Metric spaces of walks and Lipschitz duality on graphs

Since the functions \varphi\in E(G) are bounded, then \varphi\circ w is also bounded for every w\in W(G). Furthermore, since \|f\circ g\|_{Lip}\leq \|f\|_{Lip}\|g\|_{Lip} for any composition of Lipschitz functions f and g, and since \|w\|_{Lip}\leq 1 by Proposition 2.2, we get the result of (i).

On the other hand, fix \varphi \in E(G). For \varepsilon >0 there exist two different vertices v_1 and v_2 such that

Since the graph G is connected, there exists a walk w\in W(G) such that w(i)=v_1, w(j)=v_2 and d(v_1,v_2)=|i-j|. Then,

Since this holds for every \varepsilon>0 we get \|\varphi\|_{Lip}\leq sup_{w\in W(G)} \|\varphi\circ w\|_{Lip}, and by (i) we conclude (ii).

Finally, for w, u\in W(G) note that, since \varphi is Lipschitz, we get

which gives (iii).

According to the definition, we obtain

Now, for \varepsilon >0 take a pair v_1,v_2 \in V such that

Then, it is enough to consider the constant walks w_1(i)= v_1 and w_2(i)=v_2 for all i \in \mathbb N to get

From this, note that we only need to consider A= \mathbb N to conclude the result of (i):

Furthermore, taking the supremum in the inequalities at the beginning of the proof when \|\varphi\|_{Lip}\le 1, we conclude (ii):

Note that, for all n \in \mathbb{N},

and

since the graph is bounded. Moreover, for m>n,

since the series \sum_{i\in A}\tau_i d\bigl(w(i),v_0\bigr) converges.

Therefore, the sequence of partial sums (s_n(w))_{n\in\mathbb N} is Cauchy and, since \mathcal{F}(G) is complete, its limit

exists and belongs to \mathcal{F}(G). So \iota_A is well defined.

Furthermore, if w_1, w_2 \in S_1(G), we have

so the map \iota_A is Lipschitz with \mathrm{Lip}(\iota_A) \leq 1.

As shown in Lemma 3.3, the elements w\in W(G) as well as the “differences" w_1 \circleddash w_2(A) (couples of walks restricted to A \subset \mathbb N,) can be identified with molecules of the free space \mathcal F(G). Indeed, note that the evaluation

can be identified with the action \langle \iota_A(w_1)- \iota_A(w_2),\varphi \rangle, where the duality in the last formula is understood as the duality between molecules \iota_A(w_1) and \iota_A(w_2) and the elements of \operatorname{Lip}_0(V,\mathbb R).

Note also that the functions w_1 separates the elements of \operatorname{Lip}_0(V,\mathbb R) since, if \varphi_1 \ne \varphi_2, there exist a vertex v\in V such that \varphi_1(v) \ne \varphi_2(v), and v can be identified with the constant walk w_1(i)=v.

Therefore, the topology T_p is Hausdorff and weaker than the weak* topology for the duality \mathcal F(G)^*=\operatorname{Lip}_0(V,\mathbb R), providing that (B_{\operatorname{Lip}}(G), T_p) is compact. Therefore, the identity map \iota\colon(B_{\operatorname{Lip}}(G), w^*)\to(B_{\operatorname{Lip}}(G), T_p) is bijective and continous from a compact to a Hausdorff space, so we can conclude that its inverse is continous. We conclude therefore that both topologies muest coincide.

For fixed w_1,w_2, the function A \mapsto P(w_1,w_2,A) is additive. Since it is bounded with respect to d_\tau, we have that this set function is a countably additive measure. The Radon-Nikodym Theorem gives a K-bounded integrable function (a sequence) (s^{w_1,w_2}_i)_i such that

so we get the result.

First, let us recall that there is a well-known trick for extending the inequality in the statement of the theorem to the same inequality for convex combinations. It is based on the fact that we can repeat terms in the sum as many times as needed, and divide by the total number of terms, in such a way that we obtain a rational approximation for any convex combination by means of rational numbers, and thus a limit procedure yields the result. This procedure can be found well explained in [far]. Thus, from the original inequalities we obtain that, for every set \alpha_1,...,\alpha_n >0 such that \sum_{k=1}^n \alpha_k=1,

for every finite sequence of paths, and we want to get a function \varphi_0 \in B_{E(G)} such that

For all sets of walks w^1_k, w^2_k, sets A_k and parameters \alpha_k’s as above, we consider the functions

\varphi \in B_{E(G)}. We can apply a direct Hahn-Banach separation argument to obtain the result. Each such function \psi is affine and continuous on B_{E(G)}, that is compact by Theorem 3.5. The set of all such \psi defines a convex subset of the dual space C(B_{E(G)})^*, which can be separated from the origin using the Hahn-Banach separation theorem.

Therefore, there exists a function \varphi_0 \in B_{E(G)} such that

which implies

for all couples w_1, w_2 \in W(G) and every set A \in \Sigma_0 as desired.

We consider the case \Sigma= \Sigma_0. By hypothesis, there is a constant K>0 such that for every walks (w^1_1,w^2_1), ..., (w^1_n, w^2_n) and A_1,...,A_n \in \Sigma,

Applying Lemma 4.3, we find a function \varphi_0 \in C(B) such that

By Lemma 4.4, we have that the right hand side of this inequality gives a countably additive measure. Together with the inequality itself, this gives that A \mapsto P(w^1,w^2, A), which is already additive, is also countably additive. The inequality, that works for all subsets A, provides also that P(w^1,w^2, A) is countably additive with respect to the (countably additive) measure

The Radon-Nikodym Theorem (or a direct calculation) gives a \nu_{w_1,w_2}-integrable function (a sequence) s_{w_1,w_2}: \mathbb N \to \mathbb R such that

and \sup_i s^{w_1,w_2}_i \le K, what gives the result. Taking into account that the measurable space is the set of natural numbers, we obtain that

• (i) The first part is an obvious consequence of the fact that “the addition of summations over disjoint subsets of indices equals the summation over their union”. On the other hand, for every sequence of couples of walks (w^1_1,w^2_1), ..., (w^1_n, w^2_n) and sets A_1,...,A_n \in \Sigma,

  \begin{aligned} \sum_{k=1}^n P(w^1_k,w^2_k, A_k) &= \sum_{k=1}^n \sum_{i \in A_k} \tau_i \, s_i \, \big| \widehat \varphi(w^1_k(i))- \widehat \varphi(w^2_k(i)) \big| \\ &\le \sup_{i} |s_i| \, \sum_{k=1}^n \sum_{i \in A_k} \tau_i \big| \widehat \varphi(w^1_k(i))- \widehat \varphi(w^2_k(i)) \big| \\ &\le K \ \Big\| \sum_{k=1}^n \big| \langle w^1_k(A_k) \circleddash w^2_k(A_k), \cdot \rangle \big| \Big\|_{C(B)}. \end{aligned}

• (ii) If w_1,w_2 \in W(G) and A \in \Sigma, we get the Lipschits-type inequality

  \begin{aligned} \big| P(w_1,w_2,A) \big| &\le K \Big| \sum_{i \in A} \tau_i \big| \widehat \varphi(w^1(i))- \widehat \varphi(w^2(i)) \big| \\ &\le K \sum_{i \in A} \tau_i d(w_1(i), w_2(i)) \\ &= K d_\tau(w_1,w_2). \end{aligned}

• (iii) Indeed, the formula is clearly positive and symmetric with respect to w_1,w_2. On the other hand, if A \in \Sigma and w_1,w_2, w_3 \in S_1(G), we have that

  \begin{aligned} \widehat P(w_1,w_3,A) &= \sum_{i \in A} \tau_i \, s_i \, \big| \widehat \varphi(w_1(i))- \widehat \varphi(w_3(i)) \big| \\ &\le \sum_{i \in A} \tau_i \, s_i \, \big| \widehat \varphi(w_1(i))- \widehat \varphi(w_2(i)) \big| \\ & + \sum_{i \in A} \tau_i \, s_i \, \big| \widehat \varphi(w_2(i))- \widehat \varphi(w_3(i)) \big| \\ &= \widehat P(w_1,w_2,A) + \widehat P(w_2,w_3,A). \end{aligned}