High-Contrast Random Composites: Homogenisation Framework and Spectral Convergence Authors: Mikhail Cherdantsev, Kirill Cherednichenko, Igor Velčić Published: 2026 ABSTRACT AbstractWe develop a framework for multiscale analysis of elliptic operators with high-contrast random coefficients. For a general class of such operators, we provide a detailed spectral analysis of the corresponding homogenised limit operator. Under some lenient assumptions on the configuration of the random inclusions, we fully characterise the limit of the spectra of the high-contrast operators in question, which unlike in the periodic setting is shown to be different to the spectrum of the homogenised operator. Introducing a new notion of the relevant limiting spectrum, we describe the connection between these two sets. Introduction A recent drive towards the development of mathematical tools for understanding the behaviour of realistic inhomogeneous media has led to a renewal of interest to stochastic homogenisation and to an equally explosive activity in the analysis of composite media, in particular, of those with contrasting material components. In those fields the new step change is characterised by moving from qualitative analysis to quantitative results: from the general multiscale frameworks by De Giorgi and Spagnolo [22], Bakhvalov [6], Murat–Tartar [37, 44], Allaire [2], Kamotski–Smyshlyaev [33] (in the periodic setting), Yurinskii [45], Papanicolaou-Varadhan [38], Kozlov [36], Zhikov–Piatnitski [49] (in the stochastic setting), Zhikov [46, 47] (in the high-contrast periodic setting) to error bounds in appropriate functional topologies by Griso [26], Zhikov–Pastukhova [48], Birman–Suslina [9], Kenig–Lin–Shen [34] (periodic), Gloria–Otto [24], Armstrong–Smart [5], Armstrong–Kuusi–Mourrat [4] (stochastic), Cherednichenko–Cooper [17], Cherednichenko–Ershova–Kiselev [18], Cooper–Kamotski–Smyshlyaev [20] (high-contrast periodic). The high-contrast setting occupies a special place in homogenisation theory, for it enables resonant phenomena leading to micro-to-macro scale interactions, thereby bringing about new, often not naturally occurring, material properties. While the periodic high-contrast setting has now been analysed in good detail, its stochastic counterpart, which can be argued to be even more relevant to applications in material science, has not yet enjoyed the deserved attention: apart from our tentative study [16] in the bounded domain setting, we are aware of only three more works in this area—[10, 12] and [3]. Note that in [12] the authors use the notion of stochastic two-scale convergence in mean, introduced in [13], which is not convenient for the spectral analysis presented here (or in [16]). Indeed, the convergence in mean does not imply the convergence almost surely, and it is not clear what are the implications of the resolvent convergence in mean in relation to the limiting spectrum. On the other hand, the notion of stochastic two-scale convergence introduced in [49] provides necessary tools for the analysis of the limiting spectrum in the case of a bounded domain, see [16] and Theorem 4.3 below. The goal of the present paper is to provide a general framework and a comprehensive toolbox for multiscale analysis of random high-contrast media, which would lay a foundation for the related research avenue. We study the problem of homogenisation of operators of the form Aε=-∇·Aε∇ with high-contrast random (stochastic) coefficients, represented by the matrix Aε, which models a two-component material with randomly distributed and randomly shaped “soft” inclusions, whose typical size and spacing are both of order ε≪1, embedded in a “stiff” component. It is assumed that the ellipticity constants of Aε is of order 1 in the stiff component and of order ε2 in the inclusions. This scaling regime between the size of the microstructure and the ratio between the coefficients is often referred as the double porosity model. Our interest in high-contrast homogenisation problems is motivated by the band-gap structure of the spectra of the associated operators. We emphasise that this spectral phenomenon has only been observed in the homogenisation limit under the double porosity scaling. Therefore, while a more general non-uniformly elliptic setting can be very challenging, see e.g. [7], in order to obtain the specific spectral behaviour one has to work with particular geometric constraints and scaling regimes. Composites exhibiting spectral gaps are widely used for manipulating acoustic and electromagnetic waves, see e.g. [31, 35]. These were first analysed from the mathematically rigorous perspective in [46, 47] in the periodic setting. It was shown that the spectra of Aε converge in the sense of Hausdorff to the spectrum of a limit homogenised operator Ahom. The latter has a two-scale structure that captures the macro- and microscopic behaviour of the operator Aε for small values of ε. The spectrum of Ahom has a band-gap structure (in the whole-space setting), characterised with respect to the spectral parameter λ by a function β(λ), that is explicitly determined by the microscopic part of Ahom and quantifies the resonant (or anti-resonant) effect of the soft inclusions. In the case of a bounded domain the analysis of the problem and the results are very similar, in particular, the point spectrum of Ahom “populates” the bands (accumulating at their right ends) corresponding to the whole-space case. In [16] we considered the high-contrast stochastic homogenisation problem in a bounded domain. To a certain extent, our findings as well as the basic techniques (modulo replacing the reference periodicity cell with the probability space, and the standard two-scale convergence with its stochastic counterpart) were similar to those of [46] in the periodic case. Namely, we showed that the homogenised operator has a similar two-scale structure with the “macroscopic” component -∇·A1hom∇ acting in the physical space and the “microscopic” one -ΔO acting in the probability space on a prototype “inclusion” O. We proved an appropriate (i.e. stochastic two-scale) version of the resolvent convergence of Aε to Ahom and the Hausdorff convergence of their spectra. However, due to the technical challenges of the stochastic setting our understanding of the homogenised operator Ahom was limited. In particular, we were able to describe its spectrum only for a range of explicit examples. Note that the stochastic two-scale resolvent convergence of the operators (implying limε→0Sp(Aε)⊃Sp(Ahom)), proven in [16], is valid both for a bounded domain and for the whole space without any changes to the proof. Unlike for the periodic high-contrast operators, whose spectra are described by similar multiscale arguments for a bounded domain and for the whole space (leading to closely related features in the two cases, albeit resulting in spectra of different types from the operator-theoretic perspective), in the stochastic setting the situation is fundamentally different. In the present paper we show that in the case of the whole space the spectrum Sp(Ahom) is, in general, a proper subset of the limit of Sp(Aε). In fact, the situation when the spectrum of Aε occupies the whole positive half-line is not uncommon. This additional part of the limit spectrum, which does not appear in the case of a bounded domain, is attributed to the stochastic nature of the problem. At the same time, at least in the context of the applications that we have in mind, this part of the limit spectrum could (arguably) be deemed “physically irrelevant", whereas the part of the spectrum of Aε that converges to Sp(Ahom) may be thought of as “physically relevant". In terms of the spectral analysis our main objectives are to a) provide a comprehensive analysis of the homogenised operator Ahom and its spectrum; b) characterise the limit of Sp(Aε); c) understand the relation between the limit spectrum and Sp(Ahom). We next compare our results with those of [3], which also considers a random high-contrast medium in the whole space. Under the assumption that the soft inclusions are copies (up to rotations) of a finite number of C2-smooth geometric shapes, the authors of [3] construct the limit operator and prove the strong convergence of associated semigroups for any finite time. As a consequence, they show that the spectrum of the limit operator is a subset of the limit of the spectra of the original operators. Our assumptions on the types of the inclusions and their regularity are somewhat more general; in particular, we allow uncountably many inclusion shapes. As we already mentioned, we provide a full characterisation of the limiting spectrum in the case of finite-range correlations and we use different techniques (which are closer to Zhikov’s approach). The overall focus of [3] is also different. Namely, each second-order divergence-form operators is the generator of a Markov semigroup. It is well known that in the high-contrast setting the evolution of the effective limit operator exhibits memory effects, which means that corresponding process is not Markovian. The goal of their work is to equip the coordinate process with additional components so that the dynamics of the enlarged process remains Markovian in the limit. In [10] the authors consider a scattering problem for Maxwell equations with an obstacle comprising of parallel random rods with high random permittivity. In particular, in the homogenisation limit, they obtain a formula for the effective permeability, which, in the context of electromagnetism, is an analogue of Zhikov’s β-function. It should be noted that their results are obtained for a specific random model, rather than for a general class of high-contrast random composites, as in the present paper. Similar to [3], they only analyse the spectrum of the two-scale limit operator, yet do not describe the limiting spectrum, which is strictly larger in their setting. We next outline the structure of the paper and discuss our results in more detail. In Sect. 3 we recall the basic facts and definitions of the probability framework, state the main assumptions, describe the problem, and give a brief overview of the results of [16]. We have simplified the main assumption of [16], requiring only uniform Lipschitz property (more precisely, uniform minimal smoothness) and boundedness of the inclusions. In particular, these assumptions guarantee an extension property (Theorem 3.8) and a density result (Lemma A.1). In Sect. 4 we provide a complete characterisation of the spectrum of Ahom via the spectra of the operators -ΔO and -∇·A1hom∇ and the stochastic analogue of Zhikov’s β-function, defined via the solution to a resolvent problem for -ΔO, see Theorem 4.3. We then study the properties of β=β(λ) and provide a formula for its recovery from one typical realisation (in accordance with the ergodicity framework), construct and analyse the resolution of identity for the operator -ΔO, as well as characterise the point spectrum of Ahom. In the same section we improve and generalise results of [16] by providing a description of the spectra of -ΔO and Ahom under general assumptions rather than for specific examples. In particular, this description implies (again, under general assumptions) the convergence of spectra limε→0Sp(Aε)=Sp(Ahom) in the case of a bounded domain. (In what follows we will simply write limSp(Aε) instead of limε→0Sp(Aε) for brevity.) The main results of this section are a stochastic analogue of those in the periodic case. (In the case of systems the function β is matrix valued, allowing for even richer spectral behaviour, see [15].) Section 5 is concerned with the study of the limit of Sp(Aε). To that end, we introduce a function β∞(λ,ω) (where ω is an element of the probability space). We use β∞(λ,ω) to define a set G which, as we prove later, contains the limit spectrum. While β(λ) in the present setting is the stochastic analogue of its counterpart in the periodic case, β∞(λ,ω) is a new object which has no analogues in the periodic setting. An intuitive explanation of the difference between β and β∞ is as follows: the values of β(λ) are determined via the ergodic limit (in other words, from the global average distribution of inclusions), whereas the values of β∞(λ) are determined, loosely speaking, by the areas with the least dense distribution of the inclusions (non-typical areas). The first of the two main results of the section (Theorem 5.2) is that the limit of Sp(Aε) is a subset of G. We do not know whether G is the actual limit in general. However, under an additional assumption of finite range of dependence of the spacial distribution of inclusions, we establish that limSp(Aε)=G (Theorem 5.5), which is the second main result of the section. We give the proof of Theorem 5.2 in Sect. 5.3, while in Sect. 5.4 we provide auxiliary statements on the existence of cubes with approximately periodic arrangements of inclusions. Finally, in Sect. 5.5 we prove Theorem 5.5. In Sect. 5.6 we provide several examples illustrating our results: namely, we consider a (non-periodic) example where β∞=β (and, thus, limSp(Aε)=Sp(Ahom)) and those where this is not the case, i.e. limSp(Aε)≠Sp(Ahom). Analogous phenomenology can be observed for Schrödinger operators perturbed by random potentials, when the limiting spectrum is larger than the spectrum of the limit operator, see Remark 5.7. It is not surprising that this also happens in high-contrast setting, but here one has to make a significant effort to obtain the desired result, since the randomness enters the symbol of the differential operator and therefore has a more singular nature that the classical perturbation of the Schrödinger operator, which takes place within the potential. In Sect. 6 we explore the connection between the spectrum of Ahom and the limiting behaviour (as ε→0) of the spectra of Aε. For the family Aε, we introduce a notion of the relevant limiting spectrum, denoted by R-limSp(Aε)—namely, a subset of limSp(Aε) whose points are characterised by the existence of approximate eigenfunctions with significant part of their energy remaining inside a (large) fixed neighbourhood of the origin as ε→0. On the other hand, those λ that have (approximate) eigenfunctions with energy mainly concentrated in regions with non-typical (in the same sense as above) inclusion distributions, and thus located far from the origin for small ε, constitute the irrelevant limiting spectrum. We show that R-limSp(Aε) coincides with Sp(Ahom), see Theorem 6.6. We then discuss the implications of this result for the parabolic and hyperbolic evolution semigroups, cf. Corollary 6.9: namely, we show that on any finite domain one can neglect, as ε→0, the part of the initial conditions corresponding to the irrelevant limiting spectrum. A number of technical preliminaries, constructions and auxiliary statements that we use in the paper are presented in the appendices. In particular, we prove the following results: a higher regularity of the periodic homogenisation corrector for perforated domains, see Theorem 5.20, which is discussed in Appendix C; the extension property for potential vector fields in the probability space, see Proposition D.2; and a corollary of the latter, see Lemma D.3, which is important for the stochastic homogenisation corrector for perforated domains. Notation Index In order to help the reader navigate the text, we provide a short description of some important notation and indicate where it is introduced. General notation: For a Lebesgue measurable set U∈Rd, we denote by |U| its measure; For a Lebesgue measurable set U∈Rd, we denote by |U| its measure; Uc is the complement of a set U; Uc is the complement of a set U; U¯ is the closure of a set U in a relevant topology; U¯ is the closure of a set U in a relevant topology; 1U is the indicator function of a set U; 1U is the indicator function of a set U; BR(x) is the ball of radius R centred at x, and BR:=BR(0); BR(x) is the ball of radius R centred at x, and BR:=BR(0); xj is the j-th component of a vector x∈Rd; xj is the j-th component of a vector x∈Rd; Re(z) is the real part of z∈C; Re(z) is the real part of z∈C; Dom(·) is the domain of an operator or a function, depending on the context; Dom(·) is the domain of an operator or a function, depending on the context; Sp(·) stands for the spectrum of an operator. Sp(·) stands for the spectrum of an operator. Section 3: (Ω,F,P) is a complete probability space; (Ω,F,P) is a complete probability space; Tx is a dynamical system on (Ω,F,P); Tx is a dynamical system on (Ω,F,P); O∈Ω is a reference set for the set of inclusions; O∈Ω is a reference set for the set of inclusions; Oω, Oωk, Bωk are: the set of inclusions, an individual inclusion, and its extension set, respectively; Oω, Oωk, Bωk are: the set of inclusions, an individual inclusion, and its extension set, respectively; ρ,N,γ are the constants of minimal smoothness; ρ,N,γ are the constants of minimal smoothness; Dωk is the “left-bottom” vertex of the minimal cube containing Oωk; Dωk is the “left-bottom” vertex of the minimal cube containing Oωk; □ is the unit cube centred at the origin; □ is the unit cube centred at the origin; S0ε(ω), S1ε(ω), χ1ε(ω),χ0ε(ω) are the set of (ε-scaled) inclusions, its complement, and their characteristic functions; S0ε(ω), S1ε(ω), χ1ε(ω),χ0ε(ω) are the set of (ε-scaled) inclusions, its complement, and their characteristic functions; A1hom is the matrix of homogenised coefficients for the stiff component; A1hom is the matrix of homogenised coefficients for the stiff component; Aε(ω) and Ahom are the random high-contrast operator and the corresponding two-scale homogenised operator; Aε(ω) and Ahom are the random high-contrast operator and the corresponding two-scale homogenised operator; E(-∞,λ]ε and E(-∞,λ]hom are the spectral projections of Aε and Ahom, respectively; E(-∞,λ]ε and E(-∞,λ]hom are the spectral projections of Aε and Ahom, respectively; ΔO is the probabilistic “Dirichlet” Laplace operator on O; ΔO is the probabilistic “Dirichlet” Laplace operator on O; for an open set U⊂Rd we denote by ΔU the Dirichlet Laplace operator on U; for an open set U⊂Rd we denote by ΔU the Dirichlet Laplace operator on U; H and V are functions spaces in which the two-scale homogenised operator acts; H and V are functions spaces in which the two-scale homogenised operator acts; ⟨f⟩:=∫Ωf. ⟨f⟩:=∫Ωf. Section 4: β(λ) is the stochastic version of Zhikov’s β-function; β(λ) is the stochastic version of Zhikov’s β-function; Pω is an appropriately shifted inclusion which contains the origin; Pω is an appropriately shifted inclusion which contains the origin; Λs and Ψsp are eigenvalues and eigenfunctions of -ΔPω; Λs and Ψsp are eigenvalues and eigenfunctions of -ΔPω; dλ:=dist(λ,Sp(-ΔO)); dλ:=dist(λ,Sp(-ΔO)); E[0,t] is the resolution of identity for -ΔO. E[0,t] is the resolution of identity for -ΔO. Section 5: β∞(λ) is a “local–global” analogue of β(λ); β∞(λ) is a “local–global” analogue of β(λ); ℓ(x,M,λ,ω) is the local spectral average; ℓ(x,M,λ,ω) is the local spectral average; □xM is the cube of edge length M centred at x, □M is the cube of edge length M centred at the origin; □xM is the cube of edge length M centred at x, □M is the cube of edge length M centred at the origin; G is the “upper bound” for the limit of the spectra of Aε; G is the “upper bound” for the limit of the spectra of Aε; N^j, Nj, j=1,…,d, are the periodic homogenisation correctors on a (large) cube and their shift; N^j, Nj, j=1,…,d, are the periodic homogenisation correctors on a (large) cube and their shift; denotes the average value; denotes the average value; aε is the bilinear form associated with Aε+1; aε is the bilinear form associated with Aε+1; Section 6: N^jε, j=1,…,d, are the homogenisation correctors; N^jε, j=1,…,d, are the homogenisation correctors; gj,gjε,Gjε,j=1,⋯,d, are the difference of fluxes in the probability space, its realisation in physical space, and the flux corrector, respectively. gj,gjε,Gjε,j=1,⋯,d, are the difference of fluxes in the probability space, its realisation in physical space, and the flux corrector, respectively. Appendix A: ∇ω is the gradient in the probability space; ∇ω is the gradient in the probability space; C∞(Ω), C0∞(O), Wk,2(Ω),W∞,2(Ω) and W01,2(O) are the standard spaces of functions on Ω; C∞(Ω), C0∞(O), Wk,2(Ω),W∞,2(Ω) and W01,2(O) are the standard spaces of functions on Ω; Lpot2(Ω), Lsol2(Ω), Vpot2 and Vsol2 are the spaces of potential and solenoidal vector fields on Ω and their zero-mean subspaces. Lpot2(Ω), Lsol2(Ω), Vpot2 and Vsol2 are the spaces of potential and solenoidal vector fields on Ω and their zero-mean subspaces. Preliminaries and Problem Setting The setting of the problem is similar to [16]; however, we simplify the main assumption of [16] and deal mostly with the whole space Rd rather than a bounded domain. We provide some standard definitions, such as those of Sobolev spaces of functions defined on a probability space and the stochastic two-scale convergence, as well as technical statements, such as measurability properties of various mappings in Appendixes A and B. We will often use these tacitly throughout the text. Probability framework Let (Ω,F,P) be a complete probability space. We assume that the σ-algebra F is countably generated, which implies that the spaces Lp(Ω), p∈[1,∞), are separable. Definition 3.1 A family {Tx}x∈Rd of measurable bijective mappings Tx:Ω→Ω on a probability space (Ω,F,P) is called a dynamical system if: Tx∘Ty=Tx+y ∀x,y∈Rd; Tx∘Ty=Tx+y ∀x,y∈Rd; P(TxF)=P(F) ∀x∈Rd, F∈F; P(TxF)=P(F) ∀x∈Rd, F∈F; T:Rd×Ω→Ω, (x,ω)→Tx(ω) is measurable (for the standard σ-algebra on the product space, where on Rd we take the Borel σ-algebra). T:Rd×Ω→Ω, (x,ω)→Tx(ω) is measurable (for the standard σ-algebra on the product space, where on Rd we take the Borel σ-algebra). Definition 3.2 A dynamical system is called ergodic if one of the following equivalent conditions is fulfilled: f measurable, f(ω)=f(Txω)∀x∈Rd,a.e.ω∈Ω⇒fis constantP-aeω∈Ω. f measurable, f(ω)=f(Txω)∀x∈Rd,a.e.ω∈Ω⇒fis constantP-aeω∈Ω. P((TxB∪B)\(TxB∩B))=0∀x∈Rd⇒P(B)∈{0,1}. P((TxB∪B)\(TxB∩B))=0∀x∈Rd⇒P(B)∈{0,1}. Remark 3.3 Note that if the symmetric difference of TxB and B is empty, then the condition b in the above definition immediately implies P(B)∈{0,1}. It can be shown (e.g., [21]) that ergodicity is also equivalent to the following (a priori weaker) implication: Henceforth we assume that the probability space (Ω,F,P) is equipped with an ergodic dynamical system Tx. The dynamical system is used to establish a connection between the probability space (Ω,F,P) and the physical space Rd. Namely, for f∈Lp(Ω) and a.e. ω∈Ω the expression f(Txω),x∈Rd, defines an element of Llocp(Rd), which is called the realisation of f for ω. Note that the associated random field f(x,ω):=f(Txω) (considered as a function of x and ω) is an element of Llocp(Rd,Lp(Ω)). The differentiation operations in L2(Ω) as well as the associated spaces C∞(Ω), Wk,2(Ω), W∞,2(Ω) are introduced in a standard way. We also use the spaces of potential and solenoidal vector fields Lpot2(Ω) and Lsol2(Ω) and their zero-mean subspaces Vpot2 and Vsol2, see [30]. We provide the relevant definitions in Appendix A. The following theorem is the key tool in setting the basis for our analysis. Theorem 3.4 (“Ergodic Theorem”). Consider a probability space (Ω,F,P) and an ergodic dynamical system {Tx}x∈Rd on Ω. Suppose that f∈L1(Ω) and S⊂Rd is a bounded open set. Then for P-a.e. ω∈Ω one has Furthermore, for all f∈Lp(Ω), 1≤p≤∞, and a.e. ω∈Ω, the function f(x,ω)=f(Txω) satisfies f(·,ω)∈Llocp(Rd). For p<∞ one has f(·/ε,ω)=f(T·/εω)⇀∫ΩfdP weakly in Llocp(Rd) as ε→0. Note that the Ergodic Theorem implies only that the identity (1) holds P-a.e. for countably many functions simultaneously. This is why in the definition of the two-scale convergence (Definition A.2) one has to restrict to a countable (dense) subset of test functions, see also e.g. [29]. In general, statements that only require a simple convergence argument, such as passing to the limit in the resolvent problem, are valid for a set of full probability measure. However, throughout the paper we have a number of results, notably Theorem 5.5 and Proposition 5.11, where one does not invoke the ergodic theorem but uses the law of large numbers instead for the analysis of non-typical areas (as discussed in Introduction). Such results hold on a full-measure subset of Ω that in general is different from the one provided by the ergodic theorem. Thus in what follows all results can be viewed as valid on a set Ωt of full measure, for whose elements a) the Ergodic theorem holds on a suitable dense subset of L1(Ω), b) the application of the law of large numbers to a suitable collection of independent random variables (such as Theorem 5.18) is valid. We henceforth always assume that ω∈Ωt, and refer to it as ‘typical’. The main assumption and the extension property We proceed with defining the set of inclusions Oω=⋃k=1∞Oωk. As we have already mentioned, we simplify the assumptions of our previous paper [16]. It turns out that, in order to guarantee the validity of the main results, it suffices to require the uniform minimal smoothness of the inclusions, their boundedness, and uniform separation. In particular, Assumption 3.6 implies the key extension property, Theorem 3.8 (simply postulated in our previous paper [16]), and a density result, Lemma A.1. We first recall the definition of a minimally smooth set, see [42]. Definition 3.5 An open set U⊂Rd is said to be minimally smooth with constants ρ,N,γ (or (ρ,N,γ) minimally smooth), if there is a countable sequence of open sets {Ui}i=1∞ covering the boundary ∂U such that Each x∈Rd is contained in at most N of the sets Ui. Each x∈Rd is contained in at most N of the sets Ui. For any x∈∂U the ball Bρ(x) is contained in at least one Ui. For any x∈∂U the ball Bρ(x) is contained in at least one Ui. For any i∈N the portion of the boundary ∂U inside Ui is, in some Cartesian frame, the graph of a Lipschitz function whose Lipschitz semi-norm is at most γ. For any i∈N the portion of the boundary ∂U inside Ui is, in some Cartesian frame, the graph of a Lipschitz function whose Lipschitz semi-norm is at most γ. Let a set O⊆Ω be such that 00 (independent of ω) such that for all k,l∈N, k≠l, one has dist(Oωk,Oωl)>r0. There exists r0>0 (independent of ω) such that for all k,l∈N, k≠l, one has dist(Oωk,Oωl)>r0. For the extension result below we need to make the following observation: under Assumption 3.6, for every inclusion Oωk there exists an ‘extension domain’ Bωk with the following properties: Bωk is open, bounded, Bωk⊃Oωk¯, Bωk∩Oω=Oωk, and the set Bωk\Oωk¯ is (ρ,N,γ) minimally smooth with possibly different constants (ρ,N,γ) uniform in ω and k. In order to make the notation simple, we choose the constants so that all the sets Oωk, Bωk, and Bωk\Oωk¯ are (ρ,N,γ) minimally smooth uniformly in ω and k. Moreover, we can assume that appropriately translated, the domains Bωk fits into the unit cube □:=[-12,12)d. More precisely, let the vector Dωk∈Rd be defined by and denote d1/4:=(1/4,⋯,1/4)⊤, then Oωk-Dωk-d1/4⊂Bωk-Dωk-d1/4⊂□. We provide a sketch of the proof of the above claim for the 2d case, which can be easily extended to any dimension. For sufficiently small τ>0 we cover R2 with closed squares with sides τ and mutually disjoint interiors. For a given Oωk we choose the squares that have non-empty intersection with Oωk together with their “neighbours”, i.e. with those squares that share with them a common side or vertex. The union of these squares, denoted by U, is obviously connected but may fail to be Lipschitz if it contains two squares that share a single vertex and the adjacent sides form part of the boundary of U. In order to rectify this problem we add to U smaller squares with sides τ4 with centres at all vertices contained in U. Choosing τ small enough we obtain an extension domain Bωk which is minimally smooth with constants (ρ′,N′,γ′) depending only on τ. The choice of τ, in turn, is independent of k and ω and only depends on (ρ,N,γ) due to Assumption 3.6. Remark 3.7 The requirement of separation of inclusions (perforations) is standard when dealing with problems in perforated domains, see e.g. [27, 40], as it allows one to use extension techniques. This assumption is even more important for spectral analysis, as we already mentioned in the Introduction. Indeed, allowing inclusions to touch each other may have unpredictable consequences for the associated spectra, leading, in particular, to uncontrollable “pollution” of the gaps in the limiting spectrum by the spectrum arising from clusters of inclusions stuck together. Next we formulate the mentioned extension property, which is valid under Assumption 3.6. Theorem 3.8 There exist bounded linear extension operators Ek:W1,p(Bωk\Oωk)→W1,p(Bωk), p≥1, such that the extension u~:=Eku, u∈W1,p(Bωk\Oωk), satisfies the bounds where the constant Cext depends only on ρ,N,γ,p, and is independent of ω and k. Additionally, in the case p=2 the extension can be chosen to be harmonic in Oωk, Proof Inequalities (2) and (3) are direct corollaries of a classical result due to Calderón and Stein on the existence of uniformly bounded extension operators [42, Chapter IV, Theorem 5] and a uniform Poincaré inequality [27, Proposition 3.2] for minimally smooth domains. We make two observations. First, though it is not stated explicitly in [42], the norm of the extension operator depends only on the constants of the minimal smoothness assumption, which can be seen from the proof below. Second, the proof of the uniform Poincaré inequality in [27, Proposition 3.2] works for any p≥1 without amendments. Now let p=2 and fix Oωk. We look for the harmonic extension u~ in the form u~=Eku+u^, where u^∈W01,2(Oωk) (here Ek denotes the extension operator from [42]). If such u^ exists, it satisfies the identity As per discussion above, we have uniformly in ω and k. Thus the right-hand side in (4) is a bounded linear functional on W01,2(Oωk) equipped with the norm ‖∇φ‖L2(Oωk). Therefore, by the Riesz representation theorem the solution u^ of (4) exists and satisfies the bound ‖∇u^‖L2(Oωk)≤Cext‖∇u‖L2(Bωk\Oωk). The bounds (2) and (3) now follow easily. □ Remark 3.9 The above theorem can be reformulated in an obvious way for any family of domains (i.e. not related to the stochastic setting) satisfying the same minimal smoothness condition. Problem setting and an overview of the existing results For ε>0 we define The corresponding set indicator functions are denoted by χ0ε(ω) and χ1ε(ω), respectively. We consider the self-adjoint operator Aε(ω) in L2(S), where S⊂Rd denotes either a bounded domain with Lipschitz boundary or the whole space Rd, generated by the bilinear form where and A1 is a symmetric positive-definite matrix. Note that W01,2(Rd)=W1,2(Rd) when S=Rd. In what follows we assume that ω is fixed and will drop it from the notation of the operator, writing simply Aε. We recall the definition of the limit homogenised operator Ahom given in [16]. The matrix of homogenised coefficients A1hom arising from the stiff component is defined by It is well known, see [30, Chapter 8], that in order for the matrix A1hom to be positive definite it suffices to have a certain kind of extension property. It is not difficult to see that a slight adaptation of the argument in [30, Lemma 8.8] ensures that A1hom is indeed positive definite under Assumption 3.6. The variational problem (6) has a unique solution pξ in X, where X denotes the completion in L2(Ω\O) of Vpot2. This solution satisfies the identity In Appendix D we prove that under our assumptions on the regularity of the inclusions the space X can be viewed merely as the restriction of Vpot2 to Ω\O (see Lemma D.3). We define the space which is naturally embedded in L2(S×Ω), and its dense (by Lemma A.1) subspace where W01,2(O) is the Sobolev space of functions on Ω vanishing outside O, see Appendix A for the precise definition. For f∈H, we consider the following resolvent problem: find u0+u1∈V such that where ∇ω denotes the gradient in the probability space (see Appendix A). This problem gives rise to a positive definite operator Ahom in H, see [16] for details. Note that one can take an arbitrary f∈L2(S×Ω) in (8). In this case the solution coincides with the solution for the right-hand side Pf, where P is the orthogonal projection onto H, namely, we have u0+u1=(Ahom+λ)-1Pf. Furthermore, the identity (8) can be written as a coupled system Here and in what follows, unless specified otherwise, we often use the notation ⟨f⟩:=∫Ωf, tacitly assuming that functions defined only on O are extended by zero to the whole Ω. The following statement of stochastic two-scale resolvent convergence (which we will refer to simply as “resolvent convergence” for brevity), which was proved in [16] in the case of bounded S, remains valid for S=Rd with exactly the same proof. (For the notion of weak (strong) stochastic two-scale convergence, denoted by , we refer the reader to Appendix A). Theorem 3.10 Under Assumption 3.6, let λ>0 and suppose that fε is a bounded sequence in L2(S) such that Consider the sequence of solutions uε to the resolvent problem Then for a.e. ω∈Ω one has , where u0+u1 is the solution to (8). This resolvent convergence can be put in a more abstract framework of resolvent convergence in variable spaces, see [39]. It is shown that the weak resolvent convergence is equivalent to the strong one. Moreover, it entails convergence of hyperbolic semigroups and certain convergence properties for the spectra. In particular, one has “half” of the Hausdorff convergence of the spectra: where by limSp(Aε) we understand the set of all λ such that ∃{λε}, λε∈Sp(Aε), λε→λ. Finally, the following (strong stochastic two-scale) convergence for the spectral projections of Aε and Ahom holds: unless λ is an eigenvalue of Ahom. This work pursues three main goals warranted by the property (11): understanding the spectrum of Ahom, obtaining a practically useful characterisation of the limiting spectrum limSp(Aε), and investigating the nature of the difference between the two sets—limSp(Aε)\Sp(Ahom). We next report (partial) results on Sp(Ahom) obtained in [16]. Denote by -ΔO the positive definite self-adjoint operator generated by the bilinear form One has 13Sp(-ΔO)=⋃k∈NSp(-ΔOωk)¯fora.e.ω∈Ω, where -ΔOωk is the Dirichlet Laplace operator on Oωk for each ω,k. One has 13Sp(-ΔO)=⋃k∈NSp(-ΔOωk)¯fora.e.ω∈Ω, where -ΔOωk is the Dirichlet Laplace operator on Oωk for each ω,k. In the case of bounded S and under the condition that Sp(-ΔO)⊂Sp(Ahom), the strong stochastic two-scale convergence of eigenfunctions of Aε holds, i.e. if Aεuε=λεuε,∫S|uε|2=1, and if λε→λ, then uε→2u, where Ahomu=λu. In the case of bounded S and under the condition that Sp(-ΔO)⊂Sp(Ahom), the strong stochastic two-scale convergence of eigenfunctions of Aε holds, i.e. if Aεuε=λεuε,∫S|uε|2=1, and if λε→λ, then uε→2u, where Ahomu=λu. In the case of bounded S, for two explicit examples of random distribution of inclusions, it was shown that Sp(-ΔO)⊂Sp(Ahom), Zhikov’s function β(λ) was calculated, and a complete characterisation of Sp(Ahom) was given. In the case of bounded S, for two explicit examples of random distribution of inclusions, it was shown that Sp(-ΔO)⊂Sp(Ahom), Zhikov’s function β(λ) was calculated, and a complete characterisation of Sp(Ahom) was given. In the next section we will extend and improve these results to the general setting of the present work. Remark 3.11 From (13) and property 2 of Assumption 3.6 it follows, via the Poincaré inequality and the min-max theorem, that infSp(-ΔO)>0. Remark 3.12 The identity (13) was proved in [16] under an additional assumption that the normalised eigenfunctions of the operators -ΔOωk,ω∈Ω,k∈N, are bounded in the L∞-norm on any bounded spectral interval uniformly in ω and k. It is not difficult to see that this property holds under Assumption 3.6: the proof of this fact follows closely the standard argument of the De Giorgi–Nash–Moser regularity theory, see, e.g. [8, Chapter 2.3], cf. a similar result in [14, Lemma 4.3], for more details. In fact, one does not necessarily need a uniform bound of L∞-norms of the eigenfunctions of -ΔOωk in order to prove (13). Instead, it is sufficient to note that the random variable ψa,b utilised in the proof of Theorem 5.1 in [16] belongs to L2(Ω), which follows by a standard argument (cf. Proposition 4.6) via integrating its realisation in the physical space first, then integrating over Ω and using the Fubini theorem. Remark 3.13 Another, perhaps more common, approach to describing random media consists in taking Ω to be a set of locally bounded Borel measures on Rd, which is chosen to be a separable metric space (so the corresponding σ-algebra consists of Borel sets). Thus, the medium comprising a matrix medium and inclusions can be seen as a random measure. The dynamical system is then defined using the stationarity property of this measure—however, the assumption of ergodicity does not follow from this construction. In fact, ergodicity is not necessary for a suitable version of the ergodic theorem to hold. Indeed, in its absence the expectation on the right-hand side of (1) has to be replaced by a conditional expectation. The ergodicity property in the classical moderate contrast setting guarantees that the coefficients of the limit equation are deterministic. In the case when ergodicity assumption is dropped, the coefficients of the limit equation are measurable with respect the σ-algebra of the invariant sets of the dynamical system. (Under the assumption of ergodicity this σ-algebra consists of sets of measure zero and one.) Point processes such as the Poisson process and the random parking model, used in the present work in the construction of examples of a random medium, in probability theory are usually looked at as random measures. A key benefit of this approach is that one can use Palm’s theory (see [28, Section 2.10, Section 2.11]), which is especially useful in the analysis of random structures (see [49]). In the present work we do not use Palm’s theory, as we find the adopted framework sufficient for our purposes. Homogenised Operator and Its Spectrum In this section we study the operator Ahom and its spectrum. In particular, our analysis allows us to improve a number of results of [16]. We begin with two general statements about the spectrum of Ahom, Proposition 4.1 and Theorem 4.3, namely, which were proved in [16] only for certain examples. Proposition 4.1 For the spectra of the operators Ahom, defined by the problem (9)-(10), and -ΔO, defined by the form (12), one has the inclusion Sp(-ΔO)⊂Sp(Ahom). Proof Suppose that λ∈R is in the resolvent set of Ahom, so that (8) has a solution u=u0+u1 for any f∈H. First we take a non-trivial f∈L2(S)\W1,2(S). Note that in this case λu0+f does not vanish. For the corresponding solution we have where 1O is the indicator function of the set O. Note also that for two arbitrary functions w∈L2(S×O) and h∈L2(S) one has ∫Swh∈L2(O). Then, multiplying the above identity by (λu0+f)‖λu0+f‖L2(S)-2 and integrating over S, we conclude that the function solves the equation Next, we take f=gψ with arbitrary non-trivial g∈L2(S) and ψ∈L2(O). Then for the corresponding solution of (8), which we denote by u~0+u~1, we have The difference between u~1 and u^1:=λu~0ϕ solves Multiplying the last equation by g‖g‖L2(S)-2 and integrating over S, we see that is a solution of We have shown that -ΔO-λI acts onto, therefore, by the bounded inverse theorem one concludes that (-ΔO-λI)-1 is bounded.1□ Now we are in a position to define an analogue of the Zhikov β-function. For λ∉Sp(-ΔO) we set where b=b(ω,λ) is the solution to the problem Remark 4.2 The function β(λ) in (14) is an analogue of Zhikov’s β-function introduced in [46, 47] for the periodic setting. In fact, if the probability space is such that the set of inclusions is periodic, then the expression on the right-hand side of (14) coincides with the known formula for the periodic setting. As we have mentioned in the Introduction, β(λ) (to be more precise, its second term) quantifies “resonant, anti-resonant” effect of the soft inclusions. In order to clarify this connection, we analyse the properties of the physical realisations of the function b(ω,λ) in Sect. 4.1, see Remark 4.8 for further discussion. The next theorem provides a characterisation of the spectrum of Ahom. Theorem 4.3 The spectrum of the homogenised operator is fully characterised by β(λ) and the spectra of the “microscopic” -ΔO and the “macroscopic” -∇·A1hom∇ operators as follows: 16Sp(Ahom)=Sp(-ΔO)∪{λ:β(λ)∈Sp(-∇·A1hom∇)}¯. The spectrum of the homogenised operator is fully characterised by β(λ) and the spectra of the “microscopic” -ΔO and the “macroscopic” -∇·A1hom∇ operators as follows: 16Sp(Ahom)=Sp(-ΔO)∪{λ:β(λ)∈Sp(-∇·A1hom∇)}¯. If S is bounded, then for a.e. ω∈Ω we have the spectral convergence Sp(Aε)→Sp(Ahom) in the sense of Hausdorff. If S is bounded, then for a.e. ω∈Ω we have the spectral convergence Sp(Aε)→Sp(Ahom) in the sense of Hausdorff. Proof Suppose that λ∉Sp(-ΔO). In order to solve (9)–(10) we first solve (10) for u1 in terms of an arbitrary fixed u0∈W01,2(S): We then substitute the obtained expression into (9), Taking f(x,ω)=g(x)ψ(ω) with arbitrary g∈L2(S) and ψ∈dom(-ΔO), we see that the solvability of (17) is equivalent to β(λ)∉Sp(-∇·A1hom∇), and (16) follows. The second part of the statement is proved in [16]: the resolvent converges implies that the spectrum of the limit operator is contained in the limit of Sp(Aε) as ε→0, while strong two-scale convergence of eigenfunctions of ε problem implies that any limiting point of Sp(Aε) is contained in the spectrum of the limiting operator. □ Remark 4.4 In the case S=Rd one has Remark 4.5 While in the periodic case β(λ) blows up to ±∞ as λ approaches Sp(-ΔO), in the general stochastic setting this is not the case. It is not difficult to construct an example where β(λ) does not blow up near Sp(-ΔO): for instance, one can consider the example of randomly scaled inclusions presented in Sect. 5.6.4. In short, consider a model for which the inclusions are randomly scaled copies of a single set, where the random scaling parameter r takes values in some finite interval [r1,r2]. Then choosing the corresponding probability density so that it is positive on (r1,r2) and converges to zero sufficiently quickly near the ends of the interval, one gets that β(λ) converges to finite values as λ approaches Sp(-ΔO), cf. also (102) below. Properties of Zhikov’s β-function We begin by showing that the solution to (15), defined on the probability space (Ω,F,P), can be reconstructed from its physical realisations. Let ω∈O. Then by the definition of Oω there exist an inclusion containing the origin, Oω0∋0. We reserve the notation k0 for the index of such inclusion and denote Pω:=Oω0-Dω0-d1/4. For the Dirichlet Laplacian operator -ΔPω on Pω we denote by Λs=Λs(ω),s∈N, its eigenvalues arranged in the increasing order, Λs<Λs+1,s∈N, and by {Ψsp}s∈N,p=1,⋯,Ns the corresponding system of the orthonormalised eigenfunctions, Ψsp=Ψsp(·,ω), where Ns is the multiplicity of Λs. See Lemmata B.8 and B.10 in Appendix B for the measurability properties of the eigenvalues and their eigenfunctions. For ω∈O and λ∉Sp(-ΔO), we define b~(ω,λ;·)∈W01,2(Pω) as the solution to the problem We assume that b~ is extended by zero outside Pω. Note that by applying the spectral decomposition, we can write it in terms of the eigenfunctions of -ΔPω, as follows: The results of Appendix B imply that for a fixed λ the mapping ω↦b~(ω,λ;·), taking values in L2(Rd), is measurable (on the target space L2(Rd) we take Borel σ-algebra). For any locally integrable function f∈Lloc1(Rd) we denote by freg its regularisation, defined at x∈Rd by Note that the Lebesgue differentiation theorem states that freg=f a.e. Consider the function We make the following observations. First, using the regularised representative of b~ we ensure that the right-hand side of (20) is measurable. Second, for fixed ω and k one has b~reg(Txω,λ;·)=b~reg(Tyω,λ;·) for all x,y∈Oωk; therefore, fixing y∈Oωk, we have Third, the mapping x↦b˘(Txω,λ) is an element of Wloc1,2(Rd) for a.e. ω∈Ω. Finally, for a.e. ω∈O, x=D(ω) is a Lebesgue point of b~reg(ω,λ;·), i.e. Indeed, denote by N the subset of O consisting of ω for which (22) does not hold. By construction, we have for a positive R. Applying Fubini Theorem yields and the claim follows. Throughout the paper, we will use the notation Proposition 4.6 For λ∉Sp(-ΔO) the function b˘(·,λ) defined by (20) is the solution to (15), i.e. b˘ coincides with b. Proof First we show that b˘(·,λ)∈L2(O). Since b~(ω,λ,·) is the solution to (18), taking into account (13) and by the assumption |Oω0|≤2-d<1, we have It follows from Assumption 3.6 that the ball Bρ is contained in Bω0; in particular, it has no intersections with inclusions Oωk,k≠k0. Invoking (21) and (24), it follows that Integrating over Ω and using the Fubini theorem, we obtain Next we argue that b˘(·,λ)∈W1,2(O). To show this, we define the function g:Ω→Rd as follows: For a fixed ω∈O we clearly have (∇b~)reg(ω,λ;x)=∇b~reg(ω,λ;x)=∇b~(ω,λ;x) for a.e. x∈Rd. By observing that (which follows from (18)), it is not difficult to show that g∈L2(Ω;Rd), similarly to the claim that b˘(·,λ)∈L2(O). It remains to show that ∇ωb˘=g, which is equivalent to Denote by B˚R(ω) the union of the inclusions contained in the ball BR, We need the following simple assertion. Lemma 4.7 Suppose f∈L1(Ω), let U⊂Rd be an open bounded star-shaped set, and for a fixed δ>0 let (VR) be a sequence of measurable sets satisfying (R-δ)U⊂VR⊂RU. Then almost surely one has Proof The proof is straightforward and follows from the ergodic theorem and the observation that □ Applying the above lemma, we get Similarly to (21), for a fixed Oωk⊂BR and a fixed y∈Oωk we have Integrating by parts in each of the inclusions, we obtain which in conjunction with (26) implies (25). Writing (18) in the weak form, and arguing in a similar way to the above via Lemma 4.7, one easily arrives at which concludes the proof. □ Remark 4.8 The construction at the start of the section, in particular the formula (21), and Proposition 4.6 establish an explicit connection between the function b(ω,λ) and its realisation in the physical space. Namely, for a typical ω the restriction of b(Txω,λ) to an individual inclusion Oωk is the solution to the problem subject to the Dirichlet boundary condition on ∂Oωk. Writing the spectral decomposition of b(Txω,λ)|Oωk, cf. (19) and integrating it over Oωk, one obtains an expression similar to (28) below, which may be interpreted, depending on whether it is positive or negative, as the resonant or anti-resonant contribution of the individual inclusion. Applying Lemma 4.7 to the definition (14), we obtain the representation where the second term may be interpreted as the averaged resonant/anti-resonant contribution of the individual inclusions. We emphasise the importance of (27): in order to compute β(λ) one does not need to know b(·,λ) on Ω, but only its realisation b(Txω,λ) for a single typical ω. Proposition 4.9 The function β is differentiable and its derivative is uniformly positive. More precisely, β′(λ)≥1-P(O). Proof For ω∈O we have (cf. (19)) Observing that a direct calculation yields Furthermore, by construction, one has It is not difficult to see that on any closed interval [λ1,λ2]⊂Dom(β)=R\Sp(-ΔO) the right-hand side of (27) and its derivative converge uniformly as R→∞ (cf. (29) and (30) and Lemma 4.7). It follows that β(λ) is differentiable. Moreover, from (27) and (30) we have as claimed □ Remark 4.10 It was shown in [16] that almost surely. In particular, the set on the right-hand side is deterministic almost surely. For the sake of completeness, we provide a streamlined version of the proof of this statement in Proposition B.11. Resolution of identity for -ΔO Similarly to the formula (27) for β(λ), it is often useful (in examples and practical applications) to be able to characterise or reconstruct abstract probabilistic objects associated with the homogenised operator Ahom, such as -ΔO for example, from their realisations in the physical space (for a single typical ω∈Ω). For a fixed ω we can think of the family of Dirichlet Laplace operators defined on individual inclusions as the realisation of -ΔO in the physical space. More precisely, we next characterise the resolution of identity for -ΔO by using the spectral projections of these Dirichlet Laplace operators on the inclusions and pulling them back to the probability space. For 0≤t1≤t2 we define the mappings E~(t1,t2] and E(t1,t2] as follows. For all φ∈L2(O) let and The mapping E[t1,t2] is defined analogously. In what follows we extend functions from L2(O) by zero in Ω\O without mentioning it explicitly. Proposition 4.11 E[0,t] is the resolution of identity for the operator -ΔO. Proof Arguing as in the proof of Proposition 4.6, we conclude that there exists C>0, independent of t1 and t2, such that Furthermore, for all φ1,φ2∈L2(O) one has In order to prove that E[0,t] is the resolution of identity, we next verify the following properties: E[0,t] is an orthogonal projection; E[0,t] is an orthogonal projection; E[0,t1]≤E[0,t2] if t1≤t2; E[0,t1]≤E[0,t2] if t1≤t2; E[0,t] is right continuous in the strong topology; E[0,t] is right continuous in the strong topology; E[0,t]→0 if t→0 and E[0,t]→I if t→+∞ in the strong topology. E[0,t]→0 if t→0 and E[0,t]→I if t→+∞ in the strong topology. Using (32) and a representation of (E(t1,t2]φ1)(Txω) analogous to (21), it is easy to see that which implies (a). In the same way one can see that, for 0≤t1≤t2≤t3≤t4, so (b) holds. Finally, we prove (c). For φ∈L2(O) and a.e. ω∈Ω and ε=ε(ω)>0 small enough we have which follows from (31) and observing that Bρ has common points with at most one inclusion. Also, for all ε>0 one has almost surely. Integrating the right-hand side of the last inequality over Ω yields Thus, by the Lebesgue theorem on dominated convergence, we conclude that and (c) follows. Similarly we prove (d), cf. also Remark 3.11. Now, we prove that E[0,t] is the resolution of identity associated with the operator -ΔO, i.e., Consider the space By the property (d), Y is dense in L2(O). Let φ∈Y, then φ=E[0,t]φ for some t and Arguing similarly to the proof of Proposition 4.6 (i.e., using the definition of E[0,t]φ, passing to realisations, and using the ergodic theorem), one can show that Y⊂Dom(-ΔO), and for φ∈Y one has Comparing the right-hand sides of (33) and (34) on realisations, we see that on each inclusion Oωk they are linear combinations of the projections of φ(Txω) onto the eigenspaces of -ΔOωk and differ only by the coefficients mtn and Λs(Txω), respectively. In follows that Passing to the limit as R→∞ via Lemma 4.7, we get Now, passing to the limit as n→∞, we infer that Tφ=-ΔOφ, i.e. the (symmetric) restrictions of the operators T and -ΔO to Y coincide. To conclude the proof it remains to note that either of the restrictions is essentially self-adjoint. To this end, it suffices to note that the image of T|Y±i is dense in L2(O), since for all φ∈L2(O) one has where the limit is understood in the L2(O) sense. □ Remark 4.12 Similarly to Remark 4.8, one can recover the resolution of identity from realisations for a single (typical) ω via the formulae (31)–(32). Point spectrum of the homogenised operator In this section we first provide a general characterisation of the point spectrum of Ahom, arguing that it is determined completely by the point spectrum of -ΔO (in complete analogy with the periodic setting [47]). Combining it with the above construction of the resolution of identity, we then illustrate it by two classes of examples. Proposition 4.13 The point spectrum of Ahom (in the case S=Rd) consists of those eigenvalues of -ΔO whose eigenfunctions have zero mean: Proof Consider the following eigenvalue problem for Ahom (cf. (9)–(10)): where u0+u1∈V. Suppose that λ∈Spp(-ΔO) and the corresponding eigenfunction ψ has zero mean. Then the pair u0=0, u1=vψ, where v∈L2(Rd), clearly satisfies (35). Hence λ∈Spp(Ahom). Now suppose λ∈Spp(Ahom), i.e. (35) holds. We claim that u0=0. Indeed, assuming the contrary, we infer from the second equation in (35) that the problem is solvable. Note that all its solutions have the form v=v0+ψ, where v0 is a fixed solution and ψ∈E[λ,λ](L2(O)). Moreover, all elements of E[λ,λ](L2(O)) have zero mean by the solvability condition for (36). It is easy to see that the solution of the second equation in (35) has the form u1=λu0v0+φ, where φ∈L2(Rd;E[λ,λ](L2(O))). Substituting this into the first equation, we arrive at But this contradicts to the fact that the spectrum of -∇A1hom∇ in Rd has no eigenvalues, therefore u0=0, as required. Then necessarily ⟨u1⟩=0, and hence u1(x,·) is an eigenfunction of -ΔO for a.e. x∈Rd. □ Example: a finite number of inclusion shapes Assume that for a.e. ω the inclusions Oωk are (translated and rotated) copies of a finite number of sets (shapes). By (13), the spectrum of -ΔO is the union of the spectra of the Dirichlet Laplacians on these sets. In particular, Sp(-ΔO) is discrete. It is not difficult to see (cf. the construction of the resolution of identity) that the eigenspaces of -ΔO consist of those functions whose realisations are appropriate eigenfunctions on each inclusion. Thus, the point spectrum of Ahom consists of the eigenvalues of each shape whose all eigenfunctions have zero mean. Example: a continuum family of scaled copies of one shape Assume that for a.e. ω the inclusions Oωk are scaled (translated and rotated) copies of an open set O˚⊂Rd (shape), where the scaling parameter r takes values in some interval [r1,r2]⊂(0,+∞). We will write Oωk∼rO˚ if the inclusion Oωk is a translation and rotation of rO˚. Let {νj}j=1∞ be the sequence of eigenvalues for the Dirichlet Laplacian on O˚. Then the spectrum of the Dirichlet Laplacian on Oωk∼rO˚ is {r-2νj}j=1∞. Clearly, for any t∈R there are at most finitely many values of the scaling parameter satisfying t=r-2νj for some j. Consider the situation when the probability measure does not concentrate for any r∈[r1,r2]. We can make this assumption precise using the ergodicity rather than referring to an abstract probability space. Namely, assume that for a.e. ω and all r∈[r1,r2] one has equivalently, that It is then easy to see, cf. (32), that the resolution of identity E[0,t],t≥0, has no jumps, i.e., Spp(-ΔO)=∅. Convergence of the Spectra As we have already discussed in the Introduction, the spectrum of Aε and its limit behaviour in the whole space setting is very different from that in the case of a bounded domain. Many ideas we used in [16] were an adaptation of well-known techniques from periodic high-contrast homogenisation. The analysis of the whole space setting, however, calls for methods that go beyond those used in the bounded domains. The rest of the paper is devoted wholly to the former, and in what follows we will always assume that S=Rd. In the present section we analyse the limit of the spectra of Aε as ε→0. In order to understand the basic difference between the bounded domain and the whole space settings, one may think of the following simple example: the set of inclusions Oω is generated by putting balls of equal size with probability 12 at the nodes of a periodic lattice. The spectrum of Ahom then has a band-gap structure very similar to the corresponding periodic case, with the left ends of the bands moved slightly to the left. Yet, Sp(Aε)=R0+ for all ε. Indeed, due to the law of large numbers, there almost surely exists a sequence of balls BR(xR), R∈N, that is completely void of inclusions from εOω. Then, for an arbitrary λ≥0, by taking the real part of a harmonic wave Re(eik·x) with |k|2=λ, multiplying it by suitable cut-off functions and normalizing, one constructs an explicit, supported on BR(xR), R∈N, Weyl sequence for the operator Aε, thus asserting that Sp(Aε)=R+. In more advanced (and more realistic as far as applications concerned) examples, such as the random parking model, see Sect. 5.6.6 below, the spectrum of Sp(Aε) may exhibit gaps which persist in the limit. However, the general idea is the same: the additional part of the spectrum, not accounted for by Sp(Ahom) in the limit, is present due to the arbitrarily large areas of space with non-typical distribution of inclusions. In order to account for this part of the spectrum in the limit, one, somewhat surprisingly, may use a close analogue of the β-function given by the formula (27), which we will denote by β∞(λ,ω). Intuitively, one can think of is as the supremum of “local β-functions”, i.e. local averages of the expression λ+λ2b(Tx/εω,λ), whose intervals of negativity get bigger when the local volume fraction of the inclusions is relatively large, and, conversely, shrink when the local volume fraction of the inclusions is small (or completely disappear if the volume fraction tends to zero). In order to make this more precise, one needs to look at the behaviour of the distribution of inclusions in large randomly located sets. Main results For λ∉Sp(-ΔO), define where and □xM:=x+M□ is the cube of edge length M centred at x. We will use the shorthand notation □M:=M□=□0M. Note that ℓ(x,M,λ,·) is measurable, which implies the measurability of β∞(λ,·). Remark 5.1 One may interpret the term as the local averaged “resonant” (or “anti-resonant”, if it is negative) contribution from the inclusions (which from the physical point of view play a role of micro-resonators) to the term ℓ(x,M,λ,ω), see also Remark 4.8. The latter in turn may be interpreted as a “local spectral average”, a precursor for the “spectral parameter” β∞(λ,ω). By Proposition 5.11 below, the function β∞ is deterministic almost surely, that is β∞(λ,ω)=β∞(λ) for a.e. ω. The set is an “upper bound” for the limit of the spectra of Aε, as stated next. Theorem 5.2 One has limSp(Aε)⊂G almost surely. We provide the proof of this theorem in Sect. 5.3. In order to prove that G is actually the limit of Sp(Aε), we make an additional assumption of finite-range correlations, which we describe next. For a compact set K⊂Rd we denote by HK the set-valued mapping HK:ω→Oω∩K¯. By FH,K we denote the σ-algebra on the set of all compact subsets of K generated by the Hausdorff metric dH. The proof of the following lemma is completely analogous to that of Lemma B.3 and we omit it. Lemma 5.3 The set-valued mapping HK is measurable with respect to the σ-algebra FH,K. For a compact set K⊂Rd, define the σ-algebra FK by The following assumption is crucial for proving the second part of the spectral convergence. In simple words it says that the arrangements of inclusions in sets which are at least at distance κ apart are independent. Assumption 5.4 There exists κ∈R+ such that for any two compact sets K1,K2⊂Rd satisfying dist(K1,K2)>κ the σ-algebras FK1 and FK2 are independent. Theorem 5.5 Under Assumption 5.4 one has limSp(Aε)⊃G almost surely. (Hence, in view of Theorem 5.2, the two sets are equal.) The proof of the theorem is given in Sect. 5.5. Assumption 5.4 is only needed for the proof of Theorem 5.5. It is not required for any other result in this paper. (We believe that one can also relax it by allowing for sufficiently quickly (e.g., exponentially) decaying rather than finite-range correlation.) Finite range of dependence guarantees the almost sure existence of arbitrarily large cubes with approximately periodic arrangements of inclusions. This allows us to construct approximate eigenfunctions explicitly by starting from the generalised eigenfunctions of an operator with constant coefficients that represents the macroscopic component of the homogenisation limit for the corresponding periodic operator. In the absence of such (almost) periodic structures, one needs to analyse “arbitrary” sequences of operators A^ε, each capturing a specific point of the limit spectrum of Aε. Such sequences are obtained from Aε by shifting to the origin the areas of non-typical distribution of inclusions on which relevant quasimodes are supported (see the proof of Theorem 5.2 below). The limits of these sequences, understood in an appropriate sense, correspond to neither stochastic nor periodic homogenisation, and therefore in general have unpredictable spectral properties. Remark 5.6 It is well known that ergodicity implies that the spectrum of a self-adjoint operator with random coefficients is deterministic [43, Theorem 1.2.5]. In fact, this can be proved by an argument analogous to the first part of the proof of Proposition B.11. In particular, Sp(Aε(ω)) is deterministic almost surely, and so limSp(Aε) is also deterministic almost surely. Recalling the preceding discussion, the problem of characterising limSp(Aε) under the general ergodicity assumption, i.e. without an additional assumption on independence (such as Assumption 5.4), remains open. Remark 5.7 Similar observation can be made for Schrödinger operators with random perturbations of the potential. To give an example, let us consider a Schrödinger operator with periodic potential -Δ+V0, which may have a band-gap spectrum. Furthermore, let f be a standard non-negative (“single sight”) potential supported in the unit cube and qk(ω), k∈Zd, be i.i.d. random variables with the support given by [q-,q+]. Then for the operator perturbed by the random potential Vω:=∑k∈Zdqk(ω)f(·-k) one has Sp(-Δ+V0+Vω)=⋃q∈[q-,q+]Sp(-Δ+V0+qfper), where fper:=∑k∈Zdf(·-k), see e.g. [43]. Now define Vωε:=Vω(·/ε) and Hε:=-Δ+V0+Vωε. One can then argue that the spectrum of the corresponding homogenised operator (obtained via standard stochastic two-scale convergence techniques) is given by Hhom=-Δ+V0+⟨q0⟩∫Rdf, while the limiting spectrum is limSp(Hε)=⋃q∈[q-,q+]Sp(-Δ+V0+q∫Rdf) almost surely. The additional spectrum in this case is also coming from non-typical areas. Similar observations with explicit characterisation of the limiting spectrum can be made under the assumption of finite range of dependence for random variables qk. We did not find results of this kind in the literature, but they can be easily established using a simplified version of the arguments presented here (since the oscillations are present only in the lower order term). Properties of β∞(λ,ω) The main purpose of this subsection is to establish the deterministic nature of β∞(λ,ω), as well as its continuity and monotonicity. We begin with some obvious observations. Remark 5.8 (i)The use of cubes in the definition of ℓ(x,M,λ,ω) is not essential. In fact, one can utilize the scaled and translated versions of any sufficiently regular bounded open set—this will yield the same function β∞(λ,ω) (cf. also (27)). The reason we use cubes is that it will be convenient for our constructions in the proofs of both theorems. The use of cubes in the definition of ℓ(x,M,λ,ω) is not essential. In fact, one can utilize the scaled and translated versions of any sufficiently regular bounded open set—this will yield the same function β∞(λ,ω) (cf. also (27)). The reason we use cubes is that it will be convenient for our constructions in the proofs of both theorems. (ii)Clearly, one has β(λ)=limM→∞λ+λ21Md∫□Mb(Tyω,λ)dy=supx∈RdlimM→∞λ+λ21Md∫□xMb(Tyω,λ)dy≤β∞(λ,ω). Clearly, one has β(λ)=limM→∞λ+λ21Md∫□Mb(Tyω,λ)dy=supx∈RdlimM→∞λ+λ21Md∫□xMb(Tyω,λ)dy≤β∞(λ,ω). We first prove two auxiliary results which will be used throughout the remaining part of Sect. 5. The proof of the following lemma is straightforward and requires only part 2 of Assumption 3.6 and the bound (24). Lemma 5.9 Let ν>0 be fixed, and denote (cf. (38)) where SxM,ν is a measurable set satisfying □xM-ν⊂SxM,ν⊂□xM (clearly, the value of ℓν(x,M,λ,ω) depends on the choice of SxM,ν, but we do not reflect it in the notation). Denote also where □˚xM(ω):=∪Oωk⊂□xMOωk (i.e. the inclusions touching the boundary of the cube are removed). Then Note that ℓ~(x,M,λ,ω) is equal to ℓν(x,M,λ,ω) for an appropriate choice of SxM,ν. Lemma 5.10 Let {Mn}n∈N⊂R be an arbitrary sequence diverging to infinity. Then almost surely one has Proof For a fixed ω∈Ω, let a sequence Cn→∞ be such that and suppose that there exists a subsequence of Mn (still indexed by n) such that We take a sequence xn such that Clearly, there exists n0∈N and δ>0 such that, for every n≥n0, Next, we cover each cube □xnn with cubes of edge length Cn0. More precisely, for Nn:=⌊Mn/Cn0⌋Cn0 (⌊·⌋ denotes the integer part) the cube □xnn⊂□xnn is a union of ⌊Mn/Cn0⌋d disjoint cubes of edge length Cn0. On the one hand, observing that Nn/Mn→1 as n→∞ and applying Lemma 5.9 with ν=Cn0 we have On the other hand, since where xi′ denote the centres of the cubes in the partition described above, we conclude that which is a contradiction with (39). □ Finally, we present the main result of this subsection. Proposition 5.11 The function (λ,ω)↦β∞(λ,ω) is deterministic almost surely, continuous on R+\Sp(-ΔO) and strictly increasing on every interval contained in this set. Proof By Lemmata 5.9 and 5.10, we have The measurability of β∞(λ,·) now follows directly from the measurability of b(·,λ). It is easy to see from (28), (29) and (30) that almost surely the mapping λ↦ℓ~(x,M,λ,ω) has locally bounded (uniformly in x,M,ω) and positive derivative, Hence ℓ~(x,M,·,ω) is locally Lipschitz (uniformly in x,M,ω) and increasing on every open subinterval of its domain. More specifically, for λ1<λ2 contained in the same open interval from R+\Sp(-ΔO) one has The bound (40) implies a similar property for β∞. Indeed, let xn be a sequence such that β∞(λ1,ω)=limn→∞ℓ~(xn,n,λ1,ω). Passing to the limit along xn in (40), we arrive at Utilising a similar argument, one can prove the reverse inequality, concluding that β∞(λ,ω) is locally Lipschitz. Finally, for a fixed λ the function β∞(λ,·) is translation invariant and is thus constant almost surely. To conclude that the function β∞ is deterministic, i.e. β∞(λ,ω)=β∞(λ) almost surely, it suffices to take the set of probability one such that for every λ∈Q, λ∉Sp(-ΔO), β(λ,ω) is deterministic and use the almost sure continuity of λ↦β∞(λ,ω). □ Remark 5.12 Note that in general β∞(λ) is not necessarily differentiable, as can be seen from examples provided at the end of this section. Proof of Theorem 5.2 Let λ be a limit point of the spectra of Aε, i.e. λ=limε→0λε, λε∈Sp(Aε). Without loss of generality we may assume that The task at hand is to show that Since λε∈SpAε, there exists a sequence uε∈Dom(Aε), ‖uε‖L2(Rd)=1, such that with ‖fε‖L2(Rd)=:δε→0. (Such a sequence can be extracted by the diagonalisation procedure from Weyl sequences corresponding to λε for each ε.) Multiplying (43) by uε and integrating by parts, we obtain the estimate with some C>0. The sequence (uε)ε converges weakly to zero in L2(Rd) (otherwise its weak stochastic two-scale limit would be an eigenfunction of Ahom, which is impossible due to Proposition 4.13). In what follows we devise a “compactification” procedure for (uε)ε modifying it in such a way that the new sequence has a non-vanishing limit, thus allowing us to retrieve information about the relation of λ to the set G. We fix an arbitrary L>0 and cover Rd by the cubes □ξL, ξ∈LZd. A key ingredient of the “compactification” is the following obvious assertion. Lemma 5.13 Let (possibly finite) sequences aξ,bξ,cξ,dξ,ξ∈I, where I is a countable or finite set of indices, be such that Then there exists ζ∈I such that bζ+cζ+dζ≤3aζ. For each ε we apply the above lemma to the sequences ξ∈LZd, ‖uε‖L2(□ξL)≠0. Taking into account (44), we infer that for each ε there exists ξε such that Now for every ε we shift the cube □εL to the origin and re-normalise uε. Namely, we define Note that We denote by Aˇε the operator obtained from Aε by shifting its coefficients: Aˇε(x,ω):=Aε(x+ξε,ω). Analogously, Sˇ0ε,Sˇ1ε, and χˇ0ε,χˇ1ε, denote appropriately shifted set of inclusions, its complement, and the corresponding characteristic functions. Furthermore, it is convenient to denote by Oˇω and Oˇωk appropriately shifted unscaled set of inclusions and individual unscaled inclusions (which depend on ε via ξε, but we omit this dependence from the notation for brevity). From (45) we immediately have and where In order to verify (42), we need to analyse the behaviour of the sequence wLε on the soft component. To this end, we consider the decomposition where w~Lε, representing the macroscopic part, denotes the harmonic extension of wLε|Sˇ1ε to the whole Rd, as per Theorem 3.8 and a standard scaling argument, and the term captures the micro-resonant behaviour of the inclusions. Further on we will use analogous decompositions on several occasions. Since w~Lε is harmonic in the inclusions, we have The following lemma is trivial and follows from (13) by a rescaling argument. Lemma 5.14 The spectrum of the self-adjoint operator -ε2Δ on Sˇ0ε with Dirichlet boundary conditions coincides with the spectrum of -ΔO almost surely. By the above lemma, for sufficiently small ε we have (recall the notation (23)). Furthermore, using zε as a test function in (53) and taking into account (54), it is not difficult to see that Utilising Theorem 3.8 and the bounds (48), (49), (55), we infer that Next, by (48) and (54) we have Combining (41), (43) and (47), we infer that, for small enough ε, From (57) and the first bound in (56) we have that, up to a subsequence, Let ηL∈C0∞(□2L),L>0, be a family of cut-off functions satisfying 0≤ηL≤1, ηL|□3L/2=1, |∇ηL|≤C/L, with C independent of L. Using w~LεηL as a test function in (50) and integrating by parts, we get We estimate all the terms but one via (48) and (56) as follows: It remains to analyse the behaviour of the first term on the right-hand side of (59). We use the family of local averaging operators Pε on L2(Rd) defined in Lemma E.2, taking for Xkε the inclusions εOˇωk⊂Sˇ0ε and considering the decomposition where z^ε,z˚ε∈W01,2(Sˇ0ε) satisfy First, it is easy to see that Furthermore, it follows from Lemmata 5.14 and E.2 that for sufficiently small ε one has From the bound (64), (58), Lemma E.2, and the bound (48) we infer that the right-hand side of (62) vanishes in the limit as ε→0. Then, applying Lemma 5.14 again, we infer that Next, denote gλε(x):=b(Tx/ε+ξεω,λ). It follows from (24) that the sequence (gλε)ε is bounded in L2(□2L) and thus converges, up to a subsequence, weakly in L2(□2L) as ε→0 to some gλ. Integrating λ+λ2gλε over an arbitrary fixed cube contained in □2L, passing to the limit, and using the definition of β∞(λ), it is not difficult to see that Using Lemma E.2, we obtain from (61),(63) and (65) that Finally, the strong convergence of w~Lε, the relations (60), and a passage to the limit in (59) as ε→0 yield where in the last inequality we used (57) and (58). Since L is arbitrary, noting that all constants in the bounds obtained in the proof are independent of L, we conclude that β∞(λ)≥0. Existence of approximately periodic cubes In this subsection we present preparatory constructions and results necessary for the proof of Theorem 5.5. By P# we denote the push-forward of the probability measure given by the map HK on the σ-algebra FH,K (see Sect. 5.1). We use the following notation for a closed Hausdorff ball of radius r around a fixed compact set K′⊂K: Next lemma is the key assertion that implies the existence of “approximately periodic” arrangements of inclusions. Lemma 5.15 Let K⊂Rd be a compact set. There exists a subset ΩK⊂Ω of probability one such that for every ω∈ΩK and every r>0 one has Remark 5.16 One can rewrite (66) as P{ω~:dH(HK(ω),HK(ω~))0. Proof It is sufficient to prove that the set where the inequality (66) does not hold for any r>0 is a set of probability zero. Recall that the Hausdorff topology on the compact subsets of K is compact and thus separable. Consider a countable family {Km}m∈N of compact subsets of K that is dense in this topology and let Additionally, we set rm=-∞ if K is empty. Note that by the continuity of probability measure we also have P#BH,K(Km,rm)=0. Let ω be such that there exists r>0 with P#BH,K(HK(ω),r)=0. By the density of the family {Km}m, there exists Ki and r′>0 satisfying HK(ω)∈BH,K(Ki,r′)⊂BH,K(HK(ω),r). As P#BH,K(Ki,r′)=0, it follows from (67) that The right-hand side of the last equality is clearly a set of probability zero. □ Corollary 5.17 There exists a subset Ω1⊂Ω of probability one such that for every ω∈Ω1 one has Proof The equality follows from a straightforward translation argument. The existence of Ω1 follows from Lemma 5.15 and a simple observation that the intersection of a countable family of sets of probability one is a set of probability one. □ The following two theorems contain the main result of the present subsection. Though the notation is somewhat involved, their meaning can be expressed as follows: for a fixed λ, there is an arbitrarily large cube that can be divided into smaller sub-cubes with approximately periodic arrangements of inclusions (apart from a fixed size boundary layer) such that on each sub-cube the associated quantity ℓ(…,ω,λ) approximates β∞(λ). Theorem 5.18 Let ω^ be a typical element of Ω for which the statement of Corollary 5.17 is satisfied. Under Assumptions 5.4 there exists a set Ω1⊂Ω of probability one (which in general depends on the choice of ω^) such that for any ω∈Ω1, N,M∈N, x∈Rd, δ>0, and the cube □xM there exist Nd cubes □xiM, i=1,⋯,Nd, such that the cubes □xiM+κ (discarding the boundary) are mutually disjoint, their union is a cube ∪i□xiM+κ=:□x∗N(M+κ), and the following estimate holds: Proof It is direct consequence of Corollary 5.17, Assumption 5.4, and the law of large numbers that for any N,M,n∈N, x∈Qd, for almost every ω∈Ω there exist Nd cubes □xiM, i=1,⋯,Nd, such that ∪i□xiM+κ=□x∗N(M+κ) for a suitable x∗∈Rd, and We denote the set of such ω by ΩN,M,n,x(ω^) and set Clearly, Ω1(ω^) has probability one, and the statement of the theorem holds for any x∈Qd, but then, by the density argument it holds for any x∈Rd. □ Theorem 5.19 Under Assumption 5.4, there exists a set Ω1⊂Ω of probability one such that for any λ∈R\Sp(-ΔO), ω∈Ω1, δ>0, N,M0∈N there exist Nd cubes □xiM, i=1,⋯,Nd, N∋M>M0, as described in Theorem 5.18, such that Proof Let us fix λ∈Q\Sp(-ΔO). Let us fix a typical ω∈^Ω for which the statements of Lemma 5.10, Proposition 5.11 and Corollary 5.17 hold. Then from these assertions and Lemma 5.9 it follows that for all sufficiently large M∈N there exist cubes □x¯MM such that Similarly to Lemma 5.9, it is not difficult to show, via part 2 of Assumption 3.6 and the bound (24), that for all sufficiently large M and any x one has By Theorem 5.18, there exists a set of full measure Ω˚=Ω˚(ω^)⊂Ω satisfying the following: for any ω∈Ω˚ there is a sequence of cubes □xn∗N(M+κ), n∈N, such that for each n the cube □xn∗N(M+κ) is a union of mutually disjoint cubes □xinM+κ, i=1,⋯,Nd, and It is clear, by direct inspection, that the difference between ℓ~(x¯M,M,λ,ω^) and ℓ~(xin,M,λ,ω) is controlled (uniformly in M) by the maximal difference between the values of ∫Rdb~(Tx,λ,·) corresponding to the “matching” inclusions in □x¯MM and □xinM. Hence, by Lemma E.3, for sufficiently large (fixed) n one has Combining inequalities (70), (71) and (72), we obtain (69) (with xi:=xin) for each λ∈Q\Sp(-ΔO) on a set of probability one, which, in general, depends on λ (as well as on ω^, although the latter is of no importance). We define Ω1 as the intersection of these sets for all λ∈Q\Sp(-ΔO). Then by continuity of β∞(λ) we conclude that the assertion holds for all λ∈R\Sp(-ΔO). □ Proof of Theorem 5.5 We begin by outlining the main idea of the proof. In view of (11) and Theorem 4.3, we only need to consider the case when λ is such that β∞(λ)≥0. We fix ω∈Ω^ as in Theorem 5.19. Our goal is to construct an approximate solution to the spectral problem for the operator Aε. It is sufficient for this approximate “eigenfunction” to be from the operator form domain and satisfy the equation in a certain weak sense. Lifting it up to the domain of Aε can then be carried out via a simple abstract argument, see Lemma E.1. One of the key technical ingredients of our construction is the use of a periodic corrector on nearly periodic geometries. More specifically, one can always find a cube □x¯MM such that the corresponding local spectral average ℓ(x¯M,M,λ,ω^), cf. Remark 5.1, approximates β∞(λ) with a given error. According to Theorem 5.19, for any ω∈Ω1(ω^) one can find an arbitrarily large cube □x∗N(M+κ) tiled by the cubes □xiM, i=1,⋯,Nd, separated by the correlation distance κ from Assumption 5.4, such that □xiM\Oω are almost exact copies of the set □x¯MM\Oω^. Moreover, the local spectral average ℓ on the large cube □x∗N(M+κ) is also close to β∞(λ). This approximately periodic structure of the composite in □x∗N(M+κ) is important, for one then can approximate the “macroscopic” part of the the operator Aε on □εx∗εN(M+κ) by an operator with constant coefficients (obtained by homogenisation on a perfectly periodic structure) and use this as the basis for the construction of a quasimode, see (77) below. We emphasise that the set of full measure Ω1, for which the statement of Theorem 5.5 holds, depends in general on the choice of ω^ (note that Ω1, in principle, does not have to contain ω^). In what follows we introduce a number objects derived from ω^, such as the periodic correctors N^j,j=1,⋯,d, and the matrix of homogenised coefficients A^1hom, while other quantities, e.g. bε, correspond to ω∈Ω1. It is important to keep account of these dependencies. Henceforth the parameters L and M are assumed sufficiently large, and ε and δ are sufficiently small. The constants that appear in the subsequent bounds may depend on λ but are independent of ε, δ, L and M. We will emphasise this through the notation in two key bounds below. We fix δ>0 and L>0 and choose M so that (71) holds and (70) is satisfied for some cube □x¯MM (by Lemmata 5.9 and 5.10 ). Let N=N(ε) be the smallest integer such that εN(M+κ)≥L. By Theorems 5.18 and 5.19 there exist cubes □xiM+κ, i=1,⋯,Nd, (as described in Theorem 5.18) satisfying (68) and (69). In particular, their union is □x∗N(M+κ) and Clearly, the choice of the cubes (equivalently, their centres xi,x∗) depends on L,M,δ and ε, which we omit in the notation for brevity. Denote by □^MM+κ,1 the set obtained from □MM+κ by removing all the sets Oω^k whose closures are contained in □MM, Let N^j∈Wper1,2(□^MM+κ,1),j=1,⋯,d, be the solutions to the periodic corrector problems for the perforated cube □^MM+κ,1: We assume that each N^j is extended inside the inclusions in □MM+κ according to Theorem 3.8, has zero mean over □MM+κ, and extended by periodicity to the whole of Rd. Observe the following estimates with the constant depending only on A1 and the extension constant Cext, The first easily follows from the identity and the second from the Poincaré inequality (recall that M is sufficiently large, in particular M≥κ). An essential component of the construction is the higher (than L2) regularity of the correctors N^j. The proof closely follows the argument of [8] and is based on the use of special versions of two well-known results: the Poincaré-Sobolev inequality and the reverse Hölder’s inequality. In particular, the uniform scalable version of Poincaré-Sobolev inequality for perforated domains is valid under the minimal smoothness assumption. Theorem 5.20 (Higher regularity of the periodic corrector) Under Assumption 3.6 there exist p>2 and C>0 such that for a.e. ω one has uniformly in M. (Here denotes the average value.) We provide the proof of the theorem in Appendix C. As a corollary of the theorem and the bound (74), we have the bound We denote by A^1hom the matrix of homogenised coefficients associated with N^j, Take some k such that A^1homk·k=β∞(λ). Then u(x):=Re(eik·x) satisfies We define Nj∈Wper1,2(□1M+κ) as and extend it periodically to Rd. Denote by χ^1ε the characteristic function of the set □^MM+κ,1-x¯M+x1 extended periodically to Rd. Let η∈C0∞(□) be a cut-off function satisfying 0≤η≤1, η|□1/2=1, and for each L>0 define ηL by setting We have, in particular, that ηL|□L=1, |∇ηL|≤C/L. Multiplying u by ηL and normalising the resulting expression, we obtain a standard Weyl sequence for the operator -∇·A^1hom∇ (with respect to letting L→∞): Note also that Let b be the solution to (15) and denote by bε its ε-realisation, bε(x):=b(Tx/εω). We define uLε:=(1+λbε)uL. It is not difficult to see that the norm of uLε is bounded from below uniformly in ε and L, We conclude the construction of the approximate solution by introducing the corrector term We estimate the corrector term via (74)and (79), In what follows we will often use the bounds (74) and (79) without mentioning. Remark 5.21 The technical details in what follows build up to the bound (100). The general scheme for this argument is adopted from [32] and is useful in settings where one needs to establish proximity to the spectrum. We will resort to it once again in the proof of Theorem 6.6 below. Denote by aε the bilinear form associated with the operator Aε+1, We substitute uLCε into the form with an arbitrary test function v∈W1,2(Rd), To proceed with the argument we use the decomposition where v~ε∈W1,2(Rd) is the extension of v|S1ε into the set of inclusions S0ε by Theorem 3.8, and v0ε=v-v~ε∈W01,2(S0ε). The next bounds are straightforward consequence of Theorem 3.8 and the Poincaré inequality: We begin by analysing the first term on the right-hand side of (81): We first estimate the last term on the right-hand side of the latter, as follows: Next, we estimate the term containing Lemma 5.22 For sufficiently large M, δ<ρ/2, and any p∈[1,∞), one has Proof The proof consists in establishing the following straightforward facts for sufficiently small δ>0. For two bounded open connected (ρ,N,γ) minimally smooth sets U1 and U2, such that their complements are also connected, the relation dH(U1,U2)≤δ implies dH(∂U1,∂U2)≤δ. For two bounded open connected (ρ,N,γ) minimally smooth sets U1 and U2, such that their complements are also connected, the relation dH(U1,U2)≤δ implies dH(∂U1,∂U2)≤δ. The previous statement immediately implies that the symmetric difference (U1\U2)∪(U2\U1) is a subset of the δ-neighbourhood of the boundary of either of the sets. By covering the sets U1, U2 with a sufficiently fine lattice, one can easily obtain the bound |(U1\U2)∪(U2\U1)|≤Cδ, where the constant depends only on the minimal smoothness parameters and the diameter of the sets U1, U2. The previous statement immediately implies that the symmetric difference (U1\U2)∪(U2\U1) is a subset of the δ-neighbourhood of the boundary of either of the sets. By covering the sets U1, U2 with a sufficiently fine lattice, one can easily obtain the bound |(U1\U2)∪(U2\U1)|≤Cδ, where the constant depends only on the minimal smoothness parameters and the diameter of the sets U1, U2. These observations allow us to estimate the difference between χ1ε and χ^1ε in every cube □εxiεM via (68). Finally, the volume of the “boundary layers” □εxiε(M+κ)\□εxiεM relative to the volume of □εx∗L can be made arbitrarily small by choosing M sufficiently large. □ Applying Lemma 5.22, we have For the next estimate we invoke Theorem 5.20. Applying Hölder’s inequality twice and taking into account (76) and (86), we obtain Finally, we deal with the last two terms on the right-hand side of (85). The vector field is solenoidal, i.e. ∫□1M+κgj·∇φ=0 for all φ∈Wper1,2(□1M+κ), and has zero mean, by the definition of the homogenised matrix A^1hom. It is well known, see an explicit construction via Fourier series in [30, Section 1.1], that such field can be represented as the “divergence” of a skew-symmetric zero-mean field Gj, Gikj=-Gkij, Gikj∈Wper1,2(□1M+κ): which is understood in the sense that (gj)k=∂iGikj. Note that the fields Gj are commonly referred to in homogenisation theory as flux correctors. It is not difficult to see (by estimating the coefficients of the said Fourier series similarly to the proof of Lemma D.6) that where the second inequality is obtained from (74). Remark 5.23 We would like to warn the reader against confusing the flux correctors Gj, which are periodic, with the flux correctors Gjε constructed in Corollary D.5 and used in Sect. 6, which are stochastic by nature. Applying (90) (note that gj(·/ε)=ε∇·Gj(·/ε)) and integrating by parts yields where the colon denotes the Frobenius inner product of matrices. The first term in the last integral is identically zero due to the anti-symmetry of Gj. Therefore, combining (85), (87), (88), (91) and the last identity, we obtain We proceed by considering the second integral in (81). Decomposing both uLCε and v and integrating by parts one of the resulting terms, yields Since b is the solution to (15), its ε-realisation satisfies -ε2Δbε=λbε+1, thus we have The bound for the remaining terms of (94) via (74), (79), (83), (84) and Lemma D.8 is straightforward: Combining (81)–(85), (87), (88), (93)–(95) and the last bound yields where the remainder Rε satisfies Replacing v0ε with v-v~ε in (96) and recalling the definition of uLε, we rewrite (96) as follows: It remains to estimate the second term under the integral sign. Consider the piece-wise averaging operator MM+κε:L2(ε□x∗N(M+κ))→L2(ε□x∗N(M+κ)) defined by The following lemma is proved by a straightforward application of the Poincaré inequality. Lemma 5.24 Let f∈W1,2(ε□x∗N(M+κ)), then Now note that by (69) on every cube ε□xiM+κ we have Therefore, Decomposing uLv~ε=uLv~ε-MM+κε(uLv~ε)+MM+κε(uLv~ε), applying Lemma 5.24 and observing that ‖λ+λ2bε‖L2(□εx∗L)≤CLd/2, cf. (24), we obtain Combining (78), (80), (84), (97)–(99) yields where Now, resorting to Lemma E.1 we see that for sufficiently small δ and large L one has Since δ and L are arbitrary, we conclude that Remark 5.25 In the present work, in order to simplify the exposition of the main ideas, we assume that the coefficients of Aε are constant in the stiff component and are a multiple of the identity matrix in the inclusions. The coefficients Aε(·,ω) can be described as follows. Consider a sequence of random variables of the form where A1 is a positive definite matrix. Then the coefficients defined by (5) are simply the ε-realisation of Aε(ω): (Note that the operator Aε may be formally considered as the ε-realization of the operator -∇ω·Aε(ω)∇ω.) Similarly, we could analyse a more general problem where the coefficients in the constitutive parts of the composite are not assumed to be constant, namely, by setting where A0,A1∈L∞(Ω;Rd×d) are uniformly coercive. (One can look at an interesting simpler case where one does not change the geometry of the inclusions, but only the coefficients on the soft component.) We believe that our analysis can be easily adapted to this more general setting. It will manifest, in particular, in the need to analyse operator -∇ω·A0∇ω with the domain being a suitable subspace of W01,2(O) instead of the operator -ΔO. In this case one will get analogous results; in particular, the spectrum of Ahom will be characterised by a suitable version of the β-function, and the limit set G by a suitable version of β∞(λ) (cf. (14) and (37)). Moreover, the notions introduced and the results presented in Sect. 6 can also be easily adapted to this more general setting. Examples In general, constructing explicitly a probability space that would provide a “truly random” distribution of inclusions is a challenging task. We consider two general setups: inclusions randomly placed at the nodes of a periodic lattice and the random parking model. We begin with the former. Probability space setup. Let {ω~j}j∈Zd be a sequence of independent and identically distributed random vectors taking values in N0l×[r1,r2], where N0l:={0,1,…,l} and 01, r1=r2=1 (again, second component of ω~ being redundant) and assume that P{ω~0=0}=0. Since all ω~j are identically distributed, this implies that the first component of each ω~j,j∈Zd, takes value in N0l\{0}. By the law of large numbers, for each k∈N0l\{0} there exist arbitrarily large cubes containing only the shapes Yk at the lattice nodes. We have In particular, Remark 5.27 It was shown in [16] that Here, νjk and φjk denote the eigenvalues and orthonormalised eigenfunctions of -ΔYk. Randomly scaled inclusions. Take l=1 and 00, the set {ω~0∈[r-δ,r+δ]} has positive probability. It follows that for any δ>0 there exist arbitrarily large cubes containing only inclusions whose scaling parameter belongs to the interval [r-δ,r+δ]. Therefore, where νj are the eigenvalues of -ΔY1. It follows that Remark 5.28 It was shown in [16] that Here νj,r and φj,r are the eigenvalues and orthonormalised eigenfunctions of -ΔrY1. They can be obtained from those of -ΔY1 by scaling; in particular, νj,r=r-2νj. Modifications of the above examples. It is not difficult to construct explicit modifications of the above examples in which the inclusions are randomly rotated and shifted within the lattice cells (still respecting the minimal distance between them in order for the extension property to hold). As can be seen from the definition of β∞(λ) and the formula (27) for β(λ), this additional degree of freedom does not affect these functions (but only A1hom), hence both the limit set G and Sp(Ahom) (in the whole space setting) remain the same. Another possible modification is to assume that random variables {ω~j}j∈Zd have finite correlation distance. In this case, in order to be able to recover the above formulae for β∞(λ), arbitrarily large cubes containing only the inclusions of a specific type (as in the above examples) should have positive probability. If this is not the case, then only the general formula (37) is available. Random parking model. In this example we consider the random parking model described in [41] (see also [25]). The intuitive description of the model is as follows: copies of a set V arrive sequentially at random without overlapping until jamming occurs. We start by briefly recalling the graph construction of the random parking measure (RPM) in Rd. Let P be a homogeneous Poisson process with intensity one in Rd×R+, and let V⊂Rd be an open bounded set that contains the origin. An oriented graph is a special kind of directed graph in which there is no pair of vertices {x,y} for which both (x, y) and (y, x) are included as directed edges. We shall say that x is a parent of y and y is an offspring of x if there is an oriented edge from x to y. By a root of an oriented graph we mean a vertex with no parent. The construction of RPM goes as follows. Make the points of the Poisson process P on Rd×R+ into the vertices of an infinite oriented graph, denoted by Γ, by putting in an oriented edge (X,T)→(X′,T′) whenever (X′+V)∩(X+V)≠∅ and T0, consider the periodic lattice (2-4δ)Zd, denoting its points by ξi,i∈N, and consider the balls of radius δ centred at ξi,i∈N. For an arbitrary T>0 there is a positive probability of P having exactly one point (Xi,Ti) in each of the sets Bδ(ξi)×(0,T)⊂□M+4×(0,T) and no other points in the set □M+4×(0,T). Then, by the above construction, the random parking measure has exactly one point Xi inside each ball Bδ(ξi) with ξi∈□M and no other points inside □M, as no more unit cubes could fit between the cubes □Xi1. By the law of large numbers, with probability one there exist arbitrarily large cubes □xM containing only approximately periodically positioned inclusions Y1+Xi, Xi∈Bδ(ξi), ξi∈(2-4δ)Zd. Since M and δ are arbitrary, we conclude that Making an analogous construction with the periodic lattice (1+4δ)Zd, δ>0, we can see that On the other hand, since 1-periodic and 2-periodic distribution provide the greatest and the smallest possible inclusion density, it follows that equalities in (104) and (105) are attained when ∫Y1b(·,λ)<0 and ∫Y1b(·,λ)≥0, respectively. Thus Remark 5.30 The random parking model does not satisfy Assumption 5.4 (concerning finite range of dependence), and therefore we cannot apply Theorem 5.5 to it straight away. Nevertheless, the preceding argument shows that one can “recover” the values of β∞(λ) on arbitrarily large non-typical spacial regions (with approximately periodic positioning of inclusions), and then follow the argument of the proof of Theorem 5.5 in order to show that limSp(Aε)⊃G. Remark 5.31 As in the previous examples, one can elaborate the random parking model example further by allowing inclusions of different shapes, sizes, randomly rotated, through the use of the marked point processes framework, see [28]. Remark 5.32 The random parking model is a more realistic (from the point of view of applications) model of random distribution of inclusions. On one hand, it does not allow arbitrarily large inclusion-free regions. (More precisely, there exists a radius r such that any ball Br(x),x∈Rd, contains at least one inclusion.) On the other hand, the inclusions are not too close to each other, thus satisfying Assumption 3.6. If one only uses the Poisson point process in the construction, then one ends up with arbitrarily large inclusion-free regions and at the same time having arbitrarily many overlapping inclusions. The problem of overlapping inclusions can also be dealt with by using a Matérn modification of Poisson process (see e.g. [28]) which is constructed from a given point process by simultaneously erasing all points with the distance to the nearest neighbour smaller than a given constant. It can be shown that if the original process is stationary (ergodic), the resulting hardcore process is stationary (ergodic). If one starts from a Poisson process and makes a Matérn modification, this gives a point process which can be further used to construct random inclusions satisfying Assumption 3.6. However, the limit spectrum in this case would coincide with R0+, due to the existence of arbitrarily large areas inclusion-free regions. (More precisely, in this case β∞(λ)≥λ for all λ≥0.) Periodic inclusions randomly shifted and rotated within their cells. Let Y1⊂[0,1)d be a reference inclusion, as described above. In each cell of the lattice Zd we put a randomly rotated and randomly positioned (within the cell) copy of Y1. We do this in such a way that the minimal distance between the inclusions is positive. It is not difficult to describe the corresponding probability space explicitly similarly to how it was done in Sect. 5.6.1. Clearly, in this case β=β∞, and thus G=limSp(Aε)=Sp(Ahom). Relevant and Irrelevant Limiting Spectrum In this section we characterise the limiting spectrum of Aε in the gaps of Sp(Ahom). As we have seen in the previous section, in general the limit set G of Sp(Aε) is larger than Sp(Ahom). On the other hand, in the bounded domain setting we have the Hausdorff convergence of the spectra, i.e. the gaps of Sp(Ahom) are free from the spectrum of Aε in the limit. These facts indicate that the elements of the Weyl sequences for Aε corresponding to λ∉Sp(Ahom) are supported further and further away from the origin (as ε→0) and could not be contained in a bounded domain. Another observation that can be drawn from Sect. 5 is that the elements of such Weyl sequences are concentrated in the areas of space with non-typical distribution of inclusions. Both observations may be interpreted in the following way: in applications, where one deals with finite distances (finite-size medium), manufactured composites with relatively uniform distribution of inclusions, or only the bulk waves (characterised by β(λ) in the homogenisation limit) matter, the aforementioned part of the spectrum may not be relevant. In this lies our motivation to distinguish the two types of the spectrum and refer to them as relevant and irrelevant. However, it does not seem feasible to precisely demarcate the border between these two parts for finite ε: indeed, the notion of Weyl sequences “supported further and further away from the origin” is not easily quantifiable. Instead, we provide such characterisation in the limit as ε→0. In Definition 6.1 below we introduce the notions of relevant and irrelevant limiting spectra. Remarkably, but not surprisingly given the above discussion, R-limSp(Aε) coincides with Sp(Ahom). This constitutes the main result of this section, formulated in Theorem 6.6. A number of auxiliary constructions and results required in the proof of the theorem 6.6 are presented in Appendix D. Recall that the notation E(-∞,λ]ε is used for the spectral projections of the operators Aε. Obviously, one has and By P□L we denote the operator that to each u∈L2(Rd) assigns its restriction to the cube □L (extended by zero outside of □L whenever necessary). Definition 6.1 We say that λ belongs to the relevant limiting spectrum of Aε if and denote the set of such values of λ by We define the irrelevant limiting spectrum as Remark 6.2 The definition of the relevant/irrelevant limiting spectra can be generalised and applied in other settings, not necessarily stochastic. The definition of the relevant/irrelevant limiting spectra can be generalised and applied in other settings, not necessarily stochastic. In the stochastic case the relevant and irrelevant limiting spectra are deterministic almost surely due to translation invariance. In the stochastic case the relevant and irrelevant limiting spectra are deterministic almost surely due to translation invariance. Clearly, R-limSp(Aε)⊂limSp(Aε). Clearly, R-limSp(Aε)⊂limSp(Aε). Note that the sequence ‖P□LE[λ-δ,λ+δ]ε‖L2→L2>0 is increasing in L and δ. Therefore, we can replace lim supδ→0 and lim supL→∞ in (106) with limδ→0 and limL→∞, respectively. From this it is easy to see that the relevant limiting spectrum R-limSp(Aε) is a closed set. Note that the sequence ‖P□LE[λ-δ,λ+δ]ε‖L2→L2>0 is increasing in L and δ. Therefore, we can replace lim supδ→0 and lim supL→∞ in (106) with limδ→0 and limL→∞, respectively. From this it is easy to see that the relevant limiting spectrum R-limSp(Aε) is a closed set. Remark 6.3 It is not difficult to see that λ∈R-limSp(Aε) if and only if there exists a family of functions ψLε∈Dom(Aε), ‖ψLε‖L2(Rd)=1, L∈N, ε>0, satisfying and We will use this observation as an equivalent definition of relevant limiting spectrum in order to prove the main result of the section. We next discuss the relation of the relevant and irrelevant limiting spectra to the spectrum of the limit operator. Consider a family of self-adjoint operators Tε converging to T in the sense of resolvent convergence in variable spaces (such as in Theorem 3.10), cf. [39]. Assume that Definition 6.1 makes sense for Tε, i.e. the projection P□L is well defined. One may naturally expect that the inclusion R-limSp(Tε)⊃Sp(T) takes place. Indeed, let u, ‖u‖=1, be an approximate eigenfunction of the limit operator, i.e. ‖(T-λ)u‖≤δ, where δ can be made arbitrary small. Then if we can guarantee the existence of a sequence of approximate eigenfunctions uε of Tε (for the same λ) converging to u, then we automatically get that λ is in the relevant limiting spectrum of Tε. The converse statement, i.e. that R-limSp(Tε)⊂Sp(T) is not always true, and requires some additional compactness properties of the family Tε, cf. Proposition 6.4 and Remark 6.5 below. In the next subsection we show that in our setting the relevant limiting spectrum coincides with the spectrum of the limit operator Ahom. In the first four examples we provided in Sect. 5.6, there exists a non-empty set of λ such that β∞(λ)≥0>β(λ), implying that the limit spectrum is strictly larger than the spectrum of Ahom. Together with Theorem 6.6 below, this implies that the irrelevant limiting spectrum is non-empty. This (rather unsurprising) observation is in stark contrast with the periodic case, as the following assertion suggests. Proposition 6.4 Let Y0⊂□ be a reference soft inclusion and Y1:=□\Y0. Denote by Y0# and Y1# the sets obtained by extending Y0 and Y1 periodically to the whole of Rd. Consider the self-adjoint operator in L2(Rd) defined by the differential expression where A0,A1∈L∞(Rd;Rd×d) are periodic uniformly positive matrix valued functions. Then Proof To prove the result, one can follow the argument of Theorem 5.2, leading to the formula (46), and then use the obtained family of approximate eigenfunctions to construct a sequence as in Remark 6.3. We, however, will use a more direct approach utilising directly the periodicity. Consider a sequence λε∈Sp(Aε) converging to some λ. As is well known from the Floquet-Bloch theory, for every λε there exists a quasi-periodic function uε∈Hloc1(Rd) solving the equation in the distributional sense. Let us fix some (large) L>0. We normalise uε by requiring that One also has (again, by the Floquet-Bloch theory) the energy bound where the constant depends only on λ. Let us test the equation (109) with ηLv, where v∈H1(Rd) and ηL is a standard cut-off functions satisfying ηL∈C0∞(□2L), 0≤ηL≤1, ηL|□L=1, |∇ηL|≤C/L. Denoting by a#ε(·,·) the bilinear associated with the operator A#ε+1, we have Applying (110) and (111) (taking into account quasi-periodicity of uε) to the last term in the above we arrive at Applying Lemma E.1, we infer the existence of a family of functions ψLε (namely, ψLε:=(λε+1)(A#ε+1)-1ηLuε) satisfying the conditions in Remark 6.3. Thus, we conclude that λ∈R-limSp(A#ε). □ Remark 6.5 Note that in the above theorem we did not assume that the soft component Y0# is a collection of disconnected inclusions. In the case of connected Y0#, or Y0# representing a collection of infinite fibres, the limit spectrum of A#ε is larger that the spectrum of the corresponding homogenised operator A#hom. The difference limSp(A#ε)\Sp(A#hom) is attributed to all possible quasi-periodic modes supported on Y0#, see e.g. [19]. Relevant limiting spectrum Now we return to the analysis of the relevant limiting spectrum for the case of (separated) random inclusions. Theorem 6.6 The relevant limiting spectrum of Aε coincides with the spectrum of Ahom: In order to illustrate the basic idea behind Definition 6.1, we compare it to the argument presented in Theorem 5.2 (assuming that λ∉Sp(-ΔO)) via Remark 6.3. The bound (45) is sufficient for constructing from the restrictions uε|□ξεL a family of approximate eigenfunctions satisfying the condition (107), by using cut-off functions and a resolvent argument (cf. Lemma (E.1)). Most of the energy of these approximate eigenfunctions will be localised in the vicinity of the cubes □ξεL, whose location in space is random (i.e. depends on ω as well as on ε), hence the corresponding local spectral averages ℓ(ε-1ξε,ε-1L,λ,ω) (cf. Remark 5.1)) will be controlled from above only by β∞(λ). (Note that here the local spectral averages are calculated on the rescaled cubes ε-1□ξεL.) However, if we assume that for every L∈N the cubes are centred at the origin, i.e. ξε=0, as one has in Definition 6.1, then by the ergodic theorem the local spectral averages will converge to β(λ) (cf. Remark 4.8). We continue with the proof of Theorem 6.6. Proof of the inclusion R-limSp(Aε)⊂Sp(Ahom). Let λ∈R-limSp(Aε), and let ψLε be as in Remark 6.3. We denote For each L the sequences fLε and ψLε are bounded and, hence, have subsequences that converge weakly stochastically two-scale as ε→0 to some fL,ψL∈L2(Rd×Ω), respectively. In particular, By Theorem 3.10 applied to the equation (Aε+1)ψLε=(λ+1)ψLε+fLε, we infer that ψL=ψL,0+ψL,1∈V is the solution to It remains to show that ψL does not vanish as L→∞. Multiplying (112) by ψLε and integrating by parts, we easily obtain the bound where ψ~Lε is the extension of ψLε|S1ε onto Rd by Theorem 3.8. Moreover, by the same theorem we have It follows from the last two bounds we have that, up to a subsequence, ψ~Lε converges strongly in L2(K), for any bounded domain K, to some function ψ~L∈W1,2(Rd) (which a priori could be zero). Then, by the properties of stochastic two-scale convergence, cf. Proposition A.3, the convergence χ1εψ~Lε→21Ω\Oψ~L holds. On the other hand, . Thus, If λ∈Sp(-ΔO) then λ∈Sp(Ahom). Assume that λ∉Sp(-ΔO). Then, arguing in a similar way as in Sect. 5.3, cf. (52), Lemma 5.14 and (54), we infer that for small enough ε Combining this with (108), passing to the limit via (115), and taking into account (107), we infer that for large enough L, one has ‖ψL,0‖L2(□L+1)≥C>0. It follows that ‖ψL‖L2(Rd)≥C>0 and, therefore, since fL in (114) vanishes as L→∞ (cf. (113)), we conclude that λ∈Sp(Ahom). Proof of the inclusion R-limSp(Aε)⊃Sp(Ahom). Starting from a standard prototype uL of a Weyl sequence for the homogenised operator Ahom, we will construct for each element of the sequence and each ε an approximate solution uLCε to the spectral problem for the operator Aε while maintaining control of the error. In this we will follow, to a certain extent, a general scheme developed in [32]. Case λ∈Sp(Ahom)\Sp(-ΔO). A part of the argument and the initial construction we present below are similar to those contained in the proof of Theorem 5.5. Therefore, we will often refer to the relevant places in Sect. 5.5 to fill in some detail and reuse a number of formulae therein so as not be repetitive. At the same time, since we are no longer in the periodic setting of Sect. 5.5, we have no asymptotic bounds on the homogenisation correctors Njε and the terms Gjε and Bε carrying the information about the microscopic structure of the composite. Therefore, we will need to keep track of these quantities in the error bounds explicitly. By the assumption we have β(λ)≥0. Similarly to the proof of Theorem 5.5, we consider u(x):=Re(eik·x) with k such that A1homk·k=β(λ) and denote where ηL(·):=η(·/L), with the cut-off function η∈C0∞(□) satisfying 0≤η≤1, η|□1/2=1. Then uL satisfies and the bound (79). We define uLε by setting Further, let pj∈X,j=1,…,d, be the solution to the problem (7) with ξ=ej. By Lemma D.3 we can assume that pj∈Vpot2. Then for any cube □L and any ε there exists a function Njε∈W1,2(□L) such that where pjε in the ε-realisation of pj. By the ergodic theorem We define uLCε:=uLε+∂juLNjε. In what follows we will often use the bound (79) without referring to it every time. Similarly to (80), we estimate the corrector: As in the proof of Theorem 5.5, we substitute uLCε and a test function v∈W1,2(Rd) into the bilinear form associated with the operator Aε+1, namely and begin by analysing the first term on the right-hand side of (119) (recall the decomposition (82) and the bounds (83), (84)): The following bound is straightforward: In order to estimate the second term on the right-hand side of (120), we employ a similar approach as in (89)–(93). Denote by the difference of fluxes, and consider its ε-realisations By Corollary D.5, there exist skew-symmetric tensor fields Gjε∈W1,2(□L;Rd×d) (often referred to as flux correctors) such that gjε=∇·Gjε and Proceeding as in (92), we obtain Now we address the second integral on the right-hand side of (119). Analogously to (94) and by utilising the identity (95), we have We estimate the second integral on the right-hand side of (125) by employing Lemma D.8 and the bounds (83), as follows: Combining (120), (121), (124)–(126), and (84), we can rewrite the left-hand side of (119) as where the remainder term R(ε,L,λ) satisfies (cf. (117), (123)) Proceeding as in the proof of Theorem 5.5, cf. (98), and recalling the definition of β(λ), see (14), yields It remains to estimate the second term under the integral on the right-hand side of (128). Since ⟨b⟩-bε converges weakly to zero in Lloc2(Rd), by Lemma D.6 there exists a sequence of vector fields Bε∈W1,2(□L;Rd) such that ∇·Bε=⟨b⟩-bε and Integrating by parts, we obtain Combining (116), (118), (127), (129), and (130) yields where (for fixed L and λ) Now we invoke Lemma E.1: for all large enough L and small enough ε there exists ψLε∈dom(Aε) such that It follows that λ∈R-limSp(Aε) as per Remark 6.3. Case λ∈Sp(-ΔO). The argument we utilise here is similar to the one we used above but is more straightforward technically. Since R-limSp(Aε) is closed, cf. Remark 6.2 d., and due to the characterisation (13) of the spectrum of -ΔO, we can assume without loss of generality that λ∈Sp(-ΔOωk) for some k. Let φ∈W01,2(Oωk),‖φ‖L2(Oωk)=1, be the corresponding eigenfunction, We define functions φε∈W01,2(εOωk) by the formula Testing φε in the bilinear form aε(·,·) with v=v~ε+v0ε∈W1,2(Rd) (recall the decomposition (82)), and taking into account (131), we have Using the obvious identity ‖ε∇φε‖L2(εOωk)2=λ we estimate, cf. also (84), the second term in (132): In order to estimate the last term in (132), we use the Sobolev embedding for a suitable p>2: Combining (132)–(133) we obtain Arguing as at the end of the previous case, we conclude that λ∈R-limSp(Aε). Remark 6.7 Going back to the example of the Schrödinger operator in Remark 5.7, one can easily observe that the relevant limiting spectrum for the family Hε also coincides with the spectrum of the limit operator Hhom. Irrelevant limiting spectrum and semi-group convergence The next theorem has direct implications for the parabolic and hyperbolic evolution problems, see Corollary 6.9 below. It follows from Definition 6.1 that λ∈Irr-limSp(Aε) if and only if λ∈limSp(Aε) and the quantity in (106) vanishes—this property is local in the frequency domain. The property (134) below, on the other hand, is non-local in the frequency domain and cannot be obtained directly from the definition. Theorem 6.8 For any compact K⊂R\Sp(Ahom) one has almost surely Proof We assume that otherwise there is nothing to prove. Arguing by contradiction, suppose that (134) does not hold. Then there exists L and a positive constant M such that Without loss of generality, we assume that K⊂(a,b]⊂[a,b]⊂(Sp(Ahom))c. Let us fix an sufficiently small positive value and consider a partition a0=a0, where the constant C(λ) only depends on λ and dist{λ,Sp(-ΔO)}. We conclude that up to a subsequence u~ε converges weakly in W1,2(Rd) and strongly in L2(□L+1) to some u0∈W1,2(Rd) satisfying Since vε,fε are bounded in L2(Rd) they converge up to a subsequence weakly stochastically two-scale (see Proposition A.3) to v,f∈L2(Rd×Ω) respectively. Note that, since v vanishes outside Rd×O, one has It remains to pass to the limit in the equation (136) via Theorem 3.10: where the right-hand side satisfies the bound Then (137), (138) imply that which is a contradiction, since we can choose δ to be arbitrary small. □ Next we will discuss the consequences of Theorem 6.8 on parabolic and hyperbolic evolution. The first claim gives the estimate on L2-norm of the solution of parabolic problem for every time t, while the second one gives the estimate of L2-norm of hyperbolic problem in arbitrary time t with given initial velocity or initial position with additional regularity. Both of the claims imply that part of the initial condition supported in the irrelevant spectrum can be neglected on any finite domain. Corollary 6.9 Suppose that U⊂Sp(Ahom)c is closed. Then for all L,t>0, one has almost surely lim supε→0P□L∫Ue-tλdE(-∞,λ]ε(ω)L2→L2=0; lim supε→0P□L∫Ue-tλdE(-∞,λ]ε(ω)L2→L2=0; lim supε→0P□L∫Usin(tλ)λdE(-∞,λ]ε(ω)L2→L2=0; lim supε→0P□L∫Usin(tλ)λdE(-∞,λ]ε(ω)L2→L2=0; If a sequence (fε)ε>0⊂L2(Rd), ‖fε‖L2(Rd)=1, is such that limΛ→∞supε>0∫λ>Λ⟨dE(-∞,λ]ε(ω)fε,fε⟩=0, then lim supε→0P□L∫Ucos(tλ)dE(-∞,λ]ε(ω)fεL2(Rd)=0. If a sequence (fε)ε>0⊂L2(Rd), ‖fε‖L2(Rd)=1, is such that limΛ→∞supε>0∫λ>Λ⟨dE(-∞,λ]ε(ω)fε,fε⟩=0, then lim supε→0P□L∫Ucos(tλ)dE(-∞,λ]ε(ω)fεL2(Rd)=0. Proof The proof follows directly from Theorem 6.8 and from the fact that the functions e-tλ, sin(tλ)/λ, cos(tλ) are bounded on [0,+∞) and the first two vanish at infinity. □ Acknowledgements The research of Mikhail Cherdantsev was supported by the Leverhulme Trust Research Project Grant RPG-2019-240. Kirill Cherednichenko acknowledges the support of Engineering and Physical Sciences Research Council, Grants EP/L018802/2, EP/V013025/1. Igor Velčić’s work is supported by the Croatian Science Foundation under the projects number IP-2018-01-8904 and number IP-2022-10-5181, and by the NextGenerationEU framework through the project "DEEPWAVE" at the University of Zagreb Faculty of Electrical Engineering and Computing. (The views and opinions expressed are solely those of the author(s) and do not necessarily reflect those of the European Union or the European Commission. Neither the European Union nor the European Commission can be held responsible for them.) Kirill Cherednichenko and Igor Velčić’s were also supported by CONACyT CF-2019 No. 304005. The authors thank Prof. Andrey Piatnitski for useful comments. Appendices Probability Framework and Stochastic Two-Scale Convergence Let U(x):L2(Ω)→L2(Ω),x∈Rd, be the unitary group defined by The unitarity follows from the measure preserving property of the dynamical system. For each j=1,⋯,d, we denote by Dj the infinitesimal generator of the unitary group Its domain dom(Dj) is a dense linear subset of L2(Ω) and consists of f∈L2(Ω) for which the limit exists in L2(Ω). Note that iDj, j=1,⋯,d, are self-adjoint, pairwise commuting linear operators on L2(Ω). We denote ∇ω:=(D1,⋯,Dd). Furthermore, we define It is known that W∞,2(Ω) is dense in L2(Ω), the set C∞(Ω) is dense in Lp(Ω) for all p∈[1,∞) as well as in Wk,2(Ω) for all k, and W1,2(Ω) is separable (using the fact that F is countably generated). The following subspace of W1,2(Ω) plays an essential role in our analysis: Note that as a consequence of the ergodic theorem (Theorem 3.4) one has i.e. W01,2(O) consists of W1,2-functions that vanish on Ω\O. We also introduce the space It is not difficult to see that Assumption 3.6 implies that the assumptions of Lemmata 3.1 and 3.2 in [16] are satisfied, and hence the following statement holds. Lemma A.1 The space C0∞(O) is dense in L2(O) and in W01,2(O). One can equivalently define Djf as the function with the property Note that for a.e. ω∈Ω one has where the expression on the right-hand side is the distributional derivative of f(·,ω)∈Lloc2(Rd). For a random variable f∈L2(Ω), the associated random field f(x,ω)=f(Txω) is T-stationary, i.e. f(x+y,ω)=f(x,Tyω). Moreover, there is a bijection between random variables from L2(Ω) and T-stationary random fields from Lloc2(Rd,L2(Ω)). This can be extended to Sobolev spaces, where one has a higher regularity in x for realisations. Namely, the following identity holds2 (see [23, Lemma A.7] for details): Following [30], we define the spaces Lpot2(Ω) and Lsol2(Ω) of potential and solenoidal vector fields. Namely, a vector field f∈Lloc2(Rd) is called potential if it admits a representation f=∇u,u∈Wloc1,2(Rd). A vector field f∈Lloc2(Rd) is called solenoidal if A vector field f∈L2(Ω) is called potential (respectively, solenoidal), if almost all its realisations f(Txω) are potential (respectively, solenoidal) in Rd. The spaces Lpot2(Ω) and Lsol2(Ω) are closed in L2(Ω). Setting we have the following orthogonal decomposition (“Weyl’s decomposition”) Note that Vpot2 is the closure of the space {∇ωu,u∈W1,2(Ω)} in L2(Ω). We define the following notion of stochastic two-scale convergence, which is a slight variation of the definition given in [49]. We shall stay in the Hilbert setting (p=2), as it suffices for our analysis. Let S be an open Lipschitz set in Rd. Suppose that C⊂C∞(Ω) is a countable dense family of vector-functions in L2(Ω) (recall that the latter is separable) and Ωt=Ωt(C)⊂Ω is a set of probability one such that the claim of Theorem 3.4 holds for all ω∈Ωt and g∈C. Elements of Ωt are often referred to as typical, hence the subscript “t”. Definition A.2 Let {Txω}x∈Rd be a typical trajectory and (uε) a bounded sequence in L2(S). We say that (uε) weakly stochastically two-scale converges to u∈L2(S×Ω) and write if If additionally ‖uε‖L2(S)→‖u‖L2(S×Ω), we say that (uε) strongly stochastically two-scale converges to u and write uε→2u. In the next proposition we collect some properties of stochastic two-scale convergence, see [16] for the proof. Proposition A.3 Stochastic two-scale convergence has the following properties. Let (uε) be a bounded sequence in L2(S). Then there exists a subsequence (not relabelled) and u∈L2(S×Ω) such that . Let (uε) be a bounded sequence in L2(S). Then there exists a subsequence (not relabelled) and u∈L2(S×Ω) such that . If then ‖u‖L2(S×Ω)≤lim infε→0‖uε‖L2(S). If then ‖u‖L2(S×Ω)≤lim infε→0‖uε‖L2(S). If (uε)⊆L2(S) is a sequence such that uε→u in L2(S) for some u∈L2(S), then uε→2u. If (uε)⊆L2(S) is a sequence such that uε→u in L2(S) for some u∈L2(S), then uε→2u. If (vε)⊆L∞(S) is uniformly bounded by a constant and vε→v strongly in L1(S) for some v∈L∞(S), and (uε) is bounded in L2(S) with for some u∈L2(S×Ω), then . If (vε)⊆L∞(S) is uniformly bounded by a constant and vε→v strongly in L1(S) for some v∈L∞(S), and (uε) is bounded in L2(S) with for some u∈L2(S×Ω), then . Let (uε) be a bounded sequence in W1,2(S). Then on a subsequence (not relabelled) uε⇀u0 in W1,2(S), and there exists w∈L2(S,Vpot2) such that Let (uε) be a bounded sequence in W1,2(S). Then on a subsequence (not relabelled) uε⇀u0 in W1,2(S), and there exists w∈L2(S,Vpot2) such that Let (uε) be a bounded sequence in L2(S) such that ε∇uε is bounded in L2(S,Rd). Then there exists u∈L2(S,W1,2(Ω)) such that on a subsequence and Let (uε) be a bounded sequence in L2(S) such that ε∇uε is bounded in L2(S,Rd). Then there exists u∈L2(S,W1,2(Ω)) such that on a subsequence and Measurability Properties In this part of the appendix we collect a number of technical results on the measurability of various quantities being used throughout the paper. For q=(q1,⋯,qd)∈Qd we define the set Recall that we reserve the index k0 for the inclusion Oω0 containing the origin (assuming that ω∈O). Note that for a fixed ω∈O the set of points q∈Qd such that ω∈Oq is exactly Oω0∩Qd. Lemma B.1 ( [16, Lemma 5.1]) For every q∈Qd, the set Oq⊂Ω is measurable. We define the random variables Note that D~i=+∞ whenever ω∉O, and for ω∈O one has D~i=(Dω0)i. Furthermore, we denote by D=D(ω) the random vector D:=-(D~1,⋯,D~d)T-d1/4. Finally, for ω∈O we define Pω:=Oω0+D, which coincides with the definition in Sect. 4. Note that by Assumption 3.6 we have Pω⊂□. Lemma B.2 The mapping (x,ω)↦dist(x,Pω) is measurable from Rd×O to R (on Rd×O we take the product of Borel σ-algebra with F). Proof The statement follows from the representation □ Lemma B.3 The set-valued mapping H:ω↦Pω¯ is measurable, where on the closed subsets of □ we take the σ-algebra generated by the Hausdorff distance (topology) dH. Proof The Hausdorff topology on the closed subsets of a compact set is compact and thus separable. We recall that, in line with the notation introduced in Sect. 5.4, It is sufficient to prove that the set is measurable. But this can be easily seen by using Lemma B.2 and the representation □ Let {φ~l}l∈N⊂C0∞(□) be a family of functions dense in W01,2(□). Further, let ρ∈C0∞(Rd) be non-negative with ∫Rdρ=1, suppρ⊂B1, and denote ρδ(x)=δ-dρ(x/δ). We denote the characteristic function of the set Oωk0,m+D by and define a mapping from Ω taking values in W01,2(□): where ∗ stands for the usual convolution of functions. Note that for a.e. ω∈O one has suppφl,m(·,ω)⊂Pω. Lemma B.4 ([16, Lemma 5.2]) For every l,m∈N, the mapping ω↦φl,m(·,ω) taking values in W01,2(□) is measurable with respect to the Borel σ-algebra on W01,2(□). Note that by the construction we have that for a.e. ω∈Ω the family {φl,m(·,ω)}l,m∈N ⊂C0∞(Pω) is a dense subset of W01,2(Pω) (cf. Lemma A.1). Let’s introduce the sets Y[L,U]⊂O for some random variables L,U:Ω→R0+: We also define the set S[L,U]⊂W01,2(Pω) as For every r∈R and l,m∈N we define the random variable Lemma B.5 ( [16, Lemma 5.3]) Xrl,m is a measurable function for every r∈R and l,m∈N. Lemma B.6 For measurable L, U, the set Y[L,U] is measurable. Proof The claim follows by observing that Note that we use the intervals [L-1n,U+1n] in the above in order to address the case L=U. □ Lemma B.7 Let Φ:Ω→L2(Rd) and L,U:Ω→R0+ be random variables. Then the mapping is measurable. Proof The claim follows from the formula □ Recall that for ω∈O we denote by Λ1(ω)<Λ2(ω)<⋯<Λs(ω)<⋯ the eigenvalues of -ΔPω. Lemma B.8 For every s∈N the mapping Λs is measurable. Proof For s=1 we use the equality {ω:Λ1≤x}=Y[0,x],x∈R, while for s>1 one has In each case, the claim follows immediately. □ Lemma B.9 For each φ∈L2(Rd) and s∈N the mapping ω↦PS[Λs,Λs]φ, taking values in L2(Rd), where PS[Λs,Λs] is the L2-orthogonal projection on S[Λs,Λs], is measurable. Proof For every n∈N we define the random variable We also define the random variables Psn as follows: It is easy to see that for a fixed ω∈O one has PS[Λs,Λs]φ=limn→∞Psn, where the convergence on the right-hand side is in L2(Rd). □ Lemma B.10 For every ω∈O and s∈N we define Ns(ω) as the dimension of S[Λs,Λs]. Then Ns is measurable. Moreover, there exist Ψs1(ω),Ψs2(ω),⋯,Ψsn(ω),⋯, measurable, taking values in L2(Rd), such that Ψs1,⋯,Ψss is an orthonormal basis in S[Λs,Λs] (in the sense of L2(Rd)), and ΨsNs+1=ΨsNs+2=⋯=0. Proof We take the sequence φ~k and define the measurable functions Csk=PS[Λs,Λs]φ~k taking values in L2(Rd). Then we construct the sequence Ψ~sk by applying the Gramm-Schmidt orthogonalization process to the sequence Csk. Note that for every s∈N and ω∈O we have that there is at most finite number of Ψ~sk that are different from zero and they form the orthonormal basis in S[Λs,Λs]. The claim follows by rearranging the sequence Ψ~sk, taking into account that the set of all finite subsets of N is countable. □ Clearly, the sequence {Ψsp}s∈N,p=1,…,Ns in an orthonormal basis in L2(Pω). We conclude this part of the Appendix by providing a streamlined proof of [16, Corollary 4.9 and Theorem 5.6] Proposition B.11 Under Assumption 3.6 we have Proof We denote It is easy to see that the set Eig(ω) is almost surely independent of ω∈Ω, i.e. deterministic. Indeed, it is clear that for q∈Q, n∈N the set Aq,n:=ω∈Ω:(q-1n,q+1n)⊂(Eig(ω))c is translation invariant and thus is of probability zero or one. Then the set has probability one. First, we have a trivial identity valid for every ω∈Ω, Next, it is not difficult to see that the set on the left-hand side is constant for all ω∈Ω1, which proves the claim. In the rest of the proof we omit the dependence of the set Eig(ω) on ω. The inclusion (Eig)c⊂(Sp(-ΔO))c is a direct consequence of (31) and Proposition 4.11. From this we conclude that Sp(-ΔO)⊂Eig. To prove the opposite inclusion, take λ∈Eig. It is easy to see that for all δ>0 the set Iδ:={ω∈O:[λ-δ,λ+δ]∩∪s∈N{Λs(ω)}} has positive probability, since Eig is deterministic. For each l∈N we use the notation Clearly, for each δ>0 there exists l(δ)∈N such that P[λ-δ,λ+δ]l(δ)≠0, since Also from the definition it follows that almost surely we have where we used the fact that Bρ can intersect at most one inclusion in Oω, which is then necessarily contained in □4. Integrating (139) over Ω and using Fubini’s theorem, we infer that there exists C>0 such that Using the fact that δ>0 is arbitrary, we conclude that λ∈Sp(-ΔO). □ Higher Regularity of the Corrector In order to prove higher regularity of the corrector we need to recall special versions of two well-known results, namely the Poincaré-Sobolev inequality for perforated domains and the reverse Hölder’s inequality. We begin with the former. The Poincaré-Sobolev inequality provided in Lemma C.1 is valid for a more general family of perforated domains Rd\Oω than those considered in this paper. Namely, in Lemma C.1 we assume that the set Oω satisfies Assumption 3.6, but we do not make the assumption that it is generated by a dynamical system on a probability space. We still keep the notation Oω and Oωk,k∈N, for the set of inclusions and the individual inclusions, respectively. In addition, we admit sets Oω comprising a finite number of inclusions. Lemma C.1 There exist C>0 and m>1 such that for any f∈Wloc1,p(Rd\Oω), R>0, one has where p≤q≤dp/(d-p) and cR(x) is (for a given x) a constant that only depends on the values of f in BmR(x)\Oω. Proof For R>0 we introduce a rectangular box QR,γ:=(-R,R)d-1×(-γR,γR). Part 1 of Assumption 3.6 implies that there exists a number R0>0 (the largest R such that the box QR,γ is contained in the ball Bρ) such that for any inclusion Oωk and any x0∈∂Oωk there exist local coordinates y congruent to x with the origin at x0 such that the intersection of the set Rd\Oω with QR0,γ is the subgraph of a Lipschitz continuous function g with g(0)=0 and Lipschitz constant γ: Remark C.2 Obviously, the above property also holds for boxes QR,γ with 0R0/2 and, again, ∅≠BR(x)∩Oω≠BR(x). Let f~∈Wloc1,p(Rd) be the extension of f as per Theorem 3.8. We have where we use the fact that for any Oωk with Oωk∩BR(x)≠∅ the corresponding “extension set” Bωk is contained in BR+d(x) (so ). Thus (140) holds for any ball BR(x) with m=max{2m0+1,1+d/R0} and cR(x) as specified above for each case. □ We will need the following simplified version of the reverse Hölder’s inequality, see e.g. Theorem 1.10 and Remark 1.14 in [8]. Lemma C.4 Let k≥2, f∈Lq(□x0L),q>1, for some cube □x0L and assume that for some c>0 and every cube □xa contained in □x0L. Then there exists a constant ϵ=ϵ(q,d,c,k)>0 such that ∀p with q≤p0, which are assumed to be fixed for the time being. We apply Lemma C.1 with the perforated domain being given by the periodic extension of the set □M+κ,1 to the whole of Rd. Let cR(x0) be as in (140) with f replaced by N^j. Using (N^j-cR(x0))ηR2 as the test function in (73) and rearranging the terms, we obtain and hence We proceed by estimating the right-hand side of (144) term by term. For the first term, we use the Cauchy inequality: for any α>0. For the second term, applying the Cauchy inequality, (140), and then Hölder’s inequality, we obtain For the final term in (144), we first apply Hölder’s inequality with the exponents q=2d/(d-1) and p=2d/(d+1) to the integral of the product of |N^j-cR(x0)| and |∇N^j| and then the inequality (140) to the term containing |N^j-cR(x0)| with the same exponents p, q, wich yields Combining the above estimates with (144), moving the term containing α-1 to the left-hand side, choosing α large enough, and observing that □x0R/d⊂BR(x0) and B2mR(x0)⊂□x02mR, we obtain Denoting z:=|∇N^j|2d/(d+1), and dividing both sides of (145) by |□R/d|, we obtain where x0 and R>0 are arbitrary. We can now apply the reverse Hölder inequality (Lemma C.4) to infer that there exists ϵ>0 depending only on C and d such that for any μ satisfying (d+1)/d≤μ<(d+1)/d+ϵ one has Rewriting this for ∇N^j yields for all 2≤p<2+2dϵ/(d+1). Choosing □x0R/d=□M+κ, observing that the corresponding □2mR can be covered by finitely many periodically shifted copies of □M+κ, and utilising the periodicity of the corrector, we finally obtain (75). Auxiliary Results for Theorem 6.6 Denote by Hρ,N,γ(□) the family of all (ρ,N,γ) minimally smooth closed sets P⊂□1/2 whose interiors are connected and □1/2\P also connected. Furthermore, let Lpot2(□) be the space of potential vector fields on □. The direct product Lpot2(□)×Hρ,N,γ(□) is a measurable space with the product σ-algebra generated by the standard distance on L2(□;Rd) and the Hausdorff distance on Hρ,N,γ(□). For a pair (g,P)∈Lpot2(□)×Hρ,N,γ(□), let φ∈W1,2(□) be a potential of g, i.e. g=∇φ, let φ~∈W1,2(□) be the harmonic extension of φ|□\P to the whole of □, and denote g~:=∇φ~. This construction defines a mapping E~:Lpot2(□)×Hρ,N,γ(□)→Lpot2(□) by setting E~(g,P)=g~. Lemma D.1 The mapping E~ is measurable. Proof Consider a convergent sequence (gn,Pn)→(g,P) in Lpot2(□)×Hρ,N,γ(□), i.e. gn→g in L2(□;Rd) and Pn→P in the Hausdorff distance. The minimal smoothness assumption implies that (see the proof of Lemma 5.22)) We take φ,φn⊂W1,2(□),n∈N, such that g=∇φ, gn=∇φn and ∫□φ=∫□φn=0. Clearly, φn→φ in W1,2(□). By Theorem 3.8 and Remark 3.9, the harmonic extensions φ~n of φn|□\Pn to the whole of □ satisfy the uniform bound It follows that φ~n converges to some φ^∈W1,2(□) strongly in L2(□) and weakly in W1,2(□). It is not difficult to see that φ^|□\P=φ|□\P. Multiplying the equation by an arbitrary f∈C0∞(P′), where P′ is contained in the interior of P (and thus, by (146), P′ is contained in the interior of Pn for large enough n), integrating by parts and passing to the limit, we see that We infer that φ^ is harmonic in P, and, hence, φ^=φ~, where the latter denotes the harmonic extensions of φ|□\P onto □. It remains to prove the strong convergence of ∇φ~n to ∇φ~. To this end, we multiply (147) by the test function φ~n-φn and integrate by parts: Passing to the limit as n→∞ yields The last equality follows from a simple observation that ∇φn1Pn converges to ∇φ1P in L2(□) (utilising (146)). Similarly to (148), we have and thus ∇φ~n→∇φ~ in L2(□). Therefore, the mapping E~ is continuous and hence measurable. □ Proposition D.2 For every f∈Vpot2(Lpot2(Ω)) there exists an extension f~∈Vpot2(Lpot2(Ω)) of f|Ω\O onto Ω such that for a.e. realisation f~(Txω) its potential φ~∈Wloc1,2(Rd) (so ∇φ~(x)=f~(Txω)) is harmonic on the set of inclusions Oω and Proof Let f∈Lpot2(Ω). The mapping from O to Lpot2(□) that to every ω∈O assigns the vector field φω(x):=f(Tx-Dω)∈Lpot2(□) is measurable. (This can be easily checked for an arbitrary f∈L2(Ω;Rd) first by checking it for simple functions and then using the density argument.) using Lemmata B.3 and D.1, we conclude that the mapping ω↦E~(φω,Pω)∈Lpot2(□) is measurable. We define Note that thus defined f~ is measurable, since E~(φω,Pω) is harmonic and hence infinitely smooth in Pω as well as an element of Lpot2(Ω) by construction. We next prove (149). By Assumption 3.6, for a.e. ω the ball Bρ intersects at most one inclusion Oωk. At the same time, the cube □2 contains this inclusion together with its extension domain Bωk. By Theorem 3.8, one has Integrating over Ω and applying the Fubini theorem yields Note that, by the construction of f~, if φ∈Wloc1,2(Rd) is such that ∇φ(x)=f(Txω), then f~(Txω)=∇φ~(x), where φ~ is the harmonic extension of φ|Rd\Oω into the set of inclusions. It remains to show that f~∈Vpot2 if f∈Vpot2. Let φε,φ~ε∈W1,2(□) be sequences such that ∇φε(x)=f(Tx/εω) and ∇φ~ε(x)=f~(Tx/εω) in □. By the ergodic theorem, one has Using the fact that by construction φ~ε=φε on □\S0ε and integrating by parts, we obtain where n is the outward unit normal vector. Therefore, where Kε:=⋃Oωk∩∂□≠∅Oωk. It is then not difficult to see that Combining this with (151) and (150), we conclude that □ Lemma D.3 For every f∈X there exists the extension f~∈Vpot2, f~=f on Ω\O, such that for a.e. realisation f~(Txω) its potential φ~∈Wloc1,2(Rd), ∇φ~(x)=f~(Txω), is harmonic on the set of inclusions Oω, and Proof Let (fk) be a sequence in Vpot2 such that ‖f-fk‖L2(Ω\O) converges to zero. For each k let f~k∈Vpot2 be the extension of fk as in Proposition D.2. By virtue of (149), (f~k) is a Cauchy sequence and therefore it converges in L2(Ω) to some f~∈Vpot2. Clearly, one then has f~=f on Ω\O. The remaining part of the statement can be proven similarly to the argument used to show that φ^ is harmonic in P in the proof of Lemma D.1. □ In the next lemma we work with spaces of complex valued functions. Lemma D.4 Let w be a zero-mean solenoidal random vector field, w∈Vsol2(Ω). Then there exists a random tensor field Wikl,i,k,l=1,⋯,d, such that for all i and k the random field Wik·:=(Wik1,⋯,Wikn)T is an element of Vpot2, for fixed l the field Wikl is antisymmetric, i.e. Wikl=-Wkil, and wk=Wiki,k=1,…,d. (We use the usual summation convention.) Proof Since the differentiation operators iDk,k=1,⋯,d, are self-adjoint and commuting, by the spectral theorem there exists a measure space (M,μ) with a finite measure μ, a unitary operator U:L2(Ω)→L2(M), and a.e. finite real-valued functions ak, k=1,⋯,d, on M such that UiDkφ=akUφ for all φ∈L2(Ω). It is easy to see that ker∇ω consists of constants, so w⊥ker∇ω since w has mean zero. Note that f∈ker∇ω if and only if akUf=0,k=1,…,d. Therefore, the linear set U(ker∇ω)=∩kkerak consists of all L2(M) functions that vanish on M\{a=0}, where a:=(a1,⋯,ad)T. Since Uw⊥U(ker∇ω), we conclude that Furthermore, since w is solenoidal, one has and hence Since W1,2(Ω) is dense in L2(Ω), the set {Uφ:φ∈W1,2(Ω)}={ψ∈L2(M):ψa∈(L2(M))d} is dense in L2(M), and hence Define an antisymmetric in ik tensor field W~ikl with entries in L2(M) as follows (i,k,l=1,⋯,d): Note that W~iki=Uwk by (152), (153). We set Wikl:=U-1W~ikl. It only remains to show that Wik·∈Vpot2, which follows from the fact that for all φ∈Lsol2(Ω) one has a·Uφ=0 a.e. (cf. (153) with w replaced by φ): □ Corollary D.5 Let gjε, j=1,⋯,d, be as in (122). There exist skew-symmetric tensor fields Gjε∈W1,2(□L;Rd×d), (Gjε)ik=-(Gjε)ki, such that i.e. (gjε)k=∂i(Gjε)ik for all k=1,⋯,d, and Proof Applying Lemma D.4 to gj yields gj=Wiki, where we drop the index j on the right-hand side for simplicity. Since Wik·∈Vpot2 for all i, k, there exists a zero-mean function (Gjε)ik∈W1,2(□L) such that Wik·(Tx/εω)=∇(Gjε)ik(x) (i.e. Wikl(Tx/εω)=∂l(Gjε)ik). Obviously, Gjε is skew-symmetric and (154) holds. The convergence (155) follows from two observations: a) since ⟨Wikl⟩=0, the sequence ∂l(Gjε)ik converges to zero weakly in L2(□L) by the ergodic theorem; b) functions (Gjε)ik have zero mean over □L. □ Lemma D.6 Let K be a bounded domain, and suppose that a sequence (fε)⊂L2(K) is bounded and converges weakly to zero as ε→0. Then there exists a sequence of zero-mean solutions Bε∈W1,2(K;Rd) of such that Moreover, here exists C>0 such that Proof By extending fε by zero into a cube containing K and using appropriate scaling and translation arguments, it is sufficient to prove the statement of the lemma for the case K=□2π. To that end, consider the Fourier series for fε: Since fε converges weakly to zero, its Fourier coefficients also converge to zero: Define Bε as follows: By direct inspection, Bε has zero mean, satisfies the equation (156) and bound (158). Since ∑k|fkε|2 is bounded uniformly in ε, for all δ>0 there exists k0∈N such that for all ε. Furthermore, by virtue of (159) one has, for sufficiently small ε, It follows that, for sufficiently small ε, which implies (157). □ Remark D.7 In the case when fε is the ε-realisation of a zero-mean function from L2(Ω), one can prove the first part of the statement of Lemma D.6 (without the bound (158)) via an argument similar to that used in Lemma D.4 and Corollary D.5. Namely, for a zero-mean f∈L2(Ω), the potential fields Fi·:=(Fi1,⋯,Fin)T∈Vpot2,i=1,⋯,d, defined via satisfies f=Fii. Then the zero-mean field Bε=(B1ε,⋯,Bnε)T∈W1,2(K;Rd) is defined by ∇Biε=Fi·ε. Note that thus defined Bε does not necessary coincide with the one in the above lemma. We prefer the method used in the lemma since, first, it applies to a sequence which is not necessary the ε-realisation of a zero-mean function from L2(Ω), and, second, it provides the bound (158). Lemma D.8 The following bounds hold for the ε-realisations of b: Proof The rescaled function bλε(εy) is the solution to on every Oωk. By virtue of (13), we have Multiplying (160) by bλε(ε·) and integrating by parts yields The statement now follows by a rescaling argument and taking the norm over □L. □ Other Auxiliary Results Lemma E.1 Let A be a non-negative self-adjoint operator in a Hilbert space H. Let a(·,·) be the sesquilinear form associated with the operator A+1. Assume that for some u∈Dom(a), ‖u‖=1, λ∈R, and 0<ϵ<1 we have Then the exists u^∈Dom(A) such that In particular, one has Proof Defining u^:=(λ+1)(A+1)-1u, we have a(u-u^,v)=a(u,v)-(λ+1)(u,v). Substituting v=u-u^, we obtain a(u-u^,u-u^)≤ϵa(u-u^,u-u^), and hence Finally, the triangle inequality yields ‖u^‖≥1-ϵ, which, together with the obvious equality (A-λ)u^=(λ+1)(u-u^) implies the second inequality in (161). □ Lemma E.2 Let 1≤p<∞ and X⊂Rd be such that continuous functions with compact support are dense in Lp(X). Let {Xkε}k∈N,ε>0 be a family of subsets of X that are mutually disjoint for every ε and are such that limε→0supk∈NdiamXkε=0. Consider the “local averaging” operators Pε mapping f∈Lp(X) to the function Then Pεf→f strongly in Lp(X) as ε→0 for every f∈Lp(X). Proof For f∈Lp(X) we take a sequence fn∈Cc(Rd) such that fn→f strongly in Lp(X). Because of the uniform continuity of fn we obviously have Pεfn→fn strongly in Lp(X) as ε→0 for a fixed n. Fix δ>0 and n∈N such that ‖f-fn‖Lp(X)≤δ/3. Take also ε0>0 such that ‖fn-Pεfn‖Lp(X)<δ/3,0<ε<ε0. Using Hölder’s inequality on each Xkε, we obtain and hence ‖Pεfn-Pεf‖Lp(X)≤δ/3. Now, the triangle inequality yields The claim follows by the fact that δ is arbitrary. □ Lemma E.3 Let U be a bounded open set in Rd and let Un,n∈N be a sequence of open sets converging to U as n→∞ in the Hausdorff metric, such that U and Un,n∈N, are (ρ,N,γ) minimally smooth, and U,Un,Rd\U¯,Rd\Un¯,n∈N, are connected. Assume that λ is uniformly away from the spectra of the Dirichlet Laplacian operators on each of the sets -ΔU, Un,n∈N. Let f∈L2(V), where the open set V is large enough to contain U and Un,n∈N. Then the solutions ϕn∈W01,2(Un) to converges to the solution ϕ∈W01,2(U) to weakly in W1,2(V) and strongly in L2(V), after extending ϕn, ϕ by zero outside Un, U, respectively. Proof Clearly, the solutions ϕn are uniformly bounded in W1,2(V) and hence they converge (up to a subsequence) weakly in W1,2(V) and strongly in L2(V) to some ϕ^∈W1,2(V). It is not difficult to see that ϕ^∈W01,2(U). Let U′ be an arbitrary open set such that U′¯⊂U. Then, as a consequence of the minimal smoothness assumption, for sufficiently large n we have U′⊂Un (see the proof of Lemma 5.22)). Let ψ be an arbitrary test function from W01,2(U′) extended by zero outside U′. We have for sufficiently large n. Passing to the limit as n→∞, we obtain It follows from the density of smooth compactly supported in U functions in W01,2(U) that ϕ^ is the solution of (162), i.e. ϕ^=ϕ. Since we can apply the above argument to an arbitrary subsequence, the convergence holds for the entire sequence ϕn. □ Funding The research leading to these results received funding from grants funding organisations as indicated in the Acknowledgements. Data Availability There is no additional research data associated with this manuscript. Due to the nature of the research, there is no associated data. Declarations Conflict of interest The authors have no conflict of interests to declare that are relevant to the content of this article. Footnotes The argument is classical and goes as follows: Since -ΔO-λI acts onto, one can immediately show that the kernel of -ΔO-λI is trivial utilising its self-adjointness. Thus -ΔO-λI is a bijection from its domain onto L2(O). The domain of -ΔO-λI endowed with the graph norm is a Banach space. Denote it by B. The operator -ΔO-λI:B→L2(O) is bounded. Applying the the bounded inverse theorem one concludes that (-ΔO-λI)-1:L2(O)→B is bounded, hence (-ΔO-λI)-1:L2(O)→L2(O) is bounded. The equality is valid in the sense of identifying a random variable with the associated random field. Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement There is no additional research data associated with this manuscript. Due to the nature of the research, there is no associated data.