Arxiv:0806.3881V3 [Math.OA] 15 May 2009

Home , Spectral geometry

arXiv:0806.3881v3 [math.OA] 15 May 2009 2 1 prtr lcrclrssac ewr,e network, resistance electrical operator, upre yteUiest fIw eateto Mathemati of Department Iowa of University the by supported 73,6J0 08,6D5 66,8C0 11.Secondary 81Q10. 82C10, 46L60, 60D05, 60J85, 60J10, 47B32, esnegfroant ogrneodr atc models lattice order, m long-range banded ferromagnet, walk, Heisenberg random theory, graph spectral trip trees, Gel’fand system, embedding, Neumann von self-adjoint, tially partial space self-adjoint, path essentially orthogonality, paths, space, infinite boundary, t Martin Fourier kernel, Poisson analysis, harmonic function, harmonic geometry, e od n phrases. and words Key Date 2000 h oko EJwsprilyspotdb S rn DMS-045 grant NSF by supported partially was PETJ of work The PRTRTER FEETIA EITNENETWORKS RESISTANCE ELECTRICAL OF THEORY OPERATOR : ahmtc ujc Classification. Subject Mathematics a ,21.Cenversion. Clean 2019. 4, May eiae oBbPwr,i eonto fhspoern sp pioneering his of recognition in Powers, Bob to Dedicated AL .T OGNE N RNP .PEARSE J. P. ERIN AND JORGENSEN T. E. PALLE iihe om rp nry iceeptnilter,g theory, potential discrete energy, graph form, Dirichlet ff cierssac,rssac erc nqeeso curre of uniqueness metric, resistance resistance, ective rmr:0C0 57,3C0 23,4E2 62,47B25, 46F25, 46E22, 42C30, 31C20, 05C75, 05C50, Primary: smty ieroeaos none ieroeaos e operators, linear unbounded operators, linear isometry, 1 M-tts erdcn enl,fae,self-similar. frames, kernels, reproducing KMS-states, , e igdHletsae inrtasom dynamical transform, Wiener space, Hilbert rigged le, asom onayrpeetto,Mro process, Markov representation, boundary ransform, esr,wihe rps onayter,Hilbert theory, boundary graphs, weighted measure, tie,rssac om,rssac erc isotropic metric, resistance forms, resistance atrices, sNFVGEgatDMS-0602242. grant VIGRE NSF cs 13,4C5 73,5C3 22,82C41. 82C22, 52C23, 47B39, 42C25, 31C35, : 51 h oko PPwspartially was EPJP of work The 7581. irit. ahLpain transfer Laplacian, raph t,di nts, ff usion ssen- 2

Abstract. An electrical resistance network (ERN) is a weighted graph (G, c). The conductance function cxy weights the edges, which are then interpreted as resistors of possibly varying strengths. The effective resistance metric R(x, y) is the natural notion of distance between two vertices x, y in the ERN. The space of functions of finite energy (modulo constants) is a Hilbert space with inner product , which we call the energy space . The evaluation functionals on give E HE HE rise to a reproducing kernel vx for the space. Once a reference vertex o is fixed, these { } functions vx satisfy ∆vx = δx δo, where ∆ is the network Laplacian. This kernel yields − a detailed description of the structure of = in arm, where in is the closure HE F ⊕ H F of the space of finitely supported functions and arm is the closed subspace of harmonic H functions. The energy splits accordingly into a “finite part” expressed as a sum taken over E the vertices, and an “infinite part” expressed as a limit of sums. Intuitively, the latter part corresponds to an integral over some sort of boundary bdG, which is developed explicitly in 8. The kernel vx also allows us to recover easily many known (and sometimes difficult) § { } results about . As does not come naturally equipped with a natural o.n.b., we HE HE provide candidates for frames (and dual frames) when working with an infinite ERN. In particular, the presence of nonconstant harmonic functions of finite energy leads to different plausible definitions of the effective resistance metric on infinite networks. We characterize the free resistance RF (x, y) and the wired resistance RW (x, y) in terms of Neumann or Dirichlet boundary conditions on a certain operator. (In the literature, these correspond to the limit current and minimal current, resp.) We develop a library of equivalent formulations for each version. Also, we introduce the “trace resistance” RS (x, y), computed in terms of the trace of the Dirichlet form to finite subnetworks. This E provides a finite approximation which is more accurate from a probabilistic perspective, and gives a probabilistic explanation of the discrepancy between RF and RW . For R = RF or R = RW , the effective resistance is shown to be negative semidefinite, so that it induces an inner product on a Hilbert space into which it naturally embeds. We show that for (G, RF ), the resulting Hilbert space is and for (G, RW ) it is in. Under HE F the free embedding, each vertex x is mapped to the element vx of the energy kernel; under the wired embedding it is mapped to the projection fx of vx to in. This establishes as F HE the natural Hilbert space in which to study effective resistance. We obtain an analytic boundary representation for elements of arm in a sense anal- H ogous to that of Poisson or Martin boundary theory. We construct a Gel’fand triple S ⊆ S ′ and obtain a probability measure P and an isometric embedding of into H2E ⊆ HE L (S ′, P). This gives a concrete representation of the boundary in terms of the measures (1 + vx )dP S / in, where xn is a sequence tending to infinity. n ∈ ′ F { } The spectral representation for the graph Laplacian ∆ on is drastically different HE from the corresponding representation on ℓ2. Since the ambient Hilbert space is defined HE by the energy form, many interesting phenomena arise which are not present in ℓ2; we highlight many examples and explain why this occurs. In particular, we show how the deficiency indices of ∆ as an operator on indicate the presence of nontrivial boundary HE of an ERN, and why the ℓ2 operator theory of ∆ does not see this. Along the way, we prove that ∆ is always essentially self-adjoint on the ℓ2 space of functions on an ERN, and examine conditions for the network Laplacian and its associated transfer operator to be bounded, compact, essential self-adjoint, etc. We consider two approaches to measures on spaces of infinite paths in an ERN. One arises from considering the transition probabilities of a random walk as determined directly by the network, i.e., p(x, y) = cxy/ y x cxy. The other applies only to transient networks, ∼ and arises from considering the transitionP probabilities induced by a unit flow to infinity. The latter leads to the notion of forward-harmonic functions, for which we also provide a characterization in terms of a boundary representation. Using our results we establish precise bounds on correlations in the Heisenberg model for quantum spin observables, and we improve earlier results of R. T. Powers. Our focus is on the quantum spin model on the rank-3 lattice, i.e., the ERN with Z3 as vertices and with edges between nearest neighbors. This is known as the problem of long-range order in the physics literature, and refers to KMS states on the C∗-algebra of the model. OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 3

1 Contents

2 1. Introduction 5 3 1.1. Outline 5 4 1.2. What this paper is about 13 5 1.3. What this paper is not about 16 6 1.4. General remarks 20 7 1.5. Acknowledgements 20 8 2. Electrical resistance networks 22 9 2.1. The electrical resistance network model 22 10 2.2. The energy 24 11 3. Currents and potentials on electrical resistance networks 27 12 3.1. Currents on electrical resistance networks 27 13 3.2. Potential functions and their relationship to current flows. 28 14 3.3. The compatibility problem 31 15 4. The energy Hilbert space 35 16 4.1. The evaluation operator Lx and the reproducing kernel vx 37 17 4.2. The finitely supported functions and the harmonic functions 38 18 4.3. Relating the energy form to the Laplacian 40 19 5. The resistance metric 50 20 5.1. Resistance metric on finite networks 50 21 5.2. Resistance metric on infinite networks 53 22 5.3. Harmonic resistance 64 23 5.4. Comparison of resistance metric to other metrics 65 24 5.5. Generalized resistance metrics 66 25 6. von Neumann construction of the energy space 68 HE 26 6.1. von Neumann’s embedding theorem 68 27 6.2. as an invariant of G 69 HE 28 7. The boundary and boundary representation 72 29 7.1. Motivation and outline 72 30 7.2. Gel’fand triples and duality 74 31 7.3. The boundary as equivalence classes of paths 82 32 7.4. The structure of ′ 84 SG 33 8. The Laplacian on 86 HE 34 8.1. Properties of ∆ on 86 HE 35 8.2. The defect space of ∆ 90 V 36 8.3. Dual frames and the energy kernel 91 2 37 9. The ℓ theory of ∆ and the transfer operator 97 2 38 9.1. The Laplacian on ℓ (1) 97 39 9.2. The transfer operator 100 2 40 9.3. The Laplacian and transfer operator on ℓ (c) 104 41 10. The dissipation space D and its relation to 109 H HE 42 10.1. The structure of 110 HD 43 10.2. The divergence operator 113 44 10.3. Analogy with calculus and complex variables 115 45 10.4. Solving potential-theoretic problems with operators 116 46 11. Probabilistic interpretations 118 47 11.1. The path space of a general random walk 118 48 11.2. The forward-harmonic functions 122 49 12. Examples and applications 128 50 12.1. Finite graphs 128 51 12.2. Infinite graphs 132 52 13. Infinite trees 134 4 PALLEE.T.JORGENSENANDERINP.J.PEARSE

14. Lattice networks 140 1 14.1. Simple lattice networks 140 2 14.2. Noncompactness of the transfer operator 148 3 14.3. Non-simple integer lattice networks 150 4 14.4. Defect spaces 154 5 15. Application to magnetism and long-range order 157 6 2 15.1. Kolmogorov construction of L (Ω, P) 157 7 15.2. The GNS construction 159 8 15.3. Magnetism and long-range order in electrical resistance networks 161 9 15.4. KMS states 163 10 16. Future directions 166 11 Appendix A. Some functional analysis 168 12 A.1. von Neumann’s embedding theorem 168 13 Appendix B. Some operator theory 170 14 B.1. Projections and closed subspaces 170 15 B.2. Partial isometries 171 16 B.3. Self-adjointness for unbounded operators 171 17 B.4. Banded matrices 175 18 Appendix C. Navigation aids for operators and spaces 178 19 C.1. A road map of spaces 178 20 C.2. A summary of the operators on various Hilbert spaces 178 21 References 179 22 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 5

1 1. Introduction

2 This article is dedicated to the construction of an operator-theoretic context for study of 3 certain potential-theoretic function spaces on graphs, and an investigation of the resulting 4 structures. The primary object of study is an electrical resistance network (ERN), a graph 5 with weighted edges. Our foundation is the eﬀective resistance metric as the intrinsic 6 notion of distance, and we approach the analysis of the ERN by studying the space of 7 functions on the vertices which have ﬁnite (Dirichlet) energy. There is a large existing 8 literature on this subject, but ours is unique in several respects, most of which are due to 9 the following.

10 We use the effective resistance metric to find canonical Hilbert spaces of functions • 11 associated with the ERN. 12 We adhere to the intuition arising from the metaphor of electrical resistance net- • 13 works, including Kirchhoff’s Law and Ohm’s Law. 14 We apply the results of our Hilbert space construction to the isotropic Heisenberg • 15 ferromagnet and prove a theorem regarding long-range order in quantum statistical 16 mechanics for certain lattice networks. 17 It is known (see [LP09] and the references therein) that the resistance metric is • 18 unique for finite graphs and not unique for certain infinite networks. We are able 19 to clarify and explain the difference in terms of certain Hilbert space structures, 20 and also in terms of Dirichlet vs. Neumann boundary conditions for a certain 21 operator. Additionally, we introduce trace resistance, and harmonic resistance and 22 relate these to the aforementioned.

23 By using the intrinsic inner product (associated to the effective resistance) we are able to 24 obtain results which are more physically realistic than many found elsewhere in the litera- 2 25 ture. This inner product is quite different than the standard ℓ inner product for functions 26 defined on the vertices of a graph, and holds many surprises. Many of our results apply 27 much more generally than those already present in the literature. The next section elabo- 28 rates on these rather vague remarks and highlights the advantages and differences inherent 29 in our approach, in a variety of circumstances. 30 This work is uniquely interdisciplinary, and as a consequence, we have made effort to 31 address the union (as opposed to the intersection) of several disparate audiences: graph 32 theory, resistance networks, spectral geometry, fractal geometry, physics, probability, un- 33 bounded operators in Hilbert space, C*-algebras, and others. It is inevitable that parts of 34 the background material there will be unknown to some readers and so we have included 35 the appendices to mediate this. After presenting our results at various talks, we felt that the 36 inclusion of this material would be appreciated by most.

37 1.1. Outline. 38 39 2 — Electrical resistance networks. We introduce the ERN as a connected simple § 0 1 40 graph G = G , G equipped with a positive weight function c on the edges. The edges 1 0 { 0 } 41 G G G are ordered pairs of vertices, so c is required to be symmetric. Hence, ⊆ × 1 42 each edge (x, y) G is interpreted as a conductor with conductance cxy (or a resistor ∈1 43 with resistance c−xy . Heuristically, smaller conductances (or larger resistances) correspond 44 to larger distances; see the discussion of 5 just below. We make frequent use of the § 45 weight that c defines on the vertices via c(x) = y x cxy, where y x indicates that 1 ∼ ∼ 46 (x, y) G . The graphs we are most interested in areP infinite graphs, but we do not make ∈ 47 any general assumptions of regularity, group structure, etc. We require that c(x) is finite at 6 PALLEE.T.JORGENSENANDERINP.J.PEARSE

0 each x G , but we do not generally require that the degree of a vertex be finite, nor that 1 ∈ c(x) be bounded. 2 In the “cohomological” tradition of von Neumann, Birkhhoff, Koopman, and others 3 [vN32c, Koo36, Koo57], we study the ERN by analyzing spaces of functions defined on 4 it. These are constructed rigourously as Hilbert spaces in 6.1; in the meantime we collect 5 0 § some results about functions u, v : G R defined on the vertices. The network Laplacian 6 → (or discrete Laplace operator) operates on such a function by taking v(x) to a weighted 7 average of its values at neighbouring points in the graph, i.e., 8

v(x) v(y) (∆v)(x) := c (v(x) v(y)) = − , (1.1) xy − c 1 Xy x Xy x −xy ∼ ∼ 1 where x y indicates that (x, y) G . (The rightmost expression in formula (1.1) is written 9 ∼ ∈ so as to resemble the familiar difference quotients from calculus.) This is the usual second- 10 difference operator of numerical analysis, when adapted to a network. There is a large 11 literature on discrete harmonic analysis (basically, the study of the graph/network Lapla- 12 cian) which include various probabilistic, combinatoric, and spectral approaches. It would 13 be difficult to give a reasonably complete account, but the reader may find an enjoyable ap- 14 proach to the probabilistic perspective in [Spi76,Tel06], the combinatoric in [ABR07], the 15 analytic in [Fab06], and the spectral in [Chu01, GILb]. More sources are peppered about 16 the relevant sections below. Our formulation (1.1) differs from the stochastic formulation 17 often found in the literature, but the two may easily be reconciled; see (2.6). 18 Together with its associated quadratic form, the bilinear (Dirichlet) energy form 19

1 (u, v) := c (u(x) u(y))(v(x) v(y)) (1.2) E 2 xy − − xXG0 Xy x ∈ ∼ 0 acts on functions u, v : G R and plays a central role in the (harmonic) analysis on 20 → (G, c). (There is also the dissipation functional D, a twin of which acts on functions 21 1 E deﬁned on the edges G and is introduced in the following section.) The ﬁrst space of 22 functions we study on the ERN is the domain of the energy, that is, 23

0 . dom := u : G R .. (u) < . (1.3) E { → E ∞} In 6.1, we construct a Hilbert space from the resistance metric (andshowit to be a canoni- 24 § cal invariantfor (G, c) in 6.2), thereby recoveringthe familiar result that dom is a Hilbert 25 § E space with inner product . (Actually, this is not quite true, as is only a quasinorm; see 26 E E the discussion of 6.1 just below for a more accurate description.) 27 § For ﬁnite graphs, we prove the simple and folkloric key identity which relates the energy 28 and the Laplacian: 29

(u, v) = u, ∆v = ∆u, v , u, v dom , (1.4) E h i1 h i1 ∈ E 2 where u, ∆v 1 = x G0 u(x)∆v(x) indicates the standard ℓ inner product. The formula 30 h i ∈ (1.4) is extended toP inﬁnite networks in Theorem 4.41 (see (1.9) for a preliminary discus- 31 sion), where a third term appears. Indeed, understanding the mysterious third term is the 32 motivation for most of this investigation. 33

3 — Currents and potentials on electrical resistance networks. We collect several 34 § well-known and folkloric results, and reprove some variants of these results in the present 35 context. Currents are introduced as skew-symmetric functions on the edges; the intuition 36 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 7

1 is that I(x, y) = I(y, x) > 0 indicates electrical current flowing from x to y. In marked − 2 contrast to common tradition in geometric analysis [ABR07, PS07], we do not fix an ori- 3 entation. For us, an orientation is a choice of one of (x, y), (y, x) for each edge, and hence { } 4 just a notation to be redefined as convenient. In particular, any nonvanishing current de- 5 fines an orientation; one makes the choice so that I is a positive function. At this point we 6 give the definition of the dissipation, an inner product defined for functions on the edges, 7 and its associated quadratic form:

1 1 2 D(I) = c− I(x, y) . (1.5) 2 xy (x,Xy) G1 ∈ 8 Most of our results in this section are groundwork for the sections to follow; several 9 results are folkloric or obtained elsewhere in the literature. We include items which relate 10 directly to results in later sections; the reader seeking a more well-rounded background is 11 directed to [LPW08, LP09, Soa94, CdV98, Bol98] and the excellent elementary introduc- 12 tion [DS84]. After establishing the Hilbert space framework of 4, we exploit the close § 13 relationship between the two functionals and D, and use operators to translate a prob- E 14 lem from the domain of one functional to the domain of the other. We also introduce 15 Kirchhoff’s Law and Ohm’s Law, and in 3.3 we discuss the related compatibility prob- § 16 lem: every function on the vertices induces a function on the edges via Ohm’s Law, but not 17 every function on the edges comes from a function on the vertices. This is related to the 18 fact that most currents are not “efficient” in a sense which can be made clear variationally 19 (cf. Theorem 3.26) and which is important in the definition of effective resistance metric 20 in Theorem 5.2. We recover the well-known fact that the dissipation of an induced current 21 is equal to the energy of the function inducing it in Lemma 3.16; this is formalized as an 22 isometric operator in Theorem 10.12. We show that the equation

∆v = δα δω (1.6) − 23 always has a solution; we call such a function a dipole. In(1.6) and everywhere else, we 0 24 use the notation δ to indicate a Dirac mass at x G , that is, x ∈ 1, y = x, δx = δx(y) := (1.7) 0, else.  25 Proving the existence of dipoles allows us to fill gaps in [Pow76a,Pow76b] (see 1.2.1 just  § 26 below) and extend the definition of effective resistance metric in Theorem 5.2 to infinite 27 dimensions. 28 As is discussed at length in Remark 3.11, the study of dipoles, monopoles, and harmonic 29 functions is a recurring theme of this paper:

∆v = δα δω, ∆w = δω, ∆h = 0. − − 0 30 As mentioned above, for any network G and any vertices x, y G , there is a dipole in ∈ 31 dom . However, dom does not always contain monopoles or nonconstant harmonic E E 32 functions; the existence of monopoles is equivalent to transience of the network [Lyo83]; 33 we give a new criterion for transience in Lemma 4.51. In Theorem 14.5, we show that d 34 the integer lattice networks (Z , 1) support monopoles iff d 3, but in Theorem 14.17 we d ≥ 35 show all harmonic functions on (Z , 1) are linear and hence do not have finite energy. (Both 36 of these results are well known; the first is a famous theorem of Polya — we include them 37 for the novelty of method of proof.) In contrast, the binary tree in Example 13.4 support 8 PALLEE.T.JORGENSENANDERINP.J.PEARSE monopoles and nontrivial harmonic functions, both of finite energy (any network support- 1 ing nontrivial harmonic functions also supports monopoles, cf. [Soa94, Thm. 1.33]). It is 2 apparent that monopoles and nontrivial harmonic functions are sensitive to the asymptotic 3 geometry of (G, c). 4

4 — The energy Hilbert space . We use the natural Hilbert space structure on the 5 § HE space of finite-energy functions (with inner product given by ) to reinterpret previous 6 E results as claims about certain operators, and thereby clarify and generalize results from 7 2– 3. This is the energy space . 8 § § HE We construct a reproducing kernel for from first principles (i.e., via Riesz’s Lemma) 9 0 HE in 4.1. If o G is any fixed reference point, define vx to be the vector in which 10 § ∈ HE corresponds (via Hilbert space duality) to the evaluation functional Lx: 11

L u := u(x) u(o). x − Then the functions v form a reproducing kernel, and v is a solution of the discrete 12 { x} x Dirichlet problem ∆v = δ δ . Although these functions are linearly independent, they 13 x x − o are usually neither an orthonormal basis (onb) nor a frame. However, the span of v is 14 { x} dense in dom and appears naturally when the energy Hilbert space is constructed from 15 E the resistance metric by von Neumann’s method; cf. 6.1. Note that the Dirac masses 16 § δ 0 , which are the usual candidates for an onb, are not orthogonal with respect to the 17 { x}G energy inner product (1.2); cf. (2.11). In fact, Theorem 4.47 shows that δ 0 may not 18 { x}G even be dense in the energy Hilbert space! Thus, v is the only canonical choice for a 19 { x} representing set for functions of finite energy. 20 In 4.2 we use the Hilbert space structure of to better understand the role of the 21 § HE nontrivial harmonic functions. In particular, Lemma 4.22 shows that we may decompose 22 into the functions of finite support ( in) and the harmonic functions of finite energy 23 HE F ( arm): 24 H = in arm. (1.8) HE F ⊕H In 4.3, we prove a discrete version of the Gauss-Green formula (Theorem 4.41) which 25 § appears to be absent from the literature: 26

(u, v) = u(x)∆v(x) + u(x) ∂v (x), u , v (1.9) E ∂Ò ∀ ∈HE ∈ MP xXG0 xXbd G ∈ ∈ ∂v where (x) denotes the normal derivative of v, and is a space containing span vx ; see 27 ∂Ò MP { } 4.3 for precise definitions. For the moment, both the boundary and the normal derivatives 28 § are understood as limits (and hence vanish trivially for finite graphs); we will be able to 29 define these objects more concretely via techniques of Gel’fand in 7. 30 § It turns out that the boundary term (that is, the rightmost sumin(1.9)) vanishes unless 31 the network supports nontrivial harmonic functions (that is, nonconstant harmonic func- 32 tions of finite energy). More precisely, in Theorem 4.47 we prove that there exist u, v 33 ∂v ∈HE for which u , 0 if and only if the network is transient. That is, the random walk 34

bd G ∂Ò on the networkP with transition probabilities p(x, y) = cxy/c(x) is transient. We also give 35 several other equivalent conditions for transience, in 4.3.1. 36 § It is easy to prove (see Corollary 4.65) that nontrivial harmonic functions cannot lie in 37 2 0 2 0 ℓ (G ). This is why we do not require u, v ℓ (G ) in general, and why we stringently 38 ∈ avoid including such a requirement in the deﬁnition of the domain of the Laplacian. Such 39 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 9

1 a restriction would remove the nontrivial harmonic functions from the scope of our anal- 2 ysis, and we will see that they are at the core of some of the most interesting phenomena 3 appearing on an inﬁnite ERN.

4 5 — Effective resistance metric. The effective resistance metric R is foundational to § 5 our study, instead of the shortest-path metric more commonly used as graph distance. The 6 shortest-path metric on a weighted graph is usually defined to be the sum of the resistances 7 in any shortest path between two points. The effective resistance metric is also defined via 8 c, but in a more complicated way. The crucial difference is that the effective resistance 9 metric reflects both the topology of the graph and the weighting c; two points are closer to- 10 gether when there is more connectivity (more paths and/or paths with greater conductance) 11 between them. The effective resistance metric is a much more accurate way to measure 12 distance when travel from point x to point y can be accomplished simultaneously through 13 many paths, for example, flow of electrical current, fluid diffusion through porous media, 14 or data transfer over the internet. 15 In 5.1, we give a multifarious definition of the effective resistance metric R, which § 16 may be physically characterized as the voltage drop between two vertices when electrical 17 leads with a fixed current are applied to the two vertices. Most of these formulations 18 appear elsewhere in the literature, but some appear to be specific to the physics literature, 19 some to probability, and some to analysis. We collect them and prove their equivalence in 20 Theorem 5.2, including a couple new formulations that will be useful in later sections. 21 It is somewhat surprising that when these formulas are extended to an infinite network 22 in the most natural way, they are no longer equivalent. (Note that each of the six formulas 23 has both a free and wired version, but some appear much less natural in one version than F 24 in the other.) Some of the formulas lead to the “free resistance” R and others lead to the W 25 “wired resistance” R ; here we follow the terminology of [LP09]. In 5.2, we precisely § 26 characterize the types of extensions that lead to each, and explain this phenomenon in 27 terms of projections in Hilbert space, Dirichlet vs. Neumann boundary conditions, and via 28 probabilistic interpretation. Additionally, we discuss the “trace resistance” given in terms 29 of the trace of the Dirichlet form , and we study the “harmonic resistance” which is the F W E 30 difference between R and R and is not typically a metric.

31 6 — Construction of the energy space . In 6.1, we use a theorem of von Neumann § HE § 32 to give an isometric imbedding of the metric space (G, R) into ; cf. Theorem 6.1. For F W HE 33 inﬁnite networks, (G, R ) embedsinto and(G, R ) embeds into in. In 6.2 we discuss HE F § 34 how this enables one to interpret as an invariant of the original ERN. HE

35 7 — The boundary bd G and boundary representation. We study the boundary bd G in § 36 terms of the Laplacian by reinterpreting the boundary term of (1.9) as an integral over a 37 space which contains . This gives a representation of bd G as a measure space whose HE 38 structure is well-studied. 39 In Theorem 7.1 of 7.1, we observe that an important consequence of (1.9) is the fol- § 40 lowing boundary representation for the harmonic functions:

u(x) = u ∂hx + u(o), (1.10)

∂Ò Xbd G

41 for u arm, where hx = P armvx is the projection of vx to arm;see(1.8). This formula ∈H H H 42 is in the spirit of Choquet theory and the Poisson integral formula and is closely related to 43 Martin boundary theory. 10 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Unfortunately, the sum in (1.10) is only understood in a limiting sense and so provides 1 limited insight into the nature of bd G. This motivates the development of a more concrete 2 expression. We use a self-adjoint extension ∆ of ∆ to construct a Gel’fand triple G 3 ∗ V S ⊆ P = ∆ 4 G′ and a Gaussian probability measure . Here, G : dom( ∗ ∞) is a suitable dense HE ⊆ S S V (Schwartz) space of “test functions” on the ERN, and ′ is the corresponding dual space 5 SG of “distributions” (or “generalized functions”). This enables us to identify bd G as a subset 6 of , and in Corollary 7.27, we rewrite (1.10) more concretely as 7 SG′

u(x) = u(ξ)hx(ξ) dP(ξ) + u(o), (1.11) Z SG′ again for u arm and with hx = P armvx. Thus we study the metric/measure structure 8 ∈ H H of G by examining an associated Hilbert space of random variables. This is motivated in 9 part by Kolmogorov’s pioneering work on stochastic processes (see 15.1) as well as on 10 § a powerful reﬁnement of Minlos. The latter is in the context of the Gel’fand triples men- 11 tioned just above; see [Nel64] and 7.2 below. Further applications to harmonic analysis 12 § and to physics are given in 11– 15.3. 13 § §

8 — The Laplacian on . We study the operator theory of the Laplacian in some 14 § HE detail in 8.1, examining the various domains and self-adjoint extensions. We identify one 15 § domain for the Laplacian which allows for the choice of a particular self-adjoint extension 16 for the constructions in 7. Also, we give technical conditions which must be considered 17 § when the graph contains vertices of inﬁnite degree and/or the conductance functions c(x) 18 0 is unbounded on G . This results in an extension of the Royden decomposition to = 19 HE in in arm, where in is the -closure of span δ δ and in is the orthogonal 20 F 1 ⊕F 2 ⊕H F 2 E { x − o} F 1 complement of in within in. Example 14.35 shows a case where in is not dense in 21 F 2 F F 2 in. 22 F In 8.2, we study the defect space of ∆ , that is, the space spanned by solutions to 23 § V ∆u = u. In 8.2.1, we relate the boundary term of (1.9) to the the boundary form 24 − § 1 βbd(u, v) := ∆∗ u, v u, ∆∗ v , u, v dom(∆∗ ) (1.12) 2i h V iE − h V iE ∈ V of classical functional analysis; cf. [DS88, XII.4.4]. This gives a way to detect whether 25 § or not a given network has a boundary by examining the deﬁciency indices of ∆. In Theo- 26 rem 8.19, weshowthatif ∆ fails to be essentially self-adjoint, then arm , 0 . In general, 27 H { } the converse does not hold: Corollary 9.28 shows that ∆ has no defect when deg(x) < 28 ∞ and c(x) is bounded. (Thus, any homogeneous tree of degree 3 or higher with constant 29 conductances provides a counterexample to the converse.) 30 In 8.3, we study the relation between the reproducing kernel v and the spectral 31 § { x} properties of ∆ and its self-adjoint extensions. In particular, we examine the necessary 32 conditions for vx to be a frame for , and the relation between vx and δx. 33 { } HE 2 9 — The ℓ theory of ∆ and T. We consider some results for ∆ and T as operators on 34 2 § 2 ℓ (1), where the inner product is given by u, ∆v := u(x)∆v(x) and on ℓ (c), where the 35 h i1 inner product is given by u, ∆v c := c(x)u(x)∆v(x).P 36 h i 2 We prove that the Laplacian is essentiallyP self-adjoint on ℓ (1) under very mild hy- 37 potheses in 9.1. The subsequent spectral representation allows us to give a precise char- 38 § acterization of the domain of the energy functional in this context. In 9.2, we examine 39 E § boundedness and compactness of ∆ and T in terms of the decay properties of c. The space 40 2 ℓ (c) considered in 9.3 is essentially a technical tool; it allows for a proof that the terms 41 § of the Discrete Gauss-Green formula are absolutely convergent and hence independent of 42 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 11

1 any exhaustion. However, it is also interesting in its own right, and we show an interesting 1 2 connection with the probabilistic Laplacian c− ∆. Results from this section imply that ∆ is 2 3 also essentially self-adjoint on , subject to the same mild hypotheses as the ℓ (1) case. HE 4 The energy Hilbert space contains much diﬀerent information about a given inﬁnite HE 2 5 graph system (G, c) than does the more familiar ℓ sequence space, even when appropriate 6 weights are assigned. In the language of Markov processes, is better adapted to the 2 HE 7 study of (G, c) than ℓ . One reason for this is that is intimately connected with the HE 8 resistance metric R.

9 10 — and D. The dissipation space D is the Hilbert space of functions on the § HE H H 10 edges when equipped with the dissipation inner product. We solve problems in discrete 11 potential theory with the use of the drop operator d (and its adjoint d∗), where

dv(x, y) := c (v(x) v(y)). (1.13) xy − 12 The drop operator d is, of course, just an implementation of Ohm’s Law, and can be in- 13 terpreted as a weighted boundary operator in the sense of homology theory. The drop 14 operator appears elsewhere in the literature, sometimes without the weighting cxy; see 15 [Chu01, Tel06, Woe00]. However, we use the adjoint of this operator with respect to the 2 16 energy inner product, instead of the ℓ inner product used by others. This approach appears 17 to be new, and it turns out to be more compatible with physical interpretation. For exam- 2 18 ple, the displayed equation preceding [Woe00, (2.2)] shows that the ℓ adjoint of the drop 19 operator is incompatible with Kirchhoff’s node law. Since the resistance metric may be 20 defined in terms of currents obeying Kirchhoff’s laws, we elect to make this break with the 21 existing literature. Additionally, this strategy will allow us to solve the compatibility prob- 22 lem described in 3.3 in terms of a useful minimizing projection operator P , discussed § d 23 in detail in 10.4. Furthermore, we believe our formulation is more closely related to the § 24 (co)homology of the ERN as a result. 25 We decompose into the direct sum of the range of d and the currents which are HD 26 sums of characteristic functions of cycles

= ran d [χ ], (1.14) HD ⊕ ϑ 27 where ϑ is a cycle, i.e., a path in the graph which ends where it begins. In(1.14) and else- 28 where, we indicate the closed linear span of a set by [χ ] := cl span χ . From (1.8) (and ϑ { ϑ} 29 the fact that d is an isometry), it is clear that the first summand of (1.14) can be further 30 decomposed into weighted edge neighbourhoods dδx and the image of harmonic functions 31 under d in Theorem 10.8. After a first draught of this paper was complete, we discovered 32 that the same approach is taken in [LP09]. One of us (PJ) recalls conversations with Raul 33 Bott concerning an analogous Hilbert space operator theoretic approach to electrical net- 34 works; apparently attempted in the 1950s in the engineering literature. We could not find 35 details in any journals; the closest we could come is the fascinating paper [BD49] by Bott 36 et al. A further early source of influence is Norbert Wiener’s paper [WR46]. 37 In 10.4, we describe how d solves the compatibility problem and may be used to solve § ∗ 38 a large class of problems in discrete potential theory. Also, we discuss the analogy with 39 complex analysis.

40 11 — Probabilistic interpretations. In [LP09,DS84,Tel06,Woe00] and elsewhere, the § 41 randomwalk on an ERN is defined by the transition probabilities p(x, y) := cxy/c(x). In this 1 42 context, the probabilistic transition operator is P = c− T and one uses the stochastically 1 43 renormalized Laplacian ∆c := c− ∆, where c is understood as a multiplication operator; 12 PALLEE.T.JORGENSENANDERINP.J.PEARSE see Definition 2.3. This approach also arises in the discussion of trace resistance in 5.2.3 1 § and allows one to construct currents on the graph as the average motion of a random walk. 2 As an alternative to the approach described above, we discuss a probabilistic interpre- 3 tation slightly different from those typically found in the literature: we begin with a volt- 4 age potential as an initial condition, and consider the induced current I. The components 5 of such a flow are called current paths and provide a way to interpret potential-theoretic 6 problems in a probabilistic setting. We study the random walks where the transition prob- 7 ability is given by I(x, y)/ z x I(x, z). We consider the harmonic functions in this con- 8 ∼ text, which we call forward-harmonicP functions, and the associated forward-Laplacian of 9 Definition 11.19. We give a complete characterization of forward-harmonic functions as 10 cocycles, following [Jor06]. 11

12 — Examples. We collect an array of examples that illustrate the various phenomena 12 § encountered in the theory and work out many concrete examples. Some elementary finite 13 examples are given in 12.1 to give the reader an idea of the basics of electrical resistance 14 § network theory. In 12.2 we move on to infinite graphs. 15 § 13 — Trees. When the electrical resistance network has a tree as an underlying net- 16 § work, the resistance distance coincides with shortest-path metric, as there is always exactly 17 one path between any two vertices; cf. Lemma 5.46 and the preceding discussion. When 18 the tree has exponential growth, as in the case of homogeneous trees of degree 3, one 19 ≥ can always construct nontrivial harmonic functions, and monopoles of finite energy. In 20 fact, there is a very rich family of each, and this property makes this class of examples a 21 fertile testing ground for many of our theorems and definitions. In particular, these exam- 22 ples highlight the relevance and distinctions between the boundary (as we construct it), the 23 Cauchy completion, and the graph ends of [PW90, Woe00]. In particular, they enable one 24 to see how adjusting decay conditions on c affects these things. 25

d 14 — Integer lattices. The lattice electrical resistance networks (Z , c) have vertices at 26 § d the points of R which have integer coordinates, and edges between every pair of vertices 27 (x, y) with x y = 1. The case for c = 1 is amenable to Fourier analysis, and in 14.1 we 28 | − | § obtain explicit formulas for many expressions: 29

Lemma 14.4 gives a formula for the potential configuration functions v . 30 • { x} Theorem 14.7 gives a formula for the resistance distance R(x, y). 31 • Theorem 14.9 gives a formula for the resistance distance to infinity in the sense 32 • R(x, ) = limy R(x, y). 33 ∞ →∞ d Theorem 14.5 gives a formula for the solution w of ∆w = δ on Z ; it is readily 34 • − o seen that this w has finite energy (i.e., is a monopole) iff d 3. 35 ≥ In [Pól21], Pólya proved that the random walk on this graph is transient if and only if 36 d 3; see [DS84] for a nice exposition. We offer a new characterization of this dichotomy 37 ≥ d (there exist monopoles on Z if and only if d 3) which we recover in this section via a 38 ≥ new (and completely constructive) proof. In Remark 14.21 we describe how in the infinite 39 integer lattices, functions in may be approximated by functions of finite support. 40 HE d 15 — Magnetism. The integer lattice networks (Z , 1) investigated in 14 comprise the 41 § § framework of infinite models in thermodynamics and in quantum statistical mechanics. In 42 15.3 we employ these formulas in the refinement of an application to the theory of the 43 § (isotropic Heisenberg) model of ferromagnetism as studied by R. T. Powers. In addition 44 to providing an encapsulated version of the Heisenberg model, we give a commutative 45 analogue of the model, extend certain results of Powers from [Pow75, Pow76a, Pow76b, 46 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 13

1 Pow78, Pow79], and discuss the application of the resistance metric to the theory of ferro- 2 magnetism and “long-range order”. This problem was raised initially by R. T. Powers, and 3 may be viewed as a noncommutative version of Hilbert spaces of random variables. 4 Ferromagnetism in quantum statistical mechanics involves algebras of noncommutative 5 observables and may be described with the use of states on C∗-algebras. As outlined in 6 the cited references, the motivation for these models draw on thermodynamics; hence the 7 notions of equilibrium states (formalized as KMS states, see 15.4). These KMS states § 8 are states in the C∗-algebraic sense (that is, positive linear functionals with norm 1), and 9 they are indexed by absolute temperature. Physicists interpret such objects as representing 10 equilibria of infinite systems. 11 In the present case, we consider spin observables arranged in a lattice of a certain rank, 12 d = 1, 2, 3,..., and with nearest-neighbor interaction. Rigourous mathematical formula- 13 tion of phase transitions appears to be a hopeless task with current mathematical technol- 14 ogy. As an alternative avenue of enquiry, much work has been conducted on the issue of 15 long-range order, i.e., the correlations between observables at distant lattice points. These 16 correlations are measured relative to states on the C∗-algebra; in this case in the KMS states 17 for a fixed value of temperature. 18 While we shall refer to the literature, e.g. [BR79, Rue69] for formal definitions of key 19 terms from the C∗-algebraic formalism of quantum spin models, physics, and KMS states, 20 we include a minimal amount of background and terminology from the physics literature.

21 16 — Future Directions. We conclude with a brief discussion of several projects which § 22 have arisen from work on the present paper, as well as some promising new directions that 23 we have not yet had time to pursue.

24 Appendices. We give some background material from functional analysis in Appen- 25 dix A, and operator theory in Appendix B. In Appendix C, we include some diagrams to 26 help clarify the properties of the many operators and spaces we discuss, and the relations 27 between them.

28 1.2. What this paper is about. Theeffective resistance metric provides the foundation for 29 our investigations because it is the natural and intrinsic metric for an electrical resistance 30 network, as the work of Kigami has shown; see [Kig01] and the extensive list of references 31 by the same author therein. Moreover, the close relationship between diffusion geometry + 32 (i.e., geometry of the resistance metric) [MM08, SMC08, CKL 08, CM06] and random 33 walks on graphs leads us to expect/hope there will be many applications of our results to 34 several other subjects, in addition to fractals: models in quantum statistical mechanics, 35 analysis of energy forms, interplay between self-similar measures and associated energy 36 forms, certain discrete models arising in the study of quasicrystals (e.g., [BM00, BM01]), 37 and multiwavelets (e.g. [BJMP05, DJ06, DJ07, Jor06]), among others. A general theme 38 of these areas is that the underlying space is not sufficiently regular to support a group 39 structure, yet is “locally” regular enough to allow analysis via probabilistic techniques. 40 Consequently, the analysis of functions on such spaces is closely tied to Dirichlet energy 41 forms and the graph Laplacian operator associated to the graph. This appears prominently 42 in the context of the present paper as follows:

43 (1) The embedding of the metric space ((G, c), R) into the Hilbert space of func- HE 44 tions of ﬁnite energy,in such a way that the original metric may be recovered from 45 the norm, i.e., 14 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2 R(x, y) = (vx vy) = vx vy , E − k − kE where vx is the image of x under the embedding. 1 ∈HE (2) The relation of the energy form to the graph Laplacian via the equation 2

(u, v) = u(x)∆v(x) + u(x) ∂v (x), (1.15) E ∂Ò xXG0 xXbd G ∈ ∈ introduced just above in the discussion of 2. Each summation on the right hand 3 § side of (1.15) is more subtle than it appears. These details for the first sum are 4 given in Theorem 4.41, and the details for the second sum are the focus of almost 5 all of 7. 6 § (3) The presence of nonconstant harmonic functions of finite energy. These are pre- 7 W cisely the objects which support the boundary term in (1.15) and imply R (x, y) < 8 F R (x, y). They are also responsible for the boundary described in 7. 9 § (4) The solvability of the Dirichlet problem ∆w = δy, where δy is a Dirac mass 10 0 − at the vertex y G . The existence of finite-energy solutions w is equivalent 11 ∈ to the transience of the random walk on the network. Such functions are called 12 monopoles and (via Ohm’s law) they induce a unit flow to infinity as discussed 13 in [DS84, LP09, LPW08]. 14

Remark 1.1. In addition to uses in graph theory and electrical networks, the discrete Lapla- 15 cian ∆ has other uses in numerical analysis: many problemsin PDE theory lend themselves 16 to discretizations in terms of subdivisions or grids of refinements in continuous domains. 17 A key tool in applying numerical analysis to solving partial differential equations is dis- 18 cretization, and use of repeated differences; especially for using the discrete ∆ in approxi- 19 mating differential operators, and PDOs. See e.g., [AH05]. 20 One picks a grid size δ and then proceeds in steps: 21

(1) Start with a partial differential operator, then study an associated discretized oper- 22 d ator with the use of repeated differences on the δ-lattice in R . 23 (2) Solve the discretized problem for h fixed. 24 (3) As δ tends to zero, numerical analysts evaluate the resulting approximation limits, 25 and they bound the error terms. 26

When discretization is applied to the Laplace operator in d continuous variables, the 27 d result is our ∆ for the network (Z , c); see 14 for details and examples. However, when 28 § the same procedure is applied to a continuous Laplace operator on a Riemannian manifold, 29 the discretized ∆ will be the network Laplacian on a suitable inﬁnite network (G, c) which 30 d in general may have a much wilder geometry than Z . 31 This yields numerical algorithms for the solution of partial diﬀerential equations, and 32 in the case of second order PDEs, the discretized operator is the discrete Laplacian studied 33 in this investigation. 34

1.2.1. Motivation and applications. Applications to inﬁnite networks of resistors serve as 35 motivations, but our theorems have a wider scope, have other applications; and are, we 36 believe, of independent mathematical interest. Our interest originates primarily from two 37 sources. 38

(1) A series of paperswritten by Bob Powers in the 1970swhich he introduced inﬁnite 39 systems of resistors into the resolution of an important question from quantum 40 statistical mechanics in [Pow75, Pow76a, Pow76b, Pow78, Pow79]. 41 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 15

1 (2) The pioneering work of Jun Kigami on the analysis of PCF self-similar fractals, 2 viewing these objects as rescaled limits of networks; see [Kig01].

3 Indeed, our larger goal is the cross-pollination of these two areas, and we hope that the 4 results of this paper may be applicable to analysis on fractal spaces. A first step in this 5 direction is given in Theorem 16.3. To this end, a little more discussion of each of the 6 above two subjects is in order. 7 Powers was interested in magnetism and the appearance of “long-range order”, which 8 is the common parlance for correlation between spins of distant particles; see 15 for a § 9 larger discussion. Consequently, he was most interested in graphs like the integer lattice d 10 Z (with edges between vertices of distance 1, and all resistances equal to 1), or other 11 regular graphs that might model the atoms in a solid. Powers established a formulation of 12 resistance metric that we adopt and extend in 5, where we also show it to be equivalent to § 13 Kigami’s formulation(s). Also, the proofs of Powers’ originalresults on effective resistance 14 metric contain a couple of gaps that we fill. In particular, Powers does not seem to have be 15 aware of the possibility of nontrivial harmonic functions until [Pow78], where he mentions 16 them for the first time. It is clear that he realized several immediate implications of the 17 existence of such functions, but there more subtle (and just as important!) phenomena that 18 are difficult to see without the clarity provided by Hilbert space geometry. 19 Powers studied an infinite graph G by working with an exhaustion, that is, a nested se- 20 quence of finite graphs G1 G2 ... Gk G = k Gk. For example, Gk might be all d ⊆ ⊆ ⊆ ⊆ 21 the vertices of Z lying inside the ball of Euclidean radiusS k, and the edges between them. 22 Powers used this approach to obtain certain inequalities for the resistance metric, express- 23 ing the consequences of deleting small subsets of edges from the network. Although he 24 makes no reference to it, this approach is very analogous to Rayleigh’s “short-cut” meth- 25 ods, as it is called in [DS84]. 26 Powers’ use of an expanding sequence of graphs may be thought of as a “limit in the 27 large” in contrast to the techniques introduces by Kigami, which may be considered “limits 28 in the small”. Self-similarity and scale renormalization are the hallmarks of the theory of 29 fractal analysis as pursued by Kigami, Strichartz and others (see [HKK02, Kig01, Kig03, 30 Hut81, Str06, BHS05, Bea91, JP94, Jor04], for example) but these ideas do not enter into 31 Powers’ study of resistors. One aim of the present work is the development of a Hilbert 32 space framework suitable for the study of limits of networks defined by a recursive al- 33 gorithm which introduces new vertices at each step and rescales the edges via a suitable 34 contractive rescaling. As is known from, for example [JP94, Jor06, Str98a, Str06, Tep98], 35 there is a spectral duality between “fractals in the large”, and “fractals in the small”.

36 1.2.2. The significance of Hilbert spaces. A main theme in this paper is the use of Hilbert 37 space technology in understanding metrics, potential theory, and optimization on infinite 38 graphs, especially through finite-dimensional approximation. We emphasize those aspects 39 that are intrinsic to infinite electrical resistance networks, and our focus is on analytic as- 40 pects of graphs; as opposed to the combinatorial and algebraic sides of the subject, etc. 41 Those of our results stated directly in the framework of graphs may be viewed as discrete 42 analysis, yet the continuum enters via spectral theory for operators and the computation of 43 probability of sets of infinite paths. In fact, we will display a rich variety of possible spec- 44 tral types, considering the spectrum as a set (with multiplicities), as well as the associated 45 spectral measures, and representations/resolutions. 46 Related issues for Hilbert space completions form a recurrent theme throughout our pa- 47 per. Given an electrical resistance network, we primarily study three spaces of functions 16 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2 0 naturally associated with it: , D, and to a lesser extent ℓ (G ). Our harmonic anal- 1 HE H ysis of functions on G is studied via operators between the respective Hilbert spaces as 2 discussed in 10 and the Hilbert space completions of these three classes are used in an 3 § essential way. In particular, we obtain the boundary of the graph (a necessary ingredient 4 of (1.15) and the key to several mysteries) by analyzing the finite energy functions on G 5 which cannot be approximated by functions of finite support. However, this metric space 6 is naturally embedded inside the Hilbert space , which is already complete by defini- 7 HE tion/construction. Consequently, the Hilbert space framework allows us to identify certain 8 vectors as corresponding to the boundary of (G, c), and thus obtain a concrete understand- 9 ing of the boundary. 10 However, the explicit representations of vectors in a Hilbert space completion (i.e., the 11 completion of a pre-Hilbert space) may be less than transparent; see [Yoo07]. In fact, 12 this difficulty is quite typical when Hilbert space completions are used in mathematical 13 physics problems. For example, in [JO00´ ,Jor00], one beginswith a certain space of smooth 14 d functions defined on a subset of R , with certain support restrictions. In relativistic physics, 15 one must deal with reflections, and there will be a separate positive definite quadratic 16 form on each side of the “mirror”. As a result, one ends up with two startlingly different 17 2 Hilbert space completions: a familiar L -space of functions on one side, and a space of 18 distributions on the other. In [JO00´ ,Jor00], one obtains holomorphic functions on one side 19 of the mirror, and the space of distributions on the other side is spanned by the derivatives 20 of the Dirac mass, each taken at the same specific point x0. 21

1.3. What this paper is not about. Many of the topics discussed in this paper may appear 22 to have been previously discussed elsewhere in the literature, but there are certain important 23 subtleties which actually make our results quite diﬀerent. This section is intended to clarify 24 some of these. It is the opinionof the authorsthat most interesting results of this paper arise 25 primarily from three things: 26

(1) differences between finite approximations to infinite networks, and how & when 27 these differences vanish in the limit, and 28 (2) the phenomena that result when one works with a quadratic form whose kernel 29 contains the constant functions, and 30 (3) the boundary (which is not a subset of the vertices) that naturally arises when 31 a network supports nonconstant harmonic functions of finite energy, and how it 32 explains other topics mentioned above. 33

In classical potential theory, working modulo constant functions amounts to working with 34 2 the class of functions satisfying f < , but abandoning the ℓ requirement f < . 35 k ′k2 ∞ k k2 ∞ This has some interesting consequences, and the nontrivial harmonic functions play an 36 especially important role; see Remark 4.66. What would one hope to gain by removing the 37 2 ℓ condition? 38

(1) From the natural embedding of the metric space (G, R) into the Hilbert space 39 HE of functions of ﬁnite energy given by x vx, the functions vx are not generally in 40 2 7→ ℓ . See Figure 10 of Example 14.16 for an illustration. 41 2 (2) The resistance metric does not behave nicely with respect to ℓ conditions. Several 42 formulations of the resistance distance R(x, y) involve optimizing over collections 43 2 of functions which are not necessarily contained in ℓ , even for many simple ex- 44 amples. 45 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 17

2 0 1 (3) Corollary 4.65 states that nontrivial harmonic functions cannot lie in ℓ (G ). Con- 2 2 sequently, imposing an ℓ hypothesis removes the most interesting phenomena 3 from the scope of study; see Remark 4.66.

4 The inﬁnite trees studied in Examples 13.2–13.6 also provide examples of these situations.

5 While there already is a large literature on electrical networks and on graphs (see 6 e.g., [CW92, CW07, DK88, Dod06, DS84, Pow76b, CdV04, CR06, Chu07, FK07], and the 7 preprint [Str08] which we received after the first version of this paper was completed), we 8 believe that our present operator/spectral theoretic approach suggests new questions and 9 new theorems, and allows many problems to be solved in greater generality. 10 The literature on analysis on graphs breaks down into a variety of overlapping subar- 11 eas, including: combinatorial aspects, systems of resistors on infinite networks, random- 12 walk models, operator algebraic models [DJ08,Rae05], probability on graphs (e.g., infinite 13 particle models in physics [Pow79]), Brownian motion on self-similar fractals [Hut81], 14 Laplace operators on graphs, finite element-approximations in numerical analysis [BS08]; 15 and more recently, use in internet-search algorithms [FK07]. Even just the study of Laplace 16 operators on graphs subdivides further, due to recently discovered connections between 17 graphsand fractals generated by an iterated functionssystem (IFS); see e.g.,[Kig03,Str06]. 18 Other major related areas include discrete Schrödinger operators in physics, informa- 19 tion theory, potential theory, uses of the graphs in scaling-analysis of fractals (constructed 20 from infinite graphs), probability and heat equations on infinite graphs, graph C∗-algebras, 21 groupoids, Perron-Frobenius transfer operators (especially as used in models for the in- 22 ternet); multiscale analysis, renormalization, and operator theory of boundaries of infinite 23 graphs (more current and joint research between the co-authors.) The motivating appli- 24 cations from [Pow75, Pow76a, Pow76b, Pow78, Pow79] include the operator algebra of 25 electrical networks of resistors (lattice models, C∗-algebras, and their representations), and 26 more specifically, KMS-states from statistical mechanics. While working and presenting 27 our results, we learned of even more such related research directions from experts working 28 in these fields, and we are thankful to them all for taking the time to explain some aspects 29 of them to us. 30 The main point here is that the related literature is vast but our approach appears to be 31 entirely novel and our results, while reminiscent of classical theory, are also new. We now 32 elucidate certain specific differences.

33 1.3.1. Spectral theory. The spectral theory for networks contrasts sharply with that for 34 fractals, as is seen by considering the measures involved; they do not begin to become 35 similar until one considers limits of networks. The spectrum of discrete Laplacians on in- 36 finite networks is typically continuous (lattices or trees provide examples, and are worked 37 explicitly in 12). By contrast, in the analysis on fractals program of Kigami, Strichartz, § 38 and others, the Laplace operator has pure point spectrum; see [Tep98] in particular. The 39 measures used in the analysis of networks are weighted counting measures, while the mea- 40 sures used in fractal analysis are based on the self-similar measures introduced by Hutchin- 41 son [Hut81]. There is an associated and analogous entropy measure in the study of Julia 42 sets; cf. [Bea91] and the recent work on Laplacians in [RT08]. 43 Our approach differs from the extensive literature on spectral graph theory (see [Chu01] 44 for an excellent introduction, and an extensive list of further references) due to the fact 2 45 that we eschew the ℓ basis for our investigations. We primarily study ∆ as an operator 46 on , and with respect to the energy inner product. The corresponding spectral theory is HE 2 47 radically different from the spectral theory of ∆ in ℓ . Most other work in spectral graph 18 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2 theory takes place in ℓ , even implicitly when working with finite graphs: the adjoint of 1 2 the drop operator (see Definition 10.2) is taken with respect to the ℓ inner product and 2 consequently violates Kirchhoff’s laws. In fact, the discussion preceding [Woe00, (2.2)] 3 shows how this version of the adjoint is incompatible with Kirhhoff’s Law as mentioned in 4 the summary of 10 just above. Additionally, [Chu01] and others work with the spectrally 5 § 1/2 1/2 renormalized Laplacian ∆s := c− ∆c− . However, ∆s is a bounded Hermitian operator 6 (with spectrum contained in [0, 2]) and so is unsuitable for our investigationsof bdG based 7 on defect indices, etc. 8 As we have only encountered relatively few instances where the complete details are 9 worked out for spectral representations in the framework of discrete analysis, we have at- 10 tempted to provide several explicit examples. These are likely folkloric, as the geometric 11 possibilities of graphs are vast, and so is the associated range of spectral configurations. A 12 list of recent and past papers of relevance includes [Str08, Car72, Car73a, Car73b, CR06, 13 Chu07,CdV99,CdV04,Jor83], and Wigner’s original paper on the semicircle law [Wig55]. 14 The present investigation also led to a spectral analysis of the binary tree from the perspec- 15 tive of dipoles in [DJ08]; this study discovered that the spectrum of ∆ on the binary tree is 16 also given by Wigner’s semicircle law. 17 There is also a literature on infinite/transfinite networks and generalized Kirchhoff laws 18 using nonstandard analysis, etc., see [Zem91, Zem97]. However, this context allows for 19 edges with resistance 0, which we do not allow (for physical as well as theoretical rea- 20 sons). One can neglect the resistance of wires in most engineering applications, but not 21 when considering infinite networks (the epsilons add up!). The resulting theory therefore 22 divergesrapidly from the observationsof the present paper; according to our definitions, all 23 networks support currents satisfying Kirchhoff’s law, and in particular, all induced currents 24 satisfy Kirchhoff’s law. 25

1.3.2. Operator algebras. There are also recent papers in the literature which also ex- 26 amine graphs with tools from operator algebras and infinite determinants. The papers 27 [GILb, GILc, GILa] by Guido et al are motivated by questions for fractals and study the 28 detection of periods in infinite graphs with the use of the Ihara zeta function, a variant of 29 the Riemann zeta function. There are also related papers with applications to the operator 30 algebra of groupoids [Cho08,FMY05], and the papers [BM00,BM01] which apply infinite 31 graphs to the study of quasi-periodicity in solid state physics. However, the focus in these 32 papers is quite different from ours, as are the questions asked and the methods employed. 33 While periods and quasi-periods in graphs play a role in our present results, they enter 34 our picture in quite different ways, for example via spectra and metrics that we compute 35 from energy forms and associated Laplace operators. There does not seem to be a direct 36 comparison between our results and those of Guido et al. 37

1.3.3. Boundaries of graphs. There is also no shortage of papers studying boundaries of 38 inﬁnite graphs: [PW90, Saw97, Woe00] discuss the Martin boundary, [PW90, Woe00] also 39 describe the more geometrically constructed “graph ends”, and [Car72, Car73a, Car73b] 40 use unitary representations. There are also related resultsin[CdV99,CdV04] While there 41 are connections to our study, the scope is diﬀerent. 42 Martin boundary theory is really motivated by constructing a boundary for a Markov 43 process, and the geometry/topology of the boundary is rather abstract and a bit nebulous. 44 Additionally, one needs a Green’s function, and it must satisfy certain hypotheses before 45 the construction can proceed. Furthermore, the focus of Martin boundarytheory is the non- 46 negative harmonic functions. Our boundary construction is more general in that it applies 47 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 19

1 to any electrical network as in Definition 2.7 and it remains correct for all harmonic func- 2 tions of finite energy, including constant functions and harmonic functions which change 3 sign. However, it is also more restrictive in the sense that an electrical resistance network 4 may support functions which are bounded below but do not have finite energy. 5 We should also point out that our boundaryconstruction is related to, but different from, 6 the “graph ends” introduced by Freudenthal and others. The ends of a graph are the natural 7 discrete analogue of the ends of a minimal surface (usually assumed to be embedded in 3 8 R ), a notion which is closely related to the conformaltype of the surface. Starting with the 9 central book [Woe00] by Wolfgang Woess, the following references will provide the reader 10 with an introductionto the study of harmonic functions on infinite networks and the ends of 11 graphs and groups: [Woe86, Woe87, Woe89], and [Woe95] on Martin boundaries, [PW90] 12 on ends, [Woe96] on Dirichlet problems, [Woe97] on random walk. A comparison of the 13 examples in 14 and 13 illustrates that varying the resistances produces dramatic changes § § 14 in the topology of the boundary. 15 Our boundary essentially consists of infinite paths which can be distinguished by har- 16 monic functions; see 7.3 for details. It follows that transient networks with no nontrivial § 17 harmonic functions have exactly one boundary point (corresponding to the unique mono- d 18 pole). In particular, the integer lattices (Z , 1) have precisely 1 boundary point for d 3, 2 ≥ 19 and have 0 boundary points for d = 1, 2. The Martin boundary of (Z , 1) consists of two 2 20 points; similarly, (Z , 1) has two graph ends; cf. [PW90].

21 1.3.4. Measures and measure constructions. A reader glancing at our paper will notice a 22 number of incarnations of measures on infinite sample spaces: it may be a suitable space 23 of paths ( 11.1– 11.2 and 15.1) or an analogue of the Schwartz space of tempered distri- § § § 24 butions (section 7.2). The latter case relies on a construction of “Gel’fand triples” from § 25 mathematical physics. The reader may wonder why they face yet another measure con- 26 struction, but each construction is dictated by the problems we solve. Taking limits of finite 27 subsystems is a universal weapon used with great success in a variety of applications; we 28 use it here in the study of resistance distances on infinite graphs ( 5.2); boundaries, bound- § 29 ary representations for harmonic functions ( 7.2, 8.3, and 11.1– 11.2); and equilibrium § § § § 30 states and phase-transition problems in physics ( 15.1– 15.2). 2 § § 31 (1) as an L space. The central Hilbert space in this study, the energy space , HE HE 32 appears with a canonical reproducing kernel, but without any canonical basis, and there is 2 33 no obvious way to see as an L (X, µ) for some X and µ. Therefore, a major motivation HE 2 34 for our measure constructions is just to be able to work with as an L space. In 15.1, 2 HE § 35 we use a construction from probability to write = L (Ω, µ) in a way that makes the HE 36 energy kernel vx x G0 into a system of (commuting) random variables. Here, Ω is an { } ∈ 37 infinite Cartesian product of a chosen compact space S; one copy of S for each point 0 38 x G . In 15.2, we use a non-commutative version of this probability technology: rather ∈ § 39 than Cartesian products, we will use infinite tensor products of C∗-algebras , one for 0 A 40 each x G . The motivation here is an application to a problem in quantum statistical ∈ 41 mechanics. The “states” on the C∗-algebra of all observables are the quantum mechanical 42 analogues of probability measures in classical problems. Heuristically, the reader may 43 wish to think of them as non-commutative measures; see e.g., [BR97]. 44 (2) Boundary integral representation of harmonic functions. As it sometimes happens, 45 the path to bd G is somewhat circuitous: we begin with the discovery of an integral over the 46 boundary,which leads us to understand functions on the boundary, which in turn points the 47 way to a proper definition of the boundaryitself. A closely related motivation for a measure 48 is the formulation of an integral representation of harmonic functions u : ∈HE 20 PALLEE.T.JORGENSENANDERINP.J.PEARSE

u(x) = u(ξ)hx(ξ) dP(ξ) + u(o). (1.16) Z SG′ where hx = P armvx. Thus the focus of 7.2 is a formalization of the imprecise “Riemann 1 H ∂hx § sums” u(x) = u + u(o) of 4.3 as an integral of a bona fide measure. To carry this 2 bd G ∂Ò § out, we constructP a Gel’fand triple G G′ , where G is a dense subspace of and 3 S ⊆HE ⊆ S S HE G’ is its dual, but with respect to a strictly finer topology. We are then able to produce a 4 S 2 Gaussian probability measure P on G′ and isometrically embed into L ( G′ , P). In fact, 5 2 S HE S L ( G′ , P) is the second quantization of . However, the focus here is not on realizing 6 S 2 HE as an L space (or subspace), but in obtaining the boundary integral representation of 7 HE harmonic functions as in (1.16). Our aim is then to build formulasthat allow us to compute 8 values of harmonic functions u from an integral representation which yields u(x) as 9 ∈HE an integral over bd G ′ . Note that this integration in (1.16) is with respect to a measure 10 ⊆ SG depending on x just as in the Poisson and Martin representations. 11 (3) Concrete representation of the boundary. We would like to realize bd G as a measure 12 space defined on a set of well-understood elements; this is the focus of the constructions 13 in 7. The goal is a measure on the space of all infinite paths in G which yields the 14 § boundary bd G in such a way that G bd G is a compactification of G which is compatible 15 ∪ with the energy form and the Laplace operator ∆, and hence also the natural resistance 16 E metric on (G, c). This type of construction has been carried out with great success for the 17 case of bounded harmonic functions (e.g., Poisson representation and the Fatou-Primalov 18 theorem) and for nonnegative harmonic functions (e.g., Martin boundary theory), but our 19 scope of enquiry is the harmonic functions of finite energy. Finally, we would like to use 20 this Gaussian measure on to clarify bd G as a subspace of . Such a relationship is 21 SG′ SG′ a natural expectation, as the analogous thing occurs in the work of Poisson, Choquet, and 22 Martin. 23

1.4. General remarks. 24

Remark 1.2. Since we aim for several diﬀerent audiences (operator theory, analysis of 25 fractals, mathematical physics, etc), we have included more details in our exposition and 26 proofs than would otherwise be typical for a paper with a narrow focus. 27

Remark 1.3. Throughout the introductory discussion of electrical resistance networks in 28 2– 5, we discuss collections of real-valued functions on the vertices or edges of the graph 29 § § G. Such objects are most natural for the heuristics of the physical model, and addition- 30 ally allow for induced orientation/order and make certain probabilistic arguments possible. 31 However, in the latter portions of this paper, we need to incorporate complex-valued func- 32 tions into the discussion in order to make full use of spectral theory and other methods. 33

Remark 1.4. For the aid of the reader, we have included a list of symbolsand abbreviations 34 used in this document. Wherever possible, we have attempted to ensure that each symbol 35 has only one meaning. In cases of overlap, the context should make things clear. In 36 Appendix C, we also include some diagrams which we hope clarify the properties of the 37 many operators and spaces we discuss, and the relations between them. 38

1.5. Acknowledgements. 39 While working on the project, the co-authors have beneﬁtted from interaction with col- 40 leagues and students. We thank everyone for generously suggesting improvements as our 41 paper progressed. The authors are grateful for stimulating comments, helpful advice, and 42 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 21

1 valuable references from John Benedetto, Il-Woo Cho, Raul Curto, Dorin Dutkay, Alexan- 2 der Grigor’yan, Dirk Hundertmark, Richard Kadison, Keri Kornelson, Michel Lapidus, 3 Russell Lyons, Diego Moreira, Peter Mörters, Paul Muhly, Bob Powers, Marc Rieffel, 4 Karen Shuman, Sergei Silvestrov, Jon Simon, Myung-Sin Song, Bob Strichartz, Andras 5 Telcs, Sasha Teplyaev, Ivan Veselic, Lihe Wang, Wolfgang Woess, and Qi Zhang. The 6 authors are particularly grateful to Russell Lyons for several key references and examples, 7 and to Jun Kigami for several illuminating conversations and for suggesting the approach 8 in (5.41). Initially, the first named author (PJ) learned of discrete potential theory from 9 Robert T. Powers at the University of Pennsylvania in the 1970s, but interest in the subject 10 has grown exponentially since. 22 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2. Electrical resistance networks 1

In this section, we introduce the mathematical model of an electrical resistance network 2 (ERN) as a graph G whose edges are understood as resistors and whose vertices are the 3 nodes at which these resistors are connected. Thus, the resistance data is specified by a 4 function Ω, so that Ω(x, y) is the resistance of the edge (resistor) between the vertices x 5 and y. With the network data (G, Ω) fixed, we begin the study of functions defined on the 6 vertices. We define many basic terms and conceptsused throughoutthe paper, including the 7 Dirichlet energy form and the Laplace operator ∆. Additionally, we prove a key identity 8 E relating to ∆ for finite graphs: Lemma 2.13. In Theorem 4.41, this will be extended to 9 E infinite graphs, in which case it is a discrete analogue of the familiar Gauss-Green identity 10 from vector calculus. The appearance of a somewhat mysterious boundary term in the 11 Theorem 4.41 prompts several questions which are discussed in Remark 4.7. Answering 12 these questions comprises a large part of the sequel; cf. 7. In fact, Theorem 4.41 provides 13 § much of the motivation for energy-centric approach we pursue throughout our study; the 14 reader may wish to look ahead to Remark 4.66 for a preview. 15

2.1. The electrical resistance network model. This section contains the basic deﬁnitions 16 used throughout the sequel. 17

0 1 0 Deﬁnition 2.1. A graph G = G , G is given by the set of vertices G and the set of edges 18 1 0 0 { } 1 G G G . Two vertices are neighbours (or are adjacent) iﬀ there is an edge (x, y) G 19 ⊆ × ∈ 1 connecting them, and this is denoted x y. This relation is symmetric, as (y, x) G 20 1 ∼ 0 ∈ whenever (x, y) G . The set of neighboursof x G is 21 ∈ ∈ 0 . G(x) = y G .. y x . (2.1) { ∈ ∼ }

In our context, the set of edges of G will be determined by the conductance function, so 22 that all graph data is implicitly provided by c. 23

Deﬁnition 2.2. The conductance cxy is a symmetric function 24

c : G0 G0 [0, ), (2.2) × → ∞ in the sense that cxy = cyx. It is our convention that x / y if and only if cxy = 0; that is, 25 1 ∼ there is an edge (x, y) G if and only if 0 < c(x, y) < . 26 ∈ ∞ Conductance is the reciprocal of resistance, and this is the origin of the name “resistance 27 1 network”. It is important to note that c−xy gives the resistance between adjacent vertices; 28 1 this feature distinguishes c−xy from the eﬀective resistance R(x, y) discussed later, for which 29 x and y need not be adjacent. 30

0 Deﬁnition 2.3. The conductances deﬁne a measure or weighting on G by 31

c(x) := cxy. (2.3) Xy x ∼ 0 Whenever G is connected, it follows that c(x) > 0, for all x G . The notation c will also 32 ∈ be used, on occasion, to indicate the multiplication operator (cv)(x) := c(x)v(x). 33

0 0 Deﬁnition 2.4. A path γ from α G to ω G is a sequence of adjacent vertices 34 ∈ ∈ (α = x0, x1, x2,..., xn = ω), i.e., xi xi 1 for i = 1,..., n. The path is simple if any vertex 35 ∼ − appears at most once (so that a path is simply connected). 36 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 23

0 1 Definition 2.5. A graph G is connected iff for any pair of vertices α, ω G , there exists ∈ 2 a finite path γ from α to ω.

3 Remark 2.6. Note that for resistors connected in series, the resistances just add, so this 4 condition implies there is a path of ﬁnite resistance between any two points. We emphasize 5 that all graphs and subgraphs considered in this study are connected.

6 At this point, the reader may wish to peruse some of the examples of 12. § 7 Definition 2.7. An electrical resistance network (ERN) is a connected graph (G, c) whose 0 8 conductance function satisfies c(x) < for every x G . We interpret the edges as being ∞ ∈ 9 defined by the conductance: x y iff c > 0. ∼ xy 10 Note that c need not be bounded in Definition 2.7. Also, we will typically assume an 11 ERN to be simple in the sense that there are no self-loops, and there is at most one edge 12 from x to y. This is mostly for convenience: basic electrical theory says that two conductors 1 2 13 cxy and cxy connected in parallel can be replaced by a single conductor with conductance 1 2 14 cxy = cxy + cxy. Also, electric current will never flow along a conductor connecting a node 15 to itself. Nonetheless, such self-loops may be useful for technical considerations: one can 16 remove the periodicity of a random walk by allowing self-loops. This can allow one to 17 obtain a “lazy walk” which is ergodic, and hence amenable to application of tools like the 18 Perron-Frobenius Theorem. See, for example, [LPW08, LP09]. 19 We will be interested in certain operators that act on functions defined on electrical 20 resistance networks.

21 Definition 2.8. The Laplacian on G is the linear difference operator which acts on a func- 0 22 tion v : G R by → (∆v)(x) := c (v(x) v(y)). (2.4) xy − Xy x 23 ∼ 0 24 A function v : G R is called harmonic iff ∆v 0. → ≡ 25 Definition 2.9. The transfer operator on G is the linear operatorT which acts on a function 0 26 v : G R by →

(T v)(x) := cxyv(y). (2.5) Xy x ∼ 27 Hence, the Laplacian may be written ∆= c T, where (cv)(x) := c(x)v(x). − 28 We won’t worry about the domain of ∆ or T until 8. For now, consider both of these 0 § 29 operators as defined on any function v : G R. The reader familiar with the literature → 30 will note that the definitions of the Laplacian and transfer operator given here are normal- 31 ized differently than may be found elsewhere in the literature. For example, [DS84] and 32 other probabilistic references use

1 1 ∆ := c− ∆= 1 P, so (∆ v)(x) := c (v(x) v(y)), (2.6) c − c c(x) xy − Xy x ∼ 1 33 where P := c− T is the probabilistic transition operator corresponding to the transition 34 probabilities p(x, y) = cxy/c(x). For another example, [Chu01] and other spectral-theoretic 35 references use 24 PALLEE.T.JORGENSENANDERINP.J.PEARSE

1/2 1/2 1/2 1/2 cxyv(y) ∆s := c− ∆c− = 1 c− T c− , so (∆sv)(x) := v(x) . (2.7) − − Xy x c(y) ∼ p 1 However, these renormalized version are much more awkward to work with in the 2 present context; especially when dealing with the inner product and kernels of the Hilbert 3 spaces we shall study. Not only are (2.4) and (2.5) are better suited to the electrical resis- 4 tance network framework (as will be evinced by the operator theory developed in 4 and 5 § succeeding sections) but both ∆c and ∆s are bounded operators, and hence do not allow for 6 the delicate spectral analysis carried out in 7– sec:Lap-on-HE. 7 § §

2.2. The energy. In this section we study the relation between the energy and Laplacian 8 E ∆ on finite networks, as expressed in Lemma 2.13. This formula will be used prolifically, 9 as it also holds on infinite networks in many circumstances. In fact, a noticeable portion of 10 4 is devoted to determining when this is so. 11 § Definition 2.10. The graph energy of an electrical resistance network is the quadratic form 12 0 defined for functions u : G R by 13 → 1 (u) := c (u(x) u(y))2. (2.8) E 2 xy − x,Xy G0 ∈ There is also the associated bilinear energy form 14

1 (u, v) := c (u(x) u(y))(v(x) v(y)). (2.9) E 2 xy − − x,Xy G0 ∈ For both (2.8) and(2.9), note that cxy = 0 for vertices which are not neighbours, and hence 15 1 only pairs for which x y contribute to the sum; the normalizing factor of corresponds 16 ∼ 2 to the idea that each edge should only be counted once. The domain of the energy is 17

0 . dom = u : G R .. (u) < . (2.10) E { → E ∞}

The close relationship between the energy and the conductances is highlighted by the 18 simple identities 19

(δ ) = c(x), and (δ ,δ ) = c , (2.11) E x E x y − xy 0 where δ is a (unit)Dirac mass at x G . Theeasy proofis left as an exercise. A signiﬁcant 20 x ∈ upshot of (2.11) is that the Dirac masses are not orthogonal with respect to energy. 21

Remark 2.11. It is immediatefrom (2.8) that (u) = 0ifandonlyif u is a constant function. 22 E The energy form is positive semidefinite, but if we work modulo constant functions, it 23 becomes positive definite and hence an inner product. We formalize this in Definition 4.1 24 and again in 6.1. In classical potential theory (or Sobolev theory), this would amount to 25 § working with the class of functions satisfying f < , but abandoning the requirement 26 k ′k2 ∞ that f < . As a result of this, the nontrivial harmonic functions play an especially 27 k k2 ∞ important role in this paper. In particular, it is precisely the presence of nontrivial harmonic 28 functions which prevents the functions of finite support from being dense in the space of 29 functions of finite energy; see 4.2. 30 § OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 25

1 Traditionally (e.g., [Kat95,FOT94¯ ]) the study of quadratic forms would combine (u, v) E 2 and u, v 2 . In our context, this is counterproductive, and would eclipse some of our h iℓ 3 most interesting results. Some of our most intriguing questions for elements v 2 0 ∈ HE 4 involve boundary considerations, and in these cases v is not in ℓ (G ) (Corollary 4.65). 5 One example of this arises in the discrete Gauss-Green formula (Theorem 4.41); another 6 arises in study of forward-harmonic functions in 11.2. § 7 The following proposition may be found in [Str06, 1.3]or[Kig01, Ch. 2], for example. § 8 Proposition 2.12. The following properties are readily veriﬁed:

9 (1) (u, u) = (u). E E 1 10 (2) (Polarization) (u, v) = [ (u + v) (u v)]. E 4 E −E − 11 (3) (Markov property) ([u]) (u), where [u] is any contraction of u. E ≤E 12 For example, let [u] := min 1, max 0, u . The following result relates the Laplacian to { { }} 13 the graph energy on ﬁnite networks, and can be interpreted as a relation between dom 2 0 E 14 and ℓ (G ).

15 Lemma 2.13. Let G be a ﬁnite electrical resistance network. For u, v dom , ∈ E (u, v) = u(x)∆v(x) = v(x)∆u(x). (2.12) E xXG0 xXG0 ∈ ∈ 16 Proof. Direct computation yields

1 (u, v) = c u(x)v(x) u(x)v(y) u(y)v(x) + u(y)v(y) E 2 xy − − x,Xy G0 ∈ 1 1 = c(x)u(x)v(x) + c(y)u(y)v(y) 2 2 xXG0 yXG0 ∈ ∈ 1 1 u(x)T v(x) u(y)T v(y) − 2 − 2 xXG0 yXG0 ∈ n ∈ = c(x)u(x)v(x) u(x)T v(x) − xXG0 xX,y G0 ∈ ∈ = u(x) (c(x)v(x) T v(x)) − xXG0 ∈ = u(x)∆v(x). (2.13) xXG0 ∈ 17 Of course, the computation is identical for x G0 v(x)∆u(x). ∈ P 18 We include the following well-known result for completeness.

19 Corollary 2.14. On a ﬁnite electrical resistance network, all harmonic functions of ﬁnite 20 energy are constant.

21 Proof. If h is harmonic, then (h) = x G0 h(x)∆h(x) = 0. See Remark 2.11. E ∈ P 22 Connectedness is implicit in the calculations of both Lemma 2.13 and Corollary 2.14; 23 recall that all electrical resistance networks considered in this work are connected. We 24 will extend Lemma 2.13 to inﬁnite graphs in Theorem 4.41, where the formula is more 25 complicated: 26 PALLEE.T.JORGENSENANDERINP.J.PEARSE

(u, v) = u(x)∆v(x) + “boundary term” . E { } xXG0 ∈ It is shown in Theorem 4.47 that the presence of the boundary term corresponds to the 1 existence of nontrivial harmonic functions, in contrast to Corollary 2.14. In fact, one can 2 interpret Corollary 2.14 as the reason why the boundary term alluded to above vanishes on 3 ﬁnite networks. We study the interplay between and ∆ further in 9.2– 9.3. 4 E § § OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 27

1 3. Currents and potentials on electrical resistance networks

2 The potential theory for an electrical resistance network is studied via an experiment 3 in which 1 amp of current is passed through the network, inserted into one vertex and 4 extracted at some other vertex. The voltage drop measured between the two nodes is the 5 effective resistance between them, see 5. § 6 When the voltages are fixed at certain vertices, it induces a current in the network in 7 accordance with the laws of Kirchhoff and Ohm. This induced current is introduced for- 8 mally in Definition 3.17. Induced currents are important for studying flows of minimal 9 dissipation, and will also be useful in the study of forward-harmonic functions in 11.2. If § 10 a voltage drop of 1 volt is imposed between two vertices, the effective resistance between 11 these two vertices is the reciprocal of the dissipation of the induced current. 12 In Theorem 3.28 we show that there always exists an harmonic function satisfying the 13 boundary conditions implied by the above described experiment, in order to fill a gap 14 in [Pow76b]. In Theorem 3.26 and Theorem 3.26 it is shown that these harmonic functions 15 correspond to currents which minimize energy dissipation.

16 3.1. Currents on electrical resistance networks. 0 0 17 Definition 3.1. A current is a skew-symmetric function I : G G R. × → 18 Definition 3.2. An orientation is a subset of the edges which includes exactly one of each 19 pair (x, y), (y, x) . For a given current I, one may pick an orientation by requiring that { } 20 I(x, y) > 0 on every edge for which I is nonzero, and arbitrarily choosing (x, y) or(y, x) 21 outside the support of I. We refer to this as an orientation induced by the current; this will 22 be used extensively in 11.2 to study the forward-harmonic functions. § 0 23 The energyis a functionaldefined on functions v : G R which give voltages between → 24 different vertices in the network. The associated notion defined on the edges of the network 25 is the dissipation of a current.

26 Deﬁnition 3.3. The dissipation of a current may be thought of as the energy lost as a 1 27 current ﬂows through an electrical resistance network. More precisely, for I, I , I : G 1 2 → 28 R,

1 1 2 D(I) := c− I(x, y) . (3.1) 2 xy (x,Xy) G1 ∈ 29 The associated bilinear form is the dissipation form:

1 1 D(I , I ) := c− I (x, y)I (x, y). (3.2) 1 2 2 xy 1 2 (x,Xy) G1 ∈ 1 30 Again, the normalizing factor of 2 corresponds to the idea that each edge only contributes 31 once to the sum. The domain of the dissipation is

. dom D := I .. D(I) < . (3.3) { ∞} 32 Remark 3.4. When an orientation for G is chosen,it is easy to see that dom D is a Hilbert O 2 33 space under the inner product (3.2). Indeed, dom D = ℓ ( , c). O 34 Deﬁnition 3.5. A cycle ϑ is a set of n edges corresponding to a sequence of vertices 0 1 35 (x , x , x ,..., x = x ) G , for which (x , x + ) G for each k. Denote the set of 0 1 2 n 0 ⊆ k k 1 ⊆ 36 cycles in G by . L 28 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Definition 3.6. For physical realism, we often require that a current flow satisfy Kirch- 1 hoff’s node law, i.e., that the total current flowing into a vertex must equal the total current 2 flowing out of a vertex: 3

I(x, y) = 0, x G0. (3.4) ∀ ∈ Xy x ∼ This is indeed the version of Kirchhoff’s law you would find in a physics textbook; with 4 our convention I(x, y) > 0 indicates that the current flows from x to y. 5 However, if we are performing the experiment described above, then there are boundary 6 conditions at α, ω to take into account, and Kirchhoff’s node law takes the nonhomoge- 7 neous form 8

1, x = α,

I(x, y) = δα δω =  1, x = ω, (3.5) −  Xy x − ∼ 0, else,  0  where δx is the usual Dirac mass at x G .  9 ∈  Definition 3.7. A current flow from α to ω is a current I dom D that satisfies (3.5). The 10 ∈ set of all current flows is denoted (α, ω). 11 F We usually use α to denote the beginning of a flow and ω to denote its end. Shortly, we 12 will see that the currents corresponding to potentials are precisely the current flows. 13

Remark 3.8. Although trivial, it is important to note that the characteristic function of a 14 1 current path χ : G 0, 1 trivially satisfies (3.5). Also, the characteristic function 15 γ → { } of a cycle satisfies (3.4) in much the same way. As a consequence, if I (α, ω), then 16 ∈ F I + tχ (α, ω) for any t R by a brief computation. In other words, perturbation 17 ϑ ∈ F ∈ on a cycle preserves the Kirchhoff condition. However, the dissipation will vary because 18 χ D( ϑ) > 0. 19

3.2. Potential functions and their relationship to current ﬂows. From the proceeding 20 0 section, it is clear that a special role is played by functions v : G R which satisfy 21 → the equation ∆v = δα δω. Such a function is the solution to a discrete Dirichlet problem, 22 − where the “boundary”has been chosen to be α and ω (not to be confused with the boundary 23 term discussed in Remark 4.7). 24

Deﬁnition 3.9. A dipole is a function v dom which satisﬁes 25 ∈ E

∆v = δα δω (3.6) 0 − for some vertices α, ω G . The collection of all such functions is denoted (α, ω). 26 ∈ P Note that when G is ﬁnite, (α, ω) contains only a single element. This follows from 27 P Corollary 2.14 because the diﬀerence of any two solutions to (3.6) is harmonic. 28

Remark 3.10. The definition of a monopole that we give here is a heuristic definition; we 29 0 give the precise definition in Definition 4.34. A monopole at ω is a function w : G R 30 → which satisfies 31

∆w = kδω, w dom , k C. (3.7) ∈ E ∈ In the sequel, we are primarily concerned with monopoles wo, where o = ω is some fixed 32 vertex which acts as a point of reference or “origin”. Also, we typically take k = 1, as the 33 − induced current of such a monopole is a unit flow to infinity in the language of [DS84]. 34 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 29

1 Remark 3.11. The study of dipoles, monopoles, and harmonic functions is a recurring 2 theme of this paper:

∆v = δα δω, ∆w = δω, ∆h = 0. − − 3 In Theorem 3.28, we will show that (α, ω), is nonempty for any α and ω, on any network P 4 (G, c); the existence of monopoles and nontrivial harmonic functionsis a much more subtle 5 issue. 6 In Corollary 4.21,weoffer a more refinedproofof the existence of dipoles, using Hilbert 7 space techniques. Perhaps a more interesting question is when (α, ω) contains more than P 8 element; the linearity of ∆ shows immediately that any two dipoles in (α, ω) differ by a P 9 harmonicfunction. We have shown that when a connected graph is finite the only harmonic 10 functions are constant (Corollary 2.14), and therefore (α, ω) consists only of a single P 11 function, up to the addition of a constant. The situation for monopoles is similar, as the 12 difference of two monopoles at ω is also a harmonic function. 13 Not all electrical resistance networks support monopoles; the current induced by a 14 monopole is a finite flow to infinity and hence indicates that the random walk on the net- 15 work is transient, by [Lyo83]. See also [DS84, LP09] for terminology and proofs. It is 16 well-known that for a reversible Markov chain, if the random walk started at one vertex is 17 transient, then it is transient when started at any vertex. We give a very brief proof of this 18 in Lemma 3.29; and a new criterion for transience in Lemma 4.51. 19 On some networks, a monopole can be understood as the limit of a sequence of dipoles 20 v where ∆v = δ δ and x . In such a situation, a monopole can be considered xn xn xn − o n →∞ 21 as a dipole where one of the Dirac masses “sits at ”. However, this is not possible on ∞ 22 all networks, as is illustrated by the binary tree in Example 13.4. Again, the linearity of ∆ 23 shows immediately that any two monopolesat ω differ by a harmonicfunction. When these 24 monopoles correspond to a “distribution of dipoles at infinity” (i.e., a limit of sums axvx 25 where the v ’s are dipoles with x in the limit), the addition of a harmonic function x → ∞ P 26 transforms the distribution at infinity. It will take some work to make these ideas precise; 27 for now the reader can consider this remark simply as a preview of coming attractions. The 28 presence of monopoles is also extremely closely related to the existence of “long-range 3 29 order”, and the theoretical foundation of magnetism in R ; see 15.3. § 30 Furthermore, it is possible for an electrical resistance network to support monopoles but d 31 not nontrivial harmonic functions. In 14, we show that the integer lattice networks (Z , 1) § 32 support monopoles (Theorem 14.5). However, all harmonic functions are linear and hence 33 do not have finite energy; cf. Theorem 14.17. Both of these results are well-known in the 34 literature in different contexts, and/or with different terminology.

35 Lemma 3.12. The dipoles (α, ω) and the current ﬂows (α, ω) are convex sets. Fur- P F 36 thermore, if v (α, ω), then v + h (α, ω) for any harmonic function h; similarly, if ∈ P ∈ P 37 I (α, ω), then I + J (α, ω) for any function J satisfying (3.4). ∈F ∈F 38 Proof. If v (α, ω), c 0 and c = 1, then the linearity of ∆ gives i ∈P i ≥ i P

∆ civi = ci∆vi = ci(δα δω) = δα δω. X X X − − 39 The computation for the other parts is similar.

40 Theorem 3.13. obtains its minimum for some unique v (α, ω), and D obtains its E ∈ P 41 minimum for some unique I (α, ω). ∈F 30 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Proof. Each of these is a quadratic form on a convex set, by Lemma 3.12, so the result 1 is an immediate application of [Rud87, Thm. 4.10] or [Nel69], e.g. To underscore the 2 . uniqueness, suppose that (v ) = (v ). Then with ε := inf (v) .. v (α, ω) , the 3 E 1 E 2 {E ∈ P } parallelogram law gives 4

(v v ) = 2 (v ) + 2 (v ) 4ε2 = 0, E 1 − 2 E 1 E 2 − since (v ) = ε because v were chosen to be minimal. 5 E i i Deﬁnition 3.14. Ohm’s Law (V = RI) appears in the present context as 6 1 v(x) v(y) = I(x, y). (3.8) − cxy

Remark 3.15. It will shortly become evident (if it isn’t already) that current flows satis- 7 fying Kirchhoff’s law correspond to harmonic functions via Ohm’s law and that current 8 flows satisfying the nonhomogeneous Kirchhoff’s law (3.5) correspond to dipoles, that is, 9 solutions of the Dirichlet problem (3.6) with Neumann boundary conditions. To make this 10 precise, we need the notion of induced current given in Definition 3.17 and justified by 11 Lemma 3.16. 12 0 Lemma 3.16. Every function v : G R induces a unique current via I(x, y) := c (v(x) 13 → xy − v(y)), and the dissipation of this current is the energy of v: 14

D(I) = (v). (3.9) E Moreover, if v (α, ω), then I (α, ω). 15 ∈P ∈F Proof. It is clear that Ohm’s Law deﬁnes a current. The equality (3.9) is a very brief 16 calculation and follows straight from the deﬁnitions; see (2.9) and (3.1). A proofof (3.9) 17 is also given in [DS84]. 18 If v (α, ω), then ∆v = δα δω and 19 ∈P −

(δα δω)(x) = (∆v)(x) = c (v(x) v(y)) = I(x, y) (3.10) − xy − Xy x Xy x ∼ ∼ veriﬁes the nonhomogeous form of Kirchhoﬀ’s law. 20

Deﬁnition 3.17. Given v (α, ω), the induced current is deﬁned via Ohm’s Law as in 21 ∈ P the statement of Lemma 3.16. That is, 22

I(x, y) := c (v(x) v(y)). (3.11) xy − Remark 3.18. Note that (3.9) holds when the current I is induced by v. It makes no sense 23 to attempt to apply the same equality to a general current: there may be NO associated 24 potential because of the compatibility problem described just below. Nonetheless, Theo- 25 rem 10.27 provides a way to give the identity analogous to (3.9) for general currents by 26 using the adjoint of the operator implicit in (3.11). 27

Remark 3.19. If ∆v = δα δω has a solution v , then any other solution is of the form 28 − 0 v = v0 + h where h is harmonic, by linearity of ∆. So to minimize energy, one must 29 consider such perturbations: 30

d [ (v + th)] = = 0 (v , h) = 0. dt E 0 t 0 ⇐⇒ E 0 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 31

x y

Figure 1. A Dirac mass on an edge of Z2.

1 Conversely, if (v, h) = 0, then E

(v + th) = (v) + 2t (v, h) + t2 (h) (v), E E E E ≥E 2 shows that energy is minimized for t = 0. In particular, energy is minimized for v which 3 contains no harmonic component. In Lemma 4.22 this important principle is restated in 4 the language of Hilbert spaces: energy is minimized for the v which is orthogonal to the 5 space of harmonic functions with respect to . E 6 Analogous remarks hold for I which minimizes D(I). However, note that Kirchhoﬀ’s 7 Law is blind to conductances and so I (α, ω) does not imply that D(I) is minimal. In ∈ F 8 the next section, we show that induced currents are minimal with respect to D when they 9 are induced by a minimal potential v.

10 3.3. The compatibility problem. The converse to Lemma 3.16 is not always true, but a 11 partial converse is given by Theorem 3.26. Given an electrical resistance network (G, c), 12 one can always attempt to construct a Ohm’s function by fixing the value v(x0) at some 0 13 point x G , and then applying Ohm’s law to determine the value of v for other vertices 0 ∈ 14 x x . However, this attempt can fail if the network contains a cycle (see Example 12.2 for ∼ 0 15 an example) because the existence of a cycle is equivalent to the existence of two distinct 16 paths from one point to another. This phenomenonis worked out in detail for a simple case 17 in Example 3.20. 18 In general, it may happen that there are two different paths from x0 to y0, and the net 19 voltage drop v(x ) v(y ) computedalong these two paths is not equal. Such a phenomenon 0 − 0 20 makes it impossible to define v. Note that Kirchhoff’s law does not forbid this, because 21 (3.4)–(3.5) is expressed without reference to the conductances c. We refer to this as the 22 compatibility problem: a general current function may not correspond to a potential, 23 even though every potential induces a well defined current flow (see Lemma 3.16). In this 24 section we provide the following answer: for any current, there exists a unique associated 25 current which does correspond to a potential.

26 Example 3.20 (The Dirac mass on an edge). Consider a Dirac mass on an edge of the 2 27 network (Z , 1) as depicted in Figure 1. We use such a current here to illustrate the com- 28 patibility problem. To find a potential corresponding to this current, consider the following 29 dilemma: I(x, y) = 1 and I 0 elsewhere corresponds to a potential (up to a constant) ≡ 30 which has v(x) = 1 and v(y) = 0, as in Figure 2. Since I(x, w) = 0, we have v(w) = v(x), 31 and since I(y, z) = 0, we have v(z) = v(y). But then v(z) = 1 , 0 = v(z), contradicting the < 32 fact that I(w, z) = 0! ւ χ 33 Definition 3.21. A current I satisfies the cycle condition iff D(I, ϑ) = 0 for every cycle 34 ϑ . (We call χ a cycle.) ∈ L ϑ 32 PALLEE.T.JORGENSENANDERINP.J.PEARSE

x w y z

Figure 2. A failed attempt at constructing a potential to match Figure 1.

Remark 3.22. From the preceding discussion, it is clear that for a current satisfying the 1 cycle condition, voltage drop between vertices x and y may be measured by summing the 2 currents along any single path from x to y, and the result will be independentof which path 3 was chosen. In the Hilbert space interpretation of 10 the cycle condition is restated as “I 4 § is orthogonal to cycles”. The next two results must be folklore (perhaps dating back to the 5 th 19 century?) but we include them for their relevance in 10, especially the Hilbert space 6 § decomposition of Theorem 10.8 (see also Figure 4). While writing a second draft of this 7 document, the authors discovered a similar treatment in [LP09, 9]. 8 § Lemma 3.23. I is an induced current if and only if I satisﬁes the cycle condition. 9

Proof. ( ) If I is induced by v, then for any ϑ , the sum 10 ⇒ ∈ L 1 I(x, y) = (v(x) v(y)) = 0, (3.12) c − (xX,y) ϑ xy (xX,y) ϑ ∈ ∈ χ since every term v(xi) appears twice, once positive and once negative, whence D(I, ϑ) = 0. 11 ( ) Conversely, to prove that there is such a v, we must show that v(x ) v(y ) is 12 ⇐ 0 − 0 independence of the path from x0 to y0 used to compute it. In a direct analogy to basic 13 vector calculus, this is equivalent to the fact that the net voltage drop around any closed 14 cycle is 0. 15

1 (v(x) v(y)) = I(x, y) = D(I, χ ) = 0, − c ϑ (xX,y) ϑ (xX,y) ϑ xy ∈ ∈ 0 Now deﬁne v by ﬁxing v(x0) for some point x0 G , and then coherently use v(x) v(y) = 16 1 ∈ − I(x, y) to compute v at any other point. 17 cxy

The presence of cycles is not always obvious! As an exercise, we invite the reader to 18 determine the cycles involved in Example 3.20. 19

Lemma 3.24 (Resurrection of Kirchhoff’s Law). Let I be the current induced by v. Then 20 v (α, ω) if and only if I satisfies the nonhomogeneous Kirchhoff’s law. 21 ∈P Proof. ( ) Computing directly, 22 ⇒

I(x, y) = c (v(x) v(y)) = ∆v(x) = δα δω. (3.13) xy − − Xy x Xy x ∼ ∼ ( ) Conversely, to show ∆v = δα δω, 23 ⇐ −

∆v(x) = c (v(x) v(y)) defof ∆ xy − Xy x ∼ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 33

= I(x, y)(3.8) Xy x ∼ = δα δω, I (α, ω). − ∈F

1 Corollary 3.25. Let I be the current induced by v. Then v is harmonic if and only if I 2 satisﬁes the homogeneous Kirchhoﬀ’s law.

3 Proof. Mutatis mutandis, this is the same as the proof of Lemma 3.24.

4 Theorem 3.26. I minimizes D on (α, ω) if and only if I is induced by a potential v that F 5 lies in (α, ω). Moreover, v also minimizes over (α, ω). P E P 6 Proof. ( ) Since I minimizes D, we have ⇒ d [D(I + tJ)] = 0, (3.14) dt t=0 7 for any current J satisfying the homogeneous Kirchhoff’s law. From Remark 3.8, this χ 8 applies in particular to J = ϑ, where ϑ is any cycle in . d L 9 Note that D(I, χ ) = 0 iff D(I + tχ ) = 0. To see this, replace I by I + tχ in (3.1), ϑ dt ϑ t=0 ϑ χ 10 differentiate D(I + t ϑ) term-by-termh withi respect to t and evaluates at t = 0 to obtain that 11 (3.14) is equivalent to

1 1 χ I(x, y) ϑ(x, y) = I(x, y) = 0, ϑ . cxy cxy ∀ ∈ L (x,Xy) G1 (xX,y) ϑ ∈ ∈ 12 By Lemma 3.23, this shows that I is induced by some v; and by Lemma 3.24, we know 13 v (α, ω). From Lemma 3.16, it is clear that the v must also be the energy-minimizing ∈ P 14 element of (α, ω). P 15 ( ) Since I is induced by v (α, ω), the only thing we need to check is that I is min- ⇐ ∈P 16 imal with respect to any harmonic current (i.e. a current induced by a harmonic function); 17 this follows from Lemma 3.23 and the ﬁrst part of the proof. If h is any harmonic function 0 18 on G , denote the induced current by H as before. Then Lemma 3.16 gives

d d [D(I + tH)] = [ (v + th)] = 0, dt t=0 dt E t=0 19 by the minimality of v.

20 Remark 3.27. Part of the motivation for Theorem 3.26 is to ﬁx an error in [Pow76b]. The 21 author was not apparently aware of the possibility of nontrivial harmonic functions, and 22 hence did not see the need for taking the element of (α, ω) with minimal energy. This P 23 becomes especially important in Theorem 5.2. 24 Theorem 3.26 is generalized in Theorem 10.27 where we exploit certain operators to 25 obtain, for any given current I, an associated minimal current. This minimal current is 26 induced by a potential, even if the original is not, and provides a resolution to the compat- 27 ibility problem described at the beginning of 3.3, just above. § 28 In 10.4.1 we revisit this scenario and show how the minimal current may be obtained § 29 by the simple application of a certain operator, once it has been properly interpreted in 30 terms of Hilbert space theory. See Theorem 10.27 and its corollaries in particular.

31 Theorem 3.28. (α, ω) is never empty. P 34 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Proof. It is clear that (α, ω) , ∅ because one always has the characteristic function of 1 F a current path from α to ω (since we are assuming the underlying graph is connected); 2 see Definition 11.12 and Remark 3.8. From Theorem 3.13 one sees that there is always a 3 flow which minimizes dissipation. By Theorem 3.26, this minimal flow is induced by an 4 element of (α, ω). 5 P The following result is well-known in probability (see, e.g., [Str05]), but we include it 6 here for completeness and the novel method of proof. 7 0 Corollary 3.29. If the random walk on (G, c) is transient when started from y G , then 8 0 ∈ it is transient when started from any x G . 9 ∈ Proof. By [Lyo83], the hypothesis means there is a monopole w dom with ∆w = δ . 10 y ∈ E y But then by Theorem 3.28 and the linearity of ∆, v + wx is a monopole at x, for any 11 v (x, y). 12 ∈P Remark 3.30. Theorem 3.28 fills a gap in [Pow76b]. A key point is that the finite dissipa- 13 tion of the flow ensures the finite energy of the inducing voltage function, by Lemma 3.16. 14 A different proof of Theorem 3.28 is obtained in Corollary 4.21 by the application of 15 Hilbert space techniques. 16 Theorem 3.28 also follows from results of [Soa94, III.4] since the difference of two 17 § Dirac masses corresponds to a “balanced” flow, i.e., the same amount of current flows in 18 as flows out. 19 2 0 Remark 3.31. There are examples for which the elements of (α, ω) do not lie in ℓ (G ); 20 P see Figure 10 of Example 14.2. 21 0 Proposition 3.32. IfG is finite and v (α, ω), then v(ω) v(x) v(α) for all x G . 22 ∈P ≤ ≤ ∈ Proof. This is immediate from the maximum principle for harmonic functions on the finite 23 0 set G with boundary α, ω . See[LP09, 2.1], for example, or [LPW08]. 24 { } § Remark 3.33. In 5, we will see that Proposition 3.32 extends to a more general result: if v 25 § is the unique element of (α, ω) of minimal energy, then the same conclusion follows. One 26 P way to see this is to define u(x) = Px[τα < τω] (i.e., the probability that the random walk 27 W started at x reaches α before ω). By Theorem 5.18, v is defined by v(x) = u(x)R (α, ω), 28 W where R (α, ω) is the (wired) effective resistance between α and ω. 29 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 35

1 4. The energy Hilbert space

2 In this section, we study the Hilbert space of voltage functions, where the inner HE 3 product is given by the energy form. A key feature of is the presence of a family of HE 4 dipoles vx x G0 indexed by the vertices. In Corollary 4.13 we show how this family of { } ∈ 5 dipoles forms a reproducing kernel for in the sense of Aronszajn. HE 6 In Theorem 4.41, we establish a discrete version of the Gauss-Green Formula which 7 extends Lemma 2.13 to the case of infinite graphs. The appearance of a somewhat mys- 8 terious boundary term prompts several questions which are discussed in Remark 4.7. An- 9 swering these questions comprises a large part of the sequel; cf. 7. We are able to prove § 10 in Lemma 4.68 that this boundary term vanishes for finitely supported functions on G, and 2 0 11 in Corollary 4.65 that nontrivial harmonic functions cannot be in ℓ (G ). This is discussed 12 further in Remark 4.66 and provides the motivation for energy-centric approach we pursue 13 throughout our study. 14 The energy Hilbert space will facilitate our study of the resistance metric R in 5. HE § 15 In particular, it provides an explanation for an issue stemming from the “nonuniqueness of 16 currents” in certain infinite networks; see [LP09,Tho90]. This disparity leads to differences 17 between two apparently natural extensions of the effective resistance to infinite networks, 18 which are greatly clarified by the geometry of Hilbert space. Also, presents an analytic HE 19 formulation of the type problem for random walks on an electrical resistance network: 20 transience of the random walk is equivalent to the existence of monopoles, that is, finite- 21 energy solutions to a certain Dirichlet problem. In fact, this approach will readily allow us 22 to obtain explicit formulas for effective resistance on integer lattice networks in 14, with § 23 applications to a physics problem of [Pow76b] in 15. § 24 Definition 4.1. The energy form is symmetric and positive definite, and its kernel is E 25 the set of constant functions on G. Let 1 denote the constant function with value 1. Then 26 dom /R1 is a vector space with inner product and corresponding norm given by E u, v := (u, v) and u := (u, u)1/2. (4.1) h iE E k kE E 27 Upon completion with respect to this inner norm, we obtain the energy Hilbert space . 0 HE 28 Fix a reference vertex o G to act as an “origin”. It will readily be seen that all results ∈ 29 are independent of this choice.

30 Remark 4.2. In 6, we provide an alternative construction of via techniques of von § HE 31 Neumann and Schoenberg. This provides for a more explicit description of the structure 32 of and its relation to the metric geometry of (G, R), and shows that is the natural HE HE 33 Hilbert space in which to embed (G, R). However, this must be postponed until after the 34 introduction of the eﬀective resistance metric.

35 Remark 4.3 (Three warnings about ). HE 36 (1) has no canonical o.n.b.; the usual candidates δx are not orthogonal and typi- HE { } 37 cally not even dense, as we discuss further below. 38 (2) Multiplication operators are not Hermitian; see Lemma 4.4 and Remark 7.20. 39 (3) Pointwise identities should not be confused with Hilbert space identities; see Re- 40 mark 4.19 and Lemma 4.33.

41 To elaborate on the last point, note that elements of are technically equivalence classes HE 42 of functionswhich differonlybya constant; thisis whatis meantbythe notationdom /R1. E 43 In other words, if v1 = v2 + k for k C, then v1 = v2 in . When working with represen- ∈ HE 44 tatives, we typically abuse notation and use u to denote the equivalence class of u. Often, 36 PALLEE.T.JORGENSENANDERINP.J.PEARSE we choose u so that u(o) = 0 (occasionally without warning). A different but no less useful 1 choice is to pick k so that v = 0 outside a finite set when v is a function of finite support 2 (see Definition 4.16). 3

Lemma 4.4. If ϕ and Mϕ denotes the multiplication operator Mϕ : u ϕu, then 4 ∈ HE 7→ Mϕ is Hermitian if and only if ϕ = 0 in . 5 HE Proof. From the formula (2.9), 6

1 Mϕu, v = cxy(ϕ(x)u(x)v(x) ϕ(x)u(x)v(y) ϕ(y)u(y)v(x) + ϕ(y)u(y)v(y)). h iE 2 − − x,Xy G0 ∈ By comparison with the corresponding expression, this is equal to u, Mϕv iﬀ (ϕ(y) 7 h iE − ϕ(x))u(y)v(x) = (ϕ(y) ϕ(x))u(x)v(y). However, since we are free to vary u and v, it must 8 − be the case that ϕ is constant. 9

Definition 4.5. An exhaustion of G is an increasing sequence of finite and connected sub- 10 graphs G , so that G G + and G = G . 11 { k} k ⊆ k 1 k S Definition 4.6. The notation 12

:= lim (4.2) k xXG0 →∞ xXGk ∈ ∈ is used whenever the limit is independent of the choice of exhaustion G of G. We 13 { k} typically justify this independence by proving the sum to be absolutely convergent. 14

Remark 4.7. One of the main results in this section is a discrete version of the Gauss-Green 15 theorem presented in Theorem 4.41: 16

u, v = u(x)∆v(x) + u(x) ∂v (x), u, v . (4.3) h iE ∂Ò ∈HE xXG0 xXbd G ∈ ∈ This differs from the literature, where it is common to find (u, v) = u, ∆v 2 given as a 17 E h iℓ definition (of or of ∆, depending on the context), e.g. [Kig01, Str08]. After reading a 18 E preliminary version of this paper, a reader pointed out to us that a similar formula appears 19 in [DK88, Prop 1.3]; however, these authors apparently do not consider the extension of 20 this formula to infinite networks. 21 ∂v We refer to u as the “boundary term” by analogy with classical PDE theory. 22 bd G ∂Ò This terminologyP should not be confused with the notion of boundary that arises in the dis- 23 cussion of the discrete Dirichlet problem. In particular, the boundary discussed in [Kig03] 24 0 and [Kig08] refers to a subset of G . By contrast, when discussing an infinite network G, 25 our boundary bdG is never contained in G. Green’s identity follows immediately from 26 (4.3) in the form 27

(u(x)∆v(x) v(x)∆u(x)) = v(x) ∂u (x) u(x) ∂v (x) . (4.4) Ò − ∂Ò − ∂ xXG0 xXbd G ∈ ∈ Note that our deﬁnition of the Laplace operator is the negative of that often found in the 28 PDE literature, where one will ﬁnd Green’s identity written 29

(u∆v v∆u) = (u ∂ v v ∂ u). Ò ∂Ò ∂ ZΩ − Z∂Ω − OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 37

1 As the boundary term may be diﬃcult to contend with, it is extremely useful to know 2 when it vanishes. We have several results concerning this:

3 (i) Lemma 4.68 shows the boundary term vanishes when either argument of u, v has h iE 4 ﬁnite support, 5 (ii) Lemma 4.47 gives necessary and suﬃcient conditions on the electrical resistance 6 network for the boundary term to vanish for any u, v , ∈HE 7 (iii) Lemma 4.63 show the boundary term vanishes when both arguments of u, v and 2 h iE 8 their Laplacians lie in ℓ .

9 In fact, Lemma 4.47 expresses the fact that it is precisely the presence of monopoles that 10 prevents the boundaryterm from vanishing. An example with nonvanishing boundary term 11 is given in Example 13.5.

12 4.1. The evaluation operator Lx and the reproducing kernel vx. 0 13 Deﬁnition 4.8. For any vertex x G , deﬁne the linear evaluation operator Lx on by ∈ HE L u := u(x) u(o). (4.5) x − 14

0 1/2 15 Lemma 4.9. For any x G , one has L u k (u) , where k depends only on x. ∈ | x |≤ E 16 Proof. Since G is connected, we can choose an n-step path from x to o and apply the 17 Schwarz inequality to obtain

n 2 2 2 cxi,xi 1 Lxu = u(x) u(o) = − (u(xi) u(xi 1)) | | | − | c − − Xi=1 r xi,xi 1 − n n 1 2 cxi,xi 1 (u(xi) u(xi 1))    −  ≤ cxi,xi 1 − − Xi=1 −  Xi=1      k2 (u),    ≤ E 1/2 n 1 18 for k = i=1 . cxi,xi 1 P − 19 Deﬁnition 4.10. Let vx be deﬁned to be the unique element of for which HE

vx, u = u(x) u(o), for every u . (4.6) h iE − ∈HE 20 This is justiﬁed by Lemma 4.9 and the Riesz Representation Theorem.

21 Deﬁnition 4.11. Let be a Hilbert space of functions on X. An operator S on is said H H 22 to have a reproducing kernel kx x X iﬀ { } ∈ ⊆H

(Sv)(x) = kx, v , x X, v . (4.7) h iH ∀ ∈ ∀ ∈H 23 If S is projection to a subspace L , then one says k is a reproducing kernel for L. If ⊆H { x} 24 S = I, then is a reproducing kernel Hilbert space with kernel k. H 25 Theorem 4.12 (Aronszajn’s Theorem [Aro50]). Let f be a reproducing kernel for . { x} H 26 Deﬁne a sesquilinear form on the set of all ﬁnite linear combinations of these elements by

ξx fx, ηy fy := ξxηx fx(y). (4.8) *Xx Xy + Xx 38 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Then the completion of this set under the form (4.8) is again . 1 H Corollary 4.13. vx x G0 is a reproducing kernel for . Thus, span vx is dense in . 2 { } ∈ HE { } HE Proof. Choosing representatives with vx(o) = 0, it is trivial to check that vx, vy = vx(y) = 3 h iE vy(x) and then apply Aronszajn’s Theorem. 4

We note that there is a rich modern literature dealing with reproducing kernels and 5 their manifold application to both continuous analysis problems (see e.g., [AD06, AL08, 6 AAL08, BV03, Zha09]), and infinite discrete stochastic models. One of the differences 7 between these studies and our present work is the approach we take in Definition 4.9, i.e., 8 the use of “relative” reproducing kernels. 9

Definition 4.14. The family of functions vx x G0 is called the energy kernel. Note that vo 10 { } ∈ corresponds to a constant function, since vo, u = 0 for every u . Therefore, will 11 h iE ∈ HE 0 often ignore or omit this term and sometime write summations over the set G o . 12 \ { } Remark 4.15. Definition 4.14 is justified by Corollary 4.13. In this paper, the functions vx 13 will play a role analogous to fundamental solutions in PDE theory; see 10.3. 14 § The functions vx are R-valued. This can be seen by first constructing the energy kernel 15 for the Hilbert space of R-valued functions on G, and then using the decomposition of 16

a C-valued function u = u1 + iu2 into its real and imaginary parts. Alternatively, see 17 Lemma 4.29. 18 Reproducing kernels will help with many calculations and explain several of the re- 19 lationships that appear in the study of electrical resistance networks. They also extend 20 the analogy with complex function theory discussed in 10.3. The reader may ﬁnd the 21 § references [Aro50, Yoo07, Jor83] to provide helpful background on reproducing kernels. 22

4.2. The ﬁnitely supported functions and the harmonic functions. 23

Deﬁnition 4.16. Let δx be the Dirac mass at x, i.e., the class of which contains the 24 HE characteristic function of the singleton x . Then the set of “ﬁnitely supported functions” 25 { } in is 26 HE

. 0 span δ = u dom .. u(x) = k for some k, for all but finitely many x G , (4.9) { x} { ∈ E ∈ } . Then spt v H means that v span δ .. x H , and v has finite support iff H is finite. 27 ⊆ ∈ { x ∈ } Define in to be the closure of span δ with respect to . 28 F { x} E Definition 4.17. The set of harmonic functions of finite energy is denoted 29

0 . arm = arm(G ) := v .. ∆v 0 . (4.10) H H { ∈HE ≡ }

Lemma 4.18. The Dirac masses δx x G0 form a reproducing kernel for ∆. That is, for any 30 0 { } ∈ x G , one has δx, u = ∆u(x). 31 ∈ h iE Proof. Compute δx, u = (δx, u) directly from formula (2.9). 32 h iE E Remark 4.19. Note that one can take the deﬁnition of the Laplacian to be the operator A 33 deﬁned via the equation 34

δx, u = Au(x). h iE OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 39

1 This point of view is helpful, especially when distinguishing between identities in Hilbert 2 space and pointwise equations. For example, if h arm, then ∆h and the constant ∈ H 3 function 1 are identiﬁed in because u, ∆h = u, 1 = 0, for any u . However, HE h iE h iE ∈HE 4 one should not consider a (pointwise) solution of ∆u(x) = 1 to be a harmonic function.

0 5 Lemma 4.20. For any x G , ∆v = δ δ . ∈ x x − o 6 Proof. Using Lemma 4.18, ∆vx(y) = δy, vx = δy(x) δy(o) = (δx δo)(y). h iE − − 7 By applying Lemma 4.20 to vα vω, we see: − 8 Corollary 4.21. The space of dipoles (α, ω) is nonempty. P 9 Lemma 4.18 is extremely important. Since in is the closure of span δ , it implies F { x} 10 that the ﬁnitely supported functions and the harmonic functions are orthogonal. This result 11 is called the “Royden Decomposition” in [Soa94, VI] and also appears elsewhere, e.g., § 12 [LP09, 9.3]. § 13 Theorem 4.22. = in arm. HE F ⊕H 14 Proof. For all v , Lemma 4.18 gives δx, v = ∆v(x). Since in = span δx , this ∈ HE h iE F { } 15 equality shows v in whenever v is harmonic. Conversely, if δx, v = 0 for every x, ⊥ F h iE 16 then v must be harmonic. Recall that constants functions are 0 in . HE 17 Corollary 4.23. span δx is dense in iﬀ arm = 0. { } HE H 18 Remark 4.24. Corollary 4.23 is immediate from Theorem 4.22, but we wish to emphasize 19 the point, as it is not the usual case elsewhere in the literature. Part of the importance of 20 the energy kernel v arises from the fact that the Dirac masses are generally inadequate { x} 21 as a representing set for . This leads to unusual consequences, e.g., one may have HE

u , u(x)δx, in . HE xXG0 ∈ 22 More precisely, u x G u(x)δx maynottendto 0 as k , for some exhaustion Gk . k − ∈ k kE →∞ { } P 23 Deﬁnition 4.25. Let fx = P invx denote the image of vx under the projection to in. Simi- F F 24 larly, let hx = P armvx denote the image of vx under the projection to arm. H H 25 For future reference, we state the following immediate consequence of orthogonality.

26 Lemma 4.26. With fx = P invx, fx x G0 is a reproducing kernel for in, but fx arm. F { } ∈ F ⊥ H 27 Similarly, with hx = P armvx, hx x G0 is a reproducing kernel for arm, but hx in. H { } ∈ H ⊥F 28 Remark 4.27. The role of vx in with respect , is directly analogous to role of 2 HE 2h· ·iE 29 the Dirac mass δx in ℓ with respect to the usual ℓ inner product. This analogy will be 0 30 developed further when we show that v is the image of x G under a certain isometric x ∈ 31 embedding into , in 6. It is obvious that δx , and the following result shows that HE § ∈HE 32 δ is always in span v when deg(y) < . However, it is not true that v is always in y { x} ∞ y 33 span δ , or even in its closure. This is discussed further in 6. { x} § 0 34 Lemma 4.28. For any x G , δx = c(x)vx y x cxyvy. ∈ − ∼ P 35 Proof. Lemma 4.18 implies δx, u = c(x)vx y x cxyvy, u for every u , so apply 0 h iE h − ∼ iE ∈HE 36 this to u = vz, z G . Since δx, vx , it mustP also be that y x cxyvy . ∈ ∈HE ∼ ∈HE P 40 PALLEE.T.JORGENSENANDERINP.J.PEARSE

0 4.2.1. Real and complex-valued functions on G . While we will need complex-valued 1 functions for some later results concerning spectral theory, it will usually suﬃce to consider 2 R-valued functions elsewhere. 3

Lemma 4.29. The reproducing kernels vx, fx, hx are all R-valued functions. 4

Proof. Computing directly, 5

1 vz, u = (vz(x) vz(y))(u(x) u(y)) = vz, u . h iE 2 − − h iE x,Xy G0 ∈ Then applying the reproducing kernel property, 6

vz, u = u(x) u(o) = u(x) u(o) = vz, u . h iE − − h iE Thus vz, u = vz, u for every u arm, and vz must be R-valued. The same compu- 7 h iE h iE ∈ H tation applies to fz and hz. 8

Definition 4.30. A sequence of functions un converges pointwise in iff k C 9 {0 }⊆HE HE ∃ ∈ such that u (x) u(x) k, for each x G . 10 n − → ∈ Lemma 4.31. If un converges to u in , then un converges to u pointwise in . 11 { } E { } HE Proof. Define wn := un u so that wn 0. Then 12 − k kE → n wn(x) wn(o) = vx, wn vx wn →∞ 0, | − | |h iE| ≤ k kE · k kE −−−−−−→ so that lim wn exists pointwise and is a constant function. 13

4.3. Relating the energy form to the Laplacian. Before completing the extension of 14 Lemma 2.13 to inﬁnite networks, we need some deﬁnitions. 15

Deﬁnition 4.32. Let := span vx x G0 denote the vector space of ﬁnite linear combina- 16 V { } ∈ tions of dipoles, and let ∆ be the closure of the Laplacian when taken to have the dense 17 V domain . 18 V Since ∆ agrees with ∆ pointwise, we may suppress reference to the domain for ease 19 V of notation. When given a pointwise identity ∆u = v, there is an associated identity in , 20 HE but the next lemma shows that one must use the adjoint. 21

Lemma 4.33. For u, v , ∆u = v pointwise if and only if v = ∆∗ u in . 22 ∈HE V HE Proof. We show that u dom ∆∗ for simplicity, so let ϕ span vx be given by ϕ = 23 n ∈ V ∈ { } 24 i=1 aivxi . Then P n n ∆ϕ, u = ai δx δo, u = ai(∆u(xi) ∆u(o)) Lemma 4.18 h iE h i − iE − Xi=1 Xi=1 n = a (v(x ) v(o)) ∆u(x) = v(x) i i − Xi=1 n = ai vx , v (4.6) h i iE Xi=1 = ϕ, v , h iE OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 41

1 which gives the estimate ∆ϕ, u = ϕ, v ϕ v , by Schwarz, which means |h iE| |h iE| ≤ k kEk kE 2 u dom ∆∗ . The converse is trivial. ∈ V 0 3 Deﬁnition 4.34. A monopole at x G is an element wx which satisﬁes ∆wx = δx ∈ ∈ HE 4 pointwise. By Lemma 4.33, this is equivalent to

wx, ∆u = δx, u , for all u dom ∆ . (4.11) h iE h iE ∈ V 5 Let wo always denote the unique energy-minimizing monopole at the origin. When this 6 function exists (i.e., when wo ), it lies in in. In this case, we indicate the distin- v ∈ HE f F 7 guished monopoles wx := vx + wo and w x := fx + wo, where fx = P invx. It may happen v f F 8 that wx = w x; cf. Theorem 14.5. 9 The space of monopoles at x is denoted , and MPx

v f span wx, w x x G0 , wo := { } ∈ ∃ ∈HE (4.12) MP span vx , else.  { }  10 Note that a monopole need not be in dom ∆ ; see Example 13.8 or Example 14.39. V 11 However, it is always the case that wx dom ∆∗ , which is the content of the following ∈ V 12 lemma.

13 Remark 4.35. Observe that combining (4.11) with Lemma 4.18 immediately gives

wx, ∆u = ∆u(x), (4.13) h iE 14 so that a collection of monopoles wx x G0 is a reproducing kernel for ran ∆ . Note that the { } ∈ V 15 expression ∆u(x) is defined in terms of differences, so the right-hand side is well-defined 16 even without reference to another vertex, i.e., it makes sense independent of any choice of 17 representative.

18 Remark 4.36. The presence of monopoles in is equivalent to the transience of the HE 19 underlying network, that is, the transience of the simple random walk on the network with 20 transition probabilities p(x, y) = cxy/c(x). To see this, note that if wx is a monopole, 21 then the current induced by wx is a unit flow to infinity with finite energy. It was proved 22 in [Lyo83] that the network is transient if and only if there exists a unit current flow to 23 infinity; see also [LP09, Thm. 2.10]. As mentioned in Corollary 3.29, the existence of a 24 monopole at one vertex is equivalent to the existence of a monopole at every vertex.

25 Lemma 4.37. When the network is transient, contains the spaces span v , span f , MP { x} { x} 26 and span hx , where fx = P invx and hx = P armvx. { } F H v f 27 Proof. The ﬁrst two are obvious, since v = w w and f = w w by Deﬁnition 4.34. x x − o x x − o 28 For the harmonics, note that these same identities give

wv w = v = f + h = w f w + h , x − o x x x x − o x v f v f 29 which implies that h = w w . (Of course, w = w when arm = 0.) x x − x x x H 30 Corollary 4.38. arm , 0 iﬀ there is more than one monopole at x. H 31 Proof. As usual, if this is true for any x, it is true for all. Suppose contains a monopole v v HE 32 wx , wx. Then h := wx wx is a nonzero harmonic function in . − HE 42 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Deﬁnition 4.39. If H is a subgraph of G, then the boundary of H is 1

. ∁ bd H := x H .. y H , y x . (4.14) { ∈ ∃ ∈ ∼ } The interior of a subgraph H consists of the vertices in H whose neighbours also lie in H: 2

. int H := x H .. y x = y H = H bd H. (4.15) { ∈ ∼ ⇒ ∈ } \ For vertices in the boundary of a subgraph, the normal derivative of v is 3

∂v (x) := cxy(v(x) v(y)), for x bd H. (4.16) ∂Ò − ∈ Xy H ∈ Thus, the normal derivative of v is computed like ∆v(x), except that the sum extends only 4 over the neighbours of x which lie in H. 5

Definition 4.39 will be used primarily for subgraphs that form an exhaustion of G, in 6 the sense of Definition 4.5: an increasing sequence of finite and connected subgraphs G , 7 { k} so that Gk Gk+1 and G = Gk. Also, recall that bd G := limk bd G from Defini- 8 ⊆ →∞ k tion 4.40. S P P 9

Definition 4.40. A boundary sum (or boundary term) is computed in terms of an exhaus- 10 tion G by 11 { k} := lim , (4.17) k Xbd G →∞ bdXGk whenever the limit is independent of the choice of exhaustion, as in Definition 4.6. The 12 boundary bd G is examined more closely as an object in its own right in 7. 13 § The key point of the following result is that for u, v in the specified set, the two sums 14 are both finite. The decomposition is true for all u, v by taking limits of (4.19), but 15 ∈ HE is clearly meaningless if it takes the form . 16 ∞−∞ Theorem 4.41 (Discrete Gauss-Green Formula). If u and v , then 17 ∈HE ∈ MP u, v = u(x)∆v(x) + u(x) ∂v (x). (4.18) h iE ∂Ò xXG0 xXbd G ∈ ∈

Proof. It suﬃces to work with R-valued functions and then complexify afterwards. By the 18 same computation as in Lemma 2.13, we have 19

1 ∂v cxy(u(x) u(y))(v(x) v(y)) = u(x)∆v(x) + u(x) (x). (4.19) 2 − − ∂Ò xX,y Gk x Xint G x Xbd G ∈ ∈ k ∈ k Taking limits of both sides as k gives (4.18). It remainsto see that one of the sums 20 →∞ on the right-hand side is ﬁnite (and hence that both are). For this part, we work just with u 21 and polarize afterwards. Note that if v = wz is a monopole, then 22

u(x)∆v(x) = u(x)δz(x) = u(z). xXG0 xXG0 ∈ ∈ This is obviously independent of exhaustion, and immediately extends to v . 23 ∈ MP OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 43

1 Remark 4.42. It is clear that (4.18) remains true much more generally than under the 2 specified conditions. Clearly, the formula holds whenever x G0 u(x)∆v(x) < . Un- ∈ | | ∞ 3 fortunately, given any hypotheses more specific than this, theP limitless variety of infinite 4 networks almost always allow one to construct a counterexample; i.e. one cannot give a 5 conditionfor which the formula is true for all u , for all networks. To see this, suppose ∈HE 6 that v = i∞=1 aiwxi with each wxi a monopole at the vertex xi. Then P ∞ u(x)∆v(x) = aiu(xi), xXG0 Xi=1 ∈ 7 and one would need to provide a condition on sequences a that would ensure ∞ a u(x ) { i} i=1 i i 8 is absolutely convergent for all u . Such a hypothesis is not likely to be usefulP (if it is ∈HE 9 even possible to construct) and would depend heavily on the network under investigation. 10 Nonetheless, the formula remains true in many specific contexts. For example, it is 11 clearly valid whenever v is a dipole, including all those in the energy kernel. We will also 12 see that it holds for the projectionsof v to in and to arm. Consequently, for v which are x F H 13 limits of elements in , we often use this result in combination with ad hoc arguments. MP ∂u

14 Lemma 4.43. For all u dom ∆ , 0 ∆u = . Thus, the Discrete Gauss-Green G bd G Ò ∈ V − ∂ 15 formula (4.18) is independent of representatives.P P

16 Proof. On each (ﬁnite) Gk in any given exhaustion,

∂u ∆u(x) + (x) = cxy(u(x) u(y)) = 0, ∂Ò − x Xint G x Xbd G xX,y Gk ∈ k ∈ k ∈ 17 since each edge appears twice in the sum; once with each sign (orientation). For the second 18 claim, we apply the formulaof the ﬁrst to see that the result remains true when u is replaced 19 by u + k:

✟✟ ✟✟

(u + k)∆v + (u + k) ∂v = u∆v + u ∂v + k ∆v✟+ ∂v .

Ò Ò ∂Ò ∂  ✟ ∂  XG0 Xbd G XG0 Xbd G XG✟0 Xbd G  ✟    20 4.3.1. More about monopoles and the space . MP 21 Theorem 4.44 ([Soa94, Thm. 1.33]). Let u be a nonnegative function on a recurrent 22 network. Then u is superharmonic if and only if u is constant.

23 Corollary 4.45. If arm , 0, then there is a monopole in . H HE 24 Proof. If h arm and h , 0, then h = h h with h arm and h 0, by [Soa94, ∈ H 1 − 2 i ∈ H i ≥ 25 Thm. 3.72]. Since the hi cannot both be 0, Theorem 4.44 implies the network is transient. 26 Then by [Lyo83, Thm. 1], the network supports a monopole.

27 Definition 4.46. The phrase “the boundary term is nonvanishing” indicates that (4.18) 28 holds with nonzero boundary sum when applied to u, v , for every representative of u h iE 29 except the one specified by u(x) = u, w , for w x. h iE ∈ MP 30 Recall from Remark 4.36 that the network is transient iff there are monopoles in . HE 31 Theorem 4.47. The network is transient if and only if the boundary term is nonvanishing. 32 Moreover, the boundary term vanishes for the elements of ran ∆ . V 44 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Proof. ( ) If the network is transient, then as explained in Remark 4.36, thereisa w 1 ⇒ ∈HE with ∆w = δz. Now let wz := P inw so that for any u dom ∆ ,(4.18) 2 F ∈ V

∂wz u, wz = u(z) + u . h iE ∂Ò Xbd G

∂wz It is immediate that u = 0 if and only if the computation is done with the repre- 3 bd G ∂Ò sentative of u specifiedP by u(z) = u, wz . 4 h iE 0 ( ) Suppose that there does not exist w with ∆w = δz, for any z G . Then 5 ⇐ ∈ HE ∈ = span v as discussed in Definition 4.34. Therefore, it suffices to show that 6 MP { x}

u, vx = u∆vx, h iE xXG0 ∈ but this is clear because both sides are equal to u(x) u(o)by(4.6) and Lemma 4.20. 7 − For the ﬁnal claim, note that if u ran ∆ , then (4.13) gives 8 ∈ V

∂wx ∂wx

u(x) = u, wx = u∆wx + u = u(x) + u , Ò h iE ∂Ò ∂ XG0 Xbd G Xbd G so that the boundary term must vanish. 9

Remark 4.48. It follows from Theorem 4.47 that a monopole wx cannot lie in ran ∆ . 10 V However, one can have wx ran ∆∗ , as in Example 14.39. 11 ∈ V Lemma 4.49. The network is transient if and only if there is a sequence εk with εk 0 12 1 { } → and supk (εk + ∆)− δx B < . 13 k kE ≤ ∞ 1 Proof. For both directions of the proof, we let fk := (εk + ∆)− δx. 14 1 ( ) Let ∆ be any self-adjoint extension of ∆ , and let E(dλ) be the corresponding 15 ⇒ ∗ V V projection-valued measure. Then 16

1 = + ∆ 1 = ∞ Rεu (ε ∗ )− u E(dλ)u, (4.20) V Z0 ε + λ 1 where we use the notation Rε := (ε + ∆ )− for the resolvent. Note that ∆ Rε (∆ Rε)∗ = 17 ∗ V ∗ V ⊆ ∗ V ∆ = ∆ ∆ ∆ ∆ ∆ 18 ∗ ∗ Rε∗ ∗ Rε. On the other hand, ∗ ∗ and therefore Rε ∗ Rε ∗ . Combining V V V ⊆ V V ⊆ V ∆ ∆ = ∆ 19 these gives ∗ Rε Rε ∗ . Now we apply this and (4.20) to u ∗w to get V ⊆ V λ = + ∆ 1 = + ∆ 1∆ = ∆ + ∆ 1 = ∞ fk (εk ∗ )− δx (εk ∗ )− ∗ w ∗ (εk ∗ )− w E(dλ)w. V V V V V Z0 εk + λ Note that Rε is bounded, and so w dom Rε automatically. This integral implies 20 ∈

2 2 ∞ λ 2 ∞ 2 2 fk E(dλ)w E(dλ)w = w . k kE ≤ Z0 εk + λ! k kE ≤ Z0 k kE k kE 1 Thus we have supk (εk + ∆)− δx = sup fk B = w < . 21 k kE k kE ≤ k kE ∞ ( ) We show the existence of a monopole at x. Since ε f + ∆ f = δ , the bound 22 ⇐ k k k x sup fk B implies that 23 k kE ≤ 1For concreteness, one may take the Friedrichs extension, see (B.9) but this is not necessary. See also Deﬁni- tion 7.6 and 8.1 in this regard. § OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 45

∆ fk δx = εk fk εk B 0. k − kE k k≤ → 1 Let w be a weak- limit of fk . Then for ϕ dom ∆ , ∗ { } ∈ V

∆ϕ, w = lim ∆ϕ, fk = lim ϕ, ∆ fk = lim ϕ, δx εk fk = ϕ, δx , h iE k h iE k h iE k h − iE h iE →∞ →∞ →∞ 2 so that w is a monopole at x.

3 Lemma 4.50. On any network, ran ∆ in and hence arm ker ∆∗ . V ⊆F H ⊆ V 4 Proof. If v V, then clearly ∆ v in. To close the operator, we consider sequences ∈ V ∈ F 5 u V which are Cauchy in , and for which ∆u is also Cauchy in , and then include { n}⊆ E { n} E 6 u := lim un in dom ∆ by defining ∆ u := lim ∆ un. Since fn := ∆ un has finite support V V V V 7 for each n, the -limit of f must lie in in. E { n} F 8 The second claim follows upon taking orthogonal complements; note that in is closed, cℓ F 9 so we actually have (ran ∆ ) in. Alternatively, it can be proven directly, as follows. V ⊆ F 10 To see arm dom ∆∗ , we need h, ∆ v C v . However, this is trivially true H ⊆ V h V iE ≤ k kE 11 because h, ∆ v = 0 for any h arm, by Theorem 4.22. To see arm ker ∆∗ , h V iE ∈ H H ⊆ V 12 compute the value of ∆∗ h via ∆∗ h, v = h, ∆ v = 0. V h V iE h V iE cℓ 13 Theorem 4.51. The random walk on (G, c) is transient if and only if (ran ∆∗ ) = in. V F 14 Proof. ( ) If the network is transient, we actually have a monopole at every vertex by ⇒ 15 Lemma 3.29. Then any u span δx is in ran ∆∗ because the monopole wx is in dom ∆ , ∈ { } V V 16 and so in ran ∆∗ . The other inclusion is Lemma 4.50. F ⊆ V 0 17 ( ) If δx ran ∆ for some x G , then ∆ w = δx for w dom ∆ dom and so ⇐ ∈ V ∈ V ∈ V ⊆ E 18 w is a monopole. Then the induced current dw is a unit flow to infinity, and the network is 19 transient, again by [Lyo83].

20 The next result has a similar ﬂavour, but instead concerns ran(I + ∆ ), as this is the V 21 orthogonal complement of the defect space; see Deﬁnition 8.13. Example 13.9 illustrates 22 an application of Lemma 4.52.

23 Lemma 4.52. Let wz be a monopole at z. Then wz ran(I + ∆ ) if and only if there ∈HE 0 ∈ V 24 is a function u satisfying u + ∆u = 0 on G z . ∈HE \ { }

25 Proof. ( ) Since wz = v + ∆v for some v , set u := v wz. Then for x , z, it is easy ⇒ ∈HE − 26 to check u(x) + ∆u(x) = 0. ( ) Set v := w + au for a := 1/(u(z) + ∆u(z)). Then it is easy ⇐ z − 27 to check that v(x) + ∆v(x) = wz(x).

28 Definition 4.53. For an infinite graph G, we say u(x) vanishes at iff for any exhaustion ∞ 29 Gk , one can always find k and a constant C such that u(x) C < ε for all x < Gk. { } k − k∞ 30 One can always choose the representative of u so that C = 0, but this may not be ∈ HE 31 compatible with the choice u(o) = 0.

32 Deﬁnition 4.54. Say γ = (x0, x1, x2,... )isa path to iﬀ xi xi 1 for each i, and for any ∞ ∼ − 33 exhaustion G of G, { k}

k, N such that n N = x < G . (4.21) ∀ ∃ ≥ ⇒ n k 46 PALLEE.T.JORGENSENANDERINP.J.PEARSE

The next two results are almost converse to each other, although the exact converse of 1 Lemma 4.55 is false; see Figure 10 of Example 14.2. Lemma 4.55 is related to [Soa94, 2 Thm. 3.86], in which the result is stated as holding almost everywhere with respect to the 3 notion of extremal length. 4

Lemma 4.55. If u and u vanishes at , then u in. 5 ∈HE ∞ ∈F Proof. Let u = f + h vanish at . This implies that for any exhaustion Gk and any 6 ∈HE ∞ { } ε> 0, there is a k and C for which h(x) C < ε outside Gk. A harmonic function can 7 k − k∞ only obtain its maximum on the boundary, unless it is constant, so in particular, ε bounds 8 h(x) C on all of Gk. Letting ε 0, h tends to a constant function and u = f . 9 k − k∞ → 0 Lemma 4.56. If h arm is nonconstant, then from any x G , there is a path to 10 ∈ H 0 ∈ inﬁnity γ = (x0, x1,... ), with h(x j) < h(x j+1) for all j = 0, 1, 2,.... 11

cxy Proof. Since h(x) = y x c(x) h(y) supy x h(y) and h is nonconstant, we can always find 12 ∼ ≤ ∼ y x for which h(yP1) > h(x0). This follows from the maximal principle for harmonic 13 ∼ functions; cf. [LP09, 2.1], [LPW08, Ex. 1.12], or [Soa94, Thm. 1.35]. Thus, one can 14 § inductively construct a sequence which defines the desired path γ. Note that γ is infinite, 15 0 so the condition h(x j) < h(x j+1) eventually forces it to leave any finite subset of G , so 16 Definition 4.54 is satisfied. 17

The next result is the contrapositive of Lemma 4.55, but it is instructive to prove it 18 directly. 19

Lemma 4.57. Ifh arm is nonconstant, then h has at least two different limiting values 20 ∈H at . 21 ∞ 0 Proof. Choose x G for which hx = P armvx is nonconstant. Then Lemma 4.56 22 ∈ H ∈ HE gives a path to infinity γ1 along which hx is strictly increasing. Since the reasoning of 23 Lemma 4.56 works just as well with the inequalities reversed, we also get γ to along 24 2 ∞ which hx is strictly decreasing. This gives two different limiting values of hx, and hence hx 25 cannot vanish at . 26 ∞ Remark 4.58 (The harmonicregion of a network). Inorderto see whyit is necessaryto take 27 the step of choosing x for which hx is nonconstant, consider a network built by conjoining 28 a copy of the integer lattice (Z, 1) to the binary tree ( , 1), by identifying their origins. 29 T Such a network has a portion (Z) which does not support nontrivial harmonic functions, 30 and so h must be constant for any x in this region. Since h is a reproducing kernel for 31 x { x} arm, however, h cannot be constant for x in the tree portion of the network. This idea 32 H x motivates the proof of Lemma 4.57 and the definition of boundary points given in 7.3. 33 § Roughly speaking, the “harmonic portion” of a network consists of those points which can 34 be distinguished by a harmonic function, and the boundary consists of those paths to 35 ∞ which can be distinguished by harmonic functions. We will also see in Remark 5.43 that 36 the fact that a network may have nonharmonic regions (like Z in this example) contributes 37 ha of the failure of the harmonic resistance R to be a metric. 38

4.3.2. Special applications of the Discrete Gauss-Green formula. In this subsection, we 39 use Lemma 2.13 to inﬁnite networks to establish that ∆ is Hermitian when its domain is 40 correctly chosen (Corollary 4.61), and that Lemma 2.13 remains correct on inﬁnite net- 41 works for vectors in span v (Theorem 4.68). 42 { x} Lemma 4.59. If v span vx , then u, v = x G0 u(x)∆v(x). 43 ∈ { } h iE ∈ P OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 47

1 Proof. It suﬃces to consider v = vx, whence

u(y)∆vx(y) = u(y)(δx δo)(y) = u(x) u(o) = u, vx , − − h iE XG0 XG0

2 by Lemma 4.20 and the reproducing property of Corollary 4.13.

3 Theorem 4.60. For u, v span v , ∈ { x}

u, ∆v = ∆u(x)∆v(x). (4.22) h iE xXG0 ∈ 4 Furthermore, x G0 ∆u(x) = 0 for u span vx . ∈ ∈ { } P 5 Proof. Let u span v be given by the ﬁnite sum u = ξ v . Since v is a constant, we ∈ { x} x x x o 6 may assume the sum does not include o. Then P

∆u(y) = ξ ∆v (y) = ξ (δ δ )(y) = ξ . (4.23) x x x x − o y Xx Xx

7 Now we have

u, ∆u = ξxξy vx, ∆vy = ξxξy vx,δy δo . h iE h iE h − iE Xx,y Xx,y

8 Since it is easy to compute vx,δy δo = δxy + 1 (Kronecker’s delta), we have h − iE

2 2 2 2 u, ∆u = ξxξy(δxy + 1) = ξx + ξx = ∆u(x) + ∆u(x) , (4.24) h iE | | | | Xx,y Xx Xx Xx Xx

9 by (4.23). Since u span vx , ∆u span δ x δo (see (4.23)), so that u, ∆u < and ∈ { } ∈ { − } h iE ∞ 10 (4.24) is convergent. Therefore, x ∆u(x) is absolutely convergent, hence independent of 11 exhaustion. Since P

∆v (x) = 1 1 = 0 y − xXG0 ∈ 12 by Lemma 4.20, it follows that x ∆u(x) = 0, and the second sum in (4.24) vanishes. Then 13 (4.22) follows by polarizing. P

14 Corollary 4.61. The Laplacian ∆ is Hermitian and even semibounded on dom ∆ (see V V 15 Deﬁnition B.10) with

0 ∆u(x) 2 u, ∆u < . (4.25) ≤ | | ≤ h iE ∞ xXG0 ∈

16 Proof. For u, v span v , two applications of Lemma 4.60 yield ∈ { x}

∆u, v = ∆u(x)∆v(x) = ∆u(x)∆v(x) = ∆v, u . h iE h iE xXG0 xXG0 ∈ ∈ 17 This property is clearly preserved under closure of the operator. 48 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Now let u dom ∆ and choose un V with limn un u = limn ∆un 1 ∈ V { } ⊆ →∞ k − kE →∞ k − ∆u = 0. Then Fatou’s lemma [Mal95, Thm. I.7.7] yields 2 kE

2 2 ∆u(x) = lim ∆un(x) lim un, ∆un = u, ∆u , (4.26) | | | | ≤ n h iE h iE xXG0 xXG0 →∞ ∈ ∈ which gives the central inequality in (4.25) and hence semiboundedness. 3

1 2 Remark 4.62. The notation u ℓ means x G0 u(x) < and the notation u ℓ means 4 2 ∈ ∈ | | ∞ 2∈ x G0 u(x) < . When discussing an elementP u of , we say u lies in ℓ if it has a 5 ∈ | | ∞ 2 HE representative which does, i.e., if u + k ℓ for some k C. This constant is clearly 6 P ∈ ∈ necessarily unique on an inﬁnite network, if it exists. 7

The next result is a partial converse to Theorem 4.41. 8

2 Lemma 4.63. If u, v, ∆u, ∆v ℓ , then u, v = x G0 u(x)∆v(x),and u, v dom . 9 ∈ h iE ∈ ∈ E 2 1 P Proof. If u, ∆v ℓ , then u∆v ℓ , and the following sum is absolutely convergent: 10 ∈ ∈ 1 1 u(x)∆v(x) = u(x)∆v(x) + u(y)∆v(y) 2 2 xXG0 xXG0 yXG0 ∈ ∈ ∈ 1 1 = c u(x)(v(x) v(y)) c u(y)(v(x) v(y)) 2 xy − − 2 xy − xXG0 Xy x yXG0 Xx y ∈ ∼ ∈ ∼ 1 = c (u(x) u(y))(v(x) v(y)), 2 xy − − xXG0 Xy x ∈ ∼ which is (2.9). Absolute convergence justifies the rearrangement in the last equality; the 11 1 rest is merely algebra. Substituting u in for v in the identity just established, u∆u ℓ 12 ∈ shows u dom , and similarly for v. 13 ∈ E Remark 4.64. All that is required for the computation in the proof of Lemma 4.63 is that 14 1 2 u∆v ℓ , which is certainly implied by u, ∆v ℓ . However, this would not be sufficient 15 ∈ ∈ to show u or v lies in dom . 16 E We will see in Theorem 4.55 that if h arm is nonconstant, then h + k is bounded 17 ∈ H away from 0 on an infinite set of vertices, for any choice of constant k. So the next result 18 should not be surprising. 19

2 Corollary 4.65. If h is a nontrivial harmonic function, then h cannot lie in ℓ . 20 ∈HE 2 Proof. If h ℓ , then (h) = x G0 h(x)∆h(x) = x G0 h(x) 0 = 0 by Lemma 4.63. But 21 ∈ E ∈< ∈ · since h is nonconstant, (h) >P0! P 22 E ւ 2 Remark 4.66 (Restricting to ℓ misses the most interesting bit). When studying the graph 23 2 . 2 Laplacian, some authors deﬁne dom ∆ = v ℓ .. ∆v ℓ . Our philosophy is that dom 24 { ∈ ∈ }0 E is the most natural context for the study of functions on G , and this is motivated in detail 25 in 6.1. Some of the most interesting phenomena in dom are due to the presence of 26 § E nontrivial harmonic functions, as we show in this section and the examples of 13– 14. 27 § § Consequently, Corollary 4.65 shows why one loses some of the most interesting aspects of 28 2 the theory by only studying those v which lie in ℓ . Example 13.2 illustrates the situation 29 of Corollary 4.65 on a tree network. In general, if a at least two connected components of 30 2 G o are inﬁnite, then v < ℓ for vertices x in these components. 31 \ { } x OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 49

1 4.3.3. The Discrete Gauss-Green formula for vertices of infinite degree. If there are ver- 2 tices of infinite degree in the network, then it does not necessary follow that span δ { x} ⊆ 3 span v , or that span δ . However, we do have the following version of Theo- { x} { x} ⊆ MP 4 rem 4.41. When all vertices have finite degree, Theorem 4.68 follows from Theorem 4.41 5 by Lemma 4.28.

6 Definition 4.67. Let := span δx x G0 denote the vector space of finite linear combina- F { } ∈ 7 tions of Dirac masses, and let ∆ be the closure of the Laplacian when taken to have the F 8 domain . F 9 Note that is a dense domain only when arm = 0, by Corollary 4.23. Again, since ∆ F H 10 agrees with ∆ pointwise, we may suppress reference to the domain for ease of notation. F 11 The next result extends Lemma 2.13 to infinite networks.

12 Theorem 4.68. If uorvliesin dom ∆ , then u, v = x G0 u(x)∆v(x). F h iE ∈ P 13 Proof. First, suppose u dom ∆ and choose a sequence un span δx with un u ∈ F { }⊆ { } k − kE → 14 0. From Lemma 4.18, one has δx, v = ∆v(x), and hence un, v = x G0 un(x)∆v(x) h iE h iE ∈ 15 holds for each n. Deﬁne M := sup un , and note that M < , sinceP this sequence is {k kE} ∞ 16 convergent (to u ). Moreover, un, v M v by the Schwarz inequality. Since un k kE |h iE| ≤ · k kE 17 converges pointwise to u in by Lemma 4.31, this bound will allow us to apply Fatou’s HE 18 Lemma (as stated in [Mal95, Lemma 7.7], for example), as follows:

u, v = lim un, v hypothesis E n E h i →∞h i = lim un(x)∆v(x) un span δx n ∈ { } →∞ xXG0 ∈ = u(x)∆v(x). xXG0 ∈ 0 19 Note that the sum over G is absolutely convergent, as required by Deﬁnition 4.5. 20 Now suppose that v dom ∆ and observe that this implies v in also. By Theo- ∈ F ∈ F 21 rem 4.22, one can decompose u = f + h where f = P inu and h = P armu, and then F H

u, v = f, v + h, v = f, v , h iE h iE h iE h iE 22 since h is orthogonal to v. Now apply the previous argument to f, v . h iE 50 PALLEE.T.JORGENSENANDERINP.J.PEARSE

5. The resistance metric 1

We now introduce the natural notion of distance on (G, c): the resistance metric R. 2 While not as intuitive as the more common shortest-path metric, it reflects the topology of 3 the graph more accurately and is often more useful for modeling and practical applications. 4 The effective resistance is intimately related to the random walk on(G, c), the Laplacian, 5 and the Dirichlet energy form [LP09, LPW08, Soa94, Kig01, Str06, DS84]. 6 In 5.1, we give several formulationsof this metric (Theorem 5.2), each with its ownad- 7 § vantages. Many of these are familiar from the literature: (5.1) from [Pow76b] and[Per99, 8 8], (5.2) from [DS84],(5.3) from [DS84, Pow76b],(5.4)–(5.5) from [Kig01, Str06]. 9 § In 5.2, we extend these formulations to infinite networks. Due to the possible presence 10 § of nontrivial harmonic functions, some care must be taken when adjusting these formu- 11 lations. It turns out that there are two canonical extensions of the resistance metric to 12 infinite networks which are distinct precisely when arm , 0 (cf. [LP09] and the refer- 13 H ences therein): the “free” resistance and the “wired” resistance. We are able to clarify and 14 explain the difference in terms of the reproducing kernels for and for in, and also in 15 HE F terms of Dirichlet vs. Neumann boundary conditions; see Remark 5.19. We also explain 16 the discrepancy in terms of projections in and attempt to relate this to conditioning of 17 HE the random walk on the network; see Remark 5.22 and Remark 5.34. Additionally, we 18 introduce trace resistance and harmonic resistance and relate these to the free and wired 19 resistances. (Note: unlike the others, harmonic resistance is not a metric.) In the limit, the 20 trace resistance coincides with the free resistance. 21

5.1. Resistance metric on finite networks. We make the standing assumption that the 22 network is finite in 5.1. However, the results actually remain true on any network for 23 § which arm = 0. 24 H Definition 5.1. If one amp of current is inserted into the electrical resistance network at x 25 and withdrawn at y, then the (effective) resistance R(x, y) is the voltage drop between the 26 vertices x and y. 27

Theorem 5.2. The resistance R(x, y) has the following equivalent formulations: 28

.. R(x, y) = dist∆(x, y) := v(x) v(y) . ∆v = δx δy (5.1) { −. − } = dist (x, y) := (v) .. ∆v = δx δy (5.2) E {E . − } = dist (x, y) := min D(I) .. I (x, y) (5.3) D { ∈F } .. = distR(x, y) := 1/ minv dom (v) . v(x) = 1, v(y) = 0 (5.4) ∈ E{E } . 2 = distκ(x, y) := minv dom κ 0 .. v(x) v(y) κ (v) (5.5) ∈ E{ ≥ | − | ≤ E } 2 .. = dists(x, y) := sup v(x) v(y) . (v) 1 . (5.6) v dom {| − | E ≤ } ∈ E

Proof. (5.1) (5.2). We may choose v satisfying ∆v = δ δ by Theorem 3.28. Then 29 ⇐⇒ x − y

(v) = v(z)∆v(z) = v(z)(δ (z) δ (z)) = v(x) v(y), (5.7) E x − y − zXG0 zXG0 ∈ ∈ where first equality is justified by Theorem 2.13. 30 (5.2) (5.3). Note that every v (x, y) corresponds to an element I (x, y) via 31 ⇐⇒ ∈ P ∈ F Ohm’s Law by Lemma 3.16, and (v) = D(I) by the same lemma. Also, this current flow 32 E is minimal by Theorem 3.26. 33 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 51

1 (5.2) (5.4). Suppose that ∆v = δ δ . Since (v + k) = (v) and ∆(v + k) = ∆v ⇐⇒ x − y E E 2 for any constant k, we may adjust v by a constant so that v(y) = 0. Deﬁne

v v(x) u := − v(x) v(y) − 3 so that u(x) = 0 and u(y) = 1. Observe that (5.1) gives (v) = v(x) v(y), whence E −

2 1 (u) = (v)/(v(x) v(y)) = (v(x) v(y))− min (u). E E − − ≥ E 1 4 This shows (v) [min (u)]− and hence dist distR. E ≤ E E ≤ 5 For the other inequality, suppose u minimizes (u), subject to u(x) u(y) = 0. Then by E − 6 Theorem 2.13 and the same variational argument as described in Remark 3.19, we have

(ρ, u) = ρ(z)∆u(z) = 0, E zXG0 ∈ 7 for every function ρ for which ρ(x) = ρ(y) = 0. It follows that ∆u(z) = 0 for z , x, y, and 8 hence ∆u = ξδ +ηδ . Observe that (u) = ( u) = (1 u), and so the same result follows x y E E − E − 9 from minimizing with respect to the conditions u(y) = 1 and u(x) = 0. This symmetry E 1 10 forces η = ξ and we have ∆u = ξδ ξδ . Now for v = u one has ∆v = δ δ , and so − x − y ξ x − y

(u) = ξ2 (v) = ξ2(v(x) v(y)) = ξ(u(x) u(y)) = ξ, E E − − 1 11 where the second equality follows by (5.1). Then ξ = (v) = (u), whence dist distR. E E E ≥ 12 (5.4) (5.5). Starting with (5.5), it is clear that ⇐⇒

2 .. v(x) v(y) distκ(x, y) = inf κ 0 . | −(v) | κ, v dom { ≥ E ≤ ∈ E} v(x) v(y) 2 = sup | −(v) | , v dom , v nonconstant . { E ∈ E } v 13 Given a nonconstant v dom , one can substitute u := v(x) v(y) into the previous line to ∈ E | − | 14 obtain

✘✘ u(x) u(y) 2✘v(x✘) v(y) 2 dist (x, y) = sup | − | | ✘−✘ | , v dom , v nonconstant κ (u)✘v(x✘) v(y) 2 { E | − | ∈ E } 1 = sup (u) , u dom , u(x) u(y) = 1 { E ∈ E | − | } = 1/ inf (u), u dom , u(x) u(y) = 1 . {E ∈ E | − | } 15 Since we can always add a constant to u and multiply by 1 without changing the energy, ± 16 this is equivalent to letting u rangeover the subset of dom for which u(x) = 1 and u(y) = 0 E 17 and we have (5.4). 18 (5.5) (5.6). It is immediate that (5.5) is equivalent to ⇐⇒

2 v(x) v(y) . sup | − | .. (v) < . ( (v) E ∞) E 2 2 1/2 19 For any v dom , deﬁne w := v/ √ (v) so that w(x) w(y) = v(x) v(y) (v)− with ∈ E E2 | − | | − | E 20 (w) = 1. Clearly then w(x) w(y) dist (x, y). The other inequality is similar. E | − | ≤ s 52 PALLEE.T.JORGENSENANDERINP.J.PEARSE

x x x

1 1 1 R(x,z) = 1+ -1 -1 y 2 + 3 y

1 -1 -1 2 3 2 + 3 z z z

Figure 3. Eﬀective resistance as network reduction to a trivial network. This basic example uses parallel reduction followed by series reduction; see Remark 5.4.

The equivalence of (5.3) and (5.1) is shown elsewhere (e.g., see [Pow76b, II]) but 1 § the reader will find some gaps, so we have included a complete version of this proof for 2 completeness. The terminology “effective resistance metric” is common in the literature 3 (see, e.g., [Kig01]and[Str06]), where it is usually given in the form (5.4). The formulation 4 (5.5) will be helpful for obtaining certain inequalities in the sequel. It is also clear that dists 5 of (5.6) is the norm of the operator L defined by L u := u(x) u(y), see Lemma 4.9 and 6 xy xy − Theorem 5.12. 7

Remark 5.3. Taking the minimum (rather than the inﬁmum) in (5.3), etc, is justiﬁed by 8 Theorem 3.13. The same argument implies that the energy kernel on G is uniquely deter- 9 mined. 10

Remark 5.4 (Resistance distance via network reduction). Let H be a (connected) planar 11 subnetwork of a finite network G and pick any x, y H. Then H may be reduced to a 12 ∈ trivial network consisting only of these two vertices and a single edge between them via 13 the use of three basic transformations: (i) series reduction, (ii) parallel reduction, and (iii) 14 the -Y transform. Each of these transformations preserves the resistance properties of the 15 ∇ subnetwork, that is, for x, y G H, R(x, y) remains unchanged when these transformations 16 ∈ \ are applied to H. Theeffective resistance between x and y may be interpreted as the 17 resistance of the resulting single edge. An elementary example is shown in Figure 3. 18 A more sophisticated technique of network reduction is given by the Schur complement 19 construction defined in Remark 5.34. 20

The following result is not new (see, e.g. [Kig01, 2.3]), but the proof given here is 21 § substantially simpler than most others found in the literature. 22

Lemma 5.5. R is a metric. 23

Proof. Symmetry and positive deﬁniteness are immediate from (5.2),weuse(5.1) to check 24 the triangle inequality. Let v (x, y) and v (y, z). By superposition, v := v + v is 25 1 ∈P 2 ∈P 3 1 2 in (x, z). For s t, it is clear that v (s) v (t) = v (s) v (t) + v (s) v (t). By summing 26 P ∼ 3 − 3 1 − 1 2 − 2 along any path from x to z, one sees that this remains true for s / t, whence 27 ∼

R(x, z) = v (x) v (z) = v (x) v (z) + v (x) v (z) 3 − 3 1 − 1 2 − 2 v (x) v (y) + v (y) v (z) = R(x, y) + R(y, z), ≤ 1 − 1 2 − 2 where the inequality follows from Proposition 3.32. 28 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 53

1 5.2. Resistance metric on infinite networks. There are difficulties with extending the 2 results of the previous section to infinite networks. The existence of nonconstant harmonic 3 functions h dom implies the nonuniqueness of solutions to ∆u = f , and hence (5.1)– ∈ E 4 (5.3) are no longer well-defined. Two natural choices for extension lead to the free resis- F W 5 tance R and the wired resistance R . In this section, we attempt to explain the somewhat W F 6 surprising phenomenon that one may have R (x, y) < R (x, y). F 7 (1) In Theorem 5.12, we show how R corresponds to choosing solutions to ∆u = 8 δ δ from the energy kernel, and how it corresponds to currents which are x − y 9 decomposable in terms of paths. The latter leads to a probabilistic interpretation 10 which provides for a relation to the trace of the resistance discussed in 5.2.3. W § 11 (2) In Theorem 5.18, we show how R corresponds to projection to in. Since this F 12 corresponds to minimization of energy, it is naturally related to capacity.

13 See also Remark 5.22. Both of these notions are methods of specifying a unique solutions F W 14 to ∆u = f in some way. The disparity between R and R is thus explained in terms of 15 boundary conditions on ∆ as an unbounded self-adjoint operator on in Remark 5.19. HE 16 To compute effective resistance in an infinite network, we will need three notions of 17 subnetwork: free, wired, and trace. Strictly speaking, these may not actually be subnet- 18 works of the original graph; they are networks associated to a full subnetwork. Throughout 0 19 this section, we use H to denote a finite full subnetwork of G, H to denote its vertex set, F W tr 20 and H , H , and H to denote the free, wired, and trace networks associated to H (these 21 terms are defined in other sections below).

22 Deﬁnition 5.6. If H is a subnetwork of G which contains x and y, deﬁne RH(x, y)to be the 23 resistance distance from x to y as computed within H. In other words, compute RH(x, y)by 24 any of the equivalent formulas of Theorem 5.2, but extremizing over only those functions 25 whose support is contained in H.

26 We will always use the notation G ∞ to denote an exhaustion of the infinite network { k}k=1 27 G. Recall from Definition 4.5 that this means each Gk is a finite connected subnetwork of 0 0 28 G, G G + , and G = G . Since x and y are contained in all but finitely many G , we k ⊆ k 1 k k 29 may always assume that x, y G . Also, we assume in this section that the subnetworksare S ∈ k 30 full — this is not necessary, but simplifies the discussion and causes no loss of generality. 0 0 0 31 Definition 5.7. Let H G . Then the full subnetwork on H has all the edges of G for ⊆ 0 32 which both endpoints lie in H , with the same conductances.

33 5.2.1. Free resistance. 0 0 F 34 Definition 5.8. For any subset H G , the free subnetwork H is just the full subnetwork ⊆ 0 35 H. Thatis, all edgesof G with endpoints in H are edges of H, with the same conductances. F 36 Let RHF (x, y) denote the effective resistance between x and y as computed in H = H , as 37 in Definition 5.6. The free resistance between x and y is then defined to be

F R (x, y) := lim RGF (x, y), (5.8) k k →∞ 38 where G is any exhaustion of G. { k}

39 Remark 5.9. The name “free” comes from the fact that this formulation is free of any 40 boundary conditions or considerations of the complement of H, in contrast to the wired 41 and trace formulations of the next two subsections. See [LP09, 9] for further justiﬁcation § 42 of this nomenclature. 54 PALLEE.T.JORGENSENANDERINP.J.PEARSE

One can see that RHF (x, y) has the drawback of ignoring the conductivity providedby all 1 paths from x to y that pass through the complement of H. This provides some motivation 2 for the wired and trace approaches below. 3

Deﬁnition 5.10. Fix x, y G and deﬁne the linear operator Lxy on by Lxyv := v(x) 4 ∈ HE − v(y). 5

Remark 5.11. Theorem 5.12 is the free extension of Theorem 5.2 to inﬁnite networks; 6 it shows that R(x, y) = Lxy and that R(x, o) is the best possible constant k = kx in 7 k k χ Lemma 4.9. In the proof, we use the notation γ for a current which is the characteris- 8 tic function of a path, that is, a current which takes value 1 on every edge of γ Γ(x, y) 9 χ ∈ and 0 on all other edges. Then I = ξγ γ indicates that I decomposes as a sum of currents 10 supported on paths in G. P 11

F Theorem 5.12. For an inﬁnite network G, the free resistance R (x, y) has the following 12 equivalent formulations: 13

RF (x, y) = (v (x) v (y)) + (v (x) v (y)) (5.9) x − x y − y = (vx vy) (5.10) E − . = min D(I) .. I (x, y) and I = ξγχ (5.11) { ∈F γ} .. P1 = (min (v) . v(x) = 1, v(y) = 0 )− + (P arm(vx vy)) (5.12) {E } E H − . 2 = inf κ 0 .. v(x) v(y) κ (v), u dom (5.13) { ≥ | − | ≤ E ∀ ∈ E} 2 . = sup v(x) v(y) .. (v) 1, u dom (5.14) {| − | E ≤ ∀ ∈ E}

Proof. To see that (5.10) is equivalent to (5.8), ﬁx any exhaustion of G and note that 14

1 2 (vx vy) = lim cst((vx vy)(s) (vx vy)(t)) = lim RGF (x, y), E − k 2 − − − k k →∞ sX,t G →∞ ∈ k where the latter equality is from Theorem 5.2. Then for the equivalence of formulas (5.9) 15 and (5.10), simply compute 16

(vx vy) = vx vy, vx vy = vx, vx 2 vx, vy vy, vy E − h − − iE h iE − h iE − h iE and use the facts that vx is R-valued (see Remark 4.15) and vx(o) = 0. 17 To see (5.11) is equivalent to (5.8), ﬁx any exhaustion of G and deﬁne 18

.. χ (x, y) := I (x, y) . I = γ H ξγ γ . F H { ∈F ⊆ } P From (5.3), it is clearly true for each Gk that 19

. F , = .. , = ξ χ . RG (x y) min D(I) I (x y) and I γ Gk γ γ k { ∈F ⊆ } P Since (x, y) = k (x, y) , formula (5.11) follows. Note that we are justiﬁed in 20 F G F Gk using minimum insteadS of inﬁmum, since D is a quadratic form on the closed convex set 21

(x, y) . 22 F G To see (5.12), we unfortunately need to refer ahead to Theorem 5.18, which shows that 23 .. 1 ( fx fy) = (min (v) . v(x) = 1, v(y) = 0 )− , where fx = P invx. 24 E − {E } F OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 55

1 As for (5.13) and (5.14), they are both clearly equal to L (as described in Re- k xyk 2 mark 5.12) by the definition of operator norm; see [Rud87, 5.3], for example. To relate F § 3 them to R , define a subspace of consisting of those voltages whose induced currents HE 4 are supported in a finite subnetwork H by

F . = u dom .. u(x) u(y) = 0 unless x, y H . (5.15) HE H { ∈ E − ∈ } 5 This is a closed subspace, as it is the intersection of the kernels of a collection of continuous

6 linear functionals L , and so we can let Q be the projection to this subspace. Then it is k stk k 7 clear that Qk Qk+1 and that limk u Qku = 0 for all u , so ≤ →∞ k − kE ∈HE

F RG (x, y) = Lxy G C = LxyQk , (5.16) k k kHE| k → k k 8 where the ﬁrst equality follows from (5.5) (recall that Gk is ﬁnite) and therefore

F R (x, y) = lim RGF (x, y) = lim LxyQk = lim LxyQk = Lxy . k k k k k k k k →∞ →∞ →∞

F 9 Remark 5.13. In Theorem 5.12, the proofs that R is given by (5.11)or(5.13) stem from 10 essentially the same underlying martingale argument. In a Hilbert space, a martingale is 11 an increasing sequence of projections Q with the martingale property Q = Q Q + . In { k} k k k 1 12 this context, Doob’s theorem [Doo53] then states that if f is such that f = Q f { k}⊆H k k j 13 for any j k, then the following are equivalent: ≥ 14 (i) there is a f such that f = Q f for all k ∈H k k 15 (ii) sup f < . k k kk ∞ 16 Recall that conditional expectation is a projection. For (5.11), we are actually projecting to 17 subspaces of , the Hilbert space of currents introduced as the dissipation space in 10. HD § 18 In view of the previous result, the free case corresponds to consideration of only those 19 voltage functionswhose induced current can be decomposedas a sum of currentssupported 20 on paths in G. The wired case considered in the next section corresponds to considering all 21 voltages functions whose induced current flow satisfies Kirchhoff’s law (3.6); this is clear 22 from comparison of (5.11)to(5.21). See also Remark 5.20. 23 Formula (5.9) turnsoutto beusefulfor explicitcomputations;we use it to obtain explicit d 24 formulas for the effective resistance metric on Z in Theorem 14.7.

F 25 Proposition 5.14. R (x, y) is a metric.

26 Proof. One has RGF (x, z) RGF (x, y) + RGF (y, z) for any k, so take the limit. k ≤ k k

27 From Theorem 5.12, it is clear that the triangle inequality also has the formulation

(v v ) (v v ) + (v v ), x, y, z G0, E x − z ≤E x − y E y − z ∀ ∈ 28 which is easily shown to be equivalent to

v (z) + v (z) v (z) + v (y), x, y, z G0, x y ≤ z x ∀ ∈ 29 using the convention vx(o) = 0. 30 The next result is immediate from (5.13).

1 31 Corollary 5.15. Every function in is H¨older continuous with exponent . HE 2 56 PALLEE.T.JORGENSENANDERINP.J.PEARSE

It is knownfrom[Nel64] that the Gaussian measure of Brownian motion is supportedon 1 the space of such functions and this will be useful later; cf. Remark 7.28 and the beginning 2 of 7.2. It is somewhat subtle to determine if R(x, ) is in . 3 § · HE 5.2.2. Wired resistance. 4 W Definition 5.16. Given a finite full subnetwork H of G, define the wired subnetwork H 5 0 0 by identifying all vertices in G H to a single, new vertex labeled . Thus, the vertex 6 W 0 \ W ∞ set of H is H , and the edge set of H includes all the edges of H, with the same 7 ∪ {∞} 0 0 0 W conductances. However, if x H has a neighbour y G H , then H also includes an 8 ∈ ∈ \ edge from x to with conductance 9 ∞

cx := cxy. (5.17) ∞ y xX, y H∁ ∼ ∈ W Let RHW (x, y) denote the effective resistance between x and y as computed in H , as in 10 Definition 5.6. The wired resistance is then defined to be 11

W R (x, y) := lim RGW (x, y), (5.18) k k →∞ where G is any exhaustion of G. 12 { k}

Remark 5.17. The wired subnetwork is equivalently obtained by “shorting together” all 13 ∁ W vertices of H , and hence it follows from Rayleigh’s monotonicity principle that R (x, y) 14 F ≤ R (x, y); cf. [DS84, 1.4] or [LP09, 2.4]. 15 § § ∁ The justiﬁcation for (5.17) is that the identiﬁcation of vertices in Gk may result in par- 16 allel edges. Then (5.17) corresponds to replacing these parallel edges by a single edge 17 according to the usual formula for resistors in parallel. 18

Theorem 5.18. The wired resistance may be computed by any of the following equivalent 19 formulations: 20

W . R (x, y) = min v(x) v(y) .. ∆v = δx δy, v dom (5.19) v { − − ∈ E} .. = min (v) . ∆v = δx δy, v dom (5.20) v {E − ∈ E} . = min D(I) .. I (x, y), D(I) < (5.21) I { ∈F ∞} . = 1/ min (v) .. v(x) = 1, v(y) = 0 (5.22) {E } . 2 = inf κ 0 .. v(x) v(y) κ (v), v in (5.23) { ≥ | − | ≤ E ∀ ∈F } 2 . = sup v(x) v(y) .. (v) 1, v in (5.24) {| − | E ≤ ∀ ∈F }

Proof. Since (5.23) and (5.24) are both clearly equivalent to the norm of L : in C 21 xy F → (where again L u u(x) u(y) as in Remark 5.11), we begin by equating them to (5.18). 22 xy − − From Deﬁnition 4.16, we see that 23

W . := u .. spt u H (5.25) HE H { ∈HE ⊆ } is a closed subspace of . Let Q k be the projection to this subspace. Then it is clear that 24 HE W Qk Qk+1 and that limk P inu Qku = 0 for all u . Each function u on H 25 ≤ →∞ k F − kE ∈ HE corresponds to a functionu ˜ on G whose support is contained in H; simply deﬁne 26 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 57

u(x), x H, u˜(x) = ∈ u( ), x < H.  ∞ 1 It is clear that this correspondence is bijective, and that

RGW (x, y) = Lxy W C = LxyQk , k k kHE|Gk → k k 2 where the ﬁrst equality follows from (5.5) (recall that Gk is ﬁnite) and therefore

W R (x, y) = lim RGW (x, y) = lim LxyQk = LxyP in , k k k k k k F k →∞ →∞ 3 which is equivalent to (5.23). 4 To see (5.19) is equivalentto (5.20), note that the minimal energy solution to ∆u = δ δ x− y 5 lies in in, since any two solutions diﬀer by a harmonic function. Let u be a solution to F 6 ∆u = δx δy and deﬁne f = P inu. Then f in and ∆ f = δx δy implies − F ∈F −

2 f = f (z)∆ f (z) = f (z)(δx δy)(z) = f (x) f (y). (5.26) k kE − − zXG0 zXG0 ∈ ∈ 7 To see (5.19) (5.23), let κ be the optimal constant from (5.23). If u in is the unique ≤ ∈F 8 solution to ∆u = δ δ , then x − y u(x) u(y) 2 u(x) u(y) 2 κ = sup | − | | − | = u(x) u(y), u in ( (u) ) ≥ (u) − ∈F E E 9 where the last equality follows from (u) = u(x) u(y), by the same computation as in E − 10 (5.26). For the reverse inequality, note that with Lxy as just above,

u(x) u(y) 2 2 2 | − | = L u = v v , u , (u) xy (u)1/2 x − y (u)1/2 E D E EE E 11 for any u in. Note that Lemma 5.21 allows one to replace vx by fx = P invx, whence ∈F F u(x) u(y) 2 | − | ( f f ) u = ( f f ) (u) ≤E x − y E (u)1/2 E x − y E E 12 by Cauchy-Schwarz. The inﬁmum of the left-hand side over nonconstant functions u in ∈F 13 gives the optimal κ in (5.23), and thus shows that (5.23) (5.20). ≤ 14 To see (5.20) is equivalent to (5.21), apply Theorem 3.26 to I = d f , where f = P inu F 15 is the minimal energy solution to ∆u = δ δ . The equivalence of (5.22) and (5.23) is x − y 16 directly parallel to the ﬁnite case.

F W 17 Remark 5.19 (R vs. R explained in terms of boundary conditions on ∆). Observe that 18 both spaces

F . = u .. u(x) u(y) = 0 unless x, y H HE H { ∈HE − ∈ }

19 and

W . = u .. spt u H HE H { ∈HE ⊆ }

58 PALLEE.T.JORGENSENANDERINP.J.PEARSE consist of functions which have no energy outside of H. The diﬀerence is that if the 1 F complement of H consists of several connected components, then u H may take a 2 ∈ HE|W diﬀerent constant value on each one; this is not allowed for elements of H . Therefore, 3 F W HE| corresponds to Neumann boundary conditions and corresponds to Dirichlet 4 HE|H HE|H boundary conditions. That is, from the proofs of Theorem 5.12 and Theorem 5.18, we see 5

(1) RHF (x, y) = u(x) u(y) where u is the solution to ∆u = δx δy with Neumann 6 − ∁ − boundary conditions on H , and 7 (2) RHW (x, y) = u(x) u(y) where u is the solution to ∆u = δx δy under Dirichlet 8 − ∁ − boundary conditions on H . 9

Remark 5.20. While the wired subnetwork takes into account the conductivity due to all 10 paths from x to y (see Remark 5.9), it is overzealous in that it may also include paths from 11 x to y that do not correspond to any path in G (see Remark 5.13). On an infinite network, 12 this leads to current flows in which some of the current travels from x to , and then from 13 ∞ to y. Consider the example of [Mor03]: let G be Z with c , + = 1 for each n. Then 14 ∞ n n 1 define J by 15

1, n , 1 J(n, n 1) = − 0, n = 1.   If a unit current flow from 0 to 1 is defined to be a current satisfying y x I(x, y) = δx δy, 16 ∼ − then J is such a flow which “passes through ” (of course, J certainly not of finite energy). 17 ∞ P

The proof of the next result follows from the ﬁnite case, exactly as in Theorem 5.14. 18

W Theorem 5.21. R (x, y) is a metric. 19

Remark 5.22 (Relating projections in Hilbert space to conditioning of the random walk). 20 On a ﬁnite network, it is well-known that vx = R(o, x)ux, i.e., the vectors are proportional 21 with constant R(o, x), where 22

ux(y) := Py[τx < τo], (5.27) i.e., the value of ux at y should be the probability that a random walker (RW) started at y 23 reaches x before o. See, for example, [DS84, LPW08, LP09]. On inﬁnite networks, this 24 W remains true upon projecting to in, i.e., fx = R (o, x)ux where fx = P invx and ux is again 25 F F as in (5.27). 26 The probabilistic meaning of vx (when arm , 0) is less clear. Figure 7 from Exam- 27 F H F ple 13.2 appears to suggest that vx = R (o, x)ux , where 28

uF(y) := P [τ < τ RW eventually hits x or o], x y x o | that is, one conditions on the RW not meandering off to infinity. This would be in accor- 29 χ dance with the restriction I = ξγ γ of (5.11), which forbids consideration of “paths” 30 through infinity. The authors hopeP to have this matter clarified in the next version of this 31 document. One approach is to consider the trace resistance instead of the free resistance; 32 see Remark 5.34 (especially near the end). 33 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 59

1 5.2.3. Trace resistance. The third type of subnetwork takes into account the connectivity 2 of the complement of the subnetwork, but does not add anything extra. For technical 3 reasons, it is necessary to assume that the network under consideration in this section is 4 transient. However, this is already satisﬁed for networks with arm , 0 by Corollary 4.45. H 5 We believe the results of this section remain true without this assumption, and hope to ﬁnd 6 such a proof in the next version of this paper.

7 Definition 5.23. The trace of G to H is the network whose edge data is defined by the 8 Schur complement of the Laplacian of H within G. More precisely, write the Laplacian of 9 G as a matrix in block form, with the rows and columns indexed by vertices, and order the 10 vertices so that those of H appear first:

H A BT ∆= , (5.28) H∁ " B D # T 0 11 where B is the transpose of B. If ℓ(G) := f : G R , the corresponding mappings are { → }

A :ℓ(H) ℓ(H) BT : ℓ(H∁) ℓ(H) → → B :ℓ(H) ℓ(H∁) D : ℓ(H∁) ℓ(H∁). (5.29) → → 12 It turns out that the Schur complement

T 1 tr(∆, H) := A B D− B (5.30) − 13 is the Laplacian of a subnetwork with the same vertex set as that of H; cf. [Kig01, 2.1] § 14 and Remark 5.33. A formula for the conductances (and hence the adjacencies) of the trace tr 15 is given in Theorem 5.30. Denote this new subnetwork by H . 0 0 16 If H G is finite, then for x, y H, the trace of the resistance on H is denoted ⊆ ∈ 17 RHtr (x, y), and defined as in Definition 5.1. The trace resistance is then defined to be

tr R (x, y) := lim RGtr (x, y), (5.31) k k →∞ 18 where G is any exhaustion of G. { k}

19 Remark 5.24. The name “trace” is due to the fact that this approach comes by considering 20 the trace of the Dirichlet form on a subnetwork; see [FOT94¯ ]. E 1 21 Recall that ∆ = c T = c(I P), where T is the transfer operator and P = c T is the − − − 22 probabilistic transition operator deﬁned pointwise by

cxy Pu(x) = p(x, y)u(y), for p(x, y) = . (5.32) c(x) Xy x ∼ 23 The function p(x, y) gives transition probabilities, i.e., the probability that a randomwalker 24 currently at x will move to y with the next step. Since

c(x)p(x, y) = c(y)p(y, x), (5.33) 0 25 the transition operator P determines a reversible Markov process with state space G ; see 26 [LPW08, LP09]. Note that the harmonic functions (i.e., ∆h = 0) are precisely the ﬁxed 27 points of P (i.e., Ph = h). The proof of the next theorem requires a couple more deﬁnitions 60 PALLEE.T.JORGENSENANDERINP.J.PEARSE

(c) which relate P to the probability measure P on the space of paths in G. Recall from 1 Deﬁnition 2.4 that a path is a sequence of vertices x , where x = a and x x + for 2 { n}n∞=0 0 n ∼ n 1 all n. 3

0 Deﬁnition 5.25. Let Γ(a) denote the space of all paths γ beginning at the vertex a G . 4 ∈ Then Γ(a, b) Γ(a) consists of those paths that reach b, and before returning to a: 5 ⊆

. Γ(a, b) := γ Γ(a) .. b = x for some n, and x , a, 1 k n . (5.34) { ∈ n k ≤ ≤ } If a, b bd H, then we write 6 ∈

. ∁ Γ(a, b) := γ Γ(a, b) .. x H , 0 < i < τ , (5.35) H∁ { ∈ i ∈ b} 0 for the set of paths from a to b that do not pass through any vertex in H . 7

Remark 5.26. Note that if x, y bd H are adjacent, then any path of the form γ = (x, y,... ) 8 ∈ Γ , 9 is trivially in (a b) H∁ .

(c) P (γ) := p(xi 1, xi), (5.36) − Yxi γ ∈ (c) where p(x, y)isasin(5.32). The constructionof P comes by applying Kolmogorov con- 11 sistency to the natural cylinder-set Borel topology that makes Γ(a) into a compact Haus- 12 dorﬀ space; cf. 11 for further discussion. 13 § Deﬁnition 5.28. Let P[a b] denote the probability that a random walk started at a will 14 → reach b before returning to a. That is, 15

P[a b] := P(c)(Γ(a, b)). (5.37) → Note that this is equivalent to 16

P[a b] = P [τ < τ ] := P[τ < τ x = a], (5.38) → a b a b a | 0 where τa is the hitting time of a, i.e., the expected time of the ﬁrst visit to a, after leaving 17 the starting point. If a, b bd H, then we write 18 ∈

P[a b] := P(c) Γ(a, b) , (5.39) → H∁ H∁ that is, the probability that a random walk started at a will reach b via a path lying outside 19 H (except for the start and end points, of course). 20

(c) Remark 5.29. The formulation in (5.39) is conditioning P (Γ(a, b)) on avoiding H; the 21 ∁ notation is intended to evoke something like “P[a b γ H ]”. 22 → | ⊆ Theorem 5.30. If H is a subnetwork of G for which G H is transient, then the conduc- 23 \ tances in the trace are given by 24

ctr = c + c(x)P[x y] . (5.40) xy xy → H∁

OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 61

∁ 1 Proof. Using subscripts to indicate the block decomposition corresponding to H and H 2 as in (5.28), the Laplacian may be written as

c T T T ∆= A − A − B . " TBT cB TD # − − 3 Then the Schur complement is

1 tr(∆, H) = (c T ) ( T T )(c T )− ( T ) A − A − − B D − D − B 1 1 = c c P c P T (I P )− c− c P A − A A − A B − D D D B ∞ n = c c P + P T P P A − A  A B  D B Xn=0     = c (I P ).    (5.41) A − X 4 The assumption of transience ensures the matrix given as an inﬁnite sum has ﬁnite entries; 5 we return to the meaning of PX in a moment. Meanwhile, using PA(x, y) to denote the th 6 (x, y) entry of the matrix PA, we have

P[x y] = P(c) Γ(x, y) → H∁ H∁ (c) ∞ . = P γ Γ(x, y) .. γ = k  { ∈ H∁ | | } [k=1     (c) . ∞ (c) . = P γ Γ(x, y) .. γ = 1 + P γ Γ(x, y) .. γ = k { ∈ H∁ | | } { ∈ H∁ | | } Xk=2

∞ n = PA(x, y) + PBT (x, s)PD(s, t)PB(t, y) Xn=0 Xs,t

∞ k = P + P T P P (x, y). (5.42)  A B D B  Xk=0     k 7 From (5.29), it is clear that PBT k∞=0 PD PB (x, y) is the probability of the random walk ∁ ∁ 8 taking a path that steps from x PH to H , meanders through H for any finite number of ∈ 9 steps, and finally steps to y H. This justifies the probabilistic notation P in (5.41). Note ∈ X 10 Γ that PA(x, y) corresponds to the one-step path from x to y, which is trivially in (x, y) H∁ 11 = = by (5.35). Since PA(x, y) p(x, y) cxy/c(x), the desired conclusion (5.40) follows from 12 combining (5.41),(5.42), and (5.39).

13 The authors are grateful to Jun Kigami for helpful conversations and guidance regard- 14 ing the proof of Theorem 5.30. Dividing through (5.40) by c(x) immediately gives the 15 following corollary.

tr 16 Corollary 5.31. In the trace subnetwork H , the probability of the random walker moving 17 from x to y is given by

ptr(x, y) = p(x, y) + P[x y] . (5.43) → H∁

62 PALLEE.T.JORGENSENANDERINP.J.PEARSE

tr Remark 5.32. It is clear from (5.40) that the edge sets of int H and int H are identical, but 1 tr the conductance between two vertices x, y bd H is greater iﬀ there is a path from x to 2 ∈ y that does not pass through H. Indeed, if there is a path from x to y which lies entirely 3 ∁ tr in H except for the endpoints, then x and y will be adjacent in H , even if they were not 4 adjacent in H. 5

Remark 5.33 (The trace construction is valid for general subsets of vertices). While Def- 6 inition 5.23 applies to a (connected) subnetwork of G, it is essential to note that the con- 7 tr 0 0 struction of the trace H is viable for any subset H of G . The reader will note that the 8 proof of Theorem 5.30 need not be changed to incorporate this extension. In fact, it is not 9 even necessary that H be finite, although we make this assumption in Definition 5.23 to 10 remove any ambiguity. By [Kig01, Thm. 2.1.6], the Schur complement tr(∆, H) will be “a 11 Laplacian” in the sense that: it is a nonnegative definite symmetric operator with positive 12 0 diagonal entries, and has kernel consisting precisely of the constant functions on H . In 13 other words, every row (and column) of tr(∆, H) sums to 0. (This is the negative of the def- 14 inition of a Laplacian as used by Kigami and Colin de Verdiere.) Note that (5.41) can be 15 used to show the last statement of [Kig01, Thm. 2.1.6], that is, that the Schur complement 16 0 of the Laplacian on G is the Laplacian on a network whose vertices are a subset of G . 17

Remark 5.34 (Resistance distance via Schur complement). A theorem of Epifanov states 18 that every ﬁnite planar network with vertices x, y can be reduced to a single equivalent 19 conductor via the use of three simple transformations: parallel, series, and -Y;cf.[Epi66, 20 ∇ Tru89]as well as [LP09, 2.3] and [CdV98, 7.4]. More precisely, 21 § § (1) (2) (i) Parallel. Two conductors cxy and cxy connectedin parallel can be replaced by a single 22 (1) (2) conductor cxy = cxy + cxy . 23 (ii) Series. If z has only the neighbours x and y, then z may be removed from the network 24 1 1 1 and the edges cxz and cyz should be replaced by a single edge cxy = (c−xz + cyz− )− . 25 (iii) -Y. Let t be a vertex whose only neighbours are x, y, z. Then this “Y” may be 26 ∇ replaced by a triangle (“ ”) which does not include t, with conductances 27 ∇

cxtcty cytctz c c c = , c = , c = xt tz . xy c(t) yz c(t) xz c(t)

This transformation may also be inverted, to replace a with a Y and introduce a 28 ∇ new vertex. 29

It is a fun exerciseto obtainthe series and -Y formulas by applying the Schur complement 30 ∇ technique to remove a single vertex of degree 2 or 3 from a network. Indeed, these are both 31 special cases of the following: let t be a vertex of degree n, and let H be the (star-shaped) 32 subnetwork consisting only of t and its neighbours. If we write the Laplacian for just this 33 th subnetwork with the t row & column last, then 34

cx1t ... 0 cx1t . . − .  . .. . .  ∆ = . . . . H   |  0 ... cx t cx t   n − n   cx t ... cx t c(t)   − 1 − n    and the Schur complement is 35 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 63

cx1t ... 0 cx1t . . 1 . tr(∆ , H t ) =  . .. .   .  c ... c , H  . . .  c(t)  .  x1t xnt | \ { }   −   h i  0 ... cx t   cx t   n   n      1 whence the new conductance from xi to x j is given by cxitctx j /c(t). It is interesting to note 2 that the operator being subtracted corresponds to the projection to the rank-one subspace 3 spanned by the probabilities of leaving t:

cx t 1 1  .  c ... c = c(t) v v , c(t)  .  x1t xnt   h i | ih |  cx t   n    4 using Dirac’s ket-bra notation for the projection to a rank-1 subspace spanned by v where

v = p(t, x1) ... p(t, xn) . h i 5 In fact, v v = P , in the notation of (5.41). In general, the trace construction (Schur | ih | X 6 complement) has the effect of probabilistically projecting away the complement of the 7 subnetwork. 8 In Remark 5.4 we described how the effective resistance can be interpreted as the cor- 9 rect resistance for a single edge which replaces a subnetwork. The following corollary of 10 Theorem 5.30 formalizes this interpretation by exploiting the fact that the Schur comple- 11 ment construction is viable for arbitrary subsets of vertices; see Remark 5.33. In this case, 0 1 1 12 one takes the trace of the (typically disconnected) subset x, y G ; note that 1− 1 is { } ⊆ − 13 the Laplacian of the trivial 2-vertex network when the edge between them has unith conduc-i 14 tance. The following result is also an extension of [Kig01, (2.1.4)] to infinite networks.

0 15 Corollary 5.35. Let H = x, y be any two vertices of a transient network G. Then the { } 16 trace resistance can be computed via

1 1 1 T 1 tr(∆, H) = − = A B D− B. (5.44) Rtr(x, y) " 1 1 # − −

17 Proof. Take H = x, y in Theorem 5.30. As discussed in Remark 5.33, it is not necessary { } n 18 to have x y. Note that in this case, P T P P (x, y) corresponds all paths from x to ∼ B n D B 19 y that consist of more than one step: P

∞ n P[x y] = P (x, y) + P T P P (x, y) = p(x, y) + P(γ). → H∁ A  B D B  Xn=0  Xγ 2   | |≥   tr 20 Corollary 5.36. The trace resistance R (x, y) is given by

1 Rtr(x, y) = (5.45) c(x)P[x y] → 64 PALLEE.T.JORGENSENANDERINP.J.PEARSE

0 Proof. Again, take H = x, y . Then 1 { }

tr 1 HS R (x, y)− = c = c + c(x)P[x y] xy xy → H∁

= c(x) p(x, y) + P[x y] → H∁ = c(x)P[x y], → where Corollary 5.35 gives the ﬁrst equality and Theorem 5.30 gives the second. 2

Remark 5.37 (Eﬀective resistance as “path integral”). Corollary 5.36 may also be obtained 3 by the more elegant (and much shorter) approach of [LP09, 2.2], where it is stated as 4 § follows: the mean number of times a random walk visits a before reaching b is P[a 5 1 → b]− = c(a)R(a, b). We give the present proof to highlight and explain the underlying role 6 of the Schur complement with respect to network reduction; see Remarks 5.33–5.34. A 7 key point of the present approach is to emphasize the expression of eﬀective resistance 8 R(a, b)intermsof a sum over all possible paths from a to b. By Remark 5.20, it is apparent 9 tr F that this “path-integral” interpretation makes R much more closely related to R than to 10 W R , as seen by the following results. 11

Corollary 5.38 ([Kig01, Prop. 2.1.11]). Let H2 H1 be ﬁnite subnetworks of a transient 12 0 S ⊆ S network G. Then for a, b H , one hasR (a, b) = R (a, b). 13 ∈ 2 H1 H2 tr F Corollary 5.39. On any transient network, R (a, b) = R (a, b). 14

Proof. By Corollary 5.38, it is clear that R tr (a, b) = R tr (a, b) for all k. Meanwhile, any 15 Gk Gk+1 path from a to b will lie in Gk for suﬃciently large k, so it is clear by Theorem 5.36, the 16 F F tr sequence R (a, b) ∞ is monotonically decreasing with limit R (a, b) = R (a, b). 17 { Gk }k=0 Remark 5.40. Essentially, Corollary 5.38 and Corollary 5.39 both hinge on the idea that 18

1 1 RGtr (x, y) = = RGF (x, y). k c(x) (P[x y G ] + P[x y G ]∁) ≤ c(x)P[x y G ] k → | k → | k → | k where the condition on Gk indicates restriction to paths from x to y that lie entirely in Gk, 19 as in Remark 5.29. Then take the limit k . 20 →∞ 5.3. Harmonic resistance. 21

Definition 5.41. For an infinite network (G, R) define the harmonic resistance between x 22 and y by 23

1 Rha(x, y) := . RW (x, y) 1 RF(x, y) 1 − − −

Theorem 5.42. The harmonic resistance is equal to 24

ha 1 . R (x, y)− = min (h) .. h(x) = 0, h(y) = 1, h arm . (5.46) {E ∈H } ha Remark 5.43. Note: R (x, y) is nota metric. Evenif arm(G) , 0, it may not be possible 25 H to “separate” x and y by a harmonic function, in the sense of Theorem 5.42. For a basic 26 d example, conjoin a copy of (Z , 1)and thebinary tree( , 1), by identifying their origins, as 27 T in Remark 4.58. Then for x, y in the lattice portion of the network, any harmonic function 28 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 65

ha ha 1 with h(x) , h(y) will not have ﬁnite energy. Thus R (x, y) = 0. In fact, R (x, y) may not 2 even be a pseudometric.

3 5.4. Comparison of resistance metric to other metrics.

4 5.4.1. Comparison to shortest-path metric. In this section, we compare the resistance met- 5 ric to the discrete analogue of Riemannian distance. On a Riemannian manifold (Ω, g), the 6 geodesic distance is

1 1/2 . 1 distγ(x, y) := inf g(γ′(t), γ′(t)) dt .. γ(0) = x, γ(1) = y, γ C . γ (Z0 ∈ )

7 Deﬁnition 5.44. On (G, c), the geodesic distance from x to y is

. distγ(x, y) := inf r(γ) .. γ Γ(x, y) , (5.47) { ∈ } 1 8 where r(γ) := (x,y) γ c−xy (for resistors in series, the total resistance is the sum). ∈ P 9 Remark 5.45. Definition 5.44 differs from the definition of shortest path metric found in 10 the literature on general graph theory; without weights on the edges one usually defines the 11 shortest path metric simply as the minimal number of edges in a path from x to y. (This cor- 12 responds to taking c 1.) Such shortest paths always exist. According to Definition 5.44, ≡ 13 shortest paths may not exist (cf. Example 12.10). Of course, even when they do exist, they 14 are not always unique. 15 It should be observed that effective resistance is not a geodesic metric, in the usual sense 16 of metric geometry; it does not correspond to a length structurein thesenseof [BBI01, 2]. § 17 Lemma 5.46. The effective resistance is bounded above by the geodesic distance. More F 18 precisely, R (x, y) distγ(x, y) with equality if and only if G is a tree. ≤ 19 Proof. If there is a second path, then some positive amount of current will pass along it 20 (i.e., there is a positive probability of getting to y via this route). To make this precise, let 21 v = v v and let γ = (x = x , x ,..., x = y) be any path from x to y: x − y 0 1 n

RF (x, y)2 = v(x) v(y) 2 r(γ) (v), | − | ≤ E 22 by the exact same computation as in the proof of Lemma 4.9, but with u = v. The desired F 23 inequality then follows by dividing both sides by (v) = R (x, y). E 24 The other claim follows by observing that trees are characterized by the property of 0 F 25 having exactly one path γ between any x and y in G . By(5.11), R (x, y) can be found by 26 computing the dissipation of the unit current which runs entirely along γ from x to y. This 27 means that I(xi 1, xi) = 1 on γ, and I = 0 elsewhere, so −

n n F 1 2 1 R (x, y) = D(I) = I(xi 1, xi) = = r(γ). cxi 1 xi − cxi 1 xi Xi=1 − Xi=1 −

28 This type of inequality is explicitly calculated in Example 12.3.

29 Remark 5.47. It is clear from the end of the proof of Lemma 5.46 that on a tree, v v x − y 30 is locally constant on the complement of the unique path from x to y. However, this may

31 not hold for fx fy, where fx = P invx; see Example 13.2. This is an example of how the − F 32 wired resistance can “cheat” by considering currents which take a shortcut through inﬁnity; 33 compare (5.11)to(5.21). 66 PALLEE.T.JORGENSENANDERINP.J.PEARSE

5.4.2. Comparison to Connes’ metric. The formulation of R(x, y) given in (5.1) may evoke 1 Connes’maximthat a metric can be thoughtof as the inverseof a Dirac operator; cf. [Con94]. 2 This does not appear to have a literal incarnation in the current context, but we do have the 3 inequality of Lemma 5.48 in the case when c = 1. In this formulation, v is considered 4 ∈HE as a multiplication operator deﬁned on u by 5

(vu)(x) := v(x)u(x), x G0, (5.48) 2 ∀ 0∈ and both v and ∆ are considered as operators on ℓ (G dom . We use the commutator 6 ∩ E 2 notation [v, ∆] := v∆ ∆v, and [v, ∆] is understood as the usual operator norm on ℓ . 7 − k k 0 Lemma 5.48. If c = 1, then for all x, y G one has 8 ∈

2 . R(x, y) sup v(x) v(y) .. [v, ∆] √2, v dom . (5.49) ≤ {| − | k k≤ ∈ E}

Proof. We will compare (5.49)to(5.6). Writing Mv for multiplication by v, it is straight- 9 forward to compute from the deﬁnitions 10

(M ∆ ∆M )u(x) = (v(y) v(x))u(y), v − v − Xy x ∼ so that the Schwarz inequality gives 11

2 2 [Mv, ∆]u = (v(y) v(x))u(y) k k2 − xXG0 Xy x ∈ ∼

v(y) v(x) 2 u(y) 2 . ≤  | − |   | |  xXG0 Xy x  Xy x  ∈  ∼   ∼  2  0    By extending the sum of u(x) to allx G (an admittedly  crude estimate), this gives 12 2 2 | | ∈2 [v, ∆]u 2 u (v), and hence [v, ∆] 2 (v) 13 k k2 ≤ k k2E k k ≤ E 5.5. Generalized resistance metrics. In this section, we describe a notion of eﬀective 14 F W resistance between probability measures, of which R(x, y) (or R and R ) is a special case. 15 This concept is closely related to the notion of total variation of measures, and hence is 16 related to mixing times of Markov chains; cf. [LPW08, 4.1]. When the Markov chain is 17 § taken to be random walk on an ERN, the state space is just the vertices of G. 18 0 Deﬁnition 5.49. Let µ and ν be two probability measures on G . Then the total variation 19 distance between them is 20

distTV(µ, ν) := 2 sup µ(A) ν(A) . (5.50) A G0 | − | ⊆

Proposition 5.50 ([LPW08, Prop. 4.5]). Let µ and ν be two probability measures on the 21 state space Ω of a (discrete) Markov chain. Then the total variation distance between them 22 is 23

.. distTV(µ, ν) = sup u(x)µ(x) u(x)ν(x) . u 1 . (5.51)  − k k∞ ≤   Xx Ω Xx Ω   ∈ ∈      OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 67

1 Here, u := supx G0 u(x) . k k∞ ∈ | |

2 5.5.1. Eﬀective resistance between measures. If we think of µ as a linear functional acting 3 on the space of bounded functions, then it is clear that (5.51) expresses distTV(µ, ν) as the 1 4 operator norm µ ν . That is, it expresses the pairing between µ ℓ and u ℓ∞. We can k −F k ∈ ∈ 5 therefore extend R directly (see (5.13)–(5.14) and Remark 5.11).

6 Deﬁnition 5.51. The free resistance between two probability measures is

2 .. distRF (µ, ν) := sup  u(x)µ(x) u(x)ν(x) . u 1 . (5.52) − k kE ≤  xXG0 xXG0   ∈ ∈     F  7 It is clear from this deﬁnition (and Remark 5.11) that R (x, y) = distRF (δx,δy). This F 8 extension of R to measures was motivated by a question of Marc Rieﬀel in [Rie99].

9 5.5.2. Total variation spaces.

10 Deﬁnition 5.52. Since dom is a Banach space, we may deﬁne a new pairing via the E 11 bilinear form

u, µ := u(x)µ(x), (5.53) h iTV xXG0 ∈ 12 where µ is an element of

0 .. 1/2 TV := µ : G R . kµ s.t. u, µ TV kµ (u) , u dom . (5.54) { → ∃ |h i |≤ ·E ∀ ∈ E} 13 Then TV = dom u, is the dual of dom with respect to the total variation topology h ·iTV E 14 induced by (5.53). Also, the norm in TV is given by

. 1/2 µ := inf k .. u, µ k (u) , u dom . (5.55) k kTV { |h iTV|≤ ·E ∀ ∈ E}

15 Remark 5.53. Since TV is a Banach space which is the dual of a normed space, the unit 16 ball

. µ TV .. µ 1 (5.56) { ∈ k kTV ≤ } 17 is compact in the weak-⋆ topology, by Alaoglu’s theorem.

18 Lemma 5.54. The Laplacian ∆ maps into TV with ∆v TV P inv . HE k k ≤ k F kE 19 Proof. For u, v , write v = f + h with f = P inv and h = P armv, so that ∈HE F H ✟ ✟✟ u(x)∆v(x) u(x)∆ f (x) + ✟u(x✟)∆h(x) = u, f u f , ≤ ✟ |h iE| ≤ k kEk kE xXG0 xXG0 ✟ xXG0 ∈ ∈ ∈ 20 by Theorem 4.68 followed by the Schwarz inequality. The mapping is contractive relative 21 to the respective norms because v is an element of the set on the right-hand side of k kE 22 (5.55), and hence at least as big as the inﬁmum, whence ∆v TV f v . k k ≤ k kE ≤ k kE 68 PALLEE.T.JORGENSENANDERINP.J.PEARSE

6. von Neumann construction of the energy space 1 HE Studying the geometry of state space X through vector spaces of functions on X is a 2 fundamental idea and variations of it can be traced back in several areas of mathematics. 3 In the setting of Hilbert space, it originates with a suggestion of B. O. Koopman [Koo27, 4 Koo36, Koo57] in the early days of “modern” dynamical systems, ergodic theory, and the 5 systematic study of representations of groups. A separate impetus in 1932 were the two 6 2 ergodic theorems, the L variant due to von Neumann [vN32c] and the pointwise variant 7 due to G. D. Birkhoff. While Birkhoff’s version is deeper, von Neumann’s version really 8 started a whole trend: mathematical physics, quantization [vN32c], and operator theory; 9 especially the use of the adjoint operator and the deficiency indices which we find useful in 10 7– 8;cf.[vN32a,vN32b]. Further, there is an interplay between Hilbert space on the one 11 § § 2 side, and pointwise results in function theory on the other: In fact, the L -mean ergodic 12 theoremof von Neumannis really is a corollaryto the spectral theorem in its deeperversion 13 (spectral resolution via projection-valued measures) as developed in by M. H. Stone and 14 J. von Neumann in the period 1928-1932; cf. [vN32b] and [Arv76b, Ch. 2]. This legacy 15 motivates the material in this section, as well as our overall approach. 16

6.1. von Neumann’s embedding theorem. In Theorem 6.1 we show that an electrical 17 resistance network equipped with resistance metric may be embedded in a Hilbert space in 18 such a way that R is inducedfromthe innerproductof the Hilbert space. As a consequence, 19 we obtain an alternative and independent construction of the Hilbert space of finite- 20 HE energy functions. This provides further justification for as the natural Hilbert space 21 F F HE for studying the metric space (G, R ) = ((G, c), R ) and in as the natural Hilbert space 22 W F F for studying the metric space (G, R ). Although we will be interested in both (G, R ) 23 W and (G, R ), for brevity, we sometimes refer to both as (G, R) when the distinction is not 24 important. 25 When studying a metric space, it is a natural question to ask whether or not the space 26 may be represented as a Hilbert space, and von Neumann proved a general result which 27 provides an answer. The reader may wish to consult A.1 for the statement of this result 28 § (Theorem A.1) in the form it is applied below, as well as the relevant definitions. We apply 29 Theorem A.1 to the metric space (G, R) and to obtain a Hilbert space and a natural 30 H embedding (G, R) . It turns out that the Hilbert space is when the embedding is 31 →F H W HE applied with R = R and in when applied with R = R ; see Remark 6.3! Therefore, the 32 F Hilbert space is the natural place to study (G, R). The reader may find the references 33 HE [vN32a, BCR84, Ber96, Sch38b] to be helpful; see also Theorem 7.18. 34 The followingtheoremis inspiredby the work ofvon Neumannand Schoenberg [Sch38a, 35 BCR84], but is a completelynew result. One aspectof this result that contrasts sharply with 36 1/2 the classical theory is that the embedding is applied to the metric R instead of R, for each 37 F W of R = R and R = R . 38

F Theorem 6.1. (G, R ) may be isometrically embedded in a Hilbert space. 39

F Proof. According to Theorem A.1, we need only to check that R is negative semidefi- 40 0 nite (see Definition A.2). Let f : G R satisfy x G0 f (x) = 0. We must show that 41 F → ∈0 x,y F f (x)R (x, y) f (y) 0, for any finite subset F PG . From(5.10), we have 42 ∈ ≤ ⊆ P f (x)RF (x, y) f (y) = f (x) (v v ) f (y) E x − y xX,y F xX,y F ∈ ∈ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 69

= f (x) (vx) f (y) 2 f (x) vx, vy f (y) + f (x) (vx) f (y) E − h iE E xX,y F xX,y F xX,y F ∈ ∈ ∈

= 2 f (x)vx, f (y)vy − * + Xx F Xy F ∈ ∈ E 2 = 2 f (x)v 0. − x ≤ Xx F ∈ E 1 For the second equality, note that the ﬁrst two sums vanish by the assumption on f . W 2 Corollary 6.2. (G, R ) may be isometrically embedded in a Hilbert space.

3 Proof. Because the energy-minimizer in (5.20) is fx = P invx, we can repeat the proof of F 4 Theorem 6.1 with fx in place of vx to obtain the result. F 2 5 Remark 6.3. Since R (x, y) = vx vy by (5.10), Theorem A.3 shows that the embedded F k − kE 6 image of (G, R ) is unitarily equivalent to the -closure of span vx , which is . Simi- W 2 E { } HE 7 larly, R (x, y) = fx fy , where fx := P invx,by(5.20), whence the embedded image of W k − kE F 8 (G, R ) is unitarily equivalent to the -closure of span f , which is in. E { x} F Remark 6.4. One can choose any vertex o G0 to act as the “origin” and it becomes the origin of the new Hilbert space during the∈ construction outlined in A.1. As a quadratic form defined on the space of all functions v : G0 C, the energy is§ indefinite and hence allows one to define only a quasinorm. There are→ ways to deal with the fact that does not “see constant functions”. One possibility is to adjust the energy so as to obtainE a true norm, as follows: o(u, v) := (u, v) + u(o)v(o). (6.1) E E 9 The corresponding quadratic form is immediately seen to be a norm; this approach is car- 10 ried out in [FOT94¯ ], for example, and also occasionally in the work of Kigami. 11 We have instead elected to work “modulo constants” but this can be interpreted as deriv- 12 ing from the alternative approach of (6.1). The kernel of is the set of constant functions, E 13 and inspection of von Neumann’s embedding theorem (cf. (A.6)) shows that it is precisely 14 these functions which are “modded out” in von Neumann’s construction.

15 6.2. as an invariant of G. In this section, we show that may be considered as an HE HE 16 invariant of the underlying graph. Definition 6.5. Let G and H be electrical resistance networks with respective conductances cG and cH. A morphism of electrical resistance networks is a function ϕ :(G, cG) (H, cH) between the vertices of the two underlying graphs for which → cH = rcG , 0 < r < , (6.2) ϕ(x)ϕ(y) xy ∞ 0 17 for some fixed r and all x, y G . Two electrical resistance∈ networks are isomorphic if there is a bijective morphism between them. Note that this implies 1 . 1 H = (ϕ(x), ϕ(y)) .. (x, y) G . (6.3) { ∈ } Definition 6.6. A morphism of metric spaces is a homothetic map, that is, an isometry composed with a dilation: ϕ :(X, d ) (Y, d ), d (ϕ(a), ϕ(b)) = rd (a, b), 0 < r < , (6.4) X → Y Y X ∞ 18 for some fixed r and all a, b X. An isomorphism is, of course, an invertible morphism. ∈ 70 PALLEE.T.JORGENSENANDERINP.J.PEARSE

We allow for a scaling factor r in each of the previous deﬁnitions because an isomor- 1 phism amounts to a relabeling, and rescaling is just a relabeling of lengths. More formally, 2 an isomorphism in any category is an invertible mapping, and dilations are certainly invert- 3 ible for 0 < r < . 4 ∞ F W Theorem 6.7. For each of R = R , R , there is a functor :(G, c) ((G, c), R) from the 5 R → category of electrical resistance networks to the category of metric spaces. 6

Proof. One must check that an isomorphism ϕ :(G, c ) (H, c ) of electrical resistance 7 G → H networks induces an isomorphism of the corresponding metric spaces, so check that ϕ 8 preserves . We use x, y to denote vertices in G and s, t to denote vertices in H. 9 E

u ϕ, v ϕ = cxy(u ϕ(x) u ϕ(y))(v ϕ(x) v ϕ(y)) h ◦ ◦ iE ◦ − ◦ ◦ − ◦ Xx,y 1 = r− cϕ ϕ (u(ϕ(x)) u(ϕ(y)))(v(ϕ(x)) v(ϕ(y))) (x) (y) − − Xx,y 1 = r− c (u(s) u(t))(v(s) v(t)) st − − Xs,t 1 = r− u ϕ, v ϕ , (6.5) h ◦ ◦ iE where we can change to summing over s, t because ϕ is a bijection. Therefore, the re- 10 F W producing kernel vx of (G, R ) (or P invx of (G, R )) is preserved, and hence so is the 11 { } { F } metric. 12

Corollary 6.8. If π is an isomorphism of ERNs with scaling ratio r, then 13

1 ∆(v ϕ) = r− ∆(v) ϕ. (6.6) ◦ ◦

Proof. Compute ∆(v ϕ)(x) exactly as in (6.5). 14 ◦ G H Corollary 6.9. An isomorphism ϕ :(G, c ) (H, c ) of electrical resistance networks 15 → induces an isomorphism of metric spaces (where the electrical resistance networks are 16 equipped with their respective eﬀective resistance metrics). 17

We use the notation [S ] to denote the closure of the span of a set of vectors S in a 18 Hilbert space, where the closure is taken with respect to the norm of the Hilbert space. 19 The following theorem is just an application of Theorem A.3 with the quadratic form Q˜ = 20 , . 21 h· ·iK Theorem 6.10. If there is a Hilbert space = [kx] for some k : X with d(x, y) = 22 2 K → K kx ky , then there is a unique unitary isomorphism U : and it is given by 23 k − kK H → K U : ξ w ξ k . 24 x x x 7→ x x x P P Remark 6.11. Theorem 6.10 may be interpreted as the statement that there is a functor 25 from the category of metric spaces (with negative semidefinite metrics) into the category 26 of Hilbert spaces. However, we have avoided this formulation because the functor is not 27 defined for the entire category of metric spaces. For us, it suffices to note that the composi- 28 tion is a functor from electrical resistance networks to Hilbert spaces, so that = (G) 29 HE HE is an invariant of G. 30 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 71

1 Remark 6.12. To obtain a ﬁrst quantization, one would need to prove that a contractive 2 morphism between electrical resistance networks induces a contraction between the corre- 3 sponding Hilbert spaces. In other words,

f : G1 G2 = T f : (G1) (G2) → ⇒ HE →HE 4 with T f v v whenever f is contractive. The authors are currently working on this k kE ≤ k kE 5 endeavour in [JP08b]. The second quantization is discussed in Remark 7.22. 72 PALLEE.T.JORGENSENANDERINP.J.PEARSE

7. The boundary and boundary representation 1 7.1. Motivation and outline. Recall the classical result of Poisson that gives a kernel k : Ω ∂Ω R from which a bounded harmonic function can be given via × → u(x) = u(y)k(x, dy), y ∂Ω. (7.1) Z∂Ω ∈

The material of 7 is motivated by the following discrete analogue of the Poisson boundary 2 § representation of a harmonic function. 3

Theorem 7.1 (Boundary sum representation of harmonic functions). For all u arm, 4 ∈ H and hx = P armvx, 5 H

u(x) = u ∂hx + u(o). (7.2)

∂Ò Xbd G

Proof. Recall from Lemma 4.26 that hx is a reproducing kernel for arm. Therefore, 6 { }∂hx H

u(x) u(o) = hx, u = u, hx = u because 0 u∆hx = 0. Note that hx = hx by 7 bd G Ò G − h iE h iE ∂ Lemma 4.29. P P 8

Up to this point, the boundary sum in (7.2) has been understood only as a limit of 9 sums. Comparison of (7.2) and (7.1) makes one optimistic that bd G can be realized as 10 ∂hx some compact set which supports a “measure” , thus giving a nice representation of the 11 ∂Ò boundarysum of (7.2) as an integral. In Corollary 7.27, we extend Theorem 7.1 to such an 12 integral representation. 13 Boundary theory of harmonic functions can roughly be divided three ways: the bounded 14 harmonic functions (Poisson theory), the nonnegative harmonic functions (Martin theory), 15 and the finite-energy harmonic functions studied in the present paper. While Poisson the- 16 ory is a subset of Martin theory, the relationship between Martin theory and the study of 17 is more subtle. For example, there exist unbounded functions of finite energy; cf. Ex- 18 HE ample 14.30. Some details are given in [Soa94, 3.7]. 19 § Whether the focus is on the harmonic functions which are bounded, nonnegative, or 20 finite-energy, the goals of the associated boundary theory are essentially the same: 21

(1) Compactify the original domain by constructing/identifying a boundary bd . 22 D D Then = bd , where the closure is with respect to some (hopefully natural) 23 D D∪ D topology. 24 (2) Deﬁne a procedure for extending harmonic functions u from to bd . Except in 25 D D the case of Poisson theory, this extensionu ˜ is typically a measure (or other linear 26 functional) on bd ; it may not be representable as a function. 27 D

(3) Obtain a kernel k(x, β) deﬁned on bd against which one can integrate the extensionu ˜ so as to recover the valueD× of u atD a point in : D

u(x) = k(x, β)˜u(dβ), x , Zbd ∀ ∈ D D whenever u is a harmonic functions of the given class. 28

The difference between our boundary theory and that of Poisson and Martin is rooted in 29 2 our focus on rather than ℓ : both of these classical theories concern harmonic functions 30 HE with growth/decay restrictions. By contrast, provided they neither grow too wildly nor 31 oscillate too wildly, elements of may remain positive or even tend to infinity at infinity. 32 HE See Example 13.10 for a function h arm which is unbounded. 33 ∈H OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 73

1 Our boundaryessentially consists of infinite paths which can be distinguished by monopoles, 2 i.e., two paths are not equivalent iff there is a monopole w with different limiting values 3 along each path. It is an immediate consequence that recurrent networks have no boundary, 4 and transient networks with no nontrivial harmonic functions have exactly one boundary 5 point (corresponding to the fact that the monopole at x is unique). In particular, the integer d 6 lattices (Z , 1) each have 1 boundary point for d 3 and 0 boundary points for d = 1, 2. In 2 ≥ 2 7 contrast, the Martin boundary of (Z , 1) consists of two points, and (Z , 1) has two graph 8 ends; cf. [PW90]. 9 In our version of the program outlined above, we follow the steps in the order (2)-(3)- 10 (1). A brief summary is given here; further introductory material and technical details 11 appear at the beginning of each subsection.

12 For (2), we construct a Gel’fand triple G G′ to extend the energy form to a S ⊆ HE ⊆ S 13 pairing , on G G′ , and then use Ito integration to extend this new pairing h· ·iW S × S 14 to G′ . This yields a suitable class of linear functionals ξ on , and we can HE × S HE 15 extend a function u on tou ˜ on G′ by duality, i.e.,u ˜(ξ) := u, ξ . We need to HE S h iW 16 expand the scope of enquiry to include G′ because will not be sufficient; no S HE 17 infinite-dimensional Hilbert space can support a σ-finite probability measure, by 18 a theorem of Nelson. 2 19 For (3), we use the Wiener transform to isometrically embed into L ( G′ , P). Applying HE S

20 this isometry to the energy kernel v , we get a reproducing kernel k(x, dP) := { x} 21 hxdP, where hx = P armvx and P is a version of Wiener measure. In fact, P is a H 22 Gaussian probability measure on whose support is disjoint from in. SG′ F 23 For (1), we consider certain measures µx, deﬁned in terms of the kernel and the Wiener

24 measure just introduced, which are supported on G′ / in and indexed by the ver- 0 S F 25 tices x G . Then elements of the boundary bd G correspond limits of sequences ∈ 26 µ where x , modulo a suitable equivalence relation. This is the content { xn } n → ∞ 27 of 7.3. § 28 Items (2)–(3) are the content of 7.2 and the main result is Theorem 7.19 (and its corollar- § 29 ies). Due to the close relationship between the Laplacian and the random walk on a net- 30 work, there are good intuitive reasons why one would expect stochastic integrals (by which 31 we mean the Wiener transform) to be related to the boundary. “Going to the boundary” 32 of (G, c) involves a suitable notion of limit, and it is a well-known principle that suitable 2 33 limits of random walk yield Brownian motion realized in L -spaces of global measures 34 (e.g., Wiener measure). 35 However, before this program can proceed, we need a suitable dense subspace SG ⊆ 36 of “test functions” for the construction of a Gel’fand triple. The basic idea is to use HE 37 the “smooth functions”, that is, u for which ∆(... ∆(u)) , for any number of ∈ HE ∈ HE 38 applications of ∆. Making this precise requires a certain amount of attention to technical 39 details concerning the domain of ∆, and this comprises 8.1.2. Caution: when studying an § 40 operator, an important subtlety is that “the” adjoint ∆∗ depends on the choice of domain, 41 i.e., the linear subspace dom(∆) . We consider ∆ as an operator on a rather different 2 0 ⊆ H 42 Hilbert space, ℓ (G ), in 9. § 43 Finally, in 8.2.1, we examine the connection between the defect spaces of ∆ and bdG § 44 via the use of a boundary form akin to those of classical functional analysis. 45 The reader is directed to Appendix B for a brief review of some of the pertinent ideas 46 from operator theory; especially regarding the graph of an operator (Definition B.12) and 47 von Neumann’s theorem characterizing essential self-adjointness (Theorem B.18). Note: 48 in several parts of this section, we use vector space ideas that are not so common when 74 PALLEE.T.JORGENSENANDERINP.J.PEARSE discussing Hilbert spaces; e.g. finite linear span, and (not necessarily orthogonal) linear 1 independence. 2

Remark 7.2. In 11 we will return to the three-way comparison of harmonic functions 3 § which are bounded, nonnegative, or finite-energy, but for a different purpose: the construc- 4 tion of measures on spaces of (infinite) paths in (G, c). In the case of bounded harmonic 5 functions on (G, c), the associated probability space is derived directly as a space of infi- 6 nite paths in G, and the measure is constructed via the standard Kolmogorov consistency 7 method. That is, as a projective limit constructed via cylinder sets. While the present con- 8 struction is also implicitly in terms of cylinder sets (due to Minlos’ framework), the reader 9 will notice by comparison that the two probability measures and their support are quite 10 different. As a result the respective kernels take different forms. However, both techniques 11 0 yield a way to represent the values h(x) for h harmonic and x G as an integral over “the 12 ∈ boundary”. 13 2 While Doob’s martingale theory works well for harmonic functions in L∞ or L , the 14 situation for is different. The primary reason is that is not immediatelly realizable 15 2 HE HE as an L space. A considerable advantage of our Gel’fand-Wiener-Ito construction is that 16 2 is isometrically embedded into L ( G′ , P) in a particularly nice way: it corresponds to 17 HE S the polynomials of degree 1. See Remark 7.22. 18

Recall that Corollary 4.13 shows that span vx is dense in and that vx is a reproduc- 19 { } HE { } ing kernel for . Throughout 7, we will implicitly be using the version of ∆ introduced 20 HE § in Deﬁnition 4.32, which we now recall for convenience. 21

Definition 7.3. Let V := span vx x G0 denote the vector space of finite linear combinations 22 { } ∈ of dipoles. Let ∆ be the closure of the Laplacian when taken to have the dense domain V. 23 V Note that since ∆ agrees with ∆ pointwise, we may suppress reference to the domain 24 V for ease of notation. Recall from Corollary 4.61 that ∆ is Hermitian and even semi- 25 V bounded on its domain. We explore the properties of ∆ further, including its range, 26 V domain, and self-adjoint extensions, in 8. 27 § 7.2. Gel’fand triples and duality. According to the program outlined above, we would 28 like to obtain a (probability) measure space to serve as the boundary of G. It is shown 29 in [Gro67, Gro70, Min63] that no Hilbert space of functions is sufficient to support a 30 H Gaussian measure P (i.e., it is not possible to have 0 < P( ) < for a σ-finite measure). 31 H ∞ However, it is possible to construct a Gel’fand triple (also called a rigged Hilbert space): 32 a dense subspace S of with 33 H

S S ′, (7.3) ⊆H⊆ where S is dense in and S is the dual of S . Additionally, S and S must also satisfy 34 H ′ ′ some technical conditions: S is a Fréchet space in its own right but realized as dense sub- 35 space in , with density referring to the Hilbert norm in . However, S is the dual of 36 H H ′ S with respect to a Fréchet topology defined via a specific sequence of seminorms. Fi- 37 nally, it is assumed that the inclusion mapping of S into is continuous in the respective 38 H topologies. It was Gel’fand’s idea to formalize this construction abstractly using a sys- 39 tem of nuclearity axioms [GMS58ˇ , Min58, Min59]. Our presentation here is adapted from 40 quantum mechanics and the goal is to realize bd G as a subset of S ′. 41 There is a concrete situation when the Gel’fand triple construction is especially natural: 42 2 = L (R, dx) and S is the Schwartz space of functions of rapid decay. That is, each 43 H f S is C∞ smooth functions which decays (along with all its derivatives) faster than 44 ∈ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 75

1 any polynomial. In this case, S is the space of tempered distributions and the seminorms 2 deﬁning the Fr´echet topology on S are

k (n) . p ( f ) := sup x f (x) .. x R, 0 k, n m , m = 0, 1, 2,..., m {| | ∈ ≤ ≤ } (n) th 3 where f is the n derivative of f . Then S ′ is the dual of S with respect to this Fre´echet 4 topology. One can equivalently express S as

2 . 2 2 n 2 S := f L (R) .. (P˜ + Q˜ ) f L (R), n , (7.4) { ∈ ∈ ∀ } 5 where P˜ and Q˜ are the Heisenberg operators discussed in Example B.25. The operator 2 2 6 P˜ + Q˜ is most often called the quantum mechanical Hamiltonian, but some others (e.g., 7 Hida, Gross) would call it a Laplacian, and this perspective tightens the analogy with the 8 present study. In this sense, (7.4) could be rewritten S := dom ∆∞; compare to (7.8). 9 The duality between S and S allows for the extension of the inner product on to a ′ H 10 pairing of S and S ′:

, : C to , ∼ : S S ′ R. h· ·iH H×H→ h· ·iH × → 11 In other words, one obtains a Fourier-type duality restricted to S . Moreover, it is possible 12 to construct a Gel’fand triple in such a way that P(S ′) = 1 for a Gaussian probability 13 measure P. When applied to = , the construction yields two main outcomes: H HE 14 (1) The next best thing to a Fourier transform for an arbitrary graph. 2 2 15 (2) A concrete representation of as an L measure space L (S ′, P). HE HE 16 The reader may find much useful background information for this material in [Arv76a, 17 Arv76c]. As a prelude, we begin with Bochner’s Theorem, which characterizes the Fourier 18 transform of a positive finite Borel measure on the real line. The reader may find [RS75] 19 helpful for further information.

20 Theorem 7.4 (Bochner). Let G be a locally compact abelian group. Then there is a bijec- 21 tive correspondence : (G) (Gˆ), where (G) is the collection of measures on F M → PD M 22 G, and (Gˆ) is the set of positive deﬁnite functions on the dual group of G. Moreover, PD 23 this bijection is given by the Fourier transform

i ξ,x : ν ϕν by ϕν(ξ) = e h i dν(x). (7.5) F 7→ ZG

d 24 In our applications to the electrical resistance network (Z , 1) in 14, the underlying § 25 group structure allows us to apply the above version of Bochner’s theorem. Specifically, 26 in the context of group duality, Bochner’s theorem characterizes the Fourier transform of a 27 positive finite Borel measures; cf. [RS75, Ber96]. 28 For our representation of the energy Hilbert space in the case of general electri- HE 29 cal resistance network, we will need Minlos’ generalization of Bochner’s theorem from 30 [Min63, Sch73]. This important result states that a cylindrical measure on the dual of a 31 nuclear space is a Radon measure iff its Fourier transform is continuous. In this context, 32 however, the notion of Fourier transform is infinite-dimensional, and is dealt with by the 33 introduction of Gel’fand triples [Lee96].

34 Theorem 7.5 (Minlos). Given a Gel’fand triple S S , Bochner’s Theorem may be ⊆H ⊆ ′ 35 extended to yield a bijective correspondence between the positive deﬁnite functions on S 76 PALLEE.T.JORGENSENANDERINP.J.PEARSE and the Radon probability measures on S ′. Moreover, in a speciﬁc case, this correspon- 1 dence is uniquely determined by the identity 2

1 i u,ξ ˜ 2 u,u e h iH dP(ξ) = e− h iH , (7.6) ZS ′ where , is the original inner product on and , ˜ is its extension to the pairing 3 h· ·iH H h· ·iH on S S . 4 × ′ Formula (7.6) may be interpreted as defining the Fourier transform of P; the function on 5 the right-hand side is positive definite and plays a special role in stochastic integration, and 6 its use in quantization. To apply Minlos’ Theorem, we first need to construct a Gel’fand 7 triple for ; we begin by identifying a certain subspace of the domain of ∆ . Recall from 8 HE V Definition 4.32 that V := span vx x G0 . 9 { } ∈ Definition 7.6. Let ∆ be a self-adjoint extension of ∆ ; since ∆ is Hermitian and com- 10 ∗ V V V mutes with conjugation (since c is R-valued), a theorem of von Neumann’s states that such 11 an extension exists. 12 p ∆ = ∆ ∆ ∆ ∆ 13 Let ∗ u : ( ∗ ∗ ... ∗ )u be the p-fold product of ∗ applied to u . Define p V V V V V ∈ HE ∆ 14 dom( ∗ ) inductively by V p . p 1 ∆ = .. ∆ − ∆ dom( ∗ ) : u ∗ u dom( ∗ ) . (7.7) V { V ∈ V }

Deﬁnition 7.7. The (Schwartz) space of functions of rapid decay is 15

= ∆ G : dom( ∗ ∞), (7.8) p S V ∆ = ∆ R 16 where dom( ∗ ∞) : ∞p=1 dom( ∗ ) consists of all -valued functions u for which p V V ∈ HE ∆ 17 ∗ u for any p.T The space of Schwartz distributions or tempered distributions is the V ∈ HE dual space of R-valued continuous linear functionals on . 18 SG′ SG

Remark 7.8. A good choice of self-adjoint extension in Definition 7.6 is the operator ∆H 19 discussed in 8.1.2. It is critical to make the unusual step of taking a self-adjoint extension 20 § of ∆ for several reasons. Most importantly, we will need to apply the spectral theorem 21 V to extend the energy inner product , to a pairing on G G′ . In fact, it will turn 22 h· ·iE S × S p p = ∆ ∆− 23 out that for u G, v G′ , the extended pairing is given by u, v ∗ u, ∗ v , ∈ S ∈ S p p h iW h V V iE ∆ ∆− 24 where p is any integer large enough to ensure ∗ u, ∗ v . This relies crucially on V V ∈ HE the self-adjointness of the operator appearing on the right-hand side. Moreover, without 25 self-adjointness, we would be unable to prove that G is dense in ; see Lemma 7.12. 26 S HE Additionally, the self-adjoint extensions of ∆ are in bijective correspondence with the 27 V isotropic subspaces of dom(∆∗ ), and we will see that these are useful for understandingthe 28 V boundary of G in terms of defect; see Theorem 8.19. Recall that a subspace dom(∆∗ ) 29 M⊆ V is isotropic iff βbd(u, v) = 0, u, v , where βbd is as in Definition 8.17. Since dom(∆ ) 30 ∀ ∈M V is isotropic (cf. Theorem 8.18), we think of as a subspace of the quotient (boundary) 31 M space B = dom(∆∗ )/ dom(∆ ); we return to this in (7.32). 32 V V

Remark 7.9. Note that G and G′ consist of R-valued functions. This technical detail is 33

S S i u, ˜ importantbecause we do not expect the integral e h ·i dP from (7.6) to converge unless 34 S ′ W it is certain that u, is R-valued. This is the reasonR for the last conclusion of Lemma 7.14. 35 h ·i Remark 7.10. Note that G is dense in dom(∆ ) with respect to the graphnorm, by standard 36 S ∗ V spectral theory. For each p N, there is a seminorm on deﬁned by 37 ∈ SG OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 77

p u p := ∆ u . (7.9) ∗ E p k k k V k 1 ∆ N Since (dom ∗ , p) is a Hilbert space for each p , the subspace G is a Fréchet space. V k· k ∈ S 2 Definition 7.11. Let χ denote the usual indicator function of the interval [a, b] R, [a, b] ⊆ 3 and let S be the spectral transform in the spectral representation of ∆ , and let E be the as- ∗ V 4 sociated projection-valued measure. Then define En to be the spectral truncation operator 5 acting on by HE n E u := S∗χ Su = E(dt)u. n [ 1 , n] n Z1/n

6 Lemma 7.12. G is a dense analytic subspace of (with respect to ), and so G S HE E S ⊆ 7 G′ is a Gel’fand triple. HE ⊆ S

8 Proof. This essentially follows immediately once it is clear that En maps into G. For HE S 9 u , and for any p = 1, 2,... , ∈HE n ∆p 2 = 2p 2 2p 2 ∗ Enu λ E(dλ)u n u , (7.10) k V kE Z1/n k kE ≤ k kE

10 So Enu G. It follows that u Enu 0 by standard spectral theory. ∈ S k − kE →

11 Theorem 7.13. The energy form , extends to a pairing on G G′ deﬁned by h· ·iE S × S p p = ∆ ∆− u, v : ∗ u, ∗ v , (7.11) h iW h V V iE p 12 where p is any integer such that v(u) K ∆ u for all u G. | |≤ k kE ∈ S p 13 ∆ Proof. If v G′ , then there is a C and p such that s, v C ∗ s for all s G. p ∈ S |h iW| ≤ k V kE ∈ S 14 ∆ = Set ϕ( ∗ s) : s, v to obtain a continuous linear functional on (after extending to V h iW p HE 15 ∆ the orthogonal complement of span ∗ s by 0 if necessary). Now Riesz’s lemma gives a p { V } p 16 = ∆ ∆− = w for which s, v ∗ s, w for all s G and we deﬁne ∗ v : w to ∈ HE h iW h V iE ∈ S V ∈ HE 17 make the meaning of the right-hand side of (7.11) clear.

18 Lemma 7.14. The pairing on is equivalently given by SG × SG′

u, ξ = lim ξ(Enu), (7.12) W n h i →∞ 19 where the limit is taken in the topology of G′ . Moreover, u˜(ξ) = u, ξ is R-valued on G′ . S h iW S 20 Proof. En commutes with ∆ . This is a standard result in spectral theory, as En and ∆ ∗ V ∗ V 21 are unitarily equivalent to the two commuting operations of truncation and multiplication, 22 respectively. Therefore,

p p p p p p = = ∆ ∆− = ∆ ∆− = ∆ ∆− ξ(Enu) Enu, ξ ∗ En s, ∗ ξ En ∗ s, ∗ ξ ∗ s, En ∗ ξ . h iW h V V iE h V V iE h V V iE 23 Standard spectral theory also gives Env v in , so → HE

p p p p = ∆ ∆− = ∆ ∆− lim ξ(Enu) lim ∗ s, En ∗ ξ ∗ u, ∗ v . n n V V E V V E →∞ →∞h i h i 24 Note that the pairing , is a limit of real numbers, and hence is real. h· ·iW 78 PALLEE.T.JORGENSENANDERINP.J.PEARSE

˜ ˜ Corollary 7.15. En extends to a mapping En : G′ deﬁned via u, Enξ := ξ(Enu). S →HE h iE Thus, we have a pointwise extension of , to G′ given by h· ·iW HE × S u, ξ = lim u, E˜nξ . (7.13) W n E h i →∞h i

0 Lemma 7.16. If deg(x) is ﬁnite for each x G , or if c < , then one has v . 1 ∈ k k ∞ x ∈ SG

Proof. This is immediate from the technical lemma, Lemma 8.3, which we postpone for 2 now. 3

Remark 7.17. When the hypotheses of Lemma 7.16 are satisﬁed, it should be noted that span vx is dense in G with respect to , but not with respect to the Frechet topology induced{ } by the seminormsS (7.9), nor withE respect to the graph norm. One has the inclusions v s u x (7.14) (" ∆ vx #) ⊆ (" ∆ s #) ⊆ (" ∆ u #) V ∗ V ∗ V where s G and u , with the second inclusion dense and the ﬁrst inclusion not 4 ∈ S ∈ HE dense. 5

We have now obtained a Gel’fand triple G G′ , and we are ready to apply 6 S ⊆ HE ⊆ S the Minlos Theorem to a particularly lovely positive deﬁnite function on , in order that 7 SG we may obtain a particularly nice measure on G′ . Recall that we constructed from the 8 S HE resistance metric in 6 by making use of negative deﬁnite functions. We now apply this to 9 § a famous result of Schoenberg which may be found in [BCR84, SW49]. 10

Theorem 7.18 (Schoenberg). Let X bea set and let Q : X X R be a function. Then 11 × → the following are equivalent. 12

(1) Q is negative definite. 13 + tQ(x,y) (2) t R , the function p (x, y) := e is positive definite on X X. 14 ∀ ∈ t − × (3) There exists a Hilbert space and a function f : X such that Q(x, y) = 15 2 H → H f (x) f (y) . 16 k − kH 1 In the proof of the following theorem, we apply Schoenberg’s Theorem with t = 2 to the 17 F 2 resistance metric in the form R (x, y) = vx vy from (5.10). The proof of Theorem 7.19 18 k − kE also uses the notation Eξ( f ) := f (ξ) dP(ξ). 19 ′ RSG 2 Theorem 7.19. The Wiener transform : L ( G′ , P) defined by 20 W HE → S : v v˜, v˜(ξ) = v, ξ , (7.15) W 7→ h iW is an isometry. The extended reproducing kernel v˜x x G0 is a system of Gaussian random 21 { } ∈ variables which gives the resistance distance by 22

F 2 R (x, y) = Eξ((˜v v˜ ) ). (7.16) x − y Moreover, for any u, v , the energy inner product extends directly as 23 ∈HE

u, v = Eξ u˜v˜ = u˜vd˜ P. (7.17) h iE Z SG′ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 79

F 1 Proof. Since R (x, y) is negative semideﬁnite by the proof of Theorem 6.1, we may apply 1 2 2 Schoenberg’s theorem and deduce that exp( 2 u v ) is a positive deﬁnite function on − k − kE 3 . Consequently, an application of the Minlos correspondence to the Gel’fand triple HE ×HE 4 established in Lemma 7.12 yields a Gaussian probability measure P on . SG′ 5 Moreover, (7.6) gives

1 2 i u,ξ 2 u Eξ(e h iW ) = e− k kE , (7.18)

6 whence one computes

1 2 1

1 + i u, ξ u, ξ + dP(ξ) = 1 u, u + . (7.19) Z h iW − 2h iW ··· ! − 2h iE ··· SG′ 2 2 2 7 Now it follows that E(˜u ) = Eξ( u, ξ ) = u for every u G, by comparing the terms h iW k kE ∈ S 8 of (7.19) which are quadratic in u. Therefore, : G′ is an isometry, and (7.19) W HE → S 9 gives

2 2 2 Eξ( v˜x v˜y ) = Eξ( vx vy, ξ ) = vx vy , (7.20) | − | h − i k − kE 10 whence (7.16) follows from (5.10). Note that by comparing the linear terms, (7.19) implies 2 2 11 Eξ(1) = 1, so that P is a probability measure, and Eξ( u, ξ ) = 0 and Eξ( u, ξ ) = u , so h i h i k kW 12 that P is actually Gaussian. 13 Finally, use polarization to compute

1 u, v = u + v 2 u v 2 E 4 h i k kE − k − kE 1 = E u˜ + v˜ 2 E u˜ v˜ 2 by (7.20) 4 ξ ξ | | − | − | 1 = u˜ + v˜ 2 (ξ) u˜ v˜ 2 (ξ) dP(ξ) 4 Z | | −| − | SG′ = u˜(ξ)˜v(ξ) dP(ξ). Z SG′ 14 This establishes (7.17) and completes the proof.

15 It is important to note that since the Wiener transform : is an isometry, W SG → SG′ 16 the conclusion of Minlos’ theorem is stronger than usual: the isometry allows the energy

17 inner product to be extended isometrically to a pairing on G′ instead of just G G′ . HE × S S × S 2 18 Remark 7.20. With the embedding L ( G′ , P), we obtain a maximal abelian algebra HE → S 2 19 of Hermitian multiplication operators L ( ) acting on L ( , P). By contrast, see (ii) of ∞ SG′ SG′ 20 Remark 4.3.

21 Remark 7.21. The reader will note that we have taken pains to keep everything R-valued

22 in this chapter (especially the elements of G and G′ ), primarily to ensure the convergence

i u,ξ S S 23 of e h iW dP(ξ) in (7.18). However, now that we have established the fundamental ′ RS 24 identity u, v = u˜vd˜ P in (7.17) and extended the pairing , to G′ , we are h iE S′ h· ·iW HE × S 25 at liberty to complexifyR our results via the standard decomposition into real and complex

26 parts: u = u1 + iu2 with ui R-valued elements of , etc. HE 80 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2 Remark 7.22. The polynomials are dense in L ( , P). More precisely, if ϕ(t , t ,..., t ) is 1 SG′ 1 2 k an ordinary polynomial in k variables, then 2

ϕ(ξ) := ϕ u1, ξ , u2, ξ ,... uk, ξ (7.21) W W W h i h i h i is a polynomial on ′ and 3 SG

.. olyn := ϕ u1(ξ), u2(ξ),... uk(ξ) , deg(ϕ) n, . u j , ξ G′ (7.22) P { ≤ ∈HE ∈ S } is the collection of polynomialse e of degreee at most n, and olyn n∞=0 is an increasing family 4 {P } 2 whose union is all of G′ . One can see that the monomials u, ξ are in L ( G′ , P) as 5 W S h i S2n+1 follows: compare like powers of u from either side of (7.19) to see that Eξ u, ξ = 0 6 h iW and 7

(2n)! E 2n = 2n P = 2n ξ u, ξ u, ξ d (ξ) n u , (7.23) h iW Z |h iW| 2 n! k kE SG′ and then apply the Schwarz inequality. 8 2 To see why the polynomials oly should be dense in L ( , P) observe that the 9 {P n}n∞=0 SG′ 10 sequence P olyn n∞=0 of orthogonal projections increases to the identity, and therefore, { P } 2 P oly u˜ forms a martingale, for any u (i.e., for anyu ˜ L ( G′ , P)). 11 { P n } ∈HE ∈ S If we denote the “multiple Wiener integral of degree n” by 12

n .. Hn := olyn olyn 1 = cl span u, . u , n 1, P −P − {h ·iW ∈HE} ≥ and H := C1 for a vector 1 with 1 = 1. Then we have an orthogonal decomposition of 13 0 k k2 the Hilbert space 14

2 ∞ L ( ′ , P) = H . (7.24) SG n Mn=0

See [Hid80, Thm. 4.1] for a more extensive discussion. 15 2 A physicist would call (7.34) the Fock space representation of L ( , P) with “vacuum 16 SG′ vector” 1; note that Hn has a natural (symmetric) tensor product structure. Familiarity 17 with these ideas is not necessary for the sequel, but the decomposition (7.34) is helpful for 18 understanding two key things: 19 2 (i) The Wiener isometry : L ( G′ , P) identiﬁes with the subspace H1 of 20 2 W 2 HE → S HE L ( G′ , P), in particular, L ( G′ , P) is not isomorphic to . In fact, it is the second 21 S S HE quantization of . 22 HE 2 (ii) The constant function 1 is an element of L ( , P) but does not correspond to any 23 SG′ element of . In particular, constant functions in are equivalent to 0, but this is 24 H2 E HE not true in L ( ′ , P). 25 SG It is somewhat ironic that we began this story by removing the constants (via the intro- 26 duction of ), only to reintroduce them with a certain amount of eﬀort, much later. Item 27 E (ii) explains why it is not nonsense to write things like P( G′ ) = 1 dP = 1, and will be 28 S SG′ helpful when discussing boundary elements in 7.3. R 29 § OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 81

1 F i vx,ξ 2 R (o,x) 1 Corollary 7.23. For ex(ξ) := e h iW , one has Eξ(ex) = e− and hence

1 RF (x,y) Eξ(exey) = ex(ξ)ey(ξ) dP = e− 2 . (7.25) Z SG′

2 Proof. Substitute u = v or u = v v in (7.18) and apply Theorem 5.12. x x − y 3 Remark 7.24. Remark 5.37 discusses the interpretation of the free resistance as the recip- 4 rocal of an integral over a path space; Corollary 7.23 provides a variation on this theme:

F R (x, y) = 2log Eξ(exey) = 2log ex(ξ)ey(ξ) dP. (7.26) − − Z SG′ 5 Observe that Theorem 7.19 was carried out for the free resistance, but all the argu- W 6 ments go through equally well for the wired resistance; note that R is similarly negative 7 semideﬁnite by Theorem 7.18 and Corollary 6.2. Thus, there is a corresponding Wiener 2 8 transform : in L ( , P) deﬁned by W F → SG′

: v f˜, f = P inv and f˜(ξ) = f, ξ . (7.27) W 7→ F h iW 9 Again, f˜x x G0 is a system of Gaussian random variables which gives the wired resistance { } ∈ W 2 10 distance by R (x, y) = Eξ(( f˜ f˜ ) ). x − y 7.2.1. Operator-theoretic interpretation of bdG. Recall that we began this section with a comparison of the Poisson boundary representation

u(x) = u(y)k(x, dy), u bounded and harmonic on Ω Rd, (7.28) Z∂Ω ⊆

11 to the boundary representation E

∂hx u(x) = u + u(o), u arm, and hx = P armvx. (7.29) ∂Ò ∈H H Xbd G

12 Remark 7.25. For u arm and ξ , let us abuse notation and write u foru ˜. That is, ∈ H ∈ SG′ 13 u(ξ) := u˜(ξ) = u, ξ . Unnecessary tildes obscure the presentation and the similarities to h iW 14 the Poisson kernel.

Q 15 Deﬁnition 7.26. Deﬁne P to be the image measure on G′ / in induced by the standard Q 1 S F 16 projection π : / in, i.e., P (B) := P(π B), for B ( / in). SG′ → SG′ F − ∈B SG′ F Q 17 Now P is a probability measure on the quotient and Theorem 7.19 gives a correspond- 18 ing energy integral representation.

19 Corollary 7.27 (Boundary integral representation for harmonic functions).

20 For any u arm and with hx = P armvx, ∈H H

Q u(x) = u(ξ)hx(ξ) dP (ξ) + u(o). (7.30) Z / in SG′ F

21 Proof. Starting with Lemma 4.26, compute

Q u(x) u(o) = hx, u = u, hx = uhx dP , (7.31) − h iE h iE Z SG′ 82 PALLEE.T.JORGENSENANDERINP.J.PEARSE where the last equality comes by substituting v = hx in (7.17); recall from Lemma 4.29 that 1 hx = hx. Note that we are suppressing tildes as in Remark 7.25. Recall from the Gel’fand 2 triple construction that in G′ , but note that hx( f ) = hx, f = 0 for every 3 F ⊆ HE ⊆ S h iE f in, so the domain of integration passes to the quotient / in. 4 ∈F SG′ F

Remark 7.28 (Operator-theoretic interpretation of bd G). In view of Corollary 7.27, we are 5 now able to “catch” the boundary between G and G′ by using ∆ and its adjoint. The 6 S S V boundary of G may be thought of as (a possibly proper subset of) 7

bd G = ′ / in. (7.32) SG F Q Corollary 7.27 suggests that k(x, dξ) := hx(ξ)dP is the discrete analogue in of the 8 HE Poisson kernel k(x, dy), and comparison of (7.2) with (7.30) gives a way of understanding 9 a boundary integral as a limit of Riemann sums: 10

Q ∂hx uhx dP = lim u(x) (x). (7.33) k ∂Ò Z ′ SG →∞ bdXGk Q (We continue to omit the tildes as in Remark 7.25.) By a theorem of Nelson, P is fully 11 1 supported on those functions which are H¨older-continuous with exponent α = 2 , which 12 1 1 we denote by Lip( ) G′ ; see [Nel64]. Recall from Corollary 5.15 that Lip( ). 13 2 ⊆ S HE ⊆ 2 See [Arv76a, Arv76c, Min63, Nel69]. 14

7.3. The boundary as equivalence classes of paths. We are finally able to givea concrete 15 Q representation of elements of the boundary. We continue to use the measure P induced 16 on the quotient G′ / in as in Definition 7.26. Recall the Fock space representation of 17 2 S F L ( , P) discussed in Remark 7.22; if one were to reprove Theorem 7.19 for defined 18 SG′ W only on arm, then(7.19) implies that we have a correspondingFock space representation 19 H for the quotient: 20

2 G′ Q ∞ n L S , P arm⊗ . (7.34) in ! H F Mn=0 0 where arm⊗ := C1 for a unit “vacuum” vector 1 corresponding to the constant function, 21 H n and arm⊗ denotes the n-fold symmetric tensor product of with itself. Observe that 22 H HE 2 Q 1 is orthogonal to in and arm, but is not the zero element of L ( in, P ). 23 F H SG′ F Q P = 24 Lemma 7.29. For all v , / in vd 0. ∈HE G′ RS F Q Q P = P 25 Proof. The integral / in vd / in 1vd is the inner product of two elements in SG′ F SG′ F 2 G′ Q R R S 26 L ( in , P ) which lie in diﬀerent (orthogonal) subspaces; see (7.34). F Alternatively, Lemma 7.29 holds because the expectation of every odd-power monomial 27 vanishes by (7.19); see also (7.23) and the surrounding discussion of Remark 7.22. 28 Recall that by abuse of notation, we use hx to denote the extension h˜ x = hx, of hx 29 h ·iW to . Denote the measure appearing in Corollary 7.27 by 30 SG′

Q dµx := (1 + hx) dP . (7.35)

The function 1 does not show up in (7.30) because it is orthogonal to arm: 31 H OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 83

Q Q u(1 + hx) dP = udP + u, hx = u, hx , for u arm, Z / in Z / in h iE h iE ∈H SG′ F SG′ F 1 where we used Lemma 7.29. Nonetheless, its presence is necessary, for reasons that can 2 SG′ Q 2 best be explained in terms of the Fock space representation of L ( in , P ) as the second F 3 quantization of : the constant function 1 correspondsto the vaccuum vector; see [Hid80, HE 4 4] for more details. It is critical to observe that 1 = 0 in , but 1 , 0 on G′ . § 0 HE S 5 For every x G , ∈

Q Q Q 1dµx = 1(1 + hx) dP = 1 dP + 1hx dP = 1, Z / in Z / in Z / in Z / in SG′ F SG′ F SG′ F SG′ F 6 again by Lemma 7.29. We have shown that as a linear functional, µx[1] = 1. It follows by Q 7 standard functional analysis that hx 0 P -a.e. on G′ / in. In conclusion, µx is absolutely Q ≥ Q S F dµx 8 continuous with respect to P (µ P ) with Radon-Nikodym derivative = 1 + h . x ≪ dPQ x 9 Deﬁnition 7.30. Recall that a path in G is a sequence of successively adjacent vertices; 10 let γ1 = (x0, x1, x2,... ) and γ2 = (y0, y1, y2,... ) be two such paths. Deﬁne an equivalence 11 relation by

γ1 γ2 lim (h(xn) h(yn)) = 0, for every h arm. (7.36) n ≃ ⇐⇒ →∞ − ∈H

12 If β = [γ] is any such equivalence class, pick any representative γ = (x0, x1, x2,... ) and 13 consider the associated sequence of measures µ . As these are all probability measures, { xn } 14 they lie in the unit ball in the weak-⋆ topology. Alaoglu’s theorem then gives a weak-⋆ 15 limit

G′ νγ := lim µx Prob S . (7.37) n n in →∞ ∈ F 16 For any h arm, this measure satisﬁes ∈H

n ˜ →∞ ˜ h(xn) = hdµxn hdνγ. (7.38) Z / in −−−−−−→ Zbd G SG′ F 17 Thus, we deﬁne bd G to be the collection of all such β, and extend harmonic functions to 18 bdG via

h˜(β) := hd˜ νγ. (7.39) Zbd G

19 Lemma 7.31. An element β bd G deﬁnes a continuous linear functional on via ∈ HE

β(v) := lim vd˜ µx , v n Z n ∈HE →∞ SG′ 20 where γ = (x0, x1, x2,... ) is any representative of β.

. 21 Proof. We only need to check that sup β(v) .. v = 1 is bounded. However, this is { k kE } 22 immediate: 84 PALLEE.T.JORGENSENANDERINP.J.PEARSE

✟ Q ✟ Q Q v˜(1 + hx ) dP = v˜✟1 dP + vh˜ x dP = v, hx = h(xn) h(o) , n ✟ n n E Z ′ ✟Z ′ Z ′ |h i | | − | SG SG SG

2 Q where h = P armv. Note that in L ( G′ , P ), 1 is orthogonal to ; see (7.34). 1 H S HE

7.4. The structure of . The next results are structure theorems akin to those found 2 SG′ in the classical theory of distributions; see [Str03, 6.3] or [AG92, 3.5]. If G, 3 p § § HE ⊆ S = ∆ 4 then Theorem 7.32 would say G′ p ∗ ( ) (of course, this is typically false when S V HE arm , 0). 5 H S Theorem 7.32. The distribution space is 6 SG′

p . + = = ∆ .. Z G′ ξ(u) ∗ u, v u G, v , p . (7.40) S { h V iE ∈ S ∈HE ∈ }

p = ∆ 7 Proof. It is clear from the Schwarz inequality that ξ(u) ∗ u, v deﬁnes a continuous h V iE linear functional on G, for any v and nonnegative integer p. For the other direction, 8 S ∈HE we use the same technique as in Lemma 7.13. Observe that if ξ G′ , then there exists K, p 9 p ∈ S p ∆ ∆ 10 such that ξ(u) K ∗ u for every u G. This implies that the map ξ : ∗ u ξ(u) is | |≤ k V kE ∈p S. + V 7→ = ∆ .. Z 11 continuous on the subspace Y span ∗ u u , p . This can be extended to all { V ∈ HE ∈ } of by precomposing with the orthogonal projection to Y. Now Riesz’s lemma gives a 12 HE p = ∆ 13 v for which ξ(u) ∗ u, v . ∈HE h V iE p p ∆ ∆ = 14 Note that v may not lie in the domain of ∗ . If it did, one would have ∗ u, v p ∈HEp V h V iE ∆ = ∆ = = 15 u, ∗ v u, ∗ f , where f P inv. The theorem could then be written G′ h V piW h V iW F S ∆ 16 ∞p=0 ∗ ( in). However, this turns out to have contradictory implications. V F We now provide two results enabling one to recognize certain elements of . 17 S SG′ Lemma 7.33. A linear functional f : C is an element of if and only if there 18 SG → SG′ exists p N and F0, F1,... Fp such that 19 ∈ ∈HE p k f (u) = Fk, ∆ u , u . (7.41) h ∗ V iE ∀ ∈HE Xk=0

Proof. By deﬁnition, f iﬀ p, C < for which f (u) C u for every u . 20 ∈ SG′ ∃ ∞ | | ≤ k kp ∈ SG Therefore, the linear functional 21

p Φ : dom(∆k ) C by Φ(u, ∆ u, ∆2 u,... ∆p u) = f (u) k=0 ∗ ∗ V ∗ ∗ M V → V V p p is continuous and Riesz’s Lemma gives F = (Fk) with 22 k=0 ∈ k=0HE L p = ∆ ∆p = ∆k f (u) F, (u, ∗ u,... ∗ u) Fk, ∗ u . h V V i HE h V i HE L Xk=0 L

Corollary 7.34. If ∆ : is bounded, then G′ = . 23 ∗ V HE →HE S HE OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 85

Proof. We always have the inclusion ֒ G′ by taking p = 0. If ∆ is bounded, then 1 HE → S ∗ V 2 ∆ the adjoint ∗ ∗ is also bounded, and (7.41) gives V p k f (u) = (∆∗ ) Fk, u , u G. (7.42) * ∗ V + ∀ ∈ S Xk=0 HE L p k 3 = ∆ Since G is dense in by Lemma 7.12, we have f k=0( ∗ ∗ ) Fk . S HE V ∈HE P 86 PALLEE.T.JORGENSENANDERINP.J.PEARSE

8. The Laplacian on 1 HE We study the operator theory of the Laplacian in some detail, examining the various 2 domains and self-adjoint extensions. One of the primary goals in 8.1 is to determine 3 § when vx lies in the domain or range of ∆ ; this may indicate when vx lies in the Schwartz 4 V space developed in 7.2. We also identify a particular self-adjoint extension ∆ for use 5 SG § H in the constructions in 7. Also, we give technical conditions which must be considered 6 § when the graph contains vertices of infinite degree and/or the conductance functions c(x) 7 0 is unbounded on G . 8 In 8.2.1, we relate the boundary term of (1.9) to a boundary form akin to that of 9 § classical functional analysis; see Definition 8.17. In Theorem 8.19, we show that if ∆ 10 fails to be essentially self-adjoint, then arm , 0 . In general, the converse does not 11 H { } hold: any homogeneous tree of degree 3 or higher with constant conductances provides a 12 counterexample; cf. Corollary 9.28. 13 In 8.3 we consider the systems v and δ and a kind of spectral reciprocity between 14 § { x} { x} them, in terms of frame duality. In previous parts of this paper, we approximated infinite 15 networks by truncating the domain; this is the idea behind the definition of in and the use 16 F of exhaustions. This approach corresponds to a restriction to span δx x F , where F is some 17 0 { } ∈ finite subset of G . In 8.3, we consider truncations in the dual variable, i.e., restrictions to 18 § sets of the form span vx x F . Note that an element of this set generally will not have finite 19 { } ∈ support. 20 We use ran T to denote the range of the operator T, and ker T to denote its kernel 21 (nullspace). We continue to use the notation from 7: let V := span vx x G0 denote the 22 § { } ∈ vector space of finite linear combinations of dipoles. Then let ∆ be the closure of the 23 V Laplacian when taken to have the dense domain V. 24

8.1. Properties of ∆ on . 25 HE Definition 8.1. The network (G, c) satisfies the Powers bound iff c := sup c(x) < . 26 k k x G0 ∞ ∈ The Powers bound is used more in 9 (see Definition 9.8 and the surrounding discus- 27 § sion); we include it here for use in a couple of technical lemmas. 28

0 Lemma 8.2. If the Powers bound is satisfied, then ∆ maps into ℓ∞(G ). 29 HE 1/2 Proof. By Lemma 4.18 and (2.11), ∆v(x) = δx, v δx v = c(x) v . 30 | | |h iE| ≤ k kE · k kE k kE 0 Lemma 8.3. If deg(x) < for every x G , or if c < , then ran ∆ dom ∆ . 31 ∞ ∈ k k ∞ V ⊆ V 0 Proof. It suffices to show that ∆ vx = δx δo dom ∆ for every x G , and this will 32 V − ∈ V ∈ be clear if we show δx dom ∆ . By Lemma 4.28, δx = c(x)vx y x cxyvy. If deg(x) is 33 ∈ V − ∼ always finite, then we are done. If not, we need to see why y x cPxyvy dom ∆ for any 34 0 ∼ ∈ V fixed x G . P 35 ∈ 0 Fix x G and denote ϕ := y x cxyvy and ϕk := y G cxyvy. Itis clearthat ϕ ϕk 36 ∈ ∼ ∈ k k − kE → 0. We next show ∆ ϕk y Px cxy(δy δo) 0,P from which it follows that ∆ ϕk is 37 V − ∼ − E → { V } Cauchy, and that ϕ dom ∆P with ∆ ϕ = y x(δy δo): 38 ∈ V V ∼ − P 2 2

∆ ϕk cxy(δy δo) = cxy(δy δo) V − − − Xy x yXG∁ ∼ E ∈ k E

OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 87

 cxy δy δo  ≤  k − kE yXG∁   k   ∈    2 c(x) cxy δy δo ≤ k − kE yXG∁ ∈ k

2 2 = c(x)  cxy δy 2 cxy δy,δo + cxy δo  k kE − h iE k kE  ∁ ∁ ∁  yXG yXG yXG   ∈ k ∈ k ∈ k   

= c(x)  cxyc(y) + 2 cxycoy + c(o) cxy  ∁ ∁ ∁  yXG yXG yXG   ∈ k ∈ k ∈ k    c (3 c + c(o)) c ,  ≤ k k k k xy yXG∁ ∈ k ∁ 1 which tends to 0 as k gets large. Note that c < 1 for y G with k sufficiently large. oy ∈ k 2 8.1.1. Finitely supported functions and the range of ∆. In Lemma 4.50 we showed that 3 one always has ran ∆ in and hence arm ker ∆∗ . The rest of this subsection is V ⊆ F H ⊆ V 4 roughly an examination of the reverse containment, i.e., what conditions give ran ∆ = V 5 in. Determining when ran ∆ = in essentially boils down to the following technical F V F 6 question: when is span δ δ dense in in? It is curious that this never happens on { x − o} F 7 a finite network (Lemma 8.22), but is often true on an infinite network. However, see 8 Example 14.35.

9 Definition 8.4. Let in be the -closure of span δ δ and let in be the orthogonal F 2 E { x − o} F 1 10 complement of in2 in in. This extends the decomposition = in arm, in some F F HE F ⊕H 11 cases, to = in2 in1 arm. HE F ⊕F ⊕H 12 Example 14.35 shows a situation in which in is not dense in in. F 2 F 13 Lemma 8.5. Let (G, c) be an infinite network. If c < , then in = in . k k ∞ F F 2 14 Proof. It suffices to approximate the single Dirac mass δo by linear combinations of differ- (n) n 15 ences. For each n, fix n vertices xk k=1, no two of which are adjacent. Therefore, define 1 n { } 16 ϕn := n k=1(δo δx(n) ) and compute − k P n 2 n 2 2 1 1 1 (n) c δ ϕ = δ (n) = δ (n) sup c(x ) 0, o n x 2 x k k k k − kE n k n k ≤ n 1 k n ≤ n → Xk=1 Xk=1 E ≤ ≤ E 17 where the second equality comes by orthogonality; for j , k, δ (n) and δ (n) are not adjacent, xk x j 18 hence δx(n) ,δx(n) = 0by(2.11). Now it is trivial to approximate δz = (δz δo) + δo. h k j iE −

19 The idea of Lemma 8.5 is illustrated on the binary tree in Example 13.8.

20 8.1.2. Harmonic functions and the domain of ∆. Curiously, even though ∆h(x) = 0 point- 0 21 wise for every x G , it may happen that h is not in the domain of ∆. Example 13.8 ∈ 22 discusses a nontrivial harmonic function on the binary tree which does not appear to be in 23 the domain of ∆ . However, harmonic functions are always in the domain of the adjoint V 24 ∆∗ by Lemma 4.50. V 88 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Lemma 8.6. If ∆˜ is any Hermitian extension of ∆ whose domain contains arm, then 1 V V H ∆˜ h = 0 for any h arm. Moreover, ∆˜ u in for any u dom ∆˜ . 2 V ∈H V ∈F ∈ V Proof. Recall that we have the following ordering of operators: ∆ ∆˜ ∆∗ . Since 3 V ⊆ V ⊆ V ∆∗ is an extension of ∆˜ and arm ∆˜ , the ﬁrst claim follows immediately from 4 V V H ⊆ V Lemma 4.50. The secondclaim nowfollowsfromthe ﬁrst because ∆˜ v, h = v, (∆˜ )∗h = 5 h V iE h V iE 0 for every h arm, since ∆˜ (∆˜ )∗. 6 ∈H V ⊆ V We have a partial converse of Lemma 4.50. Note that if span δ δ is dense in in 7 { x − o} F (as discussed in Lemma 8.5), then Lemma 8.7 implies ker ∆∗ = arm. 8 V H Lemma 8.7. ker ∆∗ is the orthogonal complement of span δx δo . 9 V { − } Proof. Suppose u ker ∆∗ so that ∆∗ u = 0. Then 10 ∈ V V

0 = ∆∗ u, vx = u, ∆ vx = u,δx δo . h V iE h V iE h − iE This shows u is orthogonal to span δ δ . 11 { x − o} The Lemma 8.5 gives an idea of when the hypotheses of Lemma 8.7 are satisfied. In 12 fact, a weaker hypothesis will suffice: one just needs to be able to find an infinite subset of 13 nonadjacent vertices on which c(x) is bounded. 14

Definition 8.8. Define ∆H to be the extension of ∆ to the domain dom ∆ + arm by 15 V V H ∆H(v + h) := ∆ v. By abuse of notation, let ∆H denote the closure of ∆H with respect to the 16 V graph norm; see Definition B.12. 17

Lemma 8.9. ∆H is well deﬁned, Hermitian, and semibounded. 18

Proof. We must check that ∆ (0) = 0, so suppose v+h = 0 for v V and h arm. Then 19 H ∈ ∈H Lemma 4.50 gives ∆∗ (v + h) = 0, whence ∆ v = ∆∗ h = 0. 20 V V − V

Theorem 8.10. ∆H is self-adjoint. 21

Proof. Let w satisfy ∆H∗ w = w. To see that w = 0, note that w dom ∆H∗ , so 22 ∈ HE − ∈ ∆∗ w in by Lemma 8.11, just below. But then w = ∆H∗ w = ∆∗ w in, so 23 V ∈F − − V ∈F

w 2 = w, w = w∆w = w 2 0, k kE h iE − | | ≤ XG0 XG0 so that w = 0 in . This shows ∆H is essentially self-adjoint, but ∆H is closed by deﬁni- 24 HE tion, so it is self-adjoint. 25

.. Lemma 8.11. dom ∆∗H = w dom ∆∗ . ∆∗ w in . 26 { ∈ V V ∈F } Proof. For purposes of this proof, it is permissible to work with as a real vector space 27 HE and complexify afterwards. 28 ( ) Suppose that w dom ∆ , i.e., we have the estimate 29 ⊆ ∈ H∗

w, ∆H(v + h) C1 v + h , for all v V and h arm. (8.1) |h iE|≤ k kE ∈ ∈H Then for all t R, 30 ∈

2 2 2 2 2 2 2 2 w, ∆Hv C1 v + th C1 v + 2t v, h + t h , for all v V and h arm. |h iE| ≤ k kE ≤ k kE |h iE| k kE ∈ ∈H OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 89

1 This quadratic polynomial in t is nonnegative, and hence its discriminant must be nonpos- 2 itive, so that

4 2 2 2 2 2 2 C1 v, h C1 h C1 v w, ∆Hv |h iE| ≤ k kE k kE − |h iE| 2 2 v, h 2 2 = E , ∆ Phv |h i2 | v C2 w Hv k kE h ≤ k kE − |h iE| k kE 1 3 where Ph is projection to the rank-1 subspace spanned by h and C2 = . If we let hi be C1 { } 4 an ONB for arm, then H

2 2 2 2 C2 w, ∆Hv v Ph1 v , for all v V. |h iE| ≤ k kE − k kE ∈ 5 Inductively substituting v = v h , v = v (h + h ), etc, we have − 2 − 2 3

2 2 2 2 2 2 2 C2 w, ∆H(v h2) v Ph2 v Ph1 v = v Ph2 v + Ph1 v |h − iE| ≤ k − kE − k kE k kE − k kE k kE . . 2 2 2 2 2 2 C2 w, ∆H(v + ihi) v i Phi v = v P armv . |h iE| ≤ k kE − k kE k kE − k H kE P P 2 2 2 2 6 By the deﬁnition of ∆H, all the left sides are equal to C2 w, ∆Hv = C2 w, ∆ v . 2 2 |h iE| |h V iE| 7 Since P inv = v P armv , we have established k F kE k kE − k H kE

w, ∆ v C3 P inv , for all v V. |h V iE|≤ k F kE ∈ 8 Now Riesz’s lemma gives an f in such that ∈F

w, ∆ v = f, P inv , for all v V. h V iE h F iE ∈ 9 However, orthogonality allows one to remove the projection (since the ﬁrst argument is 10 already in in), whence ∆∗ w, v = f, v for all v V, and so ∆∗ w = f in. F h V iE h iE ∈ V ∈F 11 ( ) Let w be in the set on the right-hand side. To see w dom ∆ , we need the estimate ⊇ ∈ H∗ 12 (8.1), but

w, ∆H(v + h) = w, ∆ v = ∆∗ w, v = ∆∗ w, P inv , |h iE| |h V iE| h V iE h V F iE

13 where the last equality follows by the hypothesis ∆∗ w in. This gives w, ∆H(v + h) V ∈F |h iE|≤ 14 ∆∗ w P in(v + h) , but P inv = P in(v + h) v + h ,so(8.1) follows. k V kE · k F kE k F kE k F kE ≤ k kE 15 Corollary 8.12. A closed extension of ∆ is self-adjoint if and only if arm is contained V H 16 in its domain.

17 Proof. It is helpful to keep in mind the operator ordering ∆ ∆H = ∆H∗ ∆∗ . V ⊆ ⊆ V 18 ( ) Let ∆˜ be a self-adjoint extension of ∆ . If ∆H ∆˜ , then the result is obvious, and ⇒ V ⊆ 19 if ∆˜ ∆ , then again ∆ ∆ (∆˜ ) = ∆˜ , and the result is equally obvious. ⊆ H H ⊆ H∗ ⊆ ∗ ( ) If ∆˜ is a closed extension of ∆ with arm dom ∆˜ , then ∆H ∆˜ , so ⇐ V H ⊆ ⊆ clo clo clo ∆ ∆˜ (∆˜ )∗ (∆ )∗ ∆ , H ⊆ ⊆ ⊆ H ⊆ H 20 where the ﬁrst inclusion holds because ∆˜ is closed, and the last by Theorem 8.10. 90 PALLEE.T.JORGENSENANDERINP.J.PEARSE

8.2. The defect space of ∆ . Let ∆ once again denote the graph closure of the operator 1 V V ∆ on the (dense) domain V := span v as in Definition 4.32. 2 { x} Definition 8.13. Since ∆ is Hermitian on its domain by Corollary 4.61, Definition B.17 3 V and Theorem B.18 imply that the defect space of ∆ is 4 V

. Def := v dom ∆∗ .. ∆∗ v = v . (8.2) { ∈ V V − } Observe also that Def⊥ = ran(I + ∆ ). 5 V Lemma 8.14. u is a defect vector of ∆ if and only if there is a constant k such that 6 0 V ∆u(x) = u(x) + k ateach x G . 7 − ∈ Proof. Recall that the meaning of such a pointwise identity is that u dom ∆∗ and ∆∗ u = u + k in ; see Lemma 4.33. The reverse implication is obvious;∈ for theV obverseV it su− ﬃces to checkHE the claim against the (dense) energy kernel:

0 = vx, ∆∗u + u = δx δo, u + vx, u = ∆u(x) ∆u(o) + u(x) u(o), h iE h − iE h iE − − by Lemma 4.18, which proves the claim with k = ∆u(o) + u(o). 8

Remark 8.15 (Defect vectors and the Gauss-Green formula). We have introduced the defect 9 space of ∆ here to alleviate any concerns regarding the convergence of x G0 u(x)∆u(x) 10 V ∈ in (4.18); the reader will note that if u is a defect vector, then P 11

u(x)∆u(x) = u(x) 2, − | | xXG0 xXG0 ∈ ∈ 2 which must equal , since there are no defect vectors in ℓ . This is a reasonable concern, 12 −∞ as there do exist networks with nontrivial defect; see Example 14.36. However, such defect 13 vectors are proscribed by the hypotheses of Theorem 4.41, by the following lemma. 14

Lemma 8.16. dom ∆ Def = 0. 15 V ∩ Proof. Suppose u dom ∆ Def. Note that ∆∗ is an extension of ∆ , so such a u 16 ∈ V ∩ V V satisﬁes ∆ u = u. However, since ∆ is semibounded on its domain by Corollary 4.61, 17 V − V this implies 18

2 0 u, ∆ u = ∆∗ u, u = u, u = u , ≤ h V iE h V iE −h iE −k kE whence u = 0. 19

8.2.1. The boundary form. In this section, we relate the defect of ∆ to the boundary term 20 of the Discrete Gauss-Green formula (Theorem 4.41), thereby extending Theorem 4.47. 21 The reader may find [DS88, XII.4.4] to be a useful reference. 22 § Definition 8.17. Define the boundary form 23

1 βbd(u, v) := ∆∗ u, v u, ∆∗ v , u, v dom(∆∗ ). (8.3) 2i h V iE − h V iE ∈ V To see the signiﬁcance of βbd for the defect spaces, note that if ∆∗ f = z f where z C with 24 2 V ∈ Im z , 0, then βbd( f, f ) = (Im z) f . 25 k kE Lemma 8.18. The boundary form βbd(u, v) vanishes if u or v lies in dom(∆ ). 26 V OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 91

1 Proof. For v dom(∆ ), ∆∗ u, v = u, ∆ v by the deﬁnition of the adjoint, and ∈ V h V iE h V iE 2 u, ∆ v = u, ∆∗ v by the fact that ∆∗ extends ∆ . Hence, both terms of (8.3) are h V iE h V iE V V 3 equal for u, v dom(∆ ). The proof is identical if u dom(∆ ). ∈ V ∈ V 4 The following result extends Theorem 4.47.

5 Theorem 8.19. If ∆ fails to be essentially self-adjoint, then arm , 0 . V H { } 6 Proof. We prove that the boundary form β (u, v) vanishes identically whenever arm = bd H 7 0 . Since the boundary sum can only be nonzero when arm , 0 , the conclusion will { } H { } 8 follow once we show that

1 ∂u ∂v

βbd(u, v) = (∆∗ v) (∆∗ u) . (8.4)

Ò Ò 2i ∂ ∂ V − Xbd G V 9 To see this, apply Theorem 4.41 to obtain

∂v ∂v

∆∗ u, v = ∆∗ u∆ v + ∆∗ u = ∆ u∆ v + ∆ u Ò ∂Ò ∂ h V iE V V V V V V XG0 Xbd G XG0 Xbd G

10 for any u, v dom(∆∗ ). The second equality follows because ∆∗ = ∆ pointwise: ∈ V

∆∗ u(x) ∆∗ u(o) = vx, ∆∗ u = ∆ vx, u = δx δo, u = ∆u(x) ∆u(o), V − V h V iE h V iE h − iE − 11 where the last equality comes by Lemma 4.18. Also, note that u dom(∆∗ ) implies ∈ V 12 ∆∗ u , so that Theorem 4.41 applies and both terms are ﬁnite. Consequently, the two V ∈HE 0 13 sums over G cancel and the theorem follows.

14 Remark 8.20. There is an alternative, more elementary way to prove Theorem 8.19. Sup-

15 pose w , 0 is a nonzero defect vector with ∆∗ w = iw. Then we can ﬁnd a representative V 16 for w such that

∂w 2 ∂w ∂w i

w, w = w∆w + w = i w + Re w + Im w . (8.5)

Ò Ò h iE ∂Ò | | ∂ ∂ XG0 Xbd G XG0 Xbd G Xbd G 2 17 Since w = w, w is real (and strictly positive, by hypothesis), this implies the boundary k kE h iE 18 sum is nonzero and Theorem 4.47 gives the existence of nontrivial harmonic functions. 19 It also follows from (8.5) that such a nonzero defect vector satisﬁes

w 2 = Im w ∂w > 0, | | − ∂Ò XG0 Xbd G ∂w 20 so that Im w < 0.

bd G ∂Ò P 21 8.3. Dual frames and the energy kernel. In previous parts of this paper, we have approx- 22 imated infinite networks by truncating the domain; this is the idea behind the definition of 23 in in Definition 4.16, and in the use of exhaustionsfor various arguments(Definition 4.5). F 24 This approach corresponds to a restriction to span δx x F , where F is some finite subset of 0 { } ∈ 25 G . In this section, we consider truncations in the dual variable, i.e., restrictions to sets 26 of the form span vx x F . This is directly analogous to the usual time/frequency duality in { } ∈ 27 Fourier theory. 28 The energy kernel vx generally fails to be a frame for , as shown by Lemma 8.25 { } HE 29 and the ensuing remarks. However, things improve when restricting to a finite subset. We 92 PALLEE.T.JORGENSENANDERINP.J.PEARSE shall approach the infinite case via a compatible system of finite dual frames, one for each 1 0 finite subset F G o ; see Definition 8.24. In Theorem 8.29, we show that δx x F and 2 ⊆ \ { } { } ∈ vx x F form a dual frame system. 3 { } ∈ We obtain optimal frame bounds in Corollary 8.30. In Theorem 8.33, we show that the 4 boundedness of ∆ is equivalent to both the existence of a global upper frame bound (i.e., 5 V one can let F G), and the existence of a spectral gap. 6 → We begin with two lemmas whose parallels serve to underscore the theme of this section. 7

Lemma 8.21. The vectors v are linearly independent. 8 { x} Proof. Suppose that we have a (ﬁnite) linear combination ψ = , ξ v , where ξ C. 9 x o x x x ∈ Then for y , o, P 10

δy, ψ = ξx δy, vx = ξx(δy(x) δy(o)) = ξxδy(x) = ξy. h iE h iE − Xx,o Xx,o Xx,o If ψ = 0, then this calculation shows ξy = 0 for each y. 11

Lemma 8.22. The vectors δ are linearly independent. 12 { x} Proof. Suppose that we have a (ﬁnite) linear combination ψ = , ξ δ , where ξ C. 13 x o x x x ∈ Then for y , o, P 14

vy, ψ = ξx vy,δx = ξx(δx(y) δx(o)) = ξxδx(y) = ξy. h iE h iE − Xx,o Xx,o Xx,o If ψ = 0, then this calculation shows ξy = 0 for each y. 15

0 Definition 8.23. In this section we always let F G o denote a finite subset of vertices 16 . ⊆ \{ } and let V(F) = span v .. x F . Observe that elements of V(F) do not typically have 17 { x ∈ } finite support; cf. Definition 4.16 and Figure 10 of Example 14.2. Let ∆V(F) denote the 18 Laplacian when taken to have the domain V(F), even though it not dense in . 19 HE Definition 8.24. Denote D(F) := δx x F and let ∆ be the Laplacian when taken to have 20 { } ∈ F this (non-dense) domain. 21 Then D(F)isa dual frame for V(F) if there are constants 0 < A B < (called frame 22 ≤ ∞ bounds) for which 23

2 2 2 A ψ δx, ψ B ψ , ψ V(F). (8.6) k k ≤ |h iE| ≤ k k ∀ ∈ E Xx F E ∈ 2 0 Lemma 8.25. vx is a frame for if and only if ℓ (G ) and are isomorphic. 24 { } HE HE Proof. Since v is a reproducing kernel, the frame inequalities take the form 25 { x}

2 2 2 A w 2 w(x) B w 2. (8.7) k k ≤ X | | ≤ k k Each inequality indicates a (not necessarily isometric) embedding. 26

Remark 8.26. The second inequality fails if ∆ does not have a spectral gap. See also 27 Lemma 10.17. 28

Deﬁnition 8.27. Deﬁne a Hermitian F F matrix by MF := [ vx, vy ]x,y F. Let λmin := 29 | |×| | h iE ∈ min spec(MF ) and λmax := max spec(MF ). 30 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 93

Definition 8.28. For ψ , define X : ℓ(G0), where ℓ(G0) is the space of all functions on G0, by ∈ HE HE → ψ X (x) := δx, ψ . (8.8) h iE 1 By Remark 4.19, X is morally identical to the Laplacian when defined on all of ; note ψ HE 2 that X may not lie in . HE 2 3 In the proof of Theorem 8.29, the notations , 1 and 1 refer to the space ℓ (1) h· ·i k · k 2 4 discussed in 9, that is, f, g 1 = x G0 f (x)g(x) is the unweighted ℓ inner product, etc. § h i ∈ P 5 Theorem 8.29. For any finite F, one has λmin > 0 for the minimal eigenvalue of Defini- 6 tion 8.27, and δx x F is a dual frame for V(F) with frame bounds { } ∈

1 2 2 1 2 ψ δx, ψ ψ . (8.9) λ k k ≤ |h iE| ≤ λ k k max E Xx F min E ∈

7 Proof. First, to show that λmin > 0, we show that 0 is not in the spectrum of MF. By way 8 of contradiction, suppose ξ : F C such that ∃ →

MF ξ = vx, vy ξy = 0. h iE Xy

9 The vector ψ = y vx, vy ξy V(F) is nonzero by Lemma 8.21, and yet h iE ∈ P

ψ(x) ψ(o) = vx, ψ = vx, vy ξy = 0. − h iE h iE Xy < 10 Hence, ψ is constant and therefore ψ = 0 in . So 0 is not in the spectrum of MF . HE ւ 11 Then by (8.8),

2 2 ψ = δx, ψ vx = δx, ψ vx, vy δy, ψ k k h iE h iEh iEh iE E Xx F xX,y F ∈ E ∈ ψ ψ = X (x) vx, vy X (y) h iE xX,y F ∈ = Xψ, M Xψ , h F i1 ψ 2 2 ψ 2 ψ 2 12 whence λmin X 1 ψ λmax X 1, and the conclusion (8.9) follows from X 1 = k 2 k ≤ k kE ≤ k k k k 13 x F δx, ψ . ∈ |h iE| P 14 Corollary 8.30. The frame bounds in (8.9) are optimal.

15 Proof. Let ξ spec(MF ) and ξ : F C with MF ξ = λξ. The vector ξ = x F ξxvx is in ∈ → ψ ∈ 16 by the proposition and ξ = δx, ψ = X (x) for each x F by Lemma P8.21 and (8.8). HE h iE ∈ 17 Moreover,

2 2 2 ψ = ξ, MF ξ 2 = λ ξ 2 = δx, ψ . k k h i k k |h iE| E Xx F ∈ 18 We now apply this to λmin and to λmin and deduce the bounds are optimal.

19 In the next lemma, we use ∆ speciﬁcally to indicate that the Laplacian is considered 20 pointwise, and without regard to domains. 94 PALLEE.T.JORGENSENANDERINP.J.PEARSE

0 Lemma 8.31. X represents ∆ on ℓ(G ), i.e., ∆(Xψ) = X(∆ ψ) for all ψ dom ∆ . 1 V V ∈ V 0 Proof. Fix ψ dom ∆ and x G . Then 2 ∈ V ∈

∆(Xψ)(x) = c(x)Xψ(x) c Xψ(y) − xy Xy x ∼ = c(x) δx, ψ cxy δy, ψ h iE − h iE Xy x ∼ = c(x)δ c δ , ψ x − xy y * Xy x + ∼ E = ∆δx, ψ . h iE Now since δx dom ∆∗ , we have ∆(Xψ)(x) = ∆∗ δx, ψ = δx, ∆ ψ = X(∆ ψ)(x). 3 ∈ V h V iE h V iE V ψ 2 ψ 2 Lemma 8.32. For any ψ V(F), we have ψ, ∆V(F)ψ = x F X (x) + x F X (x) . 4 ∈ h iE ∈ | | ∈ P P Proof. Writing ∆ for ∆V(F), this follows from 5

ψ ψ ψ, ∆ψ = X (x)X (y) vx, ∆vy h iE h iE xX,y F ∈ = Xψ(x)Xψ(y)((δ (x) δ (o)) (δ (x) δ (o))) y − y − o − o xX,y F ∈ ψ ψ = X (x)X (y)(δy(x) + 1) (8.10) xX,y F ∈ = Xψ(x)Xψ(x) + Xψ(x) Xψ(y) ,     Xx F Xx F Xy F  ∈  ∈   ∈      where (8.10) follows because o < F.     6

Incidentally, Lemma 8.32 oﬀers a proof of Theorem 4.61. 7

Theorem 8.33. The following are equivalent: 8

(i) ∆ is a bounded operator on . 9 (ii) ThereV is a global upper frameH boundE B < in (8.9), i.e. ∞ 2 2 δx, ψ B ψ , ψ . (8.11) |h iE| ≤ k k ∀ ∈HE Xx,o E

(iii) There is a spectral gap inf spec(MF) > 0, where F runs over the set of all ﬁnite 10 0 F subsets of G o . 11 \ { } Proof. (i) = (ii). If ∆ is bounded, then by Lemma 4.18 followed by Corollary 4.61, 12 ⇒ V

2 2 2 δx, ψ = ∆ψ(x) = ψ, ∆ψ B ψ . |h iE| | | h iE ≤ k k Xx,o Xx,o E

(ii) = (i). First ﬁx ε > 0. Note that x G0 ∆ψ(x) = 0 by Corollary 4.60, so choose F 13 ⇒ ∈ so that x G0 ∆ψ(x) <ε. The hypothesisP of the global upper frame bound B gives 14 ∈ P 2 2 2 ∆ψ(x) = δx, ∆ψ B ψ , | | |h iE| ≤ k k Xx F Xx F E ∈ ∈ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 95

1 so that Lemma 8.32 implies

2 ψ, ∆ψ ∆ψ(x) 2 + ∆ψ(x) < B ψ 2 + ε, |h iE|≤ | | k k Xx F Xx F E ∈ ∈ 2 2 and we get ψ, ∆ψ B ψ as ε 0. |h iE|≤ k kE → 1 3 (i) (iii) Observe that (8.9) and Lemma 8.30 imply that B, and hence λmin(F) ⇐⇒ ≤0 4 λ (F) 1/B, F . If we have an exhaustion F F F = G o , then the min ≥ ∀ ∈F 1 ⊆ 2 ⊆··· k \ { } 5 Minimax Theorem indicates that λ (F + ) λ (F ) so min k 1 ≤ min k S

1 1 .. 2 ∆ − = sup B 0 . ψ, ∆ ψ B ψ , ψ V k Vk { ≥ h V iE ≤ k kE ∀ ∈ } = lim λmin(Fk). k →∞

6 Corollary 8.34. If δ is a dual frame for v , then the upper and lower frame bounds A { x} { x} 7 and B provide bounds on the free resistance metric:

2 2 RF(x, y) . (8.12) B ≤ ≤ A 0 8 Note that as F increases to G , one may have A 0 so that the upper bound tends to . → ∞ 9 Proof. By (5.10), we are motivated to apply the frame inequalities applied to the function 10 vx vy via Theorem 8.33: − ∈HE

2 2 2 A vx vy δz, vx vy B vx vy . k − kE ≤ |h − iE| ≤ k − kE z XG0 o ∈ \{ } 2 11 The result now follows by (5.10) upon observing that z G0 o δz, vx vy = 2. ∈ \{ } |h − iE| 0 P 12 Lemma 8.35. For ﬁnite F G , arm V(F) = ∅. A fortiori, ∆ has a spectral gap. ⊆ H ∩ V n 13 Proof. Let h = i=1 civxi . If h is harmonic, then P

0 = ∆h = ci(δxi δo) = ciδxi δo ci, X − X − X 14 which implies ci = 0 for each i, since the Dirac masses are linearly independent vectors. 15 The second claim follows because 0 is not in the point spectrum of ∆ on the ﬁnite- V 16 dimensional space V.

17 The symmetry of formula (8.13) in x and y provides another proof that ∆ is Hermitian. F 0 18 Lemma 8.36. For all x, y G , ∈

∆δx,δy = (c(x) + c(y))cxy + cxzczy. (8.13) h iE − zXx,y ∼

19 Proof. For ∆ , use z x, y to denote that z is a neighbour of both x and y, and compute F ∼

∆δx,δy = c(x)δx cxzδz,δy h iE − * Xz x + ∼ E 96 PALLEE.T.JORGENSENANDERINP.J.PEARSE

= c(x) δx,δy cxz δz,δy h iE − h iE Xz x ∼

= c(x) δx,δy cxy δy,δy cxz δz,δy h iE − h iE − Xz x h iE z∼,y

= c(x)cxy cxyc(y) + cxzczy − − Xz x z∼,y = (c(x) + c(y))c + c c . − xy xz zy zXx,y ∼ 2 2 Definition 8.37. Let cx be defined by cx(y) = cxy, so cx := cx cx := y x cxy. 1 · ∼ P In the next theorem, one would need to consider ϕ, ∆ϕ for general ϕ span vx , 2 h iE ∈ { } rather than just ϕ = vx,δx, in order to get the full spectrum [spec ∆ ]. 3 V Lemma 8.38. The spectrum of ∆ satisfies 4 V 2 c2 inf spec ∆ min inf , inf c(x) + x , V ≤ ( RF(x, o) c(x)!) 2 c2 sup spec ∆ max sup , sup c(x) + x . V ≥ ( RF(x, o) c(x)!)

Proof. We compute the action of ∆ on certain unit vectors: 5

v v 1 2 x ∆ x = = , 2 (δx(x) δx(o)) (δx(o) δo(o)) F * vx vx + vx − − − R (x, o) k kE k kE E k kE and 6

δ δ 1 c(x)2 + c2 c2 x , ∆ x = δ , ∆δ = x = + x . 2 x x c(x) * δx δx + δx h iE c(x) c(x) k kE k kE E k kE We then apply the well-known theorem that for a closed Hermitian operator S , . [spec S ] = u, Su .. u dom S, u = 1 , {h i ∈ k k } where [set] denotes the closed convex hull of set in C. Note that [spec S ] R. 7 ⊆ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 97

2 1 9. The ℓ theory of ∆ and the transfer operator

2 In this section, we discuss some results for ∆ and T when considered as operators on

2 0 .. 2 ℓ (1) := u : G C . x G0 u(x) < , (9.1) { → ∈ | | ∞} 3 with the inner product P

u, v := u(x)v(x). (9.2) h i1 xXG0 ∈ 4 The constant function 1 appears in the notation to specify the weight involved in the inner 5 product, in contrast to c. This is necessary because we will also be interested in ∆ and T as 6 operators on

2 0 .. 2 ℓ (c) := u : G C . x G0 c(x) u(x) < , (9.3) { → ∈ | | ∞} 7 with the inner product P

u, v := c(x)u(x)v(x). (9.4) h ic xXG0 8 ∈ 2 2 9 While the pointwise definition of ∆ and T remains the same on ℓ (1) and ℓ (c), they 10 are different operators with different domains and different spectra! It is important to keep 2 2 11 in mind that in general, none of , ℓ (1) or ℓ (c) are contained in any of the others. HE 12 However, we provide some conditions under which embeddings exist in 9.2.2. We give § 13 only some selected results, as this subject is well-documented elsewhere in the literature. 2 14 In 9.3, we consider a map J : ℓ (c) is the quotient map induced by the equiv- § → HE 2 15 alence relation discussed in Remark 4.3. It turns out that J is an embedding of ℓ (c) into 2 16 in, and that its range is dense in in. We will also see that P is self-adjoint on ℓ (c), even F F 2 17 though it is not even Hermitian on ℓ (1) or except when c(x) is constant. HE 2 18 9.1. The Laplacian on ℓ (1). In this section, we investigate certain properties of the 2 19 Laplacian on ℓ (1), including self-adjointness and boundedness. Dealing with unbounded 20 operators always requires a bit of care; the reader is invited to consult Appendix B.3 to 21 refresh on some principles of self-adjointness of unbounded operators. Recall that for S to 22 be self-adjoint, it must be Hermitian and satisfy dom S = dom S ∗, where

. dom S ∗ := v .. v, Su K u , u dom S . { ∈H |h i| ≤ vk k ∀ ∈ } 23 In the unbounded case, it is not unusual for dom S ( dom S ∗. Some good references for 24 this section are [Jør78, vN32a, Nel69, RS75, Rud91, DS88]. Due to Corollary 4.65, we can ignore the possibility of nontrivial harmonic functions while working in this context. Combining Theorem 4.41 with Theorem 4.47, one can relate the inner products of and ℓ2(1) by HE u, ∆v 1 = u, v , (9.5) h i h iE 2 25 for all u, v span δ . Observe that span δ is dense in ℓ (1) with respect to (9.2), and ∈ { x} { x} 26 dense in in the norm when arm = 0. Then (9.5) immediately implies that the HE E 2 H 27 Laplacian is Hermitian on ℓ (1) because, again for all u, v span δ , ∈ { x}

u, ∆v 1 = u, v = v, u = v, ∆u 1 = ∆u, v 1, (9.6) h i h iE h iE h i h i 98 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2 This may seem trivial, but it turns out that ∆ is not Hermitian on ℓ (c); cf. Lemma 9.31. 1 Theorem 9.9 shows that if c is uniformly bounded (9.14), then ∆ is a bounded operator 2 and hence self-adjoint. However, in Theorem 9.2 we are able to obtain a much stronger 3 result, without assuming any bounds: the Laplacian on any electrical resistance network 4 2 is essentially self-adjoint on ℓ (1). (Recall that ∆ is essentially self-adjoint iﬀ it has a 5 unique self-adjoint extension; cf. Deﬁnition B.15.) This is a sharp contrast to the case for 6 , as seen from Theorem 8.19. In the latter parts of this section, we also derive several 7 HE applications of Theorem 9.2. 8

9.1.1. The Laplacian as an unbounded operator. We begin with the operator ∆ deﬁned on 9 span δx , the dense domain consisting of functions with ﬁnite support. Then let ∆1 denote 10 { } 2 the closure of ∆ with respect to (9.2), that is, its minimal self-adjoint extension to ℓ (1). 11 Some good references for this section are [vN32a, Rud91, DS88]. 12

2 Lemma 9.1. The Laplacian ∆ is semibounded on dom ∆ . A fortiori,foranyu, v ℓ (1), 13 1 1 ∈

u, ∆ v = c(x)u(x)v(x) c u(x)v(y) (9.7) h 1 i1 − xy xXG0 xX,y G0 ∈ ∈

Proof. For any u, v in, a straightforward computation shows 14 ∈F

u, ∆ u = c(x) u(x) 2 c u(x)u(y), (9.8) h 1 i1 | | − xy xXG0 x,Xy G0 ∈ ∈ whence the equality in (9.7) follows by taking limits and polarizing. To see that ∆1 is 15 semibounded, apply the Schwarz inequality ﬁrst with respect to y, then with respect to x, 16 to compute 17

1/2 2 cxyu(x)u(y) c(x) u(x) cxy u(y) ≤ | |  | |  xXG0 Xy x xXG0 p Xy x  ∈ ∼ ∈  ∼  1/2  1/2   c(x) u(x) 2 c u(y) 2 ≤  | |   xy| |  xXG0  xX,y G0   ∈   ∈      =  c(x) u(x) 2,   | | xXG0 ∈ so that the diﬀerence on the right-hand side of (9.8) is nonnegative. 18

0 2 Theorem 9.2. If deg(x) < for every x G , ∆ is essentially self-adjoint on ℓ (1). 19 ∞ ∈ 1 2 Proof. Lemma 9.1 shows ∆1 is semibounded on ℓ (1), so by Theorem B.18, it suﬃces to 20 show the implication 21

∆∗v = v = v = 0, v dom ∆∗. (9.9) 1 − ⇒ ∈ 1 2 2 Suppose that v ℓ is a solution to ∆ v = v. Then clearly ∆ v ℓ , and then by 22 ∈ 1∗ − 1∗ ∈ Lemma B.23, 23

2 0 v, M∆ v = v, ∆∗v = v, v = v 0 = v = 0, ≤ h 1 i1 h 1 i1 −h i1 −k k1 ≤ ⇒ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 99

0 1 where M∆1 is the matrix of ∆1 in the ONB δx x G . To justify the first inequality, con- { } ∈ 2 sider that we may find a sequence v in with v v 0. Because the matrix { n}⊆F k − nk1 → 3 M∆ is banded, this is sufficient to ensure that M∆ v M∆ v and hence (v , M∆ v ) con- 1 1 n → 1 n 1 n 4 verges to (v, M∆ v) in the graph norm, and so v , M∆ v converges to v, M∆ v . Then 1 h n 1 ni1 h 1 i1 5 v , M∆ v = (v ) 0 for each n, and positivity is maintained in the limit (even though h n 1 ni1 E n ≥ 6 lim (v ) may not be finite). E n 7 See [Web08] for a similar result. It follows from Theorem 9.2 that the closure of the 2 8 operator ∆1 is self-adjoint on ℓ (1), and hence has a unique spectral resolution, determined 9 by a projection valued measure on the Borel subsets of the infinite half-line R+. This is in 10 sharp contrast with the continuous case; in Example B.21 we illustrate this by indicating d2 2 11 how ∆ = fails to be an essentially self-adjoint operator on the Hilbert space L (R+). 1 − dx2 2 12 Remark 9.3. The matrix for the operator ∆ on ℓ (1) is banded (cf. B.4): 1 §

c(x), y = x,

M∆ (x, y) = δx, ∆1δy 1 =  cxy, y x, (9.10) 1 h i − ∼ 0, else.  13  The bandedness of M∆1 is a crucial element of the above proof; Example B.21 shows how 14 this proof technique can fail without bandedness. See also Remark 9.3 and Example B.25 15 for what can go awry without bandedness. 16 However, bandedness is not suﬃcient to guarantee essential self-adjointness. In fact, 2 17 see Example B.25 for a Hermitian operator on ℓ which is not self-adjoint, despite having 18 a uniformly banded matrix, that is, there is some n N such that each row and column ∈ 19 has no more than n nonzero entries. The essential self-adjointness of ∆1 in this context is 20 likely a manifestation of the fact that the banding is geometrically/topologically local; the 0 21 nonzero entries correspond to the vertex neighbourhood of a point in G .

22 9.1.2. The spectral representation of ∆. It is clear from Lemma 4.63 that v dom when- 2 ∈ E 23 ever v, ∆ v ℓ (1). However, this condition is not necessary, and the precise characteriza- 1 ∈ 24 tion of dom is more subtle. E 2 1/2 25 Theorem 9.4. For all u ℓ (1) dom , u = ∆ˆ uˆ 2. Therefore, can be charac- ∈ ∩ E k kE k k HE 26 terized in terms of the spectral resolution of ∆ as

2 0 . 1/2 ℓ (1) dom = v : G C .. ∆ vˆ < , (9.11) ∩ E { → k 1 k2 ∞} 27 where vˆ is the image of v in the spectral representationb of ∆1.

28 Proof. Theorem 9.2 also gives a spectral resolution

2 ∆= λE(dλ), E : (R+) Proj(ℓ ). (9.12) Z B →

29 Applying the functional calculus to the Borel function r(x) = √x, we have

1/2 1/2 1/2 2 . 2 ∆ = λ E(dλ), dom ∆ = v ℓ .. λ E(dλ)v < . (9.13) Z { ∈ Z | | · k k ∞} 2 1/2 30 This gives v ℓ (1) dom if and only if v + k dom ∆ for some k C. However, ∈ ∩ E 1/2 ∈ ∈ 31 ∆(v + k) = ∆v, so the same is true for ∆ by the functional calculus. 100 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Remark 9.5. It is important to observe that dom is not simply the spectral transform of 1 1/2 . 2 1/2 E 2 dom ∆ˆ = vˆ .. vˆ L and ∆ˆ vˆ < . The restrictionv ˆ L must be removed because 2 { ∈ k k ∞} ∈ 2 there are many functions of ﬁnite energy which do not correspond to L functions. For an 3 elementary yet important example, see Figure 10 of Example 14.16. Indeed, recall from 4 2 Corollary 4.65 that no nontrivial harmonic function can be in ℓ ; see Example 13.2. In this, 5 example v is equal to the constant value 1 on one inﬁnite subset of the graph, and equal to 6 the constant value 0 on another. 7

d Remark 9.6. For the example of the integer lattice Z , Remark 14.21 shows quite explicitly 8 why the addition of a constant to v has no eﬀect on the spectral (Fourier) transform. 9 ∈HE In this example, one can see directly that addition of a constant k before taking the trans- 10 form corresponds to the addition of a Dirac mass after taking the transform. As the Dirac 11 mass is supported where the transform of the function vanishes,it has no eﬀect. 12

2 We can also give a reproducing kernel for ∆ on ℓ (1). Recall from (2.1) that the vertex 13 0 0 . 0 neighbourhood of x G is G(x) := y G .. y x G . Also recall from Deﬁnition 2.7 14 ∈ { ∈ ∼ }⊆ that x < G(x) and from Deﬁnition 2.3 that the conductance function is c(x) := y x cxy. 15 ∼ P Lemma 9.7. The functions ∆δx x G0 = c(x)δx c(x )χ x G0 give a reproducing kernel 16 · G(x) 2 { } ∈ { − } ∈ for ∆ on ℓ (1). 17

Proof. Since δ , u = u(x), the result follows by 18 h x i1

∆v(x) = c(x)v(x) cxyv(y) = c(x)δx, v 1 c(x )χ , v 1 = ∆δx, v 1. − h i − h · G(x) i h i Xy x ∼ 2 This is a recapitulation of (9.10). Since c(x) < , it is clear that ∆δ ℓ (1). 19 ∞ x ∈ 9.2. The transfer operator. 20

Definition 9.8. We say the graph (G, c) satisfies the Powers bound iff 21

c := sup c(x) < . (9.14) k k∞ x G0 ∞ ∈

The terminology “Powers bound” stems from [Pow76b], wherein the author uses this 22 bound to study the emergence of long-range order in statistical models from quantum me- 23 chanics. Our motivation is somewhat different, and most of our results do not require 24 such a uniform bound. However, when satisfied, it implies the boundedness of the graph 25 Laplacian (and hence its self-adjointness) and the compactness of the associated transfer 26 operator; see 9.2. 27 § 2 The fact that the Powers bound entails the inclusion ℓ (1) (see Theorem 9.18) 28 ⊆ HE illustrates how strong this assumption really is. While the Laplacian may be unbounded for 29 infinite networks in general, Theorem 9.9 gives one situation in which ∆ is always bounded. 30 To see sharpness, note that this bound is obtained in the integer lattices of Example 14.2. 31 2 t In particular, for d = 1, we have ∆ = sup 4(sin ) = 4 = 2 c . 32 k k | 2 | k k 2 Theorem 9.9. As an operator on ℓ (1), the Laplacian satisfies ∆ 2 c , and hence is 33 k k1 ≤ k k a bounded self-adjoint operator whenever the Powers bound holds. Moreover, this bound 34 is sharp. 35

Proof. Since ∆= c T, this is clear by the following lemma. 36 − OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 101

Recall from Deﬁnition 2.9 that the transfer operator T acts on an element of dom ∆ by

(T v)(x) := cxyv(y). (9.15) Xy x ∼ 1 1 One should not confuse T with the (bounded) probabilistic transition operator P = c− T; 2 recall that the Laplacian may be expressed as ∆ = c T, where c denotes the associated − 2 3 multiplication operator. Note that T = c ∆ is Hermitian on ℓ (1)by(9.6). This is a bit − 4 of a surprise, since transfer operators are not generally Hermitian. Unfortunately, T1 may 5 not be self-adjoint. In fact, the transfer operator of Example B.25 is not even essentially 6 self-adjoint; see also [vN32a, Rud91, DS88].

7 Lemma 9.10. T c . k 1 k ≤ k k 8 Proof. Recall that T1 = T pointwise. The triangle inequality and Schwarz inequality give

f, T f 1 f (x) √cxy √cxy f (y) |h i |≤ | | xXG0 Xy x ∈ ∼ 1/2 f (x) c(x)1/2 c f (y) 2 ≤ | |  xy| |  xXG0 Xy x  ∈  ∼  1/2  1/2 c(x) f (x) 2 c f (y) 2 . ≤  | |   xy| |  xXG0  x,Xy G0   ∈   ∈      /    2 1 2 9 Since the both factors above may be bounded above by c f 1 (using another ap- k k · k k 2 10 plication of Schwarz for the one on the right), we have f, T f c f . |h ic| ≤ k k · k k1 11 Remark 9.11. When d = 1, Example 14.2 (the simple integer lattice) shows that the bound 12 of Corollary 9.10 is sharp. From the proof of Lemma 14.3, one ﬁnds that

T = sup 2cos t = 2 = 1 + 1 = c(n), n Z. k k | | ∀ ∈

Deﬁnition 9.12. Let cx be deﬁned by cx(y) = cxy, so c c := c2 (9.16) x · x xy Xy x ∼ 2 13 We denote this with the shorthand c = c c . x x · x 2 2 14 Theorem 9.13. If c is bounded, then T1 : ℓ (1) ℓ (1) is bounded and self-adjoint. If T1 2 → 15 is bounded, then cx is a bounded function of x.

16 Proof. ( ) The boundedness of T is Lemma 9.10. Any bounded Hermitian operator is ⇒ 17 immediately self-adjoint; see Deﬁnition B.9. 2 18 ( ) For the converse, suppose that cx is unbounded. It follows that there is a sequence ⇐ 0 2 19 xn ∞ G with c , and a path γ passing through each xn exactly once. Consider { }n=1 ⊆ xn → ∞ 20 the orthonormal sequence δxn : T δ (x) = c δ (y) = c δ (y), and T δ 2 = c2 . 1 z xy z zy z k 1 zk1 yz Xy x Xy z Xy z ∼ ∼ ∼ 2 21 Then letting z run through the vertices of γ, it is clear that T δ . k 1 zk1 →∞ 102 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Recall from Definition 4.53 that u(x) vanishes at iff for any exhaustion G , one can 1 ∞ { k} always find k such that u(x) <ε for all x < Gk. 2 k k∞ Using a nested sequence as describe in Definition 4.53,itisnotdifficult to prove that T1 is always the weak limit of the finite-rank operators Tn defined by Tn := Pn T1 Pn, where . P is projection to G = span δ .. x G , so that n n { x ∈ n} Tn v(x) = χ (T1 v)(x) = cxyv(y). (9.17) Gn(x) Xy x y ∼G ∈ n Norm convergence does not hold without further hypotheses (see Example 14.25) but we 3 do have Theorem 9.14, which requires a lemma. 4

2 Theorem 9.14. If c ℓ and deg(x) is bounded on G, then the transfer operator T1 : 5 2 2 ∈ 2 ℓ (1) ℓ (1) is compact. If T is compact, then c vanishes at . 6 → 1 x ∞ Proof. ( ) Consider any nested sequence G of ﬁnite connected subsets of G, with 7 ⇐ { k} G = Gk, and the restriction of the transfer operator to these subgraphs, given by TN := 8 P T P , where P is projection to G . Then for D := T T , consider the operator 9 N 1S N N N N 1 − N norm 10

0 P T P D = N 1 ⊥N , (9.18) N P T P P T P k k ⊥N 1 N ⊥N 1 ⊥N where the ONB for the matrix coordinates is given by δ for some enumeration of 11 { xk } k∞=1 the vertices. Since deg(x) is bounded, the matrices for ∆1 and hence also T1 are uniformly 12 banded; whence DN is uniformly bounded with band size bN and Lemma B.24 applies. 13

Since the ﬁrst N entries of DN+bN v are 0, we have 14

2 bm 2 ∞ ∞ 2 D + v = c = c(x ) , k N bN k  mnk  m mX=N+1 Xk=1  mX=N+1   2   which tends to 0 for c ℓ (1).   15 ∈ 2 ( ) For the converse, suppose that cx does not vanish at . It follows that there is a 16 ⇒ 0 2 ∞ sequence xn ∞ G with c 1 ε > 0, and a path γ passing through each of them 17 { }n=1 ⊆ k xn k ≥ exactly once. By passing to a subsequence if necessary, is also possible to request that the 18 sequence satisﬁes 19

G(x ) G(x + ) = ∅, n, n ∩ n 1 ∀ since the sequence need not contain every point of γ. Consider the orthonormal sequence 20

δxn . We will show that T1 δxn contains no convergence subsequence: 21 T δ T δ = c δ c δ 1 xn − 1 xm xny xn − xmz xm yXxn zXxm ∼ ∼ 2 2 2 2 2 T1 δx T1 δx = T1 δx + T1 δx = c + c 2ε. k n − m k k n k k m k xny xmz ≥ yXxn zXxm ∼ ∼ There are no cross terms in the ﬁnal equality by orthogonality; xn+1 was chosen to be far 22 enough past xn that they have no neighbours in common. 23

Corollary 9.15. If c vanishes at and deg(x) is bounded, then T is compact. 24 ∞ 1 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 103

1 Proof. The proof of the forward direction of Theorem 9.14 just uses the hypotheses to 2 show that supx,y cxy can be made arbitrarily small by restricting x, y to lie outside of a 3 suﬃciently large set.

4 9.2.1. Fredholm property of the transfer operator. A stronger form of the Theorem 9.17 5 was already obtained in Corollary 4.21, but we include this brief proof for its radically 6 contrasting ﬂavour.

7 Deﬁnition 9.16. A Fredholm operator L is one for which the kernel and cokernel are ﬁnite 8 dimensional. In this case, the Fredholm index is dimker L dim ker L . Alternatively, L is − ∗ 9 a Fredholm operator if and only if Lˆ is self-adjoint in the Calkin Algebra, i.e., L = S + K, 10 where S = S ∗ and K is compact.

11 Theorem 9.17. If c vanishes at infinity, then (α, ω) is nonempty. P 12 Proof. When the Powers bound is satisfied, the previous results show ∆ is a bounded self- 13 adjoint operator, and T is compact. Consequently, ∆ is a Fredholm operator. By the Fred- 2 14 holm Alternative, ker ∆ = 0 if and only if ran ∆ = ℓ (1). Modulo the harmonic functions, 2 15 ker ∆= 0, so δα δω has a preimage in ℓ (1). − 9.2.2. Some estimates relating and ℓ2(1). In this section, we make the standing assumption that the functions underHE consideration lie in ℓ2(1). Strictly speaking, elements of are equivalence classes, but each has aH uniqueE ∩ representative in ℓ2 and it is understoodHE that we always choose this one. Our primary tool will be the identity 2 (u, v) = u, ∆v 1 from (9.5), which is valid on the intersection ℓ (1). For example, Enote that thish immediatelyi gives HE ∩ 2 1/2 2 v, ∆v = ∆v 1, and (v) = ∆ v 1, (9.19) h iE k k E k k 16 where the latter follows by the spectral theorem. Theorem 3.28 showed that (α, ω) , ∅, P 17 for any choice of α , ω. It is natural to ask other questions in the same vein. 2 18 (i) Is ℓ (1) ? No: consider the 1-dimensional integer lattice described in Exam- ⊆ HE 19 ple 14.29. 2 20 (ii) Is ℓ (1)? No: consider the function f defined on the binary tree in Exam- HE ⊆ 21 ple 13.3 which takes the value 1 on half the tree and 1 on the other half (and is 0 at − 22 o). This function has energy ( f ) = 2, but it is easily seen that there is no k for which 2 E 23 f + k ℓ (1). ∈ 2 2 24 (iii) Does ∆v ℓ imply v or v ℓ (1)? Neither of these are true, by the example in ∈ ∈HE ∈ 25 the previous item. 2 26 (iv) Is (α, ω) ℓ ? No: consider again the 1-dimensional integer lattice, with α < ω. P ⊆ 27 Then if v (α, ω), it will be constant (and equal to v(α)) for x to the left of α, and ∈P n 28 it will be constant (and equal to v(ω)) for xn right of ω. 1/2 29 Lemma 9.18. v ∆ v 1 for every v . If the Powers bound (9.14) is satisfied, 2 k kE ≤ k k·k k ∈HE 30 then ℓ (1) . ⊆HE 2 31 Proof. Since v = v, ∆v 1, this is immediate from Lemma 9.9. k kE h i 2 32 Lemma 9.19. If ∆ is bounded on ℓ (1), then it is bounded with respect to . E 2 33 Proof. The hypothesis implies ∆ is self-adjoint on ℓ , so that one can take the spectral 2 34 representation ∆ˆ on L (X, dν) and perform the following computation:

1/2 1/2 ∆v = ∆∆ v 1 ∆ˆ ∆ v 1 = ∆ˆ v . k kE k k ≤ k k∞ · k k k k∞ · k kE 104 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2 Lemma 9.20. Let v ℓ (1). Ifv 0 (or v 0), then v v 1. If v is bipartite and 1 ∈ ≥ ≤ k kE ≤ k k alternating, then v v 1. 2 k kE ≥ k k Proof. Both statements follow immediately from the equality 3

(v) = v, ∆v = v, v v, T v = v c , v(x)v(y). E h i1 h i1 − h i1 k k1 − x y Xy x ∼

2 9.3. The Laplacian and transfer operator on ℓ (c). In 9.1–9.2, we studied ∆ and T 4 2 § as operators on the unweighted space ℓ (1). In this section, we consider the renormalized 5 versions of these operators and attempt to carry over as many results as possible to the 6 2 context of ℓ (c). 7

2 0 .. 2 ℓ (c) := u : G C . x G0 c(x) u(x) < , (9.20) { → ∈ | | ∞} with the inner product P 8

u, ∆v := c(x)u(x)∆v(x). (9.21) h ic xXG0 ∈ Take the operator ∆ deﬁned on span δ , the dense domain consisting of functions with 9 { x} ﬁnite support. Then let ∆c denote the closure of ∆ with respect to (9.21), that is, its minimal 10 2 self-adjoint extension to ℓ (c). 11

2 Lemma 9.21. For u ℓ (c) and v , 12 ∈ ∈HE

u(x)∆v(x) √2 u c v . (9.22) | |≤ k k · k kE xXG0 ∈

Proof. Apply the Schwarz inequality twice, ﬁrst with respect to the x summation, then 13 with respect to y: 14

u(x)∆v(x) = √cxyu(x) √cxy(v(x) v(y)) | | − xXG0 x,Xy G0 ∈ ∈ 1/2 1/2 c u(x) 2 c v(x) v(y) 2 ≤  xy| |   xy| − |  yXG0 Xx y  Xx y  ∈  ∼   ∼   1/2   1/2 2 c u(x) 2 c v(x) v(y) 2 , ≤  xy| |  2 xy| − |  xX,y G0   x,Xy G0   ∈   ∈      and the resulting inequality retroactively justiﬁes  the implicit initial Fubination. 15

2 Deﬁnition 9.22. For u span δx ℓ (c), deﬁne a linear functional ξu on by 16 ∈ { }⊆ HE

ξu(v) := u(x)∆v(x) (9.23) xXG0 ∈ Then ξu is continuous because ξu(v) u c v by Lemma 9.21, whence Riesz’s lemma 17 | | ≤ k k · k kE 2 gives a w in for which ξu(v) = w, v holds for every v in. Let J : ℓ (c) 18 ∈ F h iE ∈ F → HE denote the map which sends u w, i.e., the map deﬁned by 19 7→ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 105

Ju, v = ξu(v). (9.24) h iE

1 Deﬁnition 9.22 allows one to see directly that Jδx = [δx]:

Jδx, v = δx(y)∆v(y) = ∆v(x) = δx, v , for all v . h iE h iE ∈HE yXG0 ∈ This idea is the reason for Definition 9.22 and also Theorem 9.23. It is also easy to see 2 2 that δx c = c(x) = δx , although the two norms c and are clearly different in k k k kE k · k k · kE general. In fact, if ϕ = x F ξxδx span δx (so F is finite), then one may easily compute ∈ ∈ { } ϕ 2 = Pc(x) ξ 2, whereas ϕ 2 = ϕ 2 c ξ ξ . k kc | | k k k kc − xy x y Xx F E Xx F Xy x ∈ ∈ ∼

2 Theorem 9.23. The map J is the quotient map induced by the equivalence relation u v 2 ≃ 3 iﬀ u v = const, and gives a continuous embedding of ℓ (c) into in with − F

2 Ju √2 u c, u ℓ (c). (9.25) k kE ≤ k k ∀ ∈ 4 Furthermore, the closure of ran J with respect to is in. E F 5 Proof. The formulation of J in (9.24) gives

Ju, vx = ξu(vx) = u(y)∆vx(y) = u(y)(δx δo)(y) = u(x) u(o). h iE − − xXG0 xXG0 ∈ ∈ 6 This shows that J is the quotient map as claimed. The bound (9.25) follows immediately 2 7 upon combining (9.22) with (9.23). Now let w = Ju for any u ℓ (c) and apply ξ to ∈ u 8 v = fx = P invx in to get F ∈F

w(x) w(o) = fx, w = ξu( fx) = u(y)(δx δo)(y) = u(x) u(o). − h iE − − yXG0 ∈ 9 Since f is thus a reproducingkernel for any element of ran J, this shows that ran J in, { x} ⊆F 10 and hence

Ju, v = ξu(v) √2 u c v = Ju √2 u c. |h iE| | |≤ k k · k kE ⇒ k kE ≤ k k 11 Because ran J contains span δ , it is easy to see that the -closure of ran J is equal to { x} E 12 in. F 2 Lemma 9.24. The adjoint map J∗ : ℓ (c) is given by HE → J∗u = u Pu, (9.26) − 13 where P is the probabilistic transition operator deﬁned in (5.32).

14 Proof. First let u span δx and v and note that Ju, v = u, ∆v 1 by (9.5). Then ∈ { } ∈HE h iE h i

1 u, ∆v = u(x)∆v(x) = u(x)c(x) v(x) c v(y) , h i1  − c(x) xy  xXG0 xXG0  Xy x  ∈ ∈  ∼    106 PALLEE.T.JORGENSENANDERINP.J.PEARSE whence Ju, v = u, (1 P)v c on the subspace span δx , which is dense in in in the 1 h iE h 2− i { } F norm and dense in ℓ (c) in the norm c. 2 k · kE k · k

Remark 9.25. Lemma 9.24 provides another proof of Theorem 9.23: 3

Alternative proof of Theorem 9.23. Suppose that v ran(J)⊥ so that Ju, v = 0 4 2 ∈ ⊆ HE h iE for all u ℓ (c). Then 5 ∈

2 Ju, v = u, J∗v c = u, v Pv c, u ℓ (c) h iE h i h − i ∀ ∈ 2 by (9.26), so that v Pv = 0 in ℓ (c). Then recall that v arm iﬀ ∆v = 0 iﬀ Pv = v. This 6 − clo ∈H shows that ran(J) = arm, and hence ran(J) = ran(J) = arm⊥ = in. 7 ⊥ H ⊥⊥ H F

Remark 9.26. Many authors use J∗ J = 1 P as the deﬁnition of the Laplace operator on 8 2 − 2 ℓ (c). It is intriguing to note that for v , one has v Pv ℓ (c), even though it is quite 9 2∈HE − ∈ possible that neither v nor Pv lies in ℓ (c) (for an extreme example, consider v arm). 10 ∈ H Furthermore, note that for every v , one has 11 ∈HE

2 2 2 c(x) v(x) Pv(x) = J∗v c √2 v < . (9.27) | − | k k ≤ k kE ∞ xXG0 ∈ 2 2 This obviously implies a bound c(x) v(x) Pv(x) B , whence 12 | − | ≤

1/2 v(x) Pv(x) Bc(x)− . | − |≤ ∁ Consequently, if Gk is any exhaustion of G, then v Pv on Gk for large k (in the sense of 13 2 { } ≈ ℓ (c)). Roughly, one can say that any v tends to being a harmonic function at , and 14 ∈HE ∞ the faster c grows, the better the approximation. 15

1 2 Corollary 9.27. If ∆u = u, then u(x) < and u(x) = O(√c(x)),as x . 16 − c(x) | | ∞ →∞ P 1 Proof. Recall that a defect vector u satisﬁes ∆u = u and hence u Pu = u. 17 − ∈ HE − − c The result follows by substituting the latter into (9.27). This immediately implies a bound 18 1 2 u(x) B, which gives the ﬁnal claim. 19 c(x) | | ≤

Recall from Deﬁnition 4.32 that ∆ denotes the closure of the Laplacian when taken to 20 V have the dense domain V := span vx x G0 o of ﬁnite linear combinations of dipoles. 21 { } ∈ \{ } 0 Corollary 9.28. If c(x) is bounded on G and deg(x) < , then ∆ is essentially self- 22 ∞ V adjoint on . 23 HE

Proof. Suppose w dom ∆∗ satisfying ∆∗w = w. This means there is a K (possibly 24 ∈ V − depending on w) such that w, ∆v K v for all v V. 25 2 |h iE|≤ k kE ∈ Then for u ℓ (c) and v span vx , set ξu(v) = x G0 u(x)∆v(x). As in the proof of 26 ∈ ∈ { } ∈ Theorem 9.23, ξu extends to a continuous linear functionalP on , so applying it to w gives 27 HE

u(x)w(x) = u(x)( w(x)) = u(x)∆w(x) = ξu(w) √2 u c w . − | |≤ k k · k kE xXG0 xXG0 xXG0 ∈ ∈ ∈

However, if c(x) is bounded by c , then 28 k k OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 107

1/2 u = c(x) u(x) 2 c 1/2 u . k kc  | |  ≤ k k k k1 xXG0   ∈  1 Combining the two displayed equations above yields the inequality

1/2 u(x)w(x) √2 c u 1 w . X ≤ k k k k · k kE 2 2 This shows u x G0 u(x )w(x) is a continuous linear functional on ℓ (c), so that Riesz’s 7→ 2∈ 2 3 lemma puts w ℓ (1). However, now that w is a defect vector in ℓ (1), Theorem 9.2 ∈P 4 applies, and hence w = 0.

1 5 Deﬁnition 9.29. Let ∆ := c ∆= 1 P = JJ denote the probabilistic Laplace operator c − − ∗ 6 on , as in (2.6). Note that we abuse notation here in the suppression of the quotient HE 2 7 map, so that 1 P denotes an operator on and a mapping ℓ (c). − HE HE → 2 8 Corollary 9.30. For any v , ∆c is contractive on and (1 P)v ℓ (c) with ∈HE HE − ∈

(1 P)v v . k − kE ≤ k kE

9 Proof. Since J is contractive, it follows that J∗ is contractive by basic operator theory; 10 this is a consequence of the polar decomposition applied to J. Then ∆c = JJ∗ is certainly 11 continuous with (1 P)v = JJ∗v √2 J∗v c 2 v . k − kE k kE ≤ k k ≤ k kE 12 Lemma 9.31. ∆c is Hermitian if and only if c(x) is a constant function on the vertices.

13 Proof. This can be seen by computing the matrix representation of ∆c with respect to the δx th 14 ONB , in which case the (x, y) entry is { √c(x) }

δx δy δx(z) c(y)δy(z) δt(z) [M∆ ]x,y = , ∆ = c(z) cty c * √c(x) c(y)+ √c(x)  c(y) − c(y) c zXG0  Xt y  ∈  ∼  p  p p  c(z)δx(z) c(y)δy(z) c(z)δx(z) δt(z) = cty  √c(x) c(y) − √c(x) c(y) zXG0  Xt y  ∈  ∼   p δt(x) p  = c(x)δ c(x) c xy − ty Xt y c(y) p ∼ pδy = c(x)δ c(x) c , xy − xy Xy x c(y) p ∼ p 15 which is not symmetric in x and y. We used ∆δy = c(y)δy t y ctyδt, which follows easily − ∼ 16 from Lemma 4.28. P 2 17 Theorem 9.32. (i) As an operator on ℓ (c), P = I J J is self-adjoint with I P I. − ∗ − ≤ ≤ 18 (ii) As an operator on , P = I JJ∗ is Hermitian if and only if c(x) is constant. HE − 19 Proof. (i) It is easy to check directlythat u, Pv and Pu, v are bothequal to u(x)c v(y). h ic h ic x,y xy 20 Then P

2 2 u, Pu c = u, (I J∗ J)u c = u, u c Ju, Ju = u c Ju . h i h − i h i − h iE k k − k kE 108 PALLEE.T.JORGENSENANDERINP.J.PEARSE

2 2 2 2 Since Ju 2 u c by (9.25), this establishes u c u, Pu c u c , and then self- 1 k kE ≤ k k −k k ≤ h i ≤ k k adjointness follows automatically because P is bounded and Hermitian. 2 (ii) P is Hermitian on iﬀ Pu, v = u, Pv for all u, v . Therefore, choose x , y, both diﬀerent fromHo,E andh use Pi=E 1 hc 1∆itoE compute ∈ HE − − 1 Pvx, vy = vx c− (δx δo), vy h iE h − − iE 1 1 = vx, vy c− (δx δo)(y) + c− (δx δo)(o) h iE − − − 1 = vx, vy . h iE − c(o) 3 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 109

1 10. The dissipation space D and its relation to H HE 2 This section is about an isometric embedding d of the Hilbert space of voltage func- HE 3 tions into the Hilbert space of current functions, and the projection P that relates d to HD d 4 its adjoint. This dissipation space will be needed for several purposes, including the HD 5 resolution of the compatibility problem discussed in 3.3 and the solution of the Dirichlet § 6 problem in the energy space via its solution in the (ostensibly simpler) dissipation space; 7 see Figure 4. The geometry of the embedding d is a key feature of our solution in The- 8 orem 10.8 to a structure problem regarding current functions on graphs. Appendix B.2 9 contains definitions of the terms isometry, coisometry, projection, initial projection, final 10 projection, and other notions used in this section. After completing a first draught of this 11 paper, we discovered several of the results of this section in[LP09] and [Soa94]. Both of 12 these texts are excellent; Lyons emphasizes connections with probability and 2 and 9 § § 13 are most pertinent to the present discussion, and Soardi emphasizes the (co)homological 14 perspective and parallels with vector calculus. 1 15 In this section, we will find it helpful to use the notation Ω(x, y) = c−xy . Definition 10.1. Considering Ω as a measure on G1, currents comprise the Hilbert space 1 . := I : G C .. I is antisymmetric and I < , (10.1) HD { → k kD ∞} 16 where the norm and inner product are given by

I := D(I)1/2 and (10.2) k kD 1 I , I := D(I , I ) = Ω(x, y)I (x, y)I (x, y). h 1 2iD 1 2 2 1 2 (x,Xy) G1 ∈

2 1 2 17 Observe that D = ℓ (G , Ω)but it is not truethat can be represented as an ℓ space H H2E 18 in such an easy manner (but see Theorem 7.19). Asan ℓ space, is obviously complete. HD 19 However, is also blind to the topology of the underlying network and this is the reason HD 20 why the space of currents is much larger than the space of potentials. This last statement is 21 made precise in Theorem 10.12, where it is shown that D is larger than by precisely H HE 22 the space of currents supported on cycles. 23 The fundamental relationship between and D is given by the following operator HE H 24 which implements Ohm’s law. It can also be considered as a boundary operator in the 25 sense of homology. Further motivation for the choice of symbology is explained in 10.3. § Deﬁnition 10.2. The drop operator d = dc : D is deﬁned by HE →H (dv)(x, y) := c (v(x) v(y)) (10.3) xy − 26 and converts potential functions into currents (that is, weighted voltage drops) by imple- 27 menting Ohm’s law. In particular, for v (α, ω), we get dv (α, ω). ∈P ∈F As Lyons comments in [LP09, 9.3], thinking of the resistance Ω(x, y) as the length of the edge (xy), d is a discrete version§ of directional derivative: v(x) v(y) ∂v (dv)(x, y) = − (x). Ω(x, y) ≈ ∂y

28 Lemma 10.3. d is an isometry.

2 2 29 Proof. Lemma 3.16 may be restated as follows: dv D = u . k k k kE 110 PALLEE.T.JORGENSENANDERINP.J.PEARSE

d d∗ / D / HE H HE in / bd / in F N F arm / ir / arm H K H yc / 0 C

Figure 4. The action of d and d∗ on the orthogonal components of and D. See HE H Theorem 10.8 and Deﬁnition 10.9.

10.1. The structure of . In Theorem 10.8, we are now able to characterize by 1 HD HD using Lemma 10.3 to extend Lemma 4.22 (the decomposition = in arm). First, 2 HE F ⊕H however, we need some terminology. Whenever we consider the closed span of a set of 3 vectors S in or D, we continue to use the notation [S ] or [S ]D to denote the closure 4 HE H E of the span in or D, respectively. 5 E Deﬁnition 10.4. Deﬁne the weighted edge neighbourhoods 6

cxy, x = z y, ηz = ηz(x, y) := dδz = ∼ (10.4) 0, else.  Then denote the space of all such currents by bd := [d in] = [η ] . This space is 7 N F D z D called ⋆ in [LP09, 2 and 9]. 8 § § Lemma 10.5. D(ηz) = deg(z). 9

Proof. Computing directly, 10

2 D(ηz) = Ω(x, y)ηz(x, y) = 1 = deg(z). (x,Xy) G1 Xy x ∈ ∼

Definition 10.6. For each h arm, we have div(dh) = 0 so that dh satisfies the homo- 11 ∈ H geneous Kirchhoff law by Corollary 3.25. Therefore, we denote ir := d arm = ker div. 12 K H Since the elements of ir are currents induced by harmonic functions, we call them har- 13 K monic currents or Kirchhoff currents. 14

Definition 10.7. Denote the space of cycles in , that is, the closed span of the charac- 15 HD teristic functions of cycles ϑ by yc := [χ ] . 16 ∈ L C ϑ D The space yc is called ^ in [LP09, 2 and 9]. 17 C § § We are now able to describe the structure of . See[LP09, (9.6)]for a different proof. 18 HD Theorem 10.8. = bd ir yc. 19 HD N ⊕ K ⊕C Proof. Lemma 3.23 expresses the fact that d is orthogonal to yc. Since d is an isom- 20 HE C etry, d = d in d arm, and the result follows from Theorem 4.22 and the definitions 21 HE F ⊕ H just above. See Figure 4. 22

Deﬁnition 10.9. Recall that a projection on a Hilbert space is by deﬁnition an operator 23 2 satisfying P = P∗ = P . The following notation will be used for projection operators: 24

P in : in P arm : arm F HE →F H HE →H OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 111

Pd : D d Pd⊥ = P yc : D yc H → HE C H →C P bd : D bd P ir : D ir. N H → N K H → K 1 Figure 4 may assist the reader with seeing how these operators relate.

2 Lemma 10.10. The adjoint of the drop operator d∗ : D is given by H →HE

(d∗I)(x) (d∗I)(y) =Ω(x, y)P I(x, y). (10.5) − d

3 Proof. Since Pd d = d and Pd = Pd∗ by deﬁnition,

dv, I = P dv, I = dv, P I h iD h d iD h d iD 1 = Ω(x, y)c (v(x) v(y))P I(x, y) by(10.3) 2 xy − d (x,Xy) G1 ∈ 1 = c (v(x) v(y))(d∗I)(x) (d∗I)(y)) by(10.5). 2 xy − − x,Xy G0 ∈

4 Remark 10.11. Observe that (10.5) only deﬁnes the function d∗I up to the addition of a 5 constant, but elements of are equivalence classes, so this is suﬃcient. Also, HE

(d∗I)(x) (d∗I)(y) =Ω(x, y)I(x, y). − 6 satisfies the same calculation as in the proof of Lemma 10.10. However, the compatibility 7 problem described in 3.3 prevents this from being a well-defined operator on all of . § HD 8 One can think of d∗ as a weighted boundary operator and d as the corresponding 9 coboundary operator; this approach is carried out extensively in [Soa94], although the 10 author does not include the weight as part of his definition.

Theorem 10.12. d andd∗ are partial isometries with initial and ﬁnal projections

d∗d = I , dd∗ = Pd . (10.6) HE 11 Furthermore, d : bd ir is unitary. HE → N ⊕ K

12 Proof. Lemma 10.3 states that d is an isometry; the ﬁrst identity of (10.6) follows imme- 13 diately. The second identity of (10.6) follows from the computation

dd∗I(x, y) = d d∗I(x) d∗I(y) − = d Ω(x, y)Pd I(x, y) by (10.5) = c Ω(x, y)P I(x, y) Ω(y, y)P I(y, y) by (10.3) xy d − d = Pd I(x, y). Deﬁnition 2.7.

14 The last claim is also immediate from the previous computation.

15 We are now able to give an proof of the completeness of which is independent of HE 16 6.1; see also Remark 6.4. § 17 Lemma 10.13. dom / constants is complete in the energy norm. E { } 112 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Proof. Let v be a Cauchy sequence. Then dv is Cauchy in by Theorem 10.12, so 1 { j} { j} HD it converges to some I (completeness of is just the Riesz-Fischer Theorem). We 2 ∈HD HD now show that v j d∗I : 3 → ∈HE

(v d∗I) = (d∗(dv I)) D(dv I) 0, E j − E j − ≤ j − → again by Theorem 10.12. 4

10.1.1. An orthonormal basis (ONB) for D. Recall from Remark 3.2 that we may always 5 1 H 1 choose an orientation on G . We use the notation ~e = (x, y) G to indicate that ~e is in the 6 ∈ orientation, and ←−e = (y, x) is not. For example, there is a term in the sum 7

D(I) = Ω(~e)I(~e)2 (10.7) ~eXG1 ∈ 1 for ~e, but there is no term for ←−e (and hence no leading coeﬃcient of 2 ). 8 Deﬁnition 10.14. For ~e = (x, y) G1, denote by ϕ the normalized Dirac mass on an edge: ∈ ~e ϕ~e := √c δ~e. (10.8)

Lemma 10.15. The weighted edge masses ϕ form an ONB for . 9 { ~e} HD 1 Proof. It is immediate that every function in in(G ) can be written as a (ﬁnite) linear 10 F 2 combination of such functions. Since D is just a weighted ℓ space (as noted in Deﬁni- 11 1 H tion 10.1), it is clear that in(G ) is dense in . To check orthonormality, 12 F HD

ϕ , ϕ = Ω(~e)ϕ (~e)ϕ (~e) = δ , h ~e1 ~e2 iD ~e1 ~e2 ~e1,~e2 ~eXG1 ∈ 13 where δ~e1,~e2 is the Kronecker delta, since ϕ~e1 (~e)ϕ~e2 (~e) = c~e iff ~e1 = ~e2, and zero otherwise. (Thereis noterm in the sum for e ; see 10.1.1.) Incidentally, the same calculation verifies 14 ←−1 § ϕ . 15 ~e ∈HD Remark 10.16. It would be nice if ϕ~e d , as this would allow us to “pull back” ϕ~e to 16 ∈ HE obtain a localized generating set for , i.e., a collection of functions with finite support. 17 HE Unfortunately, this is not the case whenever ~e is contained in a cycle, and the easiest expla- 18 nation is probabilistic. If x y, then the Dirac mass on the edge (x, y) corresponds to the 19 ∼ experiment of passing one amp from x to its neighbour y. However, there is always some 20 positive probability that current will flow from x to y around the other part of the cycle and 21 hence the minimal current will not be ϕ~e; see Lemma 5.46 for a more precise statement. 22 Of course, we can apply d∗ to obtain a nice result as in Lemma 10.17, however, d∗ ϕ~e 23 will generally not have finite support and may be difficult to compute. Nonetheless, it still 24 has a very nice property; cf. Lemma 10.17. In light of Theorem 10.12 and the previous 25 0 paragraph, it is clear that any element d ϕ is an element of (x, y) for some x G and 26 ∗ ~e P ∈ some y x. 27 ∼ Lemma 10.17. The collection d∗ϕ~e is a Parseval frame for . 28 { } HE Proof. The image of an ONB under a partial isometry is always a frame. That we have a 29 Parseval frame (i.e., a tight frame with bounds A = B = 1) follows from the fact that d is 30 an isometry: 31 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 113

2 2 2 2 d∗ϕ~e, v = ϕ~e, dv D = dv D = v . |h iE| |h i | k k k kE ~eXG1 ~eXG1 ∈ ∈ 1 We used Lemma 10.15 for the second equality and Theorem 10.12 for the third.

2 10.2. The divergence operator. In 10.4, we will see how P allows one to solve certain § d 3 potential-theoretic problems, but first we need an operator which enables us to study ∆ 4 with respect to D rather than . While the term “divergence”is standard in mathematic, H HE 5 the physics literature sometimes uses “activity” to connote the same idea, e.g., [Pow75]– 6 [Pow79]. We like the term “divergence” as it corresponds to the intuition that the elements 7 of are (discrete) vector fields. HD Definition 10.18. The divergence operator is div : D given by H →HE div(I)(x) := I(x, y). (10.9) Xy x ∼ 8 To see that div is densely defined, note that div(δ(x,y))(z) = δx(z) δy(z), and the space of 1 2 1 − 9 finitely supported edge functions in(G ) is dense in ℓ (G , Ω) = . F HD 10 Theorem 10.19. div = ∆d∗, div d = ∆, and div Pd = ∆d∗. 1 11 Proof. To compute ∆d I for a finitely supported current I in(G ), let v := d I so ∗ ∈F ∗

∆(d∗I)(x) = ∆v(x) = c (v(x) v(y)) defn ∆ xy − Xy x ∼

= cxyΩ(x, y)Pd I(x, y) defn d∗ Xy x ∼ 1 = div(Pd I)(x) cxy =Ω(x, y)− .

12 This establishes div Pd = ∆d∗, from which the result follows by Lemma 10.21. The second 13 identity follows from the ﬁrst by right-multiplying by d and applying (10.6). Then the third 14 identity follows from the second by right-multiplying by d∗ and applying (10.6) again.

15 Remark 10.20. Theorem 10.19 may be reformulatedas follows: Let u, v , and I := dv. ∈HE 16 Then ∆v = u if and only if div I = u. This result will help us solve div I = w for general 17 initial condition w in 10.4. Also, we will see in 9.1 that ∆ is essentially self-adjoint. In § § 18 that context, the results of Theorem 10.19 have a more succinct form.

19 Corollary 10.21. The kernel of div is ir yc, whence div P bd = div, div P⊥bd = 0 and K ⊕C N N 20 div( ) in. HD ⊆F 21 Proof. If I ir so that I = dh for h arm, then div I(x) = ∆h = 0 follows from ∈ K ∈ H 22 Theorem 10.19. If I = χ for ϑ , then ϑ ∈ L

div I(x) = χ (x, y) = χ (x, y) + χ (x, y) = ( 1 + 1) + 0 = 0. ϑ ϑ ϑ − Xy x (xX,y) ϑ (xX,y)<ϑ ∼ ∈ 23 To show div( ) in, it now suﬃces to consider I bd. Since div η = ∆δ by HD ⊆ F ∈ N z z 24 Theorem 10.19, the result follows by closing the span.

25 In particular, Corollary 10.21 shows that the range of div lies in , as stated in Deﬁni- HE 26 tion 10.18. The identity div P bd = div implies that the solution space (α, ω) is invariant N F 27 under minimization; see Theorem 10.30. 114 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Remark 10.22. Since div is defined without reference to c, d∗ “hides” the measure c from 1 the Laplacian. To highlight similarities with the Laplacian, recall from Definition 3.7 that 2 a current I satisfies the homogeneous or nonhomogeneous Kirchoff laws iff div I = 0 or 3 div I = δα δω, respectively. In 10.3, we consider an interesting analogy between the 4 − § previous two results and complex function theory. 5

Corollary 10.23. ∆ d div∗ and div div∗ = ∆∆ . 6 ∗ ⊇ ∗ ∗ Proof. The ﬁrst follows from Theorem 10.19 by taking adjoints, and the second follows 7 in combination with Lemma 10.19. The inclusion is for the case when ∆ may be un- 8 bounded, in which case we must be careful about domains. When T is any bounded op- 9 erator, dom T ∗S ∗ dom(S T)∗. To see this, observe that v dom T ∗S ∗ if and only if 10 ⊆ ∈ v dom S , so assume this. Then 11 ∈ ∗

STu, v = Tu, S ∗v T S ∗v u = K u , u dom S T, |h i| |h i| ≤ k k · k k · k k vk k ∀ ∈ for K = T S v . This shows v dom(S T) . 12 v k k · k ∗ k ∈ ∗ Lemma 10.24. For ﬁxed x G0, div is norm continuous in I: ∈ div(I)(x) c(x) 1/2 I . (10.10) | |≤| | k kD

Proof. Using c(x) := y x cxy as in (2.3), direct computation yields 13 ∼ P 2 2 2 div(I)(x) = I(x, y) = √cxy Ω(x, y)I(x, y) | | Xy x Xy x p ∼ ∼

c Ω(x, y) I(x, y) 2 ≤ xy | | Xy x Xy x ∼ ∼ c(x) D( I), ≤| | where we have used the Schwarz inequality and the deﬁnitions of c, D, div. 14

1/2 Corollary 10.25. For v , ∆v(x) c(x) v . 15 ∈HE | |≤ k kE Proof. Apply Theorem 10.24 to I = dv and use the second claim of Theorem 10.19. 16

One consequence of the previous lemma is that the space of functions satisfying the 17 nonhomogeneous Kirchhoﬀ condition (3.5) is also closed, as we show in Theorem 10.30. 18 In Remark 4.15, we discussed some reproducing kernels for operators on ; we now 19 HE introduce one for the divergence operator div, using the weighted edge neighbourhoods 20 η of Deﬁnition 10.4. 21 { z} Lemma 10.26. The currents η form a reproducing kernel for div. 22 { z} Proof. By Lemma 10.24, the existence of a reproducing kernel follows from Riesz’s The- 23 orem. Since it must be of the form div(I)(z) = k , I , we verify 24 h z iD

η , I = Ω(x, y)η (x, y)I(x, y) = I(z, y) = div(I)(z). h z iD z (x,Xy) G1 Xy z ∈ ∼ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 115

1 10.3. Analogy with calculus and complex variables. The material in this paper bears 2 many analogies with vector calculus and complex function theory. Several points are obvi- 3 ous, like the existence and uniqueness of harmonic functions and the discrete Gauss-Green 4 formula of Lemma 2.13. In this section, we point out a couple more subtle comparisons. 5 The drop operator d is analogous to the complex derivative

d ∂ 1 ∂ 1 ∂ ∂ = = := + ,

dz ∂z 2 ∂x i ∂y!

6 as may be seen from the discussion of the compatibility problem in 3.3. Recall from the § 7 proof of Theorem 10.8 that Lemma 3.23 expresses the fact

I, χ D = 0, ϑ v such that dv = I. h ϑi ∀ ∈L ⇐⇒ ∃ ∈HE 8 This result is analogous to Cauchy’s theorem: if v is a complex function on an open set, 9 then v = f ′ (that is, v has an antiderivative) if and only if every closed contour integral of 10 v is 0. Indeed, even the proofs of the two results follow similar methods. 11 The divergence operator div may be compared to the Cauchy-Riemann operator

∂ 1 ∂ 1 ∂ ∂¯ = := . (10.11)

∂z¯ 2 ∂x − i ∂y!

12 Indeed, in Theorem 10.19 we found that div d = c∆, which may be compared with the ¯ 1 ¯ 13 classical identity ∂∂ = 4 ∆. The Cauchy-Riemann equation ∂ f = 0 characterizes the 14 analytic functions, and div I = 0 characterizes the currents satisfying the homogeneous 15 Kirchhoff law; see Definition 3.6. 16 In 10.4, we give a solution for the inhomogeneous equation div I = w when w is § 17 given and satisfies certain conditions. The analogous problems in complex variables are as 18 follows: let W C be a domain with a smooth boundary bd W, and let ∂¯ be the Cauchy- ⊆ 19 Riemann operator (10.11). Suppose that ν is a compactly supported (0, 1) form in W. We 20 consider the boundary value problem

∂¯ f = ν, with ∂ν¯ = 0.

21 The Bochner-Martinelli theorem states that the solution f is given by the following integral 22 representation:

f (z) = f (ζ)ω(dζ, z) ν(ζ) ω(dζ, z), (10.12) Zbd W − ZW ∧ 23 where ω is the Cauchy kernel. In fact, this theorem continues to hold when W is a domain n 24 in C , if one uses the Bochner-Martinelli kernel

(n 1)! n ω ζ, = ζ¯ ζ¯ ζ ... ˆ ζ¯ ζ , ( z) − 2n ( k z¯k)d 1 d 1 ( j) d n d n (10.13) 2πi ζ z − ∧ ∧ ···∧ ∧ | − | Xk=1

25 where ˆj means that the term dζ¯ dζ has been omitted, and where j ∧ j ∂ϕ ∂ϕ ∂ϕ = dzk, and ∂ϕ¯ = dz¯k. k k X ∂zk X ∂z¯k 116 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Indeed, in Lemma 10.26, we obtain a reproducing kernel for div; this is analogous to the 1 Bochner-Martinelli kernel K(z, w); see [Kyt95] for more on the Bochner-Martinelli kernel. 2 Theorem 4.13 shows that vx is analogous to the Bergman kernel, which reproduces the holomorphic functions within L2(Ω), where Ω C is a domain. Indeed, the Bergman kernel is also associated with a metric, the Bergman⊆ metric, which is deﬁned by ∂γ dB(x, y) := inf (t) , (10.14) γ Zγ ∂t

1 where the inﬁmum is taken over all piecewise C paths γ from x to y;cf. [Kra01]. 3

10.4. Solving potential-theoretic problems with operators. We begin by discussing the 4 minimizing nature of the projections P in and P bd. Theorem 10.27 shows how d∗ solves 5 F N the compatibility problem of 3.3: Given a current ﬂow I , there does not necessarily 6 ∈HD exist a potential function v for which dv = I. Nonetheless, there is a potential 7 ∈ HE function associated to I which satisﬁes dv = P bd I, and it can be found via the minimizing 8 N projection. Consequently, Theorem 10.27 can be seen as an analogue of Theorem 3.26. 9

Theorem 10.30 shows that the solution space (α, ω) is invariant under P bd. Coupled 10 F N with the results of Theorem 10.27, this shows that if one can ﬁnd any solution I (α, ω), 11 ∈F one can obtain another solution to the same Dirichlet problem with minimal dissipation, 12 namely, P bd I. 13 N

10.4.1. Resolution of the compatibility problem. In this section we relate the projections 14

P in : in and P bd : D bd = d in F HE −→ F N H −→ N F of Deﬁnition 10.9 to some questions which arose in 3. The operators P in and P bd are 15 § F N minimizing projections because they strip away excess energy/dissipation due to harmonic 16 or cyclic functions: 17

If v (x, y), then P inv is the unique minimizer of in (x, y). 18 • ∈P F E P If I (x, y), then P bd I is the unique minimizer of D in (x, y). 19 • ∈F N F In a similar sense, Pd is also a minimizing projection. 20 Probability notions will play a key role in our solution to questions about divergence in 21 electrical networks (Deﬁnition 10.18), as well as our solution to a potential equation. The 22 divergence will be important again in 11.2 where we use it to provide a foundation for a 23 § probabilistic model which is dynamic (in contrast to other related ideas in the literature) in 24 the sense that the Markov chain is a function of a current I, which may vary. 25

Theorem 10.27. Given v , there is a unique I D which satisﬁes d∗I = v and 26 ∈ HE ∈ H minimizes I . Moreover, it is given by P I, where I is any solution of d I = v. 27 k kD d ∗

Proof. Given v , we can ﬁnd some I D for which d∗I = v, by Theorem 10.12. ∈ HE ∈ H Then the orthogonal decomposition I = Pd I + Pd⊥I gives 2 2 2 2 I = P I + P⊥I P I , (10.15) k kD k d kD k d kD ≥ k d kD so that I P I shows P I minimizes the dissipation norm. Finally, note that 28 k kD ≥ k d kD d d∗Pd I = d∗dd∗I = d∗I, by Corollary 10.10. 29

Corollary 10.28. d is a solution operator in the sense that if I is any element of then 30 ∗ HD d∗I is the unique element v for which dv = Pd I. 31 ∈HE OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 117

Corollary 10.29. DP bd = P in. Hence for I = d(vx vy), N E F − F 1/2 1/2 R (x, y) = (d∗I) = D(Pd I) , and (10.16) W E 1/2 . R (x, y) = min D(I) .. I (x, y) . { ∈F }

1 Proof. Given I D, let I0 = P bd I. Then deﬁne v by v := d∗P bd I0 = d∗P bd I. Applying ∈H N N N 2 d to both sides gives dv = P bd I by(10.6) (since P bd Pd ) so that taking dissipations and N N ≤ 3 applying Lemma 10.3 gives D(P bd I) = D(dv) = (v) = (P inv), because ran d∗P bd N E E F N ⊆ 4 in by Theorem 10.8 and Theorem 10.12. F 0 5 Theorem 10.30. For any α, ω G , the subset (α, ω) is closed with respect to and ∈ F k · kD 6 invariant under P bd. N 7 Proof. From (3.5)and(10.9), we have that I (α, ω)ifandonlydiv I = δα δω. Suppose ∈F D − 8 that I (α, ω) is a sequence of currents for which I I. Then div I = δα δω for { n}⊆F n −−−→ n − 9 every n, and from Lemma 10.24, the inequality

(div I )(x) (div I)(x) c(x) 1/2 I I | n − |≤| | k n − kD 10 gives div I(x) = δα δω. Note that x is ﬁxed, and so c(x) is just a constant in the inequality − 11 above.

12 For invariance, note that div P bd = div by Corollary 10.21. Then I (α, ω) implies N ∈F

div P bd I = div I = δα δω = P bd I (α, ω). N − ⇒ N ∈F

13 Since P bd is a subprojection of P⊥yc and P⊥ir, we have an easy corollary. N C K 0 14 Corollary 10.31. For any α, ω G , (α, ω) is invariant under Pd = P⊥yc and P⊥ir. ∈ F C K 15 Remark 10.32. Putting these tools together, we have obtained an extremely simple method 16 for solving the equation ∆v = δα δω. − 17 (1) Find any current I (α, ω). This is trivial; one can simply take the characteristic ∈F 18 function of a path from α to ω.

19 (2) Apply P bd to I to “project away” harmonic currents and cycles. N 20 (3) Apply d∗ to P bd I. Since P bd I d in, this only requires an application of Ohm’s N N ∈ F 21 law in reverse as in (10.5).

22 Then v = d∗P bd I is the desired energy-minimizing solution (since any harmonic compo- N 23 nent is removed). As a bonus, we already obtained the current Pd I induced by v. The

24 only nontrivial part of the process described above is the computation of P bd. For further N 25 analysis, one must understand the cycle space yc of G and the space ir of harmonic C K 26 currents. We hope to make progress on this problem in a future paper, see Remark 16.6. 118 PALLEE.T.JORGENSENANDERINP.J.PEARSE

11. Probabilistic interpretations 1

In 7, we constructed a measure P on G′ , where G G′ is a certain Gel’fand 2 § S S ⊆ HE ⊆ S triple. In this section, we develop a different but analogous measure on the space of infinite 3 paths in bd G. We carry out this construction for harmonic functions on (G, c) in 11.1, 4 cxy § wherethe measure is defined in terms of transition probabilities p(x, y) = c(x) of the random 5 walk, and the associated cylinder sets. When the random walk on (G, c) is transient, the 6 current induced by a monopole gives a unit flow to infinity; such a current induces an 7 1 orientation on the edges G and a new, naturally adapted, Markov chain. The state space of 8 0 this new process is also G , but the transition probabilities are now defined by the induced 9 I(x,y) current p(x, y) = . We call the fixed points of the corresponding transition operator 10 divI (x) the “forward-harmonic” functions, and carry out the analogous construction for them in 11 11.2. The authors are presently working to determine whether or not these measures can 12 § be readily related to each other or the measure P of 7.2. 13 § 11.1. The path space of a general random walk. We begin by recalling some terms 14 from 5.2.3, and providing some more detail. Let γ = (x , x , x ,..., x ) be any finite path 15 § 0 1 2 n starting at x = x0. The probability of a random walk started at x traversing this path is 16

n P(γ) := p(xk 1, xk), (11.1) − Yk=1 cxy where p(x, y) := c(x) is the probability that the walk moves from x to y as in (5.32). This 17 intuitive notion can be extended via Kolmogorov consistency to the space of all inﬁnite 18 th paths starting at x. Let Xn(γ) denote the n coordinate of γ; one can think of γ as an event 19 and Xn as the random walk (a random variable), in which case 20

Xn(γ) = location of the random walk at time n. (11.2)

Definition 11.1. Let Γ denote the space of all infinite paths γ in (G, c). Then a cylinder set 21 in Γ is specified by fixing the first n coordinates: 22

. Γ , ,..., := γ Γ .. X (γ) = x , k = 1,..., n . (11.3) (x1 x2 xn) { ∈ k k } (c) Deﬁne P on cylinder sets by 23

n (c) P (Γ(x ,x ,...,x )) := p(xi 1, xi). (11.4) 1 2 n − Yi=1

Remark 11.2. It is clear from Definition 11.1 that the probability of a random walk fol- 24 lowing the finite path γ = (x0, x1, x2,..., xn) is equal to the measure of the set of all in- 25 finite walks which agree with γ for the first n steps: combining (11.1) and (11.4) gives 26 (c) P (Γ(x1,x2,...,xn)(x)) = P(γ). Observe that (11.4) is a conditional probability: 27

(c) (c) P (Γ , ,..., (x)) = P γ Γ(x) X (γ) = x , k = 1,..., n . (11.5) (x1 x2 xn) { ∈ | k k } Remark 11.3 (Kolmogorov consistency). We use Kolmogorov’s consistency theorem to constructa measure on the space of paths beginningat vertex x G0,see[Jor06, Lem. 2.5.1] for a precise statement of this extension principle in its function∈ theoretic form and [Jor06, OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 119

Exc. 2.4–2.5] for the method we follow here. The idea is that we consider a sequence of functionals µ(n) , where µ(n) is defined on { } χ .. n := span Γ . xi xi 1, i = 1,..., n . (11.6) A { (x0,...,xn) ∼ − } Alternatively, . := f : Γ R .. f (γ ) = f (γ ) whenever X (γ ) = X (γ ) for k n . (11.7) An { → 1 2 k 1 k 2 ≤ } That is, an element of n cannot distinguish between two paths which agree for the first n steps. This means thatAµ(n) is a “simple functional” in the sense that it is constant on each cylinder set of level n: (n) (n) χ µ [ f ] = a(x0,...,xn)µ [ ]. (11.8) Γ(x0,...,xn) x0X,...,xn (n) (n+1) (n) 1 If the functionals µ are mutually consistent in the sense that µ [ f ] = µ [ f ], then 2 Kolmogorov’s consistency theorem gives a unique Borel probability measure on the space 3 of all paths. More precisely, Kolmogorov’stheorem gives the existence of a limit functional 4 which is defined for functions on paths of infinite length, and this corresponds to a measure 5 by Riesz’s Theorem; see [Jor06, Kol56].

6 In the following, we let 1 denote the constant function with value equal to 1.

(n) Theorem 11.4 (Kolmogorov). If each µ : n R is a positive linear functional satisfying the consistency condition A → µ(n+1)[ f ] = µ(n)[ f ], for all f , (11.9) ∈ An then there exists a positive linear functional µ defined on the space of functions on infinite paths such that µ[ f ] = µ(n)[ f ], f , (11.10) ∈ An 7 where f is considered as a function on an infinite path which is zero after the first n edges. (n) 8 Moreover, if we require the normalization µ [1] = 1, then µ is determined uniquely.

(c) 9 We now show that P extends to a natural probability measure on the space of infinite 10 paths Γ(x). Theorem 11.5. For (G, c), there is a unique measure P(c) defined on Γ which satisfies

(n) (c,n) 11 Proof. We must check condition (11.9) for µ = P , deﬁned by

n (c,n) χ P ( Γ ) := p(xi 1, xi) (x0,...,xn) − Yi=1 (c,n+1) 12 with (11.4) in mind. Think of V as an element of + and apply P to it: ∈ An An 1

(c,n+1) (c,n+1) χ P [V] = a(x0,...,xn+1)P ( ) Γ(x0,...,xn+1) x0,...,Xxn+1 n = a(x ,...,x ) p(xi 1, xi)p(xn, xn+1) 0 n − x0X,...,xn Xxn+1 Yi=1 120 PALLEE.T.JORGENSENANDERINP.J.PEARSE

n = a(x ,...,x ) p(xi 1, xi) p(xn, xn+1) 0 n − x0X,...,xn Yi=1 xnX+1 xn ∼ = P(c,n)[V],

1 since x + x p(xn, xn+1) = 1. For the second equality, note that f n, so we can use the n 1∼ n ∈ A sameP constant a for each (n + 2)-tuple that begins with (x0,..., xn). 2

11.1.1. A boundary representation for the bounded harmonic functions. 3 Definition 11.6. A cocycle V : Γ R is a measurable function on the infinite path space which is independent of the first finitely→ many vertices in the path: V(γ) = V(σγ), (11.12) where σ is the shift operator, i.e., if γ = (x0, x1, x2,... ), then σγ = (x1, x2, x3,... ). 4

Intuitively, a cocycle is a function on the boundary bdG; it depends only on the asymp- 5 totic trajectory of a path/random walk. A cocycle does not care where the random walk 6 began, only where it goes. More precisely, a cocycle is a special kind of martingale, as we 7 will see below. 8 The goal of this section is to show that the bounded harmonic functions are in bijective correspondence with the cocycles; see Theorem 11.9. That is, the formula

h(x) = Ex[V] (11.13) spells out a bijective correspondence between functions h arm, and cocycles V on the 9 ∈H space of infinite paths. Our present concernis the space of all bounded harmonic functions; 10 we will presently consider the class of finite-energy functions. A good reference for this 11 section is [Jor06, Thm. 2.7.1]. 12 Note that the left hand side of (11.13) involves no measure theory, in contrast to the 13 right-hand side, where the expectation refers to the integration of cocycles V against the 14 (c) (c) probability measure P . The underlying Borel probability space of P is the σ-algebra of 15 measurable sets generated by the cylinder sets in Γ, i.e., by the subsets in Γ which fix only 16 a finite number of places (in the infinite paths). 17 The condition on a measurable function V on Γ which accounts for h defined by (11.13) 18 being harmonic is that V is invariant under a finite left shift; cf. (11.12). It turns out that 19 in making the integrals Ex(V) precise, the requirement that V be measurable is a critical 20 assumption. In fact, there is a variety of non-measurable candidates for such functions V 21 on Γ. 22 Definition 11.7. For any measurable function V : Γ R, we write → (c) Ex[V] := E[V X0 = x] = V(γ) dP (11.14) | ZΓ(x) for the expected value of V, conditioned on the path starting at x. 23

Lemma 11.8. For h arm and any n = 1, 2,..., 24 ∈H

h X dP(c) = p(x, y) h X dP(c) = p(x, y)h(y) dP(c) = Ph(x), (11.16) ◦ 1 ◦ 0 ZΓx Xy x ZΓy Xy x ZΓy ∼ ∼ (c) 2 so iteration and Ph = h gives h Xn dP = h(x) for every n = 1, 2,... . Γx ◦ R Theorem 11.9. The bounded harmonic functions are in bijective correspondence with the cocycles. More precisely, if V is a cocycle, then it deﬁnes a harmonic function via

hV (x) := Ex[V]. (11.17) Conversely, if h is harmonic, then it deﬁnes a cocycle via (c) Vh(γ) := lim h(Xn(γ)), for P -a.e. γ Γ(x). (11.18) n →∞ ∈

3 Proof. ( ) Recall that ∆= c T; we will show that chV = T hV whenever V is a cocycle. ⇒ .. − 4 If Γ(x,y) := γ Γ(x) . X1(γ) = y , then Γ(x) = y x Γ(x,y) is a disjoint union and { ∈ } ∼ S (c) (c) hV (x) = Ex[V] = V(γ) dP = V(γ) dP . ZΓ(x) Xy x ZΓ(x,y) ∼ 5 For each γ Γ , , one has P(γ) = P(x, y)P(σγ) = p(x, y)P(σγ)by(11.1), whence ∈ (x y)

(c) (c) c(x)hV(x) = c(x) p(x, y)V(σγ) dP = cxy V(γ) dP = T hV(x), Xy x ZΓ(x,y) Xy x ZΓ(y) ∼ ∼ 6 where the cocycle property (11.30) is used for the second equality. 7 ( ) Now let h be a bounded harmonic function. Since ⇐

lim h(Xn(γ)) = lim h(Xn+1(γ)) = lim h(Xn(σγ)), n n n →∞ →∞ →∞ 8 the cocycle property (11.30) is obviously satisﬁed whenever the limit exists. Let Σn denote 9 the σ-algebra generated by the cylinder sets of level n, and denote Xn(γ) = xn. Then 10 X + (γ) is a neighbour y x , and n 1 ∼ n

E[h(X + ) Σ ] = E[h(X + ) Σ ] p(x , y) n 1 | n n 1 | n n yXxn ∼ = p(x , y)E[h(X + ) Σ ] n n 1 | n yXxn ∼

= E p(x , y)h(X + ) Σ  n n 1 | n yXxn   ∼    = E[h(X ) Σ ] n | n

= h(Xn). (11.19)

11 Since h is bounded, this shows h(Xn) is a bounded martingale, whence by Doob’s Theorem (c) 12 (cf. [Doo53]), it converges pointwise P -a.e. on Γ and (11.18) makes Vh well-deﬁned (c) 13 P -a.e. on Γ. 122 PALLEE.T.JORGENSENANDERINP.J.PEARSE

( ) We conclude with a proof that these two constructions correspond to inverse op- 1 ↔ erations. If V is a cocycle, we must show that limn EX (γ)[V] = V(γ). To this end, for 2 →∞ n P(c)(A ) 1 (c) A Γ, define the conditioned measure P := ∩· , so that dP = χ dP . Now for 3 ⊆ A P(c)(A) A P(c)(A) A a fixed γ Γ, let A = Γ , γ ,..., γ be the cylinder set whose first n + 1 coordinate agree 4 ∈ n (x X1( ) Xn( )) with γ. Applying the measure identity lim µ(An) = µ(An) for nested sets, we obtain 5 limn PA = δγ as a weak limit of measures. Now 6 →∞ n T

n (c) →∞ EXn(γ)[V] = V(ξ) dP (ξ) = V(ξ) dPAn (ξ) V(ξ) dδγ = V(γ). (11.20) ZΓXn(γ) ZΓ −−−−−−→ ZΓ

On the other hand, if h is harmonic, we must show Ex[Vh] = h. Then for Vh(γ) := 7 limn h(Xn(γ)), boundedness allows us to apply the dominated convergence theorem and 8 →∞ compute 9

(c) (c) Ex[Vh] = lim (h Xn(γ)) dP = lim h Xn(γ) dP . (11.21) n n ZΓx →∞ ◦ →∞ ZΓx ◦ Now the sequenceon the right-handside of (11.21) is constant by Lemma 11.8, so Ex[Vh] = 10 h(x). 11

Remark 11.10. The ( ) direction of the proof of Theorem 11.9 may also be computed 12 ⇒

h (x) = c(x)E [V] = c(x) p(x, y)E [V X = y] = c E [V] = c h (y), V x x | 1 xy y xy V Xy x Xy x Xy x ∼ ∼ ∼ where E [V X = y] = E [V] because the random walk is a Markov process. See, e.g., 13 x | 1 y [LPW08, Prop. 9.1]. 14

11.2. The forward-harmonic functions. The current passing through a given edge may 15 be interpreted as the expected value of the number of times that a given unit of charge 16 passes through it. This perspective is studied extensively in the literature; see [DS84, 17 cxy LP09] for excellent treatments. In this case, p(x, y) = c(x) helps one construct a current 18 which is harmonic, or dissipation-minimizing. However, that is not what we do here; we 19 are interested in studying current functions whose dissipation is finite but not necessarily 20 minimal. In Theorem 11.15, we show that the experiment always induces a “downstream” 21 current flow between the selected two points; that is, a path along which the potential is 22 strictly decreasing. 23 These probability notions will play a key role in our solution to questions about activity; 24 cf. Definition 11.11. We use the forward path measure again in our representation formula 25 (Theorem 11.24) for the class of forward-harmonic functions on G. The corresponding 26 Markov process is dynamically adapted to the network (and the charge on it). This repre- 27 sentation is dynamic and nonisotropic, which sets it apart from other related representation 28 formulas in the literature. 29

11.2.1. Activity of a current and the probability of a path. Given a (fixed) current, we are 30 interested in computing “how much of the current” takes any specified path from x to some 31 other (possibly distant) vertex y. This will allow us to answer certain existence questions 32 (see Theorem 11.15) and providesthe basis for the study of the forward-harmonicfunctions 33 studied in 11.2. Note that, in contrast to (11.1), the probabilistic interpretation given in 34 § Definition 11.13 (and the discussion preceding it) does not make any reference to c. In this 35 section we follow [Pow76b] closely. 36 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 123

Definition 11.11. The divergence of a current I : G1 R is the function on x G0 defined by → ∈ 1 , 2 y x I(x, y) , x α, ω divI (x) := ∼ | | (11.22) 1,P x = α, ω.   0 1 which describes the total “current traffic” passing through x G . Thus, div is an operator 1 0 ∈ 2 mapping functions on G to functions on G ; see 10.2 for details. § 3 It is intuitively obvious that for a connected graph, current should be able to flow be- 4 tween any two points. Indeed, it is a basic result in graph theory that for any connected 5 graph, one can find a path of minimal length between any two points and from Remark 2.6 6 we know that the resistance along such a path is finite. In Theorem 11.15, we will show a 7 strongerresult: that one can always find a path along which the potential function decreases 8 monotonically. In other words, there is always at least one “downstream path” between the 9 two vertices. Somewhat surprisingly, this fact is easiest to demonstrate by an appeal to a 10 basic fact about probability (Lemma 11.14), 11 Our definition of an electrical resistance network is mathematical (Definition 2.7) but is 12 motivated by engineering; modification of the conductors (c) will alter the associated prob- 13 abilities and thus change which current flows are induced, in the sense of Definition 3.17. 14 We are interested in quantifying this dependence. However, on an infinite graph the com- 15 putation of current paths involves all of G, and it is not feasible to attempt to compute these 16 paths directly. Consequently, we feel our proof of Theorem 11.15 may be of independent 17 interest. 0 18 Definition 11.12. Let v : G R be given and suppose we fix α and ω for which v(α) > → 19 v(ω). Then a current path γ (or simply, a path) is an edge path from α to ω with the extra 20 stipulation that v(xk) < v(xk 1) for each k = 1, 2,..., n. Denote the set of all current paths − 21 by Γ = Γα,ω (dependence on the initial and terminal vertices is suppressed when context 22 precludes confusion). Also, define Γα,ω(x, y) to be the subset of current paths from α to ω 1 23 which pass through the edge (x, y) G . ∈ 24 Suppose we fix a source α and sink ω and consider a single current path γ from α to I(x,y) 25 ω. With divI defined as in (11.22), one can consider as the probability that a unit of divI (x) 26 charge at x will pass to a “downstream” neighbour y. Note that I(x, y) > 0 and divI , 0, 27 since we are considering an edge of our path γ. This allows us to define a probability 28 measure on the path space Γα,ω.

Deﬁnition 11.13. If γ Γ follows the vertex path (α = x0, x1, x2,..., xn = ω), the deﬁne the probability of γ by ∈ n I(xk 1, xk) P(γ) := − . (11.23) divI (xk 1) Yk=1 − 29 This quantity gives the probability that a unit of charge at α will pass to ω by traversing 30 the path γ.

31 The following lemma is immediate from elementary probability theory, as it represents 32 the probability of a union of disjoint events, but it will be helpful. Lemma 11.14. Suppose (G, c) is an electrical resistance network and v : G0 R satisﬁes → ∆v = δα δω. Then I = dv satisﬁes − I(x, y) = P(γ). (11.24) γ ΓX(x,y) ∈ α,ω 124 PALLEE.T.JORGENSENANDERINP.J.PEARSE

The method of proof in the next proposition is a bit unusual in that it uses a probability 1 to demonstrate existence. This result ﬁlls a hole in the proof of dist∆(x, y) = distD(x, y) 2 in [Pow76b] (recall (5.1) and(5.3)). 3

Theorem 11.15. If v (α, ω), then Γα,ω , ∅. Moreover, v(α) > v(ω). 4 ∈P Proof. Theorem 3.28 ensures we can ﬁnd v (α, ω); let I be the current ﬂow associ- 5 ∈ P ated to v. Then ∆v(α) = 1 implies that there is some y α for which I(α, y) > 0. By 6 ∼ Lemma 11.14, 7

I(α, y) = P(γ) > 0, γ ΓX(α,y) ∈ α,ω which implies there must exist a positive term in the sum, and hence a γ Γα,ω. Since we 8 ∈ may now choose a path γ Γα,ω, the second claim follows. 9 ∈ 11.2.2. Forward-harmonic transfer operator. In this section we consider the functions 10 0 h : G C which are forward-harmonic, that is, functions which are harmonic with 11 → respect to a current I. We make the standing assumption that the network is transient; this 12 guarantees the existence of a monopole at every vertex, and the induced current will be a 13 unit flow to infinity; cf. Corollary 3.29. 14 We orient the edges by a fixed unit current flow I to infinity, as in Definition 3.2. The 15 forward-harmonicfunctions functions are fixed points of a transfer operator induced by the 16 flow which gives the value of h at one vertex as a convex combination of its values at its 17 downstream neighbours. 18 0 The mainidea is to constructa measureonthe space of pathsbeginning at vertex x G , 19 ∈ and then use this measure to define forward-harmonic functions. In fact, we are able to 20 produce all forward-harmonic functions from the class of functions which satisfy a certain 21 cocycle condition, see Definition 11.21. 22 In Theorem 11.24 we give an integral representation for the harmonic functions, and 23 in Corollary 11.25 we show that if I has a universal sink, then the only forward harmonic 24 functions are the constants. 25

Remark 11.16 (A current induces a direction on the resistance network.). If we fix a mini- 26 0 mal current I = Pd I, the flow gives a strict partial order on G and the flags in the resulting 27 poset are the induced current paths. Thus we say x y iff x is upstream from y, that 28 ≺ is, iff there exists a current path from x to y in the sense of Definition 11.12. Since I is 29 minimal, x y implies y ⊀ x and x ⊀ x. Transitivity is immediate upon considering the 30 ≺ concatenation of two finite paths. 31

Deﬁnition 11.17. Given a ﬁxed minimal current I = Pd I, we denote the set of all current paths in the resistance network (G, c) by . 1 Γ := γ = (x , x ,... .. (x , x + ) G , x x + . (11.25) I { 0 1 i i 1 ∈ i ≺ i 1} For n = 1, 2,... , we denote the set of all current paths of length n by (n) . 1 Γ := γ = (x , x ,..., x ) .. (x , x + ) G , x x + , (11.26) I { 0 1 n i i 1 ∈ i ≺ i 1} .. and denote the collection of paths starting at x by ΓI (x) := γ ΓI . x0 = x , and likewise 32 (n) { ∈ } for ΓI (x). 33 1 Here, the orientation is determined by I, andif I(x, y) = 0 forsome (x, y) G , then this 34 ∈ edge will not appear in any current path, and for all practical purposes it may be considered 35 as having been removed from G for the moment. 36 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 125

Definition 11.18. When a minimal current I = Pd I is fixed, the set of forward neighbours of x G0 is ∈ + 0 . nbr (x) := y G .. x y, x y . (11.27) I { ∈ ≺ ∼ } Definition 11.19. If v : G0 R, define the forward Laplacian of v by → ∆~ v(x) := c (v(x) v(y)) (11.28) xy − y nbrX+(x) ∈

1 A function h is forward-harmonic iﬀ ∆~ h = 0.

2 For Definitions 11.17–11.19, the dependence on I may be suppressed when context 3 precludes confusion. 0 Theorem 11.20. For I HD and x G , there is a unique measure Px defined on ΓI (x) which satisfies ∈ ∈ P [ f ] = P(n)[ f ], f . (11.29) x x ∈ An

4 Proof. We only need to check Kolmogorov’s consistency condition (11.9); see [Jor06, 5 Kol56]. For n < m, consider by assuming that f depends only on the ﬁrst n An ⊆ Am 6 edges of γ. (For brevity, we denote a function on n edges as a function on n + 1 vertices.) 7 Then

(m) (m) Px [ f ] = f (γ) dPx (γ) ZΓI (x)

(m) = f (x0, x1, x2,..., xn) dPx (γ) ZΓI (x)

(n) = f (x0, x1, x2,..., xn) dPx (γ) ZΓI (x) (n) = Px [ f ].

8 11.2.3. A boundary representation for the forward-harmonic functions. We now show that 9 the forward-harmonic functions are in bijective correspondence with the cocycles, when 10 deﬁned as follows.

Deﬁnition 11.21. A cocycle is a function f : ΓI R which is compatible with the probabilities on current paths in the sense that it satisﬁes→

cx0 x1 divI (x0) f (γ) = f (x0, x1, x2, x3 ... ) = + f (x1, x2, x3 ... ), (11.30) c (x0)I(x0, x1)

whenever γ = (x0, x1, x2, x3 ... ) ΓI is a current path as in Deﬁnition 11.17,and(x0, x1) is + ∈ the ﬁrst edge in γ. Also, c (x) := y nbr+(x) cxy is the sum of conductances of edges leaving ∈ x. If the operator m is given by multiplicationP by cxy divI (x) m(x, y) = , (11.31) c+(x)I(x, y) and σ denotes the shift operator, the cocycle condition can be rewritten f = m f σ. Using ek = (xk 1, xk) to denote the edges, this gives − ∞ f (e1, e2,... ) = m(e1) ... m(en) f (en+1, en+2,... ) = m(ek). (11.32) Yk=1 126 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Deﬁnition 11.22. Deﬁne the forward transfer operator TI induced by I by 1 (T f )(x) := c f (y). (11.33) I c+(x) xy y nbrX+(x) ∈

Lemma 11.23. If f : Γ(x) R is a cocycle and one deﬁnes h (x) := P [ f ], it follows that 1 → f x h f (x) is a ﬁxed point of the forward transfer operator TI . 2

Proof. With h f so deﬁned, we conﬂate the linear functional Px with the measure associated 3 to it via Riesz’s Theorem and compute 4

1 (T h )(x) = c P [ f ] defT, h I f c+(x) xy y f y nbrX+(x) ∈ cxy = + f (γ) dPy(γ) Py as a measure C (x) ZΓ y nbrX+(x) I (y) ∈ cxy = + f (σγ) dPx(γ) changeofvars ZΓ c (x) y nbrX+(x) I (x) ∈ I(x, y) cxy divI (x) = + f (σγ) dPx(γ) justalgebra divI (x) Γ c (x) I(x, y) y nbrX+(x) Z I (x) ∈ I(x, y) = f (γ) dPx(γ) by(11.30) divI (x) ZΓ y nbrX+(x) I (x) ∈ I(x, y) = Px( f ) Px as a functional divI (x) y nbrX+(x) ∈ I(x,y) = Px( f ) = 1 divI (x)

To justify the change of variables, note that if γ is a path starting atPx whose ﬁrst edge is 5 (x, y), then σ γ is a path starting at y. Moreover, since y is a downstream neighbour of x, 6 every path γ starting at y corresponds to exactly one path starting at x, namely, ((x, y), γ). 7 8 Theorem 11.24. The forward-harmonic functions are in bijective correspondence with the cocycles. More precisely, if f is a cocycle, then

h f (x) := Px[ f ] (11.34) is harmonic. Conversely, if h is harmonic, then

fh(γ) := limn h(Xn(γ)), γ Γx (11.35) →∞ ∈ th is a cocycle, where Xn(γ) is the n vertex from x along the path γ. 9 + Proof. ( ) Let f be a cocycle and deﬁne h asin(11.34) with C (x) as in Deﬁnition 11.22, 10 ⇒ f compute 11

∆~ h (x) = c (P [ f ] P [ f ]) f xy x − y y nbrX+(x) ∈ = P [ f ] c c P [ f ] x xy − xy y y nbrX+(x) y nbrX+(x) ∈ ∈ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 127

= P [ f ]C+(x) c P [ f ], x − xy y y nbrX+(x) ∈ 1 which is 0 by Lemma 11.23. 2 ( ) Let h satisfy ∆~ h = 0. Observe that X is a Markov chain with transition probability ⇐ n 3 Px at x. The above computations show that h is then a ﬁxed point of TI , and hence h(Xn) 4 is a bounded martingale. By Doob’s Theorem (cf. [Doo53]), it converges pointwise Px-ae 5 on Γ and (11.35) makes fh well-deﬁned. One can see that fh is a cocycle by the same 6 arguments as in the proof of [Jor06, Thm. 2.7.1].

7 Corollary 11.25. If I has a universal sink, in other words, if all current paths γ end at 8 some common point ω, then the only forward-harmonic function is the zero function.

9 Proof. Every harmonic function comes from a cocycle, which in turn comes from a har- 10 monic function as a martingale limit, by the previous theorem. However, formula (11.35) 11 yields

fh(γ) := lim h(Xn(γ)) = h(ω), γ Γ. n →∞ ∀ ∈ 12 Thus every cocycle is constant, and hence (11.30) implies f 0. Then (11.34) gives h ≡ 13 h 0. ≡ 128 PALLEE.T.JORGENSENANDERINP.J.PEARSE

12. Examples and applications 1

12.1. Finite graphs. 2

12.1.1. Elementary examples. 3 Example 12.1. Consider a “linear” electrical resistance network consisting of several re- 1 sistors connected in series with resistances Ωi = ci− as indicated:

Ω1 Ω2 Ω3 Ω4 Ωn α = x0 / x1 / x2 / x3 / ... / xn = ω

Construct a dipole v (α, ω) on this network as follows. Let v(x ) = V be ﬁxed. Then 4 ∈ P 0 determine v(x1)via (3.6): 5

1 ∆v(x0) = (V v(x1)) = 1 = v(x1) = V Ω1, Ω1 − ⇒ − 2 1 ∆v(x1) = (v(xk 1) v(xk)) = 0 = v(x2) = V Ω1 Ω2, Ωk − − ⇒ − − Xk=1 and so forth. Three things to notice about this extremely elementary example are (i) v 6 is fixed by its value at one point and any other dipole on this graph can differ only by 7 a constant, (ii) we recover the basic fact of electrical theory that the voltage drop across 8 resistors in series is just the sum of the resistances, and (iii) all current flows are induced 9 (this is not true of more general graphs). 10

Consider the basis e0, e1,..., eN , where ek = δxk , the unit Dirac mass at k. The Laplace operator for this model{ has the matrix} 1 1 0 ... 0 −  1 2 1 ... 0   −0 1− 2 ... 0    ∆=  . − .  . (12.1)  . ..     0 ... 1 2 1   − −   0 ... 0 1 1   −   2  One may obtain a unitary representation on ℓ (ZN) by using the diagonal matrix U(ζ) = 11

2 N 2πi/(N+1) th 1 diag(1,ζ,ζ ,...,ζ ), where ζ := e is a primitive (N + 1) root of 1, so that ζ− = 12 ζ¯. It is easy to check that for any matrix M N+1(C), one has [U(ζ)MU(ζ)∗] j,k = 13 j k ∈ M ζ − [M] j,k. Then deﬁne 14

∆(ζ) :=U(ζ)∆U(ζ)∗, and see that ∆(ζ) = U(ζ)T U(ζ) . It is clear that spec ∆= spec ∆(ζ), because 15 C− ∗

∆v = λv ∆(ζ)[U(ζ)v] = λ[U(ζ)v]. ⇐⇒ Decomposethe transfer operatorinto the sum of two shifts, sothatT = M++M , where M+ 16 − has ones below the main diagonal and zeros elsewhere, and M has ones above the diagonal 17 − and zeros elsewhere. Then we have U(ζ)M+U(ζ)∗ = ζM+ and U(ζ)M U(ζ)∗ = ζ¯M and 18 − − M = M+∗ . By induction, the characteristic polynomial can be written 19 −

pn(x) = det(xI Tn) = xpn 1(x) pn 2(x), − − − − OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 129

2 3 4 2 1 with p1 = x, p2 = x 1, p3 = x 2x, p4 = x 3x + 1, and corresponding Perron- − − √ − 1 √ 2 Frobenius eigenvalues λ1 = 0, λ2 = 1, λ3 = 2, λ4 = φ = 2 (1 + 5).

spec ∆ = 0, 1 , spec ∆ = 0, 1, 3 , spec ∆ = 0, 2, 2 √2 . 2 { } 3 { } 4 { ± }

3 Example 12.2. The correspondence (α, ω) (α, ω) described in Lemma 3.16 is not P → F 4 bijective, i.e., the converse to the theorem is false, as can be seen from the following 1 5 example. Consider the following electrical network with resistances Ωi = ci− .

Ω1 α = x0 / x1

Ω2 Ω3

Ω4 x2 / x3 = ω

6 One can verify that the following gives a current ﬂow I = I on the graph for any t [0, 1]: t ∈

x0 x1 t / 1 t t − x2 / x3 1 t − χ In fact, there are many ﬂows on this network; let ϑ be the characteristic function of the cycle ϑ = (x0, x1, x3, x2) x0 / x1 O ϑ = x2 o x3

7 so that χ (e) = 1 for each e e = (x , x ), e = (x , x ), e = (x , x ), e = (x , x ) . Then ϑ ∈ { 1 0 1 2 1 2 3 2 3 4 3 0 } 8 I + εχ will be a ﬂow for any ε R. (Although this formulation seems more awkward than t ϑ ∈ 9 simply allowing t to take any value in R, it is easier to work with characteristic functions of 10 cycles when there are many cycles in the network.) However, there will be only one value 11 of t and ε for which the above ﬂow corresponds to a potential function, and that potential 12 function is the following:

Ω2+Ω4 Ω V / V 4 Ω 1 − k=1 k P

V Ω1+Ω3 Ω (Ω1+Ω3)(Ω2+Ω4) 4 Ω 2 / V 4 Ω − k=1 k − k=1 k P P 13 This is the potential function which “balances” the ﬂow around both sides of the square; 14 it can be computed as in the previous example. These ideas are given formally in Theo- 15 rem 3.26.

2πik/N 16 Example 12.3 (Finite cyclic model). In this case, let GN have vertices given by xk = e 17 for k = 1, 2,..., N, with edges connecting each vertex to its two nearest neighbours. For 18 example, when N = 9, 19 130 PALLEE.T.JORGENSENANDERINP.J.PEARSE

x2 x3 ff NNN Ö x1 ÖÖ , x4 ,, G x = x 1 9 9 0 x5 < << x8 x ss 6 [[ x7 s

In this case, using the same basis e0, e1,..., eN , as in Example 12.1, the Laplace operator for this model has the matrix { } 2 1 0 ... 0 1 − −  1 2 1 ... 0 0   −0 1− 2 ... 0 0    ∆=  . − .  . (12.2)  . ..     0 0 0 ... 2 1   −   1 0 0 ... 1 2   − −    The Fourier transform is a spectral transform of ∆ that shows it to be unitarily equivalent 2 F 2 2πk to multiplication by 4 sin N . 3 Lemma 12.4. For S := ∗∆ and v , one has 4 F F ∈HE

1 k (Sv)(z) = (2 z z− )v(z), z = α , k = 0, 1, 2,..., N. − −

k Proof. Denote vk = v(k) and consider the Fourier transform : vk k Z vkz : 5 F { } 7→ ∈ N P k (∆v)(z) = (2vk vk 1 vk+1)z F − − − kXZ ∈ N k k k = 2 vkz vk 1z vk+1z − − − kXZ kXZ kXZ ∈ N ∈ N ∈ N 1 k = (2 z z− ) v z . − − k kXZ ∈ N 1 1 This shows (∆v)(z) = (2 z z− ) (v)(z), so that S is multiplication by (2 z z− ). 6 F − − F − −

In this case, ∆= I T with = 2I, so that and T commute. Additionally, T is the sum 7 − C C 1 2 0 of two shifts and so corresponds to multiplication by z + z− = 2cos θ on ℓ (GN ), where 8 2πk 2πk θ = N . Consequently, spec TN = 2cos N and the Perron-Frobenius eigenvalue of TN is 9 λPF = 2, which occurs for k = 0 and has eigenfunction vPF = [1, 1,..., 1]. Observe that ∆N 10 2 N 1 commutes with the cyclic shift. The eigenfunctions of the shift are v = [1, λ, λ , . . . , λ − ], 11 0 where λ G , and hence these are the other eigenfunctions of T . 12 ∈ N Proposition 12.5. The spectrum of ∆N is given by 13

2πk .. spec ∆N = 2 1 cos N . k ZN . { − ∈ } OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 131

v(x)

x0 xk xN

Figure 5. The solution v as represented on R.

1 Proof. Let U be the cyclic shift in the positive direction, i.e., U has the matrix

0 1  1 0   1 0  U =   .  . .   .. ..     1 0      2 The U and ∆ commute. For n Z , the Fourier transform is N ∈ n

1 n n vˆ(n) = z− v(z), so x v(n) = z vˆ(n). N k 7→ zXG0 nXZN ∈ N ∈

3 For k 1, 2,..., N 1 , one can ﬁnd v (k, 0) as follows: “ground” the graph with ∈ { − } ∈ P 4 v(0) = 0 and consider

... / x1 / x0 = xN o xN 1 o ... o xk / ... α α 1 α − 1 α 1 α α − − − 5 The cycle condition (the net drop of voltage around any closed cycle must be 0) yields N k k(N k) 6 v(k) v(0) = kα = (N k)(1 α), and hence α = − . This gives v(k) = − and we − − − N N 7 have

k N k − = 1, j = 0 − N − N −  k k = , < <  N N 0 0 j k ∆v( j) = δk δ0 =  − −  k + N k = 1, j = k  N N−  N k N k  − − = 0, k < j N 1.  N − N ≤ −  8 This additionally shows that if the shortest path from x to y has length k, then R(x, y) = N k 9 k N− < k. Of course, there is an easier way to get R(x, y). Since there are only two paths 10 γ1, γ2 from x to y, the laws for resistors connected in serial and parallel indicate that the 1 1 1 11 entire network can be replaced by a single edge (x, y) with resistance (Ω(γ1)− +Ω(γ1)− )− . 12 In the case of constant resistance Ω 1, this becomes ≡

1 1 1 1 1 1 1 − (N k) + k − N k (Ω(γ1)− +Ω(γ1)− )− = + = − = k − . k N k ! k(N k) ! N − − 132 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Example 12.6. Next, it is illuminating to see how things change when edges are removed 1 from the network. Consider the following example, where c = 1 and the currents are 2 as indicated, and the second network is obtained from the ﬁrst by deleting two edges, as 3 indicated: 4

5 1 11 2 α = x0 / x1 α = x0 / x1 vv vv vv 2 v 11 v{ v 1 1 4 5 2 2 x2 11 11 HH 2 HH 11 HH 1 HH 6 2 H# 11 x3 / x4 = ω x3 / x4 = ω

I1 I0

10 The dissipations are R0(α, ω) = D(I0) = 11 and R1(α, ω) = D(I1) = 1. The set of paths 5 from α to ω that don’t pass through the deleted edges contains only γ1 = (α, x1, ω) and 6 γ2 = (α, x3, ω). Then 7

4 5 9 ε = P(γ) = P(γ ) + P(γ ) = + = , 1 2 11 11 11 γXQ W ∈ \ and Powers’ inequality gives 8

2 R (α, ω) R (α, ω) ε− R (α, ω) 0 ≤ 1 ≤ 0 10 121 10 110 1 = . 11 ≤ ≤ 81 · 11 81

12.2. Inﬁnite graphs. 9

Example 12.7. Deﬁne the projections 10

Bn = χ , B⊥ = χ = χ . (12.3) Gn n G Gn ∁ \ Gn Let us denote the edge boundary between Gn and Gn+1 by 1 . 0 0 0 Edge := e = (x, y) G .. y G , x G G . (12.4) n { ∈ ∈ n ∈ \ n} We now consider the behavior of Bn⊥∆Bn , where the norm is with respect to operators 11 2 2 k k ℓ (c) ℓ (c). 12 → 2 .. Lemma 12.8. For Cn := sup y x cxy . (x, y) Edgen , { ∼ ∈ } P B⊥∆B C . (12.5) k n nk≤ n

2 0 ∁ Proof. Let v ℓ (c) and x (G ) . Then 13 ∈ ∈ n

χ χ ✘✘✘ χ (Bn⊥∆Bnv)(x) = ∁(x) cxy ✘ v(x) v(y) G Gn − Gn Xy x ∼ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 133

= c v(y), x (G0)∁. − xy ∈ n (x,y)XEdge ∈ n 0 ∁ 1 Now summing over all x (G ) , and hence over all edges in Edge , ∈ n n

2 B ∆B v 2 = c v(y) v(y) 2 c 2 C2 v 2, n⊥ n c xy xy n c k k ≤ 0 | | | | ≤ k k (x,y)XEdgen X X0 ∁ y Gn x (Gn) ∈ ∈ ∈y x ∼

2 and hence B ∆B C . k n⊥ nkc ≤ n Proposition 12.9. If the estimate 2 B⊥∆B = O(n ), as n , (12.6) k n nk →∞ 3 is satisﬁed, then ∆ has no eigenvectors at . ∞ 4 Proof. Since ∆ is semibounded, this follows from [Jør78]. (The reader may ﬁnd informa- 5 tion in [Jør76]and [Jør77] regarding the general Hermitian case.)

6 Example 12.10 (One-sided ladder model). The elementary ladder models help explain 7 the effects of adding to the network, when the new portions added to the graph are some- 8 what “peripheral” to the original graph. Also, they provide an easy example of an infinite 9 network in which a shortest path (see Remark 5.45) may not exist, and then illustrate a 10 technique for understanding what happens on a finite subgraph of interest which is embed- 11 ded in an infinite graph. Consider a graph which appears as a sideways ladder with n rungs extending to the right: α = x0 / x1 / x2 / x3 / ... xn (12.7)

ω = y0 o y1 o y2 o y3 o ... yn The one-sided ladder model furnishes a situation where no shortest path exists, as mentioned in Remark 5.45, if the resistances are deﬁned as follows:

α = x0 x1 x2 x3 ... xn (12.8) 1 1 1 1 4 16 64 4n 1 1 1 1 1 4 16 64 4n 1 1 1 1 4 16 64 4n ω = y0 y1 y2 y3 ... yn

12 To find a path from α to ω with minimal distance, one is led to consider paths stretching 13 ever further off towards infinity. It is easy to see that the shortest path metric gives

1 1 1 1 2 distγ(α, ω) = lim 2 + + + + = , " 4 16 ··· 4n ! 4n # 3 2 14 but there is no γ Γα,ω for which P(γ) = 3 . Note that the Powers bound (9.14) is violated, n ∈n n+1 15 as µ(x ) = 4 + 4 + 4 . n →∞ 134 PALLEE.T.JORGENSENANDERINP.J.PEARSE

xo xo

Figure 6. The homogeneous tree of degree 3 (left) and the binary tree from symbolic dynamics (right). The root of the tree is labeled xo. If the grey branch is pruned from the homogeneous tree, the two become isomorphic.

13. Infinite trees 1

The n-ary trees play an important role in symbolic dynamics, and they support a rich 2 family of nontrivial harmonic functions of finite energy. These graphs are essentially ho- 3 mogeneous trees; the only difference is that the homogeneous tree has one more branch 4 at the root, as can be seen from Figure 6. We use the latter examples as they are simpler 5 yet still sufficient for our purposes, and because of our related interest in symbolic dynam- 6 ics. However, almost all remarks extend to the homogeneous trees without effort; these 7 examples are well-studied because of their close relationship with group theory (especially 8 free groups). Also, they provide an excellent testbed for studying the effects of varying 9 c, and for illustrating several of our theorems. A network whose underlying graph is a 10 homogeneous tree always allows for the construction of a nontrivial harmonic function. In 11 particular, in is not dense in by Lemma 4.47 that these are equivalent. 12 F HE Remark 13.1. If the origin were removed from the binary tree, we adopt the convention 13 that vertices in one component are “positive” and indices in the other are “negative”. If the 14 vertices are indexed with binary numbers (using the empty string ∅ to denote the origin 15 o = x∅), then indices beginning with 1 are positive and indices beginning with 0 are 16 negative. 17

Example 13.2 (The reproducing kernel on the tree). Let ( , 1) be the binary tree network 18 T in Figure 6 with constant conductances. Figure 7 depicts the embedded image of a vertex 19 v , as well as its decomposition in terms of in and arm. We have chosen x to be 20 x F H adjacent to the origin o; the binary label of this vertex would be x1. 21 In Figure 7, numbers indicate the value of the function at that vertex; artistic liberties 22 have been taken. If vertices s and t are the same distance from o, then f (s) = f (t) 23 | x | | x | and similarly for hx. Note that hx provides an example of a nonconstant harmonic func- 24 2 tion in . Another key point is that hx < ℓ , see Corollary 4.65. It is easy to see that 25 HE 1 1 limz hx(z) = 2 2 , whence hx is bounded. 26 →±∞ ± (k) For fx = P invx of Figure 7, the illustration of fx in Figure 8 is the projection of vx (or F .. fx) to span δx . x Gk , where Gk consists of all vertices within k steps of o. The lines at the right side{ of each∈ ﬁgure} just indicate that the function is constant on the remainder of (k) the graph (at value 0 or 1); in particular, note that fx (y) = 0 for every vertex y which is at least k + 1 steps from the origin. Also, observe that δ δ ∆ f (k) = δ δ + s t , x x − y 2k+2 2 − 2k+2 2 s bdXG+ t bdXG ∈ k+1 − ∈ k−+1 − + where bd Gk+1 is the subset of bd Gk+1 that lies on the positive branch, etc. It is interesting 27 + + to note that if one were to identify all the vertices of bd G + = bd G bd G , then 28 k 1 k+1 ∪ k+1 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 135

1 1 1 x vx F R (x,o ) = (vx) = 1 0 0 0 0 o

1 1 1 4 8 x 16

fx W 3 R (x,o ) = (fx) = 4 1 1 1 1 8 16 2 4 o

3 4 7 15 16 1 x 8 2 H 1 hx o R (x,o ) = (hx) = 4 1 4 1 1 8 16

Figure 7. The reproducing kernel on the tree with c = 1. For a vertex x which is adjacent to the origin o, this ﬁgure illustrates the elements vx, fx = P invx, and hx = P armvx; see F H Example 13.2.

(k) (k) (k) 1 f would become harmonic at this new vertex. Observe also that h = v f is its x x x − x 2 orthogonal complement and is harmonic everywhere except on bd Gk+1.

3 Example 13.3 (A function of finite energy which is not approximable by in). We con- (k) F (k) 4 tinue to refer to in Figure 8. Since fx in and it is easy to see that fx fx 0 ∈ F k − kE → 5 and that ∆ fx = δx δy, this approximation verifies that fx = P invx. It also shows that − 1 F 6 min f in vx f = . ∈F k − kE 4 7 Example 13.4 (A monopole which is not a “dipole at infinity”). Let ( , 1) be the binary T 8 tree network in Figure 6 with constant conductances. Let x be the number of edges in the | | 9 path connecting x to o. Define a function

1 w (x) = 1 , (13.1) o x − 2| | so that essentially w = 2 h 1 for h of Example 13.2. It is easy to check that ∆w = δ o | x − 2 | x o − o so that wo is a monopole at the root o. To see that wo , ∈HE 2 ∞ n 1 1 (wo) = 2 2 1 1 E − 2n ! − − 2n+1 !! Xn=0 2 ∞ 1 = 2 2n 2n+1 ! Xn=0 1 ∞ 1 = 2 2n Xn=0 = 1. 136 PALLEE.T.JORGENSENANDERINP.J.PEARSE

1 1 1 1 1

(k) x vx

0 0 0 0 0 0 0

2k − 1 2k-j − 1 2k+2 − 2 x 2k+2 − 2 j = 0,1, ... , k 2k-1 − 1 (k) 2k+2 − 2 1 1 fx

k-j − k − 2 1 2 1 2k+2 − 2 j = 0,1, ... , k 1 2k+2 − 2 2

1 1 − 2k − 1 1 2k+2 − 2 k-j − k-1 − − 2 1 (k) − 2 1 1 2k+2 − 2 j = 0,1, ... , k hx x 1 2k+2 − 2 1 2

k-j − 2 1 j = 0,1, ... , k 2k − 1 2k+2 − 2 2k+2 − 2 0 0 y

Figure 8. Approximants to the reproducing kernel on the tree with c = 1; see Example 13.2.

However, w is not a “dipole at inﬁnity” in the sense that there is no sequence x 1 o { n} of distinct and successively adjacent vertices for which vxn converges to wo (this is in 2 d { F } contrast to the integer lattices Z , d 3). Observe that R (x, y) coincides with shortest- 3 ≥ path distance on this network (as it does on any tree). If x is a sequence tending to 4 { n} ∞ (i.e., for any N, there is an n such that xn is more than n steps from o for all n N), then 5 F ≥ (v ) = R (x , o) = n, so that w is not a limit of a sequence of dipoles. 6 E xn n o Of course, since vx is dense in , wo is the limit of linear combinations of dipoles. 7 { } . HE In fact, let bd G = x .. R(x, o) = k as before. Then 8 k { ∈ T }

vz wo(x) = lim . k 2k →∞ z Xbd G ∈ k

Example 13.5 (A function with nonvanishingboundarysum). In Theorem 4.41, we showed 9

u, v = lim u(x)∆v(x) + lim u(x) ∂v (x). h iE k k ∂Ò →∞ x Xint G →∞ x Xbd G ∈ k ∈ k 1 .. Let wo(x) = 1 x be the monopole from Example 13.4. With Gk := x . x k , we have − 2| | { | |≤ }

∂wo 1 1 1 . (x) = 1 1 = , for x bd Gk = x .. x = k . ∂Ò k k 1 k − 2 ! − − 2 − ! 2 ∈ { | | } OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 137

Since ∆w = δ , we have w (x)∆w (x) = w(o), for each k, and the energy of w is − o Gk o o − o P 1 1 1 (wo) = wo, wo = 1 + lim 1 = 1. E h iE − − 20 ! k − 2k ! 2k →∞ x Xbd G ∈ k 1 1 For h , the harmonic function with (h ) = in Example 13.2, this becomes x E x 4

∂hx 1 (hx, hx) = lim hx(x) (x) = . E k ∂Ò 4 →∞ x Xbd G ∈ k k 1 2 In fact, one can obtain this by computing the boundary term directly: each of the 2 − + k 1 3 vertices in bd Gk is connected by a single edge to Gk, and similarly for the 2 − vertices in 4 bdGk−, so

k+1 ∂hx k 1 2 1 1 k 1 1 1 1 1 hx(y) (y) = 2 − − + 2 − − = 1 . ∂Ò 2k+1 · 2k+1 2k+1 · 2k+1 4 − 2k ! y Xbd G ∈ k

5 Example 13.6 (The tree supports many nontrivial harmonic functions). We can use hx 6 of Example 13.2 to describe an infinite forest of mutually orthogonal harmonic functions 7 on the binary tree. Let z G be represented by a finite binary sequence, as discussed in ∈ 8 Remark 13.1. Define a morphism (cf. Definition 6.5) ϕ : G G by prepending, i.e., z → 9 ϕz(x) = zx. This has the effect of “rigidly” translating the the tree so that the image lies 10 on the subtree with root z. Then h := h ϕ is harmonic and is supported only on the z x ◦ z 11 subtree with root z. The supports of h and h intersect if and only if Im(ϕ ) Im(ϕ ). z1 z2 zi ⊆ z j 12 For concreteness, suppose it is Im(ϕ ) Im(ϕ ). If they are equal, it is because z = z z1 ⊆ z2 1 2 13 and we don’t care. Otherwise, compute the dissipation of the induced currents

dh , dh = 1 Ω(x, y)dh (x, y), dh (x, y). h z1 z2 iD 2 z1 z2 (x,y)Xϕ (G1) ∈ z1 , 14 Note that dhz2 (x, y) always has the same sign on the subtree with root z1 o, but dhz1 (x, y) 15 appears in the dissipation sum positively signed with the same multiplicity as it appears 16 negatively signed. Consequently, all terms cancel and 0 = dhz , dhz D = hz , hz shows h 1 2 i h 1 2 iE 17 h h . z1 ⊥ z2 18 Example 13.7 (Haar wavelets and cocycles). Example 13.6 can be heuristically described 19 in terms of Haar wavelets. Consider the boundary of the tree as a copy of the unit interval 20 with hx as the basic Haar mother wavelet; via the “shadow” cast by limn hx(xn) = 1. →±∞ ± 21 Then hz is a Haar wavelet localized to the subinterval of the support of its shadow, etc. 22 Of course, this heuristic is a bit misleading, since the boundary is actually isomorphic to N 23 0, 1 with its natural cylinder-set topology. { } 24 Example 13.8 (Why the harmonic functions may not be in the domain of ∆). We have not 25 been able to construct an example in which we can prove that arm is not contained in H 26 dom ∆, but we do have the following suggestive example, motivated by Lemma 8.5. Asin 27 Deﬁnition 4.32, let V := span vx x G0 and let ∆ = ∆ denote the closure of the Laplacian { } ∈ V 28 when taken to have the dense domain V. Let h arm. If h were an element of dom ∆ , ∈H V 29 then by (B.7), we would have a sequence

clo .. dom S := u . lim u un = lim v Sun = 0 (13.2) n H n H { →∞ k − k →∞ k − k } 138 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Again, think of the vertices of the tree as begin labelled by a word on 0, 1 , that is, a finite 1 { } binary string. If x = w, then w is the length of the word and corresponds to the number of 2 | | edges between x and o (i.e., shortest path distance to the root). Using w1 to denote the first 3 coordinate of w, define the function 4

1 h := v v . (13.3) n 2n  w − w Xw =n wX1=1 wX1=0  | |     n Since hn is a (ﬁnite) sum of all the dipoles at distance n from o, with half weighted by 2− 5 n and the other half weighted by 2− , it is clear that hn span vx . One can check that for 6 . − ∈ { } G = w .. w n , 7 n { | |≤ }

h(w), w n, = h(w) n | |≤  2− , else, ±  whence it is immediate that limn hn  h = 0. One can also check that 8 →∞ k − kE

1 ∆h = δ δ n 2n  w − w Xw =n wX1=1 wX1=0  | |     since the positive and negative weights of δ cancel out. If w, w but w = w , then they 9 0 ′ | | | ′| cannot be neighbours, and hence δ and δ are orthogonal with respect to . It is then 10 w w′ E easy to compute 11

1 1 ∆ ∆ 2 = ∆ 2 = 2 = n = , lim hn h lim hn lim n δw lim n 2 3 3 0. n k − kE n k kE n 2 k kE n 2 · · →∞ →∞ →∞ Xw =n →∞ | |

x Example 13.9. On the binary tree with c = 1, the monopole w(x) = 2−| | can be written as 12 v + ∆v for v dom ∆ : 13 ∈ V

n+1 x 1 v(x) := 2−| | √2 1 + . − √2 We leave it to the reader to check that this v satisﬁes the above equation and also 14

2(46 √2 65) 4(99 70 √2) 2 (v) = − , (∆v) = − , and v∆v = ( 23 + 17 √2). E (2 √2)2( √2 1) E (2 √2)4( √2 1) 7 − − − − − XG0

This illustrates Lemma 4.52. 15

Example 13.10 (An unbounded harmonic function of ﬁnite energy). Figure 9 is a sketch 16 of an unbounded harmonic function of ﬁnite energy on the binary tree with c = 1. To 17 construct it, pick one ray from o to , and let 18 ∞

x | | 1 h(x) = k Xk=1 for x along this ray. Then if x is in the ray and y x, ﬁx h(y) so that h is harmonic at x 19 1 ∼ (i.e., h(y) = h(x) + x ( x +1) ), and deﬁne h along the rest of this branch by 20 | | | | OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 139

n 1 1 25 + 1 Σ + 11 1 n k=1 k n+1 + 4 12 3 + 1 6 3 + 1 2 1 n(n+1) 1 +12 + 2 + 1 1 6

+1 1 + 2 0

−1

Figure 9. An unbounded harmonic function of ﬁnite energy. See Example 13.10.

h(y) h(x) h(z) = . −y z 2| − | x 1 If w denotes the monopole at o deﬁned by w(x) = 2−| | as discussed in Example 13.9 and 2 previously, then we are essentially attaching a scaled copy of w to each neighbour of the 3 chosen ray. See Figure 9. 4 It is clear that h(x) logarithmically along the chosen ray; the energy coming from →∞ 5 h(x) on this ray is

2 ∞ 1 π2 (h) ray = = . E | n! 6 Xn=1 6 The energy from each branch incident upon the ray is

1 2 1 1 1 (h) branch(n) = + (w) = + . E | n(n + 1)! n(n + 1)E n2(n + 1)2 n(n + 1) π2 7 Summing up, (h) = (h) + ∞ (h) = 2. We leave it to the reader to E E |ray n=0 E |branch(n) 2 − 8 check that h is harmonic. P

x 9 Example 13.11. On the binary tree with c = 1, the function uξ(x) = ξ−| | has energy

2 (1 ξ) 1 (uξ) = 2 − < , for ξ [0, ). E 1 2ξ2 ∞ ∈ √2 − 10 Moreover, the Discrete Gauss-Green formula applies to this example with

2ξ 1 uξ∆uξ = (1 ξ) 2 + (2ξ 1) 2 , for ξ [0, ). − − 1 2ξ ∈ √2 XG0 − 1 1 11 However, for ξ [0, 2 ), one also has G0 ∆uξ = (1 ξ) 2 + (2ξ 1) 1 2ξ . Thus, for 1 1 ∈ − − − 12 <ξ< , one has ∆uξ = , even thoughP (uξ) < and uξ∆uξ < . 2 √2 ∞ E ∞ ∞ P P 140 PALLEE.T.JORGENSENANDERINP.J.PEARSE

14. Lattice networks 1 d d The integer lattices Z R are some of the most widely-studied infinite graphs and 2 ⊆ have an extensive literature; see [DS84, Tel06], for example. We begin with some results 3 for the simple lattices; in 14.3 we consider the case when c is nonconstant. Because the 4 § case when c = 1 is amenable to Fourier analysis, we are able to compute many explicit 5 formulas for many expressions, including v and R(x, y). For d 3, we even compute 6 x ≥ R(x, ) = limy R(x, y) in Theorem 14.9 and give a formula for the monopole w. There 7 ∞ →∞ is a small amount of overlaphere with the results of [Soa94, V.2], where the focus is more 8 § on solving Poisson’s equation ∆u = f . In 15.3 we employ our formulas in the refinement 9 § of an application to the isotropic Heisenberg model of ferromagnetism. 10 In the present context, we may choose canonical representatives when working point- 11 wise: given u , we use the representative which tends to 0 at infinity. We take this as 12 ∈HE a standing assumption for this section, as it allows us to use the Fourier transform without 13 2 ambiguity or unnecessary technical details. To see that this is justified, note that ℓ (c) is 14 2 dense in in by Theorem 9.23, and hence dense in for these examples, as it is well- 15 F HE known that there are no nonconstant harmonic functions of finite energy on the integer 16 lattices (we provide a proof in Theorem 14.17 for completeness). Clearly, c = 1 implies 17 2 2 2 that ℓ (c) = ℓ (1) and that all elements of ℓ (c) vanish at . 18 ∞ Remark 14.1. As mentioned in Remark 1.1, one of the applications of the present inves- 19 tigation is to numerical analysis. Discretization of the real line amounts to considering a 20 1 . graph which is a scaled copy of the integers Gǫ = (ǫZ, 1) where ǫZ = ǫn .. n Z . After 21 ǫ { ∈ } finding the solution to a given problem, as a function of the parameter ǫ, one lets ǫ 0. 22 → Let xn denote the vertex at ǫn. 23

... x 2 x 1 x0 x1 x2 x3 ... 1/ǫ − 1/ǫ − 1/ǫ 1/ǫ 1/ǫ 1/ǫ 1/ǫ

The diﬀerence operator D acts on a function on this network by D f (xn) := f (xn) f (xn+1). 24 2 − The adjoint of D with respect to ℓ is D∗ f (xn) = f (xn) f (xn 1). Then D∗D = ∆. 25 − − 14.1. Simple lattice networks. 26

d Example 14.2 (Simple integer lattices). The lattice network (Z , c), with an edge between 27 any two vertices which are one unit apart is called simple or translation-invariant when 28 c = 1. The term “simple” originates in the literature on random walks. 29

One may compute the energy kernel directly using (5.1), that is, by ﬁnding a solution vx 30 to ∆v = δ δ as depicted in Figure 10. Then R(o, x) = v (x) v (o) = x 0 = x, which 31 x − 0 x − x − is unbounded as x . This also provides an example of a function vx for which 32 2 → ∞ ∈ HE vx < ℓ (c), as discussed in 9.2.2 and elsewhere. 33 § d In Lemma 14.4 we obtain a general formula for vx on (Z , 1). Figure 12 of Exam- 34 ple 14.16 shows how this compares to Figure 10. 35 = = χ 36 To see how the function v vx1 [1, ) may be approximated by elements of in, ∞ F deﬁne 37

k 1 n , 1 k n, un(xk) = − ≤ ≤ (14.1) 0, else.   2Technically, the embedded image of J : ℓ2(c) in is dense in in; see Deﬁnition 9.22. → F F OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 141

0 x

Figure 10. The function vx, a solution to ∆v = δx δ0 in (Z, 1). −

1 The reader can verify that u minimizes (v u)overtheset of u for which spt(u) [1, n 1] n E − ⊆ − 2 and that

n 1 1 2 − k k 1 2 2 (v u ) = (1 (1 )) + (1 ) (1 − ) + (1 (1 0)) E − n − − n − n − − n − − Xk=1 1 n 1 n = + − →∞ 0. n2 n2 −−−−−−→

3 The fact that v [ in] but limk v(xk) , limk v(x k) reﬂects that (Z, 1) has two graph ∈ F →∞ →∞ − 4 ends, unlike the other integer lattices; cf. [PW90]. Therefore, (Z, 1) also provides an ex- 5 ample of a network with more than one end which does not support nontrivial harmonic 6 functions. 7 An explicit formula is given for the potential conﬁguration functions v on the simple { x} 8 d-dimensional lattice in Lemma 14.4. By combining this formula with the dipole formula- 9 tion of resistance distance

R(x, y) = v(x) v(y), for v = v v , from (5.1), − x − y 10 we are able to compute an explicit formula for resistance distance on the translation- d 11 invariant lattice network Z in Theorem 14.7. Results in this section exploit the Fourier d d 12 duality Z T ;[Rud62] is a good reference. We are using Pontryagin duality of locally ≃ d 13 compact abelian groups; as an additive group of rank d, the discrete lattice Z is the dual d d d 14 of the d-torus T . Conversely, T is the compact group of unitary characters on Z (the op- d 15 eration in T is complex multiplication). This duality is the basis for our Fourier analysis d 16 in this context. For convenience, we view T as a d-cube, i.e., the Cartesian product of d d 17 period intervals of length 2π. In this form, the group operation in T is written additively d d 18 and the Haar measure on T is normalized with the familiar factor of (2π)− . 19 In [Pól21], Pólya proved that the random walk on the simple integer lattice is transient 20 if and only if d 3; see [DS84] for a nice introduction and a proof using electrical re- ≥ 21 sistance networks. In the present context, this can be reformulated as the statement that d 22 there exist monopoles on Z if and only if d 3. We offer a new characterization of this ≥ 23 dichotomy, which we recover in Theorem 14.5 via a new (and completely constructive) 24 proof. In Remark 14.21 we show that in the infinite integer lattices, functions in may HE 25 be approximated by functions of finite support. 26 Sometimes Pólya’s result is restated: the resistance to infinity is finite if and only if 27 d 3. There is an ambiguity in this statement which is specific to the nature of resistance ≥ 28 metric. One interpretation is that one can construct a unit flow to infinity; this is the ter- 29 minology of [DS84] for a current with div(I) = δx and it is clear that this is the induced 30 current of a monopole. Probabilistically, this definition may be rephrased: for a random 0 31 walk beginning at x G , the expected hitting time of the sphere of (shortest-path) radius ∈ 142 PALLEE.T.JORGENSENANDERINP.J.PEARSE n remains bounded as n . This approach interprets “infinity” as the “set of all points 1 → ∞ at infinity”. 2 d By contrast, we prove a much stronger result for the simple lattice networks Z in The- 3 0 orem 14.9, where we show limy R(x, y) is bounded as y , for any x G . To see 4 →∞ → ∞ ∈ the strength of this result, note that the simple (c = 1) homogeneous trees of degree d 3 5 0 ≥ have finite resistance to infinity, even though limy R(x, y) = for any x G , and any 6 →∞ ∞ ∈ choice of y . This is discussed further in Example 13.2 of the previous section. The 7 → ∞ heuristic explanation is that the resistance distance between two places is much smaller 8 when there is high connectivity between them; there is much more connectivity between x 9 and the “set of all points at infinity” than between x and a single “point at infinity”. 10 In the next result, we obtain the Fourier transform of the Laplacian; we recently noticed 11 that this corresponds almost identically to the inverse Fourier transform H of the “potential 12 kernel” of [Soa94, V.2]. 13 § d Lemma 14.3. On the electrical resistance network (Z , 1), the spectral (Fourier) transform 14 d 2 tk of ∆ is multiplication by S (t) = S (t1,..., td) = 4 k=1 sin 2 . 15 d P 2 d Proof. Each point x in the lattice Z has 2d neighbours, so we need to find the L (T ) 16 Fourier representation of 17

d ∆v(x) = (2dI T)v(x) = 2dv(x) v(x ,..., x 1,... x ). (14.2) − − 1 k ± d Xk=1 d d th Here, t = (t1,..., td) T and x = (x1,..., xd) Z . The k entry of t can be written 18 ∈ ∈ th

tk = t εk where εk = [0,..., 0, 1, 0,..., 0] has the 1 in the k slot. Then moving one step 19

i i · i x t t x t in the lattice by x x + ε corresponds to e · e k e · under the Fourier transform, and 20 7→ k 7→

d i it t ∆v(t) = 2d (e k + e− k ) vˆ(t)  −   Xk=1  c    d  = 2 (1 cos(t )) vˆ(t)  − k  Xk=1    d t  = 4 sin2 k vˆ(t). 2 Xk=1

d Lemma 14.4. Let vx x Zd be the potential conﬁguration on the integer lattice Z with 21 { d } ∈ c = 1. Then for y Z , 22 ∈ 1 cos((x y) t) cos(y t) v (y) = dt. (14.3) x d − · − · (2π) ZTd S (t)

Proof. Under the Fourier transform, Lemma 14.3 indicates that the equation ∆vx = δx δo 23

i x t − becomes S (t)ˆv = e 1, whence 24 x · −

1 ei x t 1 = iy t · . vx(y) d e− · − dt (14.4) (2π) ZTd S (t)

Since we may assume vx is R-valued, the result follows. 25 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 143

1 The following result is well-known in the literature (cf. [DS84,NW59], e.g.), but usually 2 stated in terms of the current flow induced by the monopole. d 3 Theorem 14.5. The network (Z , 1) has a monopole 1 cos(x t) = w(x) d · dt (14.5) −(2π) ZTd S (t) 4 if and only if d 3, in which case the monopole is unique. ≥ 5 Proof. As in the proof of Lemma 14.4, we use the Fourier transform to solve ∆w = δo cos t 1 1 d − 6 by converting it into S (t)ˆw(t) = 1. This gives (14.5), and since L (T ) for − S (t) ≈ S (t) ∈ 7 t 0, the integral is finite iff d 3 by the same argument as in the proof of Theorem 14.9; ≈ ≥ 8 see (14.11). It remains to check that w . Note that it follows from Theorem 14.17 that ∈HE 9 the boundary term of (4.18) vanishes, and hence we may compute the energy for d 3 via ≥

1/2 2 1 w = ∆ w 2 = S (t)ˆw(t) dt = dt < . (14.6) k kE k k ZTd ZTd S (t) ∞ 10 Uniqueness is an immediate corollary of the previous theorem; if w′ were another, then 11 ∆(w w ) = δ δ = 0 and w w is constant by Theorem 14.17. − ′ o − o − ′ 12 Remark 14.6. Upon comparing (14.5)to(14.3), it is easy to see why all networks support 13 ﬁnite-energy dipoles: the numerator in the integral for the monopole is of the order of 1 14 for t 0, while the corresponding numerator for the dipole is o(t) for t 0. ≈ ≈ d 15 Theorem 14.7. Resistance distance on the integer lattice (Z , 1) is given by

1 sin2((x y) t ) R(x, y) = − · 2 dt. (14.7) d t (2π) ZTd d 2 k k=1 sin 2 P d 16 Proof. Let vx x Zd be the potential conﬁguration on Z . Then vx vy (x, y),soby(5.2) { } ∈ − ∈P 17 wemayuse(5.9)to compute the resistance distance via R(x, y) = vx(x)+vy(y) vx(y) vy(x),

F W d i x t − − 18 since R = R on Z . Using ex = e · , substitute in the terms from (14.4) of Lemma 14.4:

1 ex(ex 1) + ey(ey 1) ex(ey 1) ey(ex 1) , = − − − − − − R(x y) d dt (2π) ZTd S (t) 1 1 e✓x + 1 ✓e✓y ey x + e✓x ey x + ✓e✓y = − − − − − − d dt (2π) ZTd S (t) 1 2 2 cos((x y) t) = , d − − · dt (14.8) (2π) ZTd S (t) 19 and the formula follows by the half-angle identity. 1 d 20 Corollary 14.8. If y x, then R(x, y) = on (Z , 1). ∼ d Proof. The symmetry of (Zd, 1) indicates that the distance from x to its neighbour will not depend on which of the 2d neighbours is chosen. For k = 1, 2,..., d, let yk be a neighbour of x in the kth direction. Then (14.3) gives 1 sin2( tk ) R(x, y ) = 2 dt. (14.9) k d t (2π) ZTd d 2 k k=1 sin 2 d P 21 Thus, k=1 R(x, yk) = 1 and R(x, yk) = R(x, y j) gives the result. P 144 PALLEE.T.JORGENSENANDERINP.J.PEARSE

d Theorem 14.9. The metric space ((Z , 1), R) is boundedif and only if d 3, in which case 1 ≥ 2 1 , = . lim R(x y) d dt ford 3 (14.10) y π d S (t) →∞ (2 ) ZT ≥

Proof. This result hinges upon the convergence properties of the integrand for R(x, y) as 2 1 d computed in Lemma 14.7. In particular, to see that 1/S (t) L (T ) one only needs to 3 ∈ check for t 0, where 4 ≈ 1 1 = O , as t 0. S (t) 2  tk  →   P  2 Switching to spherical coordinates, 1/S (ρ) = O ρ− , as ρ 0, and one requires 5 → 1 2 d 1 ρ− ρ − dS d 1 < , (14.11) Z0 | | − ∞ where dS d 1 is the usual (d 1)-dimensional spherical measure. Of course, (14.11) holds 6 − − precisely when 2+d 1 > 1, i.e., when d > 2. Similarly, the function cos((x y) t)/S (t) 7 1 d − − − − · ∈ L (T ) iff d 3. Therefore, (14.8) gives 8 ≥ 2 1 2 cos((x y) t) , = , . R(x y) d dt d − · dt for d 3 (2π) ZTd S (t) − (2π) ZTd S (t) ≥ Now replace y with a sequence of vertices tending to infinity as in Definition 4.54. By the 9 Riemann-Lebesgue lemma, the second integral vanishes and for any such y , we have 10 0 →∞ (14.10). Note that this is independentof x G , as one would expect from the translational 11 ∈ invariance of the network, since c = 1. 12

Deﬁnition 14.10. Denote R := limy R(o, y), as it is clear from the previous result that 13 ∞ →∞ the limit does not depend on the choice of y. 14 d Corollary 14.11. For d 3, there exists x Z for which R(o, x) > R . 15 ≥ ∈ ∞ Proof. From (14.7), it is clear that R(o, x) R(o, ) if and only if 16 ≤ ∞ 1 1 cos(x t) 1 1 , d − · dt d dt (2π) ZTd S (t) ≤ (2π) ZTd S (t) which is equivalent to 17

1 cos(x t) . d · 0 (14.12) (2π) ZTd S (t) ≥ d However, (14.12) cannot hold for all x Z , as such an inequality would mean that all 18 ∈ Fourier coeﬃcients of w are nonnegative, in violation of Heisenberg’s uncertainty princi- 19 ple. 20 Remark 14.12. Corollary 14.11 leads to the paradoxical conclusion that given x G0, there may be a y which is “further from x than inﬁnity”. This is the case for d = 3;∈ numerical computation of (14.10) gives lim R(x, y) 0.5054620038965394, in Z3, (14.13) y →∞ ≈ and for y = (1, 1, 1), 21 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 145

x y

Figure 11. In Z3, it may happen that R(x, y) > R(x, ), where R(x, ) = ∞ ∞ limz R(x, z). This phenomenon is represented here schematically as a “black hole”. →∞

R (o, y) 0.5334159062457338. (14.14) ≈ 1 In fact, numerical computations indicate the following extremely bizarre situation:

R(x, y ) < R(x, ) < R(x, y + ), for y := (n, n, 0). 2k ∞ 2k 1 n

2 Remark 14.13. An application of Bochner’s Theorem (see Theorem 7.4) yields a unique d 3 Radon probability measure P on T such that

it x R(o,x) d e · dP(t) = e− 2 , x Z . ZTd ∀ ∈

d 2 d 4 Corollary 14.14. For (Z , 1),v ℓ (Z ) if and only if d 3. x ∈ ≥ 5 Proof. By computations similar to those in the proof of Theorem 14.9, one can see that in

i x t 2 d 6 absolute values, the integrand (e 1)/S (t) of (14.4) is in L (T ) if and only if d 3, in · − ≥ 7 which case Parseval’s theorem applies.

d 2 d 8 Corollary 14.15. For (Z , 1), the monopole w ℓ (Z ) if and only if d 5. ∈ ≥ 9 Proof. The proof is almost identical to that of Corollary 14.14, except that the integrand is 2 d 10 1/S (t), which is in L (T ) if and only if d 5. ≥ 11 Example 14.16. To see why R is not bounded on (Z, 1), one can evaluate (14.3) explicitly 12 via the Fej´er kernel:

1 sin2((x y) t ) R(x, y) = − 2 dt (2π) ZT 2 t sin 2 1 sin2((x y) t ) = x y − 2 dt (2π) ZT | − | x y sin2 t | − | 2 1 = x y FN (t) dt | − |(2π) ZT = x y . | − | 13 where F (t) is the Fej´er kernel with N = x y ; see Figure 12. Of course, this was to be N | − | 14 expected because R coincides with shortest path metric on trees. 146 PALLEE.T.JORGENSENANDERINP.J.PEARSE

3.0

2.5

2.0 2.0

1.5 1.5

1.0 1.0 1.0

0.5 0.5 0.5

4 2 2 4 6 8 10 4 2 2 4 6 8 10 4 2 2 4 6 8 10

Figure 12. The function vx, for x = 1, 2, 3 in Z, as obtained from the Fej´er kernel. See Example 14.16.

The following result is well known; we include it for completeness and the novelty of 1 the proof. 2

d Theorem 14.17. h is a harmonic function on (Z , 1) if and only if h is linear (or aﬃne). 3 d Consequently, = [ in] for Z . 4 HE F

Proof. From ∆h = 0, the Fourier transform gives S (t)h(t) = 0. By the formula of 5 Lemma 14.3 for S (t), this means hˆ can only be supported at t = 0 and hence that hˆ is 6 a distribution which is a linear combination of derivatives of the Dirac mass at t = 0; 7 see [Rud91] for this structure theorem from the theory of distributions. 8 Denoting this by hˆ(t) = P(δo), where δo is the Dirac mass at t = 0, and P is some 9 polynomial. The inverse Fourier transform gives h(x) = P(x). If the degree of P is 2 10 or higher, then ∆h will have a constant term of 2d (cf. (14.2)) and hence cannot vanish 11 − identically. 12 d It is clear that a linear function on Z has inﬁnite energy; consequently arm is empty 13 H on this network and the second conclusion follows. 14

d d Example 14.18 (Nontrivial harmonic functions on Z Z ). Consider the disjoint union of 15 d ∪ two copies of Z , with c = 1 and d 3. Now connect the origins o , o of the two lattices 16 ≥ 1 2 with a single edge of conductance co1o2 = 1. Let w1 be a monopole on the first copy 17 d ∈HE of Z , as ensured by Theorem 14.5. We may assume w1 is normalized so that w(o1) = 1, 18 and then extend w1 to the rest of the network by lettingw ˜ 1(x) = 0 for all x in the second 19 d d copy of Z . Similarly, let w2 be a function which is a monopole on the second Z , satisfies 20 d w(o2) = 1, and extend it tow ˜ 2 by definingw ˜ 2(x) = 0 for x in the first copy of Z . Now 21 one can check that ∆w˜ 1 = ∆w˜ 2 = δo2 . Note that the unit drop inw ˜ 1 across the edge co1o2 22 − d moves the Dirac mass of ∆w1 to the second copy of Z . Now define 23

h := w w . (14.15) 1 − 2 It is easy to check that h and that h arm. 24 ∈HE ∈H Corollary 14.19. If w is the monopole on (Zd, 1),d 3, then ≥ w(x) = 1 (R(o, x) R(o, )), (14.16) 2 − ∞ and consequently limx w(x) = 0. 25 →∞

Proof. Subtract (14.10) from (14.8) and compare to (14.5). For the latter statement, one 26 can take the limit of (14.16) as x directly or apply the Riemann-Lebesgue lemma to 27 →∞ (14.5). 28 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 147

Corollary 14.20. If w is the monopole on (Zd, 1), then 1 (w) = lim R(x, y). (14.17) 2 y E →∞

1 Proof. Compare (14.17)to(14.10) and note that arm = 0 , so w is unique. H { } d 2 Remark 14.21. For (Z , 1), it is instructive to work out directly why in is dense in . F HE 3 That is, let us suppose that the boundary term vanishes for every v , and use this to ∈ HE 4 prove that every function which is orthogonal to in must be constant (and hence 0 in ). F HE 5 This shows that in is dense in in the energy norm. F HE 6 “Proof”. If v , then v = v, ∆v c < , the Fourier transform sends v vˆ(t) = E E in t ∈ H k k h i ∞ 7→ 7 Z vne · and P d v, ∆v c (2π)− vˆ(t)S (t)ˆv(t) dt < , (14.18) h i 7→ ZTd ∞

d 2 tk 8 where S (t) = 4 k=1 sin 2 , as in Lemma 14.3. Then note that the Schwarz inequality 9 gives P

2 S (t)ˆv(t) dt S (t) dt S (t)ˆv(t)2 dt, ZTd ! ≤ ZTd ZTd 1 d 10 so that S (t)ˆv(t) L (T ). From the other hypothesis, v in means that δx, v = 0 for 0 ∈ ⊥ F h iE 11 each x G , whence Parseval’s equation gives ∈

d im t 0 = δxm , v = δxm , ∆v c (2π)− e · S (t)ˆv(t) dt = 0, m. h iE h i 7→ ZTd ∀ 1 d 12 This implies that S (t)ˆv(t) = 0 in L (T ), and hencev ˆ can only be supported at t = 0. From 13 Schwartz’s theory of distributions, this means

(2) vˆ(t) = f0(t) + c0δ0(t) + c1Dδ0(t) + c2D δ0(t) + ..., 1 14 where f0 is an L function and all the other terms are derivatives of the Dirac mass at t = 0 (2) 15 (D is a diﬀerential operator of rank 2, etc.). 16 Ifv ˆ is just a function, then it is 0 a.e. and we are done. If the distribution δ0(t) is a 1 17 component ofv ˆ, then − (δ0) = 1, which is zero in . In one dimension, the distribution F 1 HE 18 δ (t) cannotbe a componentofv ˆ because (δ )(x ) = m, and this function does not have 0′ F − 0′ m 19 ﬁnite energy(the computationof the energypicks up a term of 1 on every edgeof the lattice d 20 Z ). The computation is similar for higher derivatives of δ0, but they diverge even faster. 21 For higher dimensions, note that D δ , = Dδ δ and (Dδ δ ) = (Dδ ) (δ ) (this 1 (0 0) 0 ⊗ 0 E 0 ⊗ 0 E 0 E 0 22 is a basic fact about quadratic forms on a Hilbert space), and so this devolves into same 23 argument as for the 1-dimensional case.

24 Remark 14.21 does not hold for general graphs; see Example 13.2. Also, the end of 1/2 1/2 25 the proof shows why ∆˜ (v + k) = ∆˜ vˆ, as mentioned in Remark 9.6; the addition of a 26 constant correspondson the Fourier side to the addition of a Dirac mass outside the support d 27 of χ. d 28 The case of (Z , 1), for d = 1 is a tree and hence very simple with R(x, y) = x y , and | − | 29 for d 3 may be fairly well understood by the formulas given above. However, the case ≥ 148 PALLEE.T.JORGENSENANDERINP.J.PEARSE d = 2 seems to remain a bit mysterious. It appears that R(x, y) log(1 + x y ); we now 1 ≈ | − | give two results in this direction. 2

Remark 14.22. From Theorem 14.9 it is clear that for d 3, if yn zn and both tend to , 3 ≥ ∼ 2 ∞ one has limn (R(x, yn) R(x, zn)) = 0. In fact, this remains true in Z but not Z. For Z, 4 →∞ −

n y z = R(x, y ) R(x, z ) = 1 →∞ 1 , 0. n ∼ n ⇒ | n − n | −−−−−−→ 2 A little more work is required for Z , where we work with x = o for simplicity: 5

1 cos(y t) cos(y t + t ) , , = k R(o z) R(o y) 2 · − · dt − (2π) ZT2 S (t) 1 cos(y t)(1 cos t ) + sin(y t) sin t = k k . 2 · − · dt (2π) ZT2 S (t)

1 cos tk sin tk 1 2 One can check that − , L (T ) by converting to spherical coordinates and mak- 6 S (t) S (t) ∈ ing the estimate 7

ρ 2 ρ dρ dθ< . ZT2 ρ ∞

Now the Riemann-Lebesgue Lemma shows that R(o, y) R(o, z)tendsto 0 as y (and hence 8 − also z) tends to . 9 ∞ Theorem 14.23. On (Z2, 1), the gradient of R vanishes at inﬁnity, i.e., d lim R(x, y) = 0. (14.19) y dy →∞ k

Theorem 14.24. On (Z2, 1), resistance distance is given by x R(x, y) = A(o, t) dt + ∞ B(x, t) dt, (14.20) Z0 Z0 ∂ ∂ where A(s, t) := B(s, t) := (14.21) ∂t1 ∂t2

14.2. Noncompactness of the transfer operator. 10 Example 14.25 (T may not be the uniform limit of ﬁnite-rank operators). Let G be the integers Z with edges only between vertices of distance one apart (as in Example 14.2 with + d = 1), with c 1. Then the transfer operator T := σ + σ− consists of the sum of two unilateral shifts,≡ for which the ﬁnite truncations (as described just above) are the banded matrices 0 1 0 0 ... 0 0 0  1 0 1 0 ... 0 0 0   0 1 0 1 ... 0 0 0     0 0 1 0 ... 0 0 0    TN =  ......  . (14.22)  ......     0 0 0 0 ... 0 1 0     0 0 0 0 ... 1 0 1     0 0 0 0 ... 0 1 0      OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 149

Then consider the vectors

∞ 1 π2 ξ := (0,..., 0, 1, 1 , 1 , 1 ,... ), ξ 2 = 2 = . (14.23) n 2 3 4 k nk k2 3 n zeros Xk=1 | {z } 1 Then TN does not converge to T uniformly, because for n = N,

1 1 ξ , (T T )ξ = ξ ξ + = = h n − n nic k k 1 (k n)(k n + 1) k(k + 1) Xk >n Xk >n Xk 1 | | | | − − | |≥ 1 2 π2 , ≥ k + 1! ≈ 6 Xk 1 | |≥ 2 which is bounded away from 0 as n . →∞ 3 14.2.1. The Paley-Wiener space Hs. The transfer operator is not compact in , as Ex- HE 4 ample 14.25 shows. However, by introducing the correct weights we can obtain a com- 5 pact operator, i.e., the transfer operator is compact when considered as acting on the 2 6 correct Hilbert space. To this end, we make the identification between ξ ℓ (Z) and n 2 ∈ 7 f (z) = n Z ξnz L (T) via Fourier series, so that we may use analytic continuation and ∈ ∈ 8 introduceP the following spaces. 2 n Definition 14.26. For an function f L (T) given by f (z) = n Z ξnz , we define ∈ ∈ 2 P s n 2 f := e | | ξ , (14.24) k ks  | n|  Xn Z   ∈  and consider the space.   . H := f : T T .. f < . (14.25) s { → k ks ∞} 2 2 9 For s = 0 we recover good old ℓ (c), but for s > 0, we have the subspace of ℓ (c) which . 10 consists of those functions with an analytic continuation to the annulus z .. 1 s < z < 1+s 2 { − | | } 11 about T. In general, we have Hs L (T) H s = Hs∗. ⊆ ⊆ − 12 Theorem 14.27. The transfer operator is a compact operator on Hs.

13 Proof. Using ∆= c T, we show that there exist solutions to ∆v = δα δω by construction, − − 14 using spectral theory. The Laplacian may be represented as the inﬁnite symmetric banded 15 matrix

... c(x1) c(x1, y1) ...... c(x , y ) − c(x ) c(x , y ) ...  − 1 1 2 − 2 2   ... c(x2, y2) c(x3) c(x3, y3) ...   − −   . . .   ......      16 The symmetry is immediate from the symmetry of cxy, of course. 17 Using the same notation as in Example 14.25, we must check that T T uniformly N → 18 in H , so we examine D := T T : s N − N

s n ξ, D ξ = e | |ξ ξ + h N i n n 1 Xn N | |≥ 150 PALLEE.T.JORGENSENANDERINP.J.PEARSE

1/2 1/2 s n 2 s n 2 e | | ξ e | | ξ + ≤  | n|   | n 1|   nXN   Xn N  | |≥  | |≥     1/2  1/2 s n+N 2 s n+N 2 = e | | ξ + e | | ξ + +  | n N|   | n N 1|  Xn 0  Xn 0  | |≥  | |≥     1/2  1/2 s/2 sN 2 sn 2 sn 2 = e− e ξ + e ξ e ξ (14.26)  | N | | n|   | n|   nXN+1   nXN+1   | |≥  | |≥   /     sn 2 1 2   ξ, DN ξ n N+1 e ξn N = e s/2 | |≥ | | →∞ 0. h i − 1/2 ξ s sN P2 sn 2 − −−−−−−→ k k e ξN + n N+1 e ξn | | | |≥ | | P Then T is compact in H , and in fact, T is trace class! Then cptˆ = 0, so ∆= I T implies 1 s − ∆=ˆ ˆ. 2 C

14.3. Non-simple integer lattice networks. In this section, we illustrate some of the phe- 3 nomena that may occur on integer lattices when the conductances are allowed to vary. 4 Many of these examples serve to demonstrate certain deﬁnitions or general properties dis- 5 cussed in previous sections. 6

Example 14.28 (Symmetry of the graph vs. symmetry of the network). Consider the 2- 7 dimensional integer lattice; the case d = 2 in Example 14.2, and think of these points 8

as living in the complex plane, so each vertex is m + ni, where m and n are integers and 9

i = √ 1. It is possible to deﬁne the conductances in such a way that a function v(z) has 10 −

ﬁnite energy, but v(iz) does not (this is just precomposing with a symmetry of the graph: 11 π 2 2 2 rotation by by 2 ). However, v(z) is in ℓ (1) if and only if v(iz) is in ℓ (1). Thus, ℓ (1) does 12 not see the graph. 13 Deﬁne the conductances by 14

1, y = x + 1, c = xy  Im(y) 2| |, y = x + i,   so that the conductances of horizontal edges are all 1 and the conductances of vertical 15 k edges grow like 2 . Now consider the function 16

Re(x) 2−| |, y = 0,

v(z) = v(x + iy) = 0, y , 0,   When computing the energy (v), the only contributing terms are the edges along the real 17 E axis, and the edges immediately adjacent to the real axis: 18

(v(z)) = horizontal + vertical E = 2(1/2 + 1/4 + 1/8 + ...) + 4(1/2 + 1/4 + 1/8 + ...) = 6, which is ﬁnite. However,

(v(iz)) = 2(1 + 1 + 1 + ...) + 4(1/2 + 1/4 + 1/8 + ...) = . E ∞ OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 151

2 2 0 1 Example 14.29 (An example where ℓ * ). Let Z have cn 1,n = n. Consider ℓ (G , ν) HE − 2 where ν is the counting measure. The Dirac functions δxk satisfy δxk = 1, so δxk is a 2 0 k k { } 3 bounded sequence in ℓ (G ). However, the Laplacian is

. .. ∆ =  n 2n + 1 (n + 1)   − −   ..   .     k  4 and (δx ) = δx , ∆δx = 2k + 1 →∞ . So we cannot have the bound v K v , E k h k k i −−−−−−→∞ k kE ≤ k k 5 for any constant K. 6 This is “corrected” by using the measure c instead. In this case, δ = 2k + 1 so that k kkc 7 δ is not bounded and we must use δ / √c(k) . But then the Laplacian is { k} { k } . .. ∆ =  n 1 n+1  c  2n+1 2n+1   − − .   ..      δxk 1   8 and = c(x ) (δx) = 1. E √c(xk) k E 9 Example 14.30. It is quite possible to have unboundedfunctions of finite energy. Consider 10 the network (G, c) = (Z, 1) with vertices at each integer and unit conductances to nearest n 1 11 nearest neighbours. Then it is simple to show that u(n) = i=1 n and v(n) = log 1 + n 1 π2 | | 12 = are unbounded and have finite energy — use the identity Pn∞=1 n2 6 . For v, note that 1 1 13 log 1 + n log 1 + (n 1) = log 1 + . P | |− | − | n ≤ n 1 14 Example 14.31 (An unbounded function with finite energy). Let Z have cn 1,n = 2 . Then − n 15 the function f (n) = n is clearly unbounded, but

1 1 π2 = 2 = = < . ( f ) 2 ( f (n) f (n 1)) 2 E X n − − X n 3 ∞ 16 Conclusion: it is possible to have unbounded functions of finite energy if c decays suffi- 17 ciently fast. d 18 We’ve seen that there are no harmonic functions of finite energyon(Z , c), when c = 1. 19 However, the situation is very different when c is not bounded. 1 20 Theorem 14.32. arm , 0 for (Z, c) iff c < . In this case, arm is spanned by a H −xy ∞ H 21 single bounded function. P 1 22 Proof. ( ) Fix u(0) = 0, define u(1) = and let u(n) be such that ⇒ c01 1 u(n) u(n 1) = , n. (14.27) − − cn 1,n ∀ − 23 Now u is harmonic:

∆u(n) = cn 1,n(u(n) u(n 1)) cn,n+1(u(n + 1) u(n)) − − − − − 1 1 = cn 1,n cn,n+1 = 0, − cn 1,n − cn,n+1 − 152 PALLEE.T.JORGENSENANDERINP.J.PEARSE and u is of ﬁnite energy 1

2 1 (u) = cn 1,n(u(n) u(n 1)) = < . E − − − c ∞ Xn Z Xn Z n 1,n ∈ ∈ − Note that once the values of u(0) and u(1) are ﬁxed, all the other values of u(n) are deter- 2 mined by (14.27). Therefore, arm is 1-dimensional. 3 H ( ) If ∆u(n) = cn 1,n(u(n) u(n 1)) cn,n+1(u(n + 1) u(n)) = 0 for every n, then 4 ⇐ − − − − −

cn 1,n(u(n) u(n 1)) = cn,n+1(u(n + 1) u(n)) = a, − − − − for some ﬁxed a (the amperage of a sourceless current). Then 5

2 2 1 (u) = cn 1,n(u(n) u(n 1)) = a < , (14.28) E − − − c ∞ Xn Z Xn Z n 1,n ∈ ∈ − since u arm . Note that (14.28) implies u is bounded: (u) = a n Z(u(n) u(n 6 ∈H ⊆HE E ∈ − − 1)) and n 1(u(n) u(n 1)) = limn u(n) u(0). The function u is monotonicP because 7 ≥ − − →∞ − it is harmonic,P so the sum is absolutely convergent. 8

We will now explore a specific example of this kind of network, where explicit compu- 9 tations are tractable. 10 n Definition 14.33. For a fixed constant c > 1, let (Z, c ) denote the network with integers 11 for vertices, and with geometrically increasing conductances defined by 12

max n , n 1 cn 1,n = c {| | | − |}, − so that the network under consideration is 13

c3 c2 c c c2 c3 c4 ... 2 1 0 1 2 3 ... − − We ﬁx o = 0. 14 n Lemma 14.34. On (Z, c ), the energy kernel is given by 15

0, k 0, k+1 ≤ v (k) =  1 r n > 0, n  −1 r , 1 k n,  n+1 ≤ ≤  1 −r  −1 r , k n,  − ≥  n r and similarly for n < 0. Furthermore, the function wo(n) = ar| |, a := 2(1 r) , deﬁnes a 16 − monopole, and h(n) = sgn(n)(1 w (n)) deﬁnes an element of arm. 17 − o H n n n 1 Proof. It is easy to check that ∆wo(0) = 2c(a ar) = 1, and that ∆wo(n) = c (ar ar − )+ 18 n+1 n n+1 , − r − c (ar ar ) = 0 for n 0. The reader may check that (wo) = 2(1 r) so that wo . 19 − E − ∈HE The computations for vx and h are essentially the same. See Figure 13. 20

Example 14.35 ( in not dense in in). On (Z, 2n), in = cl span δ δ is not dense F 2 F F 2 { x − o} in in (see Deﬁnition 8.4). To illustrate this, we compute the projection of δo to in2. ThisF may be accomplished by computing − F . u := projection of δ to span δ δ .. x n , n − o { x − o | |≤ } OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 153

r = 1/c r r

v1 c2 0 c0 c c2 c3 c4 -1 0 1 2 3 4

r+r 2 r+r 2 r+r 2 r

v2 c2 0 c0 c c2 c3 c4 -1 0 1 2 3 4

r+r 2+r 3 r+r 2+r 3 r+r 2 r

v3 c2 0 c0 c c2 c3 c4 -1 0 1 2 3 4

r a = w 2(1 −r) o ar ar ar 2 ar 3 ar 4 c2 c0 c c2 c3 c4 -1 0 1 2 3 4

r2 f = PF v /2 3 1 in 1 r / r4 5 2 /2 r / c2 c c c2 c3 c4 2 -1 0 1 2 3 4 − 3 r / −r2 2 /2 − r /2

n Figure 13. The functions v1, v2, and v3 on (Z, c ). Also, the monopole wo and the projection f1 = P inv1. See Lemma 14.34. F 4 2 2 4 8 16 -2 -1 0 1 2 3 4 −δ 0 −1 PF in 2

1 1 1 − − 1 − 5 5 − −1 −1 15 4 2 2 415 830 16 60 -2 -1 0 1 2 3 4

PF −δ 2 in 2( 0) −− 3

Figure 14. The projection of the Dirac mass δo onto in2; see Example 14.35 and − F also Lemma 14.34 and Lemma 14.34. 154 PALLEE.T.JORGENSENANDERINP.J.PEARSE and then taking the limit as n . The result is depicted in Figure 14. We leave the 1 → ∞ n computation of the case of general geometric conductance (Z, c ) as an exercise. 2

14.4. Defect spaces. We will construct a defect vector u satisfying ∆u = u on 3 n ∈ HE − (Z, c ), c > 1, the 1-dimensional integer lattice with geometrically growing conductances. 4 n We do this in two stages: (i) construct a defect vector on (Z+, c ), and (ii) combine two 5 n copies of this defect vector to obtain an example on (Z, c ). 6

Example 14.36 (Defect on the positive integers). We consider (Z+, c) where 7

n cn 1,n = c , n 1, − ≥ for some ﬁxed c > 1. Thus, the network under consideration is 8

c c2 c3 c4 0 1 2 3 ...

Now recursively deﬁne a system of polynomials in r = 1/c by 9

pn 1 1 1 1 1 1 0 = n n 2 2 " qn # " r 1 + r # ··· " r 1 + r # " r 1 + r # " 1 #

We will show that u(n) := q satisﬁes ∆u = u and has ﬁnite energy. It will be helpful to 10 n − note that 11

p = cn(u(n) u(n 1)), (14.29) n − − and hence 12

n+1 pn+1 = pn + qn, and qn+1 = qn + r pn+1.

Now, ∆u = u because 13 −

∆u(n) = p p + = q = u(n). n − n 1 − n − We will need the following lemma to show that u . 14 ∈HE m m m Lemma 14.37. There is an m such that p n and q (n + 1) n for n Z+. 15 n ≤ n ≤ − ∈ Proof. We prove both bounds simultaneously by induction, so assume both bounds hold 16 for n and prove 17

m pn+1 (n + 1) , and ≤ m m q + (n + 2) (n + 1) . n 1 ≤ − The estimate for pn+1 = pn + qn is immediate from the inductive hypotheses. For the qn+1 18 estimate, choose an integer m so that 19

2 2 t . 2 m(m 1) max t r .. t 0 = . − ≥ { ≥ } e log c! 2 n+1 Then (n + 1) r m(m 1) for all n, so 20 ≤ − OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 155

4.0

3.5 4.044

3.0 4.043

2.5 4.042

2.0 4.041

1.5 4.040

2 4 6 8 10 5 10 15 20 25 30

n Figure 15. A Mathematica plot of the defect vector u on (Z+, 2 ); see Example 14.36 and Lemma 14.38. The left plot shows u(x) for x = 0, 1,..., 10, and the plot on the right shows data points for u(x), x = 10, 11, 12,... .

m(m 1) n m n + 2 m 2 + rn+1 2 + − + , ≤ (n + 1)2 ≤ n + 1 n + 1! n m 1 m n+2 m 1 m 1 by using the binomial theorem to expand n+1 = 1 n+1 and n+1 = 1 + n+1 . m − 2 Multiplying by (n + 1) gives

((n + 1)m nm) + rn+1(n + 1)m (n + 2)m (n + 1)m, − ≤ − n+1 3 which is suﬃcient because the left side is an upper bound for qn+1 = qn + r pn+1.

4 Lemma 14.38. The defect vector u(n) := qn has ﬁnite energy and is bounded.

5 Proof. Applying Lemma 14.37 to the formula for yields E

∞ n 2 ∞ n 2 ∞ n 2m (u) = c (u(n) u(n 1)) = r pn r n = Li 2m(r) < , E − − ≤ − ∞ Xn=1 Xn=1 Xn=1

6 since a polylogarithm indexed by a negative integer is continuous on R, except for a single 7 pole at 1 (but recall that r (0, 1)). ∈ Lemma 14.38 ensures that the defect vector is bounded; in the example in Figure 15, the defect vector has a limiting value of 4.04468281, although the function value does not exceed 4 until x = 10. The ﬁrst few values≈ of the function are 3 17 173 3237 114325 7774837 1032268341 270040381877 140010315667637 u = 2 , 8 , 64 , 1024 , 32768 , 2097152 , 268435456 , 68719476736 , 35184372088832 ,... h[1.5, 2.125, 2.7031, 3.1611, 3.4889, 3.7073, 3.8455, 3.9296, 3.9793, 4.0080i ,... ] ≈ 8 While we are unable to provide a nice closed-form formula for the defect vector, we 9 can provide generating functions for it, using the pn = pn(r) and qn = qn(r) obtained just 10 above. Deﬁne

∞ n ∞ n P(x) = pn(r)x and Q(x) = qn(r)x . Xn=0 Xn=0 n+1 11 Multiplying both sides of p + = p + q by x and summing from n = 0 to , n 1 n n ∞

P(x) = xP(x) + xQ(x), (14.30) 156 PALLEE.T.JORGENSENANDERINP.J.PEARSE where we have used the fact that p0 = 0. Meanwhile, multiplying both sides of qn+1 = 1 n+1 n+1 q + r p + by x and summing from n = 0 to , 2 n n 1 ∞

Q(x) 1 = xQ(x) + P(rx). (14.31) − Write (14.30) in the form (1 x)P(x) = xQ(x) and substituting in (1 x)Q(x) = 1 + P(rx) 3 − (1 x)2 − from (14.31), to get 1 + P(rx) = (1 x)Q(x) = − P(x) or 4 − x

= x + x = x + x(rx) + x(rx) 2 P(x) (1 x)2 (1 x)2 P(rx) (1 x)2 (1 x)2(1 rx)2 (1 x)2(1 rx)2 P(r x) − − − − − − − n + ∞ rk x ∞ rn(n 1)/2 xn = = . ··· (1 rk x)2 n (1 rk x)2 Xn=0 Yk=0 − Xn=0 k=0 − Q k k Note that r (0, 1), so P(r x) →∞ P(0) = 0, again since p0 = 0. Now (14.30) gives 5 1 x ∈ −−−−−−→ Q(x) = −x P(x), whence 6

+ ∞ rn(n 1)/2 xn 1 Q(x) = − . n (1 rk x)2 Xn=0 k=1 − Q Example 14.39 (Defect on the integers). We consider (Z, c) as in Deﬁnition 14.33: 7

c3 c2 c c c2 c3 c4 ... 2 1 0 1 2 3 ... − − Proceeding as in Example 14.36, one uses ∆u(0) = u(0) to compute 8 −

2c(u(0) u(1)) = u(0) = u(1) = 1 + 1 u(0), − − ⇒ 2c 1 r and obtain the initial values p1 = 2 and q1 = 1 + 2 . Therefore, for Z we instead use the 9 polynomials deﬁned by 10

p 1 1 1 1 1 1 0 n = n + n 2 + 2 + 1 " qn # " r 1 r # ··· " r 1 r # " r 2 r # " 2 # n The other computations are essentially identical to those for (Z+, c ). 11 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 157

1 15. Application to magnetism and long-range order

2 The integer lattice examples studied in 14 may be applied to the theory of ferromag- § 2 3 netism. In 15.1, we construct a Hilbert space L (Ω, P) with a probability measure, follow- § 2 4 ing techniques of Kolmogorov. Since L (Ω, P) , this provides a concrete realization HE 5 of as a probability space and a commutative analogue/precursor of the Heisenberg spin HE 6 model developed in 15.3. In 15.2 we carry out the GNS construction [Arv76b] to ob- § § 7 tain a Hilbert space ϕ. Again, this will be useful for 15.3, where we recall Powers’ H § 8 approach (using a β-KMS state ϕ) and show how our results may be used to obtain certain 9 refinements of Powers’ results. 10 In [Pow76a], Powers made the first connection between the two seemingly unrelated 11 ideas: resistance distances in electrical networks, and a problem from statistical mechanics. 12 Even in the physics literature, one often distinguishes between quantum statistical model 13 as emphasizing (a) physics, or (b) rigorous mathematics. The literature for (a) is much 14 larger than it is for (b); in fact, the most basic questions (phase transition and long-range 15 order) are notoriously difficult for (b). Powers was concerned with long-range order in 16 ferromagnetic models from quantum statistical mechanics, especially Heisenberg models. 17 The notion of long-rangeorder in these models depends on a chosen Hamiltonian H,anda 18 C∗-algebra A of local observables for such models. These objects and ideas are discussed 19 in more detail in 15.3 and 15.3. The reader may also find some information on β-KMS § § 20 states in 15.4. § 21 While we shall refer to the literature, e.g. [BR79,Rue69] for formaldefinitionsof the key 22 terms from the C∗-algebra formalism of quantum spin models, physics, KMS states and the 23 like, we present a minimal amount of background and terminology from the mathematical 24 physics literature so our presentation is agreeable to a mathematical audience. A brief 25 discussion of KMS states is given in 15.4 and the reader may wish to peruse the general § 26 GNS construction is outlined in 15.2 before reading 15.2. § § 27 Here we turn to a non-commutative version of the infinite Cartesian products that went 28 into the probabilistic constructs used in sections 7 and 11 above. This is dictated directly 29 from quantum physics: Think of an algebra of observables placed on each vertex point 30 in an infinite graph, each algebra non-abelian because of quantum mechanics. The infinite 31 graphsheremay representsites froma solid state model, or a spin-model for magnetization. 32 To achieve our purpose, we will use infinite tensor products of C∗-algebras, one for each 0 33 point x G . This is dictated by our application to quantum statistical mechanics. In ∈ 34 quantum physics, the entity that corresponds to probability measures in classical problems 35 are however “states” on the algebra of all the quantum mechanical observables, a C∗- 36 algebra, but the C∗-algebra for the entire system will be an infinite tensor product C∗- 37 algebra. To gain intuition, the reader may wish to think of states as non-commutative 38 measures, and hence non-commutative probability theory (see e.g., [BR97].)

2 39 15.1. Kolmogorov construction of L (Ω, P). As a prelude to the quantum-mechanical 40 model, we ﬁrst give a probabilistic model, that is a model for classical particles, which 41 serves to illustrate the main themes. In particular, long-range order appears in this setting 42 as an estimate on correlations (in the sense of probability). 43 We consider a Brownian motion on (G, c) as a system of Gaussian random variables, 0 44 again indexed by G . For these (commutating) random variables, we will show the corre- 45 lations are given by the resistance distance R(x, y). This result is extended to the noncom- 46 mutative setting via the GNS construction in 15.2. § 158 PALLEE.T.JORGENSENANDERINP.J.PEARSE

In , we don’t really have an algebra of functions, so first we make one. Since 1 HE 0 0 E(vx, vy) := vx, vy is a positive definite form G G C, we can follow Kolmogorov’s 2 h iE × →0 construction of a measure on the space of functions on G . Denoting the Riemann sphere 3 2 by S = C , this produces a probability measure P on 4 ∪ {∞} Ω= S2, (15.1) xYG0 ∈ 0 the space of all functions on G . Also, we define 5

v˜ : Ω C byv ˜ ( f ) := f (x) f (o). (15.2) x → x − Since is a Hilbert space, is its own dual, and we can think of vx as an element of 6 HE HE or the function on deﬁned by vx, . In the latter sense,v ˜x is an extension of vx to 7 HE HE h ·iE Ω; observe that for u , 8 ∈HE

v˜x(u) = vx, u = u(x) u(o). (15.3) E 2 h i − Thus we have a Hilbert space L (Ω, P) which contains as a dense subalgebra the algebra 9 generated by v˜ . 10 { x} Another consequence of the Kolmogorov construction is that 11

E(v˜x v˜y) = v˜x v˜y dP = vx, vy . (15.4) Z h iE 2 Lemma 15.1. L (Ω, P) is unitarily equivalent to . 12 HE Proof. The mappingv ˜ v extends by linearity to an isometric isomorphism: 13 x 7→ x

2 2 E  cxv˜x  = cxvx c¯xE(v˜x vy)cy = c¯x vx, vy cy, ⇐⇒ h iE  xXG0  xXG0 x,Xy G0 xX,y G0  ∈  ∈ ∈ ∈   E which is true by (15.4 ). 14 2 Observe that one recognizes ψ L (Ω, P) as correspondingto a ﬁnite linear combination 15 ∈ 0 c v if and only if there is a ﬁnite subset F G and a function u : F C with spt u = F 16 x x ⊆ → suchP that ψ(ω) = u(ω). By Riesz’s Lemma, integration with respect to P is given by a 17 positive linear functional ϕP, i.e., the expectation is 18

E( f ) = f dP = ϕP( f ). Z Since P is a probability measure, we even have ϕP(1) = 1. Consequently, ϕP corresponds 19 to a state in the noncommutative version (cf. Deﬁnition 15.3); see Remark 15.7 and the 20 table of Figure 16. 21 d Example 15.2 (Application of Lemma 15.1 to the integer lattice network (Z , 1)). Observe 22

that Bochner’s Theorem (Theorem 7.4) gives 23

i i i x ξ y ξ (y x)ξ Eξ(e · e · ) = e − dP(ξ) = vx, vy .

ZRd h iE i i x ξ d x ξ Thus, we are obliged to setv ˜x(ξ) := e · for ξ R , whence the mapping e · vx(ξ) 24 2 d ∈ 7→ extends by Lemma 15.1 and L (R , P) . A particularly striking feature of this example 25 HE is that one sees that translation-invariance of the underlying network causes the a priori 26 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 159

d 1 inﬁnite-dimensional lion Ω to devolve into the ﬁnite-dimensional lamb R . The duals of 2 abelian groups are much tamer!

3 15.2. The GNS construction. The GNS construction takes a C∗-algebra A and a state 4 ϕ : A C (see Deﬁnition 15.3 just below), and builds Hilbert space ϕ and a rep- → H 5 resentation π : A B( ϕ). The main point is that even though a C -algebra can be → H ∗ 6 deﬁned axiomatically and without reference to any Hilbert space, one can always think 7 of a C∗-algebra as an algebra of operators on some Hilbert space. The GNS construction 8 stands for Gel’fand, Naimark, and Segal, and the literature on this construction is extensive; 9 we include a sketch of the proof, but point the reader to [Arv76b, 1.6] (for newcomers) § 10 and [BR79, 2.3.2] (for details). § 11 Following the overview of the general GNS construction, we explain how the GNS 12 construction provides a noncommutative analogue of the Kolmogorov model discussed in 13 the previous section. The Heisenberg model is built within the representation of a cer- 14 tain C∗-algebra, and we will need this framework to describe Powers’ results concerning 15 magnetism. We also provide some of the background material relevant to the applications 16 to the theory of magnetism and long-range order discussed in 15; see also the excellent § 17 references [Arv76a, Arv76b, Arv76c, BR79].

18 Deﬁnition 15.3. A state on a C -algebra A is a linear functional ϕ : A C which satisﬁes ∗ → 19 ϕ(A A) 0 and ϕ(1) = 1. ∗ ≥ 20 Theorem 15.4 (GNS construction). Given a C -algebra A, a unit vector 1 A and a state ∗ ∈ 21 ϕ, there exists

22 (1) a Hilbert space: ϕ, , ϕ, H h· ·i 23 (2) a representation π : A B( ϕ) given by A π(A)( ), and → H 3 7→ · 24 (3) a cyclic vector (the ground state ): ζ = ζϕ ϕ, ζ ϕ = 1, ∈H k k 25 for which ϕ(A) = ζ,π(A)ζ ϕ, A A. h i ∀ ∈ 4 Sketch of proof. For (1), deﬁne A, B ϕ := ϕ(A∗B). Deﬁne the kernel of ϕ in the nonstandard fashion h i . ker(ϕ) := A A .. ϕ(A∗A) = 0 . { ∈ } Intuitively, think ϕ(A A) f 2. Then one has a Hilbert space by taking the completion ∗ ↔ | | R ϕ = (A/ ker(ϕ))∼ . H 26 For (2), show that the multiplication operator π(A) : B AB is a bounded linear 7→ 27 operator on A. This follows from the computation

2 2 2 π(A)B ϕ A C B ϕ π(A)B A B k k ≤ k k ∗ k k ⇐⇒ k kϕ ≤ k kC∗ k kϕ 2 ϕ((AB)∗AB) A ϕ(B∗B), ⇐⇒ ≤ k kC∗ ϕ(B∗AB) 2 28 which is true because ϕB(A) := is a state, A∗A = A , and ϕB(A) A for ϕ(B∗B) k k k k | | ≤ k k 29 every A A. ∈ 30 For (3), start with 1 A. Then ζ = ζϕ is the image of 1 under the embedding ∈

pro jection A completion A A / / ϕ = ∼ ker(ϕ) H ker(ϕ) 3ζ is called the ground state because when ϕ is a KMS state built from the Hamiltonian H, one has Hζ = 0, i.e., the energy of ζ is 0 4This is why physicists make the inner product linear in the second variable. 160 PALLEE.T.JORGENSENANDERINP.J.PEARSE

probabilistic/classical quantum A A space Ω= x G0 S C∗-algebra = x G0 x ∈ ∈ Gaussian measureQ P state ϕ (or KMSN state ω) 2 probability space L (Ω, P) Hilbert space ϕ = GNS (A, ϕ) 2 H 0 functionv ˜ L (Ω, P) observable σ : G A, σx Ax constant function∈ 1 ground state ζ → ∈ 2 embedding W : L (Ω, P) representation πϕ : A ( ϕ) HE → →B H v v˜(1) A πϕ(A)ζ 7→ 7→ expectation E(v) = vdP measurement ϕ(σ) = ζ,πϕ(σ)ζ ϕ h i covariance E(v˜¯ v˜ ) =R v˜¯ v˜ dP correlation ϕ(σ σ ) = σ , σ ϕ x y x y ∗x · y h x yi R Figure 16. A “dictionary” between the classical and quantum aspects of this problem. In this table, S is the Riemann sphere (the one-point compactiﬁcation of C) and H is the Hamiltonian discussed by Powers. The notation σx σy is explained in (15.7). This table · is elaborated upon in Remark 15.7.

During this composition, 1 is transformed as follows: 1 1 + ker(ϕ) ζϕ. Finally, 1 7→ 7→ to verify the condition relating (1),(2),(3), use [ ]ϕ to denote an equivalence class in the 2 · quotient space and then ζϕ,π(A)ζϕ ϕ = [1]ϕ, A [1] ϕ = ϕ(1A1) = ϕ(A). 3 h i h · i Remark 15.5. When A is a commutative algebra of functions, it turns out that π( f )( ) is 4 · multiplication by f , in which case the notation is a bit heavy handed: 5

2 π( f )1P = f 1 = f, 1P L (Ω, P). · ∈ For the noncommutative case, things are diﬀerent and the full notation is really necessary. 6 2 (Note that 1P really does depend on P, in the same way that the unit in L (X,δo) is diﬀerent 7 2 from the unit of L (X, dx)). 8

Remark 15.6. In general, the resulting representation π : A B( ϕ) is a contractive 9 → H injective homomorphism, so that π(A) A . However, when A is simple (as is the case 10 k k ≤ k k in our setting), then π is actually an isometry. 11

Remark 15.7. (Kolmogorov construction vs. GNS construction) In Figure 16 we present 12 a table which gives an idea of how analogous ideas match up in the commutative and 13 noncommutative models on the same electrical resistance network (G, c). The titles of the 14 two columns in the table refer to nature of the corresponding random variables. In both 15 columns, the variables are indexed by vertices. 16 In the left column, Ω is simply a (commutative) family of measurable functions; the 17 0 collection of all measurable functions on G . One can think of this as the tensor product of 18 1-dimensional algebras C. On the right, the variablesare quantum observables, so noncom- 19 muting self-adjoint elements in a C∗-algebra. For infinite electrical resistance networks, the 20 C∗-algebra A = A(G) in question is built as an infinite tensor product of finite-dimensional 21 0 C∗-algebras Ax, x G . More specifically, A is the inductive limit C∗-algebra of algebras 22 A ∈ 0 A A (F) where F ranges over all finite subsets in G , and where each (F) = x F x is a 23 ∈ finite tensor product. N 24 A general element A A does not have a direct analogue in the left-hand column, as the 25 ∈ structure is much richer in the noncommutative case. The observable σ is a particularly 26 simple type of element of A; it is one which can be represented as a single element of 27 A A x G0 x. A general A can only be represented as a sum of such things; cf. [BR79]. 28 ∈ ∈ Q OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 161

One can think of Ω as L∞(Ω, P), as the latter is generated by the coordinate functions. Let Xx be the random variable X : Ω C by X (ω) = ω(x). x → x 0 1 (Recall that ω is any measurable function on G .) Then Xx corresponds tov ˜x = vx, , as h ·iE 2 can be seen by considering the reproducing kernel property (15.3), when representatives 3 of are chosen so that v(o) = 0. In this sense, L∞(Ω) is the commutative version of A. HE 4 Recall from Stone’s Theorem that every abelian von Neumann algebras is L∞(X) for some 5 measurable space X. In a similar vein, The Gel’fan-Naimark Theorem states that every 6 abelian C∗-algebra is C(X) for some compact Hausdorﬀ space X. Thus, C(Ω) is a dense 2 7 subalgebra of L (Ω, P) in the same way that A is a dense subalgebra of ϕ. H 8 One key point is that a state is the noncommutative version of a probability measure, 9 in the C∗-algebraic framework of quantum statistical mechanics. In the abelian case, the 10 Gaussian measure P is unique, while in the quantum statistical case, the states ϕ typically 11 are not unique. In fact, when requiring ϕ to be a KMS state as in the next section, then a 12 “phase transition” is precisely the situation of multiple β-KMS states corresponding to the 13 same value of β. KMS states are equilibrium states, so a phase transition is when more than 14 one equilibrium state (e.g., liquid and vapour) are simultaneously present; see 15.4. The § 15 physicist will recognize the table entry ϕ(A) = ζ,πϕ(A)ζ ϕ as the transition probability h i 16 from the ground state to the excited state A. d 17 In the Heisenberg model of ferromagnetism, the graph is (Z , 1), and the support of the d d 18 associated Gaussian measure P is R . Thus, the associated probability space is (R , P) and 2 d 2 d 19 the Hilbert space is L (R , P). As a result, the random variables are L -functions on R , d 20 obtained by extension from Z . See Example 15.2.

21 The use of β-KMS states is actually a crucial hypothesis in Powers’ Theorem (Theo- 22 rem 15.8), although this detail is obscured in the present exposition. The technical deﬁ- 23 nition of a KMS state is not critical for the main exposition of Powers’ problem, but his 24 results (and ours) would be unobtainable without this assumption. (The reasons for this 25 are somewhat involved, but hinge upon the stability of KMS states as equilibria.) Conse- 26 quently, we include a discussion in 15.4 outlining some key features of these states. § 27 Powers did not consider the details of the spectral representation in GNS representation 28 for the KMS states. More precisely, Powers did not consider the explicit function repre- 29 sentation (with multiplicity) of the resistance metric and the graph Laplacian ∆ as it acts 2 30 on the energy Hilbert space . The prior literature regarding ∆ has focused on ℓ , as HE 31 opposed to the drastically diﬀerent story for . HE

32 15.3. Magnetism and long-range order in electrical resistance networks. Following A A 33 [Pow75], we apply Theorem 15.4 to a β-KMS state ω and the C∗-algebra = x G0 x ∈ 34 described in Remark 15.7. The inverse temperature β will be ﬁxed throughout theN discus- 35 sion. 36 It is known that the translation-invariant ferromagnetic models do not have long-range d 37 order in Z when d = 1, 2. Powers suggested that it happens for d = 3. Below we supply 38 detailed estimates which bear out Powers’ expectations. We can now make more precise 39 the allusion which begins this section: Powers was the ﬁrst to make a connection between

40 (i) the resistance metric R(x, y), and 41 (ii) estimates of ω-correlations between observables localized at distant vertices x, y 0 ∈ 42 G . 162 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Precise estimates for (ii) are called “long-range order”; the Gestalt effect of this phenom- 1 enon is magnetism. 2 The Hamiltonian H appearing in Definition 15.13 as part of the definition of a β-KMS 3 1 state is a formal infinite sum over the edges G of the network, where the terms in the sum 4 are weighted with the conductance function µ. (An explicit formula appears just below in 5 (15.6).) The Hamiltonian H then induces a one-parameter unitary group of automorphisms 6 αt t R (as in Definition 15.12) describing the dynamics in the infinite system; and a KMS 7 { } ∈ state ω refers to this automorphism group. As mentioned above, the KMS states ω are 8 indexed by the inverse temperature β. The intrepid reader is referred to the books [BR79, 9 BR97, Rue69, Rue04] for details. 10 In this section, we discuss an application to the spin model of the isotropic Heisenberg 11 3 ferromagnet. Let G = Z and Ω 1. We consider each vertex x (or “lattice site”) to be a 12 ≡ particle whose spin is given by an observable σx which lies in the finite-dimensional C∗- 13 A A A algebra x. The C∗-algebra = x x describes the entire system. For the case when the 14 1 particles are of spin , an elementNσ A is expressed in terms of the three Pauli matrices: 15 2 x ∈

0 1 0 i 1 0 σx1 = , σx2 = − , σx3 = , (15.5) " 1 0 # " i 0 # " 0 1 # − 0 for any x G . Interaction in this isotropic Heisenberg model is given in terms of the 16 ∈ Hamiltonian 17

H = 1 c (1 σ σ ), (15.6) 2 xy − x · y x,Xy G0 ∈ where σ σ = σ σ + σ σ + σ σ . (15.7) x · y x1 ⊗ y1 x2 ⊗ y2 x3 ⊗ y3 More precisely, ω(I σ σ ) gives the amount of energy that would be required to in- 18 − x · y terchange the spins of the particle at x and the particle at y when measured in state ω, 19 and 20

ω(H) = 1 c ω(1 σ σ ) 2 xy − x · y x,Xy G0 ∈

is the weighted sum of all such interactions. The Hamiltonian H may be translated by time 21 i itH tH t into the future by A α (A) := e Ae− Aut(A). 22 7→ t ∈ Motivated by (15.11), Powers conjectures the following estimate for β-KMS states in 23 [Pow76a]: there exists a constant K (independent of G) for which 24

1 ω(1 σ σ ) Kβ− R(x, y). − x · y ≤ The following result appears in [Pow76b]. 25

Theorem 15.8 (Powers). Let ω be a β-KMS state and let H be the Hamiltonian of (15.6). 26 Then 27

ω(1 σ σ ) ω(H)R(x, y), (15.8) − x · y ≤ 3 After obtaining a bound for ω(H), the author notes that in Z , the “resistance between 28 o and inﬁnity is ﬁnite” and uses this to show that ω(1 σ σ ) = 1 ω(σ σ ) is 29 − x · y − x · y bounded. The interpretation is that correlation between the spin states of x and y remains 30 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 163

1 positive, even when x is arbitrarily far from y, and this is “long-range order” manifesting 2 as magnetism. We oﬀer the following improvement.

3 Lemma 15.9. If ω is a β-KMS state, then

1 dt ω σ σ ω . 1 liminf ( x y) 3 3 2 t (H) (15.9) y  π d k  − →∞ ≤ (2 ) ZT 2 k=1 sin ( )  2   P  3 3 2 tk 1 4 Proof. The identity limy R(x, y) = (2π)− d (2 k=1 sin ( ))− dt is shown in Theo- →∞ T 2 5 rem 14.9; a computer gives the numerical approximationR P

lim R(x, y) 0.505462 y →∞ ≈ 6 for this integral. While the limit may not exist on the left-hand side of (15.8), we can 7 certainly take the limsup, whence the result follows.

8 Remark 15.10 (Long-range order). In the model of ferromagnetism described above, con- 0 9 sider the collection of spin observables σ located at vertices x G as a system of { x} ∈ 10 non-commutative random variables. One interpretation of the previous results is that in 0 11 KMS-states, the correlations between pairs of vertices x, y G are asymptotically equal ∈ 12 to the resistance distance R(x, y). As mentioned just above, the idea is that correlation 13 between the spin states of x and y remains positive, even when x is arbitrarily far from y. 14 One interpretation of this result is that magnetism can only exist in dimensions 3 and 15 above, or else R(x, y) is unbounded and the estimate (15.9) fails. A different interpreta- 16 tion of the existence of magnetism is the existence of multiple β-KMS states for a fixed 17 temperature T = 1/kβ. This more classical view is quite different.

18 Remark 15.11. We leave it to the reader to ponder the enticing parallel:

2 1 2 u(x) u(y) R(x, y) cxy(u(x) u(y)) , u (Cor. 5.15) 2 x,y | − | ≤ X − ∀ ∈HE 1 ϕ(1 σx σy) R(x, y) cxy(ϕ(1 σx σy)), ϕ KMS states 2 x,y | − · |≤ X − · ∀ ∈ { }

19 15.4. KMS states. While the rigorous deﬁnitions provided in this mini-appendix are not 20 absolutely essential for understanding the Heisenberg model of ferromagnetism, they may 21 help the reader understand what a β-KMS state is, and hence have a better feel for the 22 discussion in the previous section. We suggest the references[BR79,BR97,Rue69,Rue04]

23 for more details. i itH tH 24 Deﬁnition 15.12. Deﬁne α : R Aut(A) by α (A) = e− Ae , for all t R and A A, → t ∈ ∈ 25 where H is a Hamiltonian (as in (15.6) below, for example). This unitary group accounts 26 for time evolution of the system, i.e.,

ψ(t), Aψ(t) = ψ(0), α (A)ψ(0) h i h t i 27 shows that measuring the time-evolved observable αt(A) in the (ground) state ψ0 = ψ(0) is 28 the same as measuring the observable A in the time-evolved state ψ(t).

29 Definition 15.13. Let ϕ be a state as in Definition 15.3. We say ϕ is a KMS state iff for all 30 A, B A, there is a function f with: ∈ 164 PALLEE.T.JORGENSENANDERINP.J.PEARSE

. (1) f is bounded and analytic on z C .. 0 < Im z < β and continuous up to the 1 { ∈ } boundary of this region; 2 (2) f (t) = ϕ(Aα (B)), for all t R; and 3 t ∈

(3) f (t + iβ) = ϕ(α (B)A), for all t R. 4 t ∈ Note that f depends on A and B. This definition is roughly saying that there is an ana- 5 lytic continuation from the graph of ϕ(Aαt(B)) to the graph of ϕ(αt(B)A), where both are 6 considered as functions of t R. 7 ∈ Definition 15.14. If A is finite-dimensional, then 8

trace(e βH A) = = − ϕ(A) ϕβ(A) : βH (15.10) trace(e− ) deﬁnes ϕβ uniquely. In this case, ϕ = ϕβ is called a β-KMS state. 9

Remark 15.15. Let δ be the inﬁnitesimal generator of the ﬂow α : R Aut(A) so that 10 tδ → αt = e . For all β-KMS states ω, Powers established the following a priori estimate 11 in [Pow76a]: 12

2 β

ω([A, B]) ω(A∗A + AA∗)ω( i[B∗,δ(B)]) (15.11) | | ≤ 2 − for all A, B A and B dom δ. 13 ∈ ∈ Remark 15.16 (Long-range order vs. phase transitions). It is excruciatingly important 14 to notice that when A is infinite-dimensional, formula (15.10) becomes meaningless, as 15 was discovered by Bob Powers in his Ph.D. Dissertation [Pow67]; see also [BR97]. The 16 reason for this is somewhat subtle: KMS states should really be formulated in terms the 17 representation of A obtained via GNS construction (see 15.2). Thus, each occurrence 18 § of A in Definition 15.14 should be replaced by πϕ(A) if we are being completely honest. 19 However, in the finite-dimensionalcase, one can use the identity representation and recover 20 (15.10) as it reads above. Unfortunately, the von Neumann algebras generated by KMS 21 states are almost always type III, i.e., the double commutant πϕ(A)′′ typically does not 22 have a trace (even though the C∗-algebra A always does). The von Neumann algebra is the 23 weak- closure of the representation (obtained via GNS construction) of the C -algebra; 24 ∗ ∗ this connection is expressed in the notation of 15.2 by the identity 25 §

ϕ(Aαt(B)) = πϕ(A∗)ζϕ,πϕ(αt(B))ζϕ

h iHϕ i it ϕ t ϕ = πϕ(A∗)ζϕ, e− H πϕ(B)e H ζϕ , h iHϕ where now ϕ in the exponent is an unbounded self-adjoint operator in the Hilbert space 26 H of the GNS representation derived from the state ϕ as in 15.2. 27 § As a consequenceof the lack of trace described just above, there is no uniqueness for ϕβ 28 in general, and this has an important physical interpretation in terms of phase transitions. 29 The parameter β is inverse temperature: β = 1/kT where k is Boltzmann’s constant and T is 30 temperature in degrees Kelvin. Whenever β is a number for which the set of β-KMS states 31 contains more than one element, one says that β corresponds to a phase transition; i.e., 32 T = 1/kβ is a temperature at which more than one equilibrium state can exist. Conversely, 33 “when the system is heated, all is vapor,” and we expect that the equilibrium state ϕ is then 34 unique for β = 1/kT 0. The lowest T for which multiplicity exceeds 1 is called the 35 ≈ critical temperature; it is found experimentally but rigorous results are hard to come by. 36 Indeed, the phase-transition problem in rigorous models is notoriously extremely diﬃcult. 37 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 165

1 Instead the related long-range order problem (as described just below) is thought to be 2 more amenable to computations.

3 To get a feel for why KMS states must exist, consider the following construction. Sup- 0 4 pose we begin with a ﬁnite set F G and the corresponding truncated Hamiltonian ⊆

H := 1 c ϕ(1 σ σ ) A(F). F 2 xy − x · y ∈ xX,y F ∈ 5 Observe that A(F) is ﬁnite-dimensional; for spin observables with spin s, for example, 6 dim(A ) = 2s + 1 for each x. Here, A is a subalgebra of the matrices + (C). Conse- x x M2s 1 7 quently,

trace(e βHF A) F = − ϕβ (A) : βH trace(e− F ) 0 8 is a well-deﬁned and unique β-KMS state. If we now let F G , then ϕβ is a β-KMS state → 9 also. However, ϕβ exists as a weak- limit and hence is not unique! ∗ 166 PALLEE.T.JORGENSENANDERINP.J.PEARSE

16. Future directions 1

Remark 16.1. We have done some groundwork in 7 for the formal construction of the 2 § boundary of an inﬁnite electrical resistance network, however there is much more to be 3 done. The development of this boundary theory is currently underway in [JP08a], where 4 we make explicit the connections between our boundary, Martin boundary, and the theory 5 of graph ends. As in this paper, the notions of dipoles, monopoles, and harmonic functions 6 play key roles. 7

Remark 16.2. In [JP08c], we attempt to apply some results of the present investigation 8 5 to the theory of fractal analysis. For now, we just show that the resistance distance as 9 defined by (5.1) extends to the context of analysis on PCF self-similar fractals. The reader 10 is referred to the definitive text [Kig01] and the excellent tutorial [Str06] for motivation 11 and definitions. 12 Suppose that is a post-critically finite (PCF) self-similar set with an approximating 13 F 0 sequence of graphs G , G ,... G ,... with G = G and is the closure of G in 14 1 m m m m F resistance metric (which is equivalent to closure in EuclidS ean metric; see [Str06, (1.6.10)]). 15 The definition of PCF can be foundin [Kig01, Def. 1.3.4 and Def. 1.3.13]. In the following 16 proof, the subscript m indicates that the relevant quantity is computed on the corresponding 17 electrical resistance network (G , R ). For example, (x, y) is the set of dipoles on G 18 m m Pm m (cf. Definition 3.6) and m(u) is the appropriately renormalized energy of a function u : 19 0 E G R (cf. [Str06, (1.3.20)]). 20 m → Theorem 16.3. For x, y , the resistance distance is given by 21 ∈F . min v(x) v(y) .. v dom , ∆v = δ δ . (16.1) { − ∈ E x − y}

Proof. By the definition of , it suffices to consider the case when x, y are junction points, 22 F that is, x, y G for some m. Then the proof follows for general x, y by taking limits. 23 ∈ m ∈F For x, y G , let v = v denote the element of (x, y) of minimal energy; the exis- 24 ∈ m m Pm tence and uniqueness of v is justified by the results of 3.2. From Theorem 5.2 we have 25 m § R (x, y) = (v ). Next, apply the harmonic extension algorithm to v to obtain v + on 26 m E m m m 1 Gm+1. By[Str06, Lem. 1.3.1], 27

v (x) v (y) = R (x, y) = (v ) = (v + ) = = (˜v), m − m m E m E m 1 ··· E wherev ˜ is the harmonic extension of v to all of G. It is clear by construction and the 28 cited results thatv ˜ minimizes (16.1). Note that we do not need to worry about the possible 29 appearance of nontrivial harmonic functions, asv ˜ is constructed as a limit of functions with 30 ﬁnite support. 31

This theorem oﬀers a practical improvement over the formulation of resistance metric 32 as found in the literature on fractals in a couple of respects: 33

(1) (16.1) providesa formula(or at least, an equationto solve) for the explicit function 34 which gives the minimum in [Str06, (1.6.1) or (1.6.2)]. 35 (2) One can compute v = vm on Gm by basic methods, i.e., Kirchhoﬀ’s law and the 36 cycle condition. To ﬁnd R(x, y), one need only evaluate v(x) v(y), and this may 37 − be done without even fully computing v on all of Gm. 38

5Finally! If you remember from the introduction, this was our initial aim! OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 167

1 Remark 16.4. The authors have uncovereda form of spectral reciprocity relating the Lapla- 2 cian to the matrix [ vx, vy ]. This topic is currently under investigation in [JP08a]. h iE 3 Remark 16.5. The metric graphsand their analysis presented in this memoir are ubiquitous, 4 and we can not do justice to the vast literature. However, the application to quantum 5 communication appears especially intriguing, and we refer to the following papers for 6 detail: [vdNB08, DB07, Fab06, GTHB05, HCDB07].

7 “... the computational power of an important class of quantum states 8 called graph states, representing resources for measurement-based quan- 9 tum computation, is reﬂected in the expressive power of (classical) formal 10 logic languages deﬁned on the underlying mathematical graphs.” from 11 [vdNB08].

12 Quantum graphs (also called cable systems in [Kig03] and graph refinements in [Tel06]) 13 are essentially a refinement of electrical resistance networks where the edges are replaced 14 by intervals and functions are allowed to vary continuously for different values of x in a 15 single edge.

16 Remark 16.6. As noted in Remark 10.32, the rank of Pd⊥ is an invariant related to the space 17 of cycles in G. However, this object is rather a blunt tool, and it would interesting to see 18 if one can obtain a more reﬁned analysis by applying extensions of the techniques Terras 19 and Stark, as in [GILc], for example. 168 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Appendix A. Some functional analysis 1

Since this memoir addresses disparate audiences, we found it helpful to organize tools 2 from functional analysis, unbounded operators, and C∗-algebras in appendix sections. A 3 recurrent theme is that the energy Hilbert space has a number of separate but unitarily 4 HE equivalent incarnations. These different facets are necessary for computing the resistance 5 metric in different ways, and several other applications. The reader may find the references 6 [Arv76b, BR79, DS88] to be helpful. 7

A.1. von Neumann’s embedding theorem. 8

Theorem A.1 (von Neumann). Suppose (X, d) is a metric space. There exists a Hilbert 9 space and an embedding w :(X, d) sending x w and satisfying 10 H →H 7→ x

d(x, y) = wx wy (A.1) k − kH 2 if and only if d is negative semideﬁnite. 11

Definition A.2. A function d : X X R is negative semidefinite iff for any f : X R 12 × → → satisfying x X f (x) = 0, one has 13 ∈ P f (x)d2(x, y) f (y) 0, (A.2) ≤ xX,y F ∈ where F is any finite subset of X. 14

von Neumann’s theorem is constructive, and provides a method for obtaining the em- 15 bedding, which we brieﬂy describe, continuing in the notation of Theorem A.1. 16

Step 1: Schwarz inequality. If d is a negative semidefinite function on X X, then 17 × define a positive semidefinite bilinear form on functions f, g : X C by 18 →

Q( f, g) = f, g := f (x)d2(x, y)g(y). (A.3) h iQ − Xx,y

One obtains a quadratic form Q( f ) := Q( f, f ), and checks that the generalized Schwarz 19 2 inequality holds Q( f, g) Q( f )Q(g) by elementary methods. 20 ≤ Step 2: The kernel of Q. Denote the collection of ﬁnitely supported functions on X 21 by in(X) and deﬁne 22 F

. in (X) := f in(X) .. f (x) = 0 . (A.4) F 0 { ∈F x } P The idea is to complete in (X) with respect to Q, but ﬁrst one needs to identify functions 23 F 0 that Q cannot distinguish. Deﬁne 24

. ker Q = f in (X) .. Q( f ) = 0 . (A.5) { ∈F 0 } It is easy to see that ker Q will be a subspace of in (X). 25 F 0 Step 3: Pass to quotient. Define Q˜ to be the induced quadratic form on the quotient 26 space in (X)/ ker Q. Onemaythenverifythat Q˜ is strictly positive definite on the quotient 27 F 0 space. As a consequence, ϕ := Q˜(ϕ) will be a bona fide norm. 28 k kHvN − OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 169

1 Step 4: Complete. When the quotient space is completed with respect to Q˜, one 2 obtains a Hilbert space

in0(X) ∼ vN := F , with ϕ, ψ = Q˜(ϕ, ψ). (A.6) H ker Q ! h iHvN −

3 Step 5: Embed (X, d) into . Fix some point o X to act as the origin; it will be HvN ∈ 4 mapped to the origin of under the embedding. Then deﬁne HvN

w :(X, d) by x w := 1 (δ δ ). →HvN 7→ x √2 x − o 5 Now w gives an embedding of (X, d) into the Hilbert space HvN, and

w w 2 = w w , w w (A.7) k x − ykvN h x − y x − yivN = w , w w , w + 1 w , w h x xivN − h x yivN 2 h y yivN = d2(x, o) + d2(x, y) d2(x, o) d2(y, o) + d2(y, o) − − = d2(x, y),

6 which verifes (A.1). The third equality follows by three computations of the form

w , w = w (a)d2(a, b)w (b) h x yivN − x y Xa,b = d2(a, b) 1 (δ (a) δ (a))(δ (b) δ (b)) − √2 x − o y − o Xa,b = = d2(x, y) d2(x, o) d2(y, o), (A.8) ··· − − 7 noting that d(a, a) = 0, etc.

8 von Neumann’s theorem also has a form of uniqueness which may be thought of as a 9 universal property.

10 Theorem A.3. If there is another Hilbert space and an embedding k : , with K H → K 11 kx ky = d(x, y) and kx x X dense in , then there exists a unique unitary isomorphism k − kK { } ∈ K 12 U : . H → K 13 Proof. We show that U : w k by U( ξ w ) = ξ k is the required isometric isomor- 7→ x x x x 14 phism. Let ξx = 0. It is conceivableP that U failsP to be well-deﬁned because of linear 15 dependency;P we show this is not the case:

2 ˜ ξxwx = ξxQ(wx, wy)ξy Xx X xX,y X ∈ ∈ 2 2 2 = ξx d (x, y) d (x, o) d (y, o) ξy by (A.8) − − xX,y X ∈ 2 2 2 = ξxξyd (x, y) ξxd (x, o) ξy ξyd (y, o) ξx − − xX,y X Xx X Xy X Xy X Xx X ∈ ∈ ∈ ∈ ∈ = ξ ξ d(x, y), (A.9) − x y xX,y X ∈ 170 PALLEE.T.JORGENSENANDERINP.J.PEARSE since x ξx = 0 by choice of ξ. However, the same computation may be applied to k with 1 the sameP result; note that (A.9) does not depend on w. Thus, w = k and U is an 2 k kH k kK isometry. Since it is an isometry from a dense set in to a dense set in , we have an 3 H K isomorphism and are ﬁnished. 4

The importance of using in (X) in the above construction is that the ﬁnitely supported 5 F 0 functions in(X) are in duality with the bounded functions B(X) via 6 F

f, β := f (x)β(x) < f in(X), β B(X). (A.10) h i ∞ ∈F ∈ Xx X ∈ The constant function β1 := 1 is a canonical bounded function. With respect to the pairing 7 in (A.10), its orthogonal complement is 8

. β⊥ = ϕ .. ϕ, β = 0 = in (X). 1 { h 1i } F 0

Appendix B. Some operator theory 9 Definition B.1. If S : is an operator on a Hilbert space , its adjoint is the H →H H operator satisfying S ∗u, v = u, Sv for every v dom S . The restriction to v dom S becomes significanth only wheni hS is unbounded,i in∈ which case one sees that the domain∈ of the adjoint is defined by . dom S ∗ := u .. u, Sv k v , v dom S , (B.1) { ∈H |h i| ≤ k k ∀ ∈ } where the constant k = k C may depend on u. 10 u ∈ An operator S is said to be self-adjoint iff S = S ∗ and dom S = dom S ∗. It is often the 11 equality of domains that is harder to check. 12

B.1. Projections and closed subspaces. Let be a complex (or real) Hilbert space, and 13 H deﬁne 14

. ( ) := A : .. A is bounded and linear B H { H→H } .. ( )sa := A ( ) . A = A∗ B H { ∈B H. } Cl( ) := V .. V is a closed linear space . H { ⊆H } 2 Definition B.2. An operator P on a Hilbert space is a projection iff it satisfies P = P = 15 H P . Denote the space of projections by Proj( ). 16 ∗ H Theorem B.3 (Projection Theorem). There is a bijective correspondence between the set 17 Cl( ) of all closed subspaces V and the set of all projections P acting on . 18 H ⊆H H Specifically, if a subspace V Cl( ) is given, there is a unique projection P Proj( ) 19 ∈ H ∈ H with range L. Conversely, if a projection P is given, set V := P . Moreover, the mapping 20 H

Proj( ) Cl( ), by P V = P H → H 7→ H is a lattice isomorphism. The ordering in Cl( ) is deﬁned by containment; and for a pair 21 H of projections P and Q we say that P Q iﬀ P = PQ. This ordering coincides with the 22 ≤ usual order on the self-adjoint elements of the algebra: 23

A B in ( ) iﬀ v, Av v, Bv , v , ≤ B H h i≤ i ∀ ∈H OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 171

1 and induces an ordering on Proj( ). Since Proj( ) is a lattice, it follows that Cl( )isa H H H 2 lattice as well.

3 Theorem B.4 ([KR97, Prop. 2.5.2]). For projections P, Q, the following are equivalent:

4 (i) P Q . H ⊆ H 5 (ii) P = PQ. 6 (iii) v, Pv v, Qv , v . h i ≤ h i ∀ ∈H 7 (iv) Pv Qv , v . k k ≤ k k ∀ ∈H 8 B.2. Partial isometries. Let and be two complex (or real) Hilbert spaces, and let H K 9 L : be a bounded linear operator. H → K 10 Deﬁnition B.5. L is a partial isometry if one (all) of the following equivalent conditions 11 is satisﬁed:

12 (i) L L is a projection in (the initial projection). ∗ H 13 (ii) LL is a projection in (the ﬁnal projection). ∗ K 14 (iii) LL∗ L = L. 15 (iv) L∗LL∗ = L∗.

16 In this case, we say that Pi := L∗L is the initial projection and P f := LL∗ is the ﬁnal 17 projection. Moreover, Pi is the projection onto ker(L)⊥ and P f is the projection onto the 18 closed subspace ran L.

19 Theorem B.6. The four conditions of Deﬁnition B.5 are equivalent. Consequently, the 20 initial and ﬁnal projections satisfy LPi = L = P f L.

21 Proof. Deﬁne P := L L. Compute that (LP L) (LP L) = 0 to deduce LP = L; the ∗ − ∗ − 22 reader can ﬁll in the rest.

23 Definition B.7. The operator L is an isometry iff Pi = I , the identity operator on . The H H 24 operator L is a coisometry iff P f = I , the identity operator on . It is clear that L is an K K 25 isometry if and only if L∗ is a coisometry.

26 B.3. Self-adjointness for unbounded operators. Throughout this section, we use to D 27 denote a dense subspace of . Some good references for this material are [vN32a,Rud91, H 28 DS88]. Deﬁnition B.8. An operator S on is called Hermitian (or symmetric or formally self- adjoint) iﬀ H u, Sv = Su, v , for every u, v . (B.2) h i h i ∈ D 29 In this case, the spectrum of S lies in R.

30 Definition B.9. An operator S on is called self-adjoint iff it is Hermitian and dom S = H 31 dom S ∗. Definition B.10. An operator S on is called semibounded iff H v, Sv 0, for every v . (B.3) h i≥ ∈ D 32 The spectrum of a semibounded operator lies in some halfline [κ, ) and the defect indices ∞ 33 of a semibounded operator always agree (see Definition B.17). The graph Laplacian ∆ 34 considered in much of this paper falls into this class. 172 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Definition B.11. An operator S on is called bounded iff there exists k R such that H ∈ v, Sv k v 2, for every v . (B.4) |h i| ≤ k k ∈ D The spectrum of a boundedoperatorlies in a compact subinterval of R. Bounded Hermitian 1 operators are automatically self-adjoint. When (G, Ω) satisfies the Powers bound (9.14), 2 the transfer operator falls into this class. 3

Deﬁnition B.12. For an operator S on the Hilbert space , the graph of S is 4 H v . (S ) := .. v , (B.5) G { Av ∈H}⊆H⊕H with the norm 5

v 2 = 2 + 2 Sv Graph : v Sv (B.6) k kH k kH and the corresponding inner product. The operator S is closed iﬀ (S ) is closed in 6 G H⊕H or closable if the closure of (S ) is the graph of an operator. In this case, the corresponding 7 clo G clo operator is S , the closure of S . The domain of S is therefore deﬁned 8

clo .. dom S := u . lim u un = lim v Sun = 0 (B.7) n H n H { →∞ k − k →∞ k − k } clo for some v and Cauchy sequence un dom S . Then one deﬁnes S u := v. If S is 9 ∈ H clo { } ⊆ Hermitian, then S will also be Hermitian, but it will not be self-adjoint in general. 10

Remark B.13. It is important to observe that an operator S is closable if and only if S ∗ has 11 dense domain. However, this is clearly satisfied when S is Hermitian with dense domain, 12 since then dom S dom S . 13 ⊆ ∗ Definition B.14. Suppose that S is a linear operator on with a dense domain dom S . Define the graph rotation operator G : Hby G(u, v) := ( v, u). It is easy H⊕H→H⊕H − to show that the graph of S ∗ is

(S ∗) = (G( (S )))⊥. (B.8) G G For any semibounded operator S on a Hilbert space, there are unique self-adjoint extensions S min (the Friedrichs extension) and S max (the Krein extension) such that S S clo S S˜ S , (B.9) ⊆ ⊆ min ⊆ ⊆ max where S˜ is any non-negative self-adjoint extension of S . For general unbounded operators, 14 these inclusions may all be strict. In (B.9), A B means graph containment, i.e., it means 15 ⊆ (A) (B), where is as in Deﬁnition B.12. The case when S = S is particularly 16 G ⊆ G G min max important. 17

Definition B.15. An operator is defined to be essentially self-adjoint iff it has a uniqueself- 18 adjoint extension. An operator is essentially self-adjoint if and only if it has defect indices 19 0,0 (see Definition B.17). A self-adjoint operator is trivially essentially self-adjoint. 20

Theorem B.16. [vN32a, Rud91, DS88] Let S be a Hermitian operator. 21 clo clo (1) S is closable, its closure S is Hermitian, and S ∗ = (S )∗. 22 (2) Every closed Hermitian extension T of S satisﬁes

S T T ∗ S ∗. ⊆ ⊆ ⊆ clo (3) S is essentially self-adjoint if and only if dom(S ) = dom S ∗. 23 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 173

1 (4) S is essentially self-adjoint precisely when both its defect indices are 0. 2 (5) S has self-adjoint extensions iﬀ S has equal defect indices.

3 The Hermitian operator S := QPQ of Example B.25 has defect indices 1,1, and yet is 4 not even semibounded.

5 Deﬁnition B.17. Let S be an operator with adjoint S . For λ C, deﬁne ∗ ∈ . Defλ(S ) := ker(λ S ∗) = v dom S ∗ .. S ∗v = λv . (B.10) − { ∈ } 6 Then Defλ(S ) is the defect space of S corresponding to λ. Elements of Defλ(S ) are called 7 defect vectors. The number dimDefλ(S ) is constant in the connected components of the 8 resolvent set C σ(S ) and is called the defect index of the component containing λ. \ 9 Note that if S is Hermitian, then its resolvent set can have at most two connected com- 10 ponents. Further, if S is semibounded, then its resolvent set can have only one con- 11 nected component, and we have only one defect index to compute: the dimension of 12 Def(S ) = Def 1(S ). These facts explain the two consequences of the following theorem, − 13 which can be found in [vN32a, Rud91, DS88].

14 Theorem B.18 (von Neumann). For a Hermitian operator S on , one has H

clo .. .. i dom S ∗ = dom S v . S ∗v = iv v . S ∗v = v , ⊕ { ∈H } ⊕ { ∈H − } 15 where the orthogonality of the direct sum on the right-hand side is with respect to the 16 graph inner product (B.6) (not the inner product of ). Consequently, S is essentially H 17 self-adjoint if and only if

S ∗v = iv = v = 0, v . (B.11) ± ⇒ ∈H 18 For a semibounded Hermitian operator S on , one has H

clo . dom S ∗ = dom S v .. S ∗v = v , ⊕ { ∈H − } 19 where again the orthogonality of the direct sum on the right-hand side is with respect to 20 the graph inner product (B.6). Consequently, S is essentially self-adjoint if and only if

S ∗v = v = v = 0, v . (B.12) − ⇒ ∈H 21 A solution v of (B.11) or(B.12) is called a defect vector (as in Deﬁnition B.17) or an 22 vector at . The idea of the proof in von Neumann’s theorem is to obtain the essential ∞ 23 self-adjointness of a Hermitian operator S by using the following stratagem: an unbounded 24 function applied to a bounded self-adjoint operator is an unbounded self-adjoint operator. 1 1 25 In this case, the function is f (x) = λ x . If we can see that(λ S ) is bounded and − − − min − 26 self-adjoint, then

1 f ((λ S )− ) = S − min min 27 is an unbounded self-adjoint operator. First, note that

[ran(λ¯ S )] = ran(λ¯ S ) = ker(λ S ∗)⊥ = Defλ(S )⊥. − − min − 1 28 Note that if λ res(S ) and (λ S ) = xE(dx) with projection-valued measures ∈ − min − R 29 E : (R) Proj( ), then R B → H 174 PALLEE.T.JORGENSENANDERINP.J.PEARSE

1 S min = (λ x− ) E(dx). ZR −

This will show S min is self-adjoint; if Defλ(S ) = 0 for “enough” λ, then S min is self-adjoint 1 and hence S is essentially self-adjoint. 2

Lemma B.19. If S is bounded and Hermitian, then it is essentially self-adjoint. 3

Proof. S is bounded iff it is everywhere defined, by the Hellinger-Toeplitz Theorem. Since 4 S ∗ is also everywhere defined in this case, it is clear the two operators have the same 5 domain. 6

Lemma B.20. For an operator S which is semibounded but not necessarily closed, 7

clo Def(S )⊥ = ran(1 + S ). (B.13)

Proof. Recall that Def(S ) = Def 1(S ). General theory gives 8 −

clo Def(S )⊥ = (ker1 + S ∗)⊥ = (ran(1 + S ))⊥⊥ = (ran(1 + S )) . clo clo It remains to check that ran(1 + S ) = (ran(1 + S )) . If S is not semibounded, one may 9 clo clo have only the containment ( ): note that u ran(1 + S ) iﬀ u = v + Sv for v dom S . 10 ⊆ ∈ ∈ Then v = lim vn for some vn dom S , and the containment is clear. For ( ), let u 11 clo ∈ ⊇ ∈ ran(1 + S ) so that u = lim u , where u = v Sv for v dom S , and note that 12 n n n n n ∈

v 2 v 2 + Sv , v + v , Sv + Sv 2 = v + Sv 2 = u 2, (B.14) k nk ≤ k nk h n ni h n ni k nk k n nk k nk where the inequality uses the fact that S is semibounded. By passing to a subsequence if 13 necessary, (B.14) implies v v u u , whence u Cauchy implies v is also 14 k n − mk ≤ k n − mk { n} { n} Cauchy and hence has a limit v. Therefore 15

n Sv = u v →∞ u v, n n − n −−−−−−→ − clo clo which allows one to deﬁne S v = u v and see u ran(1 + S ). 16 − ∈ Example B.21 (The defect of the Laplacian on (0, )). Probably the most basic example 17 ∞ of defect vectors (and how an Hermitian operator can fail to be essentially self-adjoint) is 18 d2 2 provided by the Laplace operator ∆ = on the Hilbert space = L (0, ). Exam- 19 − dx2 H ∞ ple B.25 gives an even more striking (though less simple) example. We take ∆ as having 20 the dense domain 21

. (k) (k) = f C∞(0, ) .. f (0) = lim f (x) = 0, k = 0, 1, 2,... . 0 x D { ∈ ∞ →∞ } x We always have u, ∆u 0 for u . However, ∆ u = u is satisﬁed by e . 22 h i ≥ ∈ D ∗ − − ∈H\D To see this, take any test function ϕ and compute the weak derivative via integration 23 ∈ D by parts (applied twice): 24

x ∆ = x d2 = 2 d2 x = x e− , ϕ e− , dx2 ϕ ( 1) dx2 e− , ϕ e− , ϕ . h i D − E − − D E −h i Thus the domain of ∆∗ is strictly larger than ∆˜ , and so ∆˜ fails to be self-adjoint. 25 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 175

1 One might try the approximation argument used to prove essential self-adjointness of 2 x 2 the Laplacian on ℓ (c) in Theorem 9.2: let v be a sequence with v e 0. { n} ⊆ D k n − − kc → 3 Since ∆ agrees with ∆ when restricted to , ∗ D

x x v , ∆v = v , ∆∗v e− , ∆∗e− . h n nic h n nic → h ic 4 However, there are two mistakes here. First, one does not have convergence unless the 5 original sequence is chosen so as to approximate v in the nonsingular quadratic form

u, v ∆ := u, v + u, ∆∗v . h i ∗ h i h i x 6 Second, one cannot approximate e− with respect to this nonsingular quadratic form by x 7 elements of . In fact, e is orthogonal to in this sense: D − D

x x x x x ϕ, e− ∆ = ϕ, e− + ϕ, ∆e− = ϕ, e− ϕ, e− = 0. h i ∗ h ic h ic h ic − h ic 8 Alternatively, observe that von Neumann’s theorem (Theorem B.18), a general element 9 in the domain of ∆∗ is v + ϕ , where v dom ∆ and ϕ is in the defect space. ∞ ∈ ∞ 10 Suppose we have an exhaustion with + and = (this {Hk} Hk ⊆ Hk 1 ⊆ D H Hk 11 notation indicates closed linear span of the union). W

k (x+1/x) . , = [x e− .. n k m, for n, m N]. Hn m − ≤ ≤ ∈ 12 We have that ∆ , + , , since Hn m ⊆Hn 3 m

d2 k (x+1/x) k (x+1/x) 1 2(k 1) 2 k+k2 2k x e− = x e− + − + − − + 1 . dx2 x4 x3 x2 − x x 13 However, ∆ u = u is satisﬁed by e . ∗ − − ∈H\D 14 B.4. Banded matrices.

15 Definition B.22. Consider the matrix MS corresponding to an operator S in some ONB 16 b , so the entries of M are given by { x} S M (x, y) := b , Sb . (B.15) S h x yi 17 We say MS is a banded matrix iff every row and column contains only finitely many 18 nonzero entries. A fortiori, MS is uniformly banded if no row or column has more than N 19 nonzero entries, for some N N. ∈ 20 With MS defined as in (B.15), it is immediate that MS is Hermitian whenever S is:

M (x, y) = b , Sb = Sb , b = b , Sb = M (y, x). (B.16) S h x yi h x yi h y xi S 21 Banded matrices are of interest in the present context because the graph Laplacian is 22 always a banded matrix in virtue of the fact that each vertex has only ﬁnitely many neigh- 23 bours; recall the form of M∆ given in (9.10). Since ∆ = c T, the transfer operator T − 24 is also banded. In general, the bandedness of an operator does not imply the operator is 2 25 self-adjoint. In fact, see Example B.25 for a Hermitian operator on ℓ which is not self- 26 adjoint, despite having a uniformly banded matrix. However, this property does make it 27 much easier to compute the adjoint. 176 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Lemma B.23. Let S be an unbounded Hermitian operator on with dense domain of 1 H definition = dom S . Suppose that the matrix M defines as in (B.15) is banded 2 D ⊆ H S with respect to the ONB bx x X, and define vˆ(x) := bx, v . Then v dom S ∗ and S ∗v = w 3 2 { } ∈ h i ∈ if and only if vˆ, wˆ ℓ (X) and wˆ (x) is 4 ∈

wˆ (x) = MS (x, y)ˆv(y). (B.17) Xy X ∈ Thus, S ∗ is represented by the banded matrix MS . 5

Proof. ( ) To see the form ofw ˆ in (B.17), 6 ⇒

M (x, y)ˆv(y) = b , Sb b , v = Sb , b b , v Hermitian S h x yih y i h x yih y i Xy X Xy X Xy X ∈ ∈ ∈ = Sb , v Parseval h x i = b , w S ∗v = w. h x i where the last equality is possible since v dom S ∗. It is the hypothesis of bandedness that 7 ∈ guarantees all these sums are finite, and hence meaningful. 8 Conversely, first note that it is the hypothesis of bandedness which makes the sum in 9 2 (B.17) finite, ensuringw ˆ is well-defined. suppose (B.17) holds, and thatv ˆ, wˆ ℓ (X). To 10 ∈ show v dom S , we must find a constant K < for which v, Su K u for every 11 ∈ ∗ ∞ |h i| ≤ k k u . 12 ∈ D

v, Su = v, b b , Su Parseval h i h xih x i Xx X ∈ = v, b Sb , u h xih x i Xx X ∈ = vˆ(x) MS (y, x)by, u y Xx X X ∈ = vˆ(x)MS (y, x)ˆu(y) by(B.16) Xy X Xx X ∈ ∈ 2 v, Su 2 vˆ(x)M (x, y) uˆ(y) 2 by Schwarz, |h i| ≤ S | | Xy X Xx X Xy X ∈ ∈ ∈ 2 and see that we can take K = wˆ . 13 k k 2 Lemma B.24. Let A be an operator on ℓ (Z) whose matrix MA is uniformly banded, with 14 all bands having no more than β nonzero entries. Then 15

A β sup axy . (B.18) k k≤ x,y | |

Proof. The Schwartz inequality gives 16

1/2 / 2 2 1 2 A max sup axy , sup axy x y y x k k≤ ( X | | X | | ) However, uniform bandedness gives 17 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 177

1/2 2 sup axy β sup max axy ,  | |  ≤ y | | x Xy  x   1 and similarly for the other term. 

2 Example B.25 (Two operators which are each self-adjoint, but whose product is not es- 3 sentially self-adjoint). In 7.2, we discussed the Schwartz space S of functions of rapid § 4 decay, and its dual S ′, the space of tempered distributions; cf. (7.4). Ifwe use theONB for 2 1 d 5 L (R) consisting of the Hermite polynomials, then the operators P˜ : f (x) f (x) and 7→ i dx 6 Q˜ : f (x) x f (x) have the following matrix form: 7→ 0 1  1 0 √2   √2 0 √3     . .  1  √3 .. ..  P =   , (B.19) 2  .   .. 0 √n     .   √n 0 ..     .. ..   . .     0 1  1 0 √2  −   √2 0 √3   −   . .  1  √3 .. ..  =   Q  −  . (B.20) 2i  .   .. 0 √n     .   √n 0 ..     − . .   .. ..      7 P and Q are Heisenberg’s matrices, and they satisfy the canonical commutation relation

i 2 8 PQ QP = I. P and Q provide examples of Hermitian operators on ℓ (Z) which are each − 2 9 essentially self-adjoint, but for which T = QPQ is not essentially self-adjoint. In fact, T 10 has defect indices 1,1 (cf. Deﬁnitions B.17 and B.15). These indices are found directly by 11 solving the the defect equation

T ∗ f = QPQf = i f = x(x f )′ = f ± ⇒ ± to obtain the C∞ solutions 1/x 1/x e− e x , x > 0, x , x < 0, f+(x) = f (x) = 0, x 0, − 0, x 0.  ≤  ≥ 12 Thus there is a 1-dimensional space of solutions to each defect equation, and the defect 13 indices are 1,1. To see that P and Q are actually self-adjoint, one can observe that P gen-

itx 14 erates the unitary group f (x) f (x + t) and Q generates the unitary group e . Therefore, 7→ 15 P and Q are self-adjoint by Stone’s theorem; see [DS88]. 178 PALLEE.T.JORGENSENANDERINP.J.PEARSE

Appendix C. Navigation aids for operators and spaces 1 C.1. A road map of spaces. Each arrow represents an embedding.

2 span δx / ℓ (c) / in { } F

span vx / G / / G′ { } S HO E SO

SG′ arm / in H F

C.2. A summary of the operators on various Hilbert spaces. 2 3 c unbdd c bdd c = 1 ∆ on unbdd, Herm, poss. defect unbdd, Herm, ess. s.-a. unbdd, Herm, ess. s.-a. ∆ on ℓ2H(1E) unbdd, Herm, ess. s.-a. bdd, s.-a. 2 ∆ on ℓ (c) unbdd, non-Herm non-Herm bdd, s.-a. 4 T on unbdd, Herm, poss. defect unbdd, Herm, ess. s.-a. unbdd, Herm, ess. s.-a. T on ℓ2H(1E) unbdd, Herm, ess. s.-a. bdd, s.-a. T on ℓ2(c) unbdd, non-Herm non-Herm bdd, s.-a. OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 179

1 References

2 [AAL08] Daniel Alpay, Haim Attia, and David Levanony. Une généralisation de l’intégrale stochastique de 3 Wick-Itô. C. R. Math. Acad. Sci. Paris, 346(5-6):261–265, 2008. 4 [ABR07] Ron Aharoni, Eli Berger, and Ziv Ran. Independent systems of representatives in weighted graphs. 5 Combinatorica, 27(3):253–267, 2007. 6 [AC04] Esteban Andruchow and Gustavo Corach. Differential geometry of partial isometries and partial 7 unitaries. Illinois J. Math., 48(1):97–120, 2004. 8 [AD06] D. Alpay and C. Dubi. Some remarks on the smoothing problem in a reproducing kernel Hilbert 9 space. J. Anal. Appl., 4(2):119–132, 2006. 10 [AG92] M. A. Al-Gwaiz. Theory of distributions, volume 159 of Monographs and Textbooks in Pure and 11 Applied Mathematics. Marcel Dekker Inc., New York, 1992. 12 [AH05] Kendall Atkinson and Weimin Han. Theoretical numerical analysis, volume 39 of Texts in Applied 13 Mathematics. Springer, New York, second edition, 2005. A functional analysis framework. 14 [AL08] Daniel Alpay and David Levanony. On the reproducing kernel Hilbert spaces associated with the 15 fractional and bi-fractional Brownian motions. Potential Anal., 28(2):163–184, 2008. 16 [Aro50] N. Aronszajn. Theory of reproducing kernels. Trans. Amer. Math. Soc., 68:337–404, 1950. 17 [Arv76a] William Arveson. Aspectral theorem for nonlinear operators. Bull. Amer. Math. Soc., 82(3):511–513, 18 1976. 19 [Arv76b] William Arveson. An invitation to C∗-algebras. Springer-Verlag, New York, 1976. Graduate Texts 20 in Mathematics, No. 39. 21 [Arv76c] William Arveson. Spectral theory for nonlinear random processes. In Symposia Mathematica, Vol. 22 XX (Convegno sulle Algebre C∗ e loro Applicazioni in Fisica Teorica, Convegno sulla Teoria degli 23 Operatori Indice e Teoria K, INDAM, Rome, 1975), pages 531–537. Academic Press, London, 1976. 24 [Arv02] William Arveson. A short course on spectral theory, volume 209 of Graduate Texts in Mathematics. 25 Springer-Verlag, New York, 2002. 26 [BBI01] Dmitri Burago, Yuri Burago, and Sergei Ivanov. A course in metric geometry, volume 33 of Graduate 27 Studies in Mathematics. American Mathematical Society, Providence, RI, 2001. 28 [BCR84] Christian Berg, Jens Peter Reus Christensen, and Paul Ressel. Harmonic analysis on semigroups, 29 volume 100 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1984. Theory of positive 30 definite and related functions. 31 [BD49] R. Bott and R. J. Duffin. Impedance synthesis without use of transformers. J. Appl. Phys., 20:816, 32 1949. 33 [Bea91] Alan F. Beardon. Iteration of rational functions, volume 132 of Graduate Texts in Mathematics. 34 Springer-Verlag, New York, 1991. Complex analytic dynamical systems. 35 [Ber96] Christian Berg. Moment problems and polynomial approximation. Ann. Fac. Sci. Toulouse Math. 36 (6), (Special issue):9–32, 1996. 100 ans après Th.-J. Stieltjes. 37 [BHS05] Michael Barnsley, John Hutchinson, and Orjan¨ Stenflo. A fractal valued random iteration algorithm 38 and fractal hierarchy. Fractals, 13(2):111–146, 2005. 39 [BJ99] Ola Bratteli and Palle E. T. Jorgensen. Iterated function systems and permutation representations of 40 the Cuntz algebra. Mem. Amer. Math. Soc., 139(663):x+89, 1999. r 41 [BJMP05] Lawrence Baggett, Palle Jorgensen, Kathy Merrill, and Judith Packer. A non-MRA C frame wavelet 42 with rapid decay. Acta Appl. Math., 89(1-3):251–270 (2006), 2005. 43 [BM00] Michael Baake and Robert V. Moody, editors. Directions in mathematical quasicrystals, volume 13 44 of CRM Monograph Series. American Mathematical Society, Providence, RI, 2000. 45 [BM01] Michael Baake and Robert V. Moody. Self-similarities and invariant densities for model sets. In 46 Algebraic methods in physics (Montréal, QC, 1997), CRM Ser. Math. Phys., pages 1–15. Springer, 47 New York, 2001. 48 [Bol98] Béla Bollobás. Modern graph theory, volume 184 of Graduate Texts in Mathematics. Springer- 49 Verlag, New York, 1998. 50 [BR79] Ola Bratteli and Derek W. Robinson. Operator algebras and quantum statistical mechanics. Vol. 1. 51 Springer-Verlag, New York, 1979. C∗- and W∗-algebras, algebras, symmetry groups, decomposition 52 of states, Texts and Monographs in Physics. 53 [BR97] Ola Bratteli and Derek W. Robinson. Operator algebras and quantum statistical mechanics. Vol. 2. 54 Texts and Monographs in Physics. Springer-Verlag, Berlin, second edition, 1997. Equilibrium states. 55 Models in quantum statistical mechanics. 56 [BS08] Susanne C. Brenner and L. Ridgway Scott. The mathematical theory of finite element methods, vol- 57 ume 15 of Texts in Applied Mathematics. Springer, New York, third edition, 2008. 180 PALLEE.T.JORGENSENANDERINP.J.PEARSE

[BV03] Joseph A. Ball and Victor Vinnikov. Formal reproducing kernel Hilbert spaces: the commutative 1 and noncommutative settings. In Reproducing kernel spaces and applications, volume 143 of Oper. 2 Theory Adv. Appl., pages 77–134. Birkhäuser, Basel, 2003. 3 [Car72] P. Cartier. Fonctions harmoniques sur un arbre. In Symposia Mathematica, Vol. IX (Convegno di 4 Calcolo delle Probabilità, INDAM, Rome, 1971), pages 203–270. Academic Press, London, 1972. 5 [Car73a] P. Cartier. Harmonic analysis on trees. In Harmonic analysis on homogeneous spaces (Proc. Sympos. 6 Pure Math., Vol. XXVI, Williams Coll., Williamstown, Mass., 1972), pages 419–424. Amer. Math. 7 Soc., Providence, R.I., 1973. 8 [Car73b] Pierre Cartier. Géométrie et analyse sur les arbres. In Séminaire Bourbaki, 24ème année (1971/1972), 9 Exp. No. 407, pages 123–140. Lecture Notes in Math., Vol. 317. Springer, Berlin, 1973. 10 [CdV98] Yves Colin de Verdière. Spectres de graphes, volume 4 of Cours Spécialisés [Specialized Courses]. 11 SociétéMathématique de France, Paris, 1998. 12 [CdV99] Yves Colin de Verdière. Spectre d’opérateurs différentiels sur les graphes. In Random walks and dis- 13 crete potential theory (Cortona, 1997), Sympos. Math., XXXIX, pages 139–164. Cambridge Univ. 14 Press, Cambridge, 1999. 15 [CdV04] Yves Colin de Verdière. Sur le spectre des opérateurs de type Schrödinger sur les graphes. In 16 Graphes, pages 25–52. Ed. Ec.´ Polytech., Palaiseau, 2004. 17 [Cho08] Ilwoo Cho. Measures on graphs and groupoid measures. Complex Analysis and Operator Theory, 18 2(1):1–28, 2008. 19 [Chu01] Fan Chung. Spectral Graph Theory. 2001. 20 [Chu07] Fan Chung. Random walks and local cuts in graphs. Linear Algebra Appl., 423(1):22–32, 2007. 21 + [CKL 08] R. R. Coifman, I. G. Kevrekidis, S. Lafon, M. Maggioni, and B. Nadler. Diffusion maps, reduction 22 coordinates, and low dimensional representation of stochastic systems. Multiscale Model. Simul., 23 7(2):842–864, 2008. 24 [CM06] Ronald R. Coifman and Mauro Maggioni. Diffusion wavelets. Appl. Comput. Harmon. Anal., 25 21(1):53–94, 2006. 26 [Coh80] Donald L. Cohn. Measure theory. Birkhauser, Boston, 1980. 27 [Con94] Alain Connes. Noncommutative geometry. Academic Press Inc., San Diego, CA, 1994. 28 [CR06] Fan Chung and Ross M. Richardson. Weighted Laplacians and the sigma function of a graph. In 29 Quantum graphs and their applications, volume 415 of Contemp. Math., pages 93–107. Amer. Math. 30 Soc., Providence, RI, 2006. 31 [CW92] Donald I. Cartwright and Wolfgang Woess. Infinite graphs with nonconstant Dirichlet finite harmonic 32 functions. SIAM J. Discrete Math., 5(3):380–385, 1992. 33 [CW07] Donald I. Cartwright and Wolfgang Woess. The spectrum of the averaging operator on a network 34 (metric graph). Illinois J. Math., 51(3):805–830, 2007. 35 [DB07] W. Dür and H. J. Briegel. Entanglement purification and quantum error correction. Rep. Progr. Phys., 36 70(8):1381–1424, 2007. 37 [Dev07] Robert L. Devaney. Cantor sets of circles of Sierpiński curve Julia sets. Ergodic Theory Dynam. 38 Systems, 27(5):1525–1539, 2007. 39 [Die06] Reinhard Diestel. End spaces and spanning trees. J. Combin. Theory Ser. B, 96(6):846–854, 2006. 40 [DJ06] Dorin E. Dutkay and Palle E. T. Jorgensen. Methods from multiscale theory and wavelets applied 41 to nonlinear dynamics. In Wavelets, multiscale systems and hypercomplex analysis, volume 167 of 42 Oper. Theory Adv. Appl., pages 87–126. Birkhäuser, Basel, 2006. 43 [DJ07] Dorin E. Dutkay and Palle E. T. Jorgensen. Analysis of orthogonality and of orbits in affine iterated 44 function systems. Math. Z., 256(4):801–823, 2007. 45 [DJ08] Dorin E. Dutkay and Palle E. T. Jorgensen. Spectral theory for discrete lapacians. 46 http://arxiv.org/abs/0802.2347, 2008. 47 [DK88] Jozef Dodziuk and Leon Karp. Spectral and function theory for combinatorial Laplacians. In Ge- 48 ometry of random motion (Ithaca, N.Y., 1987), volume 73 of Contemp. Math., pages 25–40. Amer. 49 Math. Soc., Providence, RI, 1988. 50 [DK03] Reinhard Diestel and Daniela Kühn. Graph-theoretical versus topological ends of graphs. J. Combin. 51 Theory Ser. B, 87(1):197–206, 2003. Dedicated to Crispin St. J. A. Nash-Williams. 52 [Dod06] Józef Dodziuk. Elliptic operators on infinite graphs. In Analysis, geometry and topology of elliptic 53 operators, pages 353–368. World Sci. Publ., Hackensack, NJ, 2006. 54 [Doo53] J. L. Doob. Stochastic processes. John Wiley & Sons Inc., New York, 1953. 55 [Doo55] J. L. Doob. Martingales and one-dimensional diffusion. Trans. Amer. Math. Soc., 78:168–208, 1955. 56 [Doo58] J. L. Doob. Probability theory and the first boundary value problem. Illinois J. Math., 2:19–36, 1958. 57 [Doo59] J. L. Doob. Discrete potential theory and boundaries. J. Math. Mech., 8:433–458; erratum 993, 1959. 58 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 181

1 [Doo66] J. L. Doob. Remarks on the boundary limits of harmonic functions. SIAM J. Numer. Anal., 3:229– 2 235, 1966. 3 [DS84] Peter G. Doyle and J. Laurie Snell. Random walks and electirc networks. Carus Monograph. Mathe- 4 matical Association of America, 1984. 5 [DS88] Nelson Dunford and Jacob T. Schwartz. Linear operators. Part II. Wiley Classics Library. John 6 Wiley & Sons Inc., New York, 1988. Spectral theory. Selfadjoint operators in Hilbert space, With 7 the assistance of William G. Bade and Robert G. Bartle, Reprint of the 1963 original, A Wiley- 8 Interscience Publication. 9 [Epi66] G. V. Epifanov. Reductin of a plane graph to an edge by star-triangle transformations. Dokl. Akad. 10 Nauk SSSR, 166:19–22, 1966. 11 [Fab06] X. W. C. Faber. Spectral convergence of the discrete Laplacian on models of a metrized graph. New 12 York J. Math., 12:97–121 (electronic), 2006. 13 [FK07] Enrico Formenti and Petr K˘urka. A search algorithm for the maximal attractor of a cellular automa- 14 ton. In STACS 2007, volume 4393 of Lecture Notes in Comput. Sci., pages 356–366. Springer, Berlin, 15 2007. 16 [FMY05] Cynthia Farthing, Paul S. Muhly, and Trent Yeend. Higher-rank graph C∗-algebras: an inverse semi- 17 group and groupoid approach. Semigroup Forum, 71(2):159–187, 2005. 18 [FOT94]¯ Masatoshi Fukushima, Y¯oichi Oshima,¯ and Masayoshi Takeda. Dirichlet forms and symmetric 19 Markov processes, volume 19 of de Gruyter Studies in Mathematics. Walter de Gruyter & Co., 20 Berlin, 1994. 21 [GILa] Daniele Guido, Tommaso Isola, and Michel L. Lapidus. Ihara zeta functions for periodic simple 22 graphs. 23 [GILb] Daniele Guido, Tommaso Isola, and Michel L. Lapidus. Ihara’s zeta function for periodic graphs and 24 its approximation in the amenable case. 25 [GILc] Daniele Guido, Tommaso Isola, and Michel L. Lapidus. A trace on fractal graphs and the ihara zeta 26 function. 27 [GMS58]ˇ I. M. Gel’fand, R. A. Minlos, and Z. Ja. Sapiro.ˇ Predstavleniya gruppy vrashcheni i gruppy Lorentsa, 28 ikh primeneniya. Gosudarstv. Izdat. Fiz.-Mat. Lit., Moscow, 1958. 29 [Gre69] Frederick P. Greenleaf. Invariant means on topological groups and their applications. Van Nostrand 30 Mathematical Studies, No. 16. Van Nostrand Reinhold Co., New York, 1969. 31 [Gro67] Leonard Gross. Potential theory on Hilbert space. J. Functional Analysis, 1:123–181, 1967. 32 [Gro70] Leonard Gross. Abstract Wiener measure and infinite dimensional potential theory. In Lectures in 33 Modern Analysis and Applications, II, pages 84–116. Lecture Notes in Mathematics, Vol. 140. 34 Springer, Berlin, 1970. 35 [GTHB05] Otfried Gühne, Géza Tóth, Philipp Hyllus, and Hans J. Briegel. Bell inequalities for graph states. 36 Phys. Rev. Lett., 95(12):120405, 4, 2005. 37 [HCDB07] L. Hartmann, J. Calsamiglia, W. Dür, and H. J. Briegel. Weighted graph states and applications to 38 spin chains, lattices and gases. J. Phys. B, 40(9):S1–S44, 2007. 39 [Hid80] Takeyuki Hida. Brownian motion, volume 11 of Applications of Mathematics. Springer-Verlag, New 40 York, 1980. Translated from the Japanese by the author and T. P. Speed. 41 [HKK02] B. M. Hambly, Jun Kigami, and Takashi Kumagai. Multifractal formalisms for the local spectral and 42 walk dimensions. Math. Proc. Cambridge Philos. Soc., 132(3):555–571, 2002. 43 [HKLW07] Deguang Han, Keri Kornelson, David Larson, and Eric Weber. Frames for undergraduates, vol- 44 ume 40 of Student Mathematical Library. American Mathematical Society, Providence, RI, 2007. 45 [Hut81] John E. Hutchinson. Fractals and self-similarity. Indiana Univ. Math. J., 30(5):713–747, 1981. 46 [JO00]´ Palle E. T. Jorgensen and Gestur Olafsson.´ Unitary representations and Osterwalder-Schrader duality. 47 In The mathematical legacy of Harish-Chandra (Baltimore, MD, 1998), volume 68 of Proc. Sympos. 48 Pure Math., pages 333–401. Amer. Math. Soc., Providence, RI, 2000. 49 [Jør76] Palle E. T. Jørgensen. Approximately reducing subspaces for unbounded linear operators. J. Func- 50 tional Analysis, 23(4):392–414, 1976. 51 [Jør77] Palle E. T. Jørgensen. Approximately invariant subspaces for unbounded linear operators. II. Math. 52 Ann., 227(2):177–182, 1977. 53 [Jør78] Palle E. T. Jørgensen. Essential self-adjointness of semibounded operators. Math. Ann., 237(2):187– 54 192, 1978. 55 [Jor83] Palle E. T. Jorgensen. An optimal spectral estimator for multidimensional time series with an infinite 56 number of sample points. Math. Z., 183(3):381–398, 1983. 57 [Jor00] Palle E. T. Jorgensen. Off-diagonal terms in symmetric operators. J. Math. Phys., 41(4):2337–2349, 58 2000. 182 PALLEE.T.JORGENSENANDERINP.J.PEARSE

[Jor04] Palle E. T. Jorgensen. Iterated function systems, representations, and Hilbert space. Internat. J. 1 Math., 15(8):813–832, 2004. 2 [Jor06] Palle E. T. Jorgensen. Analysis and probability: wavelets, signals, fractals, volume 234 of Graduate 3 Texts in Mathematics. Springer, New York, 2006. 4 [JP94] Palle E. T. Jorgensen and Steen Pedersen. Harmonic analysis and fractal limit-measures induced by 5 representations of a certain C∗-algebra. J. Funct. Anal., 125(1):90–110, 1994. 6 2 [JP98] Palle E. T. Jorgensen and Steen Pedersen. Dense analytic subspaces in fractal L -spaces. J. Anal. 7 Math., 75:185–228, 1998. 8 [JP08a] Palle E. T. Jorgensen and Erin P. J. Pearse. Boundaries of infinite electrical resistance networks. page 9 in preparation, 2008. 10 [JP08b] Palle E. T. Jorgensen and Erin P. J. Pearse. First and second quantization of electrical resistance 11 networks. page in preparation, 2008. 12 [JP08c] Palle E. T. Jorgensen and Erin P. J. Pearse. Hilbert space structure of function spaces on fractals. 13 page in preparation, 2008. 14 [Kat95] Tosio Kato. Perturbation theory for linear operators. Classics in Mathematics. Springer-Verlag, 15 Berlin, 1995. Reprint of the 1980 edition. 16 [Kel97] Johannes Kellendonk. Topological equivalence of tilings. J. Math. Phys., 38(4):1823–1842, 1997. 17 Quantum problems in condensed matter physics. 18 [Kig01] Jun Kigami. Analysis on fractals, volume 143 of Cambridge Tracts in Mathematics. Cambridge 19 University Press, Cambridge, 2001. 20 [Kig03] Jun Kigami. Harmonic analysis for resistance forms. J. Funct. Anal., 204(2):399–444, 2003. 21 [Kig08] Jun Kigami. Volume doubling measures and heat kernel estimates on self-similar sets. Mem. Amer. 22 Math. Soc., to appear:iii+94, 2008. 23 [Kol56] A. N. Kolmogorov. Foundations of the theory of probability. Chelsea Publishing Co., New York, 24 1956. Translation edited by Nathan Morrison, with an added bibliography by A. T. Bharucha-Reid. 25 [Koo27] Bernard Osgood Koopman. On rejection to infinity and exterior motion in the restricted problem of 26 three bodies. Trans. Amer. Math. Soc., 29(2):287–331, 1927. 27 [Koo36] B. O. Koopman. On distributions admitting a sufficient statistic. Trans. Amer. Math. Soc., 39(3):399– 28 409, 1936. 29 [Koo57] B. O. Koopman. Quantum theory and the foundations of probability. In Applied probability. Proceed- 30 ings of Symposia in Applied Mathematics. Vol. VII, pages pp 97–102, New York, 1957. McGraw-Hill 31 Book Co. 32 [KR97] Richard V. Kadison and John R. Ringrose. Fundamentals of the theory of operator algebras. Vol. I, 33 volume 15 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 34 1997. Elementary theory, Reprint of the 1983 original. 35 [Kra01] Steven G. Krantz. Function theory of several complex variables. AMS Chelsea Publishing, Provi- 36 dence, RI, 2001. Reprint of the 1992 edition. 37 [Kyt95] Alexander M. Kytmanov. The Bochner-Martinelli integral and its applications. Birkhäuser Verlag, 38 Basel, 1995. Translated from the Russian by Harold P. Boas and revised by the author. 39 [Lee96] Kyoung Sim Lee. Gel’fand triples associated with finite-dimensional Gaussian measure. Soochow J. 40 Math., 22(1):1–16, 1996. 41 [LP09] Russell Lyons and Yuval Peres. Probability on Trees and Graphs. Unpublished (see Lyons’ web site), 42 2009. 43 [LPW08] Daniel Levin, Yuval Peres, and Elizabeth Wilmer. Markov Chains and Mixing Times. American 44 Mathematical Society, Providence, RI, 2008. 45 [Lyo83] Terry Lyons. A simple criterion for transience of a reversible Markov chain. Ann. Probab., 46 11(2):393–402, 1983. 47 [Mal95] Paul Malliavin. Integration and probability, volume 157 of Graduate Texts in Mathematics. Springer- 48 Verlag, New York, 1995. With the collaboration of Hélène Airault, Leslie Kay and Gérard Letac, 49 Edited and translated from the French by Kay, With a foreword by Mark Pinsky. 50 [Min58] R. A. Minlos. Continuation of a generalized random process to a completely additive measure. Dokl. 51 Akad. Nauk SSSR (N.S.), 119:439–442, 1958. 52 [Min59] R. A. Minlos. Generalized random processes and their extension in measure. Trudy Moskov. Mat. 53 Obsc., 8:497–518, 1959. 54 [Min63] R. A. Minlos. Generalized random processes and their extension to a measure. In Selected Transl. 55 Math. Statist. and Prob., Vol. 3, pages 291–313. Amer. Math. Soc., Providence, R.I., 1963. 56 [MM08] M. Maggioni and H. N. Mhaskar. Diffusion polynomial frames on metric measure spaces. Appl. 57 Comput. Harmon. Anal., 24(3):329–353, 2008. 58 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 183

1 [Mor03] Ben Morris. The components of the wired spanning forest are recurrent. Probab. Theory Related 2 Fields, 125(2):259–265, 2003. 3 [Nel64] Edward Nelson. Feynman integrals and the Schrödinger equation. J. Mathematical Phys., 5:332–343, 4 1964. 5 [Nel69] Edward Nelson. Topics in dynamics. I: Flows. Mathematical Notes. Princeton University Press, 6 Princeton, N.J., 1969. 7 [Nel73a] Edward Nelson. Construction of quantum fields from Markoff fields. J. Functional Analysis, 12:97– 8 112, 1973. 9 [Nel73b] Edward Nelson. The free Markoff field. J. Functional Analysis, 12:211–227, 1973. 10 [NW59] C. St. J. A. Nash-Williams. Random walk and electric currents in networks. Proc. Cambridge Philos. 11 Soc., 55:181–194, 1959. 12 [OP96] V. V. Ostapenko and A. I. Pavlygin. Dynamic flows in networks for the generalized Kirchhoff law. 13 Kibernet. Sistem. Anal., (3):96–102, 189, 1996. 14 [Per99] Yuval Peres. Probability on trees: an introductory climb. In Lectures on probability theory and sta- 15 tistics (Saint-Flour, 1997), volume 1717 of Lecture Notes in Math., pages 193–280. Springer, Berlin, 16 1999. 17 [Phe66] Robert R. Phelps. Lectures on Choquet’s theorem. D. Van Nostrand Co., Inc., Princeton, N.J.- 18 Toronto, Ont.-London, 1966. 19 [Pól21] Georg Pólya. Uber¨ eine Aufgabe der Wahrscheinlichkeitsrechnung betreffend die Irrfahrt im Straßen- 20 netz. Math. Ann., 84(1-2):149–160, 1921. 21 [Pow67] Robert T. Powers. Representations of uniformly hyperfinite algebras and their associated von Neu- 22 mann rings. Bull. Amer. Math. Soc., 73:572–575, 1967. 23 [Pow75] Robert T. Powers. Heisenberg model and a random walk on the permutation group. Lett. Math. Phys., 24 1(2):125–130, 1975. 25 [Pow76a] Robert T. Powers. Resistance inequalities for KMS states of the isotropic Heisenberg model. Comm. 26 Math. Phys., 51(2):151–156, 1976. 27 [Pow76b] Robert T. Powers. Resistance inequalities for the isotropic Heisenberg ferromagnet. J. Mathematical 28 Phys., 17(10):1910–1918, 1976. 29 [Pow78] R. T. Powers. Resistance inequalities for the isotropic Heisenberg model. In C∗-algebras and appli- 30 cations to physics (Proc. Second Japan-USA Sem., Los Angeles, Calif., 1977), volume 650 of Lecture 31 Notes in Math., pages 160–167. Springer, Berlin, 1978. 32 [Pow79] Robert T. Powers. Resistance inequalities for the isotropic Heisenberg model. In Algèbres 33 d’opérateurs et leurs applications en physique mathématique (Proc. Colloq., Marseille, 1977), vol- 34 ume 274 of Colloq. Internat. CNRS, pages 291–299. CNRS, Paris, 1979. 35 [PS07] Gyula Pap and LászlóSzeg˝o. Matchings of cycles and paths in directed graphs. Combinatorica, 36 27(3):383–398, 2007. 37 [PW87] Massimo A. Picardello and Wolfgang Woess. Martin boundaries of random walks: ends of trees and 38 groups. Trans. Amer. Math. Soc., 302(1):185–205, 1987. 39 [PW88] Massimo A. Picardello and Wolfgang Woess. Harmonic functions and ends of graphs. Proc. Edin- 40 burgh Math. Soc. (2), 31(3):457–461, 1988. 41 [PW90] M. A. Picardello and W. Woess. Ends of infinite graphs, potential theory and electrical networks. In 42 Cycles and rays (Montreal, PQ, 1987), volume 301 of NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 43 pages 181–196. Kluwer Acad. Publ., Dordrecht, 1990. 44 [PW92] Massimo A. Picardello and Wolfgang Woess. Martin boundaries of Cartesian products of Markov 45 chains. Nagoya Math. J., 128:153–169, 1992. 46 [Rae05] Iain Raeburn. Graph algebras, volume 103 of CBMS Regional Conference Series in Mathematics. 47 Published for the Conference Board of the Mathematical Sciences, Washington, DC, 2005. 48 [Ram01] A. G. Ramm. A simple proof of the Fredholm alternative and a characterization of the Fredholm 49 operators. Amer. Math. Monthly, 108(9):855–860, 2001. 50 [Rie99] Marc A. Rieffel. Metrics on state spaces. Doc. Math., 4:559–600 (electronic), 1999. 51 [RS75] Michael Reed and Barry Simon. Methods of modern mathematical physics. II. Fourier analysis, 52 self-adjointness. Academic Press [Harcourt Brace Jovanovich Publishers], New York, 1975. 53 [RT08] Luke G. Rogers and Alexander Teplyaev. Laplacians on the basilica julia set. 54 http://arxiv.org/abs/0802.3248, 2008. 55 [Rud62] Walter Rudin. Fourier analysis on groups. Interscience Tracts in Pure and Applied Mathematics, No. 56 12. Interscience Publishers (a division of John Wiley and Sons), New York-London, 1962. 57 [Rud87] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, third edition, 1987. 184 PALLEE.T.JORGENSENANDERINP.J.PEARSE

[Rud91] Walter Rudin. Functional analysis. International Series in Pure and Applied Mathematics. McGraw- 1 Hill Inc., New York, second edition, 1991. 2 [Rue69] David Ruelle. Statistical mechanics: Rigorous results. W. A. Benjamin, Inc., New York-Amsterdam, 3 1969. 4 [Rue04] David Ruelle. Thermodynamic formalism. Cambridge Mathematical Library. Cambridge University 5 Press, Cambridge, second edition, 2004. The mathematical structures of equilibrium statistical me- 6 chanics. 7 [Sar86] Donald Sarason. Shift-invariant spaces from the Brangesian point of view. In The Bieberbach con- 8 jecture (West Lafayette, Ind., 1985), volume 21 of Math. Surveys Monogr., pages 153–166. Amer. 9 Math. Soc., Providence, RI, 1986. 10 [Saw97] Stanley A. Sawyer. Martin boundaries and random walks. In Harmonic functions on trees and build- 11 ings (New York, 1995), volume 206 of Contemp. Math., pages 17–44. Amer. Math. Soc., Providence, 12 RI, 1997. 13 [Sch38a] I. J. Schoenberg. Metric spaces and completely monotone functions. Ann. of Math. (2), 39(4):811– 14 841, 1938. 15 [Sch38b] I. J. Schoenberg. Metric spaces and positive definite functions. Trans. Amer. Math. Soc., 44(3):522– 16 536, 1938. 17 [Sch73] Laurent Schwartz. Radon measures on arbitrary topological spaces and cylindrical measures. Pub- 18 lished for the Tata Institute of Fundamental Research, Bombay by Oxford University Press, London, 19 1973. Tata Institute of Fundamental Research Studies in Mathematics, No. 6. 20 [SCW97] Laurent Saloff-Coste and Wolfgang Woess. Transition operators, groups, norms, and spectral radii. 21 Pacific J. Math., 180(2):333–367, 1997. 22 [SMC08] Arthur D. Szlam, Mauro Maggioni, and Ronald R. Coifman. Regularization on graphs with function- 23 adapted diffusion processes. J. Mach. Learn. Res., 9:1711–1739, 2008. 24 [Soa94] Paolo M. Soardi. Potential theory on infinite networks, volume 1590 of Lecture Notes in Mathemat- 25 ics. Springer-Verlag, Berlin, 1994. 26 [Spi76] Frank Spitzer. Principles of random walks. Springer-Verlag, New York, second edition, 1976. Grad- 27 uate Texts in Mathematics, Vol. 34. 28 [Str98a] Robert S. Strichartz. Fractals in the large. Canad. J. Math., 50(3):638–657, 1998. 29 2 [Str98b] Robert S. Strichartz. Remarks on: “Dense analytic subspaces in fractal L -spaces” [J. Anal. Math. 30 75 (1998), 185–228; MR1655831 (2000a:46045)] by P. E. T. Jorgensen and S. Pedersen. J. Anal. 31 Math., 75:229–231, 1998. 32 [Str03] Robert S. Strichartz. A guide to distribution theory and Fourier transforms. World Scientific Pub- 33 lishing Co. Inc., River Edge, NJ, 2003. Reprint of the 1994 original [CRC, Boca Raton; MR1276724 34 (95f:42001)]. 35 [Str05] Daniel W. Stroock. An introduction to Markov processes, volume 230 of Graduate Texts in Mathe- 36 matics. Springer-Verlag, Berlin, 2005. 37 [Str06] Robert S. Strichartz. Differential equations on fractals. Princeton University Press, Princeton, NJ, 38 2006. A tutorial. 39 [Str08] Robert S. Strichartz. Transformations of spectra of graph laplacians. 2008. 40 [SW49] Isaac J. Schoenberg and Anne Whitney. Sur la positivitédes déterminants de translation des fonctions 41 de fréquence de Pólya, avec une application àun problème d’interpolation. C. R. Acad. Sci. Paris, 42 228:1996–1998, 1949. 43 [SW91] Paolo M. Soardi and Wolfgang Woess. Uniqueness of currents in infinite resistive networks. Discrete 44 Appl. Math., 31(1):37–49, 1991. 45 [Tel06] András Telcs. The art of random walks, volume 1885 of Lecture Notes in Mathematics. Springer- 46 Verlag, Berlin, 2006. 47 [Tep98] Alexander Teplyaev. Spectral analysis on infinite Sierpiński gaskets. J. Funct. Anal., 159(2):537– 48 567, 1998. 49 [Tho90] Carsten Thomassen. Resistances and currents in infinite electrical networks. J. Combin. Theory Ser. 50 B, 49(1):87–102, 1990. 51 [Tru89] K. Truemper. On the delta-wye reduction for planar graphs. J. Graph Theory, 13(2):141–148, 1989. 52 [TW93] Carsten Thomassen and Wolfgang Woess. Vertex-transitive graphs and accessibility. J. Combin. The- 53 ory Ser. B, 58(2):248–268, 1993. 54 [vdNB08] Maarten van den Nest and Hans J. Briegel. Measurement-based quantum computation and undecid- 55 able logic. Foundations of Physics, 38(5):448–457, 2008. 56 [vN32a] J. von Neumann. Uber¨ adjungierte Funktionaloperatoren. Ann. of Math. (2), 33(2):294–310, 1932. 57 [vN32b] J. von Neumann. Uber¨ einen Satz von Herrn M. H. Stone. Ann. of Math. (2), 33(3):567–573, 1932. 58 OPERATORTHEORYOFELECTRICALRESISTANCENETWORKS 185

1 [vN32c] J. von Neumann. Zur Operatorenmethode in der klassischen Mechanik. Ann. of Math. (2), 33(3):587– 2 642, 1932. 3 [Web08] Andreas Weber. Analysis of the physical laplacian and the heat flow on a locally finite graph. 4 arXiv:0801.0812, 2008. 5 [Wig55] Eugene P. Wigner. Characteristic vectors of bordered matrices with infinite dimensions. Ann. of 6 Math. (2), 62:548–564, 1955. 7 [Woe86] Wolfgang Woess. Harmonic functions on infinite graphs. Rend. Sem. Mat. Fis. Milano, 56:51–63 8 (1988), 1986. 9 [Woe87] Wolfgang Woess. Random walks on infinite graphs. In Stochastics in combinatorial optimization 10 (Udine, 1986), pages 255–263. World Sci. Publishing, Singapore, 1987. 11 [Woe89] Wolfgang Woess. Boundaries of random walks on graphs and groups with infinitely many ends. 12 Israel J. Math., 68(3):271–301, 1989. 13 [Woe95] Wolfgang Woess. The Martin boundary for harmonic functions on groups of automorphisms of a 14 homogeneous tree. Monatsh. Math., 120(1):55–72, 1995. 15 [Woe96] Wolfgang Woess. Dirichlet problem at infinity for harmonic functions on graphs. In Potential 16 theory—ICPT 94 (Kouty, 1994), pages 189–217. de Gruyter, Berlin, 1996. 17 [Woe97] Wolfgang Woess. Harmonic functions for group-invariant random walks. In Harmonic functions on 18 trees and buildings (New York, 1995), volume 206 of Contemp. Math., pages 179–181. Amer. Math. 19 Soc., Providence, RI, 1997. 20 [Woe00] Wolfgang Woess. Random walks on infinite graphs and groups, volume 138 of Cambridge Tracts in 21 Mathematics. Cambridge University Press, Cambridge, 2000. 22 [WR46] Norbert Wiener and Arturo Rosenblueth. The mathematical formulation of the problem of conduc- 23 tion of impulses in a network of connected excitable elements, specifically in cardiac muscle. Arch. 24 Inst. Cardiol. México, 16:205–265, 1946. 25 [Yoo07] Hyun Jae Yoo. A variational principle in the dual pair of reproducing kernel Hilbert spaces and an 26 application. J. Stat. Phys., 126(2):325–354, 2007. 27 [Zem91] Armen H. Zemanian. Infinite electrical networks, volume 101 of Cambridge Tracts in Mathematics. 28 Cambridge University Press, Cambridge, 1991. 29 [Zem97] Armen H. Zemanian. Nonstandard electrical networks and the resurrection of Kirchhoff’s laws. IEEE 30 Trans. Circuits Systems I Fund. Theory Appl., 44(3):221–233, 1997. 31 [Zha09] Haizhang Zhang. Orthogonality from disjoint support in reproducing kernel Hilbert spaces. J. Math. 32 Anal. Appl., 349(1):201–210, 2009.

33 University of Iowa, Iowa City, IA 52246-1419 USA 34 E-mail address: [email protected]

35 University of Iowa, Iowa City, IA 52246-1419 USA 36 E-mail address: [email protected]