Abstract
According to the concept of typicality, an ensemble average can be accurately approximated by an expectation value with respect to a single pure state drawn at random from a high-dimensional Hilbert space. This random-vector approximation, or trace estimator, provides a powerful approach to, e.g. thermodynamic quantities for systems with large Hilbert-space sizes, which usually cannot be treated exactly, analytically or numerically. Here, we discuss the finite-size scaling of the accuracy of such trace estimators from two perspectives. First, we study the full probability distribution of random-vector expectation values and, second, the full temperature dependence of the standard deviation. With the help of numerical examples, we find pronounced Gaussian probability distributions and the expected decrease of the standard deviation with system size, at least above certain system-specific temperatures. Below and in particular for temperatures smaller than the excitation gap, simple rules are not available.
1 Introduction
Methods such as the finite-temperature Lanczos method (FTLM) [1], [2], [3], [4], [5], [6], [7] that rest on trace estimators [1], [8], [9], [10], [11], [12], [13], [14], [15], [16] and thus – in more modern phrases – on the idea of typicality [17], [18], [19], [20], approximate equilibrium thermodynamic observables with very high accuracy [2], [21]. In the canonical ensemble, the observable can be evaluated either with respect to a single random vector
or with respect to an average over R random vectors,
where numerator and denominator are averaged with respect to the same set of random vectors. The components of
In this work, we discuss the accuracy of (1) and (2), where we particularly focus on the dependence of this accuracy on the system size or, to be more precise, the dimension of the effective Hilbert space spanned by thermally occupied energy eigenstates. While it is well established that the accuracy of both equations increases with the square root of this dimension, we shed light on the size dependence from two less studied perspectives. First, we study the full probability distribution of random-vector expectation values, for the specific example of magnetic susceptibility and heat capacity in quantum spin systems on a one-dimensional lattice. At high temperatures, our numerical simulations unveil that these distributions are remarkably well described by simple Gaussian functions over several orders of magnitude. Moreover, they clearly narrow with the inverse square root of the Hilbert-space dimension towards a δ function. Decreasing temperature at fixed system size, we find the development of broader and asymmetric distributions. Increasing the system size at fixed temperature, however, distributions become narrow and symmetric again. Thus, the mere knowledge of the standard deviation turns out to be sufficient to describe the full statistics of random-vector expectation values–at least at not too low temperatures.
The second central perspective of our work is taken by performing a systematic analysis of the scaling of the standard deviation with the system size, over the entire range of temperature and in various quantum spin models including spin-1/2 and spin-1 Heisenberg chains, critical spin-1/2 sawtooth chains, as well as cuboctahedra with spins 3/2, 2 and 5/2. We show a monotonous decrease of the standard deviation with increasing effective Hilbert-space dimension, as long as the temperature is high compared to some system-specific low-energy scale. Below this scale, the scaling can become unsystematic if only a very few low-lying energy eigenstates contribute. However, when averaging according to (2) over a decent number (∼ 100) of random vectors, one can still determine the thermodynamic average very accurately in all examples considered by us. A quite interesting example constitutes the critical spin-1/2 sawtooth chain, where a single state drawn at random is enough to obtain this average down to very low temperatures.
This paper is organized as follows. In Section 2 we briefly recapitulate models, methods, as well as typicality-based estimators. In Section 3 we present our numerical examples both for frustrated and unfrustrated spin systems. The paper finally closes with a summary and discussion in Section 4.
2 Method
In this article we study several spin systems at zero magnetic field. They are of finite size and described by the Heisenberg model,
where the sum runs over ordered pairs of spins. Here and in the following operators are marked by a tilde, i.e.
Numerator and denominator of (2), the latter is the partition function, are evaluated using a Krylov-space expansion, i.e. a spectral representation of the exponential in a Krylov space with
If the Hamiltonian
In our numerical studies we evaluate the uncertainty of a physical quantity by repeating its numerical evaluation NS times. For this statistical sample we define the standard deviation of the observable in the following way:
We consider two physical quantities, the zero-field susceptibility as well as the heat capacity. Both are evaluated as variances of magnetization and energy, respectively, i.e.
We compare our results with the well-established high-temperature estimate
Here E0 denotes the ground-state energy. In general the prefactor α depends on the specific system, its structure and size, as well as on temperature [18], [19], but empirically often turns out to be a constant of order α ≈ 1 for high enough temperatures, compare [2], [6], [21]. Rigorous error bounds, see Refs. [19], [32], share the dependence on
3 Numerical Results
We now present our numerical results. First, in the following Section 3.1, the full probability distribution of random-vector expectation values is discussed for shorter spin chains, where this distribution can be easily obtained by generating a large set of different random vectors. In the remainder of Section 3 the size dependence of the standard deviation is investigated for longer spin chains of spin
3.1 Distribution of Random-Vector Expectation Values for Smaller Antiferromagnetic Spin-1/2 Chains
As a first step, in order to judge the accuracy of the single-state estimate in (1), it is instructive to study its full probability distribution p, obtained by drawing many [here
The advantage of using this equation, instead of (1), is that the mean coincides with (2), the latter should be used to correctly obtain the low-temperature average in system of finite size [21]. However, at sufficiently high temperatures or in sufficiently large systems, one might equally well use (1), as we have checked.
The single results for (9) are then collected into bins of appropriate width in order to form a “smooth” distribution p. While one might expect that p will be approximately symmetric around the respective thermodynamic average, the width of the distribution indicates how reliable a single random vector can approximate the ensemble average.
In this section, we study the probability distribution p (in the following denoted as pχ and pC) for the quantities
To begin with, in Figure 1(a), pχ is shown for different chain lengths
To proceed, Figure 1(b) again shows the probability distribution pχ, but now for the finite temperature
As a second difference compared to
In order to analyze the development of the probability distribution with respect to temperature in more detail, Figure 2(a) shows pχ for various values of
This shift of the maximum is clearly visualised also in Figure 2(b), which shows the same data, but in a different style. Moreover, Figure 2(b) additionally highlights the fact that the probability distribution pχ for a fixed (and small) value of N becomes broader (and asymmetric) for intermediate values of T. Note, that pχ might become narrower again for smaller values of T, see also discussion in Sections 3.2 and 3.3.
Eventually, in Figures 3 and 4, we present analogous results for the full probability distribution p, but now for the heat capacity
3.2 Larger Antiferromagnetic Spin-1/2 Chains
Using a Krylov-space expansion one can nowadays reach large system sizes of
Following the scaling behaviour of
That the distributions are narrow can be clearly seen by eye inspection in Figure 5 where the light blue curves depict thermal expectation values according to (1). For
Their standard deviation is provided in Figure 6. Coming from high temperatures, the typical behaviour (8) switches to a behaviour that in general depends on system (here chain) and size below a characteristic temperature (here
3.3 Antiferromagnetic Spin-1 Chains
In order to monitor an example where a vanishing excitation gap cannot be expected, not even in the thermodynamic limit, we study spin-1 chains that show a Haldane gap [36], [37], see Figure 7. The scaling formula (8) indeed suggests that for
3.4 Critical Spin-1/2 Delta Chains
As the final one-dimensional example we investigate a delta chain (also called sawtooth chain) in the quantum critical region, i.e. thermally excited above the quantum critical point (QCP) [38], [39], [40]. The QCP is met when the ferromagnetic nearest-neighbour interaction J1 and the antiferromagnetic next-nearest neighbour interaction J2 between spins on adjacent odd sites assume a ratio of
As the QCP does not depend on the size of the system and the structure of the analytically known multi-magnon flat band energy eigenstates does not either, we do not expect to find large finite-size effects when investigating the standard deviation of observables, e.g. of the heat capacity. It turns even out that by eye inspection no fluctuations are visible in Figure 9(a). The figure shows
The result is an impressive example for what it means that a quantum critical system does not possess any intrinsic scale in the quantum critical region [41], [42]. The only available scale is temperature. This means in particular that the low-energy spectrum is dense and therefore does not lead to any visible fluctuations of the estimated observables.
3.5 Antiferromagnetic Cuboctahedra with Spins 3/2, 2 and 5/2
Our last scaling analysis differs from the previous examples. The cuboctahedron is a finite-size body, that is equivalent to a kagome lattice with N = 12 [43], [44], [45]. The structure is shown in Figure 10(b). Here, we vary the spin quantum number, not the size of the system. The dimension of the respective Hilbert spaces grows considerably which leads to the expected scaling (8) above temperatures of
While the low-temperature behaviour and the standard deviation of the susceptibility are largely governed by the energy gap between singlet ground state and triplet excited state, and this does not vary massively with the spin quantum number, the heat capacity is subject to stronger changes (Fig. 11). When going from smaller to larger spin quantum numbers the strongly frustrated spin system loses some of its characteristic quantum properties while becoming more classical with increasing spin s. In particular, the low-lying singlet states below the first triplet state which dominate the low-temperature heat capacity move out of the gap for larger spin s [46], [47].
It may thus well be that the type of Hilbert space enlargement, due to growing system size which leads to the thermodynamic limit or growing spin quantum number which leads to the classical limit, is important for the behaviour of the estimators (1) and (2) at low temperatures.
4 Discussion and Conclusions
To summarize, we have studied the finite-size scaling of typicality-based trace estimators. In these approaches, the trace over the high-dimensional Hilbert space is approximated by either (i) a single random state
As a first step, we have studied the full probability distribution of expectation values evaluated with respect to single random states. As an important result, we have demonstrated that for sufficiently high temperatures and large enough system sizes (i.e. sufficiently large effective Hilbert-space dimension Zeff), the probability distributions are very well described by Gaussians [33]. In particular, for comparatively high temperatures, our numerical analysis has confirmed that the standard deviation of the probability distribution scales as
In contrast, for lower temperatures, we have shown that (i) the probability distributions can become non-Gaussian and (ii) the scaling of
A remarkable example is provided by the spin-1/2 sawtooth chain with coupling-ratio
In conclusion, we have demonstrated that typicality-based estimators provide a convenient numerical tool in order to accurately approximate thermodynamic observables for a wide range of temperatures and models. While in some cases, even a single pure state is sufficient, the accuracy of the results can always be improved by averaging over a set of independently drawn states.
Acknowledgement
This work was supported by the Deutsche Forschungsgemeinschaft DFG (397067869 (STE 2243/3-1); 355031190 (FOR 2692); 397300368 (SCHN 615/25-1)). Computing time at the Leibniz Center in Garching is gratefully acknowledged. All authors thank Hans De Raedt, Peter Prelovšek, Patrick Vorndamme, Peter Reimann, Jochen Gemmer as well as Katsuhiro Morita for valuable comments.
References
[1] J. Jaklič and P. Prelovšek, Phys. Rev. B 49, 5065 (1994).10.1103/PhysRevB.49.5065Search in Google Scholar
[2] J. Jaklič and P. Prelovšek, Adv. Phys. 49, 1 (2000).10.1080/000187300243381Search in Google Scholar
[3] I. Zerec, B. Schmidt, and P. Thalmeier, Phys. Rev. B 73, 245108 (2006).10.1103/PhysRevB.73.245108Search in Google Scholar
[4] J. Schnack and O. Wendland, Eur. Phys. J. B 78, 535 (2010).10.1140/epjb/e2010-10713-8Search in Google Scholar
[5] O. Hanebaum and J. Schnack, Eur. Phys. J. B 87, 194 (2014).10.1140/epjb/e2014-50360-5Search in Google Scholar
[6] P. Prelovšek, in: The Physics of Correlated Insulators, Metals, and Superconductors (Eds. E. Pavarini, E. Koch, R. Scalettar, and R. M. Martin), Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag Jülich, Jülich, Germany 2017.Search in Google Scholar
[7] B. Schmidt and P. Thalmeier, Phys. Rep. 703, 1 (2017).10.1016/j.physrep.2017.06.004Search in Google Scholar
[8] J. Skilling, The Eigenvalues of Mega-Dimensional Matrices, Maximum Entropy and Bayesian Methods, Kluwer, Dordrecht 1988, pp. 455–466.10.1007/978-94-015-7860-8_48Search in Google Scholar
[9] M. Hutchinson, Commun. Stat. Simulat. Comput. 18, 1059 (1989).10.1080/03610918908812806Search in Google Scholar
[10] D. A. Drabold and O. F. Sankey, Phys. Rev. Lett. 70, 3631 (1993).10.1103/PhysRevLett.70.3631Search in Google Scholar PubMed
[11] R. N. Silver and H. Röder, Int. J. Mod. Phys. C 5, 735 (1994).10.1142/S0129183194000842Search in Google Scholar
[12] G. H. Golub and U. von Matt, Tikhonov regularization for large scale problems, Tech. Rep. Stanford University 1997 technical report SCCM-97-03.Search in Google Scholar
[13] A. Weiße, G. Wellein, A. Alvermann, and H. Fehske, Rev. Mod. Phys. 78, 275 (2006).10.1103/RevModPhys.78.275Search in Google Scholar
[14] H. Avron and S. Toledo, J. ACM 58, 8 (2011).10.1145/1944345.1944349Search in Google Scholar
[15] F. Roosta-Khorasani and U. Ascher, Found. Comput. Math. 15, 1187 (2015).10.1007/s10208-014-9220-1Search in Google Scholar
[16] A. K. Saibaba, A. Alexanderian, and I. C. F. Ipsen, Numer. Math. 137, 353 (2017).10.1007/s00211-017-0880-zSearch in Google Scholar
[17] K. Inoue, Y. Maeda, H. Nakano, and Y. Fukumoto, IEEE Transact. Magn. 55, 1 (2019).10.1109/TMAG.2018.2873212Search in Google Scholar
[18] S. Sugiura and A. Shimizu, Phys. Rev. Lett. 108, 240401 (2012).10.1103/PhysRevLett.108.240401Search in Google Scholar PubMed
[19] S. Sugiura and A. Shimizu, Phys. Rev. Lett. 111, 010401 (2013).10.1103/PhysRevLett.111.010401Search in Google Scholar PubMed
[20] S. Okamoto, G. Alvarez, E. Dagotto, and T. Tohyama, Phys. Rev. E 97, 043308 (2018).10.1103/PhysRevE.97.043308Search in Google Scholar PubMed
[21] J. Schnack, J. Richter, and R. Steinigeweg, Phys. Rev. Research 2, 013186 (2020).10.1103/PhysRevResearch.2.013186Search in Google Scholar
[22] B. Collins and P. Śniady, Commun. Math. Phys. 264, 773 (2006).10.1007/s00220-006-1554-3Search in Google Scholar
[23] C. Bartsch and J. Gemmer, Phys. Rev. Lett. 102, 110403 (2009).10.1103/PhysRevLett.102.110403Search in Google Scholar PubMed
[24] P. Reimann, Nat. Commun. 7, 10821 (2016).10.1038/ncomms10821Search in Google Scholar PubMed PubMed Central
[25] H. Tal-Ezer and R. Kosloff, J. Chem. Phys. 81, 3967 (1984).10.1063/1.448136Search in Google Scholar
[26] V. V. Dobrovitski and H. De Raedt, Phys. Rev. E 67, 056702 (2003).10.1103/PhysRevE.67.056702Search in Google Scholar PubMed
[27] R. Steinigeweg, J. Gemmer, and W. Brenig, Phys. Rev. Lett. 112, 120601 (2014).10.1103/PhysRevLett.112.120601Search in Google Scholar PubMed
[28] R. Steinigeweg, J. Gemmer, and W. Brenig, Phys. Rev. B 91, 104404 (2015).10.1103/PhysRevB.91.104404Search in Google Scholar
[29] T. A. Elsayed and B. V. Fine, Phys. Rev. Lett. 110, 070404 (2013).10.1103/PhysRevLett.110.070404Search in Google Scholar PubMed
[30] J. Schulenburg, Spinpack 2.56, Magdeburg University (2017).Search in Google Scholar
[31] J. Richter and J. Schulenburg, Eur. Phys. J. B 73, 117 (2010).10.1140/epjb/e2009-00400-4Search in Google Scholar
[32] A. Hams and H. De Raedt, Phys. Rev. E 62, 4365 (2000).10.1103/PhysRevE.62.4365Search in Google Scholar
[33] P. Reimann and J. Gemmer, Phys. Rev. E 99, 012126 (2019).10.1103/PhysRevE.99.012126Search in Google Scholar PubMed
[34] H.-J. Schmidt, J. Schnack, and M. Luban, Phys. Rev. B 64, 224415 (2001).10.1103/PhysRevB.64.224415Search in Google Scholar
[35] J. Schnack, J. Schulenburg, and J. Richter, Phys. Rev. B 98, 094423 (2018).10.1103/PhysRevB.98.094423Search in Google Scholar
[36] F. D. M. Haldane, Phys. Lett. A 93, 464 (1983).10.1016/0375-9601(83)90631-XSearch in Google Scholar
[37] F. D. M. Haldane, Phys. Rev. Lett. 50, 1153 (1983).10.1103/PhysRevLett.50.1153Search in Google Scholar
[38] V. Y. Krivnov, D. V. Dmitriev, S. Nishimoto, S.-L. Drechsler, and J. Richter, Phys. Rev. B 90, 014441 (2014).10.1103/PhysRevB.90.014441Search in Google Scholar
[39] D. V. Dmitriev and V. Y. Krivnov, Phys. Rev. B 92, 184422 (2015).10.1103/PhysRevB.92.184422Search in Google Scholar
[40] A. Baniodeh, N. Magnani, Y. Lan, G. Buth, C. E. Anson, et al., NPJ Quantum Mater. 3, 10 (2018).10.1038/s41535-018-0082-7Search in Google Scholar
[41] T. Vojta, Ann. Phys. 9, 403 (2000).10.1002/1521-3889(200006)9:6<403::AID-ANDP403>3.0.CO;2-RSearch in Google Scholar
[42] M. Vojta, Rep. Prog. Phys. 66, 2069 (2003).10.1088/0034-4885/66/12/R01Search in Google Scholar
[43] I. Rousochatzakis, A. M. Läuchli, and F. Mila, Phys. Rev. B 77, 094420 (2008).10.1103/PhysRevB.77.094420Search in Google Scholar
[44] A. Honecker and M. E. Zhitomirsky, J. Phys.: Conf. Ser. 145, 012082 (2009).10.1088/1742-6596/145/1/012082Search in Google Scholar
[45] H.-J. Schmidt, A. Hauser, A. Lohmann, and J. Richter, Phys. Rev. E 95, 042110 (2017).10.1103/PhysRevE.95.042110Search in Google Scholar
[46] R. Schmidt, J. Schnack, and J. Richter, J. Magn. Magn. Mater. 295, 164 (2005).10.1016/j.jmmm.2005.01.009Search in Google Scholar
[47] J. Schnack and R. Schnalle, Polyhedron 28, 1620 (2009).10.1016/j.poly.2008.10.017Search in Google Scholar
[48] K. Morita and T. Tohyama, Phys. Rev. Research 2, 013205 (2020).10.1103/PhysRevResearch.2.013205Search in Google Scholar
[49] I. Rousochatzakis, S. Kourtis, J. Knolle, R. Moessner, and N. B. Perkins, Phys. Rev. B 100, 045117 (2019).10.1103/PhysRevB.100.045117Search in Google Scholar
© 2020 Walter de Gruyter GmbH, Berlin/Boston