microscopic origin of intra-wall pairing and near-degenerate 0- versus -phase inter-stripe channels. On the macroscopic 2D scale, however, our single-layer states show transverse pair correlations decaying within one stripe spacing () and resolve no long-range 2D superconducting order up to , in contrast to the 2D planar superconductivity observed below 16 K in LaBaCuO [6] and consistent with the weak or statistically unresolved 2D pairing in lower-energy striped Pfaffian and NN-fPEPS wavefunctions [10].
Three concrete possibilities can reconcile this tension and define a clear roadmap for future work. First, a weak 2D intra-plane condensate at the scale corresponding to —such as the small amplitude () suggested by NN-fPEPS [10]—lies at the statistical noise floor of our present distant-window estimator, and the spatially summed cancels any sign-alternating PDW component. Second, although unified same-checkpoint evaluations on the lowest-energy Table I wavefunctions at and confirm strong stripe spin order with unresolved four-bond , the common-depth multi-size series in Figure 5B,C still relies on transferred parameters on and under frame-pinned symmetry (). Independent per-size variational re-optimization at matched deep , combined with direct superfluid-stiffness and translation-restored multi-sector scaling [45, 46], will test whether full larger-torus relaxation resolves weak 2D uniform or PDW coherence. Third, if converged ground states of the pure single-band 2D –– model at remain devoid of long-range 2D order, then stabilizing planar coherence at commensurate lock-in requires ingredients beyond the minimal single-band Hubbard Hamiltonian—such as multi-orbital CuO physics, longer-range interactions, or low-temperature-tetragonal lattice couplings [2, 3, 7]—or doping away from lock-in into the regime where spin stripes soften.
References
- [1]J. M. Tranquada, B. J. Sternlieb, J. D. Axe, Y. Nakamura, S. Uchida, Evidence for stripe correlations of spins and holes in copper oxide superconductors. Nature 375, 561–563 (1995).DOI
- [2]E. Fradkin, S. A. Kivelson, J. M. Tranquada, Colloquium: Theory of intertwined orders in high temperature superconductors. Rev. Mod. Phys. 87, 457–482 (2015).DOI
- [3]B. Keimer, S. A. Kivelson, M. R. Norman, S. Uchida, J. Zaanen, From quantum matter to high-temperature superconductivity in copper oxides. Nature 518, 179–186 (2015).DOI
- [4]K. Yamada et al., Doping dependence of the spatially modulated dynamical spin correlations and the superconducting-transition temperature in La₂−xSrxCuO₄. Phys. Rev. B 57, 6165–6172 (1998).DOI
- [5]A. R. Moodenbaugh, Y. Xu, M. Suenaga, T. J. Folkerts, R. N. Shelton, Superconducting properties of La₂−xBaxCuO₄. Phys. Rev. B 38, 4596–4600 (1988).
- [6]Q. Li, M. Hücker, G. D. Gu, A. M. Tsvelik, J. M. Tranquada, Two-dimensional superconducting fluctuations in stripe-ordered La₁.₈₇₅Ba₀.₁₂₅CuO₄. Phys. Rev. Lett. 99, 067001 (2007).pubmed.ncbi.nlm.nih.gov/17930855
- [7]E. Berg, E. Fradkin, E.-A. Kim, S. A. Kivelson, V. Oganesyan, J. M. Tranquada, S. C. Zhang, Dynamical layer decoupling in a stripe-ordered high-Tc superconductor. Phys. Rev. Lett. 99, 127003 (2007).
- [8]V. J. Emery, S. A. Kivelson, O. Zachar, Spin-gap proximity effect mechanism of high-temperature superconductivity. Phys. Rev. B 56, 6120–6147 (1997).DOI
- [9]S. A. Kivelson, E. Fradkin, V. J. Emery, Electronic liquid-crystal phases of a doped Mott insulator. Nature 393, 550–553 (1998).DOI
- [10]A. Chen, J. Gray, S.-J. Du, G. K.-L. Chan, Extreme competition between stripes and superconductivity in the hole-doped Hubbard model. arXiv:2610.30321v2 (2026).arxiv.org/abs/2610.30321
- [11]P. W. Anderson, The resonating valence bond state in La₂CuO₄ and superconductivity. Science 235, 1196–1198 (1987).pubmed.ncbi.nlm.nih.gov/17818979
- [12]F. C. Zhang, T. M. Rice, Effective Hamiltonian for the superconducting Cu oxides. Phys. Rev. B 37, 3759–3761 (1988).pubmed.ncbi.nlm.nih.gov/9944993
- [13]D. J. Scalapino, A common thread: The pairing interaction for unconventional superconductors. Rev. Mod. Phys. 84, 1383–1417 (2012).DOI
- [14]D. P. Arovas, E. Berg, S. A. Kivelson, S. Raghu, The Hubbard model. Annu. Rev. Condens. Matter Phys. 13, 239–274 (2022).
- [15]M. Qin, T. Schäfer, S. Andergassen, P. Corboz, E. Gull, The Hubbard model: A computational perspective. Annu. Rev. Condens. Matter Phys. 13, 275–302 (2022).DOI
- [16]J. P. F. LeBlanc et al. (Simons Collaboration on the Many-Electron Problem), Solutions of the two-dimensional Hubbard model: Benchmarks and results from a wide range of numerical algorithms. Phys. Rev. X 5, 041041 (2015).
- [17]B.-X. Zheng et al., Stripe order in the underdoped region of the two-dimensional Hubbard model. Science 358, 1155–1160 (2017).DOI
- [18]M. Qin, C.-M. Chung, H. Shi, E. Vitali, C. Hubig, U. Schollwöck, S. R. White, S. Zhang, Absence of superconductivity in the pure two-dimensional Hubbard model. Phys. Rev. X 10, 031016 (2020).DOI
- [19]S. Sorella, Systematically improvable mean-field variational ansatz for strongly correlated systems: Application to the Hubbard model. Phys. Rev. B 107, 115133 (2023).DOI
- [20]O. K. Andersen, A. I. Liechtenstein, O. Jepsen, F. Paulsen, LDA energy bands, low-energy Hamiltonians, t′, t″, t⊥(k), and J⊥. J. Phys. Chem. Solids 56, 1573–1591 (1995).DOI
- [21]E. Pavarini, I. Dasgupta, T. Saha-Dasgupta, O. Jepsen, O. K. Andersen, Band-structure trend in hole-doped cuprates and correlation with T_c^max. Phys. Rev. Lett. 87, 047003 (2001).pubmed.ncbi.nlm.nih.gov/11461638
- [22]M. Hirayama, Y. Yamaji, T. Misawa, M. Imada, Ab initio effective Hamiltonians for cuprate superconductors. Phys. Rev. B 98, 134501 (2018).DOI
- [23]B. Ponsioen, S. S. Chung, P. Corboz, Period 4 stripe in the extended two-dimensional Hubbard model. Phys. Rev. B 100, 195141 (2019).DOI
- [24]H.-C. Jiang, T. P. Devereaux, Superconductivity in the doped Hubbard model and its interplay with next-nearest hopping t′. Science 365, 1424–1428 (2019).DOI
- [25]H.-C. Jiang, T. P. Devereaux, S. A. Kivelson, Competition between charge-density-wave and superconducting orders on eight-leg square Hubbard cylinders. npj Quantum Mater. 11, 80 (2026).DOI
- [26]H.-C. Jiang, S. A. Kivelson, Stripe order enhanced superconductivity in the Hubbard model. Proc. Natl. Acad. Sci. U.S.A. 119, e2109406119 (2022).DOI
- [27]E. W. Huang, C. B. Mendl, S. Liu, S. Johnston, H.-C. Jiang, B. Moritz, T. P. Devereaux, Numerical evidence of fluctuating stripes in the normal state of high-Tc cuprate superconductors. Science 358, 1161–1164 (2017).DOI
- [28]P. Mai, S. Karakuzu, G. Balduzzi, S. Johnston, T. A. Maier, Intertwined spin, charge and pair correlations in the two-dimensional Hubbard model in the thermodynamic limit. Proc. Natl. Acad. Sci. U.S.A. 119, e2112806119 (2022).DOI
- [29]D. Bourgund et al., Formation of individual stripes in a mixed-dimensional cold-atom Fermi–Hubbard system. Nature 637, 57–62 (2025).DOI
- [30]L. H. Kendrick et al., Pseudogap in a Fermi–Hubbard quantum simulator. Nature 656, 602–608 (2026).DOI
- [31]H. Xu, C.-M. Chung, M. Qin, U. Schollwöck, S. R. White, S. Zhang, Coexistence of superconductivity with partially filled stripes in the Hubbard model. Science 384, eadh7691 (2024).DOI
- [32]G. Carleo, M. Troyer, Solving the quantum many-body problem with artificial neural networks. Science 355, 602–606 (2017).DOI
- [33]D. Luo, B. K. Clark, Backflow transformations via neural networks for quantum many-body wave functions. Phys. Rev. Lett. 122, 226401 (2019).DOI
- [34]D. Wu et al., Variational benchmarks for quantum many-body problems. Science 386, 296–301 (2024).DOI
- [35]R. Rende, L. L. Viteritti, A. Georges, Superconductivity in the t-t′ Hubbard model from symmetry-preserving neural-network quantum states. arXiv:2608.12465 (2026).arxiv.org/abs/2608.12465
- [36]C. Roth, A. Chen, A. Sengupta, A. Georges, Superconductivity in the two-dimensional Hubbard model revealed by neural quantum states. arXiv:2511.07566 (2025).arxiv.org/abs/2511.07566
- [37]A. Chen, Z.-Q. Wan, A. Sengupta, A. Georges, C. Roth, Neural network-augmented Pfaffian wave-functions for scalable simulations of interacting fermions. Proc. Natl. Acad. Sci. U.S.A. 123, e2535288123 (2026).DOI
- [38]L. L. Viteritti, R. Rende, C. Roth, A. Sengupta, G. Carleo, A. Georges, Beyond variational bias: Resolving intertwined orders in the Hubbard model. arXiv:2604.21978 (2026).arxiv.org/abs/2604.21978
- [39]K. Loehr, B. K. Clark, Enhancing neural network backflow. arXiv:2510.26906 (2025).arxiv.org/abs/2510.26906
- [40]Y. Gu et al., Solving the Hubbard model with neural quantum states. Nat. Commun. 17, 7838 (2026).
- [41]Y. Gu, Z. Han, W. Li, Z. Xiao, T. Xiang, M. Qin, L. Wang, D. Lv, Pareto frontier of neural quantum states: Scalable, affordable, and accurate convolutional backflow for strongly correlated lattice fermions. PRX Intelligence 1, 023001 (2026).arxiv.org/abs/2604.25775
- [42]C. A. Jiménez-Hoyos, R. Rodríguez-Guzmán, G. E. Scuseria, Multi-component symmetry-projected approach for molecular ground state correlations. J. Chem. Phys. 139, 204102 (2013).
- [43]S. Zhang, J. Carlson, J. E. Gubernatis, Pairing correlations in the two-dimensional Hubbard model. Phys. Rev. Lett. 78, 4486–4489 (1997).DOI
- [44]C. N. Yang, Concept of off-diagonal long-range order and the quantum phases of liquid He and of superconductors. Rev. Mod. Phys. 34, 694–704 (1962).DOI
- [45]W. Kohn, Theory of the insulating state. Phys. Rev. 133, A171–A181 (1964).DOI
- [46]D. J. Scalapino, S. R. White, S. Zhang, Insulator, metal, or superconductor: The criteria. Phys. Rev. B 47, 7995–8007 (1993).pubmed.ncbi.nlm.nih.gov/10004809




