Catalan Numbers and Random Matrices

Lawrence Ip
[email protected]
Department of Electrical Engineering
and Computer Sciences
University of California, Berkeley
supervised by Alan Edelman
May 19 1999

1 Introduction
Catalan numbers, the number of ways of pairing n brackets, arise naturally
in some random matrix results in the computation of moments of eigenvalue
distributions.
In Section 2 we show that Catalan numbers can be represented as the
number of Dyck paths. In Section 3 we show that this representation can be
used to drastically simplify some of the calculations that Wigner used for his
derivation of the semi-circle law and in Section 4 we show that Dyck paths
can be used to interpret an expression combinatorially to prove its validity.
This last result is the main result of this project.

2 Catalan Numbers and Dyck Paths

The most common denition of the nth Catalan number is \the number of
ways of pairing n brackets". So we have C = 1, C = 2, C = 5, C = 14,
1 2 3 4
C = 42, : : : . The standard way to nd a general formula for Cn is to derive
5
a recurrence relation for it and then solve this using generating functions.

1
 However, this is not the only representation. In his new book, Stanley [2],
lists 66 combinatorial interpretations of Catalan numbers. We show that Cn
is also the number of Dyck paths of length 2n.
A Dyck path is a path on the square lattice with steps (1; 1) and (1; 1)
from (0; 0) to (2n; 0) that never falls below the x-axis. Consider a pairing
of n brackets. At each step, let i be the number of unmatched brackets to
the left of the ith bracket. So (()()) would be encoded 0,1,2,1,2,1,0. If we
now interpret the i as the heights of a Dyck path, we see that there is a
one to one correspondence between pairings of n brackets and Dyck paths of
length 2n.
We give a bijective proof of a formula for the number of Dyck paths and
hence a formula for Cn . Consider Dyck paths of length 2n. Now add a
downward step at the end of the Dyck path. There are now n + 1 downward
steps. Choose one of these steps to be special. Then we have (n + 1)Cn
possible congurations. Now cut the Dyck path into two parts, the part
before the special step and the part after it (including the special step).
Move the right part so that it starts from (0; 0) and append the left part to
the right part. We now have a special step going from (0; 0) to (1; 1) and
a path with steps (1; 1) and (1; 1) going from (1; 1) to (2n + 1; 1). We
can recover the Dyck path with the extra downward step and special step
by nding the leftmost minimum of the new path, cutting it into two parts
and reassembling. Thus this operation is a bijection. See Figure 1 for an
example.

Clearly the number of paths from (1; 1) to (2n + 1; 1) in is nn . So
2

we nd that
 
Cn = n + 1 2nn :
1

3 Wigner's Semi-circle Law

Wigner's semi-circle law essentially says that the histogram of the eigenvalues
of a random symmetric matrix are approaches a semi-circle. Wigner [3] rst
derived this for a special case where he allowed the entries to be v.
In his derivation he computed the asymptotics of the moments of the
distribution by considering the trace of H k and used this to calculate the
resulting characteristic function (Fourier transform of the distribution).

2
Figure 1: Bijection in derivation of formula for Cn.

3
 Due to the symmetry of the distribution, the terms in the expression for
the  th moment
X X
M = R 1
v i1


vi 1 ;
0 0
set i1 ;:::;i 1

are non-zero only if 0,i , i , : : : , i ,0 is a valid sequence. A valid sequence,

1 2 1
is a sequence of the form 0,i , i , : : : , i  , 0, where no two successive
1 2 2 1
members of the sequence are equal and each step jl and lj occurs an even
number of times.
Wigner then shows that the number of valid sequences of length 2 + 1
is asymptotically t (2N ) v  , where t is the number of type sequences with
2

length 2 + 1.
A type sequence of length 2 + 1 is a sequence that starts at 1 and ends
at 0, with successive members diering by 1 and all members non-negative.
By interpreting the type sequence as the heights of a Dyck path we see that
the number of type sequences t is the  th Catalan number.

4 Random Regular Graphs

McKay [1] derived the eigenvalue distribution of a large random regular
graph. Again the approach is to calculate moments of the distribution. Here
a diagonal entry of Ak has the interpretation as the number of closed walks
of length k.
McKay gave a derivation for the exact number of closed walks by using
combinatorial arguments, arriving at the following result.
Let  be a vertex of a graph X . A closed walk of length r  0 starting
0
at  is a sequence  ;  ;  ; : : : ; r of vertices of X such that r =  , and
0 0 1 2 0
i is adjacent to i for 1  i  r.
1

Lemma 2.1. Suppose X is regular of degree v. Let  be a vertex of X, and 0

suppose that the subgraph of X induced by the vertices at distance at most
r=2 from  is acyclic. Then the number of closed walks of length r in X
0

4
starting at 0 is (r), where (r) = 0 if r is odd, and
s  
(2s) =
X 2s k k vk (v 1)s k
k =1
s 2s k
s  
= v
X 2s s k (v 1)k
1

k k s =0
s  
=
X 2 s 2s 2k + 1 (v 1)k :
k k 2s k + 1
=1

Edelman and Strang later independently derived an expression for the

number of closed walks by calculating the rst few values and then passing
it to the superseeker (http://www.research.att.com/~njas/sequences/-
eisonline.html) to guess the general formula

n 1
X
mn (z) = (1 + z) n (1 + z) j z n jC
n j:
2 +1 2 +1 1+

j =0
In the notation of McKay's paper, z = v 1 and s = n +1, and the expression
becomes
s 1
X
(2s) = v s v j (v 1)s j +1 C
s j: (1)
2 1 2 1

j =1
This formula also involves Catalan numbers and it turns out that it has
a straightforward combinatorial interpretation.
Let  =  ;  ; : : : ; r be a walk of length r. We can think of this as a
0 1
traversal of the vertices of the graph. Let i be the depth of the traversal
at step i. We decrease i by one if we \backtrack" and increases i by one
otherwise. More formally,
m(d; i) = maxfj : j < i; j = dg
then
  = 0,
0

 i = i 1 if m 
+1 ( i 1 ) ;i = i ,
+1

5
 v 1
v 1
v 1 v 1 v 1
v 1 v 1 v 1
v v v v
Figure 2: Plot of i with s = 3, j = 2.
v 1
v 1 v 1 v 1
v 1 v 1
v v
Figure 3: Plot of i with s = 3, j = 1.

 i = i + 1 if m 
+1 ( i ;i
1 ) 6= i .
+1

Note that for all closed walks in the setting of Lemma 2.1 this denition is
equivalent to the denition given in McKay's paper because of the acyclic
condition of the subgraph. Thus the number of closed walks is given by the
number of walks with  s = 0.
2
Now clearly all closed walks must end with a backtrack. The number of
walks that end with a backtrack is thus v s , the rst term of Equation 1.
2 1

Now if one of these walks is not closed, then  s = 2k, where k  1,

2
because of parity considerations. In particular,  s  2. Then let 2j 1 =
2
m( s 1; 2s). For any given j there are v j ways to traverse the rst 2j 1
2
2 1

steps. The remaining 2s 2j +1 steps must have i >  j for all i > 2j 1.
2 1
Plotting the values of i we can think of this as one \upward" step followed by
a Dyck path of length 2(s j ). Now each upward step can occur in v 1 ways
and each \downward" step (corresponding to a backtrack), can occur in only
one way. So the number of ways of doing this will be (v 1)(v 1)s j Cs j .
Summing over j gives us Equation 1.

6
5 Conclusion
The representation of Catalan numbers as the number of pairings of n pairs
of brackets is the most well known but is not necessarily the most useful
one. In this project, we showed two applications where using the Dyck path
representation made it easier for the author to nd direct combinatorial in-
terpretations of various formulae. The reason Dyck paths may be easier to
work with is the graphical nature of their representation.

6 Acknowledgments
To Frank Calegari for recalling the key idea for the bijective proof of the
formula for the number of Dyck paths.

References
[1] B. D. McKay. The expected eigenvalue distribution of a large regular
graph. Linear Algebra and its Applications, 40:203{216, 1981.
[2] R. P. Stanley. Enumerative Combinatorics, volume 2. Cambridge Uni-
versity Press, 1999.
[3] E. P. Wigner. Characteristic vectors of bordered matrices with innite
dimensions. Annals of Mathematics, 62(3):548{564, November 1955.