Until now, we have considered flow problems withcj = 1for allj. In this section, we show that the flow problem can still be solved using the results we have obtained already by converting it to a problem with cj ≡1. In this section, we adopt the assumption from [33] that the speeds are linearly dependent over the field of rational numbersQ. That is, for all j= 1, . . . , m,
N
cj ∈Nfor some N ∈N. (4.31)
If cj can be written in the form cj = l1
j, where lj ∈ N for all j, we make the transformation
˜
x= (1/cj)x. Then, the flow problem becomes:
F3
∂tuj(˜x, t) =∂x˜uj(˜x, t), x˜∈[0,c1
j], t≥0, uj(x,0) =fj(x),
φ−ijcjuj(c1
j, t) =wi,jPm
k=1φ+i,k(ckuk(0, t)), ∀i.
(4.32)
If the speed cj can be written in the formcj =N/lj for some N, lj ∈ Nand j ∈ {1, . . . , m}, then we first rescale time using the transformationτ =N t. This will put the flow problem into the form where0<˜cj <1, considered above,
∂τuj(x, τ) = l1
j∂xuj(x, τ), x∈(0,1), τ ≥0, uj(x,0) =fj(x),
φ−ijcjuj(1, τ) =wi,jPm
k=1φ+i,k(ckuk(0, τ)), ∀i.
(4.33)
Notice that in this, case,˜cj = 1/lj, which is the first case we considered. Then we can apply the transformation x˜=xlj = (1/˜cj)x and this will put the problem into the form given in (4.32).
From now on, we assume that the flow problem is in the form (4.32).
We then divide the interval[0,c1
j]into unit intervals by creating artificial vertices along the edge ej. The number of artificial vertices created along ej will be (c1
j −1). We label the edges
joining these vertices as ej1, ej2, . . . , ej1/cj. In this way, we have created a larger network with n′ vertices and m′ edges and each edge is of unit length; that iss∈[0,1]with
m′ = Xm
j=1
1 cj
; n′ =n+ Xm
j=1
1 cj −1
=n+m′−m.
The density of particles on edge eji will be denoted by uji for all j = 1, . . . , m and i = 1,2, . . . ,c1
j.
For eachi= 1, . . . ,c1
j,i−1≤x˜≤i, we introduce a new variable s, which is defined on [0,1]
and depends on x˜ through the following equation:
s= ˜x−i+ 1, (4.34)
for eachiandx˜ defined above.
Remark 4.6.1. On the larger network, we shall again represent the outgoing incidence matrix as Φ− and the incoming incidence matrix as Φ+. These matrices are n′ ×m′ matrices. Every artificial vertex has exactly one incoming and outgoing edge, and the weight on the outgoing edge of an artificial vertex is 1. If vi is not an artificial vertex and ej is an outgoing edge ofvi, then the weight on ej,wij, is the same as in the original network.
The flow problem on this larger network is simply the flow problem in (4.10) with a few modi- fications:
∂tuji(s, t) =∂suji(s, t), s∈[0,1], t≥0, uji(s,0) =ϕji(s),
φ−lj
icjiuji(1, t) =wl,jiPm k=1φ+l,k
cj1
ck1
cj
uk1
cj
(0, t), ∀l= 1, . . . , n′.
(4.35)
where
ϕji(s) =
fj(˜x) if0≤x˜≤1; i= 1, fj(˜x) if1≤x˜≤2; i= 2, ... ...
fj(˜x) if c1
j −1≤x˜≤ c1j; i= c1
j
Since each artificial vertex has exactly one incoming and one outgoing edge and the speed on all the artificial edges is the same, the boundary condition on each of these artificial vertices is
φ−lj
icjiuji(1, t) =cjuji(1, t) =φ+lj
i−1cji−1uji−1(0, t) =cjuji−1(0, t) (4.36)
for all i= 2, . . . ,c1
j and j= 1,2, . . . , m. If we define the operator (A, D(A))where D(A) ={v∈W11([0,1])m′ :v(1) =C−1BCv(0)}
and Av = v′, then the flow problem on the new graph is equivalent to the abstract Cauchy problem
v′(t) =Av(t), v(0) =ϕ.
HereB is the adjacency matrix for the line graph of the new expanded network.
Define a functionv(s) = (v1(s), . . . , vm(s)), wherevj = (vj1, . . . , vj1
cj
) is defined as
vj =vji(s), s= ˜x−i+ 1,1≤i≤c−1j , 1≤j≤m. (4.37) Then we define u(x) by
uj(x) =vj(cjx)˜ (4.38)
Theorem 4.6.2. Let S:L1([0,1])m′ → L1([0,1])m be the transformation u =Sv defined by (4.38). ThenS is an isomorphism.
Proof. We note first thatS is invertible, since
vj(s) =
uj(˜x) if0≤x˜≤1; i= 1 uj(˜x) if1≤x˜≤2; i= 2 ... ...
uj(˜x) if c1
j −1≤x˜≤ c1j; i= c1
j,
(4.39)
wheresis as defined in (4.34) for all j= 1, . . . , m, defines the inverse of S. Then kSvk=
Xm
j=1
Z 1
0 |uj(x)|dx
= Xm
j=1
cj Z 1
cj
0 |vj(˜x)|d˜x
= Xm
j=1
cj
1
Xcj
i=1
Z i
i−1|vji(˜x)|d˜x
= Xm
j=1
cj
1
Xcj
i=1
Z 1
0 |vji(s)|ds
≤max{cj}kvk.
On the other hand,kS−1uk=kuk kS−1uk=
Xm
j=1
Z 1
0 |vj1(s)|ds+ Z 1
0 |vj2(s)|ds+· · ·+ Z 1
0 |vj1
cj
(s)|ds
= Xm
j=1
1
Xcj
i=1
Z 1
0 |vji(s)|ds
= Xm
j=1
1
Xcj
i=1
Z i
i−1|vj(˜x)|d˜x
= Xm
j=1
Z 1
cj
0 |vj(˜x)|d˜x= Xm
j=1
1 cj
Z 1
0 |uj(x)|dx
≤max
j
1 cj
kuk. We conclude thatS is an isomorphism.
Lemma 4.6.3. The function v∈D(A) if and only if u=Sv∈D(A).
Proof. From the preceding results (in particular Theorem 4.4.1), we know that (4.35) admits semigroup solutions. Suppose that
v(s) = v1,1(s), . . . , v1,1/c1(s), . . . , vm,1(s), vm,2(s), . . . , vm,1/cm
∈D(A).
Then Equation (4.36) gives us continuity of the flow at the artificial vertices. That is,vji(1) = vji−1(0) for all i= 2, . . . ,c1
j. So the function vj(s), defined in equation (4.37), is continuous on[0,c1
j]for all j= 1, . . . , mand Xm
j=1
Z 1
cj
0 |vj(˜x)|d˜x= Xm
j=1
Z 1
0 |vj(˜x)|d˜x+ Z 2
1 |vj(˜x)|d˜x+· · ·+ Z 1
cj 1
cj−1|vj(˜x)|d˜x
= Xm
j=1
1 cj
Z 1
0 |vj1(x)|dx+ Z 1
0 |vj2(x)|dx+· · ·+ Z 1
0 |vj1 cj
(x)|dx
<∞ since each vji(x) ∈ L1[0,1]. So vj ∈ L1([0,c1
j]) for any j = 1, . . . , m. Next, we show that vj ∈W11([0,c1
j]), j = 1, . . . , m. Since eachvji ∈W11([0,1]), there exists gji ∈L1([0,1]) such
that Z 1
0
vjiφ′ds=− Z 1
0
gjiφds, ∀φ∈C0∞([0,1]), for eachj= 1, . . . , mandi= 1, . . . ,c1
j. On the other hand, let ξ∈C0∞([0,1/cj]). Then Z 1
cj
0
vjξ′d˜x= Z 1
0
vjξ′d˜x+ Z 2
1
vjξ′d˜x+· · ·+ Z 1
cj 1 cj−1
vjξ′d˜x
From this equation, using Corollary 8.10 of [8], we get Z 1
cj
0
vjξ′d˜x=− Z 1
0
ˆ
gj1ξd˜x+ (vj1ξ)|10− Z 2
1
ˆ
gj2ξd˜x+ (vj2ξ)|10− · · · − Z 1
cj 1 cj−1
ˆ gj1
cj
ξd˜x
+ (vj1 cj
ξ)
1 cj 1 cj−1
=
− Z 1
0
ˆ
gj1ξd˜x+vj1(1)ξ(1)
+
− Z 2
1
ˆ
gj2ξd˜x+vj2(2)ξ(2)−vj2(1)ξ(1)
+· · · +
− Z 1
cj 1 cj−1
ˆ gj1
cj
ξdx˜−vj1
cj
1 cj −1
ξ
1 cj −1
,
wheregˆji(˜x) =gji(s)with s= ˜x−i+ 1for j= 1, . . . , mandi= 1, . . . ,1/cj. By the Kirchoff law at the artificial vertices and continuity of the test functions ξ on [0,c1
j], all the boundary terms in the integral cancel out. That is,
vj1(1) =vj2(1), vj2(2) =vj3(2),· · · , vj1
cj
1 cj −1
=vj1
cj−1
1 cj −1
.
Hence
Z 1
cj
0
vjξ′d˜x=− Z 1
0
ˆ
gj1ξdx˜− Z 2
1
ˆ
gj2ξd˜x− · · · − Z 1
cj 1 cj−1
ˆ gj1
cj
ξd˜x
=− Z 1
cj
0
gjvjd˜x;
where
gj(˜x) =
ˆ
gj1(˜x), if 0≤x˜≤1, ˆ
gj2(˜x), if 1≤x˜≤2, ... ...
ˆ gj1
cj
(˜x), if c1
j −1≤x˜≤ c1j. Sincegj(˜x)∈L1([0,c1
j]),vj(˜x, t)∈W11([0,c1
j])for each j= 1, . . . , m, hencevj ∈W11([0,c1
j]), j= 1, . . . , m. We have u∈W11([0,1])m, whereu is defined by (4.38). Since the Kirchoff laws at the original vertices have not been changed, we must have u∈D(A0).
Thus, the solution to (4.35) is given by
(T(t)f)(s) =C−1BnCψ(t+s−n), n∈N0, 0≤t+s−n <1s∈[0,1],
where
ψ(s) =
ϕ1,1(s) ... ϕ1,1
c1
(s) ...
ϕm,1(s) ... ϕm, 1
cm(s)
and from this, we obtain the solution to the original problem through equation (4.38) above.
Using the transformation in Theorem 4.6.2, we have
T(t)u=S−1T(N t)Su. (4.40)
That is, in order to get u1(x, t), we take the submatrix of B which contains only the first c1
1
rows of B and multiply it withϕ. To get u2(x, t), we take the next c1
2 rows by starting from the(c1
1 + 1)th row up to the(c1
1 +c1
2)th row and so on.
A Graph Theoretic Point of View
5.1 Introduction
In the preceding chapter, we proved the existence of semigroup solutions to the flow problem on the graph with no sinks. In this chapter, we extend the results of Chapter 4 to graphs with non trivial acyclic part. We will show that asymptotically, the flow will remain in certain parts of the graph with cycles and that these subgraphs where the flow remains asymptotically are those cycles with no outgoing flow. We also show that the flow on the edges in the acyclic part of the graph will be depleted in finite time while the flow in the cyclic parts of the graph with both incoming and outgoing flow will be depleted asymptotically. We start with certain useful graph descriptions which will enable asymptotic description of the flow to be easier.