Sparsest Cut

Sparsest Cut
Arash Mari Oriyad, Fatemeh Ghasemi, Mohammad Jalali
Isfahan University of Technology

arashmarioriyad@gmail.com
August 6, 2020
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Arash Mari Oriyad, Fatemeh Ghasemi, Mohammad Jalali (IUT)
Sparsest Cut August 6, 2020 1 / 34
Abstract
In this work we will obtain an approximation algorithm for the sparsest

cut problem using an interesting LP-rounding procedure in L1 spaces.
we will get as a corollary an approximate max-flow min-cut theorem
for the demands version of multicommodity.
Approximation algorithms for several other important problems will
also follow.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Overview
1 Demands multicommodity flow problem
2 Sparsest cut Problem
3 Linear programming formulation
4 Metrics, cut packings, and L1 − embeddability

Cut packings for metrics
L1 − embeddabilityofmetrics
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Demands multicommodity flow Problem
Let, we have an undirected graph G = (V, E) on n vertices, and each edge

e = (i, j) ∈ E has capacity ce (i, j) ≥ 0.
Let{(s1 , t1 ), (s2 , t2 ), ..., (sk , tk )} be a specified set of pairs of vertices. For
each commodity i, a demand,dem(i) ≥ 0, is also specified. The objective is
to maximize f, called throughput (f⋆ ), such that for each comodity i,
f.dem(i) units of this commodity can be routed simultaneously.
Consider a cut (S, S̄) where S̄ = V − S in G = (V, E). Let:
c(S) = The total capacity of edges crossing S and S̄ ∑
dem(S) = Total demand separated by this cut (S, S̄) = dem(i)
i:|{si ,ti }∩S|=1
c(S)
=⇒ f ≤ ⋆
dem(S)
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Sparsest cut Problem
In the sparsest cut problem, we have an undirected graph G = (V, E) on n

vertices, and each edge e = (i, j) ∈ E has capacity ce (i, j) ≥ 0. The sparsity
of a cut (S, S̄) is a value:
c(S)
α⋆ = min
dem(S)
Our goal in Sparsest cut problem is to find a cut with minimum α⋆ .
f⋆ ≤ α⋆
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Sparsest cut Problem-Example
Example
Consider the bipartite graph K3,2 :
c(S)
α⋆ = min = 1 and f⋆ < 1
dem(S)
This Example shows that minimum sparsity cannot be arbitrarily larger
than maximum throughput; we will show that their ratio is bounded by
O(logk).
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Linear programming formulation
Let Pi = {qij } denote the set of all paths between si and ti . Introduce
variable fij to denote the flow of commodity i sent along path qij
maximize f∑ (21.1)
subject to fij ≥ f.dem(i), i = 1, 2, ..., k
j
∑
fij ≤ ce , e∈E
e∈qij
f≥0
fij ≥ 0
Define the graph H with vertex set VH = {si , ti |1 ≤ i ≤ k} and edge set
EH = {(si , ti )|1 ≤ i ≤ k} to be the demand graph. For each edge
e = (si, ti) of H, let dem(e) = dem(i). We will show that the dual to LP
(21.1) yields a metric (V, d) satisfying:
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Theorem
Let li and de be dual variables associated with the first and second set of
inequalities of LP (21.1). We will interpret de ’s as distance label
assignments to the edges of G.The first set of inequalities ensures that for
each commodity i, li is upper bounded by the length of any path from si to
ti w.r.t. the distance label ∑
assignment. minimize ce de (21.2)
e∈E
∑
subject to de ⩾ li , qij ∈ pi , i = 1, 2, ..., k
e∈qij
∑
li dem(i) ⩾ 1,
i=1
de ⩾ 0 e∈E
li ⩾ 0 i = 1, 2, ..., k
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Example
3
For the instance given in Example 21.3, the optimal throughput is f⋆ = ;
4
this corresponds to routing the four commodities as follows:
1 1
The optimal dual solution is: de = for each edge e and li = for each
8 4
commodityi. It would be instructive for the reader to verify feasibility and
optimality of these solutions.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Claim 21.6 There is an optimal distance label assignment d for the dual
program(21.2) that is a metric on V . Furthermore, for each commodity
i,
∑ li = d(si , ti ), and the second inequality holds with equality, i.e.,
d(si , ti )dem(i) = 1.
i=1
Proof.
If for some three points u, v and w, duw > duv + dvw , then decrease duw to
duv + dvw . Since this does not decrease the shortest path between any si − ti pair,
the solution still remains feasible. Moreover, the objective function value cannot
increase by this process. Continuing in this manner, we will obtain a metric on V .
Now, the length of a shortest path from si to ti is given by the distance label
d(si , ti ). Setting li = d(si , ti ) does not change the feasibility or the objective
function value of the solution. Finally, if the second inequality holds strictly, then
we can scale down all distance labels without violating feasibility, thus
contradicting the optimality of d.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
By Claim 21.6, the dual program yields a metric (V, d) that minimizes
∑
ce de
e∈G
∑ .
dem(e)de
e∈H
By the LP-duality theorem, this equals the optimal throughput. This

proves Theorem 21.4.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Let us think of a metric (V, d) as defining the lengths of edges of the complete graph on
V. Let En denote the set of all edges in the complete graph on n vertices. Let y be a
function assigning nonnegative values to subsets of V , i.e., y : 2V −→ R+ . We will
denote the value of y on set S by yS . As before, let us say
∑that edge e feels yS if e is in
the cut (S, S̄). The amount of cut that edge e feels is y(S). Function y is called a
S:e∈δ(e)
cut packing for metric (V, d) if no edge feels more cut than its length,
∑ i.e., for each
edge. If this inequality holds with equality for each edge e ∈ En , y(S) ≤ de , then y
S:e∈δ(e)
is said to be an exact cut packing. The reason for the name “cut packing” is that
equivalently, we can think of y as assigning value y(S) + y(S̄) to each cut (S, S̄). As
shown below, in general, there may not be an exact cut packing for metric (V, d). Let us
relax this notion by allowing edges to be underpacked up to a specified extent. For
β ≥ 1, y is said to be a β-approximate cut packing if the amount of cut felt by any edge
1 de ∑
is at least fraction of its length, i.e., for each edge e ∈ En , ≤ y(S) ≤ de .
β β
S:e∈δ(e)
Clearly, the smaller β is, the better the cut packing. The following theorem shows the
importance of finding a good cut packing for (V, d).
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Theorem
Let (V, d) be the metric obtained in Theorem 21.4, and let y be a β-approximate cut packing for
′
(V, d). Among cuts with y(S) ̸= 0, let (S , S¯′ ) be the sparsest. Then, the sparsity of this cut is
at most β.f⋆
Proof.
Let y be a -approximate cut packing for metric (V, d). Then,
∑ ∑ ∑
ce de ce y(S)
e∈G e∈G S:e∈δ(e)
f = ∑
⋆
≥ ∑ ∑
dem(e)de dem(e) βy(S)
e∈H e∈H S:e∈δ(e)
∑
y(S)cS
S
= ∑
β y(S)dem(S)
S
′
1 c(S )
≥ ( ).
β dem(S′ ) . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Proof.
The first inequality follows using both the upper bound and the lower
bound on the amount of cut felt by an edge; the former in the numerator
and the latter in the denominator. The equality after that follows by
changing the order of summation. The last inequality follows from the well
known result stated below.
Proposition 21.8 For any nonnegative reals a1 , ..., an and positive reals
b1 , ..., bn and a1 , ..., an ,
∑
αi ai
i ai
∑ ≥ min
αi bi i bi
i
ai
Moreover, this inequality holds with equality iff the n values are all
bi
equal.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Corollary
If there is an exact cut packing for metric (V, d), then every cut (S, S̄)
with yS ̸= 0 has sparsity f⋆ and thus is a sparsest cut in G.
Proof.
By Theorem 21.7, the minimum sparsity cut with yS ̸= 0 has sparsity at
most f⋆ (since β = 1). Since the sparsity of any cut upper bounds f⋆ , the
sparsity of this cut equals f⋆ , and this is a sparsest cut in G. But then all
inequalities in the proof of Theorem 21.7 must hold with equality. Now, by
the second statement in Proposition 21.8, we get that every cut (S, S̄)
with yS ̸= 0 has sparsity f⋆ .
The sparsest cut in the instance specified in Example 21.3 has sparsity
strictly larger than f⋆ . By Corollary 21.9, the optimal metric for this
instance does not have an exact cut packing. However, it turns out that
every metric has an O(log n)-approximate cut packing – we will show this
using the notion of 1-embeddability of metrics.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
L1 -embeddability of metrics
A norm on the vector space Rm is a function ∥ . ∥: Rm −→ R+ , such that

for any x, y ∈ Rm , and λ ∈ R:
∥ x ∥= 0 ⇐⇒ x = 0
∥ λx ∥=| λ | . ∥ x ∥
∥ x + y ∥⩽∥ x ∥ + ∥ y ∥
For p ⩾ 1, the p-norm is defined by
∑ 1
∥ x ∥p = ( | xk |) p .
1⩽k⩽m
The associated p-metric, denoted by dLp , is defined by
dLp (x, y) =∥ x − y ∥p , ∀x, y ∈ Rm
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
L1 − embeddability of metrics
In this section, we will only consider the L1 -norm. Let σ : V −→ Rm be a
mapping, σ : V −→ Rm for some m. Let us say that ∥ σ(u) − σ(v) ∥1 is the L1
length of edge (u, v) under σ. We will say that σ is an isometric L1 -embedding for
metric (V, d) if it preserves the L1 lengths of all edges, i.e.,
∀u, v ∈ V, d(u, v) =∥ σ(u) − σ(v) ∥1 .
As shown below, in general, the metric computed by solving the dual program
may not be isometrically L1 -embeddable. Thus, we will relax this notion-we will
ensure that the mapping does not stretch any edge, but we will allow it to shrink
edges up to a specified factor. For β ⩾ 1, we will say that σ is a β-distortion
L1 -embedding for metric (V, d) if
1
∀u, v : d(u, v) ⩽∥ σ(u) − σ(v) ∥1 ⩽ d(u, v).
β
Next, we show that the question of finding an approximate cut packing for a
metric is intimately related to that of finding a low distortion L1 embedding for it.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lemma
Let σ : V −→ Rm be a mapping. There is a cut packing y : Rm −→ R+
such that each edge feels as much cut under y as its L1 length under σ.
Moreover, the number of nonzero yS ’s is at most m(n − 1).
Proof.
First consider the case when m = 1. Let the n vertices of V be mapped to
u1 ≤ u2 ≤ · · · ≤ un . Assume w.l.o.g. that the vertices are also numbered
in this order. For each i, 1 ≤ i ≤ n − 1, let yv1 ,...,vi = ui+1 − ui . Clearly,
this cut packing satisfies the required condition. For arbitrary m, we
observe that since the L1 -norm is additive, we can define a cut packing for
each dimension independently, and the sum of these packings satisfies the
required condition.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lemma
Let y : Rm −→ R+ be a cut packing with m nonzero yS ’s. There is a
mapping σ : V −→ Rm such that for each edge, its L1 length under σ is
the same as the amount of cut it feels under y.
Proof.
We will have a dimension corresponding to each set S ⊆ V such that
yS = 0. For vertices in S, this coordinate will be 0, and for vertices in S,
this coordinate will be yS . Thus, this dimension contributes exactly as
much to the L1 length of an edge as the amount of cut felt by this edge
due to yS . Hence this mapping satisfies the required condition.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Theorem
There exists a β-distortion L1 -embedding for metric (V, d) iff there exists a
β-approximate cut packing for it. Moreover, the number of nonzero cuts
and the dimension of the (V, d)-embedding are polynomially related.
Corollary
Metric (V, d) is isometrically (V, d)-embeddable iff there exists an exact
cut packing for it.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Low distortion l1 -embeddings for metrics
for metric (V, d): pick a set S ⊆ V , and define the coordinate of vertex v
to be σ(v) = mins∈S d(s, v).
Lemma
For the one-dimensional embedding given above,
∀ u, v ∈ V, |σ(u) − σ(v)| ≤ d(u, v).
Proof.
d(s1 , u) ≤ d(s2 , v).Then,
|σ(u) − σ(v)| = d(s2 , v) − d(s1 , u) ≤ d(s1 , v) − d(s1 , u) ≤ d(u, v).

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
a single edge is not overshrunk
Let us first ensure that a single edge (u, v) is not overshrunk.

define the expected contribution of set Si to the l1 length of edge(u, v) to
be E[|σi (u) − σi (v)|].
For simplicity, assume that n is a power of 2; let n = 2l . For 2 ≤ i ≤ l + 1,
set Si is formed by picking each vertex of V with probability 1/2i .
For any vertex x and nonnegative real r, let B(x, r) denote the ball of
radius r around x.
B(x, r) = {s ∈ V|d(x, s) ≤ r}
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lemma
Pr[(Si ∩ B(u, r1 ) = ∅) and (Si ∩ B(v, r2 ) ̸= ∅)] ≥ c
expected contribution of Si is ≥ c(r1 − r2 )/l.
Proof.
d(u, Si ) ≥ r1 and d(v, Si ) ≤ r2
σi (u) ≥ r1 /l and σi (v) ≤ r2 /l.Therefore, |σi (u) − σi (v)| ≥ (r1 − r2 )/l
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lemma
For 1 ≤ t ≤ l − 1, let A and B be disjoint subsets of V , such that |A| ≤ 2t
and |B| ≥ 2t−1 . Form set S by picking each vertex of V independently
with probability p = 1/(2t+1 ). Then
Pr[(S ∩ A = ∅)and(S ∩ B ̸= ∅)] ≥ (1/2)(1 − e−1/4 ).
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lemma
ρt = min{ρ ≥ 0 : |B(u, ρ)| ≥ 2t and|B(v, ρ)| ≥ 2t }
t̂ = max{t : ρt ≤ d(u, v)/2}
For 1 ≤ t ≤ t̂, the expected contribution of St+1 is at most c.ρt − ρt−1 /l,
and for t = t̂ + 1, the expected contribution of St+1 is at most:
c d(u, v)
·( − ρt−1 )
l 2
Lemma
The expected contribution of all sets S2 , ..., Sl+1 is at most:
c d(u, v)
·
2 l
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lemma
c d(u, v) c/2
Pr[contribution of all sets is ≥ ]≥
4l 2 − c/2
Proof.
the expected contribution S2 , ..., Sl+1 to the l1 length of edge (u, v) is at
most d(u, v)/2l.
d(u, v) c d(u, v) d(u, v)

p· + (1 − p) · ≥ .
l 4l l
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Ensuring that no edge is overshrunk
a single edge (u, v) with constant probability does not overshruk.

to ensure that no edge is overshrunk, we will first enhance this probability.
The key idea is to repeat the entire process several times independently
and use Chernoff bounds to bound the error
∑n probability.
Pr[Xi = 1] = p, 0 ≤ p ≤ 1, and let X = i=1 Xi clearly E[X] = np then for
0≤δ≤1
Pr[X ≤ (1 − δ)np] ≤ exp(−δ 2 np/2)
Pick sets S2 , ..., Sl+1 using probabilities specified above, independently
N = O(logn) times each. Call the sets so obtained
Sij , 2 ≤ i ≤ l + 1, 1 ≤ j ≤ N. Consider the N · l = O(log2 n) dimensional
embedding of metric (V, d) w.r.t. these N · l sets. We will prove that this
is an O(logn)-distortion l1 -embedding for metric (V, d).
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Ensuring that no edge is overshrunk
Lemma
For N = O(logn), this embedding satisfies:
Pr[||σ(u) − σ(v)||1 ≥ p c d(u,v)
4l ] ≥ 1 − 2n1 2 where p = c/(2 − c).
Proof.
A trial succeeds if the contribution of all its sets is≥ (c d(u, v))/2l.
the probability of success is at least p. Using the Chernoff bound with
δ = 1/2, If at least N p/2 trials succeed, the l1 length of edge (u, v) will be
at least p c d(u, v)/4l = d(u, v)/O(logn).
Theorem
The Nl = O(log2 n) dimensional embedding given above is an
O(logn)-distortion l1 -embedding for metric (V, d), with probability at least
1/2.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
LP − rounding − base algorithm
Let V ′ ⊆ V be the set of vertices that are sources or sinks, |V ′ | ≤ 2k. For
simplicity, assume |V ′ | is a power of 2; let |V ′ | = 2l . The sets S2 , ..., Sl + 1
will be picked from V ′ , and it is easy to verify that N = O(log k) will
suﬀice to ensure that none of the O(k2 ) edges of the demand graph is
shrunk by more than a factor of O(logk). The complete algorithm is
Solve the dual LP to obtain metric (V, d).
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Pick sets Sij , 2 ≤ i ≤ l + 1, 1 ≤ j ≤ N ,where set Sij is formed by
picking each vertex of V ′ independentl with probability 1/2i
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Obtain an l1 − embedding of (V, d) in O(log2 k)-dimensional space
w.r.t. these sets.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
w.r.t. these sets.
Obtain an approximate cut packing for (V, d) from the
l1 − embedding.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
w.r.t. these sets.
Obtain an approximate cut packing for (V, d) from the
l1 − embedding.
Output the sparsest cut used by the cut packing.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Theorem
this Algorithm achieves an approximation guarantee of O(logk) for the
sparsest cut problem.
Corollary
For a demands multicommodity flow instance with k source–sink pairs,
1 c(S) c(S)
(min ) ≤ max f ≤ min
O(log k) S⊂V dem(S) throughput f S⊂V dem(S)
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Applications
Edge expantion
Problem
Given an undirected graph G = (V, E),the edge expansion of a set S ⊂ V
with |S| ≤ n2 , is defined to be |σ(S)|, i.e., the number of edges in the cut
(S, S̄). The problem is to find a minimum expansion set.
Consider the special case of demands multicommodity flow in which we

have n(n−1)
2 distinct commodities, one for each pair of vertices. This is
called the uniform multicommodity flow problem. For this problem, the
sparsity of any cut (S, S̄) is given by
c(S)
|S|.|S̄|
Let (S, S̄), with |S||S̄|, be the cut found by the Algorithm when run on G
with uniform demands. Notice that |S| is known within a factor of 2, since
n
2 ≤ |S| ≤ n. Thus, S has expansion within an O(logn) factor of the
minimum expansion set in G. .
.
.
.
.
. . . . .
. . . .
. . . .
. . . .
. . . .
. . . . .
.
.
.
.
.
.
.
.
.

Applications
Balanced cut
Problem
Minimum b-balanced cut: Given an undirected graph G = (V, E) with
nonnegative edge costs and a rational b, 0 ≤ b ≤ 1/2 , find a minimum
capacity cut (S, S̄) such that b.n ≤ |S| ≤ (1 − b).n .
Defenition: A b-balanced cut for b = 1/2 is called a bisection cut, and

the problem of finding a minimum capacity such cut is called the minimum
bisection problem.
We will use the Theorem to obtain a pseudo-approximation algorithm for
this problem.
we will find a (1/3)-balanced cut whose capacity is within an O(logn)
factor of the capacity of a minimum bisection cut
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Lemma
The cut output by the algorithm is a (1/3)-balanced cut whose capacity is
within an O(logn) factor of the capacity of a minimum bisection cut in G
Proof.
At the end of the penultimate iteration,|U| ≤ n/3. Thus, at the beginning
′
of the last iteration, |V | ≥ 2n/3. At most half of these vertices are added
to U in the last iteration. Therefore, |VS| ≥ n/3 and n/3 ≤ |S| ≤ 2n/3.
Hence, (S, V − S) is a (1/3)-balanced cut. (T, T) be a minimum bisection
′
cut in G. Since at the beginning of each iteration, |V | ≥ 2n/3, each of
′ ′
the sets T ∩ V andT ∩ V has at least n/6 vertices.
c(U) c(T) c(T).(2n/3)

≤ O(log n) · → c(S) ≤ O(log n) ·
|U| n/6 n/6
c(S) ≤ O(log n) · c(T)
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
References
Vijay V Vazirani (Springer)

Approximation algorithms
Springer Science 3-540-65367-8.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
The End
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

Sparsest Cut

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sparsest Cut

Uploaded by

Copyright:

Available Formats

Sparsest Cut

Arash Mari Oriyad, Fatemeh Ghasemi, Mohammad Jalali

Isfahan University of Technology

In this work we will obtain an approximation algorithm for the sparsest

1 Demands multicommodity flow problem

2 Sparsest cut Problem

3 Linear programming formulation

4 Metrics, cut packings, and L1 − embeddability

Let, we have an undirected graph G = (V, E) on n vertices, and each edge

In the sparsest cut problem, we have an undirected graph G = (V, E) on n

By the LP-duality theorem, this equals the optimal throughput. This

A norm on the vector space Rm is a function ∥ . ∥: Rm −→ R+ , such that

∀u, v ∈ V, d(u, v) =∥ σ(u) − σ(v) ∥1 .

∀ u, v ∈ V, |σ(u) − σ(v)| ≤ d(u, v).

|σ(u) − σ(v)| = d(s2 , v) − d(s1 , u) ≤ d(s1 , v) − d(s1 , u) ≤ d(u, v).

Let us first ensure that a single edge (u, v) is not overshrunk.

expected contribution of Si is ≥ c(r1 − r2 )/l.

σi (u) ≥ r1 /l and σi (v) ≤ r2 /l.Therefore, |σi (u) − σi (v)| ≥ (r1 − r2 )/l

Pr[(S ∩ A = ∅)and(S ∩ B ̸= ∅)] ≥ (1/2)(1 − e−1/4 ).

d(u, v) c d(u, v) d(u, v)

a single edge (u, v) with constant probability does not overshruk.

Consider the special case of demands multicommodity flow in which we

Arash Mari Oriyad, Fatemeh Ghasemi, Mohammad Jalali (IUT)

Defenition: A b-balanced cut for b = 1/2 is called a bisection cut, and

c(U) c(T) c(T).(2n/3)

c(S) ≤ O(log n) · c(T)

Vijay V Vazirani (Springer)

You might also like