Massachusetts Institute of Technology 18.

453: Combinatorial Optimization

Michel X. Goemans May 2, 2017

6. Lecture notes on matroid intersection

One nice feature about matroids is that a simple greedy algorithm allows to optimize
over its independent sets or over its bases. At the same time, this shows the limitation of
the use of matroids: for many combinatorial optimization problems, the greedy algorithm
does not provide an optimum solution. Yet, as we will show in this chapter, the expressive
power of matroids become much greater once we consider the intersection of the family of
independent sets of two matroids.
Consider two matroids M1 = (E, I1 ) and M2 = (E, I2 ) on the same ground set E, and
consider the family of indepedent sets common to both matroids, I1 ∩ I2 . This is what is
commonly referred to as the intersection of two matroids.
In this chapter, after giving some examples of matroid intersection, we show that that
finding a largest common independent set to 2 matroids can be done efficiently, and provide a
min-max relation for the maximum value. We also consider the weighted setting (generalizing
the assignment problem), although we will not give an algorithm in the general case (although
one exists); we only restrict to a special case, namely the arborescence problem. We shall
hint an algorithm for the general case by characterizing the matroid intersection polytope
and thereby giving a min-max relation for it (an NP∩co-NP characterization). Finally, we
discuss also matroid union; a powerful way to construct matroids from other matroids in
which matroid intersection plays a central role. (The term ’matroid union’ is misleading
as it is not what we could expect after having defined matroid intersection... it does not
correspond to I1 ∪ I2 .)

6.1 Examples
6.1.1 Bipartite matchings
Matchings in a bipartite graph G = (V, E) with bipartition (A, B) do not form the indepen-
dent sets of a matroid. However, they can be viewed as the common independent sets to
two matroids; this is the canonical example of matroid intersection.
LetSMA be a partition matroid with ground set E where the partition of E is given by
E = {δ(v) : v ∈ A} where δ(v) denotes the edges incident to v. Notice that this is a
partition since all edges have precisely one endpoint in A. We also define kv = 1 for every
v ∈ A. Thus, the family of independent sets of MA is given by

IA = {F : |F ∩ δ(v)| ≤ 1 for all v ∈ A}.

In other words, a set of edges is independent for MA if it has at most one edge incident to
every vertex of A (and any number of edges incident to every vertex of b). We can similarly
define MB = (E, IB ) by

IB = {F : |F ∩ δ(v)| ≤ 1 for all v ∈ B}.

6. Lecture notes on matroid intersection May 2, 2017 2

Now observe that any F ∈ IA ∩ IB corresponds to a matching in G, and vice versa. And the
largest common independent set to IA and IB corresponds to a maximum matching in G.

6.1.2 Arborescences
Given a digraph D = (V, A) and a special root vertex r ∈ V , an r-arborescence (or just
arborescence) is a spanning tree (when viewed as an undirected graph) directed away from
r. Thus, in a r-arborescence, every vertex is reachable from the root r. As an r-arborescence
has no arc incoming to the root, we assume that D has no such arc.
r-arborescences can be viewed as sets simultaneously independent in two matroids. Let
G denote the undirected counterpart of D obtained by disregarding the directions of the arcs.
Note that if we have both arcs a1 = (u, v) and a2 = (v, u) in D then we get two undirected
edges also labelled a1 and a2 between u and v in G. Define M1 = (A, I1 ) = M (G) the graphic
matroid corresponding to G, and M2 = (A, I2 ) the partition matroid in which independent
sets are those with at most one arc incoming to every vertex v 6= r. In other words, we let

I2 = {F : |F ∩ δ − (v)| ≤ 1 for all v ∈ V \ {r}}

where δ − (v) denotes the set {(u, v) ∈ A} of arcs incoming to v. Thus, any r-arborescence is
independent in both matroids M1 and M2 . Conversely, any set T independent in both M1
and M2 and of cardinality |V | − 1 (so that it is a base in both matroids) is an r-arborescence.
Indeed, such a T being a spanning tree in G has a unique path between r and any vertex
v; this path must be directed from the root r since otherwise we would have either an arc
incoming to r or two arcs incoming to the same vertex.
In the minimum cost arborescence problem, we are also given a cost function c : A →
R and we are interested in finding the minimum cost r-arborescence. This is a directed
counterpart to the minimum spanning tree problem but, here, the greedy algorithm does not
solve the problem.

6.1.3 Orientations
Given an undirected graph G = (V, E), we consider orientations of all its edges into directed
arcs; namely, each (undirected) edge1 {u, v} is either replaced by an arc2 (u, v) from u to v,
or by an arc (v, u) from v to u. Our goal is, given k : V → N, to decide whether there exists
an orientation such that, for every vertex v ∈ V , the indegree of vertex v (the number of
arcs entering v) is at most k(v). Clearly, this is not always possible, and this problem can
be solved using matroid intersection (or network flows as well).
To attack this problem through matroid intersection, consider the directed graph D =
(V, A) in which every edge e = {u, v} of E is replaced by two arcs (u, v) and (v, u). With the
1
Usually, we use (u, v) to denote an (undirected) edge. In this section, however, we use the notation
{u, v} rather than (u, v) to emphasize that edges are undirected.
2
We use arcs in the case of directed graphs, and edges for undirected graphs.
6. Lecture notes on matroid intersection May 2, 2017 3

arc set A as ground set, we define two partition matroids, M1 and M2 . To be independent
in M1 , one can take at most one of {(u, v), (v, u)} for every (u, v) ∈ E, i.e.
I1 = {F ⊆ A : |F ∩ {(u, v), (v, u)}| ≤ 1 for all (u, v) ∈ E}.
To be independent in M2 , one can take at most k(v) arcs among δ − (v) for every v:
I2 = {F ⊆ A : |F ∩ δ − (v)| ≤ k(v) for all v ∈ V }.
Observe that this indeed defines a partition matroid since the sets δ − (v) over all v partition
A.
Therefore, there exists an orientation satisfying the required indegree restrictions if there
exists a common independent set to M1 and M2 of cardinality precisely |E| (in which case
we select either (u, v) or (v, u) but not both).

6.1.4 Colorful Spanning Trees

Suppose we have an undirected graph G = (V, E) and every edge has a color. This is
represented by a partition of E into E1 ∪ · · · ∪ Ek where each Ei represents a set of edges of
the same color i. The problem of deciding whether this graph has a spanning tree in which
all edges have a different color can be tackled through matroid intersection. Such a spanning
tree is called colorful.
Colorful spanning trees are bases of the graphic matroid M1 = M (G) which are also in-
dependent in the partition matroid M2 = (E, I2 ) defined by I2 = {F : |F ∩Ei | ≤ 1 for all i}.

6.1.5 Union of Two Forests

In Section 6.5, we show that one can decide whether a graph G has two edge-disjoint spanning
trees by matroid intersection.

6.2 Largest Common Independent Set

As usual, one issue is to find a common independent set of largest cardinality, another is to
prove that indeed it is optimal. This is done through a min-max relation.
Given two matroids M1 = (E, I1 ) and M2 = (E, I2 ) with rank functions r1 and r2
respectively, consider any set S ∈ I1 ∩ I2 and any U ⊆ E. Observe that
|S| = |S ∩ U | + |S ∩ (E \ U )| ≤ r1 (U ) + r2 (E \ U ),
since both S ∩ U and S ∩ (E \ U ) are independent in M1 and in M2 (by property (I1 ));
in particular (and this seems weaker), S ∩ U is independent for M1 while S ∩ (E \ U ) is
independent for M2 . Now, we can take the maximum over S and the minimum over U and
derive:
max |S| ≤ min [r1 (U ) + r2 (E \ U )] .
S∈I1 ∩I2 U ⊆E
Somewhat surprisingly, we will show that we always have equality:
6. Lecture notes on matroid intersection May 2, 2017 4

Theorem 6.1 (Matroid Intersection) For any two matroids M1 = (E, I1 ) and M2 =
(E, I2 ) with rank functions r1 and r2 respectively, we have:

max |S| = min [r1 (U ) + r2 (E \ U )] . (1)

S∈I1 ∩I2 U ⊆E

Before describing an algorithm for matroid intersection that proves this theorem, we
consider what the min-max result says for some special cases. First, observe that we can
always restrict our attention to sets U which are closed for matroid M1 . Indeed, if that was
not the case, we could replace U by V = spanM1 (U ) and we would have that r1 (V ) = r1 (U )
while r2 (E \ V ) ≤ r2 (E \ U ). This shows that there always exists a set U attaining the
minimum which is closed for M1 . Similarly, we could assume that E \ U is closed for M2
(but both assumptions cannot be made simultaneously).
When specializing the matroid intersection theorem to the graph orientation problem
discussed earlier in this chapter, we can derive the following.

Theorem 6.2 G = (V, E) has an orientation such that the indegree of vertex v is at most
k(v) for every v ∈ V if and only if for all P ⊆ V we have3 :
X
|E(P )| ≤ k(v).
v∈P

Similarly, for colorful spanning trees, we obtain:

Theorem 6.3 Given a graph G = (V, E) with edges of Ei colored i for i = 1, · · · , k, there
exists a colorful spanning tree if and only if deleting the edges of any c colors (for any c ∈ N)
produces at most c + 1 connected components.

We now prove Theorem 6.1 by exhibiting an algorithm for finding a maximum cardinality
independent set common to two matroids and a corresponding set U for which we have
equality in (1). For the algorithm, we will start with S = ∅ and at each step either augment
S or produce a U that gives equality. Our algorithm will rely heavily on a structure called
the exchange graph. We first focus on just one matroid.

Definition 6.1 Given a matroid M = (E, I) and an independent set S ∈ I, the exchange
graph GM (S) (or just G(S)) is the bipartite graph with bipartition S and E \ S with an edge
between y ∈ S and x ∈ E \ S if S − y + x ∈ I.

Lemma 6.4 Let S and T be two independent sets in M with |S| = |T |. Then there exists a
perfect matching between S \ T and T \ S in GM (S).

The proof is omitted. The converse to Lemma 6.4 does not hold. We next prove a
proposition that is a partial converse to the above lemma.
3
E(P ) denotes the set of edges with both endoints in P .
6. Lecture notes on matroid intersection May 2, 2017 5

Proposition 6.5 Let S ∈ I with exchange graph GM (S). Let T be a set with |T | = |S| and
such that GM (S) has a unique perfect matching between S \ T and T \ S. Then T ∈ I.

Proof: Let N be the unique matching. Orient edges in N from T \ S = {x1 , · · · , xt } to

S \ T = {y1 , · · · , yt }, and orient the rest from S \ T to T \ S. If we contract the edges of
N , observe that the resulting directed graph has no directed cycle since, otherwise, we could
find an alternating cycle prior to contraction, and this would contradict the uniqueness of
the matching. Hence the vertices of GM (S) can be numbered (by a topological ordering)
so that (i) the endpoints of the matching are numbered consecutively and (ii) all edges are
directed from smaller-numbered vertices to larger-numbered vertices. So, number S \ T and
T \ S such that N = {(y1 , x1 ), (y2 , x2 ), . . . , (yt , xt )} and such that (yi , xj ) is never an edge
for i < j.
Now suppose for the sake of contradiction that T 6∈ I. Then T has a circuit C. Take
the smallest i such that xi ∈ C (there must exist at least one element of C in T \ S since
C ⊆ T and S is independent). By construction, (yi , x) is not an edge for x ∈ C − xi .
This implies that x ∈ span(S − yi ) for all x ∈ C − xi . Hence C − xi ⊆ span(S − yi ), so
span(C − xi ) ⊆ span(span(S − yi )) = span(S − yi ). C is a cycle, so xi ∈ span(C − xi ),
and thus xi ∈ span(S − yi ). This is a contradiction, since (yi , xi ) ∈ GM (S) by assumption.
Therefore T must be in I, which proves the proposition. 4
We are now ready to describe the algorithm for proving the minmax formula. First, we
define a new type of exchange graph for the case when we are dealing with two matroids.

Definition 6.2 For S ∈ I1 ∩ I2 , the exchange graph DM1 ,M2 (S) is the directed bipartite
graph with bipartition S and E \ S such that (y, x) is an arc if S − y + x ∈ I1 and (x, y) is
an arc if S − y + x ∈ I2 .

Also define X1 := {x 6∈ S | S+x ∈ I1 }, the set of sources, and X2 := {x 6∈ S | S+x ∈ I2 },

the set of sinks. Then the algorithm is to find a path (we call it an augmenting path) from
X1 to X2 that does not contain any shortcuts (arcs that point from an earlier vertex on
the path to a non-adjacent later vertex on the path). This for example can be obtained by
selecting a shortest path from X1 to X2 . Then replace S with S4P , where P is the set of
vertices on the path. As a special case, if X1 ∩ X2 6= ∅, then we end up with a path that
consists of a singleton vertex and we can just add that element to S. If there is no such
path, then set U := {z ∈ S | z can reach some vertex in X2 in DM1 ,M2 (S)}. Alternatively,
we could define E \ U as the set of vertices which can be reached from a vertex in X1 ; this
may give a different set.
To prove that this algorithm is correct, we need to show that
1. When we stop, the sets S and U do indeed give equality in the minmax formula (1).

2. At each stage in the algorithm, S4P ∈ I1 ∩ I2 .

Proof of 1: First note that X2 ⊆ U and that X1 ∩ U = ∅ (as otherwise we could
keep running the algorithm to increase the size of S). We claim that r1 (U ) = |S ∩ U | and
6. Lecture notes on matroid intersection May 2, 2017 6

r2 (S \ U ) = |S ∩ (E \ U )|. Together, these would imply that |S| = r1 (U ) + r2 (E \ U ), which

is what we need.
Suppose first that |S ∩ U | = 6 r1 (U ). Since S ∩ U ⊆ U and S ∩ U is independent, this
would imply that |S ∩ U | < r1 (U ). Then there would have to exist some x ∈ U \ S such
that (S ∩ U ) + x ∈ I1 . As S ∈ I1 , we can repeatedly add elements of S to (S ∩ U ) + x and
thereby obtain a set of the form S + x − y for some y ∈ S \ U with S + x − y ∈ I1 . But then
(y, x) is an arc in DM1 ,M2 (S), so y ∈ U (since x ∈ U ). This is a contradiction, so we must
have |S ∩ U | = r1 (U ).
Now suppose that |S ∩ (E \ U )| =6 r2 (E \ U ). Then as before we must have |S ∩ (E \ U )| <
r2 (S \ U ). Thus there exists x ∈ (E \ U ) \ S such that (S ∩ (E \ U )) + x ∈ I2 . So, by the same
logic as before, we can find y ∈ S \ (E \ U ) such that S − y + x ∈ I2 . But S \ (E \ U ) = S ∩ U ,
so we have y ∈ S ∩ U such that S − y + x ∈ I2 . But then (x, y) is an arc in DM1 ,M2 (S), so
x ∈ U (since y ∈ U ). This is a contradiction, so we must have |S ∩ (E \ U )| = r2 (E \ U ). 4
Proof of 2: Recall that we need to show that S4P ∈ I1 ∩ I2 whenever P is a path
from X1 to X2 with no shortcuts. We first show that S4P ∈ I1 . We start by definining a
new matroid M10 from M1 as M10 := (E ∪ {t}, {J | J \ {t} ∈ I1 }. In other words, we simply
add a new element {t} that is independent from all the other elements of the matroid. Then
we know that S ∪ {t} is independent in M10 and M20 (where we define M20 analogously to
M10 ). On the other hand, if we view DM10 (S ∪ {t}) as a subgraph of DM10 ,M20 (S ∪ {t}), then
there exists a perfect matching in DM10 (S ∪ {t}) between (S ∩ P ) ∪ {t} and P \ S (given by
the arcs in P that are also arcs in DM10 (S ∪ {t}), together with the arc between {t} and the
first vertex in P ). Furthermore, this matching is unique since P has no shortcuts, so by the
proposition we know that (S ∪ {t})4P is independent in M10 , hence S4P is independent in
M1 .
The proof that S4P ∈ I2 is identical, except that this time the matching consists of the
arcs in P that are also arcs in DM20 (S ∪ {t}), together with the arc between {t} and the last
vertex in P (rather than the first). 4
So, we have proved that our algorithm is correct, and as a consequence have established
the minmax formula.

Exercise 6-1. Deduce König’s theorem about the maximum size of a matching in a bi-
partite graph from the min-max relation for the maximum independent set common to two
matroids.

6.3 Matroid Intersection Polytope

In this section, we characterize the matroid intersection polytope in terms of linear inequal-
ities, that is the convex hull of characteristic vectors of independent sets common to two
matroids. Let M1 = (E, I1 ) and M2 = (E, I2 ) be two matroids, and let

X = {χ(S) ∈ {0, 1}|E| : S ∈ I1 ∩ I2 }.

6. Lecture notes on matroid intersection May 2, 2017 7

The main result is that conv(X) is precisely given by the intersection of the matroid polytopes
for M1 and M2 .
Theorem 6.6 Let
P = {x ∈ R|E| : x(S) ≤ r1 (S) ∀S ⊆ E
x(S) ≤ r2 (S) ∀S ⊆ E .
xe ≥ 0 ∀e ∈ E}
Then conv(X) = P .
Our proof will be vertex-based. We will show that any extreme point of P is integral,
and it can then be easily seen that it corresponds to a common independent set. The proof
will rely on total unimodularity in a subtle way. Even though the overall matrix defining P
is not totally unimodular, we will show that, for every extreme point x∗ , x∗ can be seen as
the solution of a system of equations whose underlying matrix is totally unimodular. This
is a powerful approach that can apply to many settings.
Proof: Let x∗ be an extreme point of P . We know that x∗ is uniquely characterized once
we know the inequalities that are tight in the description of P . Let
Fi = {S ⊆ E : x∗ (S) = ri (S)},
for i = 1, 2. Let E0 = {e ∈ E : x∗e = 0}. We know that x∗ is the unique solution to
x(S) = r1 (S) S ∈ F1
x(S) = r2 (S) S ∈ F2
xe = 0 e ∈ E0 .
Consider the matroid polytope Pi for matroid Mi for i = 1, 2, and define the face Fi of Pi
(for i = 1, 2) to be
Fi = {x ∈ Pi : x(S) = r1 (S) ∀S ∈ Fi
.
xe = 0 ∀e ∈ E0 }
Observe that F1 ∩ F2 = {x∗ }. Also, by Theorem 4.6 of the chapter on matroid optimization,
we have that Fi can be alternatively defined by a chain Ci . Thus, x∗ is the unique solution
to
x(S) = r1 (S) S ∈ C1
x(S) = r2 (S) S ∈ C2
xe = 0 e ∈ E0 .
After eliminating all variables in E0 , this system can be written as Ax = b, where the rows
of A are the characteristic vectors of C1 ∪ C2 .
Such a matrix A is totally unimodular and this can be shown by using Theorem 3.14.
Consider any subset of rows; this corresponds to restricting our attention to chains C10 and C20 .
Consider first C10 . If we assign the largest set to R1 and then keep alternating the assignment
between R2 and R1 as we consider smaller and smaller sets, we obtain that
X X
aij − aij ∈ {0, 1},
i∈C10 ∩R1 i∈C10 ∩R2
6. Lecture notes on matroid intersection May 2, 2017 8

for all j. If for C20 we start with the largest set being in R2 , we get
X X
aij − aij ∈ {0, −1},
i∈C20 ∩R1 i∈C20 ∩R2

for all j. Combining both, we get that indeed for every j, we get a value in {0, 1, −1} showing
that the matrix is totally unimodular. As a result, x∗ is integral, and therefore corresponds
to the characteristic vector of a common independent set. 4

6.4 Arborescence Problem

The minimum cost r-arborescence is the problem of, given a directed graph D = (V, A),
a root vertex r ∈ V and a cost ca for every arc a ∈ A, finding an r-arborescence in D of
minimum total cost. This can thus be viewed as a weighted matroid intersection problem
and we could use the full machinery of matroid intersection algorithms and results. However,
here, we are going to develop a simpler algorithm using notions similar to the Hungarian
method for the assignment problem. We will assume that the costs are nonnegative.
As an integer program, the problem can be formulated as follows. Letting xa be 1 for
the arcs of an r-arborescence, we have the formulation:

X
OP T = min ca x a
a∈A

subject to:
X
xa ≥ 1 ∀S ⊆ V \ {r}
a∈δ − (S)
X
xa = 1 ∀v ∈ V \ {r}
a∈δ − (v)

xa ∈ {0, 1} a ∈ A.

In this formulation δ − (S) represents the set of arcs {(u, v) ∈ A : u ∈

/ S, v ∈ S}. One
can check that any feasible solution to the above corresponds to the incidence vector of
an r-arborescence. Notice that this optimization problem has an exponential number of
constraints. We are going to show that we can P relax both the integrality restrictions to
xa ≥ 0 and also remove the equality constraints a∈δ− (v) xa = 1 and still there will be an
r-arboresence that will be optimum for this relaxed (now linear) program. The relaxed linear
program (still with an exponential number of constraints) is:
6. Lecture notes on matroid intersection May 2, 2017 9

X
LP = min ca x a
a∈A

subject to:
X
(P ) xa ≥ 1 ∀S ⊆ V \ {r}
a∈δ − (S)
xa ≥ 0 a ∈ A.

The dual of this linear program is:

X
LP = max yS
S⊆V \{r}

subject to:
X
(D) y S ≤ ca
S:a∈δ − (S)

yS ≥ 0 S ⊆ V \ {r}.

The algorithm will be constructing an arborescence T (and the corresponding incidence

vector x with xa = 1 whenever a ∈ T and 0 otherwise) and a feasible dual solution y which
satisfy complementary slackness, and this will show that T corresponds to an optimum
solution of (P ), and hence is an optimum arborescence. Complementary slackness says:

1. yS > 0 =⇒ |T ∩ δ − (S)| = 1, and

P
2. a ∈ T =⇒ S:a∈δ− (S) yS = ca .

The algorithm will proceed in 2 phases. In the first phase, it will construct a dual feasible
solution y and a set F of arcs which has a directed path from the root to every vertex. This
may not be an r-arborescence as there might be too many arcs. The arcs in F will satisfy
condition 2 above (but not condition 1). In the second phase, the algorithm will remove
unnecessary arcs, and will get an r-arborescence satisfying condition 1.
Phase 1 is initialized with F = ∅ and yS = 0 for all S. While F does not contain a directed
path to every vertex in V , the algorithm selects a set S such that (i) inside S, F is strongly
connected (i.e. every vertex can reach every vertex) and (ii) F ∩ δ − (S) = ∅. This set S exists
since we can contract all strongly connected components and in the resulting acyclic digraph,
there must be a vertex (which may be coming from the shrinking of a strongly connected
component) with no incoming arc (otherwise tracing back from that vertex we would either
get to the root or discover a new directed cycle (which we could P shrink)). Now we increase
yS as much as possible until a new inequality, say for arc ak , S:ak ∈δ− (S) yS ≤ cak becomes
an equality. In so doing, the solution y remains dual feasible and still satisfies condition 2.
We can now add ak to F without violating complementary slackness condition 2, and then
6. Lecture notes on matroid intersection May 2, 2017 10

we increment k (which at the start we initialized at k = 1). And we continue by selecting

another set S, and so on, until every vertex is reachable from r in F . We have now such a
set F = {a1 , a2 , · · · , ak } and a dual feasible solution y satisfying condition 2.
In step 2, we eliminate as many arcs as possible, but we consider them in reverse order
they were added to F . Thus, we let i go from k to 1, and if F \ {ai } still contains a directed
path from r to every vertex, we remove ai from F , and continue. We then output the
resulting set T of arcs.
The first claim is that T is an arborescence. Indeed, we claim it has exactly |V | − 1 arcs
with precisely one arc incoming to every vertex v ∈ V \ {r}. Indeed, if not, there would be
two arcs ai and aj incoming to some vertex v; say that i < j. In the reverse delete step, we
should have removed aj ; indeed any vertex reachable from r through aj could be reached
through ai as well (unless ai is unnecessary in which case we could get rid of ai later on).
The second (and final) claim is that the complementary slackness condition 1 is also
satisfied. Indeed, assume not, and assume that we have a set S with yS > 0 and |T ∩δ − (S)| >
1. S was chosen at some point by the algorithm and at that time we added ak ∈ δ − (S) to
F . As there were no other arcs in δ − (S) prior to adding ak to F , it means that all other arcs
in T ∩ δ − (S) must be of the form aj with j > k. In addition, when S was chosen, F was
already strongly connected within S; this means that from any vertex inside S, one can go
to any other vertex inside S using arcs ai with i < k. We claim that when aj was considered
for removal, it should have been removed. Indeed, assume that aj is needed to go to vertex
v, and that along the path P to v the last vertex in S is w ∈ S. Then we could go to v
by using ak which leads somewhere in S then take arcs ai with i < k (none of which have
been removed yet as i < k < j) to w ∈ S and then continue along path P . So aj was not
really necessary and should have been removed. This shows that complementary slackness
condition 1 is also satisfied and hence the arborescence built is optimal.

6.5 Matroid Union

From any matroid M = (E, I), one can construct a dual matroid M ∗ = (E, I ∗ ).
Theorem 6.7 Let I ∗ = {X ⊆ E : E \ X contains a base of M }. Then M ∗ = (E, I ∗ ) is a
matroid with rank function
rM ∗ (X) = |X| + rM (E \ X) − rM (E).
There are several ways to show this. One is to first show that indeed the size of the
largest subset of X in I ∗ has cardinality |X| + rM (E \ X) − rM (E) and then show that rM ∗
satisfies the three conditions that a rank function of a matroid needs to satisfy (the third
one, submodularity, follows from the submodularity of the rank function for M ).
One can use Theorem 6.7 and matroid intersection to get a good characterization of when
a graph G = (V, E) has two edge-disjoint spanning trees. Indeed, letting M be the graphic
matroid of the graph G, we get that G has two edge-disjoint spanning trees if and only if
max |S| = |V | − 1.
S∈I∩I ∗
6. Lecture notes on matroid intersection May 2, 2017 11

For the graphic matroid, we know that rM (F ) = n − κ(F ) where n = |V | and κ(F ) denotes
the number of connected components of (V, F ). But by the matroid intersection theorem,
we can write:

max |S| = min [rM (E1 ) + rM ∗ (E \ E1 )]

S∈I∩I ∗ E1 ⊆E
= min [(n − κ(E1 )) + (|E \ E1 | + κ(E) − κ(E1 ))]
E1 ⊆E
= min [n + 1 + |E \ E1 | − 2κ(E1 )] ,
E1 ⊆E

where we replaced κ(E) by 1 since otherwise G would even have one spanning tree. Re-
arranging terms, we get that G has two edge-dsjoint spanning trees if and only if for all
E1 ⊆ E, we have that E \ E1 ≥ 2(κ(E1 ) − 1). If this inequality is violated for some E1 ,
we can add to E1 any edge that does not decrease κ(E1 ). In other words, if the connected
components of E1 are V1 , V2 , · · · , Vp then we can assume that E1 = E \δ(V1 , V2 , · · · Vp ) where
δ(V1 , · · · , Vp ) = {(u, v) ∈ E : u ∈ Vi , v ∈ Vj and i 6= j}. Thus we have shown:

Theorem 6.8 G has two edge-disjoint spanning trees if and only if for all partitions V1 ,
V2 , · · · Vp of V , we have
|δ(V1 , · · · , Vp )| ≥ 2(p − 1).

Theorem 6.8 can be generalized to an arbitrary number of edge-disjoint spanning trees.

This result is not proved here.

Theorem 6.9 G has k edge-disjoint spanning trees if and only if for all partitions V1 ,
V2 , · · · Vp of V , we have
|δ(V1 , · · · , Vp )| ≥ k(p − 1).

From two matroids M1 = (E, I1 ) and M2 = (E, I2 ), we can also define its union by
M1 ∪ M2 = (E, I) where I = {S1 ∪ S2 : S1 ∈ I1 , S2 ∈ I2 }. Notice that we do not impose the
two matroids to be identical as we just did for edge-disjoint spanning trees.
We can show that:

Theorem 6.10 (Matroid Union) M1 ∪ M2 is a matroid. Furthermore its rank function

is given by
rM1 ∪M2 (S) = min {|S \ F | + rM1 (F ) + rM2 (F )} .
F ⊆S

Proof: To show that it is a matroid, assume that X, Y ∈ I with |X| < |Y |. Let
X = X1 ∪ X2 and Y = Y1 ∪ Y2 where X1 , Y1 ∈ I1 and X2 , Y2 ∈ I2 . We can furthermore
assume that the Xi ’s are disjoint and so are the Yi ’s. Finally we assume that among all
choices for X1 , X2 , Y1 and Y2 , we choose the one maximizing |X1 ∩ Y1 | + |X2 ∩ Y2 |. Since
|Y | > |X|, we can assume that |Y1 | > |X1 |. Thus, there exists e ∈ (Y1 \ X1 ) such that
X1 ∪ {e} is independent for M1 . The maximality implies that e ∈ / X2 (otherwise consider
X1 ∪ {e} and X2 \ {e}). But this implies that X ∪ {e} ∈ I as desired.
6. Lecture notes on matroid intersection May 2, 2017 12

We now show the expression for the rank function. The fact that it is ≤ is obvious as an
independent set S ∈ I has size |S \ F | + |S ∩ F | ≤ |S \ F | + rM1 (F ) + rM2 (F ) and this is
true for any F .
For the converse, let us prove it for the entire ground set S = E. Once we prove that
rM1 ∪M2 (E) = min {|E \ F | + rM1 (F ) + rM2 (F )} ,
F ⊆S

the corresponding statement for any set S will follow by just restricting our matroids to S.
Let X be a base of M1 ∪ M2 . The fact that X ∈ I means that X = X1 ∪ X2 with
X1 ∈ I1 and X2 ∈ I2 . We can furthermore assume that X1 and X2 are disjoint and that
rM2 (X2 ) = rM2 (E) (otherwise add elements to X2 and possibly remove them from X1 ).
Thus we can assume that |X| = |X1 | + rM2 (E). We have that X1 ∈ I1 and also that X1 is
independent for the dual of M2 (as the complement of X1 contains a base of M2 ). In other
words, X1 ∈ I1 ∩ I2∗ . The proof is completed by using the matroid intersection theorem and
Theorem 6.7:
rM1 ∪M2 (E) = |X| = max (|X1 | + rM2 (E))
X1 ∈I1 ∩I2∗


= min rM1 (E1 ) + rM2∗ (E \ E1 ) + rM2 (E)
E1 ⊆E
= min (rM1 (E1 ) + |E \ E1 | + rM2 (E1 ) − rM2 (E) + rM2 (E))
E1 ⊆E
= min (|E \ E1 | + rM1 (E1 ) + rM2 (E1 )) ,
E1 ⊆E

as desired. 4
Since Theorem 6.10 says that M1 ∪ M2 is a matroid, we know that its rank function is
submodular. This is, however, not obvious from the formula given in the theorem.

6.5.1 Spanning Tree Game

The spanning tree game is a 2-player game. Each player in turn selects an edge. Player 1
starts by deleting an edge, and then player 2 fixes an edge (which has not been deleted yet);
an edge fixed cannot be deleted later on by the other player. Player 2 wins if he succeeds in
constructing a spanning tree of the graph; otherwise, player 1 wins. The question is which
graphs admit a winning strategy for player 1 (no matter what the other player does), and
which admit a winning strategy for player 2.
Theorem 6.11 For the spanning tree game on a graph G = (V, E), player 1 has a winning
strategy if and only if G does not have two edge-disjoint spanning trees. Otherwise, player 2
has a winning strategy.
If G does not have 2 edge-disjoint spanning trees then, by Theorem 6.8, we know that
there exists a partition V1 , · · · , Vp of V with |δ(V1 , · · · , Vp )| ≤ 2(p − 1) − 1. The winning
strategy for player 1 is then to always delete an edge from δ(V1 , · · · , Vp ). As player 1 plays
before player 2, the edges in δ(V1 , · · · , Vp ) will be exhausted before player 2 can fix p − 1 of
them, and therefore player 2 loses. The converse is the subject of exercise 6-4.
6. Lecture notes on matroid intersection May 2, 2017 13

Exercise 6-2. Derive from theorem 6.10 that the union of k matroids M1 , M2 , · · · , Mk is
a matroid with rank function

rM1 ∪M2 ∪···∪Mk (S) = min {|S \ F | + rM1 (F ) + rM2 (F ) + · · · + rMk (F )} .

F ⊆S

Exercise 6-3. Derive Theorem 6.9 from Exercise 6-2.

Exercise 6-4. Assume that G has 2 edge-disjoint spanning trees. Give a winning strategy
for player 2 in the spanning tree game.

Exercise 6-5. Find two edge-disjoint spanning trees in the following graph with 16 vertices
and 30 edges or prove that no such trees exist.

1 2 3 4

6 7
5 8

10 11
9 12

13 14 15 16