% This file is part of the Stanford GraphBase (c) Stanford University 1992
\def\title{GIRTH}
@i boilerplate.w %<< legal stuff: PLEASE READ IT BEFORE MAKING ANY CHANGES!
\let\==\equiv % congruence sign
\prerequisite{GB\_\thinspace RAMAN}
@* Introduction. This demonstration program uses graphs
constructed by the |raman| procedure in the |gb_raman| module to produce
an interactive program called \.{girth}, which computes the girth and
diameter of a class of Ramanujan graphs.
The girth of a graph is the length of its shortest cycle; the diameter
is the maximum length of a shortest path between two vertices.
A Ramanujan graph is a connected, undirected graph in which every vertex
has degree~|p+1|, with the property that every eigenvalue of its adjacency
matrix is either $\pm(p+1)$ or has absolute value $\le2\sqrt{\mathstrut p}$.
Exact values for the girth are of interest because the bipartite graphs
produced by |raman| apparently have larger girth than any other known
family of regular graphs, even if we consider graphs whose existence
is known only by non-constructive methods, except for the cubic ``sextet''
graphs of Biggs, Hoare, and Weiss [{\sl Combinatorica\/ \bf3} (1983),
153--165; {\bf4} (1984), 241--245].
Exact values for the diameter are of interest because the diameter of
any Ramanujan graph is at most twice the minimum possible diameter
of any regular graph.
The program will prompt you for two numbers, |p| and |q|. These should
be distinct prime numbers, not too large, with |q>2|. A graph is
constructed in which each vertex has degree~|p+1|. The number of
vertices is $(q^3-q)/2$, if |p| is a quadratic residue modulo~|q|, or
$q^3-q$ if |p| is not a quadratic residue. In the latter case the
graph is bipartite and it is known to have rather large girth.
If |p=2|, the value of |q| is further restricted to be of the form
$104k+(1,3,9,17,25,27,35,43,49,51,75,81)$. This means that the only
feasible values of |q| to go with |p=2| are probably 3, 17, and 43;
the next case, |q=107|, would generate a bipartite graph with
1,224,936 vertices and 3,675,808 arcs, thus requiring approximately
113 megabytes of memory (not to mention a nontrivial amount of
computer time). If you want to compute the girth and diameter
of Ramanujan graphs for large |p| and/or~|q|, much better methods are
available based on number theory; the present program is merely a
demonstration of how to interface with the output of |raman|.
Incidentally, the graph for |p=2| and |q=43| turns
out to have 79464 vertices, girth 20, and diameter~22.
The program will examine the graph, compute its girth and its diameter,
then it will prompt you for another choice of |p| and |q|.
The graphs we work with have the data type \&{Graph}, defined in |gb_graph|.
@f Graph int
@f Arc int
@f Vertex int
@ Here is the general layout of this program, as seen by the \Cee\ compiler:
@p
#include /* the system |sqrt| routine is needed */
#include "gb_graph.h" /* the standard GraphBase data structures */
#include "gb_raman.h" /* Ramanujan graph generator */
@#
@@;
main()
{
printf("This program explores the girth and diameter of Ramanujan graphs.\n");
printf("The bipartite graphs have q^3-q vertices, and the non-bipartite\n");
printf("graphs have half that number. Each vertex has degree p+1.\n");
printf("Both p and q should be odd prime numbers;\n");
printf(" or you can try p = 2 with q = 17 or 43.\n");
while (1) {
@;
g=raman(p,q,0,0);
if (g==NULL) @@;
else {
@;
@;
gb_recycle(g);
}
}
}
@ @=
Graph *g; /* the current Ramanujan graph */
int p; /* the branching factor (degree minus one) */
int q; /* cube root of the graph size */
char buffer[16]; /* place to collect what the user types */
@ @d prompt(s)
{@+printf(s);@+fflush(stdout); /* make sure the user sees the prompt */
if (fgets(buffer,15,stdin)==NULL) break;@+}
@=
prompt("\nChoose a branching factor, p: ");
if (sscanf(buffer,"%d",&p)!=1) break;
prompt("OK, now choose the cube root of graph size, q: ");
if (sscanf(buffer,"%d",&q)!=1) break;
@ @=
printf(" Sorry, I couldn't make that graph (%s).\n",
panic_code==very_bad_specs? "q is out of range":
panic_code==very_bad_specs+1? "p is out of range":
panic_code==bad_specs+5? "q is too big":
panic_code==bad_specs+6? "p is too big":
panic_code==bad_specs+1? "q isn't prime":
panic_code==bad_specs+7? "p isn't prime":
panic_code==bad_specs+3? "p is a multiple of q":
panic_code==bad_specs+2? "q isn't compatible with p=2":
"not enough memory");
@* Bounds. The theory of Ramanujan graphs allows us to predict the
girth and diameter to within a factor of 2~or~so.
In the first place, we can easily derive an upper bound on the girth
and a lower bound on the diameter, valid for any regular graph of
degree~|p+1|. Such a graph has at most $(p+1)p^{k-1}$ points at
distance~$k$ from any given vertex; this implies a lower bound
on the diameter~$d$:
$$1+(p+1)+(p+1)p+(p+1)p^2+\cdots+(p+1)p^{d-1}\;\ge\;n.$$
Similarly, if the girth $g$ is odd, say $g=2k+1$, the points at
distance~$\le k$ from any vertex must be distinct, so we have
$$1+(p+1)+(p+1)p+(p+1)p^2+\cdots+(p+1)p^{k-1}\;\le\;n;$$
and if $g=2k+2$, at least $p^k$ further points must exist at distance
$k+1$, because the $(p+1)p^k$ paths of length $k+1$ can end at
a particular vertex at most $p+1$ times. Thus
$$1+(p+1)+(p+1)p+(p+1)p^2+\cdots+(p+1)p^{k-1}+p^k\;\le\;n$$
when the girth is even.
In the following code we let $|pp|=p^{dl}$ and
$s=1+(p+1)+\cdots+(p+1)p^{dl}$.
@=
s=p+2;@+dl=1;@+pp=p;@+gu=3;
while (s2|, we can use the theory of integral quaternions to derive a lower
bound on the girth of the graphs produced by |raman|. A path of length~$g$
from a vertex to itself exists if and only if there is an integral
quaternion $\alpha=a_0+a_1i+a_2j+a_3k$ of norm $p_g$ such that
the $a$'s are not all multiples of~$p$, while
$a_1$, $a_2$, and $a_3$ are multiples of~$q$ and $a_0\not\=a_1\=a_2\=a_3$
(mod~2). This means we have integers $(a_0,a_1,a_2,a_3)$ with
$$a_0^2+a_1^2+a_2^2+a_3^2=p^g,$$ satisfying the stated properties
mod~$q$ and mod~2.
If $a_1$, $a_2$, and $a_3$ are even, they cannot all be zero so
we must have $p^g\ge1+4q^2$; if they are odd, we must have
$p^g\ge4+3q^2$. (The latter is possible only when $g$ is odd and
$p\bmod4=3$.) Since $n$ is roughly proportional to~$q^3$, this means
$g$ must be at least about ${2\over3}\log_p n$. Thus, $g$
isn't too much less than the maximum girth possible in any regular graph,
which we have shown is at most about $2\log_p n$.
When the graph is bipartite we can, in fact, prove that $g$ is
approximately ${4\over3}\log_p n$. The bipartite case occurs if and
only if $p$ is not a quadratic residue modulo~|q|; hence the
number~$g$ in the previous paragraph must be even, say $g=2r$. Then
$p^g\bmod4=1$, and $a_0$ must be odd. The congruence $a_0^2\=p^{2r}$
(mod~$q^2$) implies that $a_0\=\pm p^r$, because all numbers
relatively prime to $q^2$ are powers of a primitive root. We can
assume without loss of generality that $a_0=p^r-2mq^2$, where
$0q^2$. Conversely,
if $p^r-q^2$ can be written as a sum of three squares
$b_1^2+b_2^2+b_3^2$, then
$p^{2r}=(p^r-2q^2)^2+(2b_1q)^2+(2b_2q)^2+(2b_3q)^2$ is a
representation of the required type. If $p^r-q^2$ is a positive
integer that cannot be represented as a sum of three squares, a
well-known theorem of Legendre tells us that $p^r-q^2=4^ts$, where
$s\=7$ (mod~8). Since $p$ and $q$ are odd, we have $t\ge1$; hence
$p^r-2q^2$ is odd. If $p^r-2q^2$ is a positive odd integer, Legendre's
theorem tells us that we can write $2p^r-4q^2=b_1^2+b_2^2+b_3^2$;
hence $p^{2r}=(p^r-4q^2)^2+ (2b_1q)^2+(2b_2q)^2+(2b_3q)^2$. We
conclude that the girth is either $2\lceil\log_pq^2\rceil$ or
$2\lceil\log_p2q^2\rceil$. (This explicit calculation, which makes our
program for calculating the girth unnecessary or at best redundant in
the bipartite case, is due to G. A. Margulis and, independently, to
Biggs and Boshier [{\sl Journal of Combinatorial Theory\/ \bf B49}
(1990), 190--194].)
A girth of 1 or 2 can occur, since these graphs might have self-loops
or multiple edges if |p| is sufficiently large.
@=
if (bipartite) {@+long b=q*q;
for (gl=1,pp=p;pp<=b;gl++,pp*=p) ; /* iterate until $p^g>q^2$ */
gl+=gl;
} else {@+long b1=1+4*q*2, b2=4+3*q*q; /* bounds on $p^g$ */
for (gl=1,pp=p;pp=b2 && (gl&1) && (p&2)) break;
}
}
@ Upper bounds on the diameter of any Ramanujan graph can be derived
as shown in the paper by Lubotzky, Phillips, and Sarnak in
{\sl Combinatorica \bf8} (1988), page~275. (However, a slight correction
to their proof is necessary---their parameter~$l$ should be~odd
when $x$ and~$y$ lie in different parts of a bipartite graph.)
Their argument demonstrates that $p^{(d-1)/2}<2n$ in the
nonbipartite case and $p^{(d-2)/2}=
{@+long nn=(bipartite? n: 2*n);
double nnp=((double)nn)/sqrt((double)p);
long nnn=(long)nnp;
if ((double)nnn>nnp) nnn--; /* truncate, don't round */
for (du=0,pp=1;pp<=nnn;du+=2,pp*=p) ;
if (pp=
n=g->n;
if (n==(q+1)*q*(q-1)) bipartite=1;
else bipartite=0;
printf("The graph has %d vertices, each of degree %d, and it is %sbipartite.\n",
n,p+1,bipartite? "": "not ");
@;
printf("Any such graph must have diameter >= %d and girth <= %d;\n",
dl,gu);
@;
printf("theoretical considerations tell us that this one's diameter is <= %d",
du);
if (p==2) printf(".\n");
else {
@;
printf(",\nand its girth is >= %d.\n",gl);
}
@ We had better declare all the variables we've been using so freely.
@=
int gl,gu,dl,du; /* theoretical bounds */
long pp; /* power of $p$ */
long s; /* accumulated sum */
long n; /* number of vertices */
char bipartite; /* is the graph bipartite? */
@*Breadth-first search. The graphs produced by |raman| are symmetrical, in
the sense that there is an automorphism taking any vertex into any
other; each vertex $V$ and each edge $P$ corresponds to a $2\times2$
matrix, and the path $P_1P_2\ldots P_k$ leading from vertex~$V$ to
vertex $VP_1P_2\ldots P_k$ has the same properties as the path leading
from vertex~$U$ to vertex $UP_1P_2\ldots P_k$. Therefore we can find
the girth and the diameter by starting at any vertex $v_0$.
We will compute the number of points at distance $k$ from $v_0$ for
all $k$, by explicitly forming a linked list of all such points.
Utility field |link| will be used for the links. The lists will
terminate with a non-null |sentinel| value, so that we can also
use the condition |link==NULL| to tell if a vertex has been
encountered before. Another utility field, |dist|, will contain the
distance from the starting point; and |back| will point to a
vertex one step closer.
@d link w.v /* the field where we store links, initially |NULL| */
@d dist v.i /* the field where we store distances, initially 0 */
@d back u.v /* the field where we store backpointers, initially |NULL| */
@=
printf("Starting at any given vertex, there are\n");
{@+int k; /* current distance being generated */
int c; /* how many we've seen so far at this distance */
register Vertex *v; /* current vertex in list at distance $k-1$ */
register Vertex *u; /* head of list for distance $k$ */
Vertex *sentinel=g->vertices+n; /* nonzero link at end of lists */
int girth=999; /* length of smallest cycle found, initially infinite */
k=0;
u=g->vertices;
u->link=sentinel;
c=1;
while (c) {
for (v=u,u=sentinel,c=0,k++;v!=sentinel;v=v->link)
@;
printf("%8d vertices at distance %d%s\n", c, k, c>0? ",": ".");
}
printf("So the diameter is %d, and the girth is %d.\n",k-1,girth);
}
@ @=
{@+register Arc *a;
for (a=v->arcs;a;a=a->next) {@+register Vertex *w;
/* vertex adjacent to |v| */
w=a->tip;
if (w->link==NULL) {
w->link=u;
w->dist=k;
w->back=v;
u=w;
c++;
} else if (w->dist+kback)
girth=w->dist+k;
}
}
@* Index. Finally, here's a list that shows where the identifiers of this
program are defined and used.