betaFeedback

WorksPhysics and Astronomy

High-temperature series expansion of the dynamic Matsubara spin correlator

Abstract

The high-temperature series expansion for quantum spin models is a well-established tool to compute thermodynamic quantities and equal-time spin correlations, in particular for frustrated interactions. We extend the scope of this expansion to the dynamic Matsubara spin-spin correlator and develop an algorithm that yields exact expansion coefficients in the form of rational numbers. We focus on Heisenberg models with a single coupling constant J and spin lengths S=1/2,1. The expansion coefficients up to 12th order in J/T are precomputed on all possible ∼106\sim 10^6 graphs embeddable in arbitrary lattices and are provided in a repository. This enables calculation of static momentum-resolved susceptibilities for arbitrary site-pairs or wavevectors. We test our results for the antiferromagnetic S=1/2 chain and triangular lattice model. An important application that we discuss in a companion letter is the calculation of real-frequency dynamic structure factors. This is achieved by identifying the high-frequency expansion coefficients of the Matsubara correlator with frequency moments of the spectral function.

Introduction

The high-temperature series expansion (HTE) is an invaluable tool for the theoretical analysis of quantum spin systems in thermal equilibrium, for monographs see, e.g., Refs. 1 and 2. The HTE is oblivious to frustration and entanglement, is formulated directly in the thermodynamic limit and remains applicable in high dimensions. Up to now, the HTE can target thermodynamic quantities like entropy, heat capacity or uniform susceptibility and was also applied to obtain equal-time spin correlation functions Gii′zz=⟨SizSi′z⟩G_{ii'}^{zz} = \langle S_i^z S_{i'}^z \rangle for arbitrary site-pairs ii′ii'.

Key technical advancements of the HTE method included an extension to arbitrary spin-length SS [3, 4], flexible open-source software packages [5] or inclusion of magnetic fields [6]. As the name suggests, the main challenge of HTE is to reach temperatures much below the spin interaction T≪JT \ll J when resummation schemes of the bare series give ambiguous and thus non-reliable results. To some extent, these issues can be bypassed if qualitative information on the very low-TT behavior is available [7].

HTE has an impressive track-record of achievements: For example, it has first revealed [8] the anomalous intermediate-TT behavior of the nearest-neighbor S=1/2S=1/2 Heisenberg AFM on the triangular lattice which cannot be understood in the renormalized-classical picture suggested by the ordered ground state. The latter picture predicts a much smaller entropy and a much larger correlation length than found from HTE. The underlying physical reason of this effect is still not well understood [9, 10]. Another example is the accurate quantitative analysis of thermal phase transitions in otherwise challenging three-dimensional frustrated models [11].

However, in order to achieve a better understanding of collective phenomena arising in equilibrium quantum spin systems it is mandatory to also computationally target spin dynamics. Here, the most elementary observable is the dynamical spin-spin correlator Gii′zz(t)=⟨Siz(t)Si′z⟩G_{ii'}^{zz}(t) = \langle S_i^z(t)S_{i'}^z \rangle with Siz(t)=eiHtSize−iHtS_i^z(t) = e^{iHt}S_i^z e^{-iHt} representing a local spin operator in the Heisenberg picture [12]. The dynamical structure factor which is defined as a spatial and temporal Fourier transform of Gii′zz(t)G_{ii'}^{zz}(t) contains rich information on the dipolar excitation spectrum, the presence and stability of quasiparticles or the fractionalization of spin-flip excitations [13]. It is routinely measured using inelastic neutron scattering [14] in the solid-state context and recently also for cold-atom quantum simulators [15] via Raman spectroscopy. Although the dynamical structure factor can be calculated in various ways for different settings [16] (e.g. via exact diagonalization and linked-cluster methods [17], spin-wave theory [18] and tensor-networks [19]), it is still desirable to approach spin dynamics via the HTE and benefit from its above mentioned strengths. This is particular important for (strongly frustrated) spin liquid candidates [20] where the observables available from conventional HTE are often featureless as functions of temperature and momentum. A natural intermediate goal to this endeavor is the extension of the HTE for the dynamical spin correlation function in imaginary time t→−iτ∈iRt \to-i\tau\in i\mathbb{R} which is somewhat simpler task given the benefits of the Matsubara formalism for perturbation theory in equilibrium [12].

In this work, we develop this extension of the HTE to the dynamic imaginary frequency (Matsubara) spin-spin correlator. We term this method dynamic HTE (Dyn-HTE). We lay out the general formalism and apply it to Heisenberg models with a single coupling constant JJ and spin lengths S∈{1/2,1}S \in\{1/2,1\}. The coefficients for the expansion in powers of x≡J/Tx \equiv J/T up to order nmax⁡=12n_{\max}=12 (and in inverse frequency to order ⌈nmax⁡/2⌉\lceil n_{\max}/2 \rceil) are precomputed exactly in the form of rational numbers on all possible ∼106\sim10^6 graphs and are offered for download along with powerful tools for the creation of arbitrary lattices and efficient graph embeddings [21]. On the technical side Dyn-HTE hinges on the exploitation of the recently developed Kernel trick [22] which solves the (n+2)(n+2)-fold imaginary-time integrals (required at expansion order nn) analytically. Interestingly, the numerical cost for the expansion in order nn is only modestly increased compared to the conventional HTE for the equal-time correlator. Depending on graph topology the extra effort is by a factor of nn in the worst case.

For historic context, early developments of spin-diagrammatic schemes date back to the late 1960s [23, 24] where the Wick theorem was generalized for spin operators, see also Ref. [25] for a modern recursive formulation. Recently, these ideas have partially been revived in the development of a functional-renormalization group approach for spin systems [25, 26, 27] which in principle can also be used to generate order-by-order expansions of the Matsubara correlator [28]. However, the high expansion orders achieved by Dyn-HTE in this work have not been matched by any other approach.

One experimental observable directly available from Dyn-HTE is the static susceptibility for arbitrary site-pairs. We consider the AFM S=1/2S=1/2 Heisenberg chain and compare the static susceptibility from Dyn-HTE against error controlled quantum Monte Carlo (QMC). We also report the static susceptibility for the frustrated triangular lattice model. As an application we employ the static susceptibility of Dyn-HTE to showcase the accuracy of a simple approximate parametrization of its momentum dependence as suggested recently under the name renormalized mean-field form.

Finally, as motivated above, a main application of the Matsubara correlator is its analytical continuation to the real-frequency dynamical spin structure factor [12]. A diverse set of methods like QMC or pseudo-fermion based diagrammatic approaches [29, 30, 31, 32, 33] produce approximate numerical correlator data on a limited set of points on the Matsubara axis. In such a situation, despite recent advances [34, 35, 36, 37, 38], this analytical continuation is an ill-conditioned and error-prone procedure. For Dyn-HTE, in contrast, the obtained exact expansion can be regrouped in form of a high-frequency expansion (in inverse Matsubara frequency). In our companion work [39] we show that the associated expansion coefficients can be identified with the frequency moments of the (real-frequency) spectral function from which the dynamical structure factor can be reconstructed by standard methods. This means that Dyn-HTE allows to bypass the standard ill-defined analytical continuation procedure.

Heisenberg spin model

We consider a system of length-SS quantum spins with operators SiαS_i^\alpha where α=x,y,z\alpha=x,y,z. The subscript i=1,2,…,Ni=1,2,\ldots,N refers to the site at position ri\mathbf{r}_i of an arbitrary lattice L\mathcal{L}. The spins interact via Heisenberg exchange

characterized by a single coupling constant JJ along an arbitrary subset of all N(N−1)/2N(N-1)/2 site pairs which we call bonds (ii′)(ii'). This includes the important case of symmetry-related nearest-neighbor interactions, but also all-to-all or spatially disordered (but equal) interactions like in a lattice with vacancies. The Hamiltonian reads

H=J∑(ii′)(Si+Si′−+Si−Si′++SizSi′z)≡J∑(ii′)Vii′,(1)H=J\sum_{(ii')}\left(S_i^+S_{i'}^-+S_i^-S_{i'}^++S_i^zS_{i'}^z\right)\equiv J\sum_{(ii')}V_{ii'}, \tag*{(1)}

where i<i′i<i' (no on-site terms). The spin ladder operators are Si±=(Six±iSiy)/2S_i^\pm=(S_i^x\pm iS_i^y)/\sqrt{2}. We further assume the absence of external magnetic fields or spontaneously broken symmetries (time-reversal and spin-rotation). None of these assumptions or the restriction to the model in Eq. (1) are fundamental for Dyn-HTE and can be relaxed in future extensions in parallel to the developments in the history of conventional HTE [2].

We assume thermal equilibrium at temperature T=1/βT=1/\beta (kB=ℏ=1k_B=\hbar=1). The density matrix is ρ=e−βH/Z\rho=e^{-\beta H}/Z with Z=tr⁡e−βHZ=\operatorname{tr}e^{-\beta H} and thermal averages of operators are given by

⟨…⟩≡tr⁡[…ρ]=tr⁡[…e−βH]/Z.(2)\langle\ldots\rangle\equiv\operatorname{tr}[\ldots\rho]=\operatorname{tr}[\ldots e^{-\beta H}]/Z. \tag*{(2)}

For most applications, we assume translational invariance and place the spins on a regular lattice at positions

ri=Ri+bi(3)\mathbf{r}_i=\mathbf{R}_i+\mathbf{b}_i \tag*{(3)}

where the label ii selects both the site Ri\mathbf{R}_i of a Bravais lattice and a basis vector bi∈{b(1),…,b(Nb)}\mathbf{b}_i\in\{\mathbf{b}^{(1)},\ldots,\mathbf{b}^{(N_b)}\}. We assume NcN_c Bravais lattice sites (with periodic boundary conditions) and NbN_b basis vectors so that N=NcNbN=N_cN_b.

Matsubara spin correlator and Dyn-HTE

The imaginary-frequency (Matsubara) spin-spin correlator is defined as [12]

Giizz(iνm)=T∫0βdτdτ′eiνm(τ−τ′)⟨TSiz(τ)Siz(τ′)⟩,(4)G_{ii}^{zz}(i\nu_m)=T\int_0^\beta d\tau d\tau' e^{i\nu_m(\tau-\tau')}\langle\mathcal{T}S_i^z(\tau)S_i^z(\tau')\rangle, \tag*{(4)}

where νm=2πTm\nu_m=2\pi Tm with m∈Zm\in\mathbb{Z} is a (bosonic) Matsubara frequency and Siz(τ)=eHτSize−HτS_i^z(\tau)=e^{H\tau}S_i^ze^{-H\tau} denotes imaginary time evolution. The operator T\mathcal{T} enforces imaginary time ordering, the operator with larger imaginary time argument is moved to the left. Note that the double integral is often replaced by a single integral over the difference τ−τ′\tau-\tau' but the form in Eq. (4) will prove useful later. Due to spin rotation symmetry in the Heisenberg Hamiltonian, Eq. (4) also determines correlations for all other spin-flavor combinations, Gαα=δαα′GzzG^{\alpha\alpha}=\delta_{\alpha\alpha'}G^{zz} and the zzzz superscript of GzzG^{zz} is dropped in the following to ease notation. The spatial Fourier transform of Eq. (4) reads

Gk(iνm)=1N∑i,i′e−ik⋅(ri−ri′)Gii′(iνm).(5)G_{\mathbf{k}}(i\nu_m)=\frac{1}{N}\sum_{i,i'}e^{-i\mathbf{k}\cdot(\mathbf{r}_i-\mathbf{r}_{i'})}G_{ii'}(i\nu_m). \tag*{(5)}

The symmetries of Gii′(iνm)G_{ii'}(i\nu_m) follow from its definition (4) and the Hamiltonian (1): Hermitian conjugation leads to a reality condition Gii′(iνm)∈RG_{ii'}(i\nu_m) \in\mathbb{R} and time-reversal symmetry ensures invariance under frequency flip Gii′(iνm)=Gii′(−iνm)G_{ii'}(i\nu_m) = G_{ii'}(-i\nu_m). This also implies symmetry under exchange of site-indices Gii′(iνm)=Gi′i(iνm)G_{ii'}(i\nu_m) = G_{i'i}(i\nu_m) which leads to

Gk(iνm)=G−k(iνm)∗.(6)G_k(i\nu_m) = G_{-k}(i\nu_m)^*. \tag*{(6)}

Hence, with inversion symmetry, Gk(iνm)∈RG_k(i\nu_m) \in\mathbb{R}.

As we show in Sec. IV, the Dyn-HTE of the Matsubara correlator (4) takes the form of a double expansion in x=J/Tx = J/T and inverse Matsubara frequency 1/νm1/\nu_m. To order nmax⁡n_{\max} in xx, Dyn-HTE reads

TGii′(iνm)={pii′(0)(x):m=0∑r=1rmax⁡pii′(2r)(x)(x2πm)2r:m≠0+O(xnmax⁡+1)(7)T G_{ii'}(i\nu_m) = \begin{cases} p_{ii'}^{(0)}(x) & :m=0 \\ \sum_{r=1}^{r_{\max}}p_{ii'}^{(2r)}(x)\left(\frac{x}{2\pi m}\right)^{2r} & :m\ne0 \end{cases} + O(x^{n_{\max}+1}) \tag*{(7)}

where rmax⁡=⌊nmax⁡/2⌋r_{\max} = \lfloor n_{\max}/2\rfloor is the integer floor of nmax⁡/2n_{\max}/2. The dimensionless polynomials pii′(2r)(x)p_{ii'}^{(2r)}(x) for are of degree nmax⁡−2rn_{\max}-2r with (real) rational coefficients, here r=0,1,2,…r=0,1,2,\ldots. Depending on the lattice and site-pair ii′ii', some of the polynomial coefficients can be zero. Note that the form (7) is consistent with the symmetries of the Matsubara correlator discussed above.

Our open-source numerical implementation [21] provides the exact polynomial coefficients in Eq. (7) as rational numbers for arbitrary lattices L\mathcal{L} and all site-pairs ii′ii' therein. Currently spin lengths S∈{1/2,1}S \in\{1/2,1\} and maximum expansion order nmax⁡=12n_{\max}=12 are available.

We conclude this section with a review of the physical content [12] of the Matsubara spin correlator (4). At zero frequency, m=0m=0, the Matsubara correlator yields the static susceptibility χii′\chi_{ii'}. The latter is defined as the (negative) linear isothermal response of zz-magnetization at site ii to a static local magnetic field perturbing the Hamiltonian HH via hi′zSi′zh_{i'}^zS_{i'}^z [40],

Gii′(iνm=0)=χii′=−∂hi′z⟨Siz⟩∣hi′z=0.(8)G_{ii'}(i\nu_m=0)=\chi_{ii'}=-\left.\partial_{h_{i'}^z}\langle S_i^z\rangle\right|_{h_{i'}^z=0}. \tag*{(8)}

At finite frequency m≠0m\ne0, the Matsubara correlator does not directly represent an observable physical quantity. It is mainly considered for the simplicity of the diagrammatic field-theoretical framework which can be employed for its computation [12]. Usually, numerical data for the Matsubara correlator Gii′(iνm)G_{ii'}(i\nu_m) evaluated at a finite set of frequencies νm\nu_m is linked to its real-frequency version by analytic continuation as discussed in Sec. I. Dyn-HTE allows to take a different and more stable route and never evaluates numerical values for Gii′(iνm)G_{ii'}(i\nu_m): As

we show in our companion letter [39], the high-frequency expansion coefficients from Eq. (7) are in direct correspondence to the short-time expansion coefficients of the real-time correlator,

pii′(2r)∼∂t2r−1⟨Siz(t)Si′z⟩∣t=0.(9)p_{ii'}^{(2r)} \sim\partial_t^{2r-1}\langle S_i^z(t)S_{i'}^z\rangle\big|_{t=0}. \tag*{(9)}

The object on the right-hand side of Eq. (9) can also be identified with the (real-frequency) moments of the spectral function [41]. Using equations of motion, the moments can be unraveled as equal-time correlators of 2r−12r-1 fold nested commutators, ∼⟨[…[[Siz,H],H],…,H]Si′z⟩\sim\langle[\ldots[[S_i^z,H],H],\ldots,H]S_{i'}^z\rangle. Even for moderate rr, these are hard to compute by standard means beyond the case of one dimension and T=∞T=\infty. This explains the value of Dyn-HTE’s access to (the HTE of) pii′(2r)p_{ii'}^{(2r)} for these cases. The full spectral function and dynamical structure factor can be reconstructed based on a moderate number of frequency moments via continued fraction representations [41, 42, 43, 44] [45]. In our companion letter [39] we discuss this in detail and show benchmarks and applications centered around the dynamical structure factor. In contrast, the remainder of the current work will be concerned with the efficient evaluation of the Dyn-HTE, i.e. the polynomials pii′(2r)(x)p_{ii'}^{(2r)}(x) in Eq. (7). Here we also present, benchmark and discuss results for the Matsubara correlator in imaginary frequency.

Derivation of Dyn-HTE

In this central technical section we derive the Dyn-HTE as in Eq. (7) and provide an efficient algorithm for its calculation. To start, we split the Hamiltonian (1) into a non-interacting part H0=0H_0=0 (vanishing in the absence of a magnetic field) and an interacting part VV (here the full HH) and define the dimensionless expansion coefficients cii′(n)(iνm)c_{ii'}^{(n)}(i\nu_m) of the Matsubara correlator via

TGii′(iνm)=∑n=0∞(−x)ncii′(n)(iνm),(x=JT).(10)T G_{ii'}(i\nu_m)=\sum_{n=0}^{\infty}(-x)^n c_{ii'}^{(n)}(i\nu_m),\qquad\left(x=\frac{J}{T}\right). \tag*{(10)}

The lowest-order n=0n=0 term represents the Curie susceptibility of a free spin which is local and static (non-zero only for m=0m=0), cii′(0)(iνm)=δii′δ0mS(S+1)/3c_{ii'}^{(0)}(i\nu_m)=\delta_{ii'}\delta_{0m}S(S+1)/3.

Expansion coefficients

General perturbation theory for the Matsubara correlator [12] provides an expression for the remaining expansion coefficients in Eq. (10) at order n=1,2,…n=1,2,\ldots

cii′(n)(iνm)=Tn+2n!∑b1,…,bn∫0βeiνm(τ−τ′)dτ1…dτndτdτ′⟨TVb1(τ1)…Vbn(τn)Siz(τ)Si′z(τ′)⟩0V−con.(11)c_{ii'}^{(n)}(i\nu_m)=\frac{T^{n+2}}{n!}\sum_{b_1,\ldots,b_n}\int_0^\beta e^{i\nu_m(\tau-\tau')}d\tau_1\ldots d\tau_n d\tau d\tau'\left\langle\mathcal{T}V_{b_1}(\tau_1)\ldots V_{b_n}(\tau_n)S_i^z(\tau)S_{i'}^z(\tau')\right\rangle_0^{\mathrm{V-con}}. \tag*{(11)}

The sum is over bonds bl=(ilil′)b_l=(i_l i_l') and operator time arguments as in Vb1(τ1)V_{b_1}(\tau_1) now refer to (interaction-picture) evolution with respect to H0H_0 only. This is trivial for our case, H0=0H_0=0, e.g. Vb1(τ1)=Vb1V_{b_1}(\tau_1)=V_{b_1}. Nevertheless we must keep the time arguments because by virtue of T\mathcal{T} they determine the ordering of the string of generally non-commuting operators. These operators are thus distinguishable regardless of the chosen bonds blb_l. The ⟨…⟩0\langle\ldots\rangle_0 denotes a thermal average with respect to Hamiltonian H0H_0. Since H0=0H_0=0, the former is simply an (effectively) infinite temperature average or normalized trace over the whole NN-site spin Hilbert-space, ⟨…⟩0=tr⁡[…]/(2S+1)N\langle\ldots\rangle_0=\operatorname{tr}[\ldots]/(2S+1)^N which factorizes according to site-index, e.g. ⟨S1zS2+S2zS2−⟩0=⟨S1zS1z⟩0⟨S2+S2−⟩0\langle S_1^zS_2^+S_2^zS_2^-\rangle_0=\langle S_1^zS_1^z\rangle_0\langle S_2^+S_2^-\rangle_0 with the on-site operator order maintained.

The superscript “V − con” on the right-hand side of Eq. (11) refers to the VV-connected correlator which contains all contributions to order VnV^n, also those from the expansion of the partition function ZZ in the denominator on the right-hand side of Eq. (2). Following Ref. 46, it can be written via recursive subtractions from the ”full” correlator [without V − con, c.f. (2)],

⟨TVb1(τ1)…Vbn(τn)Siz(τ)Si′z(τ′)⟩0V−con=⟨TVb1(τ1)…Vbn(τn)Siz(τ)Si′z(τ′)⟩0−∑S⊂{1,…,n}⟨T[∏k∈SVbk(τk)]Siz(τ)Si′z(τ′)⟩0V−con⟨T∏l∈{1,…,n}∖SVbl(τl)⟩0,(12)\left\langle\mathcal{T}V_{b_1}(\tau_1)\ldots V_{b_n}(\tau_n)S_i^z(\tau)S_{i'}^z(\tau')\right\rangle_0^{\mathrm{V-con}}=\left\langle\mathcal{T}V_{b_1}(\tau_1)\ldots V_{b_n}(\tau_n)S_i^z(\tau)S_{i'}^z(\tau')\right\rangle_0-\sum_{S\subset\{1,\ldots,n\}}\left\langle\mathcal{T}\left[\prod_{k\in S}V_{b_k}(\tau_k)\right]S_i^z(\tau)S_{i'}^z(\tau')\right\rangle_0^{\mathrm{V-con}}\left\langle\mathcal{T}\prod_{l\in\{1,\ldots,n\}\setminus S}V_{b_l}(\tau_l)\right\rangle_0, \tag*{(12)}

where the sum is over true subsets SS of the set {1,…,n}\{1,\ldots,n\} (including the empty set) and the recursion terminates at ⟨TSiz(τ)Si′z(τ′)⟩0V−con≡⟨TSiz(τ)Si′z(τ′)⟩0\langle\mathcal{T}S_i^z(\tau)S_{i'}^z(\tau')\rangle_0^{\mathrm{V-con}}\equiv\langle\mathcal{T}S_i^z(\tau)S_{i'}^z(\tau')\rangle_0. We insert Eq. (12) in Eq. (11) and obtain

cii′(n)(iνm)=∑b1,…,bnTn+2n!∫0βeiνm(τ−τ′)dτ1…dτndτdτ′×{⟨TVb1(τ1)…Vbn(τn)Siz(τ)Si′z(τ′)⟩0−∑S⊂{1,…,n}⟨T[∏k∈SVbk(τk)]Siz(τ)Si′z(τ′)⟩0V−con⟨T∏l∈{1,…,n}∖SVbl(τl)⟩0}.(13)\begin{aligned} c_{ii'}^{(n)}(i\nu_m)=\sum_{b_1,\ldots,b_n}\frac{T^{n+2}}{n!}\int_0^\beta e^{i\nu_m(\tau-\tau')}d\tau_1\ldots d\tau_n d\tau d\tau' \\ &\times\left\{\left\langle\mathcal{T}V_{b_1}(\tau_1)\ldots V_{b_n}(\tau_n)S_i^z(\tau)S_{i'}^z(\tau')\right\rangle_0-\sum_{S\subset\{1,\ldots,n\}}\left\langle\mathcal{T}\left[\prod_{k\in S}V_{b_k}(\tau_k)\right]S_i^z(\tau)S_{i'}^z(\tau')\right\rangle_0^{\mathrm{V-con}}\left\langle\mathcal{T}\prod_{l\in\{1,\ldots,n\}\setminus S}V_{b_l}(\tau_l)\right\rangle_0\right\}. \tag*{(13)} \end{aligned}

Graph-based evaluation of bond-sums

Compared to the established HTE for equal-time spin-spin correlators ⟨SizSi′z⟩\langle S_i^zS_{i'}^z\rangle, the new and challenging aspect in Eq. (13) are the n+2n+2-dimensional imaginary time integrals. We relegate the integral evaluation to section IV D. Here we first focus on the bond-sums in Eq. (13), which we expose by summarizing the rest as follows,

cii′(n)(iνm)≡∑b1,…,bnFii′(n)(b1,…,bn;iνm).(14)c_{ii'}^{(n)}(i\nu_m)\equiv\sum_{b_1,\ldots,b_n}F_{ii'}^{(n)}(b_1,\ldots,b_n;i\nu_m). \tag*{(14)}

Efficient evaluation strategies for these bond-sums are well developed in the literature on conventional HTE [2]. The following discussion of the necessary elements of these strategies is self-contained and does not assume any prior knowledge. In a nutshell, our plan is to evaluate the expansion of the Matsubara correlator not directly for a full lattice L\mathcal{L}, but first for individual lattice snippets with only nn bonds which are called graphs, g(n)g^{(n)}. These are then added up as they fit into the full lattice. Formally speaking, we re-organize the sum over (potentially identical) lattice bonds ∑b1,…,bn\sum_{b_1,\ldots,b_n} as a sum over evaluations of Fii′(n)F_{ii'}^{(n)} on these graphs,

cii′(n)(iνm)=∑g(n)e(L,i,i′,g(n))cg(n)(iνm),(15)c_{ii'}^{(n)}(i\nu_m)=\sum_{g^{(n)}}e(\mathcal{L},i,i',g^{(n)})c_{g^{(n)}}(i\nu_m), \tag*{(15)}
cg(n)(iνm)=∑{e1,…,en}→{e~1,…,e~n}Fjj′(n)(e~1,…,e~n;iνm).(16)c_{g^{(n)}}(i\nu_m)=\sum_{\{e_1,\ldots,e_n\}\to\{\tilde e_1,\ldots,\tilde e_n\}}F_{jj'}^{(n)}(\tilde e_1,\ldots,\tilde e_n;i\nu_m). \tag*{(16)}

A (multi-)graph of order nn denoted by g(n)g^{(n)} can be thought of as a lattice-snippet with nn bonds, see Figure 1. Formally, it is defined by the multi-set of nn (not necessarily distinct) edges {e1,…,en}\{e_1,\ldots,e_n\} (thick black lines) connecting its arbitrarily numbered vertices (blue circles) which contain the terminal vertices j,j′j,j' (blue circles highlighted by attached red-square flags). The terminal vertices indicate the position of the external operators SjzS_j^z and Sj′zS_{j'}^z. Different graphs are distinguished by graph topology, Figure 1 shows all topological distinct graphs gt(n)g_t^{(n)}, labeled by t=1,2,…t=1,2,\ldots with up to n=3n=3 edges that are required for our purpose. As we motivate later, two graphs only differing by the exchange of terminals j↔j′j\leftrightarrow j' are considered topologically equivalent.

All required graphs with terminal vertices and terminal flags

Figure 1. All required graphs gt(n)g_t^{(n)} with n=0,1,2,3n = 0,1,2,3 edges (black lines) and arbitrary numbered vertices (blue circles). Terminal vertices j,j′j,j' (also blue circles) are indicated by their red-square terminal flags attached with thin gray lines. The symmetry factor is s[gt(n)]=1s[g_t^{(n)}] = 1 for all graphs shown except for (n,t)=(3,5)(n,t)=(3,5) where it is two (exchange of vertices 2↔32 \leftrightarrow3).

Having defined the graph g(n)g^{(n)} as a lattice snippet, we next need to place it into the full lattice L\mathcal{L}. This is done at the embedding step, formally expressed in Eq. (15). Here the embedding factor e(L,i,i′,g(n))∈{0,1,2,…}e(\mathcal{L},i,i',g^{(n)})\in\{0,1,2,\ldots\} counts the number of sub-graph isomorphisms from the graph g(n)g^{(n)} (with edge-multiplicities ignored) to the lattice L\mathcal{L} divided by the graph’s symmetry factor s[g(n)]∈{1,2,…}s[g^{(n)}]\in\{1,2,\ldots\} to avoid overcounting due to the arbitrary vertex numbering. The symmetry factor is the number of graph isomorphisms that keep the terminal vertices invariant and respect the edge-multiplicities, see the caption of Figure 1 for examples for s[g(n)]s[g^{(n)}].

For the embedding g(n)→Lg^{(n)}\to\mathcal{L} we require the assignments (i,i′)→(j,j′)(i,i')\to(j,j') or (i,i′)→(j′,j)(i,i')\to(j',j) to match the positions of the external operators. Taking into account also the latter assignment (i,i′)→(j′,j)(i,i')\rightarrow(j',j) allows us to skip graphs which are connected to another graph in the list by an exchange of j↔j′j\leftrightarrow j'. This assignment is not counted if gt(n)g_t^{(n)} is mapped to itself under j↔j′j\leftrightarrow j'. As an example, embedding factors for a number of simple lattices are given in Tab. I. For two- and in particular three-dimensional lattices the embedding step can be numerically demanding. To ensure efficiency, our numerical implementation [21] builds on lattice generating functionalities provided in the software package SpinMC.jl [47], relies on space-group symmetries for speed-up and uses advanced graph-theoretical algorithms [48, 49].

ggg1(0)g_1^{(0)}g1(1)g_1^{(1)}g1(2)g_1^{(2)}g2(2)g_2^{(2)}g3(2)g_3^{(2)}g1(3)g_1^{(3)}g2(3)g_2^{(3)}g3(3)g_3^{(3)}g4(3)g_4^{(3)}g5(3)g_5^{(3)}g6(3)g_6^{(3)}g7(3)g_7^{(3)}
dimer loc111
dimer nn111
trimer loc1221
trimer nn1111221
chain loc122
chain nn1112
chain nnn12
sq. lat. loc144
sq. lat. nn11162
tri. lat. loc1666
tri. lat. nn112110424

Table 1. Embedding factor e(L,i,i′,gt(n))e(\mathcal{L},i,i',g_t^{(n)}) for various lattices L\mathcal{L} and various separations between external sites i,i′i,i'. Here ’loc’ means local (i=i′i=i') while ’(n)nn’ stands for (next-)nearest neighbor. A missing entry means e(L,i,i′,gt(n))=0e(\mathcal{L},i,i',g_t^{(n)})=0.

Note that only graphs g(n)g^{(n)} with a single connected component provide a non-zero contribution to cii′(iνm)c_{ii'}(i\nu_m). Indeed, if there are multiple connected components, the total contribution will vanish in the course of the recursion involved in Eq. (13). Likewise, we skip graphs with (generalized) leaves, which are parts connected to the spline of the graph (the part with the terminal vertices) by a single edge. Such graphs would only contribute for finite field hh which we set to zero in the model (1). Similar considerations apply for vacuum graphs to be defined below.

Following these considerations, we have created lists of all required graphs g(n)g^{(n)} [with a potential non-zero graph-evaluation cg(n)(iνm)c_{g^{(n)}}(i\nu_m)], for n=0,1,…,12n=0,1,\ldots,12. The number of graphs per order n≳6n\gtrsim6 is roughly a factor 4 larger than for the previous order. At our highest order n=nmax⁡=12n=n_{\max}=12 we need to consider 1273854 graphs g(12)g^{(12)}.

Finally we need to consider the quantity which we wanted to evaluate on the lattice snippets defined by a selection of edges g(n)={e1,…,en}g^{(n)}=\{e_1,\ldots,e_n\}: The HTE of the Matsubara correlator. We call this HTE the graph evaluation cg(n)(iνm)c_{g^{(n)}}(i\nu_m). In Eq. (15) the correlator expansion on the full lattice is obtained by a sum over all graph evaluations weighted with the embedding factors. In Eq. (16), the individual graph evaluations are given by a sum ∑{e1,…,en}→{e~1,…,e~n}\sum_{\{e_1,\ldots,e_n\}\rightarrow\{\tilde e_1,\ldots,\tilde e_n\}} which associates the multi-set of the graph’s edges to the time-ordered bond operators in Fjj′(n)(e~1,…,e~n;iνm)F_{jj'}^{(n)}(\tilde e_1,\ldots,\tilde e_n;i\nu_m) in all n!/d[g(n)]n!/d[g^{(n)}] different ways. Here d[g(n)]d[g^{(n)}] yields the graph degeneracy which is the product of factorials of all edge weights.

As in the case of conventional HTE [2], the power of our numerical implementation of Dyn-HTE [21] rests on the fact that the lattice-specific embedding factors e(L,i,i′,g(n))e(\mathcal{L},i,i',g^{(n)}) can be quickly obtained from advanced graph theoretical algorithms [48, 49], whereas the numerically costly graph evaluations cg(n)(iνm)c_{g^{(n)}}(i\nu_m) can be pre-computed once and for all. As we will show in the following, they take the form

cg(n)(iνm)=δ0,mct;0(n)+∑l=1nct;l(n)Δ2πml,(17)c_{g^{(n)}}(i\nu_m)=\delta_{0,m}c_{t;0}^{(n)}+\sum_{l=1}^{n}c_{t;l}^{(n)}\Delta_{2\pi m}^{l}, \tag*{(17)}

where Δ2πm≡(1−δ0,m)/(2πm)\Delta_{2\pi m}\equiv(1-\delta_{0,m})/(2\pi m) is the inverse of dimensionless (non-zero) Matsubara frequencies. As we argue later, after the embedding, all odd ll vanish on the right-hand side of Eq. (15) and are thus not computed in the first place. The ct;l(n)c_{t;l}^{(n)} for l=0,2,…,rmax⁡l = 0,2,\ldots,r_{\max} are rational numbers and together with Eq. (10) they give rise to the final form (7). In Tab. II we provide the values for the evaluations ct;l(n)c_{t;l}^{(n)} for all graphs g(n)g^{(n)} with n≤3n \leq3 shown in Figure 1. The remainder of this section is concerned with the actual calculation of these graph evaluations.

(n,t)(n,t)(0,1)(0,1)(1,1)(1,1)(2,1)(2,1)(2,2)(2,2)(2,3)(2,3)(3,1)(3,1)
ct;0(n)c_{t;0}^{(n)}+1/4+1/4+1/16+1/16−1/96-1/96−1/192-1/192+1/64+1/64+1/384+1/384
ct;2(n)c_{t;2}^{(n)}0000+1/8+1/8−1/8-1/800−1/32-1/32
(n,t)(n,t)(3,2)(3,2)(3,3)(3,3)(3,4)(3,4)(3,5)(3,5)(3,6)(3,6)(3,7)(3,7)
ct;0(n)c_{t;0}^{(n)}−1/256-1/256−1/384-1/384−1/768-1/768−1/192-1/19200+1/256+1/256
ct;2(n)c_{t;2}^{(n)}+1/32+1/320000+1/16+1/16−1/32-1/3200

Table 2. Results for the coefficients ct;l(n)c_{t;l}^{(n)} in the polynomial (17) of the evaluation of all graphs gt(n)g_t^{(n)} in Figure 1 with n=0,1,2,3n = 0,1,2,3 edges.

Recursive sub-graph subtractions

For the (graph-)evaluation of the dynamic spin correlator on graph g(n)g^{(n)} denoted by cg(n)(iνm)c_{g^{(n)}}(i\nu_m) in Eq. (16) we are required to assign the multi-set of the graph’s edges {e1,…,en}\{e_1,\ldots,e_n\} to the bond-indices of the operators {Ve~1(τ1),…,Ve~n(τn)}\{V_{\widetilde e_1}(\tau_1),\ldots,V_{\widetilde e_n}(\tau_n)\} according to {e1,…,en}→{e~1,…,e~n}\{e_1,\ldots,e_n\}\to\{\widetilde e_1,\ldots,\widetilde e_n\} in all possible ways. The operators are distinguishable by their time arguments even for identical edges. However, recall the iterative definition of the V − con correlator in Eq. (13) which carries over to the graph evaluation as follows:

cg(n)(iνm)=cg(n)[full](iνm)−cg(n)[sub](iνm).(18)c_{g^{(n)}}(i\nu_m)=c_{g^{(n)}}^{[\mathrm{full}]}(i\nu_m)-c_{g^{(n)}}^{[\mathrm{sub}]}(i\nu_m). \tag*{(18)}

For the first “full” term, all n!/d[g(n)]n!/d[g^{(n)}] assignments are equivalent since the times τ1,…,τn\tau_1,\ldots,\tau_n can be relabeled and the order of operators behind T\mathcal{T} does not matter. We obtain

cg(n)[full](iνm)=Tn+2d[g(n)]∫0βeiνm(τ−τ′)dτ1…dτndτdτ′×⟨TVe1(τ1)…Ven(τn)Sjz(τ)Sj′z(τ′)⟩0V−con.(19)c_{g^{(n)}}^{[\mathrm{full}]}(i\nu_m)=\frac{T^{n+2}}{d[g^{(n)}]}\int_0^\beta e^{i\nu_m(\tau-\tau')}d\tau_1\ldots d\tau_n d\tau d\tau' \times\left\langle\mathcal{T}V_{e_1}(\tau_1)\ldots V_{e_n}(\tau_n)S_j^z(\tau)S_{j'}^z(\tau')\right\rangle_0^{\mathrm{V-con}}. \tag*{(19)}

which we will evaluate further in the next subsection.

For the second “subtracted” term in Eq. (18), we rewrite the ∑S⊊{1,…,n}\sum_{S\subsetneq\{1,\ldots,n\}} in Eq. (13) as a sum over sub-graphs g(k)⊊g(n)g^{(k)}\subsetneq g^{(n)} with k<nk<n edges. The subgraph g(k)={e1′,…,ek′}g^{(k)}=\{e'_1,\ldots,e'_k\} still needs to be connected and must contain the terminal vertices j,j′j,j'. This remaining edges g(n)\g(k)g^{(n)}\backslash g^{(k)} form a vacuum graph without external vertices which is possibly disconnected. There can be more than one ways to get to topologically equivalent g(k)g^{(k)}, this multiplicity factor is denoted by f(g(n),g(k))∈{1,2,…}f(g^{(n)},g^{(k)})\in\{1,2,\ldots\}. We

refer to Figure 2 for an example listing all possible non-zero subtractions from a particular graph g(6)g^{(6)} with n=6n=6 edges (left) shown in the 2nd and 3rd column together with their factors f(g(n),g(k))f(g^{(n)},g^{(k)}) and the g(k)g^{(k)} (top), the vacuum graph g(n)\g(k)g^{(n)}\backslash g^{(k)} is shown in the bottom. Finally, in the second line of Eq. (13) there are (nk)\binom{n}{k} possibilities to distribute the imaginary time labels between the first and second average. Hence we obtain

Subtractions from a graph

Figure 2. Subtractions from a graph g(6)g^{(6)} shown in the upper left. All connected sub-graphs g(k)g^{(k)} with k<6k<6 and non-zero graph evaluation are shown in the top row, second to last column. The vacuum graphs g(6)\g(k)≡vg^{(6)}\backslash g^{(k)}\equiv v are shown in the bottom row. These vacuum graphs are possibly disconnected, see third column. The multiplicity factors ff for the particular subtraction are also given.

cg(n)[sub](iνm)=(nk)n!∑g(k)⊆g(n)f(g(n),g(k))Tn+2∫0βeiνm(τ−τ′)dτ1…dτndτdτ′×∑g(k)={e1′,…,ek′}→{e~1′,…,e~k′}⟨TVe~1′(τ1)…Ve~k′(τk)Sjz(τ)Sj′z(τ′)⟩0V−con×∑g(n)\g(k)={ek+1′,…,en′}→{e~k+1′,…,e~n′}⟨TVe~k+1′(τk+1)…Ve~n′(τn)⟩0.(20)\begin{aligned} c_{g^{(n)}}^{[\mathrm{sub}]}(i\nu_m) ={}&\frac{\binom{n}{k}}{n!}\sum_{g^{(k)}\subseteq g^{(n)}}f(g^{(n)},g^{(k)})T^{n+2}\int_0^\beta e^{i\nu_m(\tau-\tau')}d\tau_1\ldots d\tau_n d\tau d\tau'\\ &\times\sum_{g^{(k)}=\{e'_1,\ldots,e'_k\}\to\{\widetilde e'_1,\ldots,\widetilde e'_k\}} \left\langle\mathcal{T}V_{\widetilde e'_1}(\tau_1)\ldots V_{\widetilde e'_k}(\tau_k)S_j^z(\tau)S_{j'}^z(\tau')\right\rangle_0^{\mathrm{V-con}}\\ &\times\sum_{g^{(n)}\backslash g^{(k)}=\{e'_{k+1},\ldots,e'_n\}\to\{\widetilde e'_{k+1},\ldots,\widetilde e'_n\}} \left\langle\mathcal{T}V_{\widetilde e'_{k+1}}(\tau_{k+1})\ldots V_{\widetilde e'_n}(\tau_n)\right\rangle_0. \tag*{(20)} \end{aligned}

The sum in the penultimate line assigns the multi-set of graph g(k)g^{(k)}’s edges {e1′,…,ek′}\{e'_1,\ldots,e'_k\} in all possible ways to the bond-indices of the operators Ve~1′(τ1),…,Ve~k′(τk)V_{\widetilde e'_1}(\tau_1),\ldots,V_{\widetilde e'_k}(\tau_k) and similar in the last line. We now resolve these assignment-sums.

The penultimate line in Eq. (20) appears in the graph-resolved version of Eq. (11) (which only sums over the edges of graph g(k)g^{(k)} instead of all bonds b1,…,bnb_1,\ldots,b_n of the lattice). Hence we obtain a recursive equation for the graph-evaluation cg(n)(iνm)c_{g^{(n)}}(i\nu_m)

cg(n)[sub](iνm)=∑g(k)⊆g(n)f(g(n),g(k))cg(k)(iνm)⟨g(n)\g(k)⟩0.(21)c_{g^{(n)}}^{[\mathrm{sub}]}(i\nu_m)=\sum_{g^{(k)}\subseteq g^{(n)}} f(g^{(n)},g^{(k)})c_{g^{(k)}}(i\nu_m)\left\langle g^{(n)}\backslash g^{(k)}\right\rangle_0 . \tag*{(21)}

The last term is the evaluation of the vacuum graph g(n)\g(k)g^{(n)}\backslash g^{(k)} defined from the remaining terms as

⟨g(n)\g(k)⟩0=Tn−k(n−k)!∫0βdτk+1…dτn∑g(n)\g(k)={ek+1′,…,en′}→{e~k+1′,…,e~n′}⟨TVe~k+1′(τk+1)…Ve~n′(τn)⟩0=1d[g(n)\g(k)](n−k)!∑p∈Sn−k⟨Vek+p1′′…Vek+pn−k′′⟩0.(22)\begin{aligned} \left\langle g^{(n)}\backslash g^{(k)}\right\rangle_0= \frac{T^{n-k}}{(n-k)!}\int_0^\beta\mathrm{d}\tau_{k+1}\ldots\mathrm{d}\tau_n \sum_{g^{(n)}\backslash g^{(k)}=\{e'_{k+1},\ldots,e'_n\}\to\{\tilde e'_{k+1},\ldots,\tilde e'_n\}} \left\langle\mathcal{T}V_{\tilde e'_{k+1}}(\tau_{k+1})\ldots V_{\tilde e'_n}(\tau_n)\right\rangle_0 \\ &=\frac{1}{d[g^{(n)}\backslash g^{(k)}](n-k)!}\sum_{p\in S_{n-k}}\left\langle V'_{e'_{k+p_1}}\ldots V'_{e'_{k+p_{n-k}}}\right\rangle_0 . \tag*{(22)} \end{aligned}

In the last step, time-integrals together with the time-ordering T\mathcal{T} lead to the sum over all permutations pp of the numbers 1,2,…,n−k1,2,\ldots,n-k, similar to vacuum graph evaluation for conventional HTE. For disconnected vacuum graphs g(n)\g(k)≡vg^{(n)}\backslash g^{(k)}\equiv v which separate into two connected components v=v1∪ ⁣⋅ ⁣v2v=v_1\cup\!\cdot\! v_2 we have ⟨v⟩0=⟨v1∪ ⁣⋅ ⁣v2⟩0=⟨v1⟩0⟨v2⟩0\langle v\rangle_0=\langle v_1\cup\!\cdot\! v_2\rangle_0=\langle v_1\rangle_0\langle v_2\rangle_0 and this generalizes to an arbitrary number of connected components.

In summary, the recursive formulation necessitated by the removal of V-disconnected bond operator configurations in the definition (12) works as follows: Once all the “full” graph evaluations (19) at order nn and vacuum graph evaluations (22) of orders smaller or equal to n−1n-1 have been calculated, Eqns. (18) and (21) can be used to find the nn-th order graph evaluations via the recursion

cg(n)(iνm)=cg(n)[full](iνm)−∑g(k)⊊g(n)f(g(n),g(k))cg(k)(iνm)⟨g(n)\g(k)⟩0.(23)c_{g^{(n)}}(i\nu_m)=c_{g^{(n)}}^{[\mathrm{full}]}(i\nu_m) -\sum_{g^{(k)}\subsetneq g^{(n)}}f(g^{(n)},g^{(k)})c_{g^{(k)}}(i\nu_m)\left\langle g^{(n)}\backslash g^{(k)}\right\rangle_0 . \tag*{(23)}

This expression is conceptually analogous to a formula popularized in the context of high-order perturbation theory for the fermionic Hubbard model in the diagrammatic Monte-Carlo approach (there an expansion in the Hubbard interaction UU is used). In this context, Eq. (23) is known as the connected determinant formula and was put forward by Rossi [46]. In the fermionic problem, the averages ⟨T…⟩0\langle\mathcal{T}\ldots\rangle_0 corresponding to our “full” graph evaluations (19) are taken with respect to a non-interacting fermionic (hopping) Hamiltonian H0H_0 and can thus be obtained via Wick’s theorem [12]. This amounts to a determinant evaluation that is numerically efficient. For canonical bosons, the determinant is replaced by a permanent. Then the time integrals are performed stochastically via Markov-chain Monte-Carlo.

Crucially, as we are here dealing with spin operators, Wick’s theorem is not applicable and the first term on the right-hand side of Eq. (23), cg(n)[full](iνm)c_{g^{(n)}}^{[\mathrm{full}]}(i\nu_m), given in Eq. (19), cannot be written as a determinant or permanent. On the other hand, it turns out that the time-dependence of the integrand in our case is much simpler than that encountered in the Hubbard model. These observations lead to a dedicated evaluation strategy which we discuss in the next subsection.

Kernel trick for imaginary-time integrals

The central quantity left to be computed is cg(n)[full](iνm)c_{g^{(n)}}^{[\mathrm{full}]}(i\nu_m) in Eq. (19), the full part of the graph evaluation. Its n+2n+2-fold time integral can be seen as a n+2n+2 dimensional temporal Fourier transform of an (imaginary-)time-ordered n+2n+2-point correlator, but with nn Matsubara frequencies set to zero. Recently, in Ref. 22 closed-form analytic expressions for these Fourier transforms have been found for arbitrary nn, generalizing the Lehmann representation [12] beyond 2-point functions [50] (see also [51] for an earlier solution of the nested time integrals derived in the context of perturbation theory for the Hubbard model). With the help of these “kernel functions”, the Fourier transform of a time-ordered n+2n+2-point correlator is written exactly in terms of many-body eigenstates, eigenenergies and matrix elements of the operators involved in the correlation function [22].

In a general case the Hamiltonian cannot be diagonalized and the “kernel trick” is typically not practically applicable. However, in the context of spins, as in Eq. (19), the quantum system of interest is a set of NN non-interacting spins (H0=0H_0=0) with trivial product eigenstates and all eigenenergies vanishing. This allows for a tremendous simplification of the analytical expression for Eq. (19) obtained from the kernel trick. For details about the calculation, we refer to our recent work on a complementary spin-diagrammatic scheme tailored for spin systems with close-to-mean-field physics in Ref. 52. We obtain our main exact and compact result

cg(n)[full](iνm)=1d[g(n)]∑p∈Sn+2K~n+2(pn+1,pn+2;m)⟨PVe1[p1]…Ven[pn]Sjz[pn+1]Sj′z[pn+2]⟩0.(24)c_{g^{(n)}}^{[\mathrm{full}]}(i\nu_m)=\frac{1}{d[g^{(n)}]}\sum_{p\in S_{n+2}}\widetilde K_{n+2}(p_{n+1},p_{n+2};m)\left\langle\mathcal{P}V_{e_1}^{[p_1]}\ldots V_{e_n}^{[p_n]}S_{j}^{z[p_{n+1}]}S_{j'}^{z[p_{n+2}]}\right\rangle_0 . \tag*{(24)}

where the index ordering operator P\mathcal{P} sorts the operator string according to increasing superscript index (in square-brackets). For example, in the case n=2n=2 and for a particular permutation pp it acts as PVe1[4]Ve2[1]Sjz[3]Sj′z[2]=Ve2Sj′zSjzVe1\mathcal{P}V_{e_1}^{[4]}V_{e_2}^{[1]}S_j^{z[3]}S_{j'}^{z[2]}=V_{e_2}S_{j'}^zS_j^zV_{e_1}.

The dimensionless kernel function K~n+2(pn+1,pn+2;m)\widetilde K_{n+2}(p_{n+1},p_{n+2};m) carries the dependence on external frequency via the Matsubara integer mm and depends on the positions pn+1,pn+2p_{n+1},p_{n+2} of the external operators in the operator string. It relates to the general Kernel function Kn+2K_{n+2} of Ref. 22 with n+2n+2 complex arguments Ω1,…,Ωn+2\Omega_1,\ldots,\Omega_{n+2} as

K~n+2(a+,a−;m)≡Tn+2Kn+2(0,…,0,+iνm,0,…,0,−iνm,0,…,0),(25)\widetilde{K}_{n+2}(a_+,a_-;m) \equiv T^{n+2}K_{n+2}(0,\ldots,0,+i\nu_m,0,\ldots,0,-i\nu_m,0,\ldots,0), \tag*{(25)}

where a±=1,2,…,n+2a_\pm=1,2,\ldots,n+2 mark the position of the frequency entry iν±mi\nu_{\pm m} (of arbitrary relative order a+⋛a−a_+\gtreqless a_-). In Ref. 52, we showed that

K~n+2(a+,a−;m)={1(n+2)!,m=0,(−1)∣a+−a−∣∑l=∣a+−a−∣n+2−min⁡(a+,a−)[Δ2πmi]l(sgn⁡[a−−a+])l(n+2−l)!(l−1∣a+−a−∣−1),m≠0.(26)\widetilde{K}_{n+2}(a_+,a_-;m)= \begin{cases} \frac{1}{(n+2)!}, & m=0,\\ (-1)^{|a_+-a_-|}\sum_{l=|a_+-a_-|}^{n+2-\min(a_+,a_-)} \left[\Delta_{2\pi mi}\right]^l \frac{\left(\operatorname{sgn}[a_--a_+]\right)^l}{(n+2-l)!} \binom{l-1}{|a_+-a_-|-1}, & m\ne0. \end{cases} \tag*{(26)}

which has the symmetry

K~n+2(a+,a−;m)=K~n+2(a−,a+;−m)(27)\widetilde{K}_{n+2}(a_+,a_-;m)=\widetilde{K}_{n+2}(a_-,a_+;-m) \tag*{(27)}

required by the definition (25). We remark that a somewhat similar treatment for a high-temperature expansion in the context of the fermionic Hubbard model (expansion in t/Tt/T with tt the hopping-integral) has been put forward in Ref. 44 but to lower orders and also without the geometric flexibility of the graph-based approach.

In passing, we note a variety of recently suggested approaches that have been developed to solve Matsubara sums (or, equivalently, imaginary-time integrals) that appear in bare or renormalized perturbation theory analytically. These include the discrete-Lehmann representation [53], the intermediate representation [54] and algorithmic Matsubara integration [55, 56]. However, none of these approaches are applicable to perturbation theory diagrams with general bosonic nn-point correlators as building blocks, like in the present expansion (which can be understood to be of strong-coupling type in Hubbard-model parlance).

Computational aspects for ``full" graph evaluations

We finally collect a few algorithmic tricks for efficient numerical evaluation of Eq. (24) which turns out to be the performance bottleneck in the calculation of the complete graph evaluation (23). This can be skipped by the reader unless interested in details of the numerical implementation or the reason for the absence of odd ll in the sum (17). With these tricks in place, the evaluations for all graphs up to and including order nmax⁡=12n_{\max}=12 required on the order of 500.000 core hours for spin length S=1/2S=1/2.

External operator positions: The kernel function in Eq. (24) depends only on the positions pn+1p_{n+1} and pn+2p_{n+2} of the external operators in the operator string, but not on the ordering of the edge operators VeV_e. We thus simplify Eq. (24) by splitting the sum over permutations pp into a sum over permutations of the nn edge operator positions and insert the external operator SjzS_j^z after aa edge operators and the other external operator Sj′zS_{j'}^z after another δa\delta a edge operators (and similar for the case that Sj′zS_{j'}^z appears left of SjzS_j^z),

cg(n)[full](iνm)=∑p∈Sn′∑a=0n∑δa=0n−aK~n+2(a+1,a+δa+2;m)⟨PVe1[p1]…Ven[pn]⟵a,δaSjzSj′z⟩0+∑p∈Sn′∑a=0n∑δa=0n−aK~n+2(a+δa+2,a+1;m)⟨PVe1[p1]…Ven[pn]⟵a,δaSj′zSjz⟩0.(28)c_{g(n)}^{[full]}(i\nu_m)= \sum_{p\in S_n}'\sum_{a=0}^{n}\sum_{\delta a=0}^{n-a} \widetilde{K}_{n+2}(a+1,a+\delta a+2;m) \left\langle PV_{e_1}^{[p_1]}\ldots V_{e_n}^{[p_n]} \overset{a,\delta a}{\longleftarrow}S_j^zS_{j'}^z\right\rangle_0 + \sum_{p\in S_n}'\sum_{a=0}^{n}\sum_{\delta a=0}^{n-a} \widetilde{K}_{n+2}(a+\delta a+2,a+1;m) \left\langle PV_{e_1}^{[p_1]}\ldots V_{e_n}^{[p_n]} \overset{a,\delta a}{\longleftarrow}S_{j'}^zS_j^z\right\rangle_0 . \tag*{(28)}

The kernels can be precomputed for any combination of a,δaa,\delta a and are only multiplied with the accumulated traces (∼⟨…⟩0\sim\langle\ldots\rangle_0) in the very end. In addition, the primed sum ∑p∈Sn′\sum_{p\in S_n}' excludes equivalent permutations of the multi-set {e1,…,en}\{e_1,\ldots,e_n\} which in Eq. (24) were cancelled by the factor d[gn]d[g_n].

Even kernel function: For graphs with j=j′j=j' the two traces in the first and second row of Eq. (28) are equivalent,

cg(n)[full](iνm)=∑a=0n∑δa=0n−a2K~n+2(even)(a+1,a+δa+2;m)∑p∈Sn′⟨PVe1[p1]…Ven[pn]⟵a,δaSjzSj′z⟩0,(29)c_{g(n)}^{[full]}(i\nu_m)= \sum_{a=0}^{n}\sum_{\delta a=0}^{n-a} 2\widetilde{K}_{n+2}^{(even)}(a+1,a+\delta a+2;m) \sum_{p\in S_n}'\left\langle PV_{e_1}^{[p_1]}\ldots V_{e_n}^{[p_n]} \overset{a,\delta a}{\longleftarrow}S_j^zS_{j'}^z\right\rangle_0 , \tag*{(29)}

where the even kernel

K~n+2(even)(a+,a−;m)≡[K~n+2(a+,a−;m)+K~n+2(a−,a+;m)]/2,(30)\widetilde{K}_{n+2}^{(even)}(a_+,a_-;m)\equiv \left[\widetilde{K}_{n+2}(a_+,a_-;m)+\widetilde{K}_{n+2}(a_-,a_+;m)\right]/2, \tag*{(30)}

is given by Eq. (26) with the contributions from odd ll removed and is thus real. The same simplification is also possible for all other graphs with j≠j′j \ne j'. This builds on the insight that graph evaluations cg(n)(iνm)c_g^{(n)}(i\nu_m) are of little practical relevance for themselves and are only required to find cii′(n)(iνm)c_{ii'}^{(n)}(i\nu_m) via the embedding in Eq. (15). In the latter however, embeddings with both (i,i′)→(j,j′)(i,i') \to(j,j') and (i,i′)→(j′,j)(i,i') \to(j',j) are required. Hence, the rewriting with the even kernel and the vanishing of odd powers of Δ2πm\Delta_{2\pi m} also applies to this case, but only after the embedding. These observations allow us to use Eq. (29) for arbitrary graphs.

Cyclicity of trace: Due to the cyclic nature of the trace ⟨…⟩0\langle\ldots\rangle_0 in Eq. (29), we can always equate the pp-sum of the latter for any aa by setting a→0a \to0. Then the sum over aa only has to be taken over the kernel,

K~n+2(even)(δa;m)≡∑a=0n−δaK~n+2(even)(a+1,a+δa+2;m),(31)\widetilde{K}_{n+2}^{(\mathrm{even})}(\delta a;m) \equiv\sum_{a=0}^{n-\delta a} \widetilde{K}_{n+2}^{(\mathrm{even})}(a+1,a+\delta a+2;m), \tag*{(31)}

and Eq. (29) simplifies to

cg[full](n)(iνm)=2∑δa=0nK~n+2(even)(δa;m)∑p∈Sn′⟨Sj′zPVe1[p1]…Ven[pn]←δaSjz⟩0.(32)c_g^{[full](n)}(i\nu_m)=2\sum_{\delta a=0}^{n}\widetilde{K}_{n+2}^{(\mathrm{even})}(\delta a;m)\sum_{p\in S_n}'\left\langle S_{j'}^zPV_{e_1}^{[p_1]}\ldots V_{e_n}^{[p_n]}\xleftarrow{\delta a}S_j^z\right\rangle_0. \tag*{(32)}

Anticipating the two embeddings (i,i′)→(j,j′)(i,i') \to(j,j') and (i,i′)→(j′,j)(i,i') \to(j',j) as above, we can write the more symmetric expression

cg[full](n)(iνm)=∑δa=0n[K~n+2(even)(δa,m)+K~n+2(even)(n−δa,m)]∑p∈Sn′⟨Sj′zPVe1[p1]…Ven[pn]←δaSjz⟩0.(33)c_g^{[full](n)}(i\nu_m)=\sum_{\delta a=0}^{n}\left[\widetilde{K}_{n+2}^{(\mathrm{even})}(\delta a,m)+\widetilde{K}_{n+2}^{(\mathrm{even})}(n-\delta a,m)\right]\sum_{p\in S_n}'\left\langle S_{j'}^zPV_{e_1}^{[p_1]}\ldots V_{e_n}^{[p_n]}\xleftarrow{\delta a}S_j^z\right\rangle_0. \tag*{(33)}

From this equation, also the general form of the expansions in Eq. (17) [and (7)] follow: From the definition of the kernel (26), after dropping odd powers of Δ2πm\Delta_{2\pi m} in the even kernel, graph evaluations for g(n)g^{(n)} with nn even contain Δ2πmn\Delta_{2\pi m}^{n} as the highest power. For odd nn, naively Δ2πmn+1\Delta_{2\pi m}^{n+1} can occur. However, these terms, which emerge only from the a=0a=0 contribution to K~n+2(even)(δa,m)\widetilde{K}_{n+2}^{(\mathrm{even})}(\delta a,m) mutually cancel out in the square bracket of Eq. (33).

Regarding algorithmic complexity, note that the sum in Eq. (33) has only a factor of nn more terms than the analogous expression for the conventional HTE of the equal-time correlator ⟨SjzSj′z⟩\langle S_j^zS_{j'}^z\rangle where the two external operators always stay right next to each other due to the absence of time-integrals.

Flavor sum: So far we have not considered the structure of the edge operators which according to Sec. II are Heisenberg exchange interactions, Vek=∑γkSek(1)γkSek(2)γkV_{e_k}=\sum_{\gamma_k}S_{e_k(1)}^{\gamma_k}S_{e_k(2)}^{\gamma_k}. Here the edge-flavor sum is over γk∈{+,−,z}≡{+1,−1,0}\gamma_k\in\{+,-,z\}\equiv\{+1,-1,0\} and ek(1)e_k(1) and ek(2)e_k(2) are the two vertices connected by edge eke_k. As Eq. (33) involves nn of these edge operators, resolving the VekV_{e_k} for k=1,…,nk=1,\ldots,n naively gives rise to 3n3^n flavor combinations to be summed over. As mentioned below Eq. (11), the traces ⟨…⟩0\langle\ldots\rangle_0 of strings of equal-time spin operators {S+,S−,Sz}\{S^+,S^-,S^z\} appearing in Eq. (33) factorize into traces of spin operators with the same vertex (site) index. The required on-site free spin equal-time correlators for a string of n′≤n+2n'\le n+2 spin operators are pre-computed and saved as a computationally efficient integer after multiplying by a factor (2S+1)n′(2S+1)^{n'}. The product of these factors is to be canceled by global factor 1/(2S+1)2n+21/(2S+1)^{2n+2} which is to be multiplied only at the end.

It turns out that the single free spin nn-point correlators vanish for many operator strings. For example in the case S=1/2S=1/2 and for spin-operator strings of length n′=12n'=12, only 81728172 out of the 312=5314413^{12}=531441 possible strings have a non-zero trace. Conditions for a zero trace that do not depend on the operator order are a breaking of U(1)U(1) spin rotation symmetry (unequal numbers of S+S^+ and S−S^- in the string) or a string that is odd under a π\pi-rotation around the spin-xx axis. The latter is the case for a string consisting of an odd number of SzS^z operators, SzSz…SzS^zS^z\ldots S^z. Such flavor combinations can be skipped already before entering the ∑p∈Sn′\sum_{p\in S_n}' sum. We also use the πx\pi_x-rotation symmetry in spin space that ensures equality of the traces for flavor strings connected by a global exchange of S+↔S−S^+\leftrightarrow S^- operators.

Once a given permutation p∈Snp\in S_n is selected, operator identities like (S+)a=0(S^+)^a=0 for a>2Sa>2S might render the operator string trivial. Note that this holds also if SzS^z operators are inserted in between. This is useful since it makes the condition independent of any insertions of external SzS^z operators done in the δa\delta a sum.

Results and applications

Tests of expansion coefficients

In a first step, we check the rational coefficients in the polynomials pii′(n)(x)p_{ii'}^{(n)}(x) of Eq. (7) obtained from Dyn-HTE in two settings where they are known by other means, see

Appendix A for details. In a first test, we consider the Heisenberg four spin cluster (N=4N=4) with all-to-all interactions for which the Matsubara correlator Gii′(iνm)G_{ii'}(i\nu_m) can be found analytically for S=1/2S=1/2, S=1S=1 and both for the local and non-local case. The rational coefficients of a series expansion of these exact results in x=J/Tx=J/T agrees to our graph-embedding based Dyn-HTE result. This test thus involves evaluations for all graphs with up to four vertices (as they are embeddable in the N=4N=4 all-to-all cluster). In a second test we focus on extended lattices in two and three dimensions (like kagome, triangular, pyrochlore, ...) and compute their Dyn-HTE. At each order in xx, we take the frequency- and site-sums to reproduce the conventional HTE expansion of the (purely static) uniform susceptibility χ\chi found in the conventional HTE literature.

Benchmark: Heisenberg S=1/2S=1/2 AFM chain

We proceed to compute numerical values for the Matsubara correlator (4) at finite x=J/Tx=J/T based on the Dyn-HTE expansion (7). We start in this subsection with the (infinite) nearest-neighbor S=1/2S=1/2 Heisenberg AFM chain. We study the Matsubara correlators both in real and momentum space for various distances i−i′i-i' and momenta kk, see Figure 3, left and right columns, respectively. Error controlled quantum Monte Carlo (QMC) benchmark data from the worm algorithm [57, 58] are shown as symbols, error bars are smaller than symbol size. We consider the static case m=0m=0 and the two smallest positive Matsubara frequencies m=1,2m=1,2, top to bottom row.

Matsubara correlators of the Heisenberg $S=1/2$ AFM chain

Figure 3. Matsubara correlators of the Heisenberg S=1/2S=1/2 AFM chain for real-space distances i−i′=0,1,2i-i'=0,1,2 (left) and wavevectors k/π=0.4,1k/\pi=0.4,1 (right) as obtained from Dyn-HTE (lines). The frequencies are νm=2πmT\nu_m=2\pi mT with m=0,1,2m=0,1,2 (top to bottom). Markers show benchmark results from QMC simulations of a 256-site ring with error bars smaller than the symbol size. The thin gray line denotes the evaluation of the bare Dyn-HTE series truncated at order nmax⁡=12n_{\max}=12. Symmetric Padé approximants ([6, 6] and [5, 5]) of the xx-series are shown by blue lines (full and dashed), for the transformed series in u=tanh⁡(fx)u=\tanh(fx) with f=0.205f=0.205 they are indicated in purple.

The summation of the bare Dyn-HTE series up to order n=12n=12 (gray line) starts to deviate significantly from the exact results at rather high temperature x=J/T≃2x=J/T\simeq2. This is in close analogy to conventional HTE where it is well known to signal the small radius of convergence of the xx-series in the complex plane [2]. A standard tool to extract meaningful information from the bare HTE series at smaller temperatures are Padé approximants [2], which are rational functions [K,L](x)=PK(x)/QL(x)[K,L](x)=P_K(x)/Q_L(x) with polynomials PKP_K and QLQ_L of degree KK and LL, respectively. For K+L=nmax⁡K+L=n_{\max}, their coefficients can be determined so that the series expansion of the Padé approximant agrees in the first nmax⁡n_{\max} orders with the bare (Dyn-)HTE series. For the case [K=6,L=6]≡[6,6][K=6,L=6]\equiv[6,6], denoted by blue lines in Figure 3, the agreement with QMC data already extends to x≃4x\simeq4, both for the real- and momentum-space correlator. Note that for the latter, we perform the Fourier transform before computing the Padé approximant.

For many combinations of Matsubara integer mm and distance i−i′i-i' (or momentum kk) the [6, 6] Padé approximant agrees with the error controlled QMC result to even much larger x≃8x \simeq8, most severe deviations occur for the m=0m=0 case with i−i′=1i-i'=1 or k=πk=\pi. In real applications (without available benchmark data), the quality of a prediction from Padé approximants is usually gauged by comparison of different [K,L][K,L] with K+L≤nmax⁡K+L \le n_{\max} [2]. Due to the asymptotically constant (and finite) values of TGii′(iνm)TG_{ii'}(i\nu_m) at x→∞x \to\infty, we limit ourselves to the comparison of the symmetric Padé approximants [6,6][6,6] and [5,5][5,5] (dashed lines). Indeed, the poor quality of the [6,6][6,6] approximants in the two cases mentioned above is reflected in a large disagreement between the two Padé approximants [5,5][5,5] and [6,6][6,6].

Interestingly, in these cases a transformation from the bare series in xx to a series in u=tanh⁡(fx)u=\tanh(fx) and subsequent Padé approximants can help [59]. Indeed, these u-Padé approximants shown by purple lines in Figure 3 reasonably agree with the QMC benchmark data down to the smallest temperatures (x=8x=8) considered. The free parameter f=0.205f=0.205 is determined from the condition to have the best possible agreement between the [5,5][5,5] and [6,6][6,6] u-Padés. In principle, ff could be chosen differently for each bare series obtained for the set (m,i−i′)(m,i-i') [or (m,k)(m,k)], but we picked a global value here for simplicity. Finally, we note that there are also other resummation techniques like variants of Padé, “integral” approximants [2] or conformal maps [60] that could be tested with Dyn-HTE in future work.

Heisenberg S=1/2 AFM on triangular lattice

The Heisenberg S=1/2S=1/2 AFM on the triangular lattice serves as an example for a frustrated system for which the Matsubara correlator is difficult to study with QMC due to the sign problem. We focus on the static susceptibility of Eq. (8), Gk(iνm=0)=χkG_k(i\nu_m=0)=\chi_k, which has been computed previously using the alternative bold-line diagrammatic Monte Carlo (BDMC) method [30] for spin S=1/2S=1/2. In this approach a fermionic representation is used and standard fermionic skeleton Feynman diagrams are sampled using a Monte Carlo technique. Explicitly, diagrams up to order 7 in the interaction JJ are taken into account but the self-consistency condition invokes certain diagram classes at all orders.

In Figure 4 we compare the results of Dyn-HTE (lines) and the BDMC (dots). The top panel reports the TT-dependence of the static susceptibility at the ground-state ordering wavevector at the corner of the hexagonal Brillouin zone (BZ), known as the KK-point. The convergence of different Padé approximants denoted by various blue linestyles is decent and still improves by using the uu-series (with f=0.25f=0.25, purple lines) as in the previous subsection. We are thus confident that the solid purple line (u-Padé [6,6][6,6]) gives accurate results in the temperature regime shown with only a few percent error. In contrast, for x=J/T≥2x=J/T \ge2 the BDMC obtains inconsistent and much larger values. As the BDMC study [30] does not provide convergence plots of χk=K\chi_{k=K}, error bars of the BDMC data are unfortunately not available. As a further complication, the skeleton series is now known for its possibly convergence to unphysical results [61]. In the bottom panel of Figure 4, we report the static susceptibility for a path through the BZ at various TT. Away from the KK-point already detailed in the top panel, there is good agreement between BDMC (dots) and Dyn-HTE (solid lines) for all temperatures studied.

Static susceptibility for the Heisenberg S=1/2 AFM on the triangular lattice

Figure 4. Static susceptibility χk\chi_k for the Heisenberg S=1/2S=1/2 AFM on the triangular lattice. Results from Dyn-HTE (lines) are compared against the bold-line diagrammatic Monte Carlo (BDMC) data obtained by Kulagin et al. in Ref. [30] (dots). Top: χk=K\chi_{k=K} at the KK-point in the corner of the hexagonal BZ versus TT. The convergence of various Padé approximants denoted by blue linestyles is decent and still improves by using the uu-series (for f=0.25f=0.25, purple lines, same Padé approximants as for xx-series). Bottom: χk\chi_k for a path through the BZ at various TT. The Γ\Gamma-point is the center of the BZ and MM denotes the center of the BZ-edge. Away from the KK-point there is good agreement between BDMC (dots) and Dyn-HTE (full lines). Here the [6,6][6,6] u-Padé with f=0.25f=0.25 is shown. The dashed lines represent the best fit of the renormalized mean-field form in Eq. (36) to the Dyn-HTE results. The associated fit parameters {fx,gx}\{f_x,g_x\} are given in the legend.

Renormalized mean-field form of static susceptibility

In recent work [62] two of us showed that the static susceptibility χk=Gk(iνm=0)\chi_{\mathbf{k}} = G_{\mathbf{k}}(i\nu_m = 0) of a wide selection of Heisenberg models on lattices in dimension higher than one can very well be approximated by a two-parameter function of k\mathbf{k}, the renormalized mean-field form (rMF). The standard mean-field expression for nearest-neighbor models with coupling JJ reads

[TGkMF(iνm=0)]−1=1/b1+xγ1(k),(34)[TG_{\mathbf{k}}^{\mathrm{MF}}(i\nu_m = 0)]^{-1} = 1/b_1 + x\gamma_1(\mathbf{k}), \tag*{(34)}

where b1=S(S+1)/3b_1 = S(S+1)/3 denotes the dimensionless susceptibility of a free spin and γn(k)=∑r∈nth NNexp⁡(ik⋅r)\gamma_n(\mathbf{k}) = \sum_{\mathbf{r}\in nth\ \mathrm{NN}} \exp(i\mathbf{k}\cdot\mathbf{r}) is the spatial Fourier transform of the jj-th nearest-neighbor coupling pattern [xrefEq. (34) only involves γ1\gamma_1]. In the non-symmetry broken regime, the k\mathbf{k}-dependence of the exact static susceptibility can always be expanded by using the remaining γn(k)\gamma_n(\mathbf{k}) with n>1n>1,

[TGk(iνm=0)]−1=fx+gxγ1(k)+ϵ2γ2(k)+…(35)[TG_{\mathbf{k}}(i\nu_m = 0)]^{-1} = f_x + g_x\gamma_1(\mathbf{k}) + \epsilon_2\gamma_2(\mathbf{k}) + \ldots \tag*{(35)}
≈fx+gxγ1(k).(36)\approx f_x + g_x\gamma_1(\mathbf{k}). \tag*{(36)}

The approximation in Eq. (36) is the rMF form with parameters fxf_x and gxg_x replacing 1/b11/b_1 and xx from Eq. (34), respectively. The expansion in Eq. (35) is defined by inverse Fourier transforms with respect to real-space vectors rnth NN\mathbf{r}_{nth\ \mathrm{NN}} pointing to the nnth nearest-neighbor,

fx=1NbVBZ∫BZdk[TGk(iνm=0)]−1,(37)f_x = \frac{1}{N_bV_{\mathrm{BZ}}}\int_{\mathrm{BZ}} \mathrm{d}\mathbf{k}[TG_{\mathbf{k}}(i\nu_m = 0)]^{-1}, \tag*{(37)}
gx=1NbVBZ∫BZdkeik⋅rNN[TGk(iνm=0)]−1,(38)g_x = \frac{1}{N_bV_{\mathrm{BZ}}}\int_{\mathrm{BZ}} \mathrm{d}\mathbf{k}e^{i\mathbf{k}\cdot\mathbf{r}_{\mathrm{NN}}}[TG_{\mathbf{k}}(i\nu_m = 0)]^{-1}, \tag*{(38)}
ϵn=1NbVBZ∫BZdkeik⋅rnth NN[TGk(iνm=0)]−1.(39)\epsilon_n = \frac{1}{N_bV_{\mathrm{BZ}}}\int_{\mathrm{BZ}} \mathrm{d}\mathbf{k}e^{i\mathbf{k}\cdot\mathbf{r}_{nth\ \mathrm{NN}}}[TG_{\mathbf{k}}(i\nu_m = 0)]^{-1}. \tag*{(39)}

Here VBZV_{\mathrm{BZ}} is the BZ volume and NbN_b the number of basis sites.

In Ref. [62] we could show analytically that the HTE of the beyond-rMF terms ϵn\epsilon_n for n=2,3,…n = 2, 3, \ldots start at order O(x4)O(x^4) with a particularly small prefactor. This lead us to conjecture the excellent validity of the rMF beyond the high-temperature regime which was corroborated by matching the rMF form (36) with published momentum-dependent susceptibilities obtained from the BDMC method [30] and pseudo-fermion functional renormalization group [63].

Dyn-HTE now offers an alternative and more reliable source of static susceptibility data. For χk\chi_{\mathbf{k}} of the triangular lattice model shown in Figure 4(bottom) we determined the rMF parameters fxf_x and gxg_x of Eq. (36) via least-squares fit to the Dyn-HTE data (see legend and dashed lines). The rMF form of the susceptibility describes the Dyn-HTE data remarkably well, with an almost perfect match even for the lowest temperature T/J=0.375T/J = 0.375.

A quantitative measure of the rMF form’s accuracy is provided by the ratio ∣ϵnγn(k)fx+gxγ1(k)∣\left|\frac{\epsilon_n\gamma_n(\mathbf{k})}{f_x+g_x\gamma_1(\mathbf{k})}\right| for which values much

smaller than unity indicate the applicability of the approximation. In Figure 5 we report several ratios of fxf_x, gxg_x and the numerically dominant ϵ2,ϵ3,ϵ4\epsilon_2,\epsilon_3,\epsilon_4 for the Heisenberg S=1/2S = 1/2 AFM on various lattices. They are obtained from the uu-Padé approximations for the xx-series expansions of the respective ratio found from Dyn-HTE. For the chain (top left panel), the ratios 1/fx1/f_x, gx/fxg_x/f_x and ϵn/fx\epsilon_n/f_x saturate for low TT (large xx), hence resummation of the ratios with the Padé approximant of the uu-series works very well as witnessed by comparison to QMC (dots). As ϵ2\epsilon_2 grows sizably for low TT, the validity of the rMF approximation is compromised. This was stated before in Ref. [64].

Amplitudes of parameter-ratios of the rMF form

Figure 5. Amplitudes of parameter-ratios of the rMF form (36): x/fxx/f_x, gx/fxg_x/f_x and the first few corrections ϵ2,3,4/fx\epsilon_{2,3,4}/f_x [c.f. Eq. (35)] for the nearest-neighbor S=1/2S = 1/2 Heisenberg AFM model on the chain, triangular, kagome and pyrochlore lattice from Dyn-HTE. For resummation, we used [6,6][6,6] and [5,5][5,5] u-Padé approximants. The data is multiplied by the maximum (eigenvalue) of −γ1(k)-\gamma_1(\mathbf{k}) given in the panels alongside the u-Padé parameter ff. For the chain geometry QMC data (dots) [58] matches well with the Dyn-HTE data (lines). For triangular, kagome and pyrochlore lattice ϵ2,3,4/fx\epsilon_{2,3,4}/f_x are of order 10−310^{-3} rendering the rMF an excellent approximation. Whenever the [6,6][6,6] or [5,5][5,5] Padé approximant is anomalous (with obvious poles), we use a stable lower-order Padé approximant.

However, for the two and three dimensional systems (triangular, kagome and pyrochlore lattice), the results in Figure 5 indicate that ϵ2,ϵ3,ϵ4\epsilon_2,\epsilon_3,\epsilon_4 are negligible as compared to fxf_x and gxg_x underlining the validity of the rMF approximation. Remarkably, ϵn\epsilon_n stays at least two orders of magnitude lower than gxg_x over the whole accessible temperature range. Therefore, the static susceptibility is accurately described by just gxg_x and fxf_x for these models.

This also explains the excellent fit to the rMF form in Figure 4. Other corrections ϵn\epsilon_n for n>4n > 4 (not shown) stay very small as well.

Conclusion and outlook

We have presented a dynamical extension of the high-temperature series expansion (HTE) to the Matsubara spin correlator for spin Hamiltonians (Dyn-HTE). Currently, our numerical implementation [21] features an expansion to order nmax⁡=12n_{\max}=12 and is available for S≤1S \le1 Heisenberg Hamiltonians with a single coupling constant JJ on arbitrary (in particular frustrated and high-dimensional) lattices. The real-frequency dynamic structure factor can be obtained via postprocessing of the Matsubara data as described in our companion letter [39].

For further methodological development of Dyn-HTE it would be worthwhile to reduce the number of graphs by the free-graph expansion technique (where different vertices of a graph can be assigned to the same or different lattice sites) or by the linked-cluster method [2]. In the latter, only graphs (clusters) with simple edges are required on which the observable needs to be expanded in x=J/Tx=J/T or is even determined by exact diagonalization. While this could be a viable way to higher expansion orders for small SS, it would also be interesting to extend Dyn-HTE in the form presented in this work to arbitrary SS. This could be achieved by building on closed-form expressions for the equal-time free spin correlators [22].

Moreover, possible extensions of Dyn-HTE analogous to achievements in conventional HTE include the application to Heisenberg models with more than one coupling constant, e.g. J1J_1-J2J_2 models [5], single-ion anisotropies or magnetic fields [6]. In the latter case, the kernel trick for the evaluation of the imaginary-time integrals needs to be adapted for the presence of (a few) non-zero many-body eigenenergies of H0H_0. It would also be interesting to

aim Dyn-HTE at more complex correlation functions like three-point correlators encoding higher-order response, see Ref. [28] for pioneering work. Likewise, it would be useful to improve the series convergence beyond the standard Padé approximants by implementing advanced ideas like, e.g., the homotopic action [65].

Beyond the Heisenberg case, Dyn-HTE can be extended to models with broken or reduced spin rotation symmetry like the XXZ case. Likewise, also the fermionic (or bosonic) tt-JJ-model which is prominently realized in cold-atom quantum simulation [66] should be considered. Here, existing treatments [44, 67] could be extended to higher expansion orders.

Acknowledgment

We acknowledge useful discussions with Michel Ferrero at several stages of the project and are grateful to Lode Pollet for advise on the QMC worm algorithm [58]. We further thank Nikolay Prokofiev and Boris Svistunov for useful remarks on the manuscript.

The authors acknowledge the Gauss Centre for Supercomputing e.V. (www.gauss-centre.eu) for funding this project by providing computing time through the John von Neumann Institute for Computing (NIC) on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre (JSC). The authors also acknowledge support by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through Grant No. INST 40/575-1 FUGG (JUSTUS 2 cluster). We acknowledge funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) through the Research Unit FOR 5413/1, Grant No. 465199066. B.Sch. acknowledges funding from the Munich Quantum Valley, supported by the Bavarian state government with funds from the Hightech Agenda Bayern Plus. B.Sch. and B.Sch. are supported by DFG Grant No. 524270816.

Benchmark tests for expansion of Matsubara correlator

Small spin cluster: N=4 spins with all-to-all coupling

As a first benchmark check on the expansion coefficients of Dyn-HTE in Eq. (7), we consider a cluster with N=4N=4 spins coupled by all-to-all interaction JJ, for both S=1/2S=1/2 and S=1S=1. We compute the exact local and non-local Matsubara correlators (4) in closed form using diagonalization and the spectral (Lehmann) representation [12]. For S=1/2S=1/2 they read

TG11(iνm)={5(x−1)+e2x(9x+8ex−3)8(e2x(2ex+9)+5)x=14−x232+x3128+23x42560+⋯:m=0,Δ2(ex−1)x(5Δ2x2+ex(5Δ2x2+ex(8Δ2+2)+5)+5)2(e2x(2ex+9)+5)(4Δ4x4+5Δ2x2+1)=3Δ2x28−3Δ2x332−(21Δ416+11Δ28)x4+⋯:m≠0.(40)TG_{11}(i\nu_m)=\begin{cases} \dfrac{5(x-1)+e^{2x}(9x+8e^x-3)}{8\left(e^{2x}(2e^x+9)+5\right)x}=\dfrac{1}{4}-\dfrac{x^2}{32}+\dfrac{x^3}{128}+\dfrac{23x^4}{2560}+\cdots& :m=0,\\ \dfrac{\Delta^2(e^x-1)x\left(5\Delta^2x^2+e^x(5\Delta^2x^2+e^x(8\Delta^2+2)+5)+5\right)}{2\left(e^{2x}(2e^x+9)+5\right)\left(4\Delta^4x^4+5\Delta^2x^2+1\right)}=\dfrac{3\Delta^2x^2}{8}-\dfrac{3\Delta^2x^3}{32}-\left(\dfrac{21\Delta^4}{16}+\dfrac{11\Delta^2}{8}\right)x^4+\cdots& :m\ne0. \end{cases} \tag*{(40)}
TG12(iνm)={15x+e2x(3x−8ex+3)+524(e2x(2ex+9)+5)=−x16+5x232+3x3256−221x415360+⋯:m=0,Δ2(ex−1)x(5Δ2x2+ex(5Δ2x2+ex(8Δ2+2)+5)+5)−6(e2x(2ex+9)+5)(4Δ4x4+5Δ2x2+1)=−Δ2x28+Δ2x332+(7Δ416+11Δ2384)x4+⋯:m≠0.(41)TG_{12}(i\nu_m)=\begin{cases} \dfrac{15x+e^{2x}(3x-8e^x+3)+5}{24\left(e^{2x}(2e^x+9)+5\right)}=-\dfrac{x}{16}+\dfrac{5x^2}{32}+\dfrac{3x^3}{256}-\dfrac{221x^4}{15360}+\cdots& :m=0,\\ \dfrac{\Delta^2(e^x-1)x\left(5\Delta^2x^2+e^x(5\Delta^2x^2+e^x(8\Delta^2+2)+5)+5\right)}{-6\left(e^{2x}(2e^x+9)+5\right)\left(4\Delta^4x^4+5\Delta^2x^2+1\right)}=-\dfrac{\Delta^2x^2}{8}+\dfrac{\Delta^2x^3}{32}+\left(\dfrac{7\Delta^4}{16}+\dfrac{11\Delta^2}{384}\right)x^4+\cdots& :m\ne0. \end{cases} \tag*{(41)}

Here we abbreviate Δ=Δ2πm\Delta=\Delta_{2\pi m}. The straightforward series expansions in xx (shown here to order x4x^4 for brevity) are reproduced via Dyn-HTE based on Eqs. (10), (15), (23) and (24) up to the maximum order nmax⁡=12n_{\max}=12. The same holds for the case S=1S = 1 (the full expressions for m≠0m \ne0 are too long to be shown)

TG11(iνm)={270x+288e10x+108e9x(3x+1)+20e7x(33x−7)+7e4x(66x−25)−81216(7e4x+10e7x+6e9x+e10x+3)x=23−2x29+13x354+181x41620+⋯m=0,8Δ2πm2x23−26Δ2πm2x39+(−244Δ2πm49−8Δ2πm29)x4+⋯m≠0.(42)T G_{11}(i\nu_m) = \begin{cases} \frac{270x+288e^{10x}+108e^{9x}(3x+1)+20e^{7x}(33x-7)+7e^{4x}(66x-25)-81}{216(7e^{4x}+10e^{7x}+6e^{9x}+e^{10x}+3)x} = \frac{2}{3}-\frac{2x^2}{9}+\frac{13x^3}{54}+\frac{181x^4}{1620}+\cdots& m=0,\\[6pt] \frac{8\Delta_{2\pi m}^{2}x^2}{3}-\frac{26\Delta_{2\pi m}^{2}x^3}{9}+\left(-\frac{244\Delta_{2\pi m}^{4}}{9}-\frac{8\Delta_{2\pi m}^{2}}{9}\right)x^4+\cdots& m\ne0. \end{cases} \tag*{(42)}
TG12(iνm)={810x−288e10x−108e9x(x+1)+140e7x(3x+1)+175e4x(6x+1)+81648(7e4x+10e7x+6e9x+e10x+3)x=−4x9+5x29+11x3162−3503x42430+⋯m=0,−8Δ2πm2x29+26Δ2πm2x327+(244Δ2πm427+8Δ2πm227)x4+⋯m≠0.(43)T G_{12}(i\nu_m) = \begin{cases} \frac{810x-288e^{10x}-108e^{9x}(x+1)+140e^{7x}(3x+1)+175e^{4x}(6x+1)+81}{648(7e^{4x}+10e^{7x}+6e^{9x}+e^{10x}+3)x} = -\frac{4x}{9}+\frac{5x^2}{9}+\frac{11x^3}{162}-\frac{3503x^4}{2430}+\cdots& m=0,\\[6pt] -\frac{8\Delta_{2\pi m}^{2}x^2}{9}+\frac{26\Delta_{2\pi m}^{2}x^3}{27}+\left(\frac{244\Delta_{2\pi m}^{4}}{27}+\frac{8\Delta_{2\pi m}^{2}}{27}\right)x^4+\cdots& m\ne0. \end{cases} \tag*{(43)}

Consistency with conventional HTE for uniform susceptibility

As mentioned above, the conventional HTE for equal-time correlators and thermodynamic quantities like the uniform susceptibility χ≡∑i′⟨SizSi′z⟩\chi\equiv\sum_{i'}\langle S_i^z S_{i'}^z\rangle is well established. Here we test Dyn-HTE by reproducing the expansion coefficients of χ\chi calculated to high orders for many lattices. First, we recover the equal-time correlators from the Matsubara correlators by a frequency sum,

⟨SizSi′z⟩=∑mTGii′(iνm).(44)\langle S_i^z S_{i'}^z\rangle= \sum_m T G_{ii'}(i\nu_m). \tag*{(44)}

We insert the Dyn-HTE expansion from Eq. (10) on the right-hand side and perform the frequency sum. The numerical values for the sum over even powers of Δ2πm\Delta_{2\pi m} are summarized in Table III for convenience.

ll24681012141618
∑mΔ2πml\sum_m \Delta_{2\pi m}^{l}112\frac{1}{12}1720\frac{1}{720}130240\frac{1}{30240}11209600\frac{1}{1209600}147900160\frac{1}{47900160}6911307674368000\frac{691}{1307674368000}174724249600\frac{1}{74724249600}361710670622842880000\frac{3617}{10670622842880000}438675109094217170944000\frac{43867}{5109094217170944000}

Table III. Frequency sums needed in Eq. (44).

As a first non-trivial consistency check, we calculate the uniform susceptibility χ\chi for the triangular lattice Heisenberg S=1/2S = 1/2 AFM and reproduce the expansion coefficients provided in Ref. 68 for all orders n≤nmax⁡=12n \le n_{\max} = 12. Note that this uses all graph evaluations for graphs embeddable in the triangular lattice with their complete frequency dependence [only a static contribution must survive after the sum in Eq. (44)] and thus constitutes a rather non-trivial check.

For the Kagome lattice, we also reproduce the HTE coefficients for the susceptibility [59]. The nn-th order coefficient of χ\chi in (−x)(-x) is given by an+1(−1)n/[(n+1)!4n+2]a_{n+1}(-1)^n/[(n+1)! 4^{n+2}] where ana_n is given in the right column of Tab. I in Ref. 59. There, a misprint must be corrected, a7=2711296a_7 = 2711296. This has been also noticed in Ref. 4. For S=1S = 1 we performed the same check for χ\chi against the conventional HTE [4].

Finally, for the Heisenberg S=1/2S = 1/2 AFM on the pyrochlore lattice in three dimensions, we reproduced the HTE coefficients published in Ref. [11].

References

References

  1. [2]J. Oitmaa, C. Hamer, and W. Zheng, Series Expansion Methods for Strongly Interacting Lattice Models (Cambridge University Press, Cambridge, 2006).DOI
  2. [3]H.-J. Schmidt, J. Schnack, and M. Luban, Heisenberg exchange parameters of molecular magnets from the high-temperature susceptibility expansion, Physical Review B 64, 224415 (2001).
  3. [4]A. Lohmann, H.-J. Schmidt, and J. Richter, Tenth-order high-temperature expansion for the susceptibility and the specific heat of spin-s Heisenberg models with arbitrary exchange patterns: Application to pyrochlore and kagome magnets, Physical Review B 89, 014415 (2014).DOI
  4. [5]A. Hehn, N. Van Well, and M. Troyer, High-temperature series expansion for spin-1/2 Heisenberg models, Computer Physics Communications 212, 180 (2017).
  5. [6]L. Pierre, B. Bernu, and L. Messio, High temperature series expansions of S = 1/2 Heisenberg spin models: algorithm to include the magnetic field with optimized complexity, SciPost Physics 17, 105 (2024).
  6. [7]B. Bernu and G. Misguich, Specific heat and high-temperature series of lattice models: Interpolation scheme and examples on quantum spin systems in one and two dimensions, Physical Review B 63, 134409 (2001).DOI
  7. [8]N. Elstner, R. R. P. Singh, and A. P. Young, Finite temperature properties of the spin-1/2 Heisenberg antiferromagnet on the triangular lattice, Physical Review Letters 71, 1629 (1993).
  8. [9]E. A. Ghioldi, M. G. Gonzalez, S.-S. Zhang, Y. Kamiya, L. O. Manuel, A. E. Trumper, and C. D. Batista, Dy- namical structure factor of the triangular antiferromagnet: Schwinger boson theory beyond mean field, Physical Review B 98, 184403 (2018).
  9. [10]L. Chen, D.-W. Qu, H. Li, B.-B. Chen, S.-S. Gong, J. v. Delft, A. Weichselbaum, and W. Li, Two-temperature scales in the triangular-lattice Heisenberg antiferromagnet, Physical Review B 99, 140404 (2019).DOI
  10. [11]M. G. Gonzalez, B. Bernu, L. Pierre, and L. Messio, Finite-temperature phase transitions in S=1/2 three-dimensional Heisenberg magnets from high-temperature series expansions, Physical Review B 107, 235151 (2023).
  11. [12]H. Bruus, K. Flensberg, H. Bruus, and K. Flensberg, Many-Body Quantum Theory in Condensed Matter Physics: An Introduction, Oxford Graduate Texts (Oxford University Press, Oxford, New York, 2004).
  12. [13]M. Mourigal, M. Enderle, A. Klöpperpieper, J.-S. Caux, A. Stunault, and H. M. Rønnow, Fractional spinon excitations in the quantum Heisenberg antiferromagnetic chain, Nature Physics 9, 435 (2013).DOI
  13. [14]J. Khatua, B. Sana, A. Zorko, M. Gomilšek, K. S. M. R. Rao, M. Baenitz, B. Schmidt, and P. Khuntia, Experimental signatures of quantum and topological states in frustrated magnetism, Physics Reports 1041, 1 (2023).DOI
  14. [15]M. L. Prichard, Z. Ba, I. Morera, B. M. Spar, D. A. Huse, E. Demler, and W. S. Bakr, Magnon-polarons in the Fermi–Hubbard model, Nature Physics 10.1038/s41567-025-03004-6 (2025).arxiv.org/abs/2502.06757
  15. [16]C. Lacroix, P. Mendels, and F. Mila, Introduction to Frustrated Magnetism: Materials, Experiments, Theory, 2011th ed. (Springer, Berlin Heidelberg, 2011).DOI
  16. [17]N. E. Sherman and R. R. P. Singh, Structure factors of the kagome-lattice Heisenberg antiferromagnets at finite temperatures, Physical Review B 97, 014423 (2018).DOI
  17. [18]A. Auerbach, Interacting Electrons and Quantum Magnetism, 1994th ed. (Springer, New York, 1994).
  18. [19]M. Drescher, L. Vanderstraeten, R. Moessner, and F. Pollmann, Dynamical signatures of symmetry-broken and liquid phases in an S = 1/2 Heisenberg antiferromagnet on the triangular lattice, Physical Review B 108, L220401 (2023).DOI
  19. [20]L. Savary and L. Balents, Quantum spin liquids: A review, Reports on Progress in Physics 80, 016502 (2016).DOI
  20. [21]B. Sbierski, B. Schneider, and R. Burkard, Dyn-HTSE software v1.0: 10.5281/zenodo.17234446 (2025).
  21. [22]J. Halbinger, B. Schneider, and B. Sbierski, Spectral representation of Matsubara n-point functions: Exact kernel functions and applications, SciPost Physics 15, 183 (2023).DOI
  22. [23]V. G. Yaks and A. I. Larkin, Spin waves and correlation functions in a ferromagnetic, Soviet Physics JETP 26, 647 (1968).
  23. [24]Y. A. Izyumov and Y. N. Skryabin, Statistical Mechanics of Magnetically Ordered Systems (Consultants Bureau, New York, N.Y., 1988).
  24. [25]R. Goll, D. Tarasevych, J. Krieg, and P. Kopietz, Spin functional renormalization group for quantum Heisenberg ferromagnets: Magnetization and magnon damping in two dimensions, Physical Review B 100, 174424 (2019).DOI
  25. [26]J. Krieg and P. Kopietz, Exact renormalization group for quantum spin systems, Physical Review B 99, 060403 (2019).
  26. [27]D. Tarasevych and P. Kopietz, Dissipative spin dynamics in hot quantum paramagnets, Physical Review B 104,
  27. [28]A. Rückriegel, D. Tarasevych, J. Krieg, and P. Kopietz, Recursive algorithm for generating high-temperature expansions for spin systems and the chiral nonlinear susceptibility, Physical Review B 110, 144416 (2024).
  28. [29]A. W. Sandvik, Computational Studies of Quantum Spin Systems, AIP Conference Proceedings 1297, 135 (2010).
  29. [30]S. A. Kulagin, N. Prokof’ev, O. A. Starykh, B. Svistunov, and C. N. Varney, Bold diagrammatic Monte Carlo technique for frustrated spin systems, Physical Review B 87, 024407 (2013).
  30. [31]J. Reuther and P. Wölfle, J1-J2 frustrated two-dimensional Heisenberg model: Random phase approximation and functional renormalization group, Physical Review B 81, 144410 (2010).DOI
  31. [32]N. Niggemann, B. Sbierski, and J. Reuther, Frustrated quantum spins at finite temperature: Pseudo-Majorana functional renormalization group approach, Physical Review B 103, 104431 (2021).DOI
  32. [33]T. Müller, D. Kiese, N. Niggemann, B. Sbierski, J. Reuther, S. Trebst, R. Thomale, and Y. Iqbal, Pseudo-fermion functional renormalization group for spin models, Reports on Progress in Physics 87, 036501 (2024).
  33. [34]J. Fei, C.-N. Yeh, and E. Gull, Numerical Analytical Continuation, Physical Review Letters 126, 056402 (2021).
  34. [35]K. Nogaki and H. Shinaoka, Bosonic Nevanlinna Analytic Continuation, Journal of the Physical Society of Japan 92, 035001 (2023).
  35. [36]L. Huang and C. Yue, Barycentric rational function approximation made simple: A fast analytic continuation method for Matsubara Green’s functions, Physical Review B 111, 125139 (2025).
  36. [37]M. Kliczkowski, L. Keyes, S. Roy, T. Paiva, M. Randeria, N. Trivedi, and M. M. Maska, Autoencoder-based analytic continuation method for strongly correlated quantum systems, Physical Review B 110, 115119 (2024).DOI
  37. [38]H. Shao and A. W. Sandvik, Progress on stochastic analytic continuation of quantum Monte Carlo data, Physics Reports 1003, 1 (2023).
  38. [39]R. Burkard, B. Schneider, and B. Sbierski, Dynamic correlations of frustrated quantum spins from high-temperature expansion, Physical Review Letters 136, 056501 (2026).
  39. [40]P. C. Kwok and T. D. Schultz, Correlation functions and Green functions: zero-frequency anomalies, Journal of Physics C: Solid State Physics 2, 1196 (1969).
  40. [41]V. S. Viswanath and Müller Gerhard, The Recursion Method: Application to Many-Body Dynamics, Lecture Notes in Physics Monographs No. v.23 (Springer Berlin / Heidelberg, 1994).DOI
  41. [42]H. Mori, A Continued-Fraction Representation of the Time-Correlation Functions, Progress of Theoretical Physics 34, 399 (1965).DOI
  42. [43]S. Pairault, D. Sénéchal, and A.-M. S. Tremblay, Strong-Coupling Perturbation Theory of the Hubbard Model, The European Physical Journal B 16, 85 (2000).
  43. [44]E. Perepelitsky, A. Galatas, J. Mravlje, R. Žitko, E. Khatami, B. S. Shastry, and A. Georges, Transport and optical conductivity in the Hubbard model: A high-temperature expansion perspective, Physical Review B 94, 235115 (2016).DOI
  44. [46]R. Rossi, Determinant Diagrammatic Monte Carlo in the Thermodynamic Limit, Physical Review Letters 119, 045701 (2017).
  45. [47]F. L. Buessen, fluessen/SpinMC.jl (2025).
  46. [48]Graphs.jl: A performant platform for network and graph analysis in julia (2025).
  47. [49]L. Cordella, P. Foggia, C. Sansone, and M. Vento, A (sub)graph isomorphism algorithm for matching large graphs, IEEE Transactions on Pattern Analysis and Machine Intelligence 26, 1367 (2004).
  48. [50]F. B. Kugler, S.-S. B. Lee, and J. von Delft, Multi-point Correlation Functions: Spectral Representation and Numerical Evaluation, Physical Review X 11, 041006 (2021).
  49. [51]J. Vučicević, P. Stipsić, and M. Ferrero, Analytical solution for time integrals in diagrammatic expansions: Application to real-frequency diagrammatic Monte Carlo, Physical Review Research 3, 023082 (2021).DOI
  50. [53]J. Kaye, K. Chen, and O. Parcollet, Discrete lehmann representation of imaginary time green’s functions, Phys. Rev. B 105, 235115 (2022).DOI
  51. [54]M. Wallerberger, S. Badr, S. Hoshino, S. Huber, F. Kakizawa, T. Koretsune, Y. Nagai, K. Nogaki, T. Nomoto, H. Mori, et al., sparse-ir: Optimal compression and sparse sampling of many-body propagators, SoftwareX 21, 101266 (2023).DOI
  52. [55]A. Taheridehkordi, S. H. Curnoe, and J. P. F. LeBlanc, Algorithmic matsubara integration for hubbard-like models, Phys. Rev. B 99, 035120 (2019).DOI
  53. [56]M. Burke and J. LeBlanc, TorchArmi: Generalized cpu/gpu implementation of algorithmic matsubara integration, Computer Physics Communications 308, 109437 (2025).
  54. [57]N. Sadoune and L. Pollet, Efficient and scalable path integral Monte Carlo simulations with worm-type updates for Bose-Hubbard and XXZ models, SciPost Physics Codebases , 9 (2022).arxiv.org/abs/2204.12262
  55. [58]N. Sadoune and L. Pollet, Worm algorithm code, GitHub repository (2023).
  56. [59]N. Elstner and A. P. Young, Spin-1/2 Heisenberg antiferromagnet on the kagome lattice: High-temperature expansion and exact-diagonalization studies, Physical Review B 50, 6871 (1994).
  57. [60]C. Bertrand, S. Florens, O. Parcollet, and X. Waintal, Reconstructing Nonequilibrium Regimes of Quantum Many-Body Systems from the Analytical Structure of Perturbative Expansions, Physical Review X 9, 041008 (2019).
  58. [61]E. Kozik, M. Ferrero, and A. Georges, Nonexistence of the Luttinger-Ward Functional and Misleading Convergence of Skeleton Diagrammatic Series for Hubbard-Like Models, Physical Review Letters 114, 156402 (2015).
  59. [62]B. Schneider and B. Sbierski, Taming spin susceptibilities in frustrated quantum magnets: Mean-field form and approximate nature of the quantum-to-classical correspondence, Physical Review Letters 134, 176502 (2025).DOI
  60. [63]M. G. Gonzalez, V. Noculak, A. Sharma, V. Favre, J.-R. Soh, A. Magrez, R. Bewley, H. O. Jeschke, J. Reuther, H. M. Rønnow, Y. Iqbal, and I. Živković, Dynamics of K2Ni2(SO4)3 governed by proximity to a 3D spin liquid model, Nature Communications 15, 7191 (2024).
  61. [64]S. A. Kulagin, N. Prokof’ev, O. A. Starykh, B. Svistunov, and C. N. Varney, Bold Diagrammatic Monte Carlo Method Applied to Fermionized Frustrated Spins, Physical Review Letters 110, 070601 (2013).
  62. [65]A. J. Kim, N. V. Prokof’ev, B. V. Svistunov, and E. Kozik, Homotopic Action: A Pathway to Convergent Diagrammatic Theories, Physical Review Letters 126, 257001 (2021).DOI
  63. [66]J. Koepsell, D. Bourgund, P. Sompet, S. Hirthe, A. Bohrdt, Y. Wang, F. Grusdt, E. Demler, G. Salomon, C. Gross, and I. Bloch, Microscopic evolution of doped Mott insulators from polaronic metal to Fermi liquid, Science 374, 82 (2021).DOI
  64. [67]W. Metzner, Linked-cluster expansion around the atomic limit of the Hubbard model, Physical Review B 43, 8549 (1991).

Paper details

vlnc:W_ry6428fg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

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