My postdoc mentor suggested me to think about some problems regarding the critical graph of a random graph, so I decided to sit down and read some basic words about this. I might write more on this if I want to store more summaries.
Critical group of a graph. When we consider a graph, we always mean a finite graph. Let G be a graph with n verticies. Given a vertex v, the degree \deg(v) is the number of paths ending at v in an arbitrarily small neighborhood of v. Order the vertices v_{1}, \dots, v_{n} of G. The diagonal matrix D = D(G) of G is the n \times n diagonal matrix whose diagonal entries are given by \deg(v_{1}), \dots, \deg(v_{n}). The adjacent matrix A = A(G) of G is the n \times n matrix whose entries are given by defining its (i,j) entry to be the number of edges between v_{i} and v_{j}. The Laplacian L = L(G) of G is defined to be D - A.
Convention. We will only deal with graphs whose vertices don't have self nodes. Moreover, we will assume that every pair of vertices have at max 1 edge in between. With this convention, note that
- A(G) is a (0, 1) symmetric matrix with 0's on the diagonal, and
- The entries of any row or column of L(G) = D(G) - A(G) add up to 0.
Theorem 1. If G is connected, then \ker(L(G)) is a cyclic group generated by (1, 1, \dots, 1).
Proof. Since the entries of any row of L = L(G) add up to 0, we have (1, 1, \dots, 1) \in \ker(L). Conversely, say x = (x_{1}, \dots, x_{n}) \in \ker(L) \subset \mathbb{Z}^{n}. Even without connectedness hypothesis, we have 0 = x^{T}Lx = \sum (x_{i} - x_{j})^{2}, where the sum is over 1 \leq i, j \leq n such that v_{i} and v_{j} are connected by an edge. Since G is connected, all the distinct pairs of indices (i, j) must appear in the sum, and since we work over \mathbb{Z}, we must have x_{i} = x_{j}. Thus, we have m := x_{1} = \cdots = x_{n}, so (x_{1}, \dots, x_{n}) = m (1, 1, \dots, 1) \in \mathbb{Z}(1, 1, \dots, 1), as desired. \Box
Remark. Note that if G is not connected, a similar proof to Theorem 1 shows that \ker(G) is generated by tuples that consist of 0's and 1's, where 1's are located in the coordinates corresponding to vertices that belong to each connected component.
Corollary. The rank of the cokernel \mathrm{cok}(L(G)) of L(G) is 1 so that \mathrm{cok}(L(G)) \simeq \mathbb{Z} \oplus K(G), where K(G) is a finite abelian group.
We call K(G) (up to an isomorphism) the critical group of G.
A divisor of G is an element of \mathbb{Z}^{|V(G)|}, where V(G) is the set of vertices of G. We write \mathrm{Div}(G) := \mathbb{Z}^{|V(G)|}. Writing n := |V(G)|, the degree of a divisor D = (m_{1}, \dots, m_{n}) is defined to be \deg(D) := m_{1} + \cdots + m_{n}.
Remark. It seems that the dual graph of an algebraic curve (whose vertices correspond to the irreducible components) contains quite a bit of information about the curve, which could be a motivation for studying these graph invariants for algebraic geometors. I would not try to make any analogy before I understand this story a bit better. (For example, see p.12 of this article; here, there are also graphs we do not consider in this post.)
The subgroup of divisors of degree 0 is denoted as \mathrm{Div}^{0}(G). Given a divisor (x_{1}, \dots, x_{n}) of G, there are two kinds of chip-firing moves at each vertex v_{i}:
- Fire at v_{i}: takes x_{i} \mapsto x_{i} - \deg(v_{i}) and then add 1 to any adjacent vertex.
- Borrow to v_{i}: takes x_{i} \mapsto x_{i} + \deg(v_{i}) and then subtract 1 to any adjacent vertex.
Any two divisors of G that can be achieved from one to another by a sequence of chip-firing moves are said to be equivalent, and note that this indeed defines an equivalence relation to all divisors. Note that chip-firing moves do not change the degree of a divisor, so this also defines an equivalence relation on the degree 0 divisors.
Theorem 2. If G is connected, then K(G) \simeq \mathrm{Div}^{0}(G)/\sim.
Proof. Since columns of L(G) sum up to 0, we have \mathrm{im}(L(G)) \subset \mathrm{Div}^{0}(G), so \mathrm{Div}^{0}(G)/\mathrm{im}(L(G)) \hookrightarrow \mathbb{Z}^{n}/\mathrm{im}(L(G)) \simeq \mathbb{Z} \oplus K(G). Since \mathrm{Div}^{0}(G)/\mathrm{im}(L(G)) is of rank 0, it must sit inside the subgroup of \mathbb{Z}^{n}/\mathrm{im}(L(G)) corresponding to K(G), which we may also write as K(G). This implies that \mathbb{Z} \simeq \mathbb{Z}^{n}/\mathrm{Div}^{0}(G) \simeq \frac{\mathbb{Z}^{n}/\mathrm{im}(L(G))}{\mathrm{Div}^{0}(G)/\mathrm{im}(L(G))} \simeq \mathbb{Z} \oplus \frac{K(G)}{\mathrm{Div}^{0}(G)/\mathrm{im}(L(G))}. Thus, we have \frac{K(G)}{\mathrm{Div}^{0}(G)/\mathrm{im}(L(G))} = 0, so \mathrm{Div}^{0}(G)/\mathrm{im}(L(G)) = K(G). Hence, to finish the proof, it suffices to show that \mathrm{im}(L(G)) is equal to the set of degree 0 divisors of G that are equivalent to (0, 0, \dots, 0). This immediately follows from observing that firing at v_{i} for an n-tuple (i.e., a divisor) is identical to subtracting the i-th column of L(G) from it. This finishes the proof. \Box
Smith normal forms. Recall that given an n \times n matrix L over \mathbb{Z}, there always exists U, V \in \mathrm{GL}_{n}(\mathbb{Z}) such that ULV = \mathrm{diag}(s_{1}, \dots, s_{r}, 0, 0, \dots, 0) with nonzero positive integers s_{1}, \dots, s_{r} satisfying s_{1} | \cdots | s_{r}. This implies that \mathrm{cok}(L) \simeq \mathrm{cok}(ULV) \simeq \mathbb{Z}^{n - r} \oplus \mathbb{Z}/(s_{1}) \oplus \cdots \oplus \mathbb{Z}/(s_{r}), which shows the uniqueness of s_{1}, \dots, s_{n}. We call the diagonal matrix with the entries s_{1}, \dots, s_{r}, 0, 0, \dots, 0 the Smith normal form of L. One may note that r is the rank of L and recall that U is given by the row operations on L while V is given by the column operations on L over \mathbb{Z} when it comes to achieving the diagonal matrix form.
Thus, if L = L(G) for some connected graph G, we have K(G) \simeq \mathbb{Z}/(s_{1}) \oplus \cdots \oplus \mathbb{Z}/(s_{r}).
Lemma. Keeping the notations as above, for any 1 \leq i \leq r, we have s_{1}s_{2} \cdots s_{i} = \gcd(i \times i \text{ minors of } L).
Proof. This immediately follows from noticing that the gcd of i \times i minors of L is unchanged after any row or column operation. To see the invariance under the row operations, denote by [R_{1} | \cdots | R_{i}] be an i \times i submatrix of L, where R_{1}, \dots, R_{i} are the rows. A row operation turns this matrix into a matrix of the form [R_{1} | \cdots | R_{j} + cR'_{j} | \cdots | R_{i}], and thus the corresponding minor is equal to \det[R_{1} | \cdots | R_{j} | \cdots | R_{i}] + c\det[R_{1} | \cdots | R'_{j} | \cdots | R_{i}]. The proof ends by noticing that \det[R_{1} | \cdots | R'_{j} | \cdots | R_{i}] is one of the i \times i minors before we perform the row operation. The invariance under the column operations can be similarly observed. \Box
Example. Let K_{n} be the complete graph on n vertices. To keep up with our convention, we assume that n \geq 3. We have L(K_{n}) = \begin{bmatrix} n-1 & -1 & -1 & \cdots & -1 \\ -1 & n-1 & -1 & \cdots & -1 \\ \vdots & \vdots & \vdots & \cdots & \vdots \\ -1 & -1 & -1 & \cdots & n-1 \end{bmatrix}. Using Lemma, it is immediate that s_{1} = 1. The only 2 \times 2 minors are 0, n, n(n-2), so Lemma implies that s_{2} = s_{1}s_{2} = n. We already know r = n-1 and since s_{1} | \cdots | s_{n-1}, we have n^{n-2} = s_{2}^{n-2} | s_{1}s_{2} \cdots s_{n-1}. By induction, we can compute that once we delete the first row and the last column of L(K_{n-1}), the determinant of the resulting (n - 1) \times (n - 1) matrix is \pm n^{n-2}. Lemma guarantees that this number is divisible by s_{1} \cdots s_{n-1}, so we have n^{n-2} | s_{1}s_{2} \cdots s_{n-1} | n^{n-2}. Thus, we have s_{2} \cdots s_{n-1} = s_{1}s_{2} \cdots s_{n-1} = n^{n-2}. Since s_{2} | \cdots | s_{n-1}, this implies that s_{2} = s_{3} = \cdots = s_{n-1} = n. Thus, we have K(K_{n}) \simeq (\mathbb{Z}/(n))^{n-2}.
Theorem 3. Given a connected graph G with n vertices, the critical group can be computed by taking the cokernel of the (n - 1) \times (n - 1) matrix L_{ij}(G) obtained by deleting the i-th row and the j-th column of L(G) for any 1 \leq i, j \leq n. In particular, we have \mathrm{cok}(L_{ij}(G)) = |K(G)|.
Proof. This only uses the fact that the entries of any row or column of L(G) add up to 0. If we add all the rows other then the i-th row to the i-th row of L(G), the resulting matrix will only have 0's on its i-th row. This is because the entries of each column add up to 0. Then if we add all the columns other than the j-th column to the j-th column of this matrix, the resulting matrix will only have 0's on its j-th column. This is because the entries of each row add up to 0. This finishes the proof because we already know the rank of L(G) is n-1. \Box
Theorem 4 (
Kirchhoff). If
G is connected, then
|K(G)| is the number of spanning trees of
G.
Corollary (Cayley). The number of spanning trees of K_{n} is n^{n-2}.
No comments:
Post a Comment