math0411153/gm.tex
1: \documentclass[12pt]{amsart}
2: 
3: \usepackage{amssymb,latexsym,amscd}
4: \usepackage{times}
5: 
6: \setlength{\textwidth}{6.2in}
7: \setlength{\textheight}{9.3in}
8: \setlength{\oddsidemargin}{0in}
9: \setlength{\evensidemargin}{0in}
10: \setlength{\voffset}{-0.3in}
11: \setlength{\topmargin}{-0.1in}
12: \setlength{\headsep}{0.2in}
13: 
14: \newtheorem{theorem}{Theorem}[subsection]
15: \newtheorem{corollary}[theorem]{Corollary}
16: \newtheorem{lemma}[theorem]{Lemma}
17: \theoremstyle{definition}
18: \newtheorem{definition}[theorem]{Definition}
19: \newtheorem{example}[theorem]{Example}
20: \theoremstyle{remark}
21: \newtheorem{remark}[theorem]{Remark}
22: \newtheorem{claim}[theorem]{Claim}
23: \numberwithin{equation}{section}
24: 
25: \def\deg{\operatorname{deg}}
26: \def\Diag{\operatorname{Diag}}
27: \def\sort{\operatorname{sort}}
28: \def\trace{\operatorname{trace}}
29: \def\max_degree{\operatorname{max\_deg}}
30: 
31: \title{A majorization bound for the eigenvalues 
32:        of some graph Laplacians}
33: 
34: \author{Tamon Stephen}
35: 
36: \address{Institute for Mathematics and its Applications, 
37: University of Minnesota 
38: % 400 Lind Hall, 207 Church St. S.E., 
39: % Minneapolis, MN 55455
40: }
41: \curraddr{Department of Computing and Software,
42: 1280 Main St.~West,
43: McMaster University,
44: Hamilton, Ontario,
45: Canada
46: L8S 4K1
47: }
48: \email{tamon@optlab.mcmaster.ca}
49: 
50: % \thanks{}
51: % \subjclass[2000]{05C50, 05C07}
52: 
53: \begin{document}
54: \maketitle
55: 
56: \begin{abstract}
57: It is conjectured that the Laplacian spectrum of a graph is majorized
58: by its conjugate degree sequence.  In this paper, we prove
59: that this majorization holds for a class of graphs including
60: trees.  We also show that a generalization of this conjecture to
61: graphs with Dirichlet boundary conditions is equivalent to the 
62: original conjecture.
63: \end{abstract}
64: 
65: % Setup for bibtex
66: % \nocite{*}  % bib entries to include that are not actually cited.
67: \bibliographystyle{plain}
68: % End setup for bibtex.
69: 
70: \section{Introduction}
71: One way to extract information about the structure of a graph
72: is to encode the graph in a matrix and study the invariants of 
73: that matrix, such as the spectrum.  
74: In this note, we study the spectrum of the ``Combinatorial
75: Laplacian'' matrix of a graph.
76: 
77: The {\it Combinatorial Laplacian} of a simple graph $G=(V,E)$ on 
78: the set of $n$ vertices
79: $V=\{v_1, \ldots, v_n\}$ is the $n \times n$ matrix $L(G)$ defined by:
80: $$L(G)_{ij} = \begin{cases} \deg(v_i) &\text{if \ } i=j \\ 
81:                  -1 &\text{if \ } \{i,j\} \in E \\
82:                  0 &\text{otherwise}
83: \end{cases}
84: $$
85: 
86: Here $\deg(v)$ is the {\it degree} of $v$, that is number of edges on $v$.
87: The matrix $L(G)$ is positive semidefinite, so its eigenvalues are real
88: and non-negative.
89: We list them in non-increasing order and with multiplicity:
90: $$\lambda_1(L(G)) \ge \lambda_2(L(G)) \ge \ldots \ge \lambda_{n-1}(L(G)) \ge \lambda_n(L(G)) = 0$$
91: When the context is clear, we can write $\lambda_i(G)$ or simply $\lambda_i$.
92: We abbreviate the sequence of $n$ eigenvalues as $\lambda(L(G))$.
93: 
94: We are interested in the conjecture of Grone and Merris (``GM'')
95: that the spectrum $\lambda(L(G))$
96: is majorized by the conjugate partition of the (non-increasing) 
97: sequence of vertex degrees of $G$ \cite{GM94}.
98: This question is currently being studied (see for example \cite{DR02}),
99: but has yet to be resolved.  In this paper, we extend the class of
100: graphs for which the conjecture is known to hold. 
101: We also show that if GM holds for graph Laplacians, it also holds
102: for more general ``Dirichlet Laplacians'' (cf. \cite{CL96})
103: as conjectured by Duval \cite{Duv03}.
104: 
105: \section{Background and definitions}
106: 
107: \subsection{Graphs}\label{ss:graphs}
108: Given a graph $G=(V,E)$ with $n=|V|$ vertices and $m=|E|$ edges,
109: there are several ways to represent $G$ as a matrix.  
110: There is the {\it edge--incidence
111: matrix}, a $n \times m$ matrix that records in each column the two 
112: vertices incident on a given edge.  
113: For directed graphs we can consider a signed edge--incidence matrix:
114: $$\partial(G)_{ve} = \begin{cases}
115:     1 &\text{if \ } v \text{ is the head of edge } e \\
116:     -1 &\text{if \ } v \text{ is the tail of edge } e \\
117:     0 &\text{otherwise}
118: \end{cases}
119: $$
120: 
121: There is also a $n \times n$ matrix $A(G)$ called the
122: {\it adjacency matrix} which is defined by:
123: $$A(G)_{ij} = \begin{cases} 
124:     1 &\text{if \ } (i,j) \in E \\
125:     0 &\text{otherwise}
126: \end{cases}
127: $$
128: The diagonal of $A(G)$ is zero.  
129: 
130: We can encode the (vertex) degree sequence of $G$ in non-increasing
131: order as a vector $d(G)$ of length $n$, and in an $n \times n$ matrix 
132: $D(G)$ whose diagonal is $d(G)$ and whose off-diagonal elements are 0.
133: Then the Combinatorial Laplacian $L(G)$ that we study in this
134: paper is simply $D(G)-A(G)$.  
135: It is easy to check that if we (arbitrarily) orient $G$ and 
136: consider the matrix $\partial(G)$ above, we also have 
137: $L(G)=\partial(G) \partial(G)^t$.
138: 
139: When the graph in question is clear from context, we may
140: abbreviate the above terms: $L, A, d, D$.
141: 
142: \begin{remark}\label{re:normalized} 
143: The Laplacian is sometimes defined with entries 
144: {\it normalized} by dividing by the square roots of the degrees.
145: However, we do not do that here.
146: \end{remark}
147: 
148: \subsection{Graph spectra}
149: The field of spectral graph theory is the study of the
150: structure of graphs through the spectra (eigenvalues) of
151: matrices encoding $G$.  Several surveys are available,
152: including \cite{Chu97} and \cite{CDD80}. 
153: Besides theoretical aspects of spectral graph theory,
154: these books describe a wide range of applications of the subject
155: to chemistry and physics as well as to problems in other 
156: branches of mathematics such as random walks and isoperimetric problems.
157: 
158: In the case of $L(G)$, there has been considerable effort to study 
159: the eigenvalue $\lambda_{n-1}$, which is
160: known as the {\it algebraic connectivity} of $G$.  It can be shown
161: that $\lambda_{n-1}(L(G)) = 0$ if and only if $G$ is disconnected.
162: Bounds on $\lambda_{n-1}(L(G))$ then give information on how
163: well connected a graph is, and are useful, for example, in showing
164: the existence of expander graphs.  This and other applications
165: are discussed in \cite{Chu97}.
166: 
167: Currently, little is known about the middle terms of the spectrum.
168: This is partly because it varies widely depending on the graph.
169: However, Grone and Merris \cite{GM94} conjecture that the
170: conjugate partition of the degree sequence majorizes the
171: spectrum, and showed that the majorization inequalities 
172: are tight on the class of {\it threshold} graphs.
173: This conjecture has been extended to simplicial complexes in
174: recent work by Duval and Reiner \cite{DR02}.
175: 
176: \subsection{Majorization}
177: We recall that a {\it partition} $p=p(i)$ is a non-increasing 
178: sequence of natural numbers, and its {\it conjugate} is
179: the sequence $p^T(j) := |\{i: p(i) \le j\}|$.  
180: Then $p^T$ has exactly $p(1)$ non-zero elements.
181: When convenient, we can add or drop trailing zeros in a partition.
182: For non-increasing real sequences $s$ and $t$ of length $n$, we say
183: that $s$ is {\it majorized} by $t$ (denoted $s \unlhd t$)
184: if for all $k \le n$:
185: 
186: \begin{equation}\label{eq:kst}
187: \sum_{i=1}^{k} s_i \le \sum_{i=1}^{k} t_i
188: \end{equation}
189: and 
190: \begin{equation}\label{eq:nst}
191: \sum_{i=1}^{n} s_i = \sum_{i=1}^{n} t_i
192: \end{equation}
193: 
194: The concept of majorization extends to vectors by comparing
195: the non-increasing vectors produced by sorting the
196: elements of the vector into non-increasing order.
197: Given a vector $v$, call the sorted vector $v'$ which contains 
198: the elements of $v$ sorted in non-increasing order (with multiplicity)
199: $\sort(v)$. 
200: 
201: In the context of majorization of unsorted vectors, we will
202: often want to refer to the {\it concatenation} of two vectors
203: $x$ and $y$ (ie.~the vector which contains the elements of $x$
204: followed the elements of $y$).  This is denoted $x,y$  as
205: for example in Lemma~\ref{l:xyz} below.
206: 
207: There is a rich theory of majorization inequalities which 
208: occur throughout mathematics, see for example \cite{MO81}.  
209: Matrices are an important source of majorization inequalities.
210: Notably, the relationship between the diagonal and
211: spectrum of a Hermitian matrix is characterized by
212: majorization (see for example \cite{HJ90}).  
213: 
214: We will use the following lemmas about majorization
215: which can be found in \cite{MO81}:
216: 
217: \begin{lemma}\label{l:xyP}
218: If $x$ and $y$ are vectors and $P$ is a doubly-stochastic matrix and $x=Py$, 
219: then $x \unlhd y$.
220: \end{lemma}
221: 
222: This yields two simple corollaries:
223: \begin{lemma}\label{l:xyz}
224: For any vectors $x \unlhd y$ and any vector $z$ we have: $x,z \unlhd y,z$.
225: \end{lemma}
226: 
227: \begin{lemma}\label{l:ij}
228: If $x$ and $y$ non-increasing sequences, and $x=y$ 
229: except that at indices $i<j$ we have $x_i = y_i-a$
230: and $x_j=y_j+a$ where $a \ge 0$ then $x \unlhd y$.
231: \end{lemma}
232: Lemma~\ref{l:ij} says that for
233: non-increasing sequences transferring units from lower
234: to higher indices reduces the vector in the majorization 
235: partial order.  
236: In particular,
237: if $x,x',y,y'$ are all {\it non-increasing} sequences, 
238: $x' \unlhd x$ and $y' \unlhd y$, then
239: \begin{equation}\label{eq:xyxy}
240: x' + y'  \unlhd x' + y \unlhd x+y
241: \end{equation}
242: 
243: Let $A$ and $B$ be positive semidefinite (more generally, Hermitian) 
244: matrices. Then:
245: \begin{equation}\label{eq:m1}
246: \lambda(A),\lambda(B) \unlhd \lambda(A+B)
247: \end{equation}
248: 
249: A theorem of Fan (1949) says that for positive semidefinite
250: (more generally, Hermitian) matrices $A$ and $B$:
251: \begin{equation}\label{eq:m2}
252: \lambda(A+B) \unlhd \lambda(A)+\lambda(B)
253: \end{equation}
254: 
255: Let $A$ be an $m \times n$ 0-1 (or incidence) matrix, with row sums
256: $r_1, \ldots, r_m$ and columns sums $c_1, \ldots, c_n$ both indexed
257: in non-increasing order.  Let $r^T$ be the conjugate of the 
258: partition $(r_1, \ldots, r_m)$, and $c$ be the partition
259: $(c_1, \ldots, c_n)$.  Then the Gale-Ryser theorem asserts that
260: \begin{equation}\label{eq:gr}
261: c \unlhd r^T
262: \end{equation}
263: 
264: \subsection{The Grone-Merris Conjecture}
265: The Grone-Merris conjecture (GM) is that the spectrum of
266: the combinatorial Laplacian of a graph is majorized by its
267: conjugate degree sequence, that is
268: \begin{equation}\label{eq:gm}
269: \lambda(G) \unlhd d^T(G)
270: \end{equation}
271: Note that 
272: $$ \sum_{i=1}^{n} d_i^T = \sum_{i=1}^{n} d_i = \trace(L(G))
273:    = \sum_{i=1}^{n} \lambda_i $$
274: 
275: If we ignore isolated vertices (which contribute only zero
276: entries to $\lambda$ and $d$) we will have $d^T_1=n$.
277: Using this fact, it is possible to show that
278: \begin{equation}\label{eq:1ld}
279: \lambda_1 \le d^T_1
280: \end{equation}
281: Three short proofs of this are given in \cite{DR02}.
282: The authors then continue to prove the second majorization inequality
283: \begin{equation}\label{eq:2ld}
284: \lambda_1 + \lambda_2 \le d^T_1 + d^T_2
285: \end{equation}
286: However, their proof would be difficult to extend.
287: 
288: There are several other facts which fit well with the GM
289: conjecture.  One is that if the GM conjecture holds, then 
290: the instances where (\ref{eq:gm}) holds with equality are well-understood,
291: these would be the threshold graphs of Section~\ref{ss:th}.
292: Also, since $d$ and $\lambda$ are respectively the diagonal and spectrum
293: of $L(G)$ we have $d \unlhd \lambda$.  Combining this with GM 
294: gives $d \unlhd d^T$, a fact that has been proved combinatorially.
295: We refer to \cite{DR02} for further discussion.
296: 
297: \begin{remark}[Complements]\label{re:complements}
298: Given a graph $G$, we can study its {\it complement} $\overline{G}$,
299: the graph whose edges are exactly those not included in $G$.
300: For a graph $G$ with $n$ vertices the $i$th largest vertex of $G$
301: is the $(n-i)$th largest vertex of $\overline{G}$, and we have
302: $d_i(G)=n-1-d_{n-i}(\overline{G})$.  Translating this to 
303: the conjugate partition $d^T$ yields:
304: $d_i^T(G)=n-d_{n-1-i}^T(\overline{G})$ with
305: $d_n^T(G)=d_n^T(\overline{G})=0$.
306: 
307: The relationship between $\lambda(G)$ and $\lambda(\overline{G})$
308: is the same as between $d_n^T(G)$ and $d_n^T(\overline{G})$.
309: This follows from the fact that
310: $L(G)+L(\overline{G})=n I_n-J_n$ where $J_n$ is the $n \times n$
311: matrix of ones.  We observe that the matrix $n I_n-J_n$ sends the special
312: eigenvector $e_n$ ($n$ ones) to 0, and acts as the scalar $n$
313: on $e_n^\perp$.  
314: Both $L(G)$ and $L(\overline{G})$ also send $e_n$ to 0, giving us
315: $\lambda_n(G)=\lambda_n(\overline{G})=0$.
316: Since $L(G)$ and $L(\overline{G})$ sum to $n I_n$ on $e_n^\perp$
317: they have the same set of eigenvectors on $e_n^\perp$, and 
318: and for each eigenvector the corresponding eigenvalues for 
319: $L(G)$ and $L(\overline{G})$ sum to $n$.
320: Thus 
321: $\lambda_i(G)=n-\lambda_{n-1-i}(\overline{G})$.
322: As a consequence, GM holds for $G$ if and only if GM holds for $\overline{G}$.
323: \end{remark}
324: 
325: \section{Grone-Merris on classes of graphs}
326: In this section we
327: give further evidence for the Grone-Merris conjecture by remarking that
328: it holds for several classes of graphs including threshold graphs,
329: regular graphs and trees.
330: 
331: \subsection{Threshold graphs}\label{ss:th}
332: The GM conjecture was originally formulated in the context
333: of {\it threshold} graphs, which are a class of graphs with several
334: extremal properties.  An introduction to threshold graphs is \cite{MP95}.  
335: Threshold graphs are the graphs
336: that can be constructed recursively by adding isolated vertices
337: and taking graph complements.  It turns out that they are also
338: characterized by degree sequences: the convex hull of possible 
339: (unordered) degree sequences of an $n$ vertex graph defines
340: a polytope.  
341: The extreme points of this polytope are the degree sequences that have
342: a unique labelled realization, and these are exactly the threshold graphs.
343: 
344: Threshold graphs are interesting from the point of view of
345: spectra.  Both Kelmans and Hammer \cite{HK96} and Grone
346: and Merris \cite{GM94} investigated the question of
347: which graphs have integer spectra.  They found that threshold
348: graphs are one class of graphs that have integer spectra and
349: showed for these graphs that $\lambda(G)=d^T(G)$.  
350: 
351: In the process of showing this equality for threshold graphs,
352: Grone and Merris observed that for non-threshold graphs, the 
353: majorization inequality $\lambda(G) \unlhd d^T(G)$
354: appears to hold, and made their conjecture.  
355: We could describe the conjecture as saying that threshold 
356: graphs are extreme in
357: terms of spectra, and that the these extreme spectra can be
358: interpreted as conjugate degree sequences.
359: 
360: \subsection{Regular and nearly regular graphs}
361: For some small classes of graphs, it can be easily shown
362: that the GM conjecture holds.
363: Consider a $k$-regular graph $G$ on $n$ vertices
364: (in a {\it $k$-regular} graph, all vertices have degree $k$).
365: Then the degree sequence $d(G)$ is $k$ repeated $n$ times, and its
366: conjugate $d^T(G)$ is $n$ repeated $k$ times followed by $n-k$ zeros.
367: Thus $d^T$ majorizes every non-negative sequence of sum $kn$ whose largest
368: terms is at most $n$, and in particular $\lambda \unlhd d^T$.
369: Indeed, this proof shows that GM holds for what we might call 
370: {\it nearly regular} graphs, that is graphs whose vertices have
371: degree either $k$ or $(k-1)$.
372: 
373: \subsection{Graphs with low maximum degree}
374: Using facts about the initial GM inequalities
375: we can prove that GM must hold for graphs with
376: low maximal degree.  For example, if a graph has maximum vertex
377: degree 2, then $d^T_3 = d^T_4 = \ldots = d^T_n = 0$, so for
378: $k=2,3,\ldots,n$:
379: $$\sum_{i=1}^{k} \lambda_i \le 
380:   \sum_{i=1}^{n} \lambda_i = \sum_{i=1}^{n} d^T_i = \sum_{i=1}^{k} d^T_i$$
381: More generally, the GM inequalities for $k \ge \max_degree(G)$ hold
382: trivially.  Thus GM holds for graphs
383: of maximum degree 2 by (\ref{eq:1ld}).
384: Using Duval and Reiner's result (\ref{eq:2ld}), we get that
385: GM holds for graphs of maximum degree 3.
386: 
387: \subsection{Trees and more}
388: It is tempting to try to prove GM inductively by breaking 
389: graphs into simpler components on which GM clearly holds.
390: In this section, we show that if $G$ is ``almost'' the 
391: union of two smaller graphs on which GM holds then GM holds
392: for $G$ as well.  We apply this construction to show that
393: GM holds for trees.
394: 
395: Take two graphs $A=(V_A,E_A)$ and $B=(V_B,E_B)$ 
396: on disjoint vertex sets $V_A$ and $V_B$.
397: Define their {\it disjoint sum} to be 
398: $A+B=(V_A \cup V_B, E_A \cup E_B)$.  Assuming $V_A$ and $V_B$
399: are not empty this is a disconnected graph.
400: Now take two graphs $G=(V,E_G)$ and $H=(V,E_H)$
401: on the same vertex set $V$.  
402: Define their {\it union} as $G \cup H = (V, E_G \cup E_H)$.
403: 
404: Given the spectra and conjugate degree sequences of $A$ and $B$,
405: the spectrum of $A+B$ is (up to ordering) 
406: $\lambda(A+B) = (\lambda(A), \lambda(B))$,
407: while the conjugate degree sequence of $A+B$ is $d^T(A+B) = d^T(A)+d^T(B)$
408: (taking each vector to have length $n$).  Thus by~\ref{eq:m1}
409: if $\lambda(A) \unlhd d^T(A)$ and $\lambda(B) \unlhd d^T(B)$
410: we will have $\lambda(A+B) \unlhd d^T(A+B)$.
411: 
412: In a typical situation, where neither $A$ or $B$ is very
413: small, we would expect the above majorization inequality
414: to hold with considerable slack.  We can use this slack to
415: show that if we add a few more edges to $A+B$ the majorization
416: will still hold.  
417: 
418: \begin{theorem}\label{th:abc}
419: Take graphs $A$ or $B$ on disjoint vertex sets $V_A$ and $V_B$.
420: Let $G = A+B$ and on $V = V_A \cup V_B$ let 
421: $C$ be a graph of ``new edges'' between $V_A$ and $V_B$.
422: Assume that GM holds on $A$, $B$ and $C$, ie.~that
423: $\lambda(A) \unlhd d^T(A)$, $\lambda(B) \unlhd d^T(B)$
424: and $\lambda(C) \unlhd d^T(C)$.
425: Additionally, assume that $d^T_i(C) \le d^T_i(A), d^T_i(B)$ for all $i$,
426: and that $d^T_1(B) \le d^T_m(A)$ where $m$ is the largest non-zero
427: index of $d^T(C)$ (equivalently, $m$ is the maximum vertex degree in $C$).
428: Let $H=C \cup G$.  Then:
429: \begin{equation}\label{eq:mabc}
430: \lambda(H) \unlhd d^T(H)
431: \end{equation}
432: \end{theorem}
433: 
434: \begin{proof}
435: Let $k$ be the larger of $\max_degree(A)$ and $\max_degree(B)$.  Note that 
436: $$d^T(G) = d^T(A)+d^T(B) = 
437: (d^T_1(A)+d^T_1(B), d^T_2(A)+d^T_2(B), \ldots, d^T_k(A)+d^T_k(B), 0, \ldots, 0)
438: $$
439: \begin{claim}\label{c:gc}
440: $$ d^T(H) \unrhd (d^T_1(G), d^T_2(G), \ldots, 
441:           d^T_k(G), d^T_1(C), \ldots, d^T_m(C) )$$
442: \end{claim}
443: 
444: \noindent
445: {\it Proof of Claim.}
446: The term on the right is the concatenation of two partitions,
447: $d^T(G)$ and $d^T(C)$.  The columns of $d^T(G)$ index the vertices
448: of $G$ and the length of a column gives the degree of the corresponding
449: vertex.  Since this claim is purely about the combinatorics of
450: of degree sequences, we introduce a series of
451: intermediate ``partial graphs'' where edges are allowed to have only
452: one end.  Degree sequences and their conjugates are still well defined
453: for such objects.
454: 
455: Let $G_0=G$ and $C_0=C$.  Define $G_i$ by moving one end of an edge
456: from every non-isolated vertex of $C_{i-1}$ to $G_{i-1}$, and let 
457: $C_{i}$ contain whatever is left.  Iterating this, for
458: some $l \ge 0$ we will have $G_l=H$ and $C_l$ consisting entirely
459: of isolated vertices.  
460: Then the claim will follow if we can show that:
461: $$ d^T(G_0), d^T(C_0) \unlhd d^T(G_1), d^T(C_1) \unlhd \ldots
462:  \unlhd d^T(G_l), d^T(C_l) $$
463: Compare the partitions at the $(i-1)$st majorization: we 
464: remove the first row of $d^T(C_{i-1})$ and put each element from
465: that row into a separate column (representing a distinct vertex in
466: $G$) of $d^T(G_{i-1})$.  Where there are columns of equal length
467: in $d^T(G_{i-1})$ 
468: they should be ordered so that those acquiring new elements come
469: first.
470: To see that this operation increases the partition
471: in the majorization partial order, observe that
472: after ignoring the (unchanged) contents of $d^T(C_i)$ it is
473: equivalent to sorting the new row into the partition,
474: using Lemma~\ref{l:ij} to move its final (rightmost) element 
475: to the proper column and repeating as necessary.
476: \vspace{3mm}
477: 
478: This completes the proof of the Claim~\ref{c:gc}, and gives us:
479: $$ d^T(H) \unrhd (d^T_1(A)+d^T_1(B), d^T_2(A)+d^T_2(B), \ldots, 
480:           d^T_k(A)+d^T_k(B), d^T_1(C), \ldots, d^T_m(C) )$$
481: If we sort the vector on the right into non-increasing
482: order, the first $m$ terms will remain fixed by the assumptions
483: that $d^T_m(A) \ge d^T_1(B) \ge d^T_1(C)$.  
484: Since we have assumed that $d^T_i(C) \le d^T_i(B)$ for all $i$, 
485: we can apply Lemma~\ref{l:ij} to the reordered sequence to get:
486: \begin{align*}
487: d^T(H)  \unrhd 
488:  (d^T_1(A)+d^T_1(C),& d^T_2(A)+d^T_2(C), \ldots, 
489:           d^T_m(A)+d^T_m(C), \\ & d^T_{m+1}(A), \ldots, d^T_{k}(A),
490: 	d^T_1(B), \ldots, d^T_k(B) )
491: \end{align*}
492: The right hand term decomposes as:
493: $$(d^T_1(A), \ldots, d^T_k(A), d^T_1(B), \ldots, d^T_k(B))
494:  + (d^T_1(C), \ldots, d^T_m(C),0,\ldots,0)$$
495: Since we assume $d^T_m(A) \ge d^T_1(B)$, the first $m$ entries
496: of $(d^T(A), d^T(B))$ will remain unchanged if the vector is sorted.
497: Thus:
498: \begin{equation}\label{eq:dt}
499: d^T(H) \unrhd \sort(d^T(A),d^T(B)) + d^T(C)
500: \end{equation}
501: By (\ref{eq:xyxy}) we can apply the majorizations of $\lambda$
502: by $d^T$ for $A, B, C$ to the above terms to get:
503: $$\begin{array}{cl}
504: d^T(H) & \unrhd ~ \sort(d^T(A),d^T(B)) + d^T(C) 
505:            ~ \unrhd ~ \sort(d^T(A),d^T(B)) + \lambda(C)\\
506:        & \unrhd ~ \sort(d^T(A),\lambda(B)) + \lambda(C)
507:           ~  \unrhd ~  \sort(\lambda(A),\lambda(B)) + \lambda(C)
508: \end{array}$$
509: The two terms on the right side of this equation are spectra
510: of $L(G)$ and $L(C)$ respectively.  Hence by Fan's theorem (\ref{eq:m2})
511: their sum majorizes the spectrum of L(G)+L(C)=L(H):
512: $$d^T(H) \unrhd \lambda(G)+ \lambda(C) \unrhd \lambda(H)$$
513: \end{proof}
514: 
515: More generally, we could replace the conditions in the statement 
516: of Theorem~\ref{th:abc} with the condition (\ref{eq:dt}), 
517: which can be checked combinatorially.
518: The conditions in the theorem statement and equation (\ref{eq:dt})
519: are most likely to be satisfied if $C$ is small relative to $A$ and $B$.
520: 
521: A useful case is when $C$ consists of $k$ disjoint edges. 
522: Then $m=1$ and $d^T_1(C)=2k$.  Without loss of generality
523: we can take $d_1(A) \ge d_1(B)$ and the only condition that we will
524: need to check is that $d_1(A), d_1(B) \ge d_1(C)$, ie.~both $A$ and
525: $B$ must have at least $2k$ non-isolated vertices.
526: 
527: The strategy for applying Theorem~\ref{th:abc} to show that a given 
528: graph $H$ satisfies GM is to find a ``cut'' $C$ for it that
529: contains few edges and divides $H$ into relatively large components.
530: For example we have the following result:
531: 
532: \begin{corollary}
533: The GM conjecture holds for trees.
534: \end{corollary}
535: 
536: \begin{proof}
537: Proceed by induction on the diameter of the graph.
538: If $T$ has diameter 1 or 2, then there is a vertex $v$
539: which is the neighbour of all the remaining vertices
540: and $T$ is a threshold graph.  
541: So GM holds with equality for $T$.
542: 
543: Otherwise, we can find some edge $e$ that does not have a leaf vertex.
544: Since $T$ is a tree, $e$ is a cut edge and divides $T$ into two
545: non-trivial connected components, $A$ and $B$.  We apply induction to
546: $A$ and $B$ and apply Theorem~\ref{th:abc} to $H=(A+B) \cup C$
547: where $C$ is the graph on the vertex set of $T$ containing the
548: single edge $e$.
549: \end{proof}
550: 
551: \begin{remark}[Small Graphs]  
552: The facts in this section allow us to check that GM holds for
553: some small graphs without directly computing eigenvalues.
554: For example, since the GM condition is closed under complement 
555: (see~\ref{re:complements}) for graphs on up
556: to 5 vertices it is enough to observe that either $G$ or $\overline{G}$
557: has maximum degree $\le 3$.  Out of 156 graphs on 6 vertices, 146
558: can be decomposed into smaller graphs $(A+B) \cup C$ using
559: Theorem~\ref{th:abc}.  Calculating the eigenvalues of the remaining
560: 10 does not yield a counterexample.
561: \end{remark}
562: 
563: \section{Simplices and pairs}
564: The most recent work relating to the GM conjecture has been to study the
565: spectra of more general structures than graphs, such as simplicial
566: complexes and simplicial family pairs.
567: In this section we show that the generalization of GM to graphs with
568: Dirichlet boundary conditions is equivalent to the original conjecture 
569: and may be useful in approaching GM.
570: 
571: \subsection{Simplicial complexes}
572: In \cite{DR02}, the authors look at {\it simplicial complexes},
573: which are higher dimensional analogues of simple graphs (see for example
574: \cite{Mun84}).
575: A set of faces of a given dimension $i$
576: is called an {\it i-family}.  Given a simplicial complex $\Delta$
577: we can denote the $i$-family of all faces in $\Delta$ of dimension $i$
578: as $\Delta^{(i)}$.
579: For example, a graph is a 1-dimensional complex, and its edge set is
580: the 1-family $\Delta^{(1)}$.
581: Define the degree sequence $d$ of an $i$-family to be
582: the list of the numbers of $i$-faces from the family
583: incident on each vertex, and sorted into non-increasing order.  
584: We can then define $d(\Delta, i)$ as the degree sequence of
585: $\Delta^{(i)}$, which we can abbreviate to $d(\Delta)$ or $d$
586: when the context is clear.
587: 
588: We define the {\it chain group} $C_i(\Delta)$ of formal linear
589: combinations of elements of $\Delta^{(i)}$, and generalize 
590: the signed incidence matrix $\partial$ of Section~\ref{ss:graphs}
591: to a signed boundary map 
592: $\partial_i:C_i(\Delta) \rightarrow C_{i-1}(\Delta)$.
593: This allows us to define a {\it Laplacian} on $C_i(\Delta)$, 
594: namely $L_i(\Delta) = \partial_i \partial_i^T$,
595: and study its corresponding spectrum $s(\Delta,i)$ sometimes
596: abbreviated $s(\Delta)$ or $s$.
597: 
598: Duval and Reiner \cite{DR02} looked at 
599: {\it shifted} simplicial complexes, which are a
600: generalization of threshold graphs to complexes.  
601: They showed that for a shifted complex $\Delta$ and any $i$,
602: we have $s(\Delta,i)=d^T(\Delta,i)$.  They then conjectured
603: that GM also holds for complexes, ie.~that for any complex
604: and any $i$ we have:
605: \begin{equation}\label{eq:gmc}
606: s(\Delta,i) \unlhd d^T(\Delta,i)
607: \end{equation}
608: They also show that some related facts, such as equation
609: (\ref{eq:1ld}) generalize to complexes.
610: 
611: \subsection{Simplicial pairs}
612: In \cite{Duv03}, Duval continues by studying {\it relative (family) pairs} 
613: $(K,K')$ where the set $K=\Delta^{(i)}$ for some $i$ 
614: is taken modulo a family of $(i-1)$-faces $K' \subseteq \Delta^{(i-1)}$.
615: When $K'= \emptyset$, this reduces to the situation of the previous
616: section.
617: 
618: \begin{remark}\label{re:dirchlet}
619: In the case $i=1$ this is the edge set of a graph ($K$) with a set of 
620: {\it deleted} boundary vertices $K'$.  An edge attached to a deleted
621: vertex will not be removed -- it remains as part of the pair,
622: but we now think of the edge as having a hole on one (or both) ends.
623: 
624: This type of graph with a boundary appears in conformal invariant
625: theory.  In this language, the relative Laplacian of an
626: (edge, vertex) pair is sometimes referred to as a 
627: {\it Dirichlet Laplacian} and its eigenvalues as 
628: {\it Dirichlet eigenvalues}, see for example \cite{CL96}.
629: Recently \cite{CE02} used the spectrum of the Dirichlet Laplacian in 
630: the analysis of ``chip-firing games'', which are processes on graphs
631: that have an absorbing (Dirichlet) boundary at some vertices.
632: \end{remark}
633: 
634: We can form chain groups $C_i(K)$ and $C_{i-1}(K,K')$
635: and use these to define a (signed) boundary operator on
636: the pair $\partial(K,K'): C_i(K) \rightarrow C_{i-1}(K,K')$.
637: Hence we get a Laplacian for family pairs 
638: $L(K,K')= \partial(K,K') \partial(K,K')^T$.
639: Considered as a matrix, $L(K,K')$ will be the principal submatrix
640: of $L(K)$ whose rows are indexed by the $i$-faces in $\Delta^{(i-1)}-K'$.
641: Finally, we get a spectrum $s(K,K')$ for family pairs from the 
642: eigenvalues of $L(K,K')$.
643: 
644: Duval defines the degree $d_v(K,K')$ of vertex $v$ 
645: (in the case of a graph, $v$ is allowed to be in $K'$)
646: relative to the pair $(K,K')$ as the number of faces in $K$
647: that contain $v$ such that $K-\{v\}$ is in $\Delta^{(i-1)}-K'$.
648: This allows him to define the degree sequence $d(K,K')$ for
649: pairs, and to conjecture that GM holds for relative pairs:
650: \begin{equation}\label{eq:gmr}
651: s(K,K') \unlhd d^T(K,K')
652: \end{equation}
653: 
654: 
655: \subsection{The Grone-Merris conjecture for relative pairs}
656: It turns out that at least in the case of (edge, vertex) pairs
657: that (\ref{eq:gmr}) follows from the original GM conjecture for
658: graphs.
659: 
660: \begin{theorem}\label{th:gmgmp}
661: GM for graphs $\Rightarrow$ GM for (edge, vertex) pairs.
662: \end{theorem}
663: 
664: \begin{proof}
665: Let $G=(V,E)$ be a graph with $D \subseteq V$ a set of ``deleted''
666: vertices.  Let $U=V - D$ be the remaining ``undeleted'' vertices.
667: We will assume that GM holds only on the undeleted part of the
668: graph, ie.~$G|_U$.  So we have $s(G|_U) \unlhd d^T(G|_U)$.
669: We can ignore the edges in $G|_D$ completely, since they have
670: no effect on either $s(G)$ or $d(G)$.
671: The remaining edges connect vertices in $D$ to vertices in $U$.
672: Define $G'$ to be the graph on $V$ whose edge are exactly the
673: edges of $G$ between $D$ and $U$.  Let $a$ be the degree sequence
674: of the deleted vertices in $G'$ and $b$ be the degree sequence
675: of the undeleted vertices in $G'$.
676: 
677: We can compute $d^T(E,D)$ in terms of
678: the degree sequences and spectra of $G|_U$, $G'$ and $G|_D$
679: since $d^T_i(E,D)$ is the number of vertices (deleted or not)
680: attached to at least $i$ non-deleted vertices.
681: The number of such vertices in $U$ will be $d^T_i(G|_U)$,
682: and the number in $D$ will be $d^T_i(G')=a^T$.
683: Hence $d^T(E,D) = d^T_i(G|_U) + a^T$.
684: 
685: Now consider the Laplacian $L(E,D)$.
686: This is the submatrix of $L(G)$ indexed by $U$.
687: An edge $(i,j)$ in $G|_U$ contributes to entries
688: $ii, ij, ji, jj$ in both $L(E,D)$ and $L(G)$.
689: An edge in $G'$, say from $i \in U$ to $j \in D$ contributes
690: only to entry $ii$, and an edge in $G|_D$ does not affect
691: $L(E,D)$.  So, we have $L(E,D) = L(G|_U) + \Diag(b)$, and
692: by (\ref{eq:m2}) we have:
693: \begin{equation}\label{eq:speced}
694: s(E,D) \unlhd s(G|_U) + b
695: \end{equation}
696: 
697: We complete our equivalence by appealing to the Gale-Ryser 
698: theorem (\ref{eq:gr}) to claim that $b \unlhd a^T$.
699: This follows from the fact that $a$ and $b$ are row and
700: column sums (in non-increasing order) of the $|D| \times |U|$
701: bipartite incidence matrix for $G'$.  Combining with the
702: assumption that $s(G|_U) \unlhd d^T(G|_U)$ and 
703: (\ref{eq:speced}) we get:
704: $$s(E,D) \unlhd s(G|_U) + b \unlhd d^T(G|_U) + a^T = d^T(E,D)$$
705: \end{proof}
706: 
707: This proof relies on the bipartite structure of $G'$, so
708: it is not immediately obvious how to extend it 
709: to higher dimensional complexes.  It would be interesting to
710: do this.
711: 
712: \begin{remark}
713: Because the induction used to prove Theorem~\ref{th:gmgmp}
714: requires only that the ``undeleted'' part of the graph satisfy
715: GM, it is tempting to attack the original GM conjecture by
716: showing if GM holds for a pair $(G,\{v\})$ then GM holds for $G$.
717: \end{remark}
718: 
719: \section{Acknowledgements}
720: Thanks to Vic Reiner for introducing me to this question
721: and for comments throughout.  
722: 
723: \bibliography{gm}
724: 
725: \end{document}
726: 
727: