Concatenations of Two Incidence Matrices, Spanning

Trees in Planar Graphs: Two Related Problems

Alan Bu

Under the direction of

Yuchong Pan
Department of Mathematics
Massachusetts Institute of Technology

Research Science Institute

July 29, 2023
 Abstract

We call a matrix an incidence submatrix if each row has at most one 1 and at most
one −1. In this paper, we establish an interesting connection between the concatenation
of two incidence submatrices and spanning trees in a planar multigraph, namely that the
maximum determinant of such an m × m concatenation is at least the maximum number of
spanning trees in a planar multigraph with m edges. In addition, we present some evidence
supporting our conjecture that these two quantities might indeed be equal. This includes
several algebraic results on nonplanar graphs. Finally, we give several nontrivial asymptotic
lower and upper bounds on these two quantities, showing that they both grow exponentially
in m.

Summary

Graphs are universal structures in real life abstracting pairwise relations between objects,
such as road systems, electronic circuits, and railway maps. Starting with programming
experiments, we observed and proved an interesting, undiscovered connection between two
fields of mathematics, linear algebra and graph theory. Concretely, if we concatenate two
matrices in which each row has at most one 1 and at most one −1, then the maximum
possible determinant of a matrix generated this way is at least the maximum number of
certain structures in a graph that can be drawn on a plane without edge crossings. We also
present some evidence supporting our conjecture that these two quantities might indeed be
equal. Finally, we demonstrate that these two quantities both grow exponentially in m.
 1 Introduction
Total unimodularity is an extremely powerful tool for proving integrality of a polyhedron
in combinatorial optimization. We say that a matrix is totally unimodular if the determinant
of every square submatrix is in {−1, 0, 1}. A common theme in combinatorial optimization
is to embed combinatorial objects as their characteristic vectors in the Euclidean space and
to relax combinatorial problems into linear programs, which can be solved in polynomial
time. Hence, studying properties of the resulting coefficient matrix, such as total unimodu-
larity, sometimes sheds light on properties of the corresponding polyhedron and the original
combinatorial problem. In particular, given a polyhedron P = {x : Ax ≤ b}, if the matrix
A is totally unimodular and vector b has integer components, then all of the extreme points
of P have integer components.
Given a directed multigraph D = (V, A) with V = {v1 , . . . , vn } and A = {a1 , . . . , am },
we define its incidence matrix MD to be an m × n matrix given by (aij )i,j where aij = 1 or
−1 when vertex vj is the head or tail of arc ai , respectively, and 0 otherwise. We say that a
matrix is an incidence matrix if it is the incidence matrix of some directed graph. It is well
known that incidence matrices are totally unimodular. However, given two incidence matrices
M1 and M2 , it is not immediately clear whether their concatenation M = [M1 |M2 ] is totally
unimodular. This turns out to not be true. Indeed, Goemans and Pan [2] gave a 3 × 3
submatrix of such a concatenation with determinant −3, and constructed a counterexample
from this matrix to the conjecture proposed by A. Abdi, G. Cornuéjols, and G. Zambelli in
2023 that the intersection of two crossing submodular flow systems is box-half-integral.12
Following their work, we investigate how large or small can such a determinant be.
Concretely, we study the following problem:

Problem 1.1. What is the largest (respectively, smallest) determinant of a square matrix
formed by concatenating two matrices which are submatrices of incidence matrices?

1.1 Organization of the Paper

In Section 2, we introduce definitions, notations and known theorems to be used in
subsequent sections.
In Section 3, we show that the maximum determinant max detm of an m × m matrix
formed by concatenating two matrices which are submatrices of incidence matrices is at
least the largest number max spplanar
m of spanning trees in a planar multigraph with m edges.
Furthermore, we show that the construction used in proving the above inequality is the best
possible in the sense that the maximum possible determinant of a square concatenation with
1
A polyhedron P ⊆ Rn is said to be box-half-integral if the extreme points of the intersection of P and
integer box constraints ℓ ≤ x ≤ u for some ℓ, u ∈ Zn have only half-integral components.
2
The conjecture was proposed by A. Abdi, G. Cornuéjols, and G. Zambelli in the talk titled Arc con-
nectivity and submodular flows in digraphs at the “Combinatorics and Optimization” workshop at ICERM
in 2023. The slides of the talk are available at https://app.icerm.brown.edu/assets/403/4951/4951_
3700_Abdi_032820231400_Slides.pdf, and the video of the talk is available at https://icerm.brown.
edu/video_archive/?play=3092.

1
the left incidence submatrix being the truncated incidence submatrix of a planar multigraph
G is at most the number of spanning trees in G. We conjecture that max detm = max spplanar
m .
In Section 4, we prove several theorems as evidence supporting this conjecture. We show
that the maximum possible determinant of a square concatenation with the left incidence
submatrix being the truncated incidence submatrix of a nonplanar multigraph G is at most
the number of spanning trees in G minus 18. We also show that when a nonplanar graph lies
in certain special classes, namely subdivisions of K5 or K3,3 , it always underperforms some
planar graph.
In Section 5, we prove several nontrivial asymptotic bounds on these two quantities,
showing that they grow exponentially as a function of m.

2 Preliminaries
We say that a vector is integral if all of its components are integers. Given a vector
x = (x1 , . . . , xn ) ∈ Rn , its fractionality is defined to be min{xi : i ∈ [n], xi > 0}. For a given
matrix M , we let Mi,j denote the minor of M obtained by removing the ith row and jth
column. For an m × n matrix M and subsets S1 ⊆ [m], S2 ⊆ [n], we denote the submatrix
of M whose
rows are restricted to S1 and whose columns are restricted to S2 as M [S1 , S2 ].
We let [m] n
denote the set of n-element subsets of [m]. Throughout this paper, we forbid
multigraphs from having loops.

2.1 Incidence Matrices

Given a directed multigraph D = (V, A) with V = {v1 , . . . , vn } and A = {a1 , . . . , am }, we
define its incidence matrix MD to be an m × n matrix given by (aij )i,j where aij = 1 or −1
if vertex vj is the head or tail of arc ai , respectively, and 0 otherwise. We say that a matrix
is an incidence matrix if it is the incidence matrix of some directed graph. Furthermore,
we define its truncated incidence matrix MD′ to be MD with its first column removed. Note
that any matrix M such that every row of M has at most one 1 and at most one −1 can be
written as the truncated incidence matrix of a unique directed graph along with some rows
of all 0s. To obtain this graph, we simply add a column to M with entries in {−1, 0, 1} such
that the sum of the entries in each row is equal to 0. Thus each row is either empty or has
exactly one 1 and one −1.
For an m × n matrix A and an m × k matrix B, we define their concatenation M = [A|B]
to be a m × (n + k) matrix such that for each i ∈ [m], the ith row of M is the concatenation
of the ith row of A with the ith row of B.
For every positive integer m, we define max detm to be the maximum value of det M
over all m × m matrices M which are the concatenation of two truncated incidence matrices.
Note that the minimum value of det M is simply −1 · max detm since we can flip the signs
in an arbitrary row of M to flip the sign of the determinant. It follows that max detm ≥ 0.
Given a connected multigraph G, we use max det G to denote the maximum determinant
of [MD′ |N ] over all orientations D of G and over all truncated incidence matrices N such
that [MD′ |N ] is a square matrix. If G is not connected, we define max det G = 0. Note that

2
the specific orientation of G chosen is irrelevant as given any two orientations D1 , D2 of G,
we can multiply some rows of MD1 by ±1 so that they match with the corresponding rows
of MD2 . These operations keep N as a truncated incidence matrix and do not change the
absolute value of the determinant, so we can fix an orientation D of G and define max det G
to be the largest absolute value of the determinant of [MD′ |N ] for a fixed orientation D and
over all incidence submatrices N . The chosen truncated column does not matter since the
sum of the columns of MD is the zero vector (each nonempty row has exactly one 1 and one
−1), so any MD′ can be obtained through column sum operations on another choice of MD′ .

2.2 Spanning Trees and Kirchhoff ’s Matrix-Tree Theorem

Let G = (V, E) be a multigraph. We denote the number of spanning trees of G as sp(G).
An edge contraction along an edge of G involves deleting all edges between the endpoints of
the edge and merging the two endpoints in G. For any edge e = (u, v) ∈ E, we define Gecut
and Gemerge to be the multigraphs obtained by removing and contracting edge e, respectively.
Let n = |V |. Without loss of generality, we assume that V = [n]. We define the adjacency
matrix of G to be an n × n matrix whose (i, j)-entry is the number of edges between vertex
i and vertex j for all i, j ∈ [n]. We define the Laplacian matrix of G to be an n × n matrix
L = D − A, where D is the diagonal matrix whose (i, i)-entry is the degree of vertex i for
i ∈ [n], and A is the adjacency matrix of G.
The following celebrated theorem by Kirchhoff [3] allows one to count the number of
spanning trees in any graph G algebraically.

Theorem 2.1 (Kirchhoff’s matrix-tree theorem, 1847, [3]). Let L be the Laplacian matrix
of a multigraph G. Then det L1,1 = sp(G).

In fact, the number of spanning trees in a graph G has a deletion-contraction relation

which is used in the proof of Kirchhoff’s matrix-tree theorem and is useful to us.

Theorem 2.2 (Lewin, 1982, [4]). For any multigraph G = (V, E) and edge e ∈ E, we have

sp(G) = sp(Gemerge ) + sp(Gecut ).

2.3 Planar Multigraphs

A multigraph G is said to be planar if it can be embedded into the plane in a way so
that any two edges of G can only intersect at a vertex.
For a directed planar multigraph D, we can construct its directed planar dual D∗ by
taking the faces of a planar embedding of the underlying undirected graph of D to be the
vertices of D∗ . We then rotate each edge of D by 90◦ counterclockwise to obtain a directed
graph on D∗ with a bijection between the edges of the two planar multigraphs. For ease
of notation, let the ith rows of MD∗ , MD′ ∗ correspond to the same edge as the ith rows
of MD , MD′ respectively when D∗ is constructed as the dual of D. In particular, we force
rows in MD∗ , MD′ ∗ which correspond to loops in D∗ to only have zero entries, effectively

3
removing them from consideration. We also define the undirected planar dual G∗ of G as
the underlying undirected graph of D∗ for any orientation D of G.
In particular, Euler’s formula implies that the planar dual G∗ of some graph G = (V, E)
has exactly |E| + 2 − |V | vertices. The following theorem by Tutte [7] relates the numbers
of spanning trees in a planar dual pair.
Theorem 2.3 (Tutte, 1984, [7]). Let G be a connected planar multigraph and G∗ its planar
dual. We have that sp(G) = sp(G∗ ).
If the graph obtained from contracting or deleting any edge of G (i.e., any minor of
G) is nonplanar, then G is also nonplanar. Wagner’s theorem allows us to characterize all
nonplanar graphs using this property, and is useful for analyzing general nonplanar graphs.
Theorem 2.4 (Wagner’s theorem, 1937, [8]). Given any nonplanar graph G, there exists a
sequence of edge-deletions and edge-contractions to obtain either a K3,3 or a K5 .
For any multigraph G = (V, E), we define the cycle space C(G) of G to be the space of
subgraphs of G such that every vertex of G has even degree in the subgraph. This space is
|E|
a vector space in F2 . We define a 2-basis of C(G) to be a basis of C(G) such that each edge
appears in at most two elements in the basis. Mac Lane [5] gave the following alternative
characterization of planar graphs:
Theorem 2.5 (Mac Lane’s planarity criterion, 1936, [5]). A multigraph G is planar if and
only if its cycle space has a 2-basis.

3 Proving max detm ≥ max spplanar

m

Using a computer program, we computed max detm , max spplanar m and max spgeneral
m for
m ∈ [10], and the results are given in Table 1. It is surprising that the values of max detm and
max spplanar
m match up for m ∈ [10]. This result suggests that there might be some connection
between the number of spanning trees on planar multigraphs with the determinant of the
concatenation of incidence submatrices. In this section, we prove that max detm is always
lower bounded by max spplanar
m . First, we prove the following theorem:
Theorem 3.1. Let G be a connected planar multigraph. Let D be an orientation of G. Then
det MD′ MD′ ∗ = ± sp(G).
 

Proof. It is clear that G has at least one spanning tree. Suppose G has m edges and n
vertices. Notice that by Euler’s formula, MD′ ∗ is an m × (m − n + 1) matrix, so [MD′ |MD′ ∗ ] is
a square matrix. Let ci instead denote the ith column vector of MD and let vi be the vertex
corresponding to ci . Now, consider the product LG = MDT MD . Let the element in the ith
row and jth column be denoted lij . Then lij = ci · cj for all 1 ≤ i, j ≤ n. Let cij denote the
jth element of ci . Then
m
X X
lii = c2ij = 1 = deg vi .
j=1 cij ̸=0

4
In any row, if cik , cjk ̸= 0 and i ̸= j then cik · cjk = −1 since cik , cjk cannot both be 1 or both
be −1. So for any i ̸= j we have
m
X X X
lij = cik · cjk = cik · cjk = −1,
k=1 k∈[m] k∈[m]
cik ,cjk ̸=0 cik ,cjk ̸=0

which is exactly the negative of the number of edges between vi and vj in G. Hence LG is
exactly the Laplacian matrix of G.
Similarly, the Laplacian matrix LG∗ of G∗ is equal to MDT ∗ MD∗ . Suppose the truncated
column of MD is exactly its first column. Then the Kirchhoff matrix-tree theorem [1] implies
that det MD′T MD′ = det L1,1 = sp(G) ̸= 0, so it follows that the column vectors of MD′ are
linearly independent. From Theorem 2.3, we know that the number of spanning trees of G
and G∗ are equal, so by symmetry it follows that det MD′T∗ MD′ ∗ = sp(G∗ ) = sp(G) ̸= 0 so the
column vectors of MD′ ∗ are also linearly independent.
Now, we claim that MD′ and MD′ ∗ are orthogonal complements of one another. It suffices
to show that for any column vector c1 of MD′ and column vector c2 of MD′ ∗ , we have that
c1 · c2 = 0. Let v ∈ V be the vertex of G corresponding to c1 and f be the face of G which
is dual to the vertex in G∗ corresponding to c2 . Multiplying both the ith row of MD′ and
the ith row of MD′ ∗ by −1 does not change the fact that the two directed graphs are duals
of one another.
Without loss of generality, take an orientation D such that all edges incident to v have
their head at v. If there is no edge e such that both v, f are incident to e, then c1 · c2 = 0.
Otherwise, there exists an edge e such that v lies on e and e lies on the face f . It follows
that the vertex v lies on face f . The edges touching face f form a cycle on G, so v, f share
exactly two edges. It follows from the definition of the directed planar dual that one of them
has its tail at f and the other has its head at f . Hence c1 · c2 = 1 · 1 + 1 · −1 = 0 as desired.
Now, consider the following equality:
2 T  ′
det MD′ MD′ ∗ = det MD′ MD′ ∗ MD MD′ ∗
  
" #
MD′T  ′
MD MD′ ∗

= det ′T
MD∗
" #
MD′T MD′ 0
= det
0 MD′T∗ MD′ ∗
= det MD′T MD′ · det MD′T∗ MD′ ∗
= sp(G)2 .
It follows that det MD′ MD′ ∗ = ± sp(G) as desired.
 

The above proof requires the use of a planar dual, which is ill-defined for nonplanar
graphs and is thus hard to use to understand the behavior of nonplanar graphs. We also
present a proof through the cycle space of G in Appendix B, which exists regardless of the
planarity of G, and is potentially more generalizable.

5
Corollary 3.2. For any positive integer m, we have that max detm ≥ max spplanar
m .
Furthermore, we can ask whether it is possible to construct a square matrix M formed by
concatenating the truncated incidence matrix of planar multigraph G with another incidence
submatrix to obtain a larger determinant. It turns out that Theorem 3.1 gives the best
possible bound:
Theorem 3.3. Given a multigraph G, we have that max det G ≤ sp(G).
Proof. Suppose G has m edges and n vertices. Consider any matrix M = [MD′ |N ] such that
D is an orientation of G. Let aij denote the element in the ith row and jth column of M .
Consider X Y
det M = sgn(σ) aiσ(i) .
σ∈Sm i
[m]


Let A ∈ n−1 , B = [m] \ A denote the two disjoint sets {σ(1), σ(2), . . . , σ(n − 1)} and
{σ(n), σ(n+1), . . . , σ(m)}, respectively. For any two such sets, let σ(A, B) denote the unique
permutation such that A = {σ(1), σ(2), . . . , σ(n − 1)}, B = {σ(n), σ(n + 1), . . . , σ(m)},
σ(1) < σ(2) < · · · < σ(n − 1) and σ(n) < σ(n + 1) < · · · < σ(m). Furthermore, let A(i)
denote the ith smallest element of A and let B(i) denote the ith smallest element of B.
Then we can decompose σ as σ(A, B) ◦ σ1 ◦ σ2 where σ1 , σ2 are permutations on the first
n − 1 elements and last m − n + 1 elements of σ,
respectively. Let [n, m] denote the set
[m]
{n, n + 1, . . . , m}. Then summing over all A ∈ n−1 , we have

XXX n−1
Y m−n+1
Y
det M = sgn(σ) aiσ1 (A(i)) a(n+j−1)σ2 (B(j))
A σ1 σ2 i=1 j=1
n−1
! m−n+1
!
X X Y X Y
= sgn(σ(A, B)) sgn(σ1 ) aiσ1 (A(i)) sgn(σ2 ) a(n+i−1)σ2 (B(j))
A σ1 i=1 σ2 i=1
X
= sgn(σ(A, B)) det M [A, [n − 1]] det M [B, [n, m]].
A

Since the incidence matrix of a directed multigraph is totally unimodular, it follows that
det M [A, [n − 1]], det M [B, [n, m]] ∈ {−1, 0, 1}, so we have
X X
det M ≤ | det M [A, [n − 1]]| = (det M [A, [n − 1]])2 .
[m] [m]
A∈(n−1 ) A∈(n−1 )

Now by the Cauchy-Binet formula, we have the following equality:

X X
det MD′T MD′ = det M T [[n − 1], S1 ] det M [S1 , [n − 1]] = (det M [S1 , [n − 1]])2 .
S1 S1

The Kirchhoff matrix-tree theorem implies that det MD′T MD′ = sp(G), so we have shown the
desired inequality.

6
 This theorem shows that given that one of the incidence submatrices involved in the
concatenation is the truncated incidence matrix of a particular planar multigraph G, the
construction given in the statement of Theorem 3.1 creates the largest possible determinant
out of all choices for the other incidence matrix.

Corollary 3.4. For any connected planar multigraph G, we have max det G = sp(G).

4 Bounding the Determinant on Nonplanar Graphs

Given that we can construct the upper bound of Theorem 3.3 for all planar graphs, one
might expect there to be a similar construction for nonplanar graphs as well that may be
more complex. It turns out that this is not possible. We start by proving a few lemmas.

Lemma 4.1. Let c1 , c2 , c3 , . . . , ck be the columns of a incidence submatrix M . Then we have

that c1 , c2 , . . . , ck−2 , ck−1 + ck form the columns of another incidence submatrix.

Proof. Let the jth element of column i of the original matrix be denoted cij . We proceed
to show that the elements of the jth row of elements of the new matrix, consisting of
{cij |1 ≤ i ≤ k − 2} ∪ {c(k−1)j + ckj }, has at most one 1, one −1, and the remaining elements
are all equal to 0. We perform casework on the jth row for all j ∈ [m]:
Case 1. One of c(k−1)j and ckj is equal to 0.
Since we’ve deleted one 0 from row j, the total number of 1 and −1 entries remains
unchanged, so the desired condition holds.
Case 2. Both c(k−1)j and ckj are nonzero.
In this case, since we must have that c(k−1)j = −ckj = ±1, adding columns k − 1 and
k deletes both nonzero entries from row j and replaces them with zero. Hence the desired
condition still holds.
Thus every row of the matrix formed from c1 , c2 , . . . , ck−2 , ck−1 +ck has entries in {−1, 0, 1}
and at most one 1 and at most one −1, so it is an incidence submatrix.

Lemma 4.2. Let A be an incidence submatrix such that the last row of A contains only
0s, and let B be the incidence submatrix created from removing the last row of A. Then
the maximum value of det[A|M ] over all incidence submatrices M is at most the maximum
value of det[B|N ] over all incidence submatrices N .

Proof. Suppose A is an m × (n − 1) matrix. Let K = [A|M ], and let kij denote the (i, j)-
entry of K. We show that there always exists K ′ = [B|N ] such that | det K ′ | ≥ | det K|. We
perform casework based on the number of nonzero entries in the mth row of K.
Notice that kmi = 0 for all i ≤ n − 1, so it follows that there is at most one 1 and at most
one −1 in row m. We divide into cases based on the number of nonzero entries in row m.
Case 1. There are no nonzero entries in row m.

7
 All entries in row m must equal 0 so det K = 0. It follows that any choice of N satisfies
| det K ′ | ≥ | det K| as desired.
Case 2. There is exactly one nonzero entry in row m.
Suppose that the nonzero entry appears in column s. Then by Laplace expansion we
have m
X
det K = (−1)1+j kmj det K1,j = (−1)m+s kms det Km,s ≤ | det Km,s |.
j=1

Notice that no elements of A are in column s since s ≥ n, so the leftmost n−1 columns of
Km,s correspond to B. Furthermore, no new 1s or −1s are added to each row of K in Km,s ,
so the rightmost m − n columns of Km,s form an incidence submatrix. Taking K ′ = Km,s
completes the proof.
Case 3. There are exactly two nonzero entries in row m.
Let the 1 and −1 entries appear in columns a and b respectively. Since swapping
columns does not affect the absolute value of the determinant, without loss of generality
suppose a = m, b = m − 1. Denote the ith column vector of K as ci . The column vectors
cn , cn+1 , . . . , cm−2 , cm−1 + cm form an incidence submatrix by Lemma 4.1. Let K ′ be obtained
by adding the mth column of K to its (m − 1)th column. We can perform Laplace expansion
on K ′ as follows:
m
X
det K = det K ′ = ′
(−1)1+j kmj det Km,j ′
= (−1)2m kmm det Km,m ′
= det Km,m .
j=1

′
Notice that the rightmost m − n columns of Km,m are exactly cn , cn+1 , . . . , cm−1 + cm with
row m removed. Since cn , cn+1 , . . . , cm−1 + cm form an incidence submatrix, the rightmost
m − n columns of Km,m also form an incidence submatrix, so taking K ′ = Km,m completes
the proof.
Thus we have shown the desired result in all cases.

Remark 4.3. In the above proof, N can always be constructed by taking column sum opera-
tions or removing a column from M . Thus if M has no nonzero entries in a given row i, we
can construct N such that | det[A|M ]| ≤ | det[B|N ]| and N has no nonzero entries in row i.

This result additionally shows that any extra rows of zeroes in the left submatrix do not
affect the maximal determinant which can be achieved.

Lemma 4.4 (Edgecut Lemma). Given a multigraph G and an edge e of G, we have that
max det G ≤ max det Gemerge + max det Gecut .

Proof. Suppose G has m edges and n vertices. Let max det G be obtained by the m × m
matrix [MD′ |N ], where N is an incidence submatrix and D is an orientation of G. Without
loss of generality, suppose the row of MD′ corresponding to edge e is the first row. Denote

8
the entry of [MD′ |N ] in the ith row and jth column as aij . Denote the ith row vector of
[MD′ |N ] as ri .
Let r1,merge = (a1 , a2 , . . . , am ) and r1,cut = (b1 , b2 , . . . , bm ) where
( (
a1i if i ≤ n − 1 0 if i ≤ n − 1
ai = and bi =
0 if i ≥ n a1i if i ≥ n.

Then we have r1 = r1,merge + r1,cut . Let Mmerge = [r1,merge r2 · · · rm ] and Mcut =

[r1,cut r2 · · · rm ]. Since the determinant is a multilinear function with respect to its row
vectors, we have that det M = det Mmerge + det Mcut .
We wish to show that det Mmerge ≤ max det Gemerge , so we divide into three cases depend-
ing on the number of nonzero entries in r1,merge .
Case 1. There are no nonzero entries in r1,merge .
Since r1,merge = 0, we must have det Mmerge = 0.
Case 2. There is exactly one nonzero entry in r1,merge .
We can perform Laplace expansion along the first row of Mmerge . Let the nonzero entry
appear in column s, then we have
m
X
det Mmerge = (−1)1+j a1j det M1,j = (−1)1+s a1s det M1,s ≤ | det M1,s |.
j=1

In particular, the remaining n − 2 columns of M1,s correspond to columns in MD′ , so they

form an incidence submatrix A. Let B be the incidence submatrix obtained by removing all
of the rows of zeroes from A, then det Mmerge ≤ | det M1,s | ≤ max det[A|N1 ] ≤ max det[B|N2 ]
over all incidence submatrices N1 , N2 by repeated application of Lemma 4.2 on the empty
rows of M1,s .
Now we claim that B is exactly the truncated incidence submatrix of Gemerge . In particular,
for each edge e′ ∈ G such that e′ ̸= e such that e has a nonzero entry in column s, if the
edge was incident to both v0 and vs , then it is deleted since its row becomes a row of 0s.
Otherwise, if its head was at vs , then the new head of the edge becomes v0 . On the other
hand, if its tail was at vs , then the new tail of the edge becomes v0 . All other edges remain
unchanged, and this is exactly the construction of Gemerge .
Thus it follows that max det[B|N ] = max det Gemerge , which completes the proof.
Case 3. There are exactly two nonzero entries in r1,merge .
Suppose the 1 and −1 entries in row 1 appear in columns a and b respectively. Without
loss of generality, suppose a = n − 1, b = n − 2.
′
Then let Mmerge be the matrix created by adding column n − 1 to column n − 2 in
Mmerge . Let the ith column vector of Mmerge be denoted by ci . It follows that the leftmost
′
n − 2 columns of Mmerge are exactly c1 , c2 , . . . , cn−3 , cn−2 + cn−1 . By Lemma 4.1, these
column vectors form an incidence submatrix. Furthermore, the rows of all zeroes in this new

9
 ′
submatrix are exactly those rows such that cn−2 = −cn−1 ̸= 0 in Mmerge . Let Mi,j denote the
′ ′
minor of Mmerge created by removing the ith row and jth column and let aij be the entry of
′ ′
Mmerge in row i and column j. By Laplace expansion along the first row of Mmerge , we have
m
X
′
det Mmerge = det Mmerge = (−1)1+j a′1j det M1,j
′ ′
≤ | det M1,n−1 |.
j=1

′
Notice that the leftmost n − 2 columns of M1,n−1 still form an incidence submatrix
A. Let B be obtained by removing the rows of zeroes from A. We have det Mmerge ≤
′
| det M1,n−1 | ≤ max det[A|N1 ] ≤ max det[B|N2 ] over all incidence submatrices N1 , N2 by
repeated application of Lemma 4.2.
We claim that B is exactly the truncated incidence matrix of Gmerge . First, for each edge
e ∈ G such that e′ ̸= e, if e′ is not incident to vn−1 then it is unaffected. If e′ between vn−2
′
′
and vn−1 , then it is mapped to a row of zeroes in M1,n−1 , so it is deleted in B. Otherwise, if
′ ′ ′
e has its head at vn−1 , then the new head of e is mapped to vn−2 in Mmerge . Similarly, if e′
′ ′
has its tail at vn−1 then the new tail of e is mapped to vn−2 in Mmerge . This is exactly the
construction for the desired contraction, proving the claim.
Hence it follows that max det[B|N2 ] = max det Gemerge , showing the desired claim in all
cases.
Now, notice that the leftmost n−1 columns of Mcut are exactly the leftmost n−1 columns
of M , except row 1 becomes a row of zeroes. By Lemma 4.2, we can remove this row to
obtain an incidence submatrix A such that | det Mcut | ≤ | max det[A|N ]| over all incidence
submatrices N . In particular, since only the row corresponding to edge e is removed from
MD′ to obtain A, it follows that A is exactly the truncated incidence matrix of Gecut , so
det Mcut ≤ max det Gcut .
Thus, it follows that max det G = det M = det Mmerge + det Mcut ≤ max det Gemerge +
max det Gecut , completing the proof.

Theorem 4.5. For any nonplanar multigraph G, we have that max det G ≤ sp(G) − 18.
Proof. One can check using brute force that the value of max det K3,3 = 63 and the value of
max det K5 = 100.
Recall that we have that for any multigraph G, we have that sp(G) = sp(Gemerge ) +
sp(Gecut ). Let us define the excess exc(G) of a multigraph G to be equal to sp(G)−max det G.
It follows from Lemma 4.4 that

exc(G) = sp(G) − max det G

≥ sp(Gemerge ) − max det Gemerge + (sp(Gecut ) − max det Gecut )

= exc(Gemerge ) + exc(Gecut )

and from Theorem 3.3 that exc(G) ≥ 0 for all multigraphs G and edges e ∈ G.
It follows that exc(G) ≥ exc(Gemerge ) and exc(G) ≥ exc(Gecut ). Thus the excess of a
multigraph G is at least that of any multigraph that can be obtained from a sequence of

10
edge contractions and deletions on G. Thus, since Wagner’s theorem asserts that one can
obtain either K3,3 or K5 from a series of edge contractions and deletions on any nonplanar
multigraph G, it follows that exc(G) ≥ min{exc(K3,3 ), exc(K5 )} = 18, giving the desired
result.
This result shows that we cannot find a similar construction for nonplanar multigraphs,
which shows that a determinant equal to the number of spanning trees cannot be achieved for
nonplanar multigraphs. However, it does not rule out the possibility that a large nonplanar
multigraph could have a large number of spanning trees with a positive but small excess,
resulting in a bigger determinant than the maximal planar multigraph. We rule out this
possibility for a special class of graphs:

Theorem 4.6. If G is a subdivision of K3,3 or K5 with m edges, then max det G ≤ max spplanar
m .

The proof for this theorem is in Appendix C.

We also hope to generalize this type of result to arbitrary nonplanar multigraphs, which
would prove that the full equality holds:

Conjecture 4.7. For all positive integers m, we have max detm = max spplanar
m .

5 Asymptotic Bounds
In this section, we study asymptotic behaviors of max detm and max spplanar
m as functions
of m. In particular, we give several lower and upper bounds on these two quantities.

5.1 Lower Bounds

First, we give the following lower bound on max detm , which is tight for m ∈ [6] ∪ {8}.

Theorem 5.1. For all m ∈ N, there exists an m × m matrix M that is the concatenation of
two incidence submatrices with
(
Fm+1 if m is odd
det M =
Fm+1 + Fm−1 − 2 if m is even,

where Fi denotes the ith Fibonacci number.

Proof. Throughout this proof, scripts are taken modulo m. Consider the even case first. We
define M = (aij )i,j where


 1 if j ≡ i − 1 (mod m)

1 if j ≡ i (mod m)
aij =


 −1 if j ≡ i + 1 (mod m)

0 otherwise.

11
Since exchanging two columns preserves the absolute value of the determinant, the sets of
odd-indexed and even-indexed columns partition M into two incidence submatrices
P Qof M .
We show that det M = Fm+1 +Fm−1 −2. We use the formula det M = σ sgn(σ) i aiσ(i) .
Fix a permutation σ corresponding to a nonzero term of the summation. Mark each row i
with ℓi = U, C, L depending on whether σ(i) ≡ i + 1, i, i − 1 (mod m), respectively.
We now perform casework on the permutation σ. If ℓi = ℓi+1 = U for some i ∈ [m] it
follows that ℓi = U for all i and σ = (1 2 3 · · · m). Similarly, if ℓi = ℓi+1 = L for some
i ∈ [m], then ℓi = L for all i. These two cases contribute sgn(1 2 3 · · · m)(−1)m = −1 and
sgn((1 2 3 · · · m)−1 )1m = −1 to the determinant, respectively.
For all the remaining permutations, if ℓi = U then σ(i) = i + 1 so ℓi+1 ̸= C. Since there
are no two Us in a row, we have ℓi+1 = L . Conversely, if ℓi = L then σ(i) = i−1 so ℓi−1 ̸= C.
Thus ℓi−1 = U.
Hence σ consists of blocks of C and blocks of UL in a cycle of m letters. In particular, for
a permutation σ with k blocks of UL, the term contributes sgn(σ)(−1)k to the determinant.
However, since σ can be written as the product of k adjacent transpositions, with one between
the two elements of each UL block, it follows that sgn(σ) = (−1)k , so the permutation
contributes exactly 1 to the determinant.
Finally, to count the total number of such blocks, we proceed with casework on whether
1 and n are together in a UL block. If not, the problem becomes equivalent to the number
of ways to partition n consecutive numbers into consecutive blocks of sizes 1 and 2, which
is equal to Fn+1 (this can be proved either by an inductive argument or by generating
functions). Otherwise, the problem becomes equivalent to the number of ways to partition
the n − 2 remaining numbers 2, 3, . . . , n − 1 into consecutive blocks of sizes 1 and 2, which
is equal to Fn−1 . Thus, adding all of the cases show that det M = Fn+1 + Fn−1 − 2.
For the odd case, we instead define


 1 if j = i − 1

1 if j = i
aij =


 −1 if j = i + 1

0 otherwise.

Again, the sets of odd-indexed and even-indexed columns partition M into two incidence
submatrices of M .
We now define ℓi = U, C, L for each row i in the same way, with the only difference being
that ℓ1 ̸= L and ℓm ̸= U. The only permutations which are removed from the cases are the
case where all ℓi = L, the case where all ℓi = U, and the case when n, 1 are in a UL block
together. It still holds that each permutation formed from C and UL blocks contributes 1
to the determinant. Thus, we have det M = Fn+1 .
√
1+ 5
Corollary 5.2. We have that max detm = Ω(ϕm ) where ϕ = 2
≈ 1.618.

In combinatorial optimization, one might be interested in fractionality of extreme points

of a polyhedron. For instance, Pritchard [6] showed that there exist extreme points of the

12
Held-Karp relaxation (also known as the subtour elimination relaxation) of the symmetric
traveling salesman problem with fractionality exponentially small in the number of vertices.
We show that the minimal fractionality of an extreme point of a polyhedron P = {x : M x ≤
b} where M is the concatenation of two incidence submatrices and b is an integral vector can
be exponentially small in the dimension of M by slightly modifying the above construction.

Theorem 5.3. The minimum fractionality of a vector x = M −1 b over all invertible matrices
M which can be written as the concatenation of two incidence submatrices and over all
1
integral vectors b is at most Fn+1 .

The proof of this theorem proceeds similarly to that of Theorem 5.1 and is in Appendix D.
We can also use Theorem 3.1 to obtain lower bounds on max detm from lower bounds on
max spplanar
m . In particular, the bound from Theorem 5.1 can also be obtained by using the
directed planar dual construction on the wheel graph with m edges. Furthermore, a classical
result from statistical physics on the spanning tree entropy of square lattices implies a better
lower bound on max spplanar
m .
Let C denote Catalan’s constant 112 − 312 + 512 − 712 + · · · . Let L(Z2 ) denote the regular
square lattice, which has vertex set Z and rook-adjacent edges. A grid graph is defined to
be a finite connected subgraph of L(Z2 ) and is always planar.

Theorem 5.4 (Wu, 1977, [9]). For all m ∈ N, there exists a grid graph G with m edges such
that sp(G) = Ω(e2Cm/π ).

Corollary 5.5. We have that max detm = Ω(e2Cm/π ) where e2C/π ≈ 1.791.

5.2 Upper Bounds

Unlike the lower bounds, upper bounds on max spplanarm cannot be directly transformed
into upper bounds on max detm without proving Conjecture 4.7. In this subsection, we
provide several upper bounds on max detm , which automatically give upper bounds on
max spplanar
m . First, we have the following trivial upper bound:
m


Theorem 5.6. For any positive integer m, we have that max detm ≤ ⌊m/2⌋ .

Proof. From Theorem 3.3, we know that max detm ≤ sp(G) over all multigraphs G with m
edges. Suppose G has n vertices, then since each

spanning

tree consists of exactly n−1 edges,
m m m


we have that sp(G) ≤ n−1 . Since maxn n−1 = ⌊m/2⌋ , we obtain the desired result.
m


Since ⌊m/2⌋ ∼ 2m , this is an exponential upper bound. According to a user JimT on
Mathematics Stack Exchange, he claimed to have found an elementary proof that max spplanar
m ≤
m 3
1.93 . We were unable to find their proof online or in the literature, but we were able to
obtain an exponentially stronger result by bounding the number of nonzero entries in the
matrix:
3
See https://math.stackexchange.com/questions/2832917/maximum-number-of-spanning-trees-
of-a-planar-graph-with-a-fixed-number-of-edges/3969689.

13
 √
Theorem 5.7. For any positive integer m, we have that max detm ≤ ( 3 7)m ≈ 1.913m .

Proof. We proceed by strong induction. Note that the result holds for m = 1 trivially. Given
that the claim is true for all m = 1, 2, . . . , k − 1, we show it is true for m = k.
Suppose max detm is achieved by the absolute value of the determinant of the matrix
M = [MD′ |MF′ ] = (aij )i,j where D, F are orientations of multigraphs G, H, respectively, each
with m edges. Let n be the number of vertices of G. Consider the m × (m + 2) rectangular
matrix [MD |MF ]. This matrix has at most 4m nonzero entries since each row has at most
two nonzero entries in MD , MF respectively. Since there are m + 2 columns, there exists a
column with at most ⌊ m+2 4m
⌋ = 3 nonzero entries. If this column is a column of [MD′ |MF′ ]
then M has a column with at most three nonzero entries.
Otherwise, without loss of generality suppose the column is the truncated column of the
incidence matrix of G. Let c1 , c2 , . . . , cm be the column vectors of M . Since the sum of
the column vectors of MD is equal to 0, it follows that the truncated column c0 is equal to
−1 · (c1 + c2 + · · · + cn−1 ). It follows that
 
det M = det c1 c2 . . . cm
 
= det c1 + c2 + · · · + cn−1 c2 . . . cm
 
= − det c0 c2 . . . cm .

Thus, we can choose a different matrix M ′ formed by replacing column c1 with c0 in M such
that | det M ′ | = det maxm and such that there is a column of M ′ with at most three nonzero
entries. Replacing M with M ′ , we can assume M has a column with at most three nonzero
entries.
Since we can switch the order of the rows and columns without affecting the absolute
value of the determinant, assume without loss of generality that the first column of M has at
most three nonzero entries and these k ≤ 3 nonzero entries appear in the rows 1, 2, . . . , k. Let
ri denote the ith row vector of M , and let r1,0 = (a1 , a2 , . . . , am ) and r1,1 = (b1 , b2 , . . . , bm )
where ( (
a1i if i ≤ n − 1 0 if i ≤ n − 1
ai = and bi =
0 if i ≥ n a1i if i ≥ n.
Analogously define r2,0 , r2,1 , r3,0 , and r3,1 . Then from the multilinearity of the determinant,
we have
X  
det M = det r1,i r2,j r3,k r4 . . . rm .
i,j,k∈{0,1}

However, since the matrix formed by r1,1 , r2,1 , r3,1 , r4 , . . . , rm has no nonzero entries in
its first column, it has determinant 0. Furthermore, for any of the other matrices in the
summation, each row has the property that it either has no nonzero entries in MA′ or has
no nonzero entires in MB′ , so it can be removed using Lemma 4.2 without decreasing the
absolute value of the determinant and without losing the same property in the other rows
by Remark 4.3.

14
 Thus for all i, j, k ∈ {0, 1} and (i, j, k) ̸= (1, 1, 1) we can apply Lemma 4.2 to the first
three rows in succession to obtain a reduction to an (m − 3) × (m − 3) matrix Ni,j,k such that
Ni,j,k can be written as the concatenation of two incidence submatrices and | det Ni,j,k | ≥
| det(r1,i , r2,j , r3,k , r4 , . . . , rm )|. Hence,
X √ √
Ni,j,k ≤ 7 · ( 7)m−3 = ( 7)m ,
3 3
det M =
i,j,k∈{0,1},(i,j,k)̸=(1,1,1)

as desired.

6 Conclusion and Future Directions

In Theorem 3.1, we establish an interesting connection between max detm and max spplanar m .
In addition, we provide some evidence supporting Conjecture 4.7 that the reverse direction
also holds, including Theorems 4.5 and 4.6, and Table 1 in the appendix. If this conjecture
holds, then it implies an algebraic characterization for computing the maximum number of
spanning trees in a planar multigraph.
There are several potential approaches to prove Conjecture 4.7. One possible approach is
to devise an algorithm that transforms a square matrix [MD′ 1 |MD′ 2 ] with D1 and D2 nonplanar
to some other square matrix [MD′ 1 |MD′ ′ ] such that D2′ has fewer spanning trees than D2 , while
2
preserving the determinant. Another possible method is to extend Theorem 4.6 to show that
any nonplanar graph G underperforms some planar graph with the same number of edges in
terms of max det G. In addition, the application of Lemma 4.4 in the proof of Theorem 4.5
is extremely loose. For most nonplanar graphs, one can obtain many copies of K5 or K3,3 as
minors, which could potentially be used to prove stronger versions of Theorem 4.5.
For asymptotic bounds on max detm and max spplanar m , we have established that c1 1.791m ≤
planar m
max spm ≤ max detm ≤ c2 1.913 for some constants c1 , c2 > 0. It would be interesting to
improve these bounds and to close the gap. Matching lower and upper bounds would also
imply Conjecture 4.7.

Acknowledgments
I’d like to thank my mentor, Yuchong Pan, for providing the guidance, support, and
the resources that led to this research project. I’d like to thank Prof. Michel X. Goemans
for proposing my research topic and his talks on combinatorial optimization. I’d like to
thank Dr. Tanya Khovanova for her invaluable advice along the way. I’d like to thank Peter
Gaydarov, Allen Lin, and an anonymous reviewer for their helpful suggestions. I’d like to
thank RSI, CEE, and MIT for making all of this possible. I’d also like to thank all of the
sponsors of RSI.

15
 References
[1] Seth Chaiken and Daniel J. Kleitman. Matrix tree theorems. Journal of combinatorial
theory, Series A, 24(3):377–381, 1978.

[2] Michel X. Goemans and Yuchong Pan. A counterexample to box-half-integrality of the

intersection of crossing submodular flow systems. 2023.

[3] Gustav Kirchhoff. Ueber die auflösung der gleichungen, auf welche man bei der unter-
suchung der linearen vertheilung galvanischer ströme geführt wird. Annalen der Physik,
148(12):497–508, 1847.

[4] Mordechai Lewin. A generalization of the matrix-tree theorem. Mathematische

Zeitschrift, 181:55–70, 1982.

[5] Saunders Mac Lane. A combinatorial condition for planar graphs. Seminarium Matemat.,
1936.

[6] David Pritchard. k-edge-connectivity: approximation and LP relaxation. In Approxima-

tion and Online Algorithms: 8th International Workshop, WAOA 2010, Liverpool, UK,
September 9-10, 2010. Revised Papers 8, pages 225–236. Springer, 2011.

[7] W. T. Tutte. Graph theory, encyclopedia of mathematics and its applications, 1984.

[8] Klaus Wagner. Über eine eigenschaft der ebenen komplexe. Mathematische Annalen, 114
(1):570–590, 1937.

[9] F. Y. Wu. Number of spanning trees on a lattice. Journal of Physics A: Mathematical

and General, 10(6):L113, 1977.

16
 A Exact Values of max detm , max spplanar
m and max spgeneral
m

m max detm max spplanar

m max spgeneral
m
1 1 1 1
2 2 2 2
3 3 3 3
4 5 5 5
5 8 8 8
6 16 16 16
7 24 24 24
8 45 45 45
9 75 75 81
10 130 130 135

Table 1: Values of max detm , max spplanar

m and max spgeneral
m for m ∈ [10].

B Alternative Proof of Theorem 3.1

Proof. Let C(G) be a cycle basis for the circulation space of G. Since G is planar, by Mac
Lane’s planarity criterion there exists a 2-basis K for C(G). For each cycle in this basis, we
can define a fixed direction of flow around the cycle such that every vertex is the source of
one edge flow and the sink of another edge flow of the cycle. In particular, if G has m edges
and n vertices then K can be written as an m × (m − n + 1) matrix where each column
corresponds to a basis element and an entry in the row corresponding to a given directed
edge of the chosen orientation D of G is 1 if it points along the flow of the basis element,
−1 if it points against the flow, and 0 if it does not belong to the basis element.
Let c1 , c2 , . . . , cn−1 be the column vectors of MD′ and let d1 , d2 , . . . , dm−n+1 be the columns
of K. Importantly, we know that K is an orthogonal complement of G since for any i ∈
[n − 1], j ∈ [m − n + 1], the value of ci · dj represents the total amount of accumulation of the
flow corresponding to dj at vertex i, which is always equal to 0 for cycles. As in the previous
proof,
2 T  ′
det MD′ K = det MD′ K
  
MD K
" #
MD′T  ′ 
= det M D K
KT
" #
MD′T MD′ 0
= det
0 KT K
= det MD′T MD′ · det K T K.

17
 Linearly independent sets of rows of K correspond to linearly independent edges in the
directed multigraph D′ which has K as its incidence submatrix, and det MD′T′ MD′ ′ is exactly
the number of such sets (i.e. the number of spanning trees of D′ ). Hence it suffices to show
that there is a bijection between the spanning trees of G and the linearly independent sets
of K in order to show that det K T K = det MD′T MD′ .
In K, I claim that the rows corresponding to some set of r = m + 1 − n edges are linearly
independent if and only if we can remove any r − 1 of them from the graph, and the last
remaining edge is always part of a cycle. This claim is equivalent to the assertion that for
each of those r edges, there is an element of the column space of K such that the entries of
the other r − 1 rows are 0 and not the specified edge.
It is clear that this is algebraically equivalent to the linearly independence of those r
rows of K. Furthermore, looking at the edges of G which are not one of these r edges, this
condition becomes equivalent to enumerating the sets of n − 1 edges of G with the property
that the addition of any other edge of G creates a cycle containing the new edge in the
resulting subgraph. I claim that this holds if and only if those n − 1 edges form a spanning
tree. If they do not form a spanning tree then the resulting subgraph of G obtained by
picking an edge between any two disjoint components would violate the condition. On the
other hand, if the n − 1 edges form a spanning tree, then there is a path between any two
vertices of G, so simply combining the new edge with the path between the endpoints of the
new edge gives the desired cycle. Hence the bijection holds, so it follows that K has the
same number of linearly independent sets as spanning trees of G, which finishes.

C Proof of Theorem 4.6

Lemma C.1. Let A be a m × m matrix with row vectors r1 , r2 , . . . , rm . If the first column
vector of A is equal to (1, −1, 0, 0, · · · , 0) then det A is equal to the determinant of the
(m − 1) × (m − 1) matrix formed by the rows r1 + r2 , r3 , r4 , . . . , rm with their first entries
removed.

Proof. Let A = (aij )i,j . Let the matrix B be obtained by adding the first row to the second
row of A. The only nonzero entry in the first column of B is a11 = 1. We can perform
Laplace expansion along the first column of B as follows:
m
X
det A = (−1)i+1 ki1 det Ki,1 = (−1)1+1 k11 det K1,1 = det K1,1 .
i=1

Notice that K1,1 is exactly the (m − 1) × (m − 1) matrix formed by the rows r1 +

r2 , r3 , r4 , . . . , rm with their first entries removed, finishing.
Now we have the proof of Theorem 4.6:
Proof. First, we consider the case when G is a subdivision of K3,3 . For each edge e of the
original K3,3 , let Ge denote a tuple of edges (Ge1 , Ge2 , . . . , Gek ) such that Ge1 , . . . , Gek are the
edges created from subdivisions on the original edge e and such that Gei is incident to Gei+1

18
for all i ∈ [k − 1]. Since K3,3 has 6 vertices and 9 edges, and since each subdivision adds
a vertex and an edge, it follows that G has m − 3 vertices. Suppose that max det G is
attained by the matrix M = [MD′ |N ] = (aij )i,j where N is the truncated incidence matrix of
another multigraph G′ and D is some orientation of G. Without loss of generality, choose
the orientation so that the head of Gei is the tail of Gei+1 for all i ∈ [k − 1]. Then G′ has
m + 2 − (m − 3) = 5 vertices.
Now each pair of edges Gei , Gei+1 in G, there exists a vertex v ∈ G such that deg v = 2
and v is incident to Gei , Gei+1 . We call this vertex the ith child of edge e. Let this vertex
correspond to column i in MD′ , and let the columns of M be c1 , c2 , . . . , cm . Let n be the
number of vertices of G. If column i is the truncated column, then since the sum of the
columns of MD is 0, we know c0 = −1·(c1 +c2 +·  · ·+cn−1 ) is the column
 corresponding to the
missing vertex. We can then replace M with c0 c2 c3 · · · cm , whose determinant has
the same absolute value since it can be written as the composition of column sum operations
and scaling columns by −1. Thus we can assume this column is not the truncated column
and can freely choose which vertex corresponds to the truncated column.
Now for each edge e of the original K3,3 , let the truncated column be a vertex of the
original K3,3 which is not an endpoint of e. Suppose k is the number of elements of Ge .
Rearrange the columns of MD′ so that the ith child of e is in the ith column for all i = [k − 1].
Then rearrange the rows of MD′ so that Gei lies in the ith row for all i ∈ [k]. Let
r1,0 = (a1 , a2 , . . . , am ) and r1,1 = (b1 , b2 , . . . , bm ) where
( (
a1i if i ≤ n − 1 0 if i ≤ n − 1
ai = and bi =
0 if i ≥ n a1i if i ≥ n.

Analogously define ri,0 , ri,1 for all i ∈ m.

We can now repeatedly apply Lemma C.1 to M for k − 1 times to obtain a (m − k +
1) × (m − k + 1) matrix M ′ . Note that each time the operation is applied, a child of e is
effectively contracted into one of its neighbors, so the first row of M ′ corresponds to the
edge e of the original K3,3 . Now notice that the first row of the resulting matrix is exactly
Pk Pk
i=1 ri,0 + i=1 ri,1 , which is linear in terms of the sum of all ri,1 . Since the determinant
is a multilinear function, it follows that since M is maximal, that we can assume that ri,1
is equal for all i since we are maximizing a linear function on the sum of all ri,1 . This does
not change the left submatrix of M . Thus, by setting ri,1 for i ∈ [k] to be identical in M , we
can obtain another matrix M = [MD′ |N ′ ] which attains the maximal possible | det M | such
that every edge in Ge has identical entries in submatrix N .
We can then repeat this algorithm for all edges e of the original K3,3 to obtain a matrix
M with maximal | det M | and such that there are at most 9 distinct values of ri,1 : one for
each e. Thus it follows that the underlying simple graph of G′ has at most 9 edges and 5
vertices, so it must be planar. Hence G′ is a planar graph, so we have that max det G =
| det M | ≤ sp(G′ ) ≤ max spplanarm by Theorem 3.3, finishing.
Now, consider the case when G is a subdivision of K5 . Identically to the proof of K3,3 ,
we define Ge = (Ge1 , Ge2 , . . . , Gek ) for each edge e of the original K5 with the same properties.

19
Again, let max det G be attained by the matrix M = [MD′ |N ] = (aij )i,j where N is the trun-
cated incidence matrix of another multigraph G′ and D is some orientation of G. Without
loss of generality, choose the orientation so that the head of Gei is the tail of Gei+1 for all
i ∈ [k − 1]. Then G′ has m + 2 − (m − 5) = 7 vertices.
Using the same application of Lemma C.1 as in the K3,3 case, we can choose a matrix
M such that for every edge e of the original K5 , all rows i such that their corresponding
edge is in Ge have identical values for ri,1 . We will denote this shared value as re . Thus, the
underlying simple graph G′′ of the multigraph G′ whose truncated incidence matrix is N has
at most 10 edges on 7 vertices.
If G′′ is planar then G′ is planar so we have max det G = | det M | ≤ sp(G′ ) ≤ Pmax spplanar
m
by Theorem 3.3. Otherwise, G′′ must be nonplanar. Furthermore, since 2|E| = v deg v, it
follows by the pigeonhole principle that there exists a vertex v with degree at most ⌊ 20 7
⌋=2
′′ ′′
in G . We can then divide into cases based on the degree of this vertex in G .
If the degree of v in G′′ is 0, then its degree in G′ is also 0, so it follows that G′ is
disconnected. Hence det M ≤ sp(G′ ) = 0, finishing trivially.
If the degree of v in G′′ is 1, then removing v from G′′ gives a graph on 6 vertices with at
most 9 edges. This graph must be nonplanar, so it must equal exactly K3,3 . Hence it follows
that every edge e of the original K5 maps to a distinct edge of G′′ . Let e be the edge of the
original K5 which maps to the only edge incident to v. As in the K3,3 case, rearrange the
columns of MD′ so that the ith child of e is in the ith column for all i = [k − 1] where k is
the number of elements of Ge . Then rearrange the rows of MD′ so that Gei lies in the ith row
for all i ∈ [k]. Define ri,0 and ri,1 analogously as in the K3,3 case.
We can then apply Lemma C.1 to M for k − 1 times in succession to obtain an (m − k +
1) × (m − k + 1) matrix M ′ such that the entries of the first row of M ′ in the left submatrix
corresponds to the directed edge between the tail of Ge1 and the head of Gek , and the entries
of the first row of M ′ in the right submatrix N is equal to k · re . Notice furthermore that
the only nonzero entry in the column of N corresponding to vertex v lies in row 1, so the
entries of the first row in the left submatrix do not affect the determinant since they do not
appear in the Laplace expansion along the specified column v.
Thus, we can change the head of Gek to be another vertex of the original K5 which is
not an endpoint of e without affecting | det M |. This has the effect of changing the left
submatrix to become the truncated incidence matrix of a different multigraph H such that
H is a subdivision of some multigraph on 5 vertices with 10 edges that is not K5 . This
multigraph is necessarily planar, so it follows that max det G ≤ det M ≤ max det H ≤
sp(H) ≤ max spplanar
m by Theorem 3.3, finishing.
Finally, consider the case when the degree of v in G′′ is 2. It follows that G′′ is the
subdivision of another graph H on 6 vertices and at most 9 edges. Since planarity is preserved
under the subdivision operation, H must be K3,3 with exactly 9 edges, so no two edges of
e map to the same edge in G′′ . So let e1 , e2 be the two edges of the original K5 which map
to the two edges incident to v in G′′ . Since we can freely arrange the columns and rows of
M , as long as the columns of MD′ and N are preserved, so we can in fact apply Lemma C.1
on an column with two nonzero entries such that one entry is 1 and the other is −1. Thus,

20
we can repeatly perform Lemma C.1 on all of the children of edges e1 , e2 . Let k1 , k2 be the
number of elements of Ge1 , Ge2 respectively.
After all of these operations along with row swaps, we obtain a (m − k1 − k2 + 2) × (m −
k1 − k2 + 2) matrix M ′ such that the ith row for i ≥ 3 corresponds to a row of M with
the entries corresponding to the children of e1 , e2 removed. Furthermore, the entries of the
first row of M ′ in the left submatrix correspond to a directed edge from the tail of Ge11 to
the head of Gek11 , while the entries of the first row of M ′ in the right submatrix are equal to
k1 · re1 . Similarly, the entries of the second row of M ′ in the left submatrix correspond to a
directed edge from the tail of Ge12 to the head of Gek22 , while the entries of the second row of
M ′ in the right submatrix are equal to k2 · re2 . Let s1 , s2 , . . . , st denote the row vectors of
M ′ , where t = m − k1 − k2 + 2. Define si,0 , si,1 for all i analogously to ri,0 and ri,1 .
We also know that the only two nonzero entries in the column of M ′ corresponding to
the vertex v of G′ belong in s1 , s2 . Since we can scale s1 , s2 by −1 by flipping the direction
of all of the edges in Ge1 , Ge2 in M , respectively, we can assume without loss of generality
that the nonzero entries in s1 , s2 are k1 , −k2 respectively. So we have:

det M ′ = k1 · k2 · ks11 ks22 s3 s4 · · · st

 

We can then apply Lemma C.1 on ks11 ks22 s3 s4 · · · st so that the determinant of
 

this matrix is equal to the determinant of the matrix formed by ks11 + ks22 , s3 , s4 , . . . , st after
removing the entries of each vector corresponding to vertex v of G′ . Notice that the vector
s1
k1
+ ks22 is linear in sk1,0
1
+ sk2,0
2
. From the multilinearity of the determinant it follows that if A
is the matrix obtained by replacing s1 with s2,0 + s1,1 in M ′ and B is the matrix obtained
by replacing s2 with s1,0 + s2,1 in M ′ , then

det A + det B
det M ′ =
2
. So without loss of generality, assume that | det A| ≥ | det M ′ |.
We can obtain A instead of M ′ from the same applications of Lemma C.1 on M by
modifying M into K by replacing the tail of Ge11 with the tail of Ge12 and the head of Gek11
with the head of Gek22 . This changes only the value of s1,0 to equal the value of s2,0 in the
matrix M ′ , which is exactly the change to create matrix A.
Hence | det K| = | det A| ≥ | det M ′ | = | det M |. However, from the construction of K,
we know that the multigraph H whose truncated incidence matrix is the left submatrix of K
can be expressed as the subdivision of a graph with 5 vertices and 10 edges with some pair of
edges sharing the same endpoints. Thus H is the subdivision of a graph which is planar, and
must also be planar. Hence we have det max G = | det M | ≤ | det K| ≤ sp(H) ≤ max spplanarm
by Theorem 3.3, finishing.
Thus we have exhausted all cases, so we are done.

21
 D Proof of Theorem 5.3
Proof. For all n ∈ N, we define


 1 if j = i − 1

1 if j = i
aij =


 −1 if j = i + 1

0 otherwise.

Fix a permutation σ corresponding to a nonzero term of the summation. Mark each row
i with ℓi = U, C, L depending on whether σ(i) ≡ i + 1, i, i − 1 (mod m), respectively. We
compute the determinant of the minors M1,1 and M1,2 , respectively. The case of ℓi = U for all
i and the case of ℓi = L for all i contribute zero to the summation. Hence, we assume that σ
is composed of blocks of C and UL over a cycle of m elements. Note that computing det M1,1
is equivalent to fixing ℓ1 = C, and freely placing rows 2, 3, . . . , n into blocks of sizes 1 and 2,
yielding Fn solutions each contributing 1 to the determinant. This shows that det M1,1 = Fn .
Computing det M1,2 is equivalent to fixing rows 1, 2 into a UL block, and freely placing rows
3, 4, . . . , n into blocks of sizes 1 and 2, yielding −Fn−1 for the determinant.
Note that gcd(Fn , Fn−1 ) = gcd(Fn−1 + Fn−2 , Fn−1 ) = gcd(Fn−2 , Fn−1 ). By an inductive
argument on n, it follows that gcd(Fn , Fn−1 ) = gcd(1, 1) = 1. By Bezout’s theorem, there
exist r, s ∈ Z such that rFn − sFn−1 = 1. Now, let b ∈ Nn such that b1 = r, b2 = s, and
bk = 0 for k ≥ 3. By Cramer’s rule, the first coordinate x1 of the solution to M x = b satisfies

det(M1 ) rFn − sFn−1 1

x1 = = = .
det(M ) Fn+1 Fn+1

This completes the proof.

22