Abstract
An extremely useful tool and viewpoint for the study of random walks is Kirchhoff’s theory of electric networks. Our treatment here roughly follows, we also refer the reader for an in-depth comprehensive study.
You have full access to this open access chapter, Download chapter PDF
An extremely useful tool and viewpoint for the study of random walks is Kirchhoff’s theory of electric networks. Our treatment here roughly follows [69, Chapter 8], we also refer the reader to [61] for an in-depth comprehensive study.
A network is a connected graph G = (V, E) endowed with positive edge weights, {c e}e ∈ E (called conductances ). The reciprocals r e = 1∕c e are called resistances .
In Sects. 2.1–2.4 below we discuss finite networks. We extend our treatment to infinite networks in Sect. 2.5.
2.1 Harmonic Functions and Voltages
Let G = (V, E) be a finite network. In physics classes it is taught that when we impose specific voltages at fixed vertices a and z, then current flows through the network according to certain laws (such as the series and parallel laws). An immediate consequence of these laws is that the function from V to \(\mathbb {R}\) giving the voltage at each vertex is harmonic at each x ∈ V ∖{a, z}.
Definition 2.2
A function \(h:V\to \mathbb {R}\) is harmonic at a vertex x if
Instead of starting with the physical laws and proving that voltage is harmonic, we now take the axiomatically equivalent approach of defining voltage to be a harmonic function and deriving the laws as corollaries.
Definition 2.3
Given a network G = (V, E) and two distinct vertices a, z ∈ V , a voltage is a function \(h:V\to \mathbb {R}\) that is harmonic at any x ∈ V ∖{a, z}.
We will show in Claim 2.8 and Corollary 2.7 that for any \(\alpha , \beta \in \mathbb {R}\), there is a unique voltage h such that h(a) = α and h(z) = β (this assertion is true only when the network is finite).
Claim 2.4
If h 1, h 2 are harmonic at x then so is any linear combination of h 1, h 2.
Proof
Let \(\bar {h}=\alpha h_1+\beta h_2\) for some \(\alpha ,\beta \in \mathbb {R}\). It holds that
□
Claim 2.5
If \(h:V\to \mathbb {R}\) is harmonic at all the vertices of a finite network, then it is constant.
Proof
Let M =supx h(x) be the maximum value of h. Let A = {x ∈ V : h(x) = M}. Since G is finite, A ≠ ∅. Given x ∈ A, we have that h(y) ≤ h(x) for all neighbors y of x. By harmonicity, h(x) is the weighted average of the values of h(y) at the neighbors; but this can only happen if all neighbors of x are also in A. Since G is connected we obtain that A = V implying that h is constant. □
We now show that a voltage is determined by its boundary values, i.e., by its values at a, z.
Claim 2.6
If h is a voltage satisfying h(a) = h(z) = 0, then h ≡ 0.
Proof
Put M =maxx h(x) (which is attained since G is finite) and let A = {x ∈ V : h(x) = M}. As before, by harmonicity, if x ∈ A ∖{a, z} then all of its neighbors are also in A. Since G is connected, there exists a simple path from x to either a or z such that only its endpoint is in {a, z}. Since h(a) = h(z) = 0 we learn that M = 0, that is, h is non-positive. Similarly, one proves that h is non-negative, thus h ≡ 0. □
Corollary 2.7 (Voltage Uniqueness)
For every \(\alpha ,\beta \in \mathbb {R}\) , if h, h′ are voltages satisfying h(a) = h′(a) = α and h(z) = h′(z) = β, then h ≡ h′.
Proof
By Claim 2.4, the function h − h′ is a voltage, taking the value 0 at a and z, hence by Claim 2.6 we get h ≡ h′. □
Claim 2.8
For every \(\alpha ,\beta \in \mathbb {R}\), there exists a voltage h satisfying h(a) = α,h(z) = β.
Proof 1
We write n = |V |. Observe that a voltage h with h(a) = α and h(z) = β is defined by a system of n − 2 linear equations of the form (2.1) in n − 2 variables (which are the values h(x) for x ∈ V ∖{a, z}). Corollary 2.7 guarantees that the matrix representing that system has empty kernel, hence it is invertible. □
We present an alternative proof of existence based on the random walk on the network. Consider the Markov chain {X n} on the state space V with transition probabilities
This Markov chain is a weighted random walk (note that if c xy are all 1 then the described chain is the so-called simple random walk ). We write \(\mathbb {P}_x\) and \(\mathbb {E}_x\) for the probability and expectation, respectively, conditioned on X 0 = x. For a vertex x, define the hitting time of x by
Proof 2
We will find a voltage g satisfying g(a) = 0 and g(z) = 1 by setting
Indeed, g is harmonic at x≠a, z, since by the law of total probability and the Markov property we have
For general boundary conditions α, β we define h by
By Claim 2.4, h is a voltage, and clearly h(a) = α and h(z) = β, concluding the proof. □
This proof justifies the equality between simple random walk probabilities and voltages that was discussed at the start of this chapter: since the function \(x \mapsto \mathbb {P}_x( \tau _z < \tau _a )\) is harmonic on V ∖{a, z} and takes values 0, 1 at a, z respectively, it must be equal to the voltage at x when voltages 0, 1 are imposed at a, z.
Claim 2.9
If h is a voltage with h(a) ≤ h(z), then h(a) ≤ h(x) ≤ h(z) for all x ∈ V .
Furthermore, if h(a) < h(z) and x ∈ V ∖{a, z} is a vertex such that x is in the connected component of z in the graph G ∖{a}, and x is in the connected component of a in the graph G ∖{z}, then h(a) < h(x) < h(z).
Proof
This follows directly from the construction of h in Proof 2 of Claim 2.8 and the uniqueness statement of Corollary 2.7. Alternatively, one can argue as in the proof of Claim 2.6 that if M =maxx h(x) and m =minx h(x), then the sets A = {x ∈ V : h(x) = M} and B = {x ∈ V : h(x) = m} must each contain at least one element of {a, z}.
To prove the second assertion, we note that by Claim 2.8 and Corollary 2.7 it is enough to check when h is the voltage with boundary values h(a) = 0 and h(z) = 1. In this case, the condition on x guarantees that the probabilities that the random walk started at x visits a before z or visits z before a are positive. By proof 2 of Claim 2.8 we find that h(x) ∈ (0, 1). □
2.2 Flows and Currents
For a graph G = (V, E), denote by \(\vec {E}\) the set of edges of G, each endowed with the two possible orientations. That is, \((x,y)\in \vec {E}\) iff {x, y}∈ E (and in that case, \((y,x)\in \vec {E}\) as well).
Definition 2.10
A flow from a to z in a network G is a function \(\theta :\vec {E}\to \mathbb {R}\) satisfying
-
1.
For any {x, y}∈ E we have θ(xy) = −θ(yx) (antisymmetry ), and
-
2.
∀x∉{a, z} we have ∑y:y∼x θ(xy) = 0 (Kirchhoff’s node law ).
Claim 2.11
If θ 1, θ 2 are flows then, so is any linear combination of θ 1, θ 2.
Proof
Let \(\bar {\theta }=\alpha \theta _1+\beta \theta _2\) for some \(\alpha ,\beta \in \mathbb {R}\). It holds that
and for x≠a, z,
□
Definition 2.12
Given a voltage h, the current flow θ = θ h associated with h is defined by θ(xy) = c xy(h(y) − h(x)).
In other words, the voltage difference across an edge is the product of the current flowing along the edge with the resistance of the edge. This is known as Ohm’s law. According to this definition, the current flows from vertices with lower voltage to vertices with higher voltage. We will use this convention throughout, but the reader should be advised that some other sources use the opposite convention.
Claim 2.13
The current flow associated with a voltage is indeed a flow.
Proof
The current flow is clearly antisymmetric by definition. To show that it satisfies the node law, observe that for x≠a, z, since h is harmonic,
□
Claim 2.14
The current flow associated with a voltage h satisfies Kirchhoff’s cycle law , that is, for every directed cycle \(\vec {e}_1,\ldots ,\vec {e}_m\),
Proof
Write \(\vec {e}_i=(x_{i-1},x_i)\), and observe that x 0 = x m. We have that
□
For examples of a flow which does not satisfy the cycle law and a current flow, see Fig. 2.1.
Claim 2.15
Given a flow θ which satisfies the cycle law, there exists a voltage h = h θ such that θ is the current flow associated with h. Furthermore, this voltage is unique up to an additive constant.
Proof
For every vertex x, let \(\vec {e}_1,\ldots ,\vec {e}_k\) be a path from a to x, and define
Note that since θ satisfies the cycle law, the right hand side of (2.3) does not depend on the choice of the path, hence h(x) is well defined. Let x ∈ V , and consider a given path \(\vec {e}_1,\ldots ,\vec {e}_k\) from a to x (if x = a we take the empty path). To evaluate h(y) for y ∼ x, consider the path \(\vec {e}_1,\ldots ,\vec {e}_k,xy\) from a to y, so h(y) = h(x) + r xy θ(xy). It follows that h(y) − h(x) = r xy θ(xy), hence θ(xy) = c xy(h(y) − h(x)), meaning that θ is indeed the current flow associated with h.
Since θ(xy) = c xy(h(y) − h(x)) for any x ∼ y, the node law of immediately implies that h is a voltage. To show that h is unique up to an additive constant, suppose that \(g:V\to \mathbb {R}\) is another voltage such that r xy θ(xy) = g(y) − g(x). It follows that g(y) − h(y) = g(x) − h(x) for any x ∼ y. Since G is connected it follows that g − h is the constant function on V . □
Definition 2.16
The strength of a flow θ is
Claim 2.17
For every flow θ,
Proof
We have that
where the first equality is due to antisymmetry, and the third equality is due to the node law. The claim follows again by antisymmetry. □
Claim 2.18
If θ 1, θ 2 are flows satisfying the cycle law and ∥θ 1∥ = ∥θ 2∥, then θ 1 = θ 2.
Proof
Let \(\bar {\theta }=\theta _1-\theta _2\). According to Claim 2.11, \(\bar {\theta }\) is a flow. It also satisfies the cycle law, as for every cycle \(\vec {e}_1,\ldots ,\vec {e}_m\),
Observe in addition that \(\|\bar {\theta }\|=\|\theta _1\|-\|\theta _2\|=0\). Now, let \(h=h_{\bar {\theta }}\) be the voltage defined in Claim 2.15, chosen so that h(a) = 0. Note that it is harmonic at a, since
Similarly, using Claim 2.17 it is also harmonic at z. Since h is harmonic everywhere, it is constant by Claim 2.5, and thus h ≡ 0, hence \(\bar {\theta }\equiv 0\) and so θ 1 = θ 2. □
This last claim prompts the following useful definition.
Definition 2.19
The unit current flow from a to z is the unique current flow from a to z of strength 1.
2.3 The Effective Resistance of a Network
Suppose we are given a voltage h on a network G with fixed vertices a and z. Scaling h by a constant multiple causes the associated current flow to scale by the same multiple, while adding a constant to h does not change the current flow at all. Therefore, the strength of the current flow is proportional to the difference h(z) − h(a).
Claim 2.20
For every non-constant voltage h and a current flow θ corresponding to h, the ratio
is a positive constant which does not depend on h.
Proof
Let h 1, h 2 be two non-constant voltages, and let θ 1, θ 2 be their associated current flows. For i = 1, 2, let \(\bar {h}_i=h_i/\|\theta _i\|\) and let \(\bar {\theta }_i\) be the current flow associated with \(\bar {h}_i\) (note that since h i is non-constant ∥θ i∥≠ 0). Thus, \(\|\bar {\theta }_i\|=1\). By Claim 2.18 we get \(\bar {\theta }_1=\bar {\theta }_2\) and therefore \(\bar {h}_1=\bar {h}_2+c\) for some constant c by Claim 2.15. It follows that \(\bar {h}_1(z)-\bar {h}_1(a)=\bar {h}_2(z)-\bar {h}_2(a)\).
To see that this constant is positive, it is enough to check one particular choice of a voltage. By Claim 2.8, let h be the voltage with h(a) = 0 and h(z) = 1. By Claim 2.9 and since G is connected, we have that h(x) > 0 for at least one neighbor x of a. Thus, the corresponding current flow θ has ∥θ∥ > 0 making (2.4) positive. □
Claim 2.20 is the mathematical manifestation of Ohm’s law which states that the voltage difference across an electric circuit is proportional to the current through it. The constant of proportionality is usually called the effective resistance of the circuit.
Definition 2.21
The number defined in (2.4) is called the effective resistance between a and z in the network, and is denoted \(\mathcal {R}_{\mathrm {eff}}(a \leftrightarrow z)\). We call its reciprocal the effective conductance between a and z and is denoted \(\mathcal {C}_{\mathrm {eff}}(a \leftrightarrow z):=\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z)^{-1}\).
For examples of computing the effective resistances of networks, see Fig. 2.2.
Notation
In most cases we write \(\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z)\) and suppress the notation of which network we are working on. However, when it is important to us what the network is, we will write \(\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z ; G)\) for the effective resistance in the network G with unit edge conductances and \(\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z ; (G,\{r_e\}))\) for the effective resistance in the network G with edge resistances {r e}e ∈ E. Furthermore, given disjoint subsets A and Z of vertices in a graph G, we write \(\mathcal {R}_{\mathrm {eff}}(A \leftrightarrow Z)\) for the effective resistance between a and z in the network obtained from the original network by identifying all the vertices of A into a single vertex a, and all the vertices of Z into a single vertex z.
Probabilistic Interpretation
For a vertex x we write \(\tau _x^+\) for the stopping time
where X t is the weighted random walk on the network, as defined in (2.2). Note that if X 0 ≠ x then \(\tau _x = \tau _x^+\) with probability 1.
Claim 2.22
Proof
Consider the voltage h satisfying h(a) = 0 and h(z) = 1, and let θ be the current flow associated with h. Due to uniqueness of h (Corollary 2.7) we have that for x≠a, z,
hence
□
Network Simplifications
Sometimes a network can be replaced by a simpler network, without changing the effective resistance between a pair of vertices.
Claim 2.23 (Parallel Law)
Conductances add in parallel. Suppose e 1, e 2 are parallel edges between a pair of vertices, with conductances c 1 and c 2, respectively. If we replace them with a single edge e′ with conductance c 1 + c 2, then the effective resistance between a and z is unchanged.
A demonstration of the parallel law appears in Fig. 2.3.
Proof
Let G′ be the graph where e 1 and e 2 are replaced with e′ with conductance c 1 + c 2. Then it is immediate that if h is any voltage function on G, then it remains a voltage function on the network G′. The claim follows. □
Claim 2.24 (Series Law)
Resistances add in series. Suppose that u∉{a, z} is a vertex of degree 2 and that e 1 = (u, v 1) and e 2 = (u, v 2) are the two edges touching u with edge resistances r 1 and r 2, respectively. If we erase u and replace e 1 and e 2 by a single edge e′ = (v 1, v 2) of resistance r 1 + r 2, then the effective resistance between a and z is unchanged.
The series law is depicted in Fig. 2.4.
Proof
Denote by G′ the graph in which u is erased and e 1 and e 2 are replaced by a single edge (v 1, v 2) of resistance r 1 + r 2. Let θ be a current flow from a to z in G, and define a flow θ′ from a to z in G′ by putting θ′(e) = θ(e) for any e ≠ e 1, e 2 and θ′(v 1, v 2) = θ(v 1, u). Since u had degree 2, it must be that θ(v 1, u) = θ(u, v 2). Thus θ′ satisfies the node law at any x∉{a, z} and ∥θ∥ = ∥θ′∥. Furthermore, since θ satisfies the cycle law, so does θ′. We conclude θ′ is a current flow of the same strength as θ and the voltage difference they induce is the same. □
The operation of gluing a subset of vertices S ⊂ V consists of identifying the vertices of S into a single vertex and keeping all the edges and their conductances. In this process we may generate parallel edges or loops.
Claim 2.25 (Gluing)
Gluing vertices of the same voltage does not change the effective resistance between a and z.
Proof
This is immediate since the voltage on the glued graph is still harmonic. □
Example: Spherically Symmetric Tree
Let Γ be a spherically symmetric tree , that is, a rooted tree where all vertices at the same distance from the root have the same number of children. Denote by ρ the root of the tree, and let \(\{d_n\}_{n\in \mathbb {N}}\) be a sequence of positive integers. Every vertex at distance n from the root ρ has d n children. Denote by Γn the set of all vertices of height n. We would like to calculate \(\mathcal {R}_{\mathrm {eff}}(\rho \leftrightarrow \Gamma _n)\). Due to the tree’s symmetry, all vertices at the same level have the same voltage and therefore by Claim 2.25 we can identify them. Our simplified network has now one vertex for each level, denoted by \(\{v_i\}_{i\in \mathbb {N}}\) (where ρ = v 0), with | Γn+1| edges between v n and v n+1. Using the parallel law (Claim 2.23), we can reduce each set of | Γn| edges to a single edge with resistance \(\frac {1}{|\Gamma _n|}\), then, using the series law (Claim 2.24) we get
see Fig. 2.5.
By Claim 2.22 we learn that
where τ n is the hitting time of Γn for the random walk on Γ. Observe that
so by (2.6) we reach an interesting dichotomy. If \(\sum _{i=1}^\infty \frac {1}{d_1 \cdots d_i} = \infty \), then the random walker returns to ρ with probability 1, and hence returns to ρ infinitely often almost surely. If \(\sum _{i=1}^\infty \frac {1}{d_1 \cdots d_i} < \infty \), then with positive probability the walker never returns to ρ, and hence visits ρ only finitely many times almost surely.
The former graph is called a recurrent graph and the latter is called transient . We will get back to this dichotomy in Sect. 2.5.
2.3.1 The Commute Time Identity
The following lemma shows that the effective resistance between a and z is proportional to the expected time it takes the random walk starting at a to visit z and then return to a, in other words, the expected commute time between a and z. We will use this lemma only in Chap. 6 so the impatient reader may skip this section and return to it later.
Lemma 2.26 (Commute Time Identity)
Let G = (V, E) be a finite network and a ≠ z two vertices. Then
Proof
We denote by \(G_z:V\times V \to \mathbb {R}\) the so-called Green function
and note that
It is straightforward to show that the function ν(x) = G z(a, x)∕π x is harmonic in V ∖{a, z}. Also, we have that G z(a, z) = 0 and \(G_z(a,a) = \frac {1}{\mathbb {P}_a(\tau _z < \tau _a)} = \pi _a\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z)\). Thus, ν is a voltage function with boundary conditions ν(z) = 0 and \(\nu (a) = \mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z)\) which satisfies
Similarly, the same analysis for \(\mathbb {E}_z[\tau _a]\) yields the same result, with the voltage function η which has boundary conditions \(\eta (z) = \mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z)\) and η(a) = 0. Therefore, η(x) = ν(a) − ν(x) for all x ∈ V since both sides are harmonic functions in V ∖{a, z} that receive the same boundary values. This implies that
Summing these up gives
□
2.4 Energy
So far we have seen how to compute the effective resistance of a network via harmonic functions and current flows. However, in typical situations it is hard to find a flow satisfying the circle law. Luckily, an extremely useful property of the effective resistance is that it can be represented by a variational problem. Our intuition from highschool physics suggests that the energy of the unit current flow is minimal among all unit flows from a to z. The notion of energy can be made precise and will allow us to obtain valuable monotonicity properties. For instance, removing any edge from an electric network can only increase its effective resistance. Hence, any recurrent graph remains recurrent after removing any subset of edges from it. Two variational problems govern the effective resistance, Thomson’s principle, which is used to bound the effective resistance from above, and Dirichlet’s principle, used to bound it from below.
Definition 2.27
The energy of a flow θ from a to z, denoted by \(\mathcal {E}(\theta )\), is defined to be
Note that in the second sum we sum over undirected edges, but since θ(xy)2 = θ(yx)2, this is well defined.
Theorem 2.28 (Thomson’s Principle)
and the unique minimizer is the unit current flow.
Proof
First, we will show that the energy of the unit current flow is the effective resistance. Let I be the unit current flow, and h the corresponding (Claim 2.15) voltage function.
Observe that in the second term of the right hand side, for every x≠a, z the sum over all y ∼ x is 0 due to the node law, hence the entire term equals \(\frac {1}{2}(h(a)-h(z))\). From antisymmetry of I, the first term on the right hand side equals \(-\frac {1}{2}(h(a)-h(z))\), hence the right hand side equals altogether \(h(z)-h(a)=\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z)\).
We will now show that every other flow J with ∥J∥ = 1 has \(\mathcal {E}(J) \geq \mathcal {E}(I)\). Let J be such flow and write J = I + (J − I). Set θ = J − I and note that ∥θ∥ = 0. We have
Now,
where the last inequality follows from the same reasoning as before. We conclude that \(\mathcal {E}(J) \geq \mathcal {E}(I)\) as required and that equality holds if and only if \(\mathcal {E}(\theta )=0\), that is, if and only if J = I. □
Corollary 2.29 (Rayleigh’s Monotonicity Law)
If {r e}e ∈ E and \(\{r^{\prime }_e\}_{e\in E}\) are edge resistances on the same graph G so that \(r_e \leq r^{\prime }_e\) for all edges e ∈ E, then
Proof
Let θ be a flow on G, then
This inequality is preserved while taking infimum over all flows with strength 1. Applying Theorem 2.28 finishes the proof. □
Corollary 2.30
Gluing vertices cannot increase the effective resistance between a and z.
Proof
Denote by G the original network and by G′ the network obtained from gluing a subset of vertices. Then every flow θ on G (viewed as a function on the edges) is a flow on G′. Hence the infimum in Theorem 2.28 taken over flows in G′ is taken over a larger subset of flows. □
Definition 2.31
The energy of a function \(h:V\to \mathbb {R}\), denoted by \(\mathcal {E}(h)\), is defined to be
Compare the following lemma with Thomson’s principle (Theorem 2.28).
Lemma 2.32 (Dirichlet’s Principle)
Let G be a finite network with source a and sink z. Then
Proof
The infimum is obtained when h is the harmonic function taking 0 and 1 at a, z respectively. The reason is that if there exists v≠a, z with
then we can change the value of h at v to be the right hand side of (2.7) and the energy will only decrease. One way to see this is that if X is a random variable with a second moment, then the value \(\mathbb {E}(X)\) minimizes the function \(f(x)=\mathbb {E}\left ( (X-x)^2 \right )\).
Let h be that harmonic function and let I be its current flow, so I(xy) = c xy(h(y) − h(x)). Write \(\hat {I}=\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z)\cdot I\), so \(\|\hat {I}\|=1\). By Thomson’s principle,
hence
□
2.5 Infinite Graphs
Let G = (V, E) be an infinite connected graph with edge resistances {r e}e ∈ E. We assume henceforth that this network is locally finite , that is, for any vertex x ∈ V we have ∑y:y∼x c xy < ∞. Let {G n} be a sequence of finite subgraphs of G such that \(\bigcup _{n\in \mathbb {N}} G_n = G\) and G n ⊂ G n+1; we call such a sequence an exhaustive sequence of G. Identify all vertices of G ∖ G n with a single vertex z n.
Claim 2.33
Given an exhaustive sequence {G n} of G, the limit
exists.
Proof
The graph G n ∪{z n} can be obtained from G n+1 ∪{z n+1} by gluing the vertices in G n+1 ∖ G n with z n+1 and labeling the new vertex z n. By Corollary 2.30, the effective resistance \(\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow z_n ; G_n \cup \{z_n\})\) is increasing in n. □
Claim 2.34
The limit in (2.8) does not depend on the choice of exhaustive sequence {G n}.
Proof
Indeed, let {G n} and \(\{G^{\prime }_n\}\) be two exhaustive sequences of G. We can find subsequences {i k}k≥1 and {j k}k≥1 such that
Since \(\{ G_{i_1}, G^{\prime }_{j_1}, G_{i_2}, \ldots \}\) is itself an exhaustive sequence of G, the limit of effective resistances for this sequence exists and equals the limits of effective resistances for the subsequences \(\{ G_{i_k} \}\) and \(\{ G^{\prime }_{j_k} \}\). In turn, these are equal to the limits of effective resistances for the original sequences {G n} and \(\{ G^{\prime }_n \}\), respectively. □
Definition 2.35
In an infinite network, the effective resistance from a vertex a and ∞ is
We are now able to address the question of recurrence versus transience of a graph systematically. Recall the definition of \(\tau _x^+\) in (2.5). In an infinite network we define \(\tau _a^+=\infty \) when there is no time t ≥ 1 such that X t = a.
Definition 2.36
A network (G, {r e}e ∈ E) is called recurrent if \(\mathbb {P}_a(\tau _a^+=\infty ) = 0\), that is, if the probability of the random walker started at a never returning to a is 0. Otherwise, it is called transient .
Observe that since G is connected, if \(\mathbb {P}_a(\tau _a^+=\infty )=0\) for one vertex a, then it holds for all vertices in the network. As we have seen, if n is large enough so that a ∈ G n, then
Since \(\bigcap _n \{\tau _{G\setminus G_{n}} < \tau _a^+\} = \{\tau _a^+=\infty \}\) we have
with the convention that 1∕0 = ∞.
Definition 2.37
Let G be an infinite network. A function \(\theta :E(G)\to \mathbb {R}\) is a flow from a to ∞ if it is anti-symmetric and satisfies the node law on each vertex v ≠ a.
The following follows easily from Theorem 2.28, we omit the proof.
Theorem 2.38 (Thomson’s Principle for Infinite Networks)
Let G be an infinite network, then
Corollary 2.39
Let G be an infinite graph. The following are equivalent:
-
1.
G is transient.
-
2.
There exists a vertex a ∈ V such that \(\mathcal {R}_{\mathrm {eff}}(a\leftrightarrow \infty ) < \infty \) . Hence all vertices satisfy this.
-
3.
There exists a vertex a ∈ V and a unit flow θ from a to ∞ with \(\mathcal {E}(\theta ) <\infty \) . Hence all vertices satisfy this.
We will now develop a useful method for bounding effective resistances from below. This will lead us to a popular sufficient criterion for recurrence in Corollary 2.43.
Definition 2.40
A cutset Γ ⊆ E(G) separating a from z is a set of edges such that every path from a to z must use an edge from Γ.
Claim 2.41
Let θ be a flow from a to z in a finite network, and let Γ a cutset separating a from z. Then
Proof
Denote by Z the set of vertices separated from a by Γ. Denote by G′ the network where Z is identified to a single vertex x and all edges having both endpoints in Z are removed. Now, the restriction of θ to the edges of the new network is a flow from a to x. By Claim 2.17, we have ∑y:y∼x θ(yx) = ∥θ∥. Also, all edges incident to x must be in Γ, since otherwise x is not separated from a by Γ. Therefore
□
Theorem 2.42 (Nash-Williams Inequality)
Let { Γ n} be disjoint cutsets separating a from z in a finite network. Then
Proof
Let θ be a flow from a to z with ∥θ∥ = 1. From Cauchy-Schwarz, for each n we have
Also, since Γn is a cutset, the flow passing through Γn is at least ∥θ∥, by Claim 2.41. So
Combining them, we get that
Summing over all n gives
Applying Thomson’s principle (Theorem 2.28) yields the result. □
Consider now an infinite network G = (V, E). We say that Γ ⊂ E is a cutset separating a from ∞ if any infinite simple path from a must intersect Γ.
Corollary 2.43
In any infinite network, if there exists a collection { Γ n} of disjoint cutsets separating a from ∞ such that
then the network is recurrent.
Example 2.44 (\(\mathbb {Z}^2\) is Recurrent)
Define Γn as the set of vertical edges {(x, y), (x, y + 1)} with |x|≤ n and \(\min \{ |y|, |y+1| \} = n\) along with the horizontal edges {(x, y), (x + 1, y)} with |y|≤ n and \(\min \{ |x|, |x+1| \} =n\), see Fig. 2.6. Then { Γn} is a collection of disjoint cutsets separating 0 from ∞. Also, | Γn| = 4(2n + 1) and therefore \(\sum _n \left (\sum _{e\in \Gamma _n} c_e\right )^{-1} = \infty \). We deduce by Corollary 2.43 that \(\mathbb {Z}^2\) is recurrent.
Remark 2.45
There are recurrent graphs for which there exists M < ∞ such that for every collection { Γn} of disjoint cutsets, \(\sum _n \left (\sum _{e\in \Gamma _n} c_e\right )^{-1} \leq M\). Therefore, the Nash-Williams inequality is not sharp. See Exercise 4 of this chapter.
2.6 Random Paths
We now present the method of random paths, which is one of the most useful methods for generating unit flows on a network and bounding their energy. In fact, it is possible to show that the electric flow can be represented by such a random path. Suppose G is a network with fixed vertices a, z and μ is a probability measure on the set of paths from a to z.
Claim 2.46
For a path γ sampled from μ, let
where by \(\vec {e}\) and we mean the two orientations of an edge e of G. Set
Then θ is a flow from a to z with ∥θ∥ = 1.
Proof
θ is antisymmetric since θ γ is antisymmetric for every γ, and it satisfies the node law since θ γ satisfies the node law. Similarly, the “strength” of θ γ (i.e., ∑x∼a θ γ(ax)) is 1, hence ∥θ∥ = 1. □
An example of the use of this method is the following classical result.
Theorem 2.47
\(\mathbb {Z}^3\) is transient.
Proof
For R > 0 denote by B R = {(x, y, z) : x 2 + y 2 + z 2 ≤ R 2} the ball of radius R in \(\mathbb {R}^3\). Put \(V_R=B_R\cap \mathbb {Z}^3\) and let ∂V R be the external vertex boundary of V R, that is, the set of vertices not in V R which belong to an edge with an endpoint in V R.
We construct a random path μ from the origin 0 to ∂V R by choosing a uniform random point p in ∂B R = {(x, y, z) : x 2 + y 2 + z 2 = R 2}, drawing a straight line between 0 and p in \(\mathbb {R}^3\), considering the set of distance at most 10 in \(\mathbb {R}^3\) from the line, and then choosing (in some arbitrary fashion) a path in \(\mathbb {Z}^3\) which is contained inside this set. The non-optimal constant 10 was chosen in order to guarantee that such a discrete path exists for any point p ∈ ∂B R.
By Claim 2.46, the measure μ corresponds to a flow from 0 to ∂V R. To estimate the energy of this flow, we note that if \(\vec {e}\) is an edge at distance r ≤ R from the origin, then the probability that it is traversed by a path drawn by μ is O(r −2). Furthermore, there are O(r 2) such edges. Hence the energy of the flow is at most
for some constant C < ∞ which does not depend on R. By Claim 2.46 and Theorem 2.28 we learn that \(\mathcal {R}_{\mathrm {eff}}(\mathbf {0}\leftrightarrow \partial V_R)\le C\) for all R, and so by Corollary 2.39 we deduce that \(\mathbb {Z}^3\) is transient. □
2.7 Exercises
-
1.
Let G z(a, x) be the Green’s function, that is,
$$\displaystyle \begin{aligned}G_z(a,x) = \mathbb{E}_a\big [ \# \mbox{visits to }x\mbox{ before visiting }z \big ] \, .\end{aligned}$$Show that the function h(x) = G z(a, x)∕π(x) is a voltage.
-
2.
Show that the effective resistance satisfies the triangle inequality. That is, for any three vertices x, y, z we have
$$\displaystyle \begin{aligned} \mathcal{R}_{\mathrm{eff}}(x \leftrightarrow z) \leq \mathcal{R}_{\mathrm{eff}}(x \leftrightarrow y) + \mathcal{R}_{\mathrm{eff}}(y \leftrightarrow z) \, .\end{aligned} $$(2.9) -
3.
Let a, z be two vertices of a finite network and let τ a, τ z be the first visit time to a and z, respectively, of the weighted random walk. Show that for any vertex x
$$\displaystyle \begin{aligned}{\mathbf{P}}_x( \tau_a < \tau_z ) \leq {\mathcal{R}_{\mathrm{eff}}(x \leftrightarrow \{a,z\} ) \over \mathcal{R}_{\mathrm{eff}}(x \leftrightarrow a)} \, .\end{aligned}$$ -
4.
Consider the following tree T. At height n it has 2 n vertices (the root is at height n = 0) and if \((v_1, \ldots , v_{2^n})\) are the vertices at level n we make it so that v k has 1 child at level n + 1 and if 1 ≤ k ≤ 2n−1 and v k has 3 children at level n + 1 for all other k.
-
(a)
Show that T is recurrent.
-
(b)
Show that for any disjoint edge cutsets Πn we have that ∑n| Πn|−1 < ∞. (So, the Nash-Williams criterion for recurrence is not sharp)
-
(a)
-
5.
-
(a)
Let G be a finite planar graph with two distinct vertices a ≠ z such that a, z are on the outer face. Consider an embedding of G so that a is the left most point on the real axis and z is the right most point on the real axis. Split the outer face of G into two by adding the ray from a to −∞ and the ray from z to + ∞. Consider the dual graph G ∗ of G and write a ∗ and z ∗ for the two vertices corresponding to the split outer face of G. Assume that all edge resistances are 1. Show that
$$\displaystyle \begin{aligned}\mathcal{R}_{\mathrm{eff}}(a \leftrightarrow z; G) = {1 \over \mathcal{R}_{\mathrm{eff}}(a^* \leftrightarrow z^*; G^*)} \, .\end{aligned}$$ -
(b)
Show that the probability that a simple random walk on \(\mathbb {Z}^2\) started at (0, 0) has probability 1∕2 to visit (0, 1) before returning to (0, 0).
-
(a)
-
6.
Let G = (V, E) be a graph so that \(V=\mathbb {Z}\) and the edge set E = ∪k≥0 E k where \(E_0 = \{ (i,i+1) : i \in \mathbb {Z}\}\) and for k > 0
$$\displaystyle \begin{aligned}E_k = \Big \{ \big ( 2^k(n-1/2), 2^k(n+1/2) \big ) : n \in \mathbb{Z} \Big \} \, .\end{aligned}$$Is G recurrent or transient?
References
R. Lyons, Y. Peres, Probability on trees and networks, in Cambridge Series in Statistical and Probabilistic Mathematics, vol. 42 (Cambridge University Press, New York, 2016). MR3616205
Y. Peres, Probability on trees: an introductory climb, in Lectures on Probability Theory and Statistics (Saint-Flour, 1997) (Springer, Berlin, 1999), pp. 193–280. MR1746302
Author information
Authors and Affiliations
Rights and permissions
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
Copyright information
© 2020 The Author(s)
About this chapter
Cite this chapter
Nachmias, A. (2020). Random Walks and Electric Networks. In: Planar Maps, Random Walks and Circle Packing. Lecture Notes in Mathematics(), vol 2243. Springer, Cham. https://doi.org/10.1007/978-3-030-27968-4_2
Download citation
DOI: https://doi.org/10.1007/978-3-030-27968-4_2
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-27967-7
Online ISBN: 978-3-030-27968-4
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)