betaFeedback

WorksPhysics and Astronomy

Quasi-long-range order in the quantum XY model on Sierpiński gaskets

Abstract

Fractal geometry has recently been shown to enhance the amplitude of the superconducting order parameter, but the fate of the superconducting phase on such lattices is far less clear. Classical XY models on finitely ramified fractals have no finite-temperature transition, as the energy of a vortex stays bounded as the system grows. However, that statement concerns an infinite gasket and thermal vortices, while real fractal structures are finite by construction. It also leaves open what happens at T=0T = 0, where the phase competes not with temperature but with its own quantum fluctuations. In this work we address whether superconducting phase coherence survives at T=0T=0. To that end, we study the spin-1/21/2 XY ferromagnet, the phase sector of a Josephson network with at most one Cooper pair per island, thus capturing the maximum effect of quantum phase fluctuations, on both triangular and honeycomb Sierpiński gaskets. By using vision-transformer neural quantum states optimized by variational Monte Carlo, we reach generations up to G=5G = 5, that is 366366 sites for the triangular gasket and 609609 for the honeycomb one, well beyond the reach of exact diagonalization. We find that the transverse two-point correlator decays algebraically, Cxy(r)∼r−ηC_{xy}(r) \sim r^{-η} with η≃0.12 ⁣− ⁣0.32η\simeq 0.12\!-\!0.32, independent of distance measured along the graph, along the fractal boundary, or in the embedding plane. Quantum phase fluctuations thus leave quasi-long-range order intact in the ground state. We further discuss the implication of these results for the superconducting response of a finite gasket at low but nonzero temperature.

Introduction

Quantum dynamics on artificial fractal lattices has recently become an experimentally accessible subject. Fractal arrays of artificial atoms have been assembled by positioning CO molecules on Cu(111) [1] or grown by molecular self-assembly [2], as well as realized as bismuth nanostructures hosting topological edge and corner states [3]. Mesoscopic superconducting realizations go back even further: Josephson junction arrays and superconducting wire networks patterned as Sierpiński gaskets have been measured since the late 1980s [4, 5, 6], and fractal Josephson weak links still attract attention [7].

Recent calculations have also sharpened the case for superconductivity in fractal systems. Mean-field Bogoliubov-de Gennes (BdG) treatment of the attractive Hubbard model on the Sierpiński gasket has been shown to give a superconducting critical temperature well above that of the regular triangular lattice, with a superfluid stiffness large enough to support macroscopic phase coherence [8, 9]. The Sierpiński carpet shows no such effect, a difference attributed to the ramification properties of the two fractals [8]. Furthermore, for order parameters of non-trivial symmetry, the fractal geometry reshuffles the competition between pairing channels [9], while related boundary-driven enhancements appear on hyperbolic lattices [10]. The underlying mechanism is spectral: a finitely ramified fractal compresses the single-particle eigenvalues into narrow, strongly degenerate clusters, which favors superconducting pairing.

The above results are, however, mean-field statements about the pairing amplitude, while in low dimensions the superconducting transition temperature is usually controlled not by the superconducting amplitude, but by its phase. On a fractal lattice the phase is the problematic part. Early results by Vallat, Korshunov, and Beck have shown that the energy of a vortex on a two-dimensional Sierpiński gasket does not grow with the system size but saturates, such that free vortices are present at arbitrarily low temperatures and thus no Berezinski–Kosterlitz–Thouless (BKT) transition into the superconducting state can occur [11].

Monte Carlo simulations of the classical XY model have confirmed the lack of a BKT transition for the Sierpiński gasket [12] and for the Sierpiński pyramid [13]. The mechanism is the same in both cases: on a finitely ramified structure an arbitrarily large region can be disconnected by cutting a finite number of bonds, such that an excitation that destroys coherence costs an energy that does not grow with the system size, while the entropy of placing it does. This is in fact the argument that forbids ordering in the one-dimensional Ising model, transplanted to a fractal [14, 15], thus putting the transition temperature at zero. Interestingly, the Menger sponge, whose order of ramification is infinite, escapes this reasoning, yet Monte Carlo finds no transition there either [16]. The quantum case has been addressed only for antiferromagnetic exchange, where exact diagonalization of the s=1/2s = 1/2 model on a gasket consisting of N=42N = 42 sites has found a disordered ground state with a very short correlation length [17]. The disorder there comes from frustration rather than from the fractal geometry: antiferromagnetic bonds on the odd loops of the gasket cannot be satisfied simultaneously. The Josephson mapping that we consider here, however, gives the ferromagnetic sign, for which the same triangles are un-

Examples of different generations $G=1$, $2$, $4$ for the triangular and honeycomb Sierpiński gasketsVector (PDF)

caption: Examples of different generations G=1G=1, 22, 44 for the (a) triangular and (b) honeycomb Sierpiński gaskets.

:::

frustrated.

The mean-field BdG and classical XY-model results discussed above still leave an important question unsolved. While the mean-field results of the pairing amplitude are encouraging for superconductivity, the classical results on phase fluctuations are not, but neither addresses the question of the ground state at zero temperature. That is, on a fractal geometry, does the superconducting phase have a coherent ground state at all, or do quantum fluctuations disorder it already at T=0T=0? Note that the classical arguments concern thermal vortex proliferation and say nothing about T=0T=0, where the phase instead competes with the charging energy rather than with temperature. However, if quantum fluctuations destroy off-diagonal order in the ground state, the mean-field enhancement of the gap is of no practical use.

Here we address this question numerically. We compute the ground state of the spin-1/2 XY ferromagnet, the extreme quantum limit of a Josephson network with at most one Cooper pair per island, on several different generations of both the triangular and honeycomb Sierpiński gaskets, see Fig. 1. To achieve this, we use a vision-transformer (ViT) neural quantum state (NQS) optimized with variational Monte Carlo. We find that the transverse correlator decays as a power law with an exponent η≃0.12\eta\simeq0.12–0.320.32, which depends only weakly on the gasket generation and hardly at all on how distance is exactly defined. Thus, quantum fluctuations leave quasi-long-range order (QLRO) intact. We then discuss what this implies for the finite-temperature physics of realistic structures, which are finite by construction and for which the thermodynamic limit is of limited relevance. This uncovers an anticipated design trade-off where increasing gasket generation enhances the pairing amplitude, while phase coherence favors smaller systems.

Model and mapping

From a Josephson network to the quantum XY model

We start by considering a network of superconducting islands coupled by weak Josephson links, forming a graph. When the pairing gap on each island, labeled by ii, is large compared with both the island charging energy ECE_C and the Josephson coupling EJE_J, the magnitude of the pairing amplitude is fixed and the low-energy dynamics is carried by the superconducting phases θi\theta_i, described by the quantum phase (rotor) model [18, 19]

Hrotor=HC+HJ=EC2∑i(n^i−ng)2−EJ∑⟨i,j⟩cos⁡(θi−θj),(1)H_{\mathrm{rotor}}=H_C+H_J=\frac{E_C}{2}\sum_i(\hat{n}_i-n_g)^2-E_J\sum_{\langle i,j\rangle}\cos(\theta_i-\theta_j), \tag*{(1)}

where n^i\hat{n}_i counts the Cooper pairs on island ii, and ngn_g is the offset charge set by an external gate potential. We work in the charge eigenbasis, n^i∣n⟩=n∣n⟩\hat{n}_i|n\rangle=n|n\rangle with n∈Zn\in\mathbb{Z}, in which HCH_C is diagonal and the phase enters through the unitary shift eiθi=∑n∣n+1⟩⟨n∣e^{i\theta_i}=\sum_n|n+1\rangle\langle n|. The Josephson term

cos⁡(θi−θj)=12(eiθie−iθj+h.c.)(2)\cos(\theta_i-\theta_j)=\frac{1}{2}\left(e^{i\theta_i}e^{-i\theta_j}+\mathrm{h.c.}\right) \tag*{(2)}

therefore transfers a single Cooper pair between neighboring islands. For convenience we take ng=1/2n_g=1/2, at which the two charge states ni=0,1n_i=0,1 are degenerate, with E(0)=E(1)=EC/8E(0)=E(1)=E_C/8, while every other charge state lies at least ECE_C higher. When EC≫EJE_C\gg E_J, the Josephson term is too weak to populate those higher energy states and the low-energy Hilbert space is spanned by the on-site states {∣0⟩i,∣1⟩i}\{|0\rangle_i,|1\rangle_i\}, where ∣1⟩i|1\rangle_i carries one Cooper pair of charge 2e2e relative to ∣0⟩i|0\rangle_i. Each island thus holds at most one pair, although the ground state within this space is not simply a product of on-site states.

Next, let Pi=∣0⟩i⟨0∣i+∣1⟩i⟨1∣iP_i=|0\rangle_i\langle0|_i+|1\rangle_i\langle1|_i project onto these two states, with P=∏iPiP=\prod_iP_i. Restricted to the image of PP, the charging term HCH_C is the constant NEC/8NE_C/8, which we henceforth drop. Acting on the low-energy space, eiθie^{i\theta_i} takes ∣0⟩i→∣1⟩i|0\rangle_i\to|1\rangle_i and ∣1⟩i→∣2⟩i|1\rangle_i\to|2\rangle_i, but the latter is removed by the projection, leaving

PieiθiPi=∣1⟩i⟨0∣i≡bi†,Pie−iθiPi=∣0⟩i⟨1∣i≡bi.(3)P_i e^{i\theta_i}P_i=|1\rangle_i\langle0|_i\equiv b_i^\dagger,\qquad P_i e^{-i\theta_i}P_i=|0\rangle_i\langle1|_i\equiv b_i. \tag*{(3)}

For i≠ji\ne j, the two exponentials act on distinct tensor factors, so the projection of their product factorizes exactly, Peiθie−iθjP=bi†bjPe^{i\theta_i}e^{-i\theta_j}P=b_i^\dagger b_j, and thus

Pcos⁡(θi−θj)P=12(bi†bj+bj†bi).(4)P\cos(\theta_i-\theta_j)P=\frac{1}{2}\left(b_i^\dagger b_j+b_j^\dagger b_i\right). \tag*{(4)}

The operators bi,bi†b_i,b_i^\dagger obey (bi†)2=0(b_i^\dagger)^2=0 and {bi,bi†}=1\{b_i,b_i^\dagger\}=1 on a given site ii, while commuting on different sites. They are thus hard-core bosons with bi†bi=∣1⟩i⟨1∣ib_i^\dagger b_i=|1\rangle_i\langle1|_i counting the Cooper pairs on island ii. Transferring a pair between an occupied and an empty island stays within the low-energy space, so the projected Josephson term is nonzero and gives

Hhc=−EJ2∑⟨i,j⟩(bi†bj+bj†bi).(5)H_{\mathrm{hc}}=-\frac{E_J}{2}\sum_{\langle i,j\rangle}(b_i^\dagger b_j+b_j^\dagger b_i). \tag*{(5)}

Processes that leave the low-energy space, such as moving a pair onto an already occupied island, cost ECE_C and also generate a nearest-neighbor density-density interaction of order EJ2/ECE_J^2/E_C, negligible in the limit considered here. For ∣ng−1/2∣<1\lvert n_g-1/2\rvert<1, but where the same two charge states still remain lowest, detuning ngn_g adds a chemical potential −μ∑ibi†bi-\mu\sum_i b_i^\dagger b_i with μ=(EC/2)(1−2ng)\mu=(E_C/2)(1-2n_g). This term commutes with Eq. (5) and thus at most shifts which pair-number sector lies lowest, and therefore we drop it.

The next step is to identify that hard-core bosons are equivalent to spin-1/2 operators. Under the Matsubara-Matsuda identification [20]

Si+=bi†,Si−=bi,Siz=n^i−12,(6)S_i^+=b_i^\dagger,\qquad S_i^-=b_i,\qquad S_i^z=\hat n_i-\frac{1}{2}, \tag*{(6)}

the hard-core constraint becomes the two-dimensionality of the on-site Hilbert space of spin operators, ng=1/2n_g=1/2 corresponds to ∑iSiz=0\sum_i S_i^z=0, and the pair hopping turns into a transverse exchange,

bi†bj+bj†bi=Si+Sj−+Si−Sj+=2(SixSjx+SiySjy).(7)b_i^\dagger b_j+b_j^\dagger b_i=S_i^+S_j^-+S_i^-S_j^+=2(S_i^xS_j^x+S_i^yS_j^y). \tag*{(7)}

All this results in Eq. (5) transforming into the spin-1/2 XY ferromagnet

H=−J∑⟨i,j⟩(SixSjx+SiySjy),J=EJ>0,(8)H=-J\sum_{\langle i,j\rangle}(S_i^xS_j^x+S_i^yS_j^y),\qquad J=E_J>0, \tag*{(8)}

with Six,y=12σix,yS_i^{x,y}=\frac{1}{2}\sigma_i^{x,y}.

We here stress that the word “island” carries no commitment to a particular length scale. In our earlier work on fractal superconductors [8, 9], the lattice was atomistic, with each island being a single atom of an artificial crystal of the kind assembled by STM manipulation [1] or molecular self-assembly [2]. The Josephson description of Eq. (1) is also directly relevant to mesoscopic arrays of lithographic junctions [4, 5, 6]. An effective Josephson lattice has even been derived directly from the Hubbard model for the cuprates, with the coupling obtained from microscopic correlation functions [21]. What is required is simply that the energy cost of adding a Cooper pair to a site exceeds the coupling between sites, and this holds at the atomic scale as much as for micron-sized grains. In a single orbital, it is the Pauli principle forbidding a second pair outright, so the effective charging energy is naturally large. The relevant parameters are thus the ratio EJ/ECE_J/E_C and also the connectivity of the graph, not the physical size of an island. (8) therefore describes the phase sector of a fractal superconductor at both extremes, both an atomistic lattice and a lithographic array, and the ratio EJ/ECE_J/E_C is what interpolates between the nearly classical and the strongly quantum regimes.

For the spin-1/2 XY ferromagnet in Eq. (8), the in-plane magnetization ⟨Six⟩\langle S_i^x\rangle translates to the condensate order parameter ⟨bi⟩\langle b_i\rangle in Eq. (5), while the transverse correlator

Cxy(i,j)=⟨σixσjx+σiyσjy⟩∝⟨bi†bj+h.c.⟩(9)C_{xy}(i,j)=\langle\sigma_i^x\sigma_j^x+\sigma_i^y\sigma_j^y\rangle\propto\langle b_i^\dagger b_j+\mathrm{h.c.}\rangle \tag*{(9)}

is the off-diagonal one-body density matrix. The long-distance behavior of the latter is the standard diagnostic: a nonzero limit means off-diagonal long-range order, algebraic decay means a quasi-condensate, while exponential decay is an insulator. As Eq. (8) is invariant under global U(1)U(1) rotations about the zz axis, ⟨Six⟩\langle S_i^x\rangle vanishes identically in any finite system. The correlator, or equivalently the squared in-plane magnetization

⟨M⊥2⟩=∑i≠j⟨σixσjx+σiyσjy⟩=∑i≠jCxy(i,j),(10)\langle M_\perp^2\rangle=\sum_{i\ne j}\langle\sigma_i^x\sigma_j^x+\sigma_i^y\sigma_j^y\rangle=\sum_{i\ne j}C_{xy}(i,j), \tag*{(10)}

is therefore the appropriate order parameter [22].

Geometries and distance measures

We consider the triangular Sierpiński gasket of generation GG, which has N=3(3G+1)/2N=3(3^G+1)/2 sites (N=6,15,42,123,366N=6,15,42,123,366 for G=1,…,5G=1,\ldots,5) and 3G+13^{G+1} bonds, alongside its decorated variant, which has N=(5⋅3G+3)/2N=(5\cdot3^G+3)/2 sites (N=9,24,69,204,609N=9,24,69,204,609) and an identical bond count, see Fig. 1. To measure the correlator in Eq. (10), we need a notion of distance. We therefore measure the separation between sites ii and jj in two ways: by the graph distance dijd_{ij}, i.e. the number of bonds along the shortest path, and by the Euclidean distance rijr_{ij} in the embedding plane. In each case, we average over all pairs at a given separation,

Cxy(d)=1Nd∑dij=dCxy(i,j),(11)C_{xy}(d)=\frac{1}{\mathcal{N}_d}\sum_{d_{ij}=d}C_{xy}(i,j), \tag*{(11)}

where Nd\mathcal{N}_d is the number of such pairs, and analogously for Cxy(r)C_{xy}(r) with rijr_{ij} in place of dijd_{ij}. In addition, we consider pairs lying on the outer boundary between the extremal corners, where the two measures coincide. These are the longest paths in the structure and the ones most exposed to disconnection, so comparing their decay with the bulk tests whether the algebraic behavior is a property of the state as a whole.

Vision-transformer ansatz and variational optimization

Beyond the smallest generations of the considered fractal structures, the Hilbert space is far out of reach of exact methods, and so we turn to a variational approach. We therefore parametrize the many-body ground state wave function by a vision-transformer neural quantum state [23, 24, 25], optimized by variational Monte Carlo within NetKet [26]. The gaskets have no translational symmetry to exploit and no natural one-dimensional ordering, so neither symmetric convolutional architectures nor matrix-product states apply directly, whereas the attention mechanism handles an unstructured graph without imposing locality. Working with the wave function itself also gives access to all-to-all correlators and to fidelities between independently optimized states, which we use in Appendix § 1.1.1 to assess the reproducibility of the optimization.

The ansatz consists of three stages, shown in Figure 2: (a) the embedding of the spin configuration into a sequence of patch representations, (b) a stack of transformer encoder layers, and (c) an output layer that maps the result onto the complex logarithm of the wave function, log⁡Ψ(σ)\log\Psi(\sigma) [23], for some spin configuration σ\sigma. The patches in Figure 2(a) are the smallest building blocks of the gasket, indicated by the dashed regions and, unlike in conventional ViT implementations [25, 27], here they are allowed to overlap. Each patch is linearly projected into a DD-dimensional embedding space. The resulting sequence of patch embeddings, x=(x1,…,xnp)\boldsymbol{x} = (x_1,\ldots,x_{n_p}), one DD-dimensional token per patch, is the input to the encoder.

Figure 2. Vision-transformer architecture. (a) The dashed blue boxes indicate the patches into which the fractal is tokenized. The arrows denote some spin configuration σ\boldsymbol{\sigma}. (b) The tokens are processed by nℓn_\ell transformer encoder layers. (c) The output layer maps the result onto the log-amplitude log⁡Ψ(σ)\log\Psi(\boldsymbol{\sigma}).

Each encoder layer in Figure 2(b) consists of a prenormalized [28] factored multi-head spatial attention block, followed by a feed-forward network (FFN), with residual connections [29] around both; we omit the layer index. Here, after layer normalization (LN), the patch representations are passed through a learnable value projection and split among nhn_h heads, giving vj(h)∈RD/nh\boldsymbol{v}_j^{(h)} \in\mathbb{R}^{D/n_h}. The output of head hh for patch ii is

Ai(h)=∑jαij(h)wijvj(h),(12)\boldsymbol{A}_i^{(h)} = \sum_j \alpha_{ij}^{(h)} w_{ij} \boldsymbol{v}_j^{(h)}, \tag*{(12)}

where the attention weights αij(h)\alpha_{ij}^{(h)} are learnable parameters, and wijw_{ij} is a distance-dependent kernel, which softly biases the attention weights [30],

wij=(1+rij)−γ∑k(1+rik)−γ,(13)w_{ij} = \frac{(1+r_{ij})^{-\gamma}}{\sum_k (1+r_{ik})^{-\gamma}}, \tag*{(13)}

with rijr_{ij} the Euclidean distance between the centers of patches ii and jj. The exponent γ\gamma is learnable and shared by all heads within a layer, while each head has its own attention matrix α(h)\alpha^{(h)} and acts on its own D/nhD/n_h-dimensional value subspace, so that different heads can learn different spatial correlations. The head outputs are then concatenated into a single vector Ai\boldsymbol{A}_i per patch and linearly projected, which mixes information between heads. The input representation is then added back through the residual connection. The second sub-block is a position-wise FFN, a two-layer multilayer perceptron (MLP) with a Gaussian Error Linear Unit (GELU) activation. Here, after the second layer normalization the hidden dimension is expanded from DD to 4D4D and projected back to DD, and a second residual connection gives the output of that layer. This is repeated for nln_l encoder layers, the output of layer ℓ\ell serving as the input of layer ℓ+1\ell+1.

After the last encoder layer, the patch representations are summed and layer-normalized into a single global vector, shown in Figure 2(c). This vector is passed through two independent output branches, each a MLP followed by its own layer normalization, which produce the real and imaginary feature vectors ζR\zeta_R and ζI\zeta_I, combined into ζ=ζR+iζI\zeta= \zeta_R + i\zeta_I. The logarithm of the variational wave function is then finally given as

log⁡Ψ(σ)=∑i=1Dlog⁡cosh⁡ζi.(14)\log\Psi(\boldsymbol{\sigma}) = \sum_{i=1}^{D} \log\cosh\zeta_i. \tag*{(14)}

The network parameters are optimized by variational Monte Carlo (VMC) [31], with spin configurations sampled from ∣Ψ(σ)∣2|\Psi(\sigma)|^2. The energy is minimized by stochastic reconfiguration (SR) [32] in its minSR formulation [33], which makes SR affordable at large parameter counts. We use Ns=213N_s = 2^{13} Monte Carlo samples spread over as many Markov chains, combined with the SPRING momentum scheme [34] with a momentum of 0.8. The optimization of the variational state is performed in two stages. In the first stage the total magnetization is not fixed: configurations are sampled with local Metropolis updates, and the state is optimized for 750 iterations with an initial learning rate of 10−210^{-2} with exponential decay and a diagonal shift of 5×10−45 \times10^{-4}. In the second stage the total magnetization is fixed to Sz=0S^z = 0 for an even number of sites and to Sz=1/2S^z = 1/2 for an odd number, and configurations are sampled with Metropolis exchange updates, which preserve SzS^z. The network is warm-started from the converged states of the first stage and optimized for a further 400 iterations. The resulting state is then symmetrized under the C3vC_{3v} point group of the gasket by quantum-number projection [35] and optimized for another 20 iterations. For seeds whose variance deviates strongly from the others, the second stage is repeated. Hyperparameters tuned for one generation or gasket type need not transfer to another, so we select the network and learning-rate hyperparameters by a grid search minimizing the average energy variance at G=3G = 3 and G=4G = 4.

Convergence of the variational state towards the ground state is assessed by monitoring the energy and the energy variance per site, which we combine into the V-score [36]

V=NVar⁡(E)(E−E∞)2,(15)V = \frac{N \operatorname{Var}(E)}{(E-E_\infty)^2}, \tag*{(15)}

where E∞=Tr⁡H/dim⁡HE_\infty= \operatorname{Tr} H/\dim\mathcal{H} is the infinite-temperature energy, which vanishes for the XY Hamiltonian. Convergence of the energy at an early stage does not guarantee that the variational quantum state is free from excited-state contributions. A decreasing energy variance per site indicates that the wavefunction is approaching an eigenstate of the Hamiltonian. The simultaneous convergence of these two quantities, with further corroboration of the high fidelity between different optimized states depicted in Appendix § 1.1.1, provides strong evidence that the variational quantum state approaches the ground state.

Figure 3. V-score over the last 200 iterations for the (a) triangular and (b) honeycomb Sierpiński gaskets. Dashed vertical lines mark the iteration from which the C3vC_{3v} symmetry projection is enforced. V-scores for G=1G = 1 are omitted, as they repeatedly drop below O(10−5)O(10^{-5}) and recover during the VMC iterations.

Results

Using the ViT NQS to calculate the ground state, we extract the transverse two-point shell-averaged correlator in Eq. (11), averaged over 5 independently optimized variational states, to understand phase coherence in the rotor model. We plot the result in Figure 4 for both the triangular (a,c) and honeycomb (b,d) Sierpiński gaskets. For each gasket, we report both the correlator as a function of network distance dd and Euclidean distance rr. For comparison, we also include the pristine lattice, which is not a fractal. On a double-logarithmic scale, all data fall on straight lines over the full accessible range of distances, for every generation of both fractals and for both

Two-point shell-averaged correlators

Figure 4. Two-point shell-averaged correlators CxyC_{xy} for the triangular Sierpiński gasket (a,c) and the honeycomb Sierpiński gasket (b,d) as a function of graph distance dd (a,b) and Euclidean distance rr (c,d), for different generations GG of each fractal. Straight lines on the double-logarithmic scale correspond to algebraic decay fits to Eq. (16).

metrics. This leads to the fits

Cxy(r)=Ar−η(16)C_{xy}(r) = A r^{-\eta} \tag*{(16)}

with the exponents in Table I. Throughout, η\eta lies between 0.12 and 0.32. Thus, finite and positive, so there is no true long-range order, but still well below unity, so the decay is slow and the state is therefore a quasi-condensate. Notably, the two metrics give the same exponent to within a few percent, and so do boundary pairs, which follow the longest paths and are the most exposed to disconnection. The algebraic decay is thus a property of the state as a whole rather than of a particular projection of the geometry. The amplitude AA stays close to unity in the graph metric but falls with GG in the Euclidean one, as expected from the growing mismatch between chemical and embedding distance as the gasket is refined. Note that five generations span about two decades in linear size. This is enough to establish the algebraic character of the decay and to bracket the exponent. However, it is not enough to extract η\eta in the thermodynamic limit or to interpret its mild drift with GG. If we instead attempt an exponential fit, Cxy=Be−r/ξC_{xy} = B e^{-r/\xi}, it performs worse throughout and returns correlation lengths comparable to the system size, as expected for an exponential fit to algebraic data.

To summarize, within the stated numerical limits, our conclusion is unambiguous and holds for both geometries, all generations, and both metrics: the ground state of the quantum XY model on the Sierpiński gasket supports quasi-long-range order. In terms of superconductivity, this means that the extreme quantum limit of a fractal

graphEuclideanboundary
AAη\etaAAη\etaAAη\eta
Triangular gasket
G=1G = 11.230.1821.220.1811.290.210
G=2G = 21.130.2851.070.2441.150.292
G=3G = 31.100.2781.020.2551.100.290
G=4G = 41.070.2501.000.2421.070.278
G=5G = 51.000.1960.9450.1901.030.232
Pristine1.070.1761.070.1821.090.210
Honeycomb gasket
G=1G = 11.190.2650.9800.2681.210.290
G=2G = 21.140.2980.8900.2861.150.315
G=3G = 31.100.2920.8450.2801.100.314
G=4G = 41.040.2570.8080.2451.050.286
G=5G = 50.8800.1580.6920.1210.8950.180
Pristine1.050.1860.9090.1891.050.216

Table 1. Fit parameters of Eq. (16) for the triangular and honeycomb Sierpiński gaskets across three different distance measures.

Josephson network is a quasi-condensate.

Discussion of finite structures and finite temperature

Having established quasi-long-range order in the ferromagnetic spin-1/2 XY model for finite fractals, we come back to the actual question of superconducting phase coherence at finite temperatures relevant for experimentally accessible fractal superconductors. First of all, the classical results cited in the introduction concern the thermodynamic limit. They follow the helicity modulus (the response of the free energy to a twist of the boundary phases, that is the phase stiffness of the classical model) as the gasket generation is increased without bound [11, 12, 13]. For a bulk material, this limit of ever-increasing gasket generations is the natural limit. However, for current experimental structures it is not: an STM-assembled fractal contains dozens to a few hundred atoms, a lithographic gasket array is fabricated at low generation, such that the sample is always finite by construction. Thus the relevant question must be how a gasket of finite generation GG behaves at the temperatures and length scales at which it is operated.

Second, one might be tempted to read algebraic correlations with η≃1/4\eta\simeq1/4 found in this work as evidence of BKT physics. Here η=1/4\eta= 1/4 follows from η=kBT/2πKR\eta= k_{\mathrm{B}}T/2\pi K_{\mathrm{R}} at the universal jump of the helicity modulus, KR=2/πK_{\mathrm{R}} = 2/\pi [37, 38, 39], and therefore requires a finite TBKTT_{\mathrm{BKT}}. However, on a finitely ramified fractal the classical transition temperature vanishes, and the same relation gives η→0\eta\to0. This is consistent with the classical ground state, which on any connected graph is fully aligned and has true long-range order rather than algebraic decay. The algebraic correlations we find are thus of quantum origin, and the thermal mechanism that fixes 1/4 cannot produce them.

Our exponents are instead set by the lattice geometry and by the quantum fluctuations of the phase, and their proximity to 1/4 does not, by itself, indicate an underlying connection to BKT criticality.

Thus, posed this way, the problem at finite temperatures in finite fractals separates into a quantum part and a classical part. The quantum part is settled by this work: at T=0T = 0, and at the maximal strength of quantum phase fluctuations compatible with superconductivity, the finite gasket carries algebraic off-diagonal correlations with an exponent well below unity. This result complements the mean-field results of Ref. [8], where the superfluid stiffness of these gaskets was found to be finite and comparable to that of the regular lattice. Since our limit of EC≫EJE_{\mathrm{C}} \gg E_{\mathrm{J}} is the extreme quantum limit of the rotor model, this is the least favorable case for phase coherence. At any finite EC/EJE_{\mathrm{C}/E_{\mathrm{J}}} the phase fluctuations are weaker and the classical XY limit is even recovered as EC/EJ→0E_{\mathrm{C}}/E_{\mathrm{J}} \to0. Quantum fluctuations of the phase, then, are not what limits the existence of a finite fractal superconductor.

The classical part of the problem is the finite-size behavior at T>0T > 0, which we intend to address in depth in future work, though the outline of the answer is already clear. A gasket of generation GG has a linear size 2G2^G. Since the vortex energy is bounded, free vortices are present at any T>0T > 0 with a density that sets a mean spacing ℓv(T)\ell_{\mathrm{v}}(T) between them, and the phase stays coherent over distances shorter than ℓv\ell_{\mathrm{v}}. A finite structure therefore retains a superconducting response as long as ℓv(T)≳2G\ell_{\mathrm{v}}(T) \gtrsim2^G, which defines a crossover temperature that depends on GG, but that the thermodynamic-limit analysis integrates out. The existing experiments are consistent with this picture of existing phase coherence. In periodic arrays of gaskets of order GG joined at their vertices, where the geometry is fractal below the gasket size and Euclidean above it, a BKT transition is recovered and shows up as a sharp drop of the inverse sheet magnetoinductance, with transition temperatures following τc,G=(3/5)Gτc,0\tau_{\mathrm{c},G} = (3/5)^G \tau_{\mathrm{c},0}, where τc,0=kBT/EJ≃1.5\tau_{\mathrm{c},0} = k_{\mathrm{B}}T/E_{\mathrm{J}} \simeq1.5 for a regular triangular array of the same junctions [6]. Therefore, higher generations of fractals show diminishing phase coherence, though it remains measurable in smaller systems.

This illustrates an anticipated design trade-off. Increasing the gasket generation sharpens the spectral clustering that enhances the pairing amplitude [8], while the phase stiffness, and with it phase coherence, moves the other way; in the periodic arrays of Ref. [6] it even falls by a factor of 3/5 per generation. Since the actual superconducting transition temperature TcT_{\mathrm{c}} is bounded by the smaller of the two scales [40], the optimum necessarily lies at a finite, and presumably rather low, generation. The figure of merit can thus be thought of as the product of amplitude and stiffness rather than the gap enhancement alone. Note however that how strongly the stiffness of a single gasket falls with GG is a separate question from the array measurement, and a harder one. The helicity modulus is defined through a twist imposed across the system, which the periodic arrays of Ref. [6] realize but an isolated gasket with open boundaries does not. A response can still be defined, either by threading flux through the loops of the gasket or from the current-current correlator of Eq. (8), and our variational states give access to both; which of these is the appropriate finite-size analogue of the stiffness we leave for future work.

Conclusion

We compute the ground state of the ferromagnetic spin-1/21/2 XY model, which is the phase sector of a Josephson network in the limit of at most one Cooper pair per island, i.e. the extreme quantum limit, on both triangular and honeycomb Sierpiński gaskets up to generation G=5G=5, using vision-transformer neural quantum states. This architecture reaches system sizes of several hundred sites on a lattice with no translational symmetry and no natural one-dimensional ordering, where exact methods can at most handle a few dozen sites.

We find that the transverse correlations decay algebraically with η≃0.12−0.32\eta\simeq0.12 - 0.32, the same for different metrics and for both geometries, and the exponent drifts only weakly with generation. Quantum phase fluctuations therefore do not destroy quasi-long-range order on a finitely ramified fractal, not even when they are at their strongest. Together with the recent mean-field prediction of an enhanced pairing amplitude on the same geometry [8, 9], this locates the only remaining obstacle to fractal superconductivity in the classical, finite-temperature sector. Notably, the theorems that forbid ordering on an infinite gasket have limited bearing on structures of a few hundred sites, and the careful finite-size classical analysis is the natural next step. Our reasoning suggests a trade-off in gasket size, with larger gasket generations yielding a higher pair amplitude but lower phase coherence. Two clear questions still remain. The phase stiffness of the ground state would turn the trade-off between pairing amplitude and phase coherence into a quantitative statement about the optimal gasket generation, but

as discussed above, the standard definition through an imposed twist does not carry over to an isolated fractal with open boundaries, and settling which response takes its place is part of the problem. A second open question is the rotor model at finite EC/EJE_C/E_J, which interpolates between the case treated here and the classical limit; we expect coherence to survive throughout, but the crossover has not been mapped.

Acknowledgments

R.C., M.I.K., and A.A.B. were supported by the research program “Materials for the Quantum Age” (QuMat). This program (registration number 024.005.006) is part of the Gravitation program financed by the Dutch Ministry of Education, Culture and Science (OCW). A.A.B. was also supported by NWO grant OCENW.M.23.044. J.v.P and A.M.B.S. acknowledge funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (ERC-2022-CoG, Grant agreement No. 101087096). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. The computations were enabled by resources in project UPPMAX 2026/1-2 provided by Uppsala University at UPPMAX, as well as the National Academic Infrastructure for Supercomputing in Sweden (NAISS), partially funded by the Swedish Research Council through grant agreement no. 2022-06725.

Data availability

The code and data required to reproduce the results of this work are openly available [41].

Fidelity analysis

To check that the optimization converges to the same state regardless of initialization, we performed five independent optimizations for each gasket geometry at generations G=1G=1–55, differing only in the random seed used to initialize the variational parameters. For each pair of optimized states we computed the infidelity 1−F1-F, as implemented in NetKet [26]. For the smallest systems (triangular G≤2G \le2, honeycomb G=1G=1) it was evaluated exactly by summing over the full SzS^z sector; for larger systems it was estimated by Monte Carlo sampling. The results, shown in Figure 5, confirm that different runs converge to the same state. Where energy differences are statistically resolved, they remain below 10−410^{-4} per site and reflect small differences in convergence rather than distinct states.

Figure 5. The lower triangle shows the pairwise infidelities 1−F1-F, and the upper triangle the absolute energy difference per site, between independently optimized runs for the triangular (top row) and honeycomb (bottom row) gaskets, showing that all runs converge to the same state. Dots mark pairs whose energy difference exceeds 2σ2\sigma.

References

  1. [1]S. N. Kempkes, M. R. Slot, S. E. Freeney, S. J. M. Zevenhuizen, D. Vanmaekelbergh, I. Swart, and C. Morais Smith, Nature Physics 15, 127 (2019).
  2. [2]J. Shang, Y. Wang, M. Chen, J. Dai, X. Zhou, J. Kuttner, G. Hilt, X. Shao, J. M. Gottfried, and K. Wu, Nature Chemistry 7, 389 (2015).
  3. [3]R. Cañellas, C. Liu, R. Arouca, L. Eek, G. Wang, Y. Yin, D. Guan, Y. Li, S. Wang, H. Zheng, C. Liu, J. Jia, and C. Morais Smith, Nature Physics 20, 1421 (2024).
  4. [4]J. M. Gordon, A. M. Goldman, J. Maps, D. Costello, R. Tiberio, and B. Whitehead, Phys. Rev. Lett. 56, 2280 (1986).
  5. [5]R. Meyer, J. L. Gavilan, B. Jeanneret, R. Théron, C. Leemann, H. Beck, and P. Martinoli, Phys. Rev. Lett. 67, 3022 (1991).
  6. [6]R. Meyer, S. E. Korshunov, C. Leemann, and P. Martinoli, Phys. Rev. B (2002), arXiv:cond-mat/0205105.DOI
  7. [7]M. Amundsen, V. Juričić, and J. A. Ouassou, Appl. Phys. Lett. 125, 092601 (2024).
  8. [8]A. A. Iliasov, R. Canyellas, M. I. Katsnelson, and A. A. Bagrov, J. Phys.: Condens. Matter 38, 205601 (2026).
  9. [9]R. Canyellas, M. I. Katsnelson, and A. A. Bagrov, arXiv preprint (2026), arXiv:2603.17362.arxiv.org/abs/2603.17362
  10. [10]V. Bashmakov, A. Iliasov, T. Bzdušek, and A. A. Bagrov, arXiv preprint (2025), arXiv:2509.09330.arxiv.org/abs/2509.09330
  11. [11]A. Vallat, S. E. Korshunov, and H. Beck, Phys. Rev. B 43, 8482 (1991).
  12. [12]B. Mitrović and S. K. Bose, Phase Transitions 83, 572 (2010).
  13. [13]M. Przedborski and B. Mitrović, Phase Transitions 84, 85 (2011).
  14. [14]Y. Gefen, B. B. Mandelbrot, and A. Aharony, Phys. Rev. Lett. 45, 855 (1980).
  15. [15]Y. Gefen, A. Aharony, Y. Shapir, and B. B. Mandelbrot, J. Phys. A: Math. Gen. 17, 435 (1984).
  16. [16]M. A. Przedborski and B. Mitrović, Physica A (2014), arXiv:1310.4065.DOI
  17. [17]A. Voigt, W. Wenzel, J. Richter, and P. Tomczak, Eur. Phys. J. B 38, 49 (2004).
  18. [18]A. Dutta and J. Bhattacharjee, Physical Review B 64, 184106 (2001).
  19. [19]M. Medvidović and D. Sels, PRX Quantum 4, 040302 (2023).
  20. [20]T. Matsubara and H. Matsuda, Progress of Theoretical Physics 16, 569 (1956).
  21. [21]M. Harland, S. Brener, A. I. Lichtenstein, and M. I. Katsnelson, Phys. Rev. B 100, 024510 (2019).
  22. [22]A. W. Sandvik, AIP Conf. Proc. 1297, 135 (2010).
  23. [23]G. Carleo and M. Troyer, Science 355, 602 (2017).
  24. [24]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Advances in neural information processing systems 30, 10.48550/arXiv.1706.03762 (2017).
  25. [25]L. L. Viteritti, R. Rende, and F. Becca, Phys. Rev. Lett. 130, 236401 (2023).
  26. [26]F. Vicentini, D. Hofmann, A. Szabó, D. Wu, C. Roth, C. Giuliani, G. Pescia, J. Nys, V. Vargas-Calderón, N. Astrakhantsev, and G. Carleo, SciPost Phys. Codebases 7 (2022).
  27. [27]L. L. Viteritti, R. Rende, A. Parola, S. Goldt, and F. Becca, Physical Review B 111, 134411 (2025).
  28. [28]R. Xiong, Y. Yang, D. He, K. Zheng, S. Zheng, C. Xing, H. Zhang, Y. Lan, L. Wang, and T. Liu, in Proceedings of the 37th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 119, edited by H. D. III and A. Singh (PMLR, 2020) pp. 10524–10533.DOI
  29. [29]K. He, X. Zhang, S. Ren, and J. Sun, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) pp. 770–778.
  30. [30]L. L. Viteritti, R. Rende, S. Sachdev, and G. Carleo, arXiv preprint arXiv:2602.02665 10.48550/arXiv.2602.02665 (2026).arxiv.org/abs/2602.02665
  31. [31]F. Becca and S. Sorella, Quantum Monte Carlo approaches for correlated systems (Cambridge University Press, 2017).DOI
  32. [32]S. Sorella, Phys. Rev. B 71, 241103(R) (2005).
  33. [33]A. Chen and M. Heyl, Nature Physics 20, 1476 (2024).
  34. [34]G. Goldshlager, N. Abrahamsen, and L. Lin, Journal of Computational Physics 516, 113351 (2024).
  35. [35]Y. Nomura, Journal of Physics: Condensed Matter 33, 174003 (2021).
  36. [36]D. Wu, R. Rossi, F. Vicentini, N. Astrakhantsev, F. Becca, X. Cao, J. Carrasquilla, F. Ferrari, A. Georges, M. Hibat-Allah et al., Science 386, 296 (2024).
  37. [37]J. M. Kosterlitz and D. J. Thouless, J. Phys. C: Solid State Phys. 6, 1181 (1973).
  38. [38]D. R. Nelson and J. M. Kosterlitz, Phys. Rev. Lett. 39, 1201 (1977).
  39. [39]Y. Komura and Y. Okabe, J. Phys. Soc. Jpn. 81, 113001 (2012).
  40. [40]V. J. Emery and S. A. Kivelson, Nature 374, 434 (1995).
  41. [41]R. Canyellas and J. van Poppelen, FractalXY, release v1.0, https://github.com/jannesvoppelen/FractalXY/releases/tag/v1.0 (2026).

Paper details

vlnc:W_58x8zey5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents