2 is an integer. 'oaip
which is just the definition of a, Similarly, (2.2) is equivalent to
We now have the theorem:
If l(n) is the cardinality of the largest monotone increasing subsequence of
.p,(n 1)} then
i i
(1.2) lim inf~,.n 2 l(n) = 21, lim sup,,,.n 2 l(n) forp = 2,
2
and
(1.3)
2
2
lim infn,.n1 l(n) = 2(1 p lim sup,,..n l(n) = pi forp > 2.
This result is as precise an analogue to (1.1) as one could realistically expect.
Also, the LoganShepp lower bound on c makes it particularly noteworthy that
.L
(1.4) limp..lim.inf,,,.n 2 l(n) = 2.
For digestibility, the proof of these results has been divided into four parts. First we obtain a geometrical characterization of the monotone subsequences of van der Corput's sequence. Upper and lower bounds are then obtained for l(n). Finally the required limits are identified.
i i i
1
PSEUDORANDOM HA~RSLEY LAW 269
2. The geometrical representation. Our first goal is to obtain a representation
from which detailed information about {opp(O), 9?P(I), op
,,(n I)) can be de
duced. We define a,, to be the unique permutation of {% 1, n 1} which
puts 9)p(i), 0 ~Q i %Q n 1 in increasing order, i.e.,
qgp(a.(0)) < 9)p(a.(1)) < ... < (pp(a.(n
This permutation will be written as a sequence
a. = (a.(0), a.(1), . . , an(n 1))
and our reason for introducing an is the elementary fact that the length of the longest monotone increasing subsequence of an is equal to 1(n).
We also have the following law of formation of a .. P. which will be crucial.
LEmmA 2. 1. One obtains cr.. from ap. by replacing the entry ap.(i) of up. by the sequence ap.(i) + p no..
REmARK. Here and in the following if x is a sequence then a + bx is a sequence with the same domain as x defined by (a + bx)(i) = a + bx(i).
PROOF OF LEmmA. Notice that there are two things we must show:
+ pna.(j)) < 99
P(a
(2.1) P.(i) + pn,.(j +
and
,.(i) +p nCF.(j)) < (p
P(a
(2.2) P.(i + 1) + pn,.(f)),
for anyiJ.
We first suppose j ~pn and k is any integer. Setting j = l"i0'i2ip', k
E"PnakP 1n we havej + p"k W,_0apl, and consequently
(2.3) pp(j + p"k) 2 T,Oaip (" 1)
Enlap 1) + p n (i+ 1)
10 2 78 n aiP
pp(j) + pnpp(k) for j
pn., np
P(Ujj))
(2.4) fp"(op.(i)) + p nrp,(0.(i)) < Ipp(ap.(i + 1)) + p nqgp(a.(j,)).
{O, I'. pn _ 1),.p
To check (2.4) we note that, as i runs through P(i) MnS
through {O, p n, 2p n, . . . ' (p n _ 1)p n } so that, in fadt, qpp(ap.(i + 1)) = p n +
Tp(up.(i)). The proof of the lemma is thus complete.
In many ways it is easier to work with the permutation matrix An associated with an. We recall that An is an n x n matrix defined by
An(iJ) = 1 if i = an(j) =D otherwise.
270 A. DEL SUNCO AND J. MICHAEL STEELE
To rephrase Lemma 2.1 in terms of matrices we define m different p" X mp"
matrices Ap, ' lp'~y APT by
A,,.(i,jm~') if mli
1 otherwise and
1)mod m).
Ap'.(iJ) (j
p
Intuitively, we getP_. by inserting m 1 columns of O's after each column of Ap.
p
and then shift 1~ by 1 places to the right to get
Finally we can phrase Lemma 2.1 by the following formula for A.. as a block matrix:
p
'T>._ 1 (2)
(2.5) Anw
This matrix representation of the permutation permits a geometrical view of the decreasing subsequences. Formally an increasing subsequence of a,, is a sequence of integers P = {jl
3. The lower bound lemma. The main result of this section is the following:
LEmmA 3. 1. For P"2 ;Q m r% p"~' we have
(3.1) 1(mp n) >p n 1 + m(p 1) (3p 1),
andforpn3 < M Q pn2 we have
(3.2) 1(Mpn) >pn3(p _ 1) + MP _ (P2 + 2p _ l).
PROOF. First we will consider m such that pn2 ,Q M 4Q p n 1 and construct an explicit path through A... By setting m, = (pn1 m)l(p 1) and M2 = M M, we have m,p2 + m2p =p n. For 0 < i < [m,] we can define a path P, through Ap. of length 2p 1 and domain [P2i,p2(i + 1) 11 by '
pi = (P2i, p~ + p, p2i + 2p, _ ' p2(i + 1) p,
p 2(i + 1) _,p + 1, 'p2G + l) l).
For 0 Q i < [m21 we then define a path through Ap. with length p and
domain 1 M1]p2 + ip, [M1]p2 + (i + 1)p 1] by
P[m,l+i = Q M1]p2 + ip, [M1]pI + ip + I'. . . ' [M1]p2 + (i + 1)p l).
i i
i i
i
PSEUDORANDOM HA~RSLEY LAW 271
Now we consider the paths F, in ~ '(') which correspond to Pi. These paths are
1 or
defined by mP, + aj 1(i). Finally we define our desired path P through A...
as the concatenation of the
P = POPI ... P[.j+[%]].
To make certain that the path P is properly defined we have to check that for each k e dom lisi~ , 1 and 1 E= dorn Ai we have k > 1. This condition will be abbreviated by domFi+l > domF,, and it follows immediately from the corresponding fact that dom P,+ 1 > dom P,. By our construction we finally have
1(Mpn) > l(P) = [M1](2P 1) + [M2]P
> M1(2P 1) + M2P OP 1)
= P n1 + m(p 1) (3p 1)
which completes the proof of (3.1).
We now consider M'pn3.Q M Q pn2 and begin by setting m, = (p`2
m)l(p 1) and M2 = m m,. This time we have mp 3 + M2p2 = pn. As in our previous construction there is a path of length p 2 through Ap3, so for 0 < i < [m,] there is a path Pi of length p 2 through Ap. with domain [ip3, (i + 1)p3 1 ]. For 0 Q i < [m21 there is a path P[mj+i of length 2p 1 through Ap. with domain 11 M1]p3 + ip2, [M1]p3 + (i + 1)p2 1]. Next we set A, = mPi + ajl(i) and obtain a path through by letting
p =
Finally we calculate as before:
1(,,n) > 1(p) MI ] p2 + [ M2](2P 1)
> M1P 2 + M2 (2p (p2 + 2p 1)
= p n2(p + nP (P2 + 2p 1)
which completes the proof of (3.2).
REmARK. The preceding proof probably appears more involved than an examination of an example would indicate. By calculating A P. for p = 3 and several small values of m and n one can easily find the path P and see how it evolves as m and n increase. Similarly in the arguments which follow one should keep the example p = 3 clearly in mind, perhaps by keeping a small table of the A,..
4. The upper bound lenuna. The recursive nature of A,,. provides the key to the'following, Lemma 4. 1. For any positive integer m we have
(4.1) 1(mp",) Q P` + M(P
and
(4.2) l(Wn) < pn 1 _ pn2 +
i
i i i
1
i
i
1
272 A. DEL "CO AND J. MICHML STEELE
PROOF. For any path P through A... we can decompose P as a concatenation
P = P0P1 ... Pm 1
where P, is a path through AP5 (1). Here one should note that the range of Ai is a subset of [ip.", (i + 1)p" 11.
We now let P, be the corresponding path in Ap. of F,. That is, we let Pi m m l(Ai a,~ 1 (i)). We do not have dom. Pi < dom. P,,, but we obviously do have dom P, < dom P,+, so the two paths have at most one common point. If dom P, n dom Pi+l :,'= 0 we shall say Pi and P,+, are linked. Since I(P) ~ W1 P) the
i0 1( lemma would follow from E' '1(Pi) < min(p` 1 + m(p 1), p, 1 p" 2 + i0
mp). This fact will be proved using only the inequality dom Pi < dom. Pi + 1.
First we decompose Ap. into p" blocks of height p` and width p. We note that each block has exactly one 1, and the pattern of I's in each column of blocks is the same as the pattern in Ap.
The block width w(P) of a path P is defined as the number of distinct columns of blocks which intersect P. The block height h(P) is defined correspondingly. Further, we let g equal the number of integers i such that Pi is linked to P,,,.
If P, and P,+ 1 are linked we define their linkage as the concatenation PiPi*, 1 where Pi*+ 1 denotes P, with its first entry removed. By forming the successive linkage of the P, we obtain m tt paths P11A, Pm, which are unlinked and for which
(4.3) + li~i(Pi)
Next we form all possible concatenations of successive P, to get s paths P,', P2, P.,' which allow no further concatenation. One then has s < m tt and
(4.4)
1 I(Pil) mi p,
Moreover, one notes that no two P,' share a common column of blocks for the simple reason that dom, Pi < dom P,, , and the 1 in any block is to the right of the 1 in any block above it.
The crucial observation is that for each path Pi' we have
,(P,') < h(Pi') + w(Pi') 1 < p + w(Pi') 1
which by summing over i yields
(4.5) E~ 1(Pi') < ES1w(p,') + S(P I).
1 1 i
Since no two Pi' share a column of blocks we have
13_ IW(pi,) < pn 1.
(4.6) 1
By the inequality (4.5) we obtain
(4.7) 2:~1 1(pi,) < pn1 + S(p
which by settings = m tt X with X > 0 using (4.3) and (4.4) becomes
Em~ 11(pi) < pn1 + (M X)(p 1) tt(p 2).
(4.8) i0
i
i
1
1
PSEUIDORANDOM HA~RSLEY LAW 273
This completes the proof of (4. 1) since X > 0 and tt > 0. On the other hand (4.6) also yields s < p n 1 so tt < p ` 1 (m X) which by (4.8) implies
EM1 + (M X)(P 1) + (pn 1 (m X))(p 2)
i_ol(pi) Q pn p n p n1 +
This implies 1(Mpn) .1~ p" p n` + m and by replacing n by n 1 and m by nw we obtain the proof of (4.2). This completes our proof of the upper bound lemma.
By applying the preceding lemmas, we should note that in certain cases the inequalities can be made to provide equality. By inspecting the proof of (3.1) in the cases of M = p n 1 and M = p n 1 one sees that it is not necessary to estimate greatest integers so that the 3p 1 disappears from the right of (3.1). Combining this with (4. 1) we have the identities
(4.9) 1(p2nl)=pn and I(P2n 2) =p n2 (2p 1).
By similar analysis for p 2 we also have the identity
(4.10) 1(m2n) 2` + m for 2n2 < M < 2n1
2` + 2m for 2` Q m ,Q r'.
5. Identification of the limits. To complete the proof of our theorem, we first
1 S = (,,n : pn1 %Q M identify lim supn~.N 21(N) where N. is restricted to
p n1 ).
LEmmA 5. 1. Setting K = max(2(1 p )12, p 2L, 2 p') and k = min(2(1
1).L, 1,
p 2 P2 2 p'), we have
(5.1) "M SUPN c: sN I(N) K,
and
(5.2) lim infN c: sN 21 I(N) k.
PRooF. As in Section 3 we deal with p n 2 ,;;
. M .~ pn 1 (case 1) and pn3 ,Q M 4Q p n2 (case 2) separately. This time we begin with case 2. By (3.2) and (4.2) we have
g.(M) (p2 + 2p + 1)(.n)12 ~ 1(npn)(Mpn)21,Q gn(M)
2(p 2
where gn(x) = (p` _ 1) + Xp)(Xp n) ~ 1. Consequently we have
liM inf I(Mp.)(Mpn)L = lim inf gn(m),
(5.3)
where the lim inf is taken over the set S, = {,,n :pn3 ,Q M :,Q pn2, 1 < n). One has the same equation as (5.3) for, the lim sup.
Next observe g,(m) has its minimum at a,, =p n 3(p 1) and gn is decreasing
3 <
for p` Q x ,' an and increasing for an . X Q pn2. One easily checks that
.L g,(pn3) = p21, g,(pn2) . 2 p', and gn(a.) = 2(1 p)2 so we have
(5.4) lim infw~ s, g,,(m) = k and lim sup,,p. s, gn(m) = K.
274 A. DEL "CO AND J. MICHAEL STEELE
To prove the comparable identities for case 1 one sets f,(x) = (p n 1 + X(p
1))(xp")'2L and obtains from (3.1) and (4.1) that
f,.,= S'l(np n)(Mp n) 1 = liM SUp P.'S2f
lim in 2 "(M)
Where S2 = {Mpn : pn2 < ,n 1, n > 1}. The same equality holds for the limsup, and we note as before that f. has a minimum at p n 1(p l)` = bn, decreases for p n2 ,Q X '~ bn and increases for bn Q X ,Q pn '. Finally we note fn(bn) = 2(1
)I, i, f 1)
, f 2
p n(pn~ 1) = p and n(pn 2 p Hence. we have
(5.5) lim infmp.,=sfn(m) > k and lim supw.,=s~fn(m) ~Q K.
One has inequality in (5.5) since bn is not an integer. Actually equality can be
proved but is not required for the rest of the proof. By (5.3), (5.4) and (5.5) we have
completed the proof of (5. 1) and (5.2) as required by the lemma.
The proof of our theorem can now be completed in a routine way. For an arbitrary integer N we write
Mp n pn3 Q M
(5.6) N = + r with . pn 1 and 0 4Q r < pn+2.
Letting 1(i,j) denote the cardinality of the largest increasing subsequence of
(pp(i + 1), fpp(j 1)) we see
(5.7) 1(ipk, (i + 1)pk) = 1(pk)
as a consequence of (2.3). There is also the obvious fact that l(iJ) is subadditive,
1(i, k) < l(iJ) + 1(j, k) for i ~jQk.
By choosing an integer t such that tp"12 <.. np" and ,n +pn+2 (t + 2)p +2 one has by (5.6) that
l(N) ,Q I(Mpn + pn+2)
< I(Mpn) + I(Mpn, Mpn + pn+2)
.Q I(Mpn) + l(tpn+2, (t + 2)pn+2)
I(Mpn) + 21(pn+2).
Hence we have
(5.8) 1(Mpn) :Q I(N) Q I(Mpn) + 21(pn +2).
1(pn+2)1(Mpn)i
Since for large N one has NImpn is near 1 and 2 is near 0, the
inequality (5.8) together with Lemma 5 completes the proof of the theorem.
REFERENCES
[11 1)AER, R. M. and BRocK, P. (1968). Natural sorting over permutation spaces. Math. Conw. 22 385410.
[2] HA~RsLEY, J. M. (1972). A few seedlings of rewarch. Proc. Sixth Berkeley Synw. Math. Statist. Probability 1345394. Univ. of California Press.
[3] H~RsLEY, J. M. and ~Scoms, D. C. (1964). Monte Carlo Metho*. Methuen, London.
i i i i i
i
1
1
PSEUDORANDOM HAMNERSLEY LAW 275
[4] KiNomAN, J. F. C. (1973). Subadditive ergodic theory. Ann. Probability 1883899.
[5] KNum, D. E. (1969). The Art of ConWuter Progra~hg, 2 (Semintonerical Algorithm). AddisonWeslay, Palo Alto.
[6] LooAN, B. F. and SHEPP, L. A. (1975). A variational problem for random Young tableaux. Advances in Math. 26 206222.
[71 VERsHiK, A. M. and KERov, S. V. (1977). Asymptotics of the Plancherel measure of the symmetric group and the limiting form of Young tables. Soviet Math. Doki. 18 527531.
DEPARTMENT OF MATHEMATICS DEPARTMENT OF STATISTICS
OHIO STATE UNWERSITY STANFORD UNIVERSIrY
CoLumBus, 0wo, 43210 STANFORD, CALiFoRmA 94305
i
i
1