Abstract
We show that the local density of states (LDOS) of a wide class of tightbinding models has a weak bodyorder expansion. Specifically, we prove that the resulting bodyorder expansion for analytic observables such as the electron density or the energy has an exponential rate of convergence both at finite Fermitemperature as well as for insulators at zero Fermitemperature. We discuss potential consequences of this observation for modelling the potential energy landscape, as well as for solving the electronic structure problem.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
An atomistic potential energy landscape (PEL) is a mapping assigning energies \(E(\varvec{r})\), or local energy contributions, to atomic structures \(\varvec{r} = \{\varvec{r}_\ell \}_{\ell \in \Lambda } \in ({\mathbb {R}}^d)^\Lambda \), where \(\Lambda \) is a general (possibly infinite) index set. Highfidelity models are provided by the Born–Oppenheimer PEL associated with ab initio electronic structure models such as tightbinding, Kohn–Sham density functional theory (DFT), Hartree–Fock, or even lower level quantum chemistry models [38, 48, 54, 58, 73, 94]. Even now, however, the high computational cost of electronic structure models severely limits their applicability in material modelling to thousands of atoms for static and hundreds of atoms for longtime dynamic simulations.
There is a long and successful history of using surrogate models for the simulation of materials, devised to remain computationally tractable but capture as much detail of the reference ab initio PEL as possible. Empirical interatomic potentials are purely phenomenological and are able to capture a minimal subset of desired properties of the PEL, severely limiting their transferability [23, 86]. The rapid growth in computational resources, increased both the desire and the possibility to match as much of an ab inito PEL as possible. A continuous increase in the complexity of parameterisations since the 1990s [6, 7, 36] has over time naturally led to a new generation of “machinelearned interatomic potentials” employing universal approximators instead of empirical mechanistic models. Early examples include symmetric polynomials [11, 80], artificial neural networks [8] and kernel methods [5]. A striking case is the Gaussian approximation potential for Silicon [4], capturing the vast majority of the PEL of Silicon of interest for material applications.
The purpose of the present work is, first, to rigorously evaluate some of the implicit or explicit assumptions underlying this latest class of interatomic potential models, as well as more general models for atomic properties. Specifically, we will identify natural modelling parameters as approximation parameters and rigorously establish convergence. Secondly, our results indicate that nonlinearities are an important feature, highlighting some superior theoretical properties. Finally, unlike existing nonlinear models, we will identify explicit lowdimensional nonlinear parameterisations yet prove that they are systematic. In addition to justifying and supporting the development of new models for general atomic properties, our results establish generic properties of ab initio models that have broader consequences, e.g. for the study of the mechanical properties of atomistic materials [15, 17, 32, 93]. The application of our results to the construction and analysis of practical parameterisations (approximation schemes) that exploit our results will be pursued elsewhere.
Our overarching principle is to search for representations of properties of ab initio models in terms of simple components, where “simple” is of course highly contextspecific. To illustrate this point, let us focus on modelling the potential energy landscape (PEL), which motivated this work in the first place. Pragmatically, we require that these simple components are easier to analyse and manipulate analytically or to fit than the PEL. For many materials (at least as long as Coulomb interaction does not play a role), the first step is to decompose the PEL into site energy contributions,
where one assumes that each \(E_\ell \) is local, i.e., it depends only weakly on atoms far away. In previous works we have made this rigorous for the case of tightbinding models of varying complexity [14, 16, 17, 93]. In practise, one may therefore truncate the interaction by admitting only those atoms \({\varvec{r}}_k\) with \(r_{\ell k} :=\varvec{r}_k  \varvec{r}_\ell  < r_\mathrm{cut}\) as arguments. Typical cutoff radii range from 5Å to 8Å, which means that on the order 30 to 100 atoms still make important contributions. Thus the site energy \(E_\ell \) is still an extremely highdimensional object and short of identifying lowdimensional features it would be practically impossible to numerically approximate it, due to the curse of dimensionality.
A classical example that illustrates our search for such lowdimensional features is the embedded atom model (EAM) [23], which assigns to each atom \(\ell \in \Lambda \) a site energy
While the site energy \(E^\mathrm{eam}_\ell \) remains highdimensional, the representation is in terms of three onedimensional functions \(\phi , \rho , F\) which are easily represented for example in terms of splines with relatively few parameters. Such a lowdimensional representation significantly simplifies parameter estimation, and vastly improves generalisation of the model outside a training set. Unfortunately, the EAM model and its immediately generalisations [6] have limited ability to capture a complex ab initio PEL. Still, this example inspires our search for representations of the PEL involving parameters that are

lowdimensional,

shortranged.
Following our work on locality of interaction [14, 16, 17, 93] we will focus on a class of tightbinding models as the ab initio reference model. These can be seen either as discrete approximations to density functional theory [38] or alternatively as electronic structure toy models sharing many similarities with the more complex Kohn–Sham DFT and Hartree–Fock models.
To control the dimensionality of representations, a natural idea is to to consider a bodyorder expansion,
where \(\varvec{r}_{\ell k} :=\varvec{r}_k  \varvec{r}_\ell \) and we say that \(V_n(\varvec{r}_{\ell k_1}, \dots , \varvec{r}_{\ell k_n})\) is an \((n+1)\)body potential modelling the interaction of a centre atom \(\ell \) and n neighbouring atoms \(\{k_1,\dots ,k_n\}\). This expansion was traditionally truncated at bodyorder three (\(N = 2\)) due to the exponential increase in computational cost with N. However, it was recently demonstrated by Shapeev’s moment tensor potentials (MTPs) [80] and Drautz’ atomic cluster expansion (ACE) [25] that a careful reformulation leads to models with at most linear Ndependence. Indeed, algorithms proposed in [2, 80] suggest that the computational cost may even be Nindependent, but this has not been proven. Even more striking is the fact that the MTP and ACE models which are both linear models based on a bodyordered approximation, currently appear to outperform the most advanced nonlinear models in regression and generalisation tests [66, 106].
These recent successes are in stark contrast with the “folklore” that bodyorder expansions generally converge slowly, if at all [10, 25, 27, 46, 86]. The fallacy in those observations is typically that they implicitly assume a vacuum cluster expansion (cf. § 2.2). Indeed, our first set of main results in § 2.4 will be to demonstrate that a rapidly convergent bodyorder approximation can be constructed if one accounts for the chemical environment of the material. We will precisely characterise the convergence of such an approximation as \(N \rightarrow \infty \), in terms of the Fermitemperature and the bandgap of the material.
In the simplest scheme we consider, we achieve this by considering atomic properties \([O(\mathcal {H})]_{\ell \ell }\), where \(\mathcal {H}\) is a tightbinding Hamiltonian and O an analytic function. Approximating O by a polynomial on the spectrum \(\sigma (\mathcal {H})\) results in an approximation of the atomic property \([p(\mathcal {H})]_{\ell \ell }\), which is naturally “bodyordered”. To obtain quasioptimal approximation results, naive polynomial approximation schemes (e.g. Chebyshev) are suitable only in the simplest scenarios. For the insulating case we leverage potential theory techniques which in particular yield quasioptimal approximation rates on unions of disconnected domains. Our main results are obtained by converting these into approximation results on atomic properties, analysing their qualitative features, and taking care to obtain sharp estimates in the zeroFermitemperature limit.
These initial results provide strong evidence for the accuracy of a linear bodyorder approximation in relatively simple scenarios, and would for example be useful in a study of the mechanical response of single crystals with a limited selection of possible defects. However, they come with limitations that we discuss in the main text. In response, we consider a much more general framework, generalizing the theory of bond order potentials [55], that incorporates our linear bodyordered model as well as a range of nonlinear models. We will highlight a specific nonlinear construction with significantly improved theoretical properties over the linear scheme.
For both the linear and nonlinear bodyordered approximation schemes we prove that they inherit regularity, symmetries and locality of the original quantity of interest.
Finally, we consider the case of selfconsistent tightbinding models such as DFTB [33, 59, 78]. In this case the highly nonlinear chargeequilibration leads in principle to arbitrarily complex intermixing of the nuclei information, and thus arbitrarily high bodyorder. However, our results on the bodyordered approximations for linear tightbinding models mean that each iteration of the selfconsistent field (SCF) iteration can be expressed in terms of a low bodyordered and local interaction scheme. This leads us to propose a selfsimilar compositional representation of atomic properties that is highly reminiscent of recurrant neural network architectures. Each “layer” of this representation remains “simple” in the sense that we specified above.
2 Results
2.1 Preliminaries
2.1.1 Tight binding model
We suppose \(\Lambda \) is a finite or countable index set. For \(\ell \in \Lambda \), we denote the state of atom \(\ell \) by \({\varvec{u}}_\ell = (\varvec{r}_\ell , v_\ell , Z_\ell )\) where \(\varvec{r}_\ell \in {\mathbb {R}}^d\) denotes the position, \(v_\ell \) the effective potential, and \(Z_\ell \) the atomic species of \(\ell \). Moreover, we define \(\varvec{r}_{\ell k} :=\varvec{r}_k  \varvec{r}_\ell \), \(r_{\ell k}:=\varvec{r}_{\ell k}\), and \(\varvec{u}_{\ell k} :=( \varvec{r}_{\ell k}, v_\ell , v_k, Z_\ell , Z_k)\). For functions f of the relative atomic positions \({\varvec{u}}_{\ell k}\), the gradient denotes the gradient with respect to the spatial variable: \(\nabla f({\varvec{u}}_{\ell k}) :=\nabla \big ( \xi \mapsto f(( \xi , v_\ell , v_k, Z_\ell , Z_k ))\big )\big _{\xi = \varvec{r}_{\ell k}}\). The whole configuration is denoted by \({\varvec{u}}= (\varvec{r}, v, Z) = (\{\varvec{r}_\ell \}_{\ell \in \Lambda }, \{v_\ell \}_{\ell \in \Lambda }, \{Z_\ell \}_{\ell \in \Lambda })\).
For a given configuration \(\varvec{u}\), the tight binding Hamiltonian takes the following form:
(TB) For \(\ell , k \in \Lambda \) and \(N_\mathrm {b}\) atomic orbitals per atom, we suppose that
where h and t have values in\({\mathbb {R}}^{N_\mathrm {b}\times N_\mathrm {b}}\), are independent of the effective potential v, and are continuously differentiable with
for some \(h_0,\gamma _0>0\).
Moreover, we suppose the Hamiltonian satisfies the following symmetries:

\(h({\varvec{u}}_{\ell k}) = h({\varvec{u}}_{k\ell })^\mathrm {T}\) and \(t({\varvec{u}}_{\ell m}, {\varvec{u}}_{km}) = t({\varvec{u}}_{k m}, {\varvec{u}}_{\ell m})^\mathrm {T}\) for all \(\ell , k, m\in \Lambda \),

For orthogonal transformations \(Q \in {\mathbb {R}}^{d \times d}\), there exist orthogonal \(D^\ell (Q) \in {\mathbb {R}}^{N_\mathrm {b}\times N_\mathrm {b}}\) such that \(\mathcal {H}( Q {\varvec{u}}) = D(Q) \mathcal {H}({\varvec{u}}) D(Q)^\mathrm {T}\) where \(D(Q) = \mathrm {diag}( \{ D^\ell (Q) \}_{\ell \in \Lambda } )\) and \(Q{\varvec{u}}:=( \{Q \varvec{r}_\ell \}_{\ell \in \Lambda }, v, Z)\).
Remark 1
(i) The constants in (2.2)(2.3) are independent of the atomic sites \(\ell , k, m \in \Lambda \).
(ii) Pointwise bounds on \(h(\varvec{u}_{\ell k})\) and \(t(\varvec{u}_{\ell m}, \varvec{u}_{km})\) are normally automatically satisfied since most linear tight binding models impose finite cutoff radii. Moreover, the assumption on the derivatives \(\nabla h(\varvec{u}_{\ell k})\) and \(\nabla t(\varvec{u}_{\ell m}, \varvec{u}_{km})\) states that there are no long range interactions in the model. In particular, we are assuming that Coulomb interactions have been screened, a typical assumption in many practical tight binding codes [20, 68, 71].
(iii) The Hamiltonian is symmetric and thus the spectrum is real.
(iv) The operators \(\mathcal {H}({\varvec{u}})\) and \(\mathcal {H}( Q{\varvec{u}})\) are similar, and thus have the same spectra.
(v) The symmetry assumptions [84] of (TB) are justified in [16, Appendix A].
(vi) The entries of \(\mathcal {H}({\varvec{u}})_{\ell k} \in {\mathbb {R}}^{N_\mathrm {b} \times N_\mathrm {b}}\) will be denoted \(\mathcal {H}({\varvec{u}})_{\ell k}^{ab}\) for \(1 \leqslant a,b \leqslant N_\mathrm {b}\). When clear from the context, we drop the argument \(({\varvec{u}})\) in the notation.
The assumptions (TB) define a general threecentre tight binding model, whereas, if \(t \equiv 0\), a simplification made in the majority of tight binding codes, we say (TB) is a twocentre model [38].
The choice of potential in (TB) defines a hierarchy of tight binding models. If \(v = \mathrm{const}\), (TB) defines a linear tight binding model, a simple yet common model [14, 16, 17, 70]. In this case, we implicitly assume that the Coulomb interactions have been screened, a typical assumption made in practice for a wide variety of materials [20, 68, 71, 72]. Supposing that v is a function of a selfconsistent electronic density, we arrive at a nonlinear model such as DFTB [33, 59, 78]. Abstract variants of these nonlinear models have been analysed, for example, in [93, 99]. Through much of this article we will treat \(\varvec{r}, v\) as independent inputs into the Hamiltonian, but will discuss their connection and selfconsistency in § 2.7.
For a finite system \({\varvec{u}}\) (that is, with \(\Lambda \) a finite set), we consider analytic observables of the density of states [14, 93]: for functions \(O :{\mathbb {R}} \rightarrow {\mathbb {R}}\) that can be analytically continued into an open neighbourhood of \(\sigma \big ( \mathcal {H}( {\varvec{u}}) \big )\), we consider that
where \((\lambda _s, \psi _s)\) are normalised eigenpairs of \(\mathcal {H}({\varvec{u}})\). Many properties of the system, including the particle number functional and Helmholtz free energy, may be written in this form [14, 16, 70, 93]. By distributing these quantities amongst atomic positions, we obtain a wellknown spatial decomposition [14, 16, 35, 38],
For infinite systems, we may define \(O_\ell ({\varvec{u}})\) through the thermodynamic limit [14, 16] or via the holomorphic functional calculus; see § 4.1.2 for further details.
When discussing derivatives of the local observables, we will simplify notation and write
2.1.2 Local observables
Although the results in this paper apply to general analytic observables, our primary interest is in applying them to two special cases. A local observable of particular importance is the electron density; for inverse Fermitemperature \(\beta \in (0,+\infty ]\) and fixed chemical potential \(\mu \), we use the notation of (2.4) to define
Throughout this paper \(F^\beta ({\varvec{u}}) :=\big (F_\ell ^\beta ({\varvec{u}})\big )_{\ell \in \Lambda }\) will denote a vector and so (2.6) reads \(\rho = F^\beta ({\varvec{u}})\).
In § 2.7, we consider the case where the effective potential is a function of the electron density (2.6) (that is, \(v = {w(\rho )}\) for some \(w :{\mathbb {R}}^\Lambda \rightarrow {\mathbb {R}}^\Lambda \)) which leads to the selfconsistent local observables
where \({\varvec{u}}(\rho ) :=\big ( \varvec{r}, w(\rho ), Z\big )\).
Remark 2
All the results of this paper also hold for the offdiagonal entries of the density matrix (\(\rho _{\ell k} :=\mathrm {tr} \, F^\beta \big (\mathcal {H}({\varvec{u}})\big )_{\ell k}\)) without any additional work. This fact will be clear from the proofs. It is likely though that additional properties related to the offdiagonal decay (nearsightedness) and spatial regularity further improve the “sparsity” of the density matrix. A complete analysis would go beyond the scope of this work.
The second observable we are particularly interested in is the site energy, which allows us to decompose the total potential energy landscape into localised “atomic” contributions. In the grand potential model for the electrons, which is appropriate for large or infinite condensed phase systems [14], it is defined as
The total grand potential is defined as \(\sum _\ell G^\beta _\ell ({\varvec{u}})\) [14, 70].
For \(\beta < \infty \), the functions \(F^\beta (\,\cdot \,)\) and \(G^\beta (\,\cdot \,)\) are analytic in a strip of width \(\pi \beta ^{1}\) about the real axis [17, Lemma 5.1]. To define the zero Fermitemperature observables, we assume that \(\mu \) lies in a spectral gap (\(\mu \not \in \sigma \big ( \mathcal {H}({\varvec{u}}) \big )\); see § 2.1.3). In this case, \(F^\beta (\,\cdot \,)\) and \(G^\beta (\,\cdot \,)\) extend to analytic functions in a neighbourhood of \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big )\) for all \(\beta \in (0,\infty ]\).
In order to describe the relationship between the various constants in our estimates and the inverse Fermitemperature or spectral gap (in the case of insulators), we will state all of our results for \(O^\beta = F^\beta \) or \(G^\beta \). Other analytic quantities of interest can be treated similarly with constants depending, e.g., on the region of analyticity of the corresponding function \(z\mapsto O(z)\).
2.1.3 Metals, insulators, and defects
As we can see from (2.4), the structure of the spectrum \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big )\) will have a key role in the analysis. Firstly, by (TB), \(\mathcal {H}({\varvec{u}})\) is a bounded selfadjoint operator on \(\ell ^2(\Lambda \times \{1,\dots ,N_\mathrm {b}\})\) and thus the spectrum is real and contained in some bounded interval. In order to keep the mathematical results general, we will not impose any further restrictions on the spectrum. However, to illustrate the main ideas, we briefly describe typical spectra seen in metals and insulating systems.
In the case where \(\varvec{u}\) describes a multilattice in \({\mathbb {R}}^d\) formed by taking the union of finitely many shifted Bravais lattices, the spectrum \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big )\) is the union of finitely many continuous energy bands [57]. That is, there exist continuous functions, \(\varepsilon ^\alpha :\overline{\mathrm {BZ}} \rightarrow {\mathbb {R}}\), on the Brillouin zone \(\mathrm {BZ}\), a compact connected subset of \({\mathbb {R}}^d\), such that
In particular, in this case, \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big ) = \sigma _{\mathrm {ess}}\big ( \mathcal {H}({\varvec{u}}) \big )\) is the union of finitely many intervals on the real line. The band structure \(\{\varepsilon ^\alpha \}\) relative to the position of the chemical potential, \(\mu \), determines the electronic properties of the system [89]. In metals \(\mu \) lies within a band, whereas for insulators, \(\mu \) lies between two bands in a spectral gap. Schematic plots of these two situations are given in Figure 1.
We now consider perturbations of a reference configuration \({\varvec{u}}^\mathrm {ref} = (\varvec{r}^\mathrm {ref}, v^\mathrm {ref}, Z^\mathrm {ref})\) defined on an index set \(\Lambda ^\mathrm {ref}\).
Proposition 2.1
(Perturbation of the Spectrum) For \(\delta , R_\mathrm {def} > 0\), there exists \(\delta _0 > 0\) such that if \({\varvec{u}}= (\varvec{r}, v, Z)\) is a configuration defined on some index set \(\Lambda \) satisfying \(\Lambda \setminus B_{R_\mathrm {def}} = \Lambda ^\mathrm {ref} \setminus B_{R_\mathrm {def}}\), \(\Lambda \cap B_{R_\mathrm {def}}\) is finite, \(Z_k = Z_k^\mathrm {ref}\) for all \(k \in \Lambda \setminus B_{R_\mathrm {ref}}\), and \(\sup _{k \in \Lambda \setminus B_{R_\mathrm {def}}} \big [ \varvec{r}_k  \varvec{r}_k^\mathrm {ref} +  v_k  v_k^\mathrm {ref}  \big ] \leqslant \delta _0\), then
In particular, if \({\varvec{u}}^\mathrm {ref}\) describes a multilattice, then, since local perturbations in the defect core are of finite rank, the essential spectrum is unchanged and we obtain finitely many eigenvalues bounded away from the spectral bands. Moreover, a small global perturbation can only result in a small change in the spectrum. Again, a schematic plot of this situation is given in Figure 2.
For the remainder of this paper, we consider the following notation:
Definition 1
Suppose that \({\varvec{u}}^\mathrm {ref}\) is a general reference configuration defined on \(\Lambda ^\mathrm {ref}\) and \({\varvec{u}}\) is a configuration arising due to Proposition 2.1. Then, we define \(I_\) and \(I_+\) to be compact intervals and \(\{\lambda _j\}\) to be a finite set such that
and \(\max I_ \leqslant \mu \leqslant \min I_+\). Moreover, we define
The constants in Definition 1 are also displayed in Figure 2. The constant \(\mathsf {g}\) in Definition 1 is slightly arbitrary in the sense that as long as \(B_\delta \big ( \sigma \big (\mathcal {H}({\varvec{u}}^\mathrm {ref})\big ) \big ) \subset I_ \cup I_+\) (where \(\delta \) is the constant from Proposition 2.1), then there exists a finite set \(\{\lambda _j\}\) as in (2.9). Choosing smaller \(\mathsf {g}\) reduces the size of the set \(\{\lambda _j\}\).
2.2 Vacuum cluster expansion
For a system of M identical particles \(X_1,\dots ,X_M\), a maximal bodyorder N, and a permutation invariant energy \(E = E(\{X_1,\dots ,X_M\})\), we may consider the vacuum cluster expansion,
where the nbody interaction potentials \(V^{(n)}\) are defined by considering all isolated clusters of \(j \leqslant n\) atoms:
The expansion (2.12) is exact for \(N = M\). The vacuum cluster expansion is the traditional and, arguably, the most natural manybody expansion of a potential energy landscape. However, in many systems, it converges extremely slowly with respect to the bodyorder N and is thus computationally impractical. An intuitive explanation for this slow convergence is that, when defining the bodyorder expansion in this way, we are building an interaction law for a condensed or possibly even crystalline phase material from clusters in vacuum where the bonding chemistry is significantly different. Although this observation appears to be “common knowledge” we were unable to find references that provide clear evidence for it. However, some limited discussions and further references can be found in [10, 25, 27, 46, 86].
Our own approach employs an entirely different mechanism, which in particular incorporates environment information and leads to an exponential convergence of an Nbody approximation. Technically, our approximation is not an expansion, that is, the nbody terms \(V^{(n)}\) of the classical cluster expansion are replaced by terms that depend also on the highest bodyorder N. We will provide a more technical discussion contrasting our results with the vacuum cluster expansion in § 2.6.
2.3 A general framework
Before we consider two specific bodyordered approximations, we present a general framework which both incorporates many (linearscaling) electronic structure methods from the literature (e.g. the kernel polynomial method (KPM) [82], bondorder potentials (BOP) [26, 39, 55, 74], and quadraturebased methods [69, 87, 88]), and illustrates the key features needed for a convergent scheme: To that end, we introduce the local density of states (LDOS) [38] which is the (positive) measure \(D_\ell \) supported on \(\sigma (\mathcal {H})\) such that
Existence and uniqueness follows from the spectral theorem for normal operators (e.g. see [1, Theorem 6.3.3] or [92]). In particular, (2.4) may be written as the integral \( O_\ell ({\varvec{u}}) = \int O \,\mathrm {d}D_\ell \).
Then, on constructing a (possibly signed) unit measure \(D_\ell ^N\) with exact first N moments (that is, \(\int x^n \mathrm {d}D_\ell ^N(x) = \mathrm {tr}[\mathcal {H}^n]_{\ell \ell }\) for \(n = 1,\dots , N\)), we may define the approximate local observable \(O_\ell ^N({\varvec{u}}) :=\int O \, \mathrm {d}D_\ell ^N\), and obtain the general error estimates
where \({\mathcal {P}}_N\) denotes the set of polynomials of degree at most N, and \(\Vert \,\cdot \,\Vert _{\mathrm {op}}\) is the operator norm on a function space \(({\mathcal {S}}, \Vert \,\cdot \,\Vert _\infty )\). For example, we may take \({\mathcal {S}}\) to be the set of functions analytic on an open set containing \({\mathscr {C}}\), a contour encircling \(\mathrm {supp}\big ( D_\ell  D_\ell ^N \big )\), and consider
Alternatively, we may consider \({\mathcal {S}} = L^\infty \big ( \mathrm {supp}(D_\ell  D_\ell ^N) \big )\) leading to the total variation operator norm.
Equation (2.14) highlights the key generic features that are crucial ingredients in obtaining convergence results:

Analyticity. The potential theory results of § 4.1.5 connect the asymptotic convergence rates for polynomial approximation to the size and shape of the region of analyticity of O.

Spectral Pollution. While \(\mathrm{supp} D_\ell \subset \sigma (\mathcal {H})\), this need not be true for \(D_\ell ^N\). Indeed, if \(\mathrm{supp} D_\ell ^N\) introduces additional points within the band gap, this may significantly slow the convergence of the polynomial approximation; cf. § 2.6.

Regularity of \(D_\ell ^N\). Roughly speaking, the first term of (2.14) measures how “wellbehaved” \(D_\ell ^N\) is. In particular, if \(D_\ell ^N\) is positive, then this term is bounded independently of N, whereas, if \(D_\ell ^N\) is a general signed measure, then this factor contributes to the asymptotic convergence behaviour.
In the sections to follow, we introduce linear (§ 2.4) and nonlinear (§ 2.5) approximation schemes that fit into this general framework. Moreover, in § 2.6, we also write the vacuum cluster expansion as an integral against an approximate LDOS. In order to complement the intuitive explanation for the slow convergence of the vacuum cluster expansion, we investigate which of the requirements listed above fail.
In the appendices, we review other approximation schemes that fit into this general framework such as the quadrature method (Appendix D), numerical bond order potentials (Appendix E), and the kernel polynomial method (Appendix F).
2.4 Linear bodyordered approximation
We will construct two distinct but related manybody approximation models. To construct our first model we exploit the observation that polynomial approximations of an analytic function correspond to bodyorder expansions of an observable.
An intuitive approach is to write the local observable in terms of its Chebyshev expansion and truncate to some maximal polynomial degree. The corresponding projection operator is a simple example of the kernel polynomial method (KPM) [82] and the basis for analytic bond order potentials (BOP) [74]. We discuss in Appendix F that these schemes put more emphasis on the approximation of the local density of states (LDOS) and, in particular, exploit particular features of the Chebyshev polynomials to obtain a positive approximate LDOS. Since our focus is instead on the approximation of observables, we employ a different approach that is tailored to specific properties of the band structure and leads to superior convergence rates for these quantities.
For a set of \(N+1\) interpolation points \(X_N = \{x_j\}_{j=0}^N\), and a complexvalued function O defined on \(X_N\), we denote by \(I_{X_N}O\) the degree N polynomial interpolant of \(x\mapsto O(x)\) on \(X_N\). This gives rise to the bodyordered approximation
We may connect (2.15) to the general framework in § 2.3 by defining
and \(\ell _j\) are the node polynomials corresponding to \(X_N = \{x_j\}_{j=0}^N\) (that is, \(\ell _j\) are the polynomials of degree N with \(\ell _j(x_i) = \delta _{ij}\)).
Proposition 2.2
\(I_{X_N} O_\ell ({\varvec{u}})\) has bodyorder at most 2N. More specifically, there exists \((n+1)\)body potentials \(V_{nN}\) for \(n = 0,\dots ,2N1\) such that
Proof
(Sketch of the Proof.) Since (2.15) is a linear combination of the monomials \([\mathcal {H}^n]_{\ell \ell }\), it is enough to show that, for each \(n \in {\mathbb {N}}\),
has finite body order.
Each term in (2.18) depends on the central atom \(\ell \), the \(n1\) neighbouring sites \(\ell _1, \dots , \ell _{n1}\), and the at most n additional sites arising from the threecentre summation in the tight binding Hamiltonian (TB). In particular, (2.15) has body order at most 2N. See § 4.2 for a complete proof including an explicit definition of the \(V_{nN}\). \(\square \)
If one uses Chebyshev points as the basis for the bodyordered approximation (2.15), the rates of convergence depend on the size of the largest Bernstein ellipse (that is, ellipses with foci points \(\pm 1\)) contained in the region of analyticity of \(z \mapsto O(z)\) [95]. This leads to a exponentially convergent bodyorder expansion in the metallic finitetemperature case (see § 4.1.4 for the details).
However, the resulting estimates deteriorate in the zerotemperature limit. Instead, we apply results of potential theory to construct interpolation sets \(X_N\) that are adapted to the spectral properties of the system (see § 4.1.5 for examples) and (i) do not suffer from spectral pollution, and (ii) (asymptotically) minimise the total variation of \(D_{\ell }^{N,\mathrm {lin}}\) which, in this context, is the Lebesgue constant [95] for the interpolation operator \(I_{X_N}\). This leads to rapid convergence of the bodyorder approximation based on (2.15). The interpolation sets \(X_N\) depend only on the intervals \(I_, I_+\) from Definition 1 (see also Figure 2) and can be chosen independently of \({\varvec{u}}^\mathrm {ref}\) as long as \(B_\delta \big ( \sigma \big ( \mathcal {H}({\varvec{u}}^\mathrm {ref}) \big ) \big ) \subset I_ \cup I_+\).
Theorem 2.3
Suppose \(\varvec{u}^\mathrm {ref}\) is given by Definition 1. Fix \(0 < \beta \leqslant \infty \) and suppose that, either \(\beta < \infty \) or \(\mathsf {g}>0\). Then, for all \(N\in {\mathbb {N}}\), there exist constants \(\gamma _N > 0\) and interpolation sets \(X_N = \{x_j\}_{j=0}^N \subset I_ \cup I_+\) satisfying (2.17) such that
where \(O^\beta = F^\beta \) or \(G^\beta \) and \(C_1, C_2, \eta >0\) are independent of N. The asymptotic convergence rate \(\gamma :=\lim _{N\rightarrow \infty } \gamma _N\) is positive and exhibits the asymptotic behaviour
In this asymptotic relation, we assume that the limit \(\mathsf {g}\rightarrow 0\) is approached symmetrically about the chemical potential \(\mu \).
Remark 3
Higher derivatives may be treated similarly under the assumption that higher derivatives of the tight binding Hamiltonian (TB) exist and are short ranged.
2.4.1 The role of the point spectrum
We now turn towards the important scenario when a localised defect is embedded within a homogeneous crystalline solid. Recall from § 2.1.3 (see in particular Fig. 2) that this gives rise to a discrete spectrum, which “pollutes” the band gap [70]. Thus, the spectral gap is reduced and a naive application of Theorem 2.3 leads to a reduction in the convergence rate of the bodyordered approximation. We now improve these estimates by showing that, away from the defect, we obtain improved preasymptotics, reminiscent of similar results for locality of interaction [17].
In that follow, we fix \({\varvec{u}}\) satisfying Definition 1. While improved estimates may be obtained by choosing \(\{\lambda _j\}\) as interpolation points, leading to asymptotic exponents that are independent of the defect, in practice, this requires full knowledge of the point spectrum. Since the point spectrum within the spectral gap depends on the whole atomic configuration, the approximate quantities of interest corresponding to these interpolation operators would no longer satisfy Proposition 2.2.
Remark 4
This phenomenon has been observed in the context of Krylov subspace methods for solving linear equations \(Ax = b\) where outlying eigenvalues delay the convergence by O(1) steps without affecting the asymptotic rate [30]. Indeed, since the residual after n steps may be written as \(r_n = p_n(A) r_0\) where \(p_n\) is a polynomial of degree n, there is a close link between polynomial approximation and convergence of Krylov methods.
On the other hand, we may use the exponential localisation of the eigenvectors corresponding to isolated eigenvalues to obtain prefactors that decay exponentially as \(\varvec{r}_\ell  \rightarrow \infty \).
Theorem 2.4
Suppose \({\varvec{u}}\) satisfies Definition 1 with \(\mathsf {g}>0\). Fix \(0<\beta \leqslant \infty \) and suppose that, if \(\beta = \infty \), then \(\mathsf {g}^\mathrm {def}>0\), and let \(C_1,C_2, \gamma _N, \gamma , \eta \), and \(X_N = \{x_j\}_{j=0}^N \subset I_\cup I_+\) be given by Theorem 2.3. Then, there exist \(\gamma _\mathrm {CT},\gamma _N^{\mathrm {def}}>0\) such that
where \(O^\beta = F^\beta \) or \(G^\beta \) and \(C_3, C_4>0\) are independent of N. The asymptotic convergence rate \(\gamma ^\mathrm {def} :=\lim _{N\rightarrow \infty } \gamma ^\mathrm {def}_N\) is positive and we have
In these asymptotic relations, we assume that the limits \(\mathsf {g}^\mathrm {def},\mathsf {g}\rightarrow 0\) are approached symmetrically about the chemical potential \(\mu \).
In practice, Theorem 2.4 means that, for atomic sites \(\ell \) away from the defectcore, the observed preasymptotic error estimates may be significantly better than the asymptotic convergence rates obtained in Theorem 2.3.
Remark 5
(Locality) (i) By Theorem 2.4, and the locality estimates for the exact observables \(O_\ell ^\beta \) [17], we immediately obtain corresponding locality estimates for the approximate quantities:
(ii) We investigate another type of locality in Appendix B where we show that various truncation operators result in approximation schemes that only depend on a small atomic neighbourhood of the central site. An exponential rate of convergence as the truncation radius tends to infinity is obtained.
Remark 6
(Connection to the general framework) The fact that the exponents in Theorem 2.4 depend on the discrete eigenvalues of \(\mathcal {H}({\varvec{u}})\) can be seen from the general estimate (2.14) applied to the approximate LDOS \(D_\ell ^{N,\mathrm {lin}}\) from (2.16):

Spectral Pollution. We choose the interpolation points so that the support of \(D_\ell ^{N,\mathrm {lin}}\) lies within \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big )\) and so spectral pollution does not play a role,

Regularity of \(D_\ell ^{N.\mathrm {lin}}\). The total variation of \(D_\ell ^{N,\mathrm {lin}}\) can be estimated by the Lebesgue constant [95] for the interpolation operator \(I_{X_N}\):
$$\begin{aligned} \Vert D_\ell ^{N,\mathrm {lin}} \Vert _{\mathrm {TV}}&:=\sup _{\Vert f\Vert _{L^\infty (\sigma (\mathcal {H}))} = 1}  I_{X_N} f(\mathcal {H})_{\ell \ell }  \leqslant \sup _{\Vert f\Vert _{L^\infty (\sigma (\mathcal {H}))} = 1} \sup _{x \in \sigma (\mathcal {H})}  I_{X_N} f(x) \\&= \sup _{x \in \sigma (\mathcal {H})} \sum _j \ell _j(x). \end{aligned}$$This quantity depends on the discrete eigenvalues within the band gap.
2.5 A nonlinear representation
The method presented in § 2.4 approximates local quantities of interest by approximating the integrand \(O :{\mathbb {C}} \rightarrow {\mathbb {C}}\) with polynomials. As we have seen, this leads to approximation schemes that are linear functions of the spatial correlations \(\{[\mathcal {H}^n]_{\ell \ell }\}_{n \in {\mathbb {N}}}\). In this section, we construct a nonlinear approximation related to bondorder potentials (BOP) [26, 39, 55] and show that the added nonlinearity leads to improved asymptotic error estimates that are independent of the discrete spectra lying within the band gap. In this way, the nonlinearity captures “spectral information” from \(\mathcal {H}\) rather than only approximating \(O :{\mathbb {C}} \rightarrow {\mathbb {C}}\) without reference to the Hamiltonian.
Applying the recursion method [49, 50], a reformulation of the Lanczos process [61], we obtain a tridiagonal (Jacobi) operator T on \(\ell ^2({\mathbb {N}}_0)\) whose spectral measure is the LDOS \(D_\ell \) [91] (see § 4.3.1 for the details). We then truncate T by taking the principal \(\frac{1}{2}(N+1) \times \frac{1}{2}(N+1)\) submatrix \(T_{\frac{1}{2}(N1)}\) and define
where \(D_\ell ^{N,\mathrm {nonlin}} = \sum _s [\psi _s]_{0}^2 \delta (\,\cdot  \lambda _s)\) is a spectral measure for \(T_{\frac{1}{2}(N1)}\) (that is, \((\lambda _s,\psi _s)\) are normalised eigenpairs of \(T_{\frac{1}{2}(N1)}\)). By showing that the first N moments of \(D_\ell ^{N,\mathrm {nonlin}}\) are exact, we are able to apply (2.14) to obtain the following error estimates. The asymptotic behaviour of the exponent in these estimates follows by proving that the spectral pollution of \(D_\ell ^{N,\mathrm {nonlin}}\) in the band gap is sufficiently mild.
Theorem 2.5
Suppose \({\varvec{u}}\) satisfies Definition 1. Fix \(0 < \beta \leqslant \infty \) and suppose that, if \(\beta = \infty \), then \(\mathsf {g}>0\). Then, for N odd, there exists an open set \(U \subset {\mathbb {C}}^{N}\) such that (2.24) extends to an analytic function \(\Theta _N :U \rightarrow {\mathbb {C}}\), such that
where \(O^\beta = F^\beta \) or \(G^\beta \). The asymptotic convergence rate \(\gamma :=\lim _{N\rightarrow \infty } \gamma _N\) is positive and \(\gamma \sim \mathsf {g} + \beta ^{1}\) as \(\mathsf {g} + \beta ^{1} \rightarrow 0\).
Remark 7
It is important to note that \({\Theta _N} :U \rightarrow {\mathbb {C}}\) can be constructed without knowledge of \(\mathcal {H}\) because, as we have seen, if the discrete eigenvalues are known a priori, then Theorem 2.5 is immediate from Theorem 2.4 by adding finitely many additional interpolation points on the discrete spectrum.
In particular, the fact that \({\Theta _N}\) is a materialagnostic nonlinearity has potentially farreaching consequences for material modelling.
Remark 8
(Connection to the general framework) The fact that the exponents in Theorem 2.5 are independent of the discrete eigenvalues of \(\mathcal {H}({\varvec{u}})\) can be seen from the general estimate (2.14) applied to the approximate LDOS \(D_\ell ^{N,\mathrm {nonlin}}\) from (2.24):

Spectral Pollution. We show that \(\big \mathrm {supp}\,D_\ell ^{N,\mathrm {nonlin}} \setminus \mathrm {supp}\,D_\ell \big \) remains bounded independently of N and so spectral pollution only slows the convergence by at most O(1) steps,

Regularity of \(D_\ell ^{N,\mathrm {nonlin}}\). Since \(D_\ell ^{N,\mathrm {nonlin}}\) is a positive unit measure, we have the bound \(\Vert D_\ell  D_\ell ^{N,\mathrm {nonlin}}\Vert _{\mathrm {TV}} \leqslant 2\).
Remark 9
(Quadrature Method) Alternatively, we may use the sequence of orthogonal polynomials [40] corresponding to \(D_\ell \) as the basis for a Gauss quadrature rule to evaluate local observables. This procedure, called the Quadrature Method [51, 69], is a precursor of the bond order potentials. Outlined in Appendix D, we show that it produces an alternative scheme also satisfying Theorem 2.5.
The linearscaling spectral Gauss quadrature (LSSGQ) method [87] is based upon this idea, albeit in the context of finite difference approximations to the DFT Hamiltonian. However, since the resulting discrete Hamiltonian in [87] is banded, the analysis of the present work may be readily applied. Therefore, Theorem 2.5 provides rigorous justification for the exponential rate of convergence for increasing bodyorder (number of quadrature points), complementing the intuitive explanations and numerical experiments of [87].
Since the convergence results are independent of system size, we obtain a linearscaling method, a result that complements the intuitive explanation [87, (56)], and numerical evidence [87, Fig. 5].
Remark 10
(Convergence of Derivatives) In this more complicated nonlinear setting, obtaining results such as (2.21) is more subtle. We require an additional assumption on \(D_\ell \), which we believe maybe be typically satisfied, but we currently cannot justify it and have therefore postponed this discussion to Appendix C. We briefly mention, however, that if \(D_\ell \) is absolutely continuous (e.g., in periodic systems), we obtain
2.6 The vacuum cluster expansion revisited
For \(\ell \in \Lambda \), we denote by \(\mathcal {H}\big _{\ell ;K}\) the Hamiltonian matrix corresponding to the finite subsystem \(\{\ell \}\cup K \subset \Lambda \): for \(k_1,k_2 \in \{\ell \} \cup K\),
For an observable O, the vacuum cluster expansion as detailed in § 2.2 is constructed as follows:
Therefore, on defining the spectral measure \( D_{\ell ;K} :=\sum _s \delta \big ( \,\cdot \,  \lambda _s(K) \big ) [\psi _s(K)]_{\ell }^2 \) where \(\big (\lambda _s(K), \psi _s(K) \big )\) the are normalised eigenpairs of \(\mathcal {H}\big _{\ell ;K}\), we may write the vacuum cluster expansion as in § 2.3:
While \(D_{\ell }^{N,\mathrm {vac}}\) is a generalised signed measure (with values in \({\mathbb {R}} \cup \{\pm \infty \}\)), all moments are finite. More specifically, if we absorb the effective potential and two centre terms into the three centre summation by writing \(\mathcal {H}_{k_1 k_2} = \sum _m \mathcal {H}_{k_1 k_2 m}\), see (4.16), we have
Equation (2.30) follows from the proof of Proposition 2.2, see (4.19). In particular, the first N moments of \(D_\ell ^{N,\mathrm {vac}}\) are exact. Therefore, we may apply the general error estimate (2.14) and describe the various features of \(D_\ell ^{N,\mathrm {vac}}\) which provide mathematical intuition for the slow convergence of the vacuum cluster expansion:

Spectral Pollution. When splitting the system up into arbitrary subsystems as is the case in the vacuum cluster expansion, one expects significant spectral pollution in the band gaps, leading to a reduction in the convergence rate,

Regularity of \(D_\ell ^{N,\mathrm {vac}}\). The approximate LDOS is a linear combination of countably many Dirac deltas and does not have bounded variation. Moreover, \(D_\ell ^{N,\mathrm {vac}}\) has values in \({\mathbb {R}} \cup \{\pm \infty \}\).
2.7 Selfconsistency
Throughout this section, we suppose that the effective potential is a function of a selfconsistent electron density: that is, (2.6) becomes the following nonlinear equation:
where \({\varvec{u}}(\rho ) :=\big (\varvec{r}, {w(\rho )}, Z\big )\). We shall assume that the effective potential satisfies the following:
(EP) We suppose that \({w} :{\mathbb {R}}^\Lambda \rightarrow {\mathbb {R}}^\Lambda \) is twice continuously differentiable with
for some \(\gamma _v > 0\).
Remark 11
(i) For a smooth function \(\widetilde{w} :{\mathbb {R}} \rightarrow {\mathbb {R}}\), the effective potential \({w(\rho )_\ell :=\widetilde{w}(\rho _\ell )}\) satisfies (EP). This leads to the simplest abstract nonlinear tight binding models discussed in [93, 99].
(ii) The (shortranged) Yukawa potential defined by \({w(\rho )}_\ell :=\sum _{m \not = \ell } \frac{\rho _m  Z_m}{r_{\ell m}} e^{\tau \, r_{\ell m}}\) (for some \(\tau > 0\)) also fits into this general framework. This setting already covers many important modelling scenarios and also serves as a crucial stepping stone towards charge equilibration under full Coulomb interaction, which goes beyond the scope of the present work.
The main result of this section is the following: if there exists a selfconsistent solution \(\rho ^\star \) to (2.31), then we can approximate \(\rho ^\star \) with selfconsistent solutions to the following approximate selfconsistency equation:
for sufficiently large N. The operator \(I_{X_N} F^\beta \) is a linear bodyordered approximation of the form we analyzed in detail in § 2.4.
To do this, we require a natural stability assumption on the electronic structure problem, which was employed for example in [93, 99, 100]:
(STAB) The stability operator \({\mathscr {L}}(\rho )\) is the Jacobian of \(\rho \mapsto F^\beta \big ( {\varvec{u}}(\rho ) \big )\). We say electron densities \(\rho ^\star \) solving (2.31) are stable if \(I  {\mathscr {L}}(\rho ^\star )\) is invertible as a bounded linear operator \(\ell ^2 \rightarrow \ell ^2\).
Remark 12
(Stability) (i) The stability condition of Theorem 2.6 is a minimal starting assumption that naturally arises from the analysis [93, 99, 100].
For example, if \(\rho \) is a stable selfconsistent electron density, then there exists \(\phi ^{(m)} \in \ell ^2(\Lambda )\) such that [93]:
(ii) As noted in [99] (in a slightly simpler setting), the stability condition of Theorem 2.6 is automatically satisfied for multilattices with \(\nabla w\) positive semidefinite. In fact, in this case the stability operator is negative semidefinite.
Theorem 2.6
For \({\varvec{u}}\) satisfying Definition 1, suppose that \(\rho ^\star \) is a corresponding stable selfconsistent electron density.
Then, for N sufficiently large, there exist selfconsistent solutions \(\rho _{N}\) of (2.32) such that
where \(\gamma _N\) are the constants from Theorem 2.3 applied to \({\varvec{u}}(\rho ^\star )\).
Corollary 2.7
Suppose that \(\rho ^\star \) and \(\rho _N\) are as in Theorem 2.6 and denote by \(O^\mathrm {sc}_\ell (\varvec{u}) :=O_\ell \big ( {\varvec{u}}(\rho ^\star ) \big )\) a selfconsistent local observable as in (2.7). Then,
where \(\gamma _N\) are the constants from Theorem 2.3 applied to \({\varvec{u}}(\rho ^\star )\).
In order for this result to be of any practical use, we need to solve the nonlinear equation (2.32) for the electron density via a selfconsistent field (SCF) procedure. Supposing we have the electron density \(\rho ^{i}\) and corresponding state \({\varvec{u}}^{i} :={\varvec{u}}(\rho ^{i})\) after i iterations, we diagonalise the Hamiltonian \(\mathcal {H}({\varvec{u}}^{i})\) and hence evaluate the output density \(\rho ^\mathrm {out} = I_{X_N}{F}^\beta ({\varvec{u}}^{i})\). At this point, since the simple iteration \(\rho ^{i+1} = \rho ^{\mathrm {out}}\) does not converge in general, a mixing strategy, possibly combined with Anderson acceleration [19], is used in order to compute the next iterate. The analysis of such mixing schemes is a major topic in electronic structure and numerical analysis in general and so we only present a small step in this direction.
Proposition 2.8
(Stability) The approximate electron densities \(\rho _N\) arising from Theorem 2.6 are stable in the following sense: \(I  {\mathscr {L}}_N(\rho _N) :\ell ^2 \rightarrow \ell ^2\) is an invertible bounded linear operator where \({\mathscr {L}}_N\) is the Jacobian of \(\rho \mapsto I_{X_N} F^\beta \big ( {\varvec{u}}(\rho ) \big )\). Moreover, \(\big ( I  {\mathscr {L}}_N(\rho _N) \big )^{1}\) is uniformly bounded in N in operator norm.
Theorem 2.9
For \({\varvec{u}}\) satisfying Definition 1, suppose that \(\rho _{N}\) is a corresponding approximate selfconsistent electron density stable in the sense of Proposition 2.8. For fixed \(\rho ^0\), we define \(\{\rho ^i\}_{i = 0}^\infty \) via the Newton iteration
Then, for \(\Vert \rho ^0  \rho _{N}\Vert _{\ell ^\infty }\) sufficiently small, the Newton iteration converges quadratically to \(\rho _{N}\).
A more thorough treatment of these SCF results is beyond the scope of this work. See [12, 53, 63] for recent results in the context of HartreeFock and KohnSham density functional theory. For a recent review of SCF in the context density functional theory, see [101].
Remark 13
It is clear from the proofs of Theorems 2.6 and 2.9 that as long as the approximate scheme \(F^{\beta ,N}\) satisfies
then we may approximate (2.31) with approximate selfconsistent solutions \(\rho _N = F^{\beta ,N}\big ({\varvec{u}}(\rho _N)\big )\). In particular, as long as we have the estimate from Remark 10 (see Appendix C for the technical details), then we may use the nonlinear approximation scheme \({\Theta _N}\) from Theorem 2.5 in Theorems 2.6 and 2.9 . In this case, we obtain error estimates that are (asymptotically) independent of the discrete spectrum.
Remark 14
In the linearscaling spectral Gauss quadrature (LSSGQ) method [87], a selfconsistent field iteration analogous to (2.32) is proposed. In particular, with the caveats outlined in Remark 13 taken into consideration, Theorem 2.6 goes some way to rigorously justify the exponential rate of convergence observed numerically in [87, Fig. 4].
3 Conclusions and Discussion
The main result of this work is a sequence of rigorous results about bodyordered approximations of a wide class of properties extracted from tightbinding models for condensed phase systems, the primary example being the potential energy landscape. Our results demonstrate that exponentially fast convergence can be obtained, provided that the chemical environment is taken into account. In the spirit of our previous results on the locality of interaction [16, 17, 93], these provide further theoretical justification—albeit qualitative—for widely assumed properties of atomic interactions. More broadly, our analysis illustrates how to construct general lowdimensional but systematic representations of highdimensional complex properties of atomistic systems. Our results, as well as potential generalisations, serve as a starting point towards a rigorous endtoend theory of multiscale and coarsegrained models, including but not limited to machinelearned potential energy landscapes.
In the following paragraphs we will make further remarks on the potential applications of our results, and on some apparent limitations of our analysis.
3.1 Representation of atomic properties
Our initial motivation for studying the bodyorder expansion was to explain the (unreasonable?) success of machinelearned interatomic potentials [5, 8, 80], and our remarks will focus on this topic, however in principle they apply more generally.
Briefly, given an ab initio potential energy landscape (PEL) \(E^\mathrm{QM}\) for some material one formulates a parameterised interatomic potential
and then “learn” the parameters \({\varvec{\theta }}\) by fitting them to observations of the reference PEL \(E^\mathrm{QM}\). A great variety of such parameterisations exist, including but not limited to neural networks [8], kernel methods [5] and symmetric polynomials [2, 25, 80]. Symmetric polynomials are linear regression schemes where each basis function has a natural bodyorder attached to it. It is particularly striking that for very low bodyorders of four to six these schemes are able to match and often outperform the more complex nonlinear regression schemes [66, 80, 106]. Our analysis in the previous sections provides a partial explanation for these results, by justifying why one may expect that a reference ab initio PEL intrinsically has a low bodyorder. Moreover, classical approximation theory can now be applied to the bodyordered components as they are finitedimensional to obtain new approximation results where the curse of dimensionality is alleviated.
Our results on nonlinear representations are less directly applicable to existing MLIPs, but rather suggest new directions to explore. Still, some connections can be made. The BOPtype construction of § 2.5,
points towards a blending of machinelearning and BOP techniques that have not been explored to the best of our knowledge. A second interesting connection is to the overlapmatrix based fingerprint descriptors (OMFPs) introduced in [105] where a global spectrum for a small subcluster is used as a descriptor, while (3.1) can be understood as taking the projected spectrum as the descriptor. Thus, Theorem 2.5 suggests (1) an interesting modification of OMFPs which comes with guaranteed completeness to describe atomic properties; and (2) a possible pathway towards proving completeness of the original OMFPs.
Finally, our selfconsistent representation of § 2.7 motivates how to construct compositional models, reminiscent of artificial neural networks, but with minimal nonlinearity that is moreover physically interpretable. Although we did not pursue it in the present work, this is a particularly promising starting point to incorporate meaningful electrostatic interaction into the MLIPs framework.
3.2 Linear bodyordered approximation: the preasymptotic regime
Possibly the most significiant limitation of our analysis of the linear bodyordered approximation scheme is that the estimates deteriorate when defects cause a pollution of the point spectrum. Here, we briefly demonstrate that this appears to be an asymptotic effect, while in the preasymptotic regime this deterioration is not noticable.
To explore this we choose a union of intervals \(E \supseteq \sigma (\mathcal {H})\) and a polynomial \(P_N\) of degree N and note
We then construct interpolation sets (Fejér sets) such that the corresponding polynomial interpolant gives the optimal asymptotic approximation rates (for details of this construction, see §4.1.5§4.1.8). We then contrast this with a best \(L^\infty (E)\)approximation, and with the nonlinear approximation scheme from Theorem 2.5. We will observe that the nonlinearity leads to improved asymptotic but comparable preasymptotic approximation errors.
As a representative scenario we consider the FermiDirac distribution \(F^\beta (z) = (1 + e^{\beta z})^{1}\) with \(\beta = 100\) and both the “defectfree” case \(E_1 :=[1,a]\cup [b,1]\) and \(E_2 :=[1,a] \cup [c,d] \cup [b,1]\) with the parameters \(a = 0.2, b = 0.2, c = 0.06\), and \(d = 0.03\). Then, for fixed polynomial degree N and \(j \in \{1,2\}\), we construct the \((N+1)\)point Fejér set for \(E_j\) and the corresponding polynomial interpolant \(I_{j,N}F^\beta \). Moreover, we consider a polynomial \(P^\star _{j,N}\) of degree N minimising the right hand side of (3.2) for \(E = E_j\). Then, in Figure 3, we plot the errors \(\Vert F^\beta  I_{j,N} F^\beta \Vert _{L^\infty (E_j)}\) and \(\Vert F^\beta  P_{j,N}^\star \Vert _{L^\infty (E_j)}\) for both \(j = 1\) (Fig. 3a) and \(j=2\) (Fig. 3b) against the polynomial degree N together with the theoretical asymptotic convergence rates for best \(L^\infty (E_j)\) polynomial approximation (4.15).
What we observe is that, as expected, introducing the interval [c, d] into the approximation domain drastically affects the asymptotic convergence rate and the errors in the approximation based on interpolation. While the best approximation errors follow the asymptotic rate for larger polynomial degree, it appears that, preasymptotically, the errors are significantly reduced. We also see that the approximation errors are significantly better than the general error estimate \(\Vert F^\beta  \Pi _N F^\beta \Vert _{L^\infty } \lesssim e^{\pi \beta ^{1} N}\) where \(\Pi _N\) is the Chebyshev projection operator (see § 4.1.4).
Moreover, in Figure 3b, we plot the errors when using a nonlinear approximation scheme satisfying Theorem 2.5. In this simple experiment, we consider the Gauss quadrature rule \(\Theta _{N} :=\int I_{X_{\frac{1}{2}(N1)}} F^\beta \mathrm {d}D_\ell \) where \(X_{\frac{1}{2}(N1)}\) is the set of zeros of the degree \(\frac{1}{2}(N+1)\) orthogonal polynomial (see Appendix D) with respect to \( \mathrm {d}D_\ell (x) :=\big ( \chi _{E_1}(x) + \sum _{j} \delta (x  \lambda _j) \big ) \mathrm {d}x \) where \(\{\lambda _j\} = \{c,\tfrac{1}{2}(c+d),d\} \subset [c,d]\). While \(D_\ell \) does not correspond to a physically relevant Hamiltonian, the same procedure may be carried out for any measure supported on \(E_1\) with \(\mathrm {supp}\, D_\ell \cap [c,d]\) finite. Then plotting the errors \(  F^\beta _\ell  \Theta _N\), we observe improved asymptotic convergence rates that agree with that of the “defectfree” case from Figure 3a. However, the improvement is only observed in the asymptotic regime which corresponds to bodyorders never reached in practice.
4 Proofs
4.1 Preliminaries
Here, we introduce the concepts needed in the proofs of the main results.
4.1.1 Hermite integral formula
For a finite interpolation set \(X\subset {\mathbb {C}}\), we let \(\ell _X(z) :=\prod _{x\in X}(z  x)\) be the correpsonding node polynomial.
For fixed \(z\in {\mathbb {C}} \setminus X\), we suppose that O is analytic on an open neighbourhood of \(X\cup \{z\}\). Then, for a simple closed positively oriented contour (or system of contours) \({\mathscr {C}}\) contained in the region of analyticity of O, encircling X, and avoiding \(\{z\}\), we have
If, in addition, \({\mathscr {C}}\) encircles \(\{z\}\), then
The proof of these facts is a simple application of Cauchy’s integral formula, [3, 95].
4.1.2 Resolvent calculus
Given a configuration \({\varvec{u}}\), we consider the Hamiltonian \(\mathcal {H}= \mathcal {H}({\varvec{u}})\) and functions O analytic in some neighbourhood of the spectrum \(\sigma (\mathcal {H})\). We define \(O(\mathcal {H})\) via the holomorphic functional calculus [1]:
where \({\mathscr {C}}\) is a simple closed positively oriented contour (or system of contours) contained in the region of analyticity of O and encircling the spectrum \(\sigma (\mathcal {H})\).
The following Combes–Thomas resolvent estimate [21] will play a key role in the analysis:
Lemma 1
(CombesThomas) Suppose that \({\varvec{u}}\) satisfies Definition 1 and \(z \in {\mathbb {C}}\) is contained in a bounded set with \(\mathrm {dist}\left( z, \sigma \big (\mathcal {H}({\varvec{u}})\big )\right) \!>\!0\) and \(\mathfrak {d}\!:=\! \mathrm {dist}\left( z, \sigma \big (\mathcal {H}({\varvec{u}}^\mathrm {ref})\big )\right) > \delta \).
Then, there exists a constant \(C>0\) such that
and \(\gamma _\mathrm {CT} :=c \min \{1, \mathfrak {d}\}\) and \(c>0\) depends on \(h_0, \gamma _0, d\) and \(\min _{\ell \not =k} r_{\ell k}\).
Proof
A proof with \(\gamma _{\mathrm {CT}}\) depending instead on \(\mathrm {dist}\left( z, \sigma \big (\mathcal {H}({\varvec{u}})\big )\right) \) can be found in [16]. A lowrank update formula leads to the improved “defectindependent” result [17] where the exponent only depends on the distance between z and the reference spectrum. See [93] for an explicit description of \(\gamma _\mathrm {CT}\) in terms of the constants \(\gamma _0, d\) and the noninterpenetration constant \(\min _{\ell \not =k} r_{\ell k}\). \(\square \)
A key observation for arguments involving forces (or more generally, derivatives of the analytic quantities of interest) is that the CombesThomas estimate allows us to bound derivatives of the resolvent operator:
Lemma 2
Suppose that \(z \in {\mathbb {C}}\) with \(\mathfrak {d}:=\mathrm {dist}\left( z, \sigma \big (\mathcal {H}({\varvec{u}})\big )\right) > 0\). Then,
where \(\gamma _\mathrm {CT}\) is the CombesThomas constant from Lemma 1 and \(\gamma _0\) is the constant from (TB).
Proof
This result can be found in the previous works [14, 16, 17], but we give a brief sketch for completeness.
Derivatives of the resolvent have the following form:
The result follows by applying the CombesThomas resolvent estimates together with the fact that the Hamiltonian is shortranged (TB).
Assuming that the Hamiltonian has higher derivatives that are also shortranged, higher order derivatives of the resolvent can be treated similarly [16]. \(\square \)
4.1.3 Local observables
Firstly, we note that \(F^\beta (\,\cdot \,)\) is analytic away from the simple poles at \(\pi \beta ^{1} (2\mathbb {Z} + 1)\). Moreover, \(G^\beta (\,\cdot \,)\) can be analytically continued onto the open set \({\mathbb {C}} \setminus \left\{ \mu + i r :r\in {\mathbb {R}}, r \geqslant \pi \beta ^{1} \right\} \) [17]. Therefore, we may consider (4.3) with \(O = F^\beta \) or \(G^\beta \) and a contour \({\mathscr {C}}_\beta \) encircling \(\sigma \left( \mathcal {H}\right) \) and avoiding \({\mathbb {C}} \setminus \left\{ \mu + i r :r\in {\mathbb {R}}, r \geqslant \pi \beta ^{1} \right\} \). Therefore, we may choose \({\mathscr {C}}_\beta \) so that the constant \(\mathfrak {d}\), from Lemma 1, is proportional to \(\beta ^{1}\). Moreover, if there is a spectral gap, the constant \(\mathfrak {d}\) is uniformly bounded below by a positive constant multiple of \(\mathsf {g}\) as \(\beta \rightarrow \infty \).
In the case of insulators at zero Fermitemperature, we take \({\mathscr {C}}_\infty \) encircling \(\sigma \left( \mathcal {H}({\varvec{u}})\right) \cap (\infty , \mu )\) and avoiding the rest of the spectrum. Therefore, we may choose \({\mathscr {C}}_\infty \) so that the constant \(\mathfrak {d}\), from Lemma 1, is proportional to \(\mathsf {g}\).
Following [16, Lemma 4], we can conclude that \(\sigma (\mathcal {H}) \subset [\underline{\sigma },\overline{\sigma }]\) for some \(\underline{\sigma }, \overline{\sigma }\) depending on \(h_0, \gamma _0, v, d\) and \(\min _{\ell \not =k} r_{\ell k}\). This means that, the contours \({\mathscr {C}}_\beta \) can be chosen to have finite length and, when applying Lemma 1, we have \(\gamma _\mathrm {CT} = c \min \{1, \max \{ \beta ^{1}, \mathsf {g} \}\}\).
Moreover, for all \(0< \mathsf {b} < \pi \) and bounded sets \(A_\beta \subset A \subset {\mathbb {C}}\) such that
both \(F^\beta (\,\cdot \,)\) and \(G^\beta (\,\cdot \,)\) are uniformly bounded on \(A_\beta \) independently of \(\beta \) [17, Lemma 5.2].
4.1.4 Chebyshev Projection and Interpolation in Chebyshev Points
We denote by \(\{T_n\}\) the Chebyshev polynomials (of the first kind) satisfying \(T_n(\cos \theta ) = \cos n\theta \) on \([1,1]\) and, equivalently, the recurrence \(T_{0} = 1, T_1 = x\), and \(T_{n+1}(x) = 2x T_n(x)  T_{n1}(x)\).
For O Lipshitz continuous on \([1,1]\), there exists an absolutely convergent Chebyshev series expansion: there exists \(c_n\) such that \(O(z) = \sum _{n=0}^\infty c_n T_n(z)\). For maximal polynomial degree N, the corresponding projection operator is denoted \(\Pi _N O(z) :=\sum _{n=0}^N c_n T_n(z)\). This approach is a special case of the Kernel Polynomial Method (KPM) which we briefly review in Appendix F.
On the other hand, supposing that the interpolation set is given by the Chebyshev points \(X = \{ \cos \frac{j\pi }{N} \}_{0 \leqslant j \leqslant N}\), we may expand the polynomial interpolant \(I_N O :=I_{X} O\) in terms of the Chebyshev polynomials: there exists \(c_n^\prime \) such that \(I_N O(z) = \sum _{n=0}^N c_n^\prime T_n(z)\).
For functions O that can be analytically continued the Bernstein ellipse \(E_\rho :=\{ \frac{1}{2}( z + z^{1} ) :z = \rho \}\) for \(\rho > 1\), the corresponding coefficients \(\{c_n\}\), \(\{c^\prime _n\}\) decay exponentially with rate \(\rho \). This leads to the following error estimates
For \(O^\beta = F^\beta \) or \(G^\beta \), these estimates give an exponential rate of convergence with exponent depending on \(\sim \beta ^{1}\). Indeed, after scaling \(\mathcal {H}\) so that the spectrum is contained in \([1,1]\), we obtain
and we conclude by directly applying (4.5). The same estimate also holds for \(I_N\) (or any polynomial).
For full details of all the statements made in this subsection, see [95].
4.1.5 Classical logarithmic potential theory
In this section, we give a very brief introduction to classical potential theory in order to lay out the key notation. For a more thorough treatment, see [75] or [37, 62, 76, 95].
It can be seen from the Hermite integral formula (4.2) that the approximation error for polynomial interpolation may be determined by taking the ratio of the size of the node polynomial \(\ell _X\) at the approximation points to the size of \(\ell _X\) along an appropriately chosen contour. Logarithmic potential theory provides an elegant mechanism for choosing the interpolation points so that the asymptotic behaviour of \(\ell _X\) can be described.
We suppose that \(E \subset {\mathbb {C}}\) is a compact set. We will see that choosing the interpolation nodes as to maximise the geometric mean of pairwise distances provides a particularly good approximation scheme:
Any set \({\mathcal {F}}_n \subset E\) attaining this maximum is known as a Fekete set. It can be shown that the quantities \(\delta _n(E)\) form a decreasing sequence and thus converges to what is known as the transfinite diameter: \(\tau (E) :=\lim \limits _{n\rightarrow \infty } \delta _n(E)\).
We let \(\ell _n(z)\) denote the node polynomial corresponding to a Fekete set and note that
Therefore, rearranging (4.8), we obtain \(\lim _{n\rightarrow \infty } \Vert \ell _n\Vert _{L^\infty (E)}^{1/n} \leqslant \tau (E)\). In fact, this inequality can be replaced with equality, showing that Fekete sets allow us to describe the asymptotic behaviour of the node polynomials on the domain of approximation.
To extend these results, it is useful to recast the maximisation problem (4.7) into the following minimisation problem, describing the minimal logarithmic energy attained by n particles lying in E with the repelling force \(1/z_i  z_j\) between particles i and j lying at positions \(z_i\) and \(z_j\), respectively:
Fekete sets can therefore be seen as minimal energy configurations and described by the normalised counting measure \(\nu _n :=\frac{1}{n} \sum _{j = 1}^n \delta _{z_j}\) where \({\mathcal {F}}_n = \{z_j\}_{j = 1}^n\).
The minimisation problem (4.9) may be extended for general unit Borel measures \(\mu \) supported on E by defining the logarithmic potential and corresponding total energy by
The infimum of the energy over the space of unit Borel measures supported on E, known as the Robin constant for E, will be denoted \(\infty < V_E \leqslant +\infty \). The capacity of E is defined as \(\mathrm {cap}(E):=e^{V_E}\) and is equal to the transfinite diameter [34]. Using a compactness argument, it can be shown that there exists an equilibrium measure \(\omega _E\) with \(I(\omega _E) = V_E\) and, in the case \(V_E <\infty \), by the strict convexity of the integral, \(\omega _E\) is unique [77]. Moreover, if \(V_E < \infty \) (equivalently, if \(\mathrm {cap}(E) > 0\)), then \(U^{\omega _E}(z) \leqslant V_E\) for all \(z \in {\mathbb {C}}\), with equality holding on E except on a set of capacity zero (we say this property holds quasieverywhere).
Moreover, if \(\mathrm {cap}\,E > 0\), then it can be shown that the normalised counting measures, \(\nu _n\), corresponding to a sequence of Fekete sets weak\(\star \) converges to \(\omega _E\). Since \(U^{\nu _n}(z) = \frac{1}{n}\log \frac{1}{\ell _n(z)}\), the weak\(\star \) convergence allows one to conclude that
uniformly on compact subsets of \({\mathbb {C}} \setminus E\). Here, we have defined the Green’s function \(g_E(z) :=V_E  U^{\omega _E}(z)\), which describes the asymptotic behaviour of the node polynomials corresponding to Fekete sets. We therefore wish to understand the Green’s function \(g_E\).
4.1.6 Construction of the Green’s function
Now we restrict our attention to the particular case where \(E \subset {\mathbb {R}}\) is a union of finitely many compact intervals of nonzero length.
It can be shown that the Green’s function \(g_E\) satisfies the following Dirichlet problem on \({\mathbb {C}} \setminus E\) [75]:
In fact, it can be shown that (4.11) admits a unique solution [75] and thus (4.11) is an alternative definition of the Green’s function. Using this characterisation, it is possible to explicitly construct the Green’s function \(g_E\) as follows. In the upper half plane, \(g_E(z) = \mathrm {Re}(G_E(z))\) where \(G_E:\{ z \in {\mathbb {C}} :\mathrm {Im}(z) \geqslant 0 \} \rightarrow \{ z\in {\mathbb {C}} :\mathrm {Re}(z) \geqslant 0, \mathrm {Im}(z) \in [0,\pi ]\}\) is a conformal mapping on \(\{z :\mathrm {Im}(z) > 0\}\) such that \(G_E(E) = i[0,\pi ]\), \(G_E( \min E ) = i \pi \), and \(G_E( \max E ) = 0\). Using the symmetry of E with respect to the real axis, we may extend \(\mathrm {Re}(G_E(z))\) to the whole complex plane via the Schwarz reflection principle. Then, one can easily verify that this analytic continuation satisfies (4.11). Since the image of \(G_E\) is a (generalised) polygon, \(z \mapsto G_E(z)\) is an example of a Schwarz–Christoffel mapping [29]. See Figure 4 for the case \(E = [1,\varepsilon ]\cup [\varepsilon ,1]\).
We shall briefly discuss the construction of the Schwarz–Christoffel mapping \(G_E\) for \(E = [1,\varepsilon _]\cup [\varepsilon _+,1]\). We define the prevertices \(z_1 = 1, z_2 = \varepsilon _, z_4 = \varepsilon _+, z_5 = 1\) and wish to construct a conformal map \(G_E\) with \(G_E(z_k) = \omega _k\) as in Figure 4. For simplicity, we also define \(z_0 :=\infty \) and \(z_6 :=\infty \) and observe that because the image is a polygon, \(\mathrm {arg} \, G_E^\prime (z)\) must be constant on each interval \((z_{k1}, z_k)\) and
where \(z_k^ \in (z_{k1}, z_k)\), \(z_k^+ \in (z_{k}, z_{k+1})\), and \(\alpha _k \pi \) is the interior angle of the infinite slit strip at vertex \(\omega _k\) (that is, \(\alpha _1 = \alpha _2 = \alpha _4 = \alpha _5 = \frac{1}{2}\) and \(\alpha _3 = 2\)). After defining \(z^\alpha :=z^\alpha e^{i \alpha \,\mathrm {arg}\,z}\) where \(\mathrm {arg}\,z \in (\pi ,\pi ]\), we can see that for \(z \in (z_{k1}, z_k)\), we have \(\mathrm {arg} \prod _{j = k}^5 (z  z_j)^{\alpha _j  1} = \sum _{j = k}^{5} (\alpha _j  1)\pi \) and so the jump in the argument of \(z \mapsto \prod _{j = 1}^5 (z  z_j)^{\alpha _j  1}\) is \((1  \alpha _k)\pi \) at \(z_k\) as in (4.12). Therefore, integrating this expression, we obtain
Since \(G_E(1) = A\), we take \(A = 0\) (to ensure (4.11c) holds). Moreover, since the real part of the integral is \(\sim \log z\) as \(z \rightarrow \infty \), we apply (4.11b) to conclude \(B = 1\). Finally, we can choose \(z_3\) such that \(\mathrm {Re}\,G_E(z) = 0\) for all \(z \in E\); that is,
For more details, see [37]. We use the Schwarz–Christoffel toolbox [29] in matlab to evaluate (4.13) and plot Figure 5.
For the simple case \(E :=[1,1]\), by the same analysis, we can disregard \(z_2,z_3,z_4\) and \(\omega _2,\omega _3,\omega _4\) and integrate the corresponding expression to obtain the closed form \(G_{[1,1]}(z) = \log ( z + \sqrt{z  1}\sqrt{z + 1})\).
A similar analysis allows one to construct conformal maps from the upper half plane to the interior of any polygon. For further details, rigorous proofs and numerical considerations, see [31].
4.1.7 Interpolation nodes
The only difficulty in obtaining (4.10) in practice is the fact that Fekete sets are difficult to compute. An alternative, based on the Schwarz–Christoffel mapping \(G_E\), are Fejér points. For equally spaced points \(\{\zeta _j\}_{j=1}^{n}\) on the interval \(i [0,\pi ]\), the \(n^\text {th}\) Fejér set is defined by \(\{ G_E^{1}(\zeta _j)\}_{j=1}^n\). Fejér sets are also asymptotically optimal in the sense that (4.10) is satisfied where \(\ell _n\) is now the node polynomial corresponding to npoint Fejér set.
Another approach is to use Leja points which are generated by the following algorithm: for fixed \(z_1,\dots ,z_n\), the next interpolation node \(z_{n+1}\) is constructed by maximising \(\prod _{j = 1}^n z_j  z\) over all \(z \in E\). Sets of this form are also asymptotically optimal [90] for any choice of \(z_1 \in E\). Since we have fixed the previous nodes \(z_1,\dots ,z_{n}\), the maximisation problem for constructing \(z_{n+1}\) is much simpler than that of (4.7).
More generally, if the normalised counting measure corresponding to a sequence of sets \(\{z_j\}_{j=1}^n \subset E\) weak\(\star \) converges to the equilibrium measure \(\omega _E\), then the corresponding node polynomials satisfy (4.10).
For the simple case where \(E=[1,1]\), many systems of zeros or maxima of sequences of orthogonal polynomials are asymptotically optimal in the sense of (4.10). In fact, since the equilibrium measure for \([1,1]\) is the arcsine measure [76]
any sequence of sets with this limiting distribution is asymptotically optimal. An example of particular interest are the Chebyshev points \(\{ \cos \frac{j\pi }{n} \}_{0 \leqslant j \leqslant n}\) given by the \(n+1\) extreme points of the Chebyshev polynomials defined by \(T_n( \cos \theta ) = \cos n \theta \).
4.1.8 Asymptotically optimal polynomial approximations
Suppose that E is the union of finitely many compact intervals of nonzero length and \(O:E \rightarrow {\mathbb {C}}\) extents to an analytic function in an open neighbourhood of E. On defining \({\mathscr {C}}_\gamma :=\{ z \in {\mathbb {C}} :g_E(z) = \gamma \}\), we denote by \(\gamma ^\star \) the maximal constant for which O is analytic on the interior of \({\mathscr {C}}_{\gamma ^\star }\). We let \(P_N^\star \) be the best \(L^\infty (E)\)approximation to O in the space of polynomials of degree at most N and suppose that \(I_N\) is a polynomial interpolation operator in \(N+1\) points satisfying (4.11). Then, the Green’s function \(g_E\) determines the asymptotic rate of approximation for not only polynomial interpolation, but also for best approximation:
For a proof that the asymptotic rate of best approximation is given by the Green’s function see [76]. The result for polynomial interpolation uses the Hermite integral formula and (4.10), see (4.20) and (4.22), below.
4.2 Linear bodyorder approximation
In this section, we use the classical logarithmic potential theory from § 4.1.5 to prove the approximation error bounds for interpolation. However, we first show that polynomial approximations lead to bodyorder approximations:
Proof of Proposition 2.2
We first simplify the notation by absorbing the effective potential and twocentre terms into the threecentre summation:
Now, supposing that \(I_XO(z) = \sum _{j = 0}^{X1} c_j z^j\), we obtain
there the first two terms in the outer summation are \(c_0\) and \(c_1 \mathcal {H}_{\ell \ell }\). Now, for a fixed bodyorder \((n+1)\), and \(k_1< \dots < k_{n}\) with \(k_l \not =\ell \), we construct \(V_{nN}({\varvec{u}}_\ell ; {\varvec{u}}_{\ell k_1}, \dots , {\varvec{u}}_{\ell k_n})\) by collecting all terms in (4.17) with \(0 \leqslant j \leqslant X1\) and \(\{ \ell , \ell _1, \dots , \ell _{j1}, m_1, \dots , m_j\} = \{\ell , k_1, \dots , k_n\}\). In particular, the maximal bodyorder in this expression is \(2(X1)\) for threecentre models and \(X1\) in the twocentre case.
More explicitly, using the notation (2.26), we have that
Here, we have applied an inclusionexclusion principle to ensure that we are not only summing over sites in \(\{k_1,\dots ,k_n\}\) but we select at least one of each site in this set. Indeed, if we choose \(\ell _1, \dots , \ell _{j1}, m_1, \dots , m_j\) such that \(\{\ell ,\ell _1, \dots , \ell _{j1}, m_1, \dots , m_j\} = \{\ell \}\cup K_0\), then the expression \(\mathcal {H}_{\ell \ell _1 m_1} \mathcal {H}_{\ell _1 \ell _2 m_2} \cdots \mathcal {H}_{\ell _{j1} \ell m_j}\) appears in each term of (4.19) with \(K \supseteq K_0\) exactly once (with a ± sign). Therefore, the number of times \(\mathcal {H}_{\ell \ell _1 m_1} \mathcal {H}_{\ell _1 \ell _2 m_2} \dots \mathcal {H}_{\ell _{j1} \ell m_j}\) appears is exactly
That is, (4.19) only contains the terms in the summation (4.18). \(\square \)
Proof of Theorem 2.3
We let \(\ell _N(x) :=\prod _{j} (x  x_j^N)\) be the node polynomial for \(X_N :=\{x_j^N\}_{j=0}^N\). Again, we fix the configuration \({\varvec{u}}\) and consider \(\mathcal {H}:=\mathcal {H}({\varvec{u}})\).
Supposing that \({\mathscr {C}}\) is a simple closed positively oriented contour encircling \(\sigma (\mathcal {H})\), we apply the Hermite integral formula (4.2) to obtain that
where
At this point we apply standard results of classical logarithmic potential theory (see, § 4.1.5 or [62]) and conclude by noting that if the interpolation points are asymptotically distributed according to the equilibrium distribution corresponding to \(E:=I_ \cup I_+\), then after applying (4.10), we have that
Here, the equilibrium distribution and the Green’s function \(g_E(z)\) are concepts introduced in § 4.1.5 and § 4.1.6.
Therefore, by choosing the contour \({\mathscr {C}} :=\{ \xi \in {\mathbb {C}} :g_E(\xi ) = \gamma \}\) for \(0< \gamma < g_E(\mu + i\pi \beta ^{1})\), the asymptotic exponents in the approximation error is \(\gamma \).
The maximal asymptotic convergence rate is given by \(g_E(\mu + i\pi \beta ^{1})\) since \({\mathscr {C}}\) must be contained in the region of analyticity of \(O^\beta \) and the first singularity of \(O^\beta \) is at \(\mu + i\pi \beta ^{1}\) (for \(O^\beta = F^\beta \) or \(G^\beta \)).
Examples of the equipotential level sets \({\mathscr {C}}\) are given in Figure 5.
Using the Green’s function results of § 4.1.6, \(g_E( \mu + i \pi \beta ^{1} ) = \mathrm {Re}\, G_E( \mu + i \pi \beta ^{1} )\) where \(G_E\) is the integral (4.13). The asymptotic behaviour of this maximal asymptotic convergence rate for the separate \(\beta \rightarrow \infty \) and \(\mathsf {g}\rightarrow 0\) limits can be found in [37, 81]. Here, we consider the \(\beta ^{1} + \mathsf {g} \rightarrow 0\) limit where the gap remains symmetric about the chemical potential \(\mu \).
To simplify the notation we consider \(I_ \cup I_+ = [1,\varepsilon _] \cup [\varepsilon _+,1]\) where \(\varepsilon _\pm = \mu \pm \frac{1}{2}\mathsf {g}\). By choosing to integrate (4.13) along the contour composed of the intervals \([1,\mu ]\) and \([\mu , \mu + i\pi \beta ^{1}]\), we obtain
Since \(g_E(\mu ) \sim \mathsf {g}\) as \(\mathsf {g} \rightarrow 0\) [37], we only consider the remaining term in (4.23).
For \(\zeta \in \mu + i [0,\pi \beta ^{1} ]\), we have \(c^{1} \leqslant \sqrt{\zeta \pm 1} \leqslant c\), and so the integral in (4.23) has the same asymptotic behaviour as
where we have used the change of variables \(\widetilde{\zeta } = \frac{\zeta  \varepsilon _}{\varepsilon _+\varepsilon _}\).
Since the integrands are uniformly bounded along the domain of integration, (4.24) is \(\sim \beta ^{1}\) as \(\beta \rightarrow \infty \).
The constant prefactor in (4.21) is inversely proportional to the distance \(\mathrm {dist}\big ( {\mathscr {C}}, \sigma (\mathcal {H}) \big )\) between the contour \({\mathscr {C}} = \{ g_E = \gamma \}\) and the spectrum \(\sigma (\mathcal {H})\). In particular, since \(g_E\) is uniformly Lipschitz with constant \(L>0\) on the compact region bounded by \({\mathscr {C}}\), we have: there exists \(\lambda \in \sigma (\mathcal {H})\) and \(\xi \in {\mathscr {C}}\) such that
Therefore, choosing \(\gamma \) to be a constant multiple of \(g_E(\mu + i\pi \beta ^{1})\), we conclude that the constant prefactor C satisfies \(C \sim (\mathsf {g} + \beta ^{1})^{1}\) as \(\mathsf {g} + \beta ^{1} \rightarrow 0\).
To extend the bodyorder expansion results to derivatives (in particular, to forces), we write the quantities of interest using resolvent calculus, apply Lemma 2 to bound the derivatives of the resolvent, and use the Hermite integral formula (4.20) to conclude: for \({\mathscr {C}}_1\), \({\mathscr {C}}_2\) simple closed positively oriented contours encircling the spectrum \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big )\) and \({\mathscr {C}}_1\), respectively, we have
We conclude by choosing appropriate contours \({\mathscr {C}}_l = \{g_E = \gamma _l\}\) for \(l=1,2\) and applying (4.22). \(\square \)
4.2.1 The role of the point spectrum
To begin this section, we sketch the proof of Proposition 2.1.
Proof of Proposition 2.1
(i) Supnorm perturbations. We suppose that \(\sup _k \big [ \varvec{r}_k  \varvec{r}^{\mathrm {ref}}_k + v_k  v_k^{\mathrm {ref}} \big ] \leqslant \delta \) for \(\delta >0\) sufficiently small such that
where, \(\xi _{\ell k} \in [\varvec{r}_{\ell k}, \varvec{r}_{\ell k}^{\mathrm {ref}}]\), \(\xi ^{(l)}_{\ell m} \in [\varvec{r}_{\ell m}, \varvec{r}_{\ell m}^{\mathrm {ref}}]\), and \(\zeta ^{(l)}_{k m} \in [\varvec{r}_{km}, \varvec{r}_{km}^{\mathrm {ref}}]\). Therefore, if \(\psi \in \ell ^2\), we have
Therefore, applying standard results from perturbation theory [56, p. 291], we obtain
(ii) Finite rank perturbations. The finite rank perturbation result has been presented in [70] in a slightly different setting. We sketch the main idea here for completeness.
Since the essential spectrum is stable under compact (in particular, finite rank) perturbations [56], the set
is both compact and discrete and therefore finite. \(\square \)
Proof of Theorem 2.4
Suppose that \({\mathscr {C}}\) is a simple closed contour encircling the spectrum \(\sigma \big (\mathcal {H}({\varvec{u}})\big )\) and \((\lambda _s, \psi _s)\) are normalised eigenpairs corresponding to the finitely many eigenvalues outside \(I_ \cup I_+\). Therefore, we have that
The first term of (4.27) may be treated in the same way as in the proof of Theorem 2.3. Moreover, derivatives of this term may be treated in the same way as in (4.25). It is therefore sufficient to bound the remaining term and its derivative.
Firstly, we note that the eigenvectors corresponding to isolated eigenvalues in the spectral gap have the following decay [17]: for \({\mathscr {C}}^\prime \) a simple closed positively oriented contour (or system of contours) encircling the \(\{\lambda _s\}\), we have that
where \(\gamma _\mathrm {CT}\) is the CombesThomas constant from Lemma 1 with \(\mathfrak {d} = \mathrm {dist}\big ({\mathscr {C}}^\prime , \sigma (\mathcal {H}({\varvec{u}}))\big )\). The constant prefactor in (4.28) depends on the distance between the contour and the defect spectrum \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big )\). Similar estimates hold for the derivatives. For full details on the derivation of (4.28), see [17, (5.18)–(5.21)].
Therefore, combining (4.28) and the Hermite integral formula, we conclude as in the proof of Theorem 2.3. \(\square \)
4.3 Nonlinear bodyorder approximation
In this section, we prove Theorem 2.5 by applying the recursion method to reformulate the problem into a semiinfinite linear chain and replacing the farfield with vacuum.
4.3.1 Recursion method
In that follows, we briefly introduce the recursion method [49, 50], a reformulation of the Lanczos process [61], which generates a tridiagonal (Jacobi) operator T [91] whose spectral measure is \(D_\ell \) and the corresponding sequence of orthogonal polynomials [40]. This process provides the basis for constructing approximations to the LDOS giving rise to nonlinear approximation schemes satisfying Theorem 2.5.
Recall that \(D_\ell \) is the LDOS satisfying (2.13). We start by defining \(p_0 :=1\), \(a_0 :=\int x \mathrm {d}D_\ell (x)\) and \(b_1 p_1(x) :=x  a_0\) where \(b_1\) is the normalising constant to ensure \(\int p_1(x)^2 \mathrm {d}D_\ell (x) = 1\). Then, supposing we have defined \(a_0, a_1, b_1, \dots , a_n, b_{n}\) and the polynomials \(p_0(x), \dots , p_n(x)\), we set
Then, \(\{p_n\}\) is a sequence of orthogonal polynomials with respect to \(D_\ell \) (i.e. \(\int p_n p_m \mathrm {d}D_\ell = \delta _{nm}\)) and we have that
(see Lemma D.1 for a proof). Moreover, we denote by T the infinite symmetric tridiagonal matrix on \({\mathbb {N}}_0\) with diagonal \((a_n)_{n \in {\mathbb {N}}_0}\) and offdiagonal \((b_n)_{n\in {\mathbb {N}}}\).
Remark 15
It will also prove convenient for us to renormalise the orthogonal polynomials by defining \(P_n(x) :=b_n p_n(x)\) and \(b_0 :=1\); that is,
One advantage of this formulation is that it explicitly defines the coefficients \(\{b_n\}\).
Therefore, if we have the first \(2N+1\) moments \(\mathcal {H}_{\ell \ell }, \dots , (\mathcal {H}^{2N+1})_{\ell \ell }\), it is possible to evaluate \(Q_{2N+1}(\mathcal {H})_{\ell \ell }\) (that is, \(\int Q_{2N+1} \mathrm {d}D_\ell \)) for all polynomials \(Q_{2N+1}\) of degree at most \(2N+1\), and thus compute \(T_N\). In particular, for a fixed observable of interest O, we may write
Remark 16
In Appendix E we introduce more complex bond order potential (BOP) schemes based on the recursion method and show that they also satisfy Theorem 2.5.
4.3.2 Error estimates
Equation (4.35) states that the nonlinear approximation scheme given by \(\Theta _{2N+1}\) simply approximates the LDOS with the spectral measure of \(T_N\) corresponding to \(\varvec{e}_0 :=(1,0,\dots ,0)^\mathrm {T}\). We now show that \([(T_N)^n]_{00} = [T^n]_{00} = [\mathcal {H}^n]_{\ell \ell }\) for all \(n \leqslant 2N+1\) and thus we may apply (2.14) to conclude.
By the orthogonality, we have \([T^0]_{ij} = \int p_i(x) x^0 p_j(x) \mathrm {d}D_\ell (x) = \delta _{ij}\). Therefore, assuming \([T^n]_{ij} = \int p_i(x) x^n p_j(x) \mathrm {d}D_\ell (x)\), we can conclude that
Here, we have applied (4.29) directly. In particular, if \(i=j=0\), we obtain \([T^n]_{00} = [\mathcal {H}^n]_{\ell \ell }\).
4.3.3 Analyticity
To conclude the proof of Theorem 2.5, we show that \(\Theta _{2N+1}\) as in (4.35) extents to an analytic function on some open set \(U \subset {\mathbb {C}}^{2N+1}\). Throughout this section, we use the rescaled orthogonal polynomials \(\{P_n\}\) from Remark 15.
For a polynomial \(P(x) = \sum _{j=0}^m c_j x^j\), we use the notation \(\mathcal {L} P(z_1,\dots ,z_m) :=c_0 + \sum _{j=1}^m c_j z_j\) for the linear function satisfying \(P(x) = \mathcal {L}P(x,x^2,\dots ,x^m)\). To extend the recurrence coefficients from (4.32), we start by defining
To simplify the notation, we write \(\varvec{z}_{1:m}\) for the mtuple \((z_1,\dots ,z_m)\). Given \(a_0(z_1), \dots , a_n( \varvec{z}_{1:2n+1})\) and \(b_1(\varvec{z}_{1:2}), \dots , b_n(\varvec{z}_{1:2n})\), we define \(P_{n+1}(x;\varvec{z}_{1:2n+1})\) to be the polynomial in x satisfying the same recursion as (4.32) but as a function of \(\varvec{z}_{1:2n+1}\):
With this notation, we define
Since \(P_{n+1}(x) = P_{n+1}(x;\mathcal {H}_{\ell \ell },\dots ,[\mathcal {H}^{2n+1}]_{\ell \ell })\), we have extended the definition of the recursion coefficients (4.34) to functions of multiple complex variables.
We now show that \(a_n(\varvec{z}_{1:2n+1})\) and \(b_n^2(\varvec{z}_{1:2n})\) are rational functions. As a preliminary step, we show that both \(P_{n+1}^2\) and \(\frac{P_{n+1} P_n}{b_n}\) are polynomials in x with coefficients given by rational functions of \(a_n,b_n^2\) and all previous recursion coefficients. This statement is clearly true for \(n = 0\): \(P_1^2 = (x  a_0)^2\) and \(\frac{P_1 P_0}{b_0} = x  a_0\). Therefore, by induction and noting that
we can conclude. Therefore, by (4.38) and (4.39) and (4.40), we can apply another induction argument to conclude that \(a_{n+1}(\varvec{z}_{1:2n+3})\) and \(b_{n+1}^2(\varvec{z}_{1:2(n+1)})\) are rational functions.
We fix N and define the following complex valued tridiagonal matrix
If \(z_j = [\mathcal {H}^j]_{\ell \ell }\) for each \(j = 1,\dots ,N\), (4.43) is similar to \(T_N\) from (4.31).
Now, on defining \(U :=\{ \varvec{z} \in {\mathbb {C}}^{2N+1} :b_n^2(\varvec{z}_{1:2n}) \not = 0 \,\, \forall n=1,\dots ,N\}\), the mapping \(U \rightarrow {\mathbb {C}}^{(N+1)\times (N+1)}\) given by \(\varvec{z} \mapsto T_N(\varvec{z})\) is analytic. Therefore, for appropriately chosen contours \({\mathscr {C}}_{\varvec{z}}\) encircling \(\sigma \big (T_N(\varvec{z}) \big )\), we have that
In particular, \(\Theta _{2N+1}\) is an analytic function on
Remark 17
Since \({\mathbb {C}}^{2N+1} \setminus U\) is the zero set for some (nonzero) polynomial P in \(2N+1\) variables, it has \((2N+1)\)dimensional Lebesgue measure zero [45].
Remark 18
In Appendix D we show that the eigenvalues of \(T_N(\varvec{z})\) are distinct for \(\varvec{z}\) in some open neighbourhood, \(U_0 \subset U\), of \({\mathbb {R}}^{2N+1}\), which leads to the following alternative proof. On \(U_0\), the eigenvalues and corresponding left and right eigenvectors can be chosen to be analytic: there exist analytic functions \(\varepsilon _j, \psi _{j}, \phi _{j}^\star \) for \(j = 0,\dots ,N\) such that
(More precisely, we apply [44, Theorem 2] to obtain analytic functions \(\psi _j,\phi ^\star _j\) of each variable \(z_0,\dots ,z_{2N+1}\) separately and then apply Hartog’s theorem [60] to conclude that \(\psi _j,\phi ^\star _j\) are analytic as functions on \(U \subset {\mathbb {C}}^{2N+1}\).) Therefore, the nonlinear method discussed in this section can also be written in the form
which is an analytic function on \(\{ \varvec{z} \in U_0 :O \text { analytic at } \varepsilon _j(\varvec{z}) \text { for each } j\}\) (as it is a finite combination of analytic functions only involving products, compositions and sums).
4.4 Selfconsistent tight binding models
We start with the following preliminary lemma:
Lemma 3
Suppose that \(T :\ell ^2(\Lambda ) \rightarrow \ell ^2(\Lambda )\) is an invertible bounded linear operator with matrix entries \(T_{\ell k}\) satisfying \( \big  T_{\ell k} \big  \leqslant c_T e^{\gamma _T r_{\ell k}} \) for some \(c_T, \gamma _T > 0\).
Then, there exists an invertible bounded linear operator \(\overline{T} :\ell ^\infty (\Lambda ) \rightarrow \ell ^\infty (\Lambda )\) extending \(T:\ell ^2(\Lambda ) \rightarrow \ell ^2(\Lambda )\) (that is, \(\overline{T}\big _{\ell ^2(\Lambda )} = T\)).
Proof
First, we denote the inverse of T and its matrix entries by \(T^{1}:\ell ^2(\Lambda ) \rightarrow \ell ^2(\Lambda )\) and \(T^{1}_{\ell k}\), respectively. Then, applying the CombesThomas estimate to T yields the offdiagonal decay estimate \(T^{1}_{\ell k} \leqslant C e^{\gamma _{\mathrm {CT}} r_{\ell k}}\) for some \(C, \gamma _\mathrm {CT} > 0\) [93].
Due to the offdiagonal decay properties of the matrix entries, the operators \(\overline{T}, \overline{T}^{1} :\ell ^\infty (\Lambda ) \rightarrow \ell ^\infty (\Lambda )\) given by
are well defined bounded linear operators with norms \(\sup _\ell \sum _{k \in \Lambda } T_{\ell k}\) and \(\sup _\ell \sum _{k \in \Lambda } T^{1}_{\ell k}\), respectively. To conclude, we note that
and so \(\overline{T}^{1}\) is the inverse of \(\overline{T}\). Here, we have exchanged the summations over k and m by applying the dominated convergence theorem: \(\big \sum _k T_{\ell k}T^{1}_{km} \phi _m\big  \leqslant C e^{\frac{1}{2}\min \{\gamma _T, \gamma _\mathrm {CT}\} r_{\ell m}} \Vert \phi \Vert _{\ell ^\infty }\) is summable over \(m \in \Lambda \). \(\square \)
Throughout the following proofs, we denote by \(B_r(\rho )\) the open ball of radius r about \(\rho \) with respect to the \(\ell ^\infty \)norm. Moreover, we briefly note that the stability operator can be written as the product \({\mathscr {L}}(\rho ) :={\mathscr {F}}(\rho ) \nabla {w(\rho )}\), where [93]
where \({\mathscr {C}}\) is a simple closed contour encircling the spectrum \(\sigma \big ( \mathcal {H}({\varvec{u}}(\rho )) \big )\).
Proof of Theorem 2.6
Since \(\rho \mapsto F^\beta ({\varvec{u}}(\rho ))\) is \(C^2\), and \(\big (I  {\mathscr {L}}(\rho ^\star )\big )^{1}\) is a bounded linear operator, we necessarily have that \(\big (I  {\mathscr {L}}(\rho )\big )^{1}\) is a bounded linear operator for all \(\rho \in B_r(\rho ^\star )\) for some \(r > 0\).
By applying Theorem 2.3, together with the assumption (EP), we obtain
for all \(\rho \in B_r(\rho ^\star )\). As a direct consequence, we have \(\Vert {{\mathscr {L}}}(\rho )  {\mathscr {L}}_N(\rho ) \Vert _{\ell ^2 \rightarrow \ell ^2} \leqslant C e^{\frac{1}{2}\gamma _N N}\) and we may choose N sufficiently large such that \(\Vert {\mathscr {L}}(\rho )  {\mathscr {L}}_N(\rho ) \Vert _{\ell ^2 \rightarrow \ell ^2} < \Vert (I  {\mathscr {L}}(\rho ))^{1} \Vert ^{1}_{\ell ^2 \rightarrow \ell ^2}\). In particular, for such N, the operator \(I  {{\mathscr {L}}}_N(\rho ) :\ell ^2 \rightarrow \ell ^2\) is invertible with inverse bounded above in operator norm independently of N.
We now show that \(I  {\mathscr {L}}_N(\rho )\) satisfies the assumptions of Lemma 3. Using (4.47) and (EP), together with the CombesThomas estimate (Lemma 1), we conclude that
for all \(\rho \in B_r(\rho ^\star )\). In particular, \(I  {\mathscr {L}}_N(\rho )\) extends to a invertible bounded linear operator \(\ell ^\infty \rightarrow \ell ^\infty \) and thus its inverse \(\big (I  {\mathscr {L}}_N(\rho )\big )^{1} :\ell ^\infty \rightarrow \ell ^\infty \) is bounded.
Now, the mapping \(\rho \mapsto \rho  I_{X_N}F^\beta \big ({\varvec{u}}(\rho )\big )\) between \(\ell ^\infty \rightarrow \ell ^\infty \) is continuously differentiable on \(B_r(\rho ^\star )\) and the derivative at \(\rho ^\star \) is invertible (i.e. \(\big (I  {\mathscr {L}}_N(\rho ^\star )\big )^{1}:\ell ^\infty \rightarrow \ell ^\infty \) is a well defined bounded linear operator). Since the map \(\rho \mapsto I_{X_N}{F}^\beta \big ({\varvec{u}}(\rho )\big )\) is \(C^2\), its derivative \({{\mathscr {L}}}_N\) is locally Lipschitz about \(\rho ^\star \) and so there exists \(L>0\) such that
Moreover, by Theorem 2.3, we have that
where \(b_{N} \lesssim e^{\gamma _N N}\). In particular, we may choose N sufficiently large such that \(2b_{N} L < 1\) and \(t^\star _{N} :=\frac{1}{L}( 1  \sqrt{1  2 b_{N} L} ) < r\).
Thus, the Newton iteration with initial point \(\rho ^0 :=\rho ^\star \), defined by
converges to a unique fixed point \(\rho _N = I_{X_N}{F}^\beta ({\varvec{u}}(\rho _N))\) in \(B_{t^\star _{N}}(\rho ^\star )\) [102, 104]. That is, \(\Vert \rho _N  \rho ^\star \Vert _{\ell ^\infty } \leqslant t_N^\star \leqslant 2 b_{N}\). Here, we have used the fact that \(1  \sqrt{1  x} \leqslant x\) for all \(0 \leqslant x \leqslant 1\).
Since \(\rho _N \in B_r(\rho ^\star )\), we have \(I  {\mathscr {L}}_N(\rho _N) :\ell ^2 \rightarrow \ell ^2\) is invertible and thus Lemma 2.8 also holds. \(\square \)
Proof of Proposition 2.9
We proceed in the same way as in the proof of Theorem 2.6. In particular, since \(\rho _N\) is stable, if \(\Vert \rho ^0  \rho _N\Vert _{\ell ^\infty }\) is sufficiently small, \((I  {\mathscr {L}}_N(\rho ^0))^{1}\) is a bounded linear operator on \(\ell ^2\). Moreover, by the exact same argument as in the proof of Theorem 2.6, \(I  {\mathscr {L}}_N(\rho ^0):\ell ^\infty \rightarrow \ell ^\infty \) defines an invertible bounded linear operator. Also, \(I  {\mathscr {L}}_N(\rho )\) is Lipschitz in a neighbourhood about \(\rho ^0\) and
Here, we have used that
Therefore, as long as \(\Vert \rho ^0  \rho _N\Vert _{\ell ^\infty }\) is sufficiently small, we may apply the Newton iteration starting from \(\rho ^0\) to conclude. \(\square \)
Proof of Corollary 2.7
As a direct consequence of (4.51), we have that
Here, we have applied the standard convergence result (Theorem 2.3) with fixed effective potential. \(\square \)
References
Aupetit, B.: A Primer on Spectral Theory. Springer, Berlin (1991)
Bachmayr, M., Csanyi, G., Drautz, R., Dusson, G., Etter, S., van der Oord, C., Ortner, C.: Atomic cluster expansion: Completeness, efficiency and stability, ArXiv eprints arXiv:1911.03550 (2019).
Bak, J., Newman, D.J.: Complex Analysis. Springer, Berlin (2010)
Bartók, A.P., Kermode, J., Bernstein, N., Csányi, G.: Machine learning a generalpurpose interatomic potential for silicon. Phys. Rev. X 8, 041048, 2018
Bartók, A.P., Payne, M.C., Kondor, R., Csányi, G.: Gaussian approximation potentials: the accuracy of quantum mechanics, without the electrons. Phys. Rev. Lett. 104, 136403, 2010
Baskes, M.I.: Modified embeddedatom potentials for cubic materials and impurities. Phys. Rev. B: Condens. Matter 46, 2727–2742, 1992
Bazant, M.Z., Kaxiras, E., Justo, J.F.: Environmentdependent interatomic potential for bulk silicon. Phys. Rev. B Condens. Matter 56, 8542–8552, 1997
Behler, J., Parrinello, M.: Generalized neuralnetwork representation of highdimensional potentialenergy surfaces. Phys. Rev. Lett. 98, 146401, 2007
Benzi, M., Boito, P., Razouk, N.: Decay properties of spectral projectors with applications to electronic structure. SIAM Rev. 55, 3–64, 2013
Biswas, R., Hamann, D.R.: New classical models for silicon structural energies. Phys. Rev. B 36, 6434–6445, 1987
Braams, B.J., Bowman, J.M.: Permutationally invariant potential energy surfaces in high dimensionality. Int. Rev. Phys. Chem. 28, 577–606, 2009
Cancès, É., Kemlin, G., Levitt, A.: Convergence analysis of direct minimization and selfconsistent iterations. SIAM J. Matrix Anal. Appl. 42, 243–274, 2021
Cancès, E., Ehrlacher, V., Maday, Y.: Periodic schrödinger operators with local defects and spectral pollution. SIAM J. Numer. Anal. 50, 3016–3035, 2012
Chen, H., Lu, J., Ortner, C.: Thermodynamic limit of crystal defects with finite temperature tight binding. Arch. Ration. Mech. Anal. 230, 701–733, 2018
Chen, H., Nazar, F.Q., Ortner, C.: Geometry equilibration of crystalline defects in quantum and atomistic descriptions. Math. Models Methods Appl. Sci. 29, 419–492, 2019
Chen, H., Ortner, C.: QM/MM methods for crystalline defects. Part 1: locality of the tight binding model. Multiscale Model. Simul. 14, 232–264, 2016
Chen, H., Ortner, C., Thomas, J.: Locality of interatomic forces in tight binding models for insulators. ESAIM Math. Model. Numer. Anal. 54, 2295–2318, 2020
Chen, J., Lu, J.: Analysis of the divideandconquer method for electronic structure calculations. Math. Comput. 85, 2919–2938, 2016
Chupin, M., Dupuy, M.S., Legendre, G., Séré, É.: Convergence analysis of adaptive DIIS algorithms with application to electronic ground state calculations, ArXiv eprints arXiv:2002.12850 (2020).
Cohen, R.E., Mehl, M.J., Papaconstantopoulos, D.A.: Tightbinding totalenergy method for transition and noble metals. Phys. Rev. B 50, 14694–14697, 1994
Combes, J., Thomas, L.: Asymptotic behavior of eigenfunctions for multiparticle Schrödinger operators. Commun. Math. Phys. 34, 251–270, 1973
CyrotLackmann, F.: On the electronic structure of liquid transitional metals. Adv. Phys. 16, 393–400, 1967
Daw, M.S., Baskes, M.I.: Embeddedatom method: derivation and application to impurities, surfaces, and other defects in metals. Phys. Rev. B Condens. Matter 29, 6443–6453, 1984
Denisov, S.A., Simon, B.: Zeros of orthogonal polynomials on the real line. J. Approx. Theory 121, 357–364, 2003
Drautz, R.: Atomic cluster expansion for accurate and transferable interatomic potentials. Phys. Rev. B 99, 014104, 2019
Drautz, R.: From electrons to interatomic potentials for materials simulations. In: Pavarini, E., Koch, E. (eds.) Topology, Entanglement, and Strong Correlations. Forschungszentrum Jülich GmbH, Institute for Advanced Simulation, Berlin (2020)
Drautz, R., Fähnle, M., Sanchez, J.M.: General relations between manybody potentials and cluster expansions in multicomponent systems. J. Phys. Condens. Matter 16, 3843–3852, 2004
Driscoll, T.A.: Algorithm 756: a MATLAB toolbox for Schwarz–Christoffel mapping. ACM T. Math. Softw. 22, 168–186, 1996
Driscoll, T.A.: Schwarz–Christoffel toolbox. https://github.com/tobydriscoll/sctoolbox (2007)
Driscoll, T.A., Toh, K.C., Trefethen, L.N.: From potential theory to matrix iterations in six steps. SIAM Rev. 40, 547–578, 1998
Driscoll, T.A., Trefethen, L.N.: Schwarz–Christoffel Mapping. Cambridge University Press, Cambridge (2002)
Ehrlacher, V., Ortner, C., Shapeev, A.V.: Analysis of boundary conditions for crystal defect atomistic simulations. Arch. Ration. Mech. Anal. 222, 1217–1268, 2016
Elstner, M., Seifert, G.: Density functional tight binding. Philos. Trans. R. Soc. A 372, 20120483, 2014
Embree, M., Trefethen, L.N.: Green’s functions for multiply connected domains via conformal mapping. SIAM Rev. 41, 745–761, 1999
Ercolessi, F.: Tightbinding molecular dynamics and tightbinding justification of classical potentials, lecture notes, 2005
Ercolessi, F., Adams, J.B.: Interatomic potentials from firstprinciples calculations: the forcematching method. EPL 26, 583, 1994
Etter, S.: Polynomial and Rational Approximation for Electronic Structure Calculations, Ph.D. thesis. University of Warwick, UK (2019)
Finnis, M.: Interatomic Forces in Condensed Matter. Oxford University Press, Oxford (2003)
Finnis, M.: Bondorder potentials through the ages. Prog. Mater Sci. 52, 133–153, 2007
Freud, G.: Orthogonal Polynomials. Elsevier, Amsterdam (2014)
Glanville, S., Paxton, A.T., Finnis, M.W.: A comparison of methods for calculating tightbinding bond energies. J. Phys. F Metal Phys. 18, 693–718, 1988
Goedecker, S.: Linear scaling electronic structure methods. Rev. Mod. Phys. 71, 1085–1123, 1999
Grafakos, L.: Classical Fourier Analysis. Springer, Berlin (2016)
Greenbaum, A., Li, R.C., Overton, M.L.: Firstorder perturbation theory for eigenvalues and eigenvectors. SIAM Rev. 62, 463–482, 2020
Gunning, R.C., Rossi, H.: Analytic Functions of Several Complex Variables. PrenticeHall, New York (1965)
Haliciogli, T., Pamuk, H.O., Erkoc, S.: Interatomic potentials with multibody interactions. Phys. Status Solidi (b) 149, 81–92, 1988
Hammerschmidt, T., Seiser, B., Ford, M., Ladines, A., Schreiber, S., Wang, N., Jenke, J., Lysogorskiy, Y., Teijeiro, C., Mrovec, M., Cak, M., Margine, E., Pettifor, D., Drautz, R.: BOPfox program for tightbinding and analytic bondorder potential calculations. Comput. Phys. Commun. 235, 221–233, 2019
Hartree, D.R.: The wave mechanics of an atom with a nonCoulomb central field. Part I: theory and methods. Math. Proc. Camb. Philos. Soc. 24, 89–110, 1928
Haydock, R., Heine, V., Kelly, M.J.: Electronic structure based on the local atomic environment for tightbinding bands. J. Phys. C Solid State Phys. 5, 2845–2858, 1972
Haydock, R., Heine, V., Kelly, M.J.: Electronic structure based on the local atomic environment for tightbinding bands, II. J. Phys. C Solid State Phys. 8, 2591–2605, 1975
Haydock, R., Nex, C.M.M.: Comparison of quadrature and termination for estimating the density of states within the recursion method. J. Phys. C Solid State Phys. 17, 4783–4789, 1984
Haydock, R., Nex, C.M.M.: A general terminator for the recursion method. J. Phys. C Solid State Phys. 18, 2235–2248, 1985
Herbst, M.F., Levitt, A.: Blackbox inhomogeneous preconditioning for selfconsistent field iterations in density functional theory. J. Phys. Condens. Matter 33, 085503, 2020
Hohenberg, P., Kohn, W.: Inhomogeneous electron gas. Phys. Rev. 136, B864–B871, 1964
Horsfield, A.P., Bratkovsky, A.M., Fearn, M., Pettifor, D.G., Aoki, M.: Bondorder potentials: theory and implementation. Phys. Rev. B 53, 12694–12712, 1996
Kato, T.: Perturbation Theory for Linear Operators, 2nd edn. Springer, Berlin (1995)
Kittel, C.: Introduction to Solid State Physics, 8th edn. Wiley, New York (2004)
Kohn, W., Sham, L.J.: Selfconsistent equations including exchange and correlation effects. Phys. Rev. 140, A1133–A1138, 1965
Koskinen, P., Mäkinen, V.: Densityfunctional tightbinding for beginners. Comput. Mater. Sci. 47, 237–253, 2009
Krantz, S.: Function Theory of Several Complex Variables. American Mathematical Society, Providence (2001)
Lanczos, C.: An iteration method for the solution of the eigenvalue problem of linear differential and integral operators. Journal of Research of the National Bureau of Standards, 1950
Levin, E., Saff, E.B.: Potential theoretic tools in polynomial and rational approximation. In Harmonic Analysis and Rational Approximation. Lecture Notes in Control and Information Science, pp. 71–94. Springer, Berlin (2006)
Levitt, A.: Screening in the finitetemperature reduced Hartree–Fock model. Arch. Rat. Mech. Anal. 238, 901–927, 2020
Lewin, M., Séré, É.: Spectral pollution and how to avoid it. Proc. Lond. Math. Soc. 100, 864–900, 2009
Luchini, M.U., Nex, C.M.M.: A new procedure for appending terminators in the recursion method. J. Phys. C Solid State Phys. 20, 3125–3130, 1987
Lysogorskiy, Y., van der Oord, C., Bochkarev, A., Menon, S., Rinaldi, M., Hammerschmidt, T., Mrovec, M., Thompson, A., Csanyi, G., Ortner, C., Drautz, R.: Performant implementation of the atomic cluster expansion (PACE): application to copper and silicon, ArXiv eprints arXiv:2103.00814 (to appear in NPJ Computational Materials) (2021)
Mead, L.R., Papanicolaou, N.: Maximum entropy in the problem of moments. J. Math. Phys. 25, 2404–2417, 1984
Mehl, M.J., Papaconstantopoulos, D.A.: Applications of a tightbinding totalenergy method for transition and noble metals: elastic constants, vacancies, and surfaces of monatomic metals. Phys. Rev. B 54, 4519–4530, 1996
Nex, C.M.M.: Estimation of integrals with respect to a density of states. J. Phys. A Math. Gen. 11, 653–663, 1978
Ortner, C., Thomas, J.: Point defects in tight binding models for insulators. Math. Models Methods Appl. Sci. 30, 2753–2797, 2020
Papaconstantopoulos, D.A.: Handbook of the Band Structure of Elemental Solids, From Z = 1 To Z = 112. Springer, New York (2015)
Papaconstantopoulos, D.A., Mehl, M.J., Erwin, S.C., Pederson, M.R.: Tightbinding Hamiltonians for carbon and silicon. Symposium R  Tight Binding Approach to Comput. Mater. Sci. 491, 221, 1997
Parr, R.G., Weitao, Y.: DensityFunctional Theory of Atoms and Molecules. Oxford University Press, Oxford (1994)
Pettifor, D.: New manybody potential for the bond order. Phys. Rev. Lett. 63, 2480–2483, 1989
Ransford, T.: Potential Theory in the Complex Plane. Cambridge University Press, Cambridge (1995)
Saff, E.B.: Logarithmic potential theory with applications to approximation theory. Surveys in Approximation Theory 5, 165–200, 2010
Saff, E.B., Totik, V.: Logarithmic Potentials with External Fields. Springer, Berlin (1997)
Seifert, G., Joswig, J.O.: Densityfunctional tight bindingan approximate densityfunctional theory method. Wiley Interdiscip. Rev. Comput. Mol. Sci. 2, 456–465, 2012
Seiser, B., Pettifor, D.G., Drautz, R.: Analytic bondorder potential expansion of recursionbased methods. Phys. Rev. B 87, 094105, 2013
Shapeev, A.: Moment tensor potentials: a class of systematically improvable interatomic potentials. Multiscale Model. Simul. 14, 1153–1173, 2016
Shen, J., Strang, G., Wathen, A.J.: The potential theory of several intervals and its applications. Appl. Math. Opt. 44, 67–85, 2001
Silver, R., Roder, H.: Densities of states of megadimensional Hamiltonian matrices. Int. J. Mod. Phys. C 05, 735–753, 1994
Silver, R., Roeder, H., Voter, A., Kress, J.: Kernel polynomial approximations for densities of states and spectral functions. J. Comput. Phys. 124, 115–130, 1996
Slater, J.C., Koster, G.F.: Simplified LCAO method for the periodic potential problem. Phys. Rev. 94, 1498–1524, 1954
Stahl, H., Totik, V.: General Orthogonal Polynomials, Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge (1992)
Stillinger, F.H., Weber, T.A.: Computer simulation of local order in condensed phases of silicon. Phys. Rev. B Condens. Matter 31, 5262–5271, 1985
Suryanarayana, P., Bhattacharya, K., Ortiz, M.: Coarsegraining Kohn–Sham density functional theory. J. Mech. Phys. Solids 61, 38–60, 2013
Suryanarayana, P., Pratapa, P.P., Sharma, A., Pask, J.E.: SQDFT: spectral quadrature method for largescale parallel O(N) Kohn–Sham calculations at high temperature. Comput. Phys. Commun. 224, 288–298, 2018
Sutton, A.P.: Electronic Structure of Materials. Oxford University Press, Oxford (1993)
Taylor, R., Totik, V.: Lebesgue constants for Leja points. IMA J. Num. Anal. 30, 462–486, 2008
Teschl, G.: Jacobi operators and completely integrable nonlinear lattices, vol. 72. Mathematical Surveys and Monographs, Providence (2000)
Teschl, G.: Mathematical Methods in Quantum Mechanics. The American Mathematical Society, Providence (2014)
Thomas, J.: Locality of interatomic interactions in selfconsistent tight binding models. J. Nonlinear Sci. 30, 3293–3319, 2020
Thomas, L.H.: The calculation of atomic fields. Math. Proc. Camb. Philos. Soc. 23, 542–548, 1927
Trefethen, L.N.: Approximation Theory and Approximation Practice, Extended. SIAM, Philadelphia (2019)
Tsing, N.K., Fan, M.K., Verriest, E.I.: On analyticity of functions involving eigenvalues. Linear Algebra Appl. 207, 159–180, 1994
Turchi, P., Ducastelle, F., Treglia, G.: Band gaps and asymptotic behaviour of continued fraction coefficients. J. Phys. C Solid State Phys. 15, 2891–2924, 1982
Voter, A.F., Kress, J.D., Silver, R.N.: Linearscaling tight binding from a truncatedmoment approach. Phys. Rev. B 53, 12733–12741, 1996
Weinan, E., Lu, J.: Electronic structure of smoothly deformed crystals: Cauchy–Born rule for the nonlinear tightbinding model. Commun. Pure Appl. Math. 63, 1432–1468, 2010
Weinan, E., Lu, J.: The Kohn–Sham equation for deformed crystals. Mem. Am. Math. Soc. 221, 1, 2012
Woods, N.D., Payne, M.C., Hasnip, P.J.: Computing the selfconsistent field in Kohn–Sham density functional theory. J. Phys. Condens. Matter 31, 453001, 2019
Yamamoto, T.: A convergence theorem for Newton’s method in Banach spaces. Jpn. J. Appl. Math. 3, 37–52, 1986
Yang, W.: Direct calculation of electron density in densityfunctional theory. Phys. Rev. Lett. 66, 1438–1441, 1991
Zhengda, H.: A note on the Kantorovich theorem for Newton iteration. J. Comput. Appl. Math. 47, 211–217, 1993
Zhu, L., Amsler, M., Fuhrer, T., Schaefer, B., Faraji, S., Rostami, S., Ghasemi, S.A., Sadeghi, A., Grauzinyte, M., Wolverton, C., Goedecker, S.: A fingerprint based metric for measuring similarities of crystalline structures. J. Chem. Phys. 144, 034203, 2016
Zuo, Y., Chen, C., Li, X., Deng, Z., Chen, Y., Behler, J., Csányi, G., Shapeev, A.V., Thompson, A.P., Wood, M.A., Ong, S.P.: Performance and cost assessment of machine learning interatomic potentials. J. Phys. Chem. A 124, 731–745, 2020
Acknowledgements
We gratefully acknowledge stimulating discussions with Gábor Csányi, Simon Etter, and Jianfeng Lu.
JT is supported by EPSRC Grant EP/HO23364/1 (as part of the MASDOC DTC) and by EPSRC Grant EP/W522594/1. HC is supported by the Natural Science Foundation of China under grant 11971066. CO is supported by EPSRC Grant EP/R043612/1, Leverhulme Research Project Grant RPG2017191 and by the Natural Sciences and Engineering Research Council of Canada (NSERC) [funding reference number GR019381].
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by C. Le Bris.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A. Notation
Here we summarise the key notation:

\(\Lambda \) : finite or countable index set,

\({\varvec{u}}_\ell = (\varvec{r}_\ell , v_\ell , Z_\ell )\) : state of atom \(\ell \) where \(\varvec{r}_\ell \in {\mathbb {R}}^d\) denotes the atomic position, \(v_\ell \) the effective potential, and \(Z_\ell \) the atomic species,

\({\varvec{u}}= \{{\varvec{u}}_\ell \}_{\ell \in \Lambda }\) : configuration,

\(\varvec{r}_{\ell k} :=\varvec{r}_k  \varvec{r}_\ell \) and \(r_{\ell k} :=\varvec{r}_{\ell k}\) : relative atomic positions,

\(\delta _{ij}\) : Kronecker delta (\(\delta _{ij} = 0\) for \(i\not =j\) and \(\delta _{ii} = 1\)),

\(\mathrm {Id}_n\) : \(n \times n\) identity matrix,

\(\,\cdot \,\) : absolute value on \({\mathbb {R}}^d\) or \({\mathbb {C}}\),

\(\,\cdot \,\) : Frobenius matrix norm on \({\mathbb {R}}^{n\times n}\),

\(\nabla h = (\nabla h_{ab})_{1 \leqslant a,b \leqslant n}\) : gradient of \(h :{\mathbb {R}}^d \rightarrow {\mathbb {R}}^{n\times n}\),

\(M^\mathrm {T}\) : transpose of the matrix M,

\(\mathrm {Tr}\) : trace of an operator,

\(f \sim g\) as \(x \rightarrow x_0 \in {\mathbb {R}}\cup \{\pm \infty \}\) or \({\mathbb {C}} \cup \{\infty \}\) if there exists an open neighbourhood N of \(x_0\) and positive constants \(c_1,c_2 > 0\) such that \(c_1 g(x) \leqslant f(x) \leqslant c_2 g(x)\) for all \(x \in N\),

C : generic positive constant that may change from one line to the next,

\(f \lesssim g\) : \(f \leqslant C g\) for some generic positive constant,

\({\mathbb {N}}_0 = \{ 0, 1, 2, \dots \}\) : Natural numbers including zero,

\(\delta (\,\cdot \,)\) : Dirac delta, distribution satisfying \(\int f(x) \mathrm {d}\delta (x) = f(0)\),

\(\Vert f\Vert _{L^\infty (X)} :=\sup _{x\in X} f(x)\) : supnorm of f on X,

\(\mathrm {dist}(z, A) :=\inf _{a\in A} z  a\) : distance between \(z\in {\mathbb {C}}\) and the set \(A\subset {\mathbb {C}}\),

\(a + b S :=\{a + bs :s \in S\}\),

\([\psi ]_\ell \) : the \(\ell ^\text {th}\) entry of the vector \(\psi \),

\(\Vert \psi \Vert _{\ell ^2}:=\left( \sum _{k} [\psi ]_k^2 \right) ^{1/2}\) : \(\ell ^2\)norm of \(\psi \),

\(\mathrm {tr}\,M :=\sum _\ell M_{\ell \ell }\) : trace of matrix M,

\(\Vert M\Vert _\mathrm {max} :=\max _{\ell , k} M_{\ell k}\) : maxnorm of the matrix M,

\(\sigma (T)\) : the spectrum of the operator T,

\(\sigma _\mathrm {disc}(T) \subset \sigma (T)\) : isolated eigenvalues of finite multiplicity,

\(\sigma _\mathrm {ess}(T) :=\sigma (T) \setminus \sigma _\mathrm {disc}(T)\) : essential spectrum,

\(\Vert T\Vert _{X \rightarrow Y} :=\sup _{x \in X, \Vert x\Vert _X = 1} \Vert Tx\Vert _Y\) : operator norm of \(T :X \rightarrow Y\),

\(\nabla v\) : Jacobian of \(v :{\mathbb {R}}^\Lambda \rightarrow {\mathbb {R}}^\Lambda \),

\([a, b] :=\{ (1t) a + t b :t \in [0,1] \}\) : closed interval between \(a,b \in {\mathbb {R}}^d\) or \(a,b \in {\mathbb {C}}\),

\(\int _a^b :=\int _{[a,b]}\) : integral over the interval [a, b] for \(a,b \in {\mathbb {C}}\),

\(\mathrm {len}({\mathscr {C}})\) : length of the simple closed contour \({\mathscr {C}}\),

\(\mathrm {supp} \, \nu \) : support of the measure \(\nu \), set of all x for which every open neighbourhood of x has nonzero measure,

\(\mathrm {conv}\, A :=\{ t a + (1t) b :a,b \in A, t \in [0,1] \}\) : convex hull of A.
Appendix B. Locality: Truncation of the Atomic Environment
We have seen that analytic quantities of interest may be approximated by bodyorder approximations. However, each polynomial depends on the whole atomic configuration \({\varvec{u}}\). In this section, we consider the truncation of the approximation schemes to a neighbourhood of the central site \(\ell \) and prove the exponential convergence of the corresponding sparse representation.
1.1 B.1. Banded Approximation
One intuitive approach is to restrict the interaction range globally and consider the following banded approximation:
Therefore, approximating \(O_\ell ({\varvec{u}})\) with a function depending on the first N moments \([\widetilde{\mathcal {H}}^{r_\mathrm {c}}]_{\ell \ell }\) (e.g. applying Theorem 2.4 or 2.5 to \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\)) results in an approximation scheme depending only on finitely many atomic sites in a neighbourhood of \(\ell \). This can be seen from the fact that
Moreover, we obtain appropriate error estimates by combining Theorem 2.4 or 2.5 with the following estimate:
Proposition B.1
Suppose \({\varvec{u}}\) satisfies Definition 1. Fix \(0 < \beta \leqslant \infty \) and suppose that, if \(\beta = \infty \), then \(\mathsf {g}, \mathsf {g}^\mathrm {def} >0\). Then, we have
Suppose \(\gamma _N(r_\mathrm {c})\) and \(\gamma _N^\mathrm {def}(r_\mathrm {c})\) are the rates of approximation from Theorems 2.4 and 2.5 when applied to \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\). Then \(\gamma _N(r_\mathrm {c}) \rightarrow \gamma _N\) and \(\gamma _N^\mathrm {def}(r_\mathrm {c}) \rightarrow \gamma _N^\mathrm {def}\) as \(r_\mathrm {c} \rightarrow \infty \), with an exponential rate.
Proof
We first note that
Therefore, applying (TB), we obtain
To conclude we choose a suitable contour \(\mathscr {C}\) and apply the CombesThomas estimate (Lemma 1) together with (B.4):
As a direct consequence of (B.4), we have also have \(\Vert \mathcal {H}({\varvec{u}})  \widetilde{\mathcal {H}}^{r_\mathrm {c}}({\varvec{u}}) \Vert _{\ell ^2 \rightarrow \ell ^2} \lesssim e^{\frac{1}{2}\gamma _0 r_{\mathrm {c}}}\) and so \(\mathrm {dist}\big ( \sigma (\mathcal {H}), \sigma (\widetilde{\mathcal {H}}^{r_\mathrm {c}}) \big ) \lesssim e^{\frac{1}{2}\gamma _0 r_{\mathrm {c}}}\) [56]. This means that for sufficiently large \(r_\mathrm {c}\), we obtain the same rates of approximation when applying Theorems 2.4 and 2.5 to \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\). \(\square \)
1.2 B.2. Truncation
One downside of the banded approximation is that the truncation radius depends on the maximal polynomial degree (e.g. see (B.2)). In this section, we consider truncation schemes that only depend on finitely many atomic sites independent of the polynomial degree:
where the restriction of the Hamiltonian has been introduced in (2.26).
On defining the quantities
where the operators \(I_{X_N}\) are given by Theorem 2.3, we obtain a sparse representation of the Nbody approximation depending only on finitely many atomic sites, independently of the maximal bodyorder N.
Proposition B.2
Suppose \({\varvec{u}}\) satisfies Definition 1. Fix \(0 < \beta \leqslant \infty \) and suppose that, if \(\beta = \infty \), then \(\mathsf {g}, \mathsf {g}^\mathrm {def} >0\). Then,
where \(O^\beta = F^\beta \) or \(G^\beta \) and \(\gamma _\mathrm {CT}\) is the constant from Lemma 1 applied to \(\mathcal {H}({\varvec{u}})\).
Proof
Applying the Hermite integral formula (4.1) directly, we conclude that \(I_{X_N}O^\beta (z)\) is bounded uniformly in N along a suitably chosen contour \({\mathscr {C}} :=\{ g_E = \gamma \}\) (examples of such contours are given in Figure 5). It is important to note that the contour \({\mathscr {C}}\) must be chosen to encircle both \(\sigma (\mathcal {H})\) and \(\sigma ( \widetilde{\mathcal {H}}^{r_\mathrm {c}} )\).
In the following, we let \(\gamma _\mathrm {CT}\) be the CombesThomas exponent from Lemma 1 corresponding to \(\mathcal {H}\).
Similarly to (B.7), we obtain
This concludes the proof. \(\square \)
The fact that the exponents of Proposition B.2 are independent of the defect states within the band gap is in the same spirit to the improved locality estimates of [17].
Remark 19
(Divideandconquer Methods) This truncation scheme is closely related to the divideandconquer method for solving the electronic structure problem [103]. In this context the system is split into many subsystems that are only related through a global choice of Fermi level. In our notation, this method consists of constructing \(N_{\mathrm {DAC}}\) smaller Hamiltonians \(\widetilde{\mathcal {H}}^{r_\mathrm {c},\ell _j}\) centred on the atoms \(\ell _j\) (for \(j = 1,\dots ,N_{\mathrm {DAC}}\)) and approximating the quantities \(O_\ell ({\varvec{u}})\) for \(\ell \) in a small neighbourhood of \(\ell _j\) by calculating \(\mathrm {tr}\, O\big ( \widetilde{\mathcal {H}}^{r_\mathrm {c},\ell _j} \big )_{\ell \ell }\). That is, the eigenvalue problem for the whole system is approximated by solving \(N_\mathrm {DAC}\) smaller eigenvalue problems in parallel. In particular, this method leads to linear scaling algorithms [42]. Theorem B.2 then ensures that the error in this approximation decays exponentially with the distance between \(\ell \) and the exterior of the subsystem centred on \(\ell _j\).
A similar error analysis in the context of divideandconquer methods in KohnSham density functional theory can be found in [18].
Remark 20
(General truncation operators) It should be clear from the proof of Proposition B.2 that more general truncation operators may be used. Indeed, Proposition B.2 is satisfied for all truncation operators \(\widetilde{\mathcal {H}}^{r_\mathrm {c}} = \widetilde{\mathcal {H}}^{r_\mathrm {c}}({\varvec{u}})\) satisfying the following conditions:

(T1)
For every polynomial p, the quantity \(p\big (\widetilde{\mathcal {H}}^{r_\mathrm {c}}\big )_{\ell \ell }\) depends on at most finitely many atomic sites depending on \(r_\mathrm {c}\) but not p,

(T2)
For all \(k,m\in \Lambda \), we have \([\widetilde{\mathcal {H}}^{r_\mathrm {c}}]_{km} \rightarrow \mathcal {H}_{km}\) as \(r_\mathrm {c} \rightarrow \infty \),

(T3)
There exists \(c_0 > 0\) such that for all \(\gamma , r_\mathrm {c} > 0\),
$$\begin{aligned} \sum _{km} e^{\gamma r_{\ell k}} \left \big [ \mathcal {H} \widetilde{\mathcal {H}}^{r_\mathrm {c}} \big ]_{k m} \right \leqslant C e^{ c_0 \min \{\gamma _0, \gamma \} \, r_{\mathrm {c}} } \end{aligned}$$for some \(C>0\) depending on \(\gamma \) but not on \(r_\mathrm {c}\).
Due to the exponential weighting of the summation, (T3) states that \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\) captures the behaviour of the Hamiltonian in a small neighbourhood of the site \(\ell \). Moreover, when making the approximation \(I_{X_N}O\big ( \mathcal {H}\big )_{\ell \ell }\approx I_{X_N}O\big ( \widetilde{\mathcal {H}}^{r_\mathrm {c}} \big )_{\ell \ell }\), the number of atomic sites involved is finite by (T1).
Remark 21
(Nonlinear schemes) One may be tempted to approximate the Hamiltonian with the truncation, \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\), and then apply the nonlinear scheme of Theorem 2.5. In doing so, we obtain the following error estimates:
A problem with this analysis is that the constant \(\widetilde{\gamma }_N(r_\mathrm {c})\) in (B.13) arises by applying Theorem 2.5 to \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\) rather than the original system \(\mathcal {H}\). In particular, this means that \(\widetilde{\gamma }_N(r_\mathrm {c})\) depends on the spectral properties of \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\) rather than \(\mathcal {H}\). Since spectral pollution is known to occur when applying naive truncation schemes [64], the choice of \(\widetilde{\mathcal {H}}^{r_\mathrm {c}}\) is important for the analysis. In particular, it is not clear that \(\widetilde{\gamma }_N(r_\mathrm {c}) \rightarrow \gamma _N\) in general. This is in contrast the the result of Proposition B.1.
Appendix C. Convergence of Derivatives in the Nonlinear Approximation Scheme
As mentioned in Remark 10, the results of this section depend on the “regularity” properties of \(D_\ell \):
Definition 2
(Regular \(n^\text {th}\)root Asymptotic Behaviour) For a unit measure \(\nu \) with compact support \(E :=\mathrm {supp}\,\nu \subset {\mathbb {R}}\), we say \(\nu \) is regular and write \(\nu \in \mathbf {Reg}\) if the corresponding sequence of orthonormal polynomials \(\{p_n(\,\cdot \,;\nu )\}\) satisfy
locally uniformly on \({\mathbb {C}} \setminus \mathrm {conv}(E)\).
Remark 22
The regularity condition says that the \(n^\text {th}\)root asymptotic behaviour of \(p_n(z;\nu )\) is minimal: in general, we have [85, Theorem 1.1.4]
where \(g_\nu \geqslant g_E\) is the minimal carrier Green’s function of \(\nu \) [85].
Under the regularity condition of Definition 2, we obtain results analogous to (2.21):
Theorem C.1
Suppose that \(\varvec{u}\) satisfies Definition 1 and \(\ell \in \Lambda \) is such that \(D_\ell \in \mathbf {Reg}\). Then, with the notation of Theorem 2.5, we in addition have
More generally, if the regularity assumption is not satisfied, it may still be the case that Theorem C.1 holds but with reduced locality exponent \(\eta \). To formulate this result, we require the notion of minimal carrier capacity:
Definition 3
(Minimal carrier capacity) For arbitrary Borel sets C, the capacity of C is defined as
where \(\mathrm {cap}(K)\) is defined as in § 4.1.5.
For a unit measure \(\nu \) with compact support \(E :=\mathrm {supp}\,\nu \subset {\mathbb {R}}\), the set of carriers of \(\nu \) and the minimal carrier capacity are defined as
respectively.
Under these definitions, we have the following [85, p. 8–10]:
Remark 23
For a unit measure \(\nu \) with compact support \(E :=\mathrm {supp}\,\nu \subset {\mathbb {R}}\), we have
(i) The set of minimal carriers \(\Gamma _0(\nu ) :=\{ C \in \Gamma (\nu ) :\mathrm {cap}(C) = c_\nu , C \subset E\}\) is nonempty,
(ii) If \(c_\nu >0\), then there exists a minimial carrier equilibrium distribution \(\omega _\nu \), a (uniquely defined) unit measure with \(\mathrm {supp} \,\omega _\nu \subset E\) satisfying
(iii) \(g_\nu \equiv g_E\) if and only if \(c_\nu = \mathrm {cap}(E)\),
(iv) In particular, if \(c_\nu = \mathrm {cap}(E)\), then \(\nu \in \mathbf {Reg}\) (although the converse is false [85, Example 1.5.4]),
(v) Suppose \(c_\nu > 0\). Then, on defining \(\nu _n\) to be the discrete unit measure giving equal weight to each of the zeros of \(p_n(\,\cdot \,;\nu )\), the condition that
where \(\omega _E\) is the equilibrium distribution for E, is equivalent to \(\nu \in \mathbf {Reg}\) [85, Thm. 3.1.4]. In particular, this justifies (4.10).
We therefore arrive at the corresponding result for \(\ell \in \Lambda \) for which the corresponding LDOS has positive minimal carrier capacity:
Proposition C.2
Suppose that \(\varvec{u}\) satisfies Definition 1 and \(\ell \in \Lambda \) such that \(c_{D_\ell } > 0\). Then, with the notation of Theorem 2.5, we in addition have
where \(\eta _\ell > 0\),
and \(\eta > 0\) is the constant from Theorem C.1.
The proofs of Theorem C.1 and Proposition C.2 follow from the following estimates on the derivatives of the recursion coefficients \(\{a_n, b_n\}\), and the locality of the tridiagonal operators \(T_N\), together with the asymptotic upper bounds (i.e. Definition 2 or Remark 22).
Lemma C.3
Suppose \(\varvec{u}\) satisfies Definition 1. Then, for a simple closed positively oriented contour \({\mathscr {C}}^\prime \) encircling the spectrum \(\sigma \big ( \mathcal {H}({\varvec{u}}) \big )\), there exists \(\eta = \eta ({\mathscr {C}}^\prime ) > 0\) such that
where \(\eta \sim \mathfrak {d}\) as \(\mathfrak {d} \rightarrow 0\) where \(\mathfrak {d} :=\mathrm {dist}\big ( {\mathscr {C}}^\prime , \sigma \big ( \mathcal {H}({\varvec{u}}^\mathrm {ref}) \big ) \big )\).
In the following, we denote by \(T_\infty \) the infinite symmetric matrix on \({\mathbb {N}}_0\) with diagonal \((a_n)_{n \in {\mathbb {N}}_0}\) and offdiagonal \((b_n)_{n\in {\mathbb {N}}}\).
Lemma C.4
Fix \(N \in {\mathbb {N}} \cup \{\infty \}\). Suppose that \(z \in {\mathbb {C}}\) with \(\mathfrak {d}_N :=\mathrm {dist}\big ( z, \sigma (T_N) \big ) > 0\). Then, for each \(i, j \in {\mathbb {N}}_0\), we have
(i) For each \(r \in {\mathbb {N}}\), we have \(\gamma _{r,N} \sim \mathfrak {d}_N\) as \(\mathfrak {d}_N \rightarrow 0\).
(ii) We have \(\lim _{r\rightarrow \infty } \gamma _{r,\infty } = \lim _{N\rightarrow \infty } \gamma _{N,N} = g_{\sigma (T_\infty )}(z)\) where \(g_{\sigma (T_\infty )}\) is the Green’s function for the set \(\sigma (T_\infty )\) as defined in (4.10).
Remark 24
The fact that \(g_{\sigma (T_\infty )}\) does not depend on the discrete eigenvalues of \(T_\infty \) means that asymptotically the locality estimates do not depend on defect states in the band gap arising due to perturbations satisfying Proposition 2.1, for example. Indeed, this has been shown more generally for operators with offdiagonal decay [17]. We show an alternative proof using logarithmic potential theory.
We will assume Lemmas C.4 and C.3 for now and return to their proofs below.
We first add on a constant multiple of the identity, cI, to the operators \(\{T_N\}\) so that the spectra are contained in an interval bounded away from \(\{0\}\). Moreover, we translate the integrand by the same constant: \(\widetilde{O}(z) :=O(z  c)\). Then, we extend \(T_N\) to an operator on \(\ell ^2(\mathbb {N}_0)\) by defining \([T_N \psi ]_i = \sum _{j=0}^N [T_{N}]_{ij} \psi _j\) for \(0\leqslant i \leqslant N\) and \([T_N \psi ]_i =0\) otherwise. We therefore choose a simple closed contour (or system of contours) \({\mathscr {C}}\) encircling \(\bigcup _{N} \sigma (T_N)\) so that
Therefore, applying Lemma C.3, a simple calculation reveals that
where \(\gamma _{r,N} = \gamma _{r,N}({\mathscr {C}})\) is the constant from Lemma C.4. We therefore may conclude by choosing \({\mathscr {C}}^\prime :=\{ g_E = \gamma \}\) if \(D_\ell \in \mathbf {Reg}\) and \({\mathscr {C}}^\prime :=\{ g_{D_\ell } = \gamma \}\) otherwise for some constant \(\gamma > 0\) sufficiently small such that the summation in the square brackets converges.
Proof of Lemma C.3
The proof follows from the following identities:
To do this, it will be convenient to renormalise the orthogonal polynomials as in Remark 15 (that is, we consider \(P_n(x) :=b_n p_n(x)\)). Moreover, we define \(b_{1} :=1\). Using the shorthand \(\partial :=\frac{\partial }{\partial {\varvec{u}}_m}\), we therefore obtain: \(\partial b_{1} = \partial b_0 = 0\), \(\partial P_{1}(x) = \partial P_0(x) = 0\), and
for all \(n \geqslant 0\).
By noting \(\partial P_1(x) =  \partial a_0\) and applying (C.8), we can see that \(\partial P_{n}\) is a polynomial of degree \(n1\) for all \(n \geqslant 0\). Therefore, since \(P_{n}\) is orthogonal to all polynomials of degree \(n1\), we have
which concludes the proof of (C.7).
To prove a similar formula for the derivatives of \(a_{n}\), we first state a useful identity which will be proved after the conclusion of the proof of (C.7):
Therefore, we have that
Applying (C.11) for \(k \leqslant n1\), we can see that \(\partial a_n\) can be written as
for some coefficients \(d_{1,k}, d_{0,k}\). Using (C.11) and assuming the result for \(k \leqslant n1\), we have
for all \(k \leqslant n1\). \(\square \)
Proof of (C.11)
We have that
where \(\mathrm {l.o.t.}\) (“lower order term”) denotes a polynomial of degree strictly less than n that changes from one line to the next. That is, since \(c_{11} = \partial a_0 = \partial \big ( \frac{a_0}{b_0}\big ) b_0\), we apply an inductive argument to conclude that
\(\square \)
Proof of Lemma C.4
The first statement is the CombesThomas resolvent estimate (Lemma 1) for tridiagonal operators (which, in particular, satisfy the offdiagonal decay assumptions of Lemma 1).
To obtain the asymptotic estimates of (ii), we apply a different approach based on the banded structure of the operators. Since \(T_N\) is tridiagonal, \([(T_N)^n]_{ij} = 0\) if \(i  j > n\). Therefore, for any polynomial P of degree at most \(ij1\), we have [9]
We may apply the results of logarithmic potential theory (see (4.15)), to conclude. Here, it is important that \(\left \sigma (T_\infty ) \setminus \sigma (T_N)\right \) remains bounded independently of N so that, asymptotically, (C.19) has exponential decay with exponent \(g_{\sigma (T_\infty )}\).
The proof that \(\left \sigma (T_\infty ) \setminus \sigma (T_N)\right \) is uniformly bounded can easily be shown when considering the sequence of orthogonal polynomials generated by \(T_\infty \). A full proof is given in parts (ii) and (iv) of Lemma D.1. \(\square \)
Appendix D. Quadrature Method
The quadrature method as outlined in this section was introduced in [69] to approximate the LDOS. For a comparison of various nonlinear approximation schemes, see [51] and [41]. The former is a practical comparison of quadrature and BOP methods, while the later also discusses the maximum entropy method [67].
We now give an alternative proof of Theorem 2.5 by introducing the quadrature method [69].
Recall that \(D_\ell \) is the local density of states (LDOS) satisfying (2.13) and \(\{p_n\}\) is the corresponding sequence of orthogonal polynomials generated via the recursion method:
(see the proof of Lemma D.1, below).
We use the set of zeros of \(p_{N+1}\), denoted by \(X_N = \{ \varepsilon _0,\dots ,\varepsilon _N \}\), as the basis for the following quadrature rule:
Here, \(\ell _j\) is the polynomial of degree N with \(\ell _j(\varepsilon _i) = \delta _{ij}\).
The following lemma highlights the fundamental properties of Gauss quadrature and allows us to show that the approximation scheme given by
satisfies Theorem 2.5.
Lemma D.1
Suppose that \(\{p_n\}\) is the sequence of polynomials generated by the recursion method (4.29), \(X_N\) is the set of zeros of \(p_{N+1}\), and \(\{w_j\}\) are the weights satisfying \(\int I_{X_N} O(x) \mathrm {d}D_\ell (x) = \sum _{j=0}^N w_j O(\varepsilon _j)\). Then,

(i)
\(\{p_n\}\) is orthonormal with respect to \(D_\ell \): \(\int p_n(x) p_m(x) \mathrm {d}D_\ell (x) = [p_n(\mathcal {H})p_m(\mathcal {H})]_{\ell \ell } = \delta _{nm}\),

(ii)
\(X_N = \sigma (T_N)\) where \(T_N\) is given by (4.31),

(iii)
\(X_N\subset {\mathbb {R}}\) is a set of \(N+1\) distinct points,

(iv)
If \([a,b] \cap \mathrm {supp}\, D_\ell = \emptyset \), then the number of points in \(X_N \cap [a,b]\) is at most one,

(v)
If \(P_{2N+1}\) is a polynomial of degree at most \(2N+1\), then \( P_{2N+1}(\mathcal {H})_{\ell \ell } = \sum _{j=0}^N w_j P_{2N+1}(\varepsilon _j), \)

(vi)
The weights \(\{w_j\}\) are positive and sum to one.
Proof
The idea behind the proofs are standard in the theory of Gauss quadrature (e.g. see [40]) but, for the convenience of the reader, they are collected together in D.3. \(\square \)
Remark 25
The quadrature rule discussed in this section can be seen as the exact integral with respect to the following approximate LDOS
This measure has unit mass by Lemma D.1 (vi), and, by Lemma D.1 (v), the first \(2N+1\) moments of \(D_\ell ^{2N+1,\mathrm {q}}\) are given by \([\mathcal {H}^n]_{\ell \ell }\) for \(n = 1,\dots , 2N+1\).
In the following two sections we prove error estimates and show that the functional form is analytic on an open set containing \(\big (\mathcal {H}_{\ell \ell }, \dots , [\mathcal {H}^{2N+1}]_{\ell \ell }\big )\).
1.1 D.1. Error Estimates.
Applying Remark 25, together with (2.14), we have: for every polynomial \(P_{2N+1}\) of degree at most \(2N+1\),
Now, since \(\sigma \big (\mathcal {H}\big ) \subset I_ \cup \{\lambda _j\} \cup I_+\) where \(\{\lambda _j\}\) is a finite set, we may apply part (iv) of Lemma D.1 to conclude that the number of points in \(X_N \setminus \big ( I_ \cup I_+ \big )\) is bounded independently of N. Accordingly, we may apply (4.15) with \(E = I_ \cup I_+\), to obtain the following asymptotic bound
In particular, we obtain the stated asymptotic behaviour.
Remark 26
(Spectral pollution) While \(\sigma (\mathcal {H}) \subset \liminf _{N\rightarrow \infty } \sigma (T_N)\), we do not claim that the sequence \(\sigma (T_N)\) is free from spurious eigenvalues. That is, there may exist sequences \(\lambda _N \in \sigma (T_N)\) such that \(\lambda _N \rightarrow \lambda \) along a subsequence and \(\lambda \not \in \sigma (\mathcal {H})\). Indeed, there exist measures supported on a union of disjoint intervals \([a,b]\cup [c,d]\) for which the corresponding sequences of orthogonal polynomials suffer from spurious eigenvalues at every point of the gap (b, c) [24, 85]. In this paper, we only require the much milder property that the number of eigenvalues in the gap remains uniformly bounded in the limit \(N\rightarrow \infty \).
For a more general discussion of spectral pollution, see [13, 64].
1.2 D.2. Analyticity.
To conclude the proof of Theorem 2.5, we show that \(\Theta ^\mathrm {q}_{2N+1}\) as defined in (D.2) is analytic in a neighbourhood of \((\mathcal {H}_{\ell \ell }, [\mathcal {H}^2]_{\ell \ell }, \dots , [\mathcal {H}^{2N+1}]_{\ell \ell })\). Recall that in (4.43) we have extended the definition of \(T_N\) to an analytic function on \(U :=\{ \varvec{z} \in {\mathbb {C}}^{2N+1} :b_{n}^2(\varvec{z}_{1:2n}) \not = 0 \,\, \forall n = 1,\dots ,N \}\).
We define \(X_N(\varvec{z})\) to be the set of eigenvalues of \(T_N(\varvec{z})\). Since \(X_N = X_N\big (\mathcal {H}_{\ell \ell }, \dots , [\mathcal {H}^{2N+1}]_{\ell \ell }\big )\) is a set of \(N+1\) distinct points (Lemma D.1 (iii)), there exists a continuous choice of eigenvalues \(X_N(\varvec{z}) = \{ \varepsilon _0(\varvec{z}), \dots , \varepsilon _N(\varvec{z})\}\) such that \(X_N(\varvec{z})\) is a set of \(N+1\) distinct points in a neighbourhood, \(U_0\), of \((\mathcal {H}_{\ell \ell },\dots ,[\mathcal {H}^{2N+1}]_{\ell \ell })\in U\) and each \(\varepsilon _j\) is analytic on \(U_0\) [56, 96]. With this in hand, we define \(\Theta ^\mathrm {q}_{2N+1} :U_0 \rightarrow {\mathbb {C}}\) by
which is analytic on \(\{ \varvec{z} \in U_0 :O \text { analytic at } \varepsilon _j(\varvec{z}) \,\,\forall j=0,\dots ,N\}\).
1.3 D.3. Proof of Lemma D.1
Proof of (i). First note that \(\int p_0 p_1 \mathrm {d}D_\ell = 0\). We assume that \(p_0, \dots , p_n\) are mutually orthogonal with respect to \(D_\ell \), and note that,
Therefore, we conclude by noting
and applying (D.5). Equation (D.5) also justifies the tridiagonal structure (4.31).
Proof of (ii). We may rewrite the recurrence relation (4.29) as \(x \varvec{p}(x) = T_N \varvec{p}(x) + b_{N+1}p_{N+1}(x) \varvec{e}_N\) where \(\varvec{p}(x) :=\big (1, p_1(x), \dots , p_{N}(x) \big )^T\), \([\varvec{e}_N]_j = \delta _{jN}\), and \(T_N\) is the tridiagonal matrix (4.31). In particular, each \(\varepsilon _j \in X_N\) is an eigenvalue of \(T_N\) (with eigenvector \(\varvec{p}(\varepsilon _j)\)).
Proof of (iii). Since \(T_N\) is symmetric, the spectrum is real. Now, for each \(\varepsilon _j \in X_N = \sigma (T_N)\), the matrix \((T_N  \varepsilon _j)_{\lnot N\lnot 0}\) formed by removing the \(N^\text {th}\) row and \(0^\text {th}\) column is lowertriangular with diagonal \((b_1,\dots ,b_N)\). Since each \(b_i > 0\), \((T_N  \varepsilon _j)_{\lnot N\lnot 0}\) has full rank and thus \(\varepsilon _j\) is a simple eigenvalue of \(T_N\).
Proof of (iv). Suppose that (after possibly relabelling) \(\varepsilon _0, \varepsilon _1 \in X_N \cap [a,b]\). After defining \(R(x) :=\prod _{j = 2}^N (x  \varepsilon _j)\), a polynomial of degree \(N1\), and noting \((x  \varepsilon _0) (x  \varepsilon _1) > 0\) on \(\mathrm {supp}\,D_\ell \), we obtain
contradicting part (i).
Proof of (v). We may write \(P_{2N+1} = p_{N+1} q_N + r_N\) where \(q_N, r_N\) are polynomials of degree at most N and note that \([p_{N+1}(\mathcal {H})q_N(\mathcal {H})]_{\ell \ell } = 0\) by (i) and \(P_{2N+1}(\varepsilon _j) = r_N(\varepsilon _j)\) since X is the set of zeros of \(p_{N+1}\). Therefore,
In (D.9) we have used the fact that polynomial interpolation in \(N+1\) distinct points is exact for polynomials of degree at most N.
Proof of (vi). \(\ell _j(x)^2\) is a polynomial of degree 2N and so, by (v), we have
Moreover, \(\sum _{j=0}^N \ell _j(x)\) is a polynomial of degree N equal to one on \(X_N\) (a set of \(N+1\) distinct points) and so \(\sum _{j=0}^N \ell _j(x) \equiv 1\). Finally, \(\sum _{j=0}^N w_j = \int \big ( \sum _{j=0}^N \ell _j(x) \big ) \mathrm {d}D_\ell (x) = 1\).
Appendix E. Numerical BondOrder Potentials (BOP)
In mathematical terms, the idea behind BOP methods is to replace the local density of states (LDOS) with an approximation using only the information from the truncated tridiagonal matrix \(T_N\) (and possibly additional hyperparameters). Since the first N coefficients contain the same information as the first \(2N+1\) moments \(\mathcal {H}_{\ell \ell }, \dots , [\mathcal {H}^{2N+1}]_{\ell \ell }\), this approach is closely related to the method of moments [22].
Equivalently, the resolvent \([(z  \mathcal {H})^{1}]_{\ell \ell }\), which can be written conveniently as the continued fraction expansion
is replaced with an approximation \(G_\ell ^N\) only involving the coefficients from \(T_N\). For example, for fixed terminator \(t_\infty \), we may define
Truncating (E.1) to level N, which is equivalent to replacing the farfield of the linear chain with vacuum and choosing \(t_\infty = 0\), results in a rational approximation to the resolvent and thus a discrete approximation to the LDOS. We have seen that truncation of the continued fraction in this way leads to an approximation scheme satisfying Theorem 2.5.
Alternatively, the farfield may be replaced with a constant linear chain with \(a_{N+j} = a_\infty \) and \(b_{N+j} = b_{\infty }\) for all \(j \geqslant 1\) leading to the square root terminator \(t_\infty (z) = \frac{b_\infty ^2}{za_\infty  t_\infty (z)}\) [38, 49, 97].
More generally, one may choose any “approximate” local density of states \(\widetilde{D}_\ell \) and construct a corresponding terminator that encodes the information from \(\widetilde{D}_\ell \) [52, 65]. For example, \(\widetilde{D}_\ell (x) :=\frac{1}{b_\infty \pi } \sqrt{1  \big (\frac{x  a_\infty }{2b_\infty }\big )^2}\) results in the square root terminator. While we are unaware of any rigorous results, there is numerical evidence [52] to suggest that the error in the approximation scheme is related to the smoothness of the difference \(D_\ell  \widetilde{D}_\ell \).
Equivalently, we may choose any bounded symmetric tridiagonal (Jacobi) operator \(\widetilde{T}_N\) with diagonal \(a_0, a_1, \dots , a_N, \widetilde{a}_{N+1}, \dots \) and offdiagonal \(b_1,\dots ,b_N,\widetilde{b}_{N+1},\dots \). That is, we may evaluate the recursion method exactly to level N and append the farfield boundary condition \(\{\widetilde{a}_n, \widetilde{b}_n\}_{n \geqslant N+1}\) to the semiinfinite linear chain. This approach also includes the case \(t_\infty = 0\) as in § 4.3 by choosing \(\widetilde{a}_n = \widetilde{b}_n = 0\) for all n.
With this in hand, we define
where \(\widetilde{D}_\ell ^{{2N+1},\mathrm {BOP}}\) is the appropriate spectral measure corresponding to \(\widetilde{T}_N\).
1.1 E.1. Error estimates
Since \([(\widetilde{T}_N)^n]_{00} = [(T_N)^n]_{00} = [(T_\infty )^n]_{00}\) is independent of the farfield coefficients \(\{\widetilde{a}_j, \widetilde{b}_j\}\) for all \(n \leqslant 2N+1\), we can immediately see that the first \(2N+1\) moments of \(\widetilde{D}_\ell ^{{2N+1},\mathrm {BOP}}\) agree with those of \(D_\ell \). In particular, we may immediately apply (2.14) to obtain error estimates that depend on \(\mathrm {supp}\big ( D_\ell  \widetilde{D}_\ell ^{2N+1,\mathrm {BOP}}\big )\).
Therefore, as long as the farfield boundary condition is chosen so that there are only finitely many discrete eigenvalues in the band gap independent of N, the more complicated BOP schemes converge at least as quickly as the \(t_\infty = 0\) case. Intuitively, if the farfield boundary condition is chosen to capture the behaviour of the LDOS (e.g. the type and location of bandedge singularities), then the integration against the signed measure \(D_\ell  \widetilde{D}_\ell ^{2N+1,\mathrm {BOP}}\) as in (2.14) may lead to improved error estimates. A rigorous error analysis to this affect is left for future work.
1.2 E.2. Analyticity
Since \(\widetilde{T}_N\) is bounded and symmetric, the spectrum \(\sigma (\widetilde{T}_N)\) is contained in a bounded interval of the real line. In particular, we can apply the same arguments as in (4.44) to conclude that (E.3) defines a nonlinear approximation scheme given by an analytic function on an open subset of \({\mathbb {C}}^{2N+1}\).
Appendix F. Kernel Polynomial Method & Analytic Bond Order Potentials
We first introduce the Kernel Polynomial Method (KPM) for approximating the LDOS [82, 83, 98]. In this section, we scale the spectrum and assume that \(\sigma (\mathcal {H}) \subset [1,1]\).
For a sequence of kernels \(K_N(x,y)\), we define the approximate quantities of interest
Under the choice \(K_N(x,y) :=\frac{2}{\pi } \sqrt{1y^2} \sum _{n=0}^N U_n(x) U_n(y)\) (where \(U_n\) denotes the \(n^\text {th}\) Chebyshev polynomial of the second kind), we arrive at a projection method similar to that discussed in § 4.1.4: if \(O(x) = \sum _{m=0}^\infty c_m U_m(x)\), then
Equivalently, we may consider the corresponding approximate LDOS
However, truncation of the Chebyshev series in this way leads to artificial oscillations in the approximate LDOS known as Gibbs oscillations [43]. Moreover, without damping these oscillations, the approximate LDOS need not be positive. However, on defining
we obtain a positive approximate LDOS [98] where the damping coefficients \(d_n :=(1  \tfrac{n}{N})\) reduce the effect of Gibbs ringing. In practice, one may instead choose the Jackson kernel [47].
The problem with the above analysis in practice is that the damping factors that we have introduced mean that more moments \([\mathcal {H}^n]_{\ell \ell }\) are required in order to obtain good approximations to the LDOS. Instead, analytic BOP methods [74, 79] compute the first N rows of the tridiagonal operator \(T_\infty \), thus obtaining the first \(2N+1\) moments exactly. Then, a farfield boundary condition (such as a constant infinite linear chain) is appended to form a corresponding Jacobi operator \(\widetilde{T}_N\) as in Appendix E. Now, since higher order moments of \(\widetilde{T}_N\) can be efficiently computed, we may evaluate the following approximate LDOS
where \(d_n\) are damping coefficients and \(M > 2N+1\). The damping is chosen so that the lower order moments which are computed exactly and are more important for the reconstruction of the LDOS are only slightly damped. With this choice of kernel, the approximate quantities of interest take the form
Efficient implementation of analytic BOP methods can be carried out using the BOPfox program [47].
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Thomas, J., Chen, H. & Ortner, C. BodyOrdered Approximations of Atomic Properties. Arch Rational Mech Anal 246, 1–60 (2022). https://doi.org/10.1007/s0020502201809w
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s0020502201809w