1995 Putnam exam. Here you will find
I. The questions
II. Answers I posted
III. Completions of the problems, and other comments, posted or
emailed to me by others.
You will note that A6 and B6 are the hardest, generating a lot of
interesting commentary, extensions, and differing solutions; these
discussions occupy over half this file. There was also a fair amount
of commentary on A4 and B5.
I. First the questions
From: rusin@washington.math.niu.edu (Dave Rusin)
Newsgroups: sci.math
Subject: Putnam exam: list of questions
Date: 3 Dec 1995 08:01:37 GMT
56th Putnam Exam, 12/2/95 [You have 6 hours, good luck.]
A1. Let S be a set of real numbers which is closed under multiplication
(that is, if a and b are in S then so is ab). Let T and U be
disjoint subsets of S whose union is S. Given that the product of any
_three_ (not necessarily distinct) elements of T is in T and that the
product of any three elements of U is in U, show that at least one of
the two subsets T, U is closed under multiplication.
A2. For what pairs (a,b) of positive real numbers does the improper
integral
\integral_b^\infty ( ... ) dx
converge?
[sorry for loss of clarity; integrand is
sqrt( sqrt(x+a) - sqrt(x) ) - sqrt( sqrt(x) - sqrt(x-b) )
--djr ]
A3. The number d_1 d_2 ... d_9 has nine (not necessarily distinct) decimal
digits. The number e_1 e_2 ... e_9 is such that each of the nine
9-digit numbers formed by replacing just one of the digits d_i in
d_1 d_2 ... d_9 by the corresponding digit e_i (1 <= i <= 9) is
divisible by 7. The number f_1 f_2 ... f_9 is related to e_1 e_2...e_9
in the same way: that is, each of the nine numbers formed by replacing
one of the e_i by the corresponding f_i is divisible by 7. Show that,
for each i, d_i-f_i is divisible by 7.
(For example, if d_1 d_2 ... d_9 = 199501996, then e_6 may be 2 or 9
since 199502996 and 199509996 are multiplies of 7.)
A4. Suppose we have a necklace of n beads. Each bead is labeled with
an integer and the sum of all these labels is n-1. Prove that we can
cut the necklace to form a string whose consecutive labels x_1, x_2, ...
x_n satisfy
Sum_{i=1}^k x_i <= k-1 for k=1, 2, ..., n
A5. Let x_1, x_2, ..., x_n be differentiable (real-valued) functions of a
single variable t which satisfy
dx_1/dt = a_11 x_1 + a_12 x_2 + ... + a_1n x_n
dx_2/dt = a_21 x_1 + a_22 x_2 + ... + a_2n x_n
... ... ...
dx_n/dt = a_n1 x_1 + a_n2 x_2 + ... + a_nn x_n
for some constants a_ij >= 0. Suppose that for all i, x_i(t) -> 0 as
t --> oo. Are the functions x_1, x_2, ..., x_n necessarily linearly
dependent?
A6. Suppose that each of n people writes down the numbers 1, 2, 3
in random order in one column of a 3 x n matrix, with all orders
equally likely and with the orders for different columns independent of
each other. Let the row sums a, b, c of the resulting matrix be
rearranged (if necessary) so that a <= b <=c. Show that for some
n >= 1995, it is at least four times as likely that both b=a+1
and c=a+2 as that a=b=c.
==============================================================================
B1. For a partition pi of {1, 2, 3, 4, 5, 6, 7, 8, 9}, let pi(x) be
the number of elements in the part containing x. Prove that for any
two partitions pi and pi', there are two distinct numbers x and y
in {1, 2, 3, 4, 5, 6, 7, 8, 9} such that pi(x)=pi(y) and
pi'(x)=pi'(y).
(A _partition_ of a set S is a collection of disjoint subsets (parts)
whose union is S.)
B2. An ellipse, whose semi-axes have lengths a and b, rolls without
slipping on the curve y = c sin(x/a). How are a, b, c related, given that
the ellipse completes one revolution when it traverses one period of the curve?
B3. To each positive integer with n^2 decimal digits we associate the
determinant of the matrix obtained by writing the digits in order across
the rows. For example, for n=2, to the integer 8617 we associate
det([[8,6],[1,7]]) = 50. Find, as a function of n, the sum of all the
determinants associated with n^2-digit integers. (Leading digits are
assumed to be nonzero; for example, for n=2, there are 9000 determinants.)
B4. Evaluate
________________________________
/ 1
\ 8 / 2207 - ------------------------ .
\ / 1
\ / 2207 - ----------------
v 2207 - ...
Express your answer in the form (a + b sqrt(c))/d, where a, b, c, d
are integers.
B5. A game starts with four heaps of beans, containing 3, 4, 5 and 6 beans.
The two players move alternately. A move consists of taking _either_
a. one bean from a heap, provided at least two beans are left
behind in that heap, or
b. a complete heap of two or three beans.
The player who takes the last heap wins. To win the game, do you want to move
first or second? Give a winning strategy.
B6. For a positive real number alpha, define
S(alpha) = { [n alpha] : n = 1, 2, 3, ...}.
Prove that {1, 2, 3, ...} cannot be expressed as the disoint union of three
sets S(alpha), S(beta), and S(gamma). (As usual, [x] is the greatest
integer <= x . )
==============================================================================
II. Now the answers as I first posted them. There are some corrections and
comments in the next section; I'll flag them here.
From: rusin@washington.math.niu.edu (Dave Rusin)
Newsgroups: sci.math
Subject: Putnam spoilers -- first go-round
Date: 3 Dec 1995 06:25:17 GMT
Well, it's probably legal to post answers to the Putnam exam now. I didn't
quite finish problem A-6, but I think I have reasonably complete (and
usually nice) answers to the other. I thought the exam was pretty fair this
year -- many of the questions seem like the kind of thing one might use
to challenge students within a particular course.
I await the usual corrections and slick improvements :-).
dave
[Added in a later post:
By the way, while I have the floor (screen?) let me acknowledge that some
of the A-questions were the subject of conversation at lunch time, so
that the thoughts in the answers I wrote down may not be mine alone.
In particular, Yining Xia had some helpful comments on several problems.
dave
]
A1. Suppose not. Then there exist t1, t2 in T and u1, u2 in U such
that t1*t2 lies in U and u1*u2 lies in T. Thus we have by
assumption that t1*t2*(u1*u2) lies in T, and yet also
u1*u2*(t1*t2) lies in U. The usual axioms of multiplication apply
in R so these products are equal, contradicting the assumption that
U and T are disjoint.
A2. This is the kind of problem undergraduates should learn to do
as soon as they begin to work with improper integrals: one learns to
estimate by finding the dominant terms.
If a=b, the integral converges, otherwise not.
Indeed, the integrand is x^(1/4) times
sqrt(sqrt(1+a/x)-1) - sqrt(1-sqrt(1-b/x)). Using Taylor series, say,
one can show that for large x, sqrt(1+a/x) = 1 + a/(2x) - a^2/(8x^2) * C,
where C approaches 1 as x--> oo. Thus the first half of the integrand
is sqrt(a/2x - a^2/(8x^2)*C) = sqrt(a/(2x))*sqrt(1-a/4x*C) =
sqrt(a/2x) -sqrt(a/2x)*(a/8x)*C'. Likewise, the second half of the integrand
is sqrt(b/2x) +sqrt(b/2x)*(b/8x)*C''.
So if a=b, the leading terms drop out. What remains is a constant
multiple of x^(-3/2), multiplied by a term which approaches 1 as x-->0.
In particular, we can find a constant D such that the entire integrand
is less than D x^(-5/4) for sufficiently large x, making the
integral over [x1, x2] be less than D(x1^(-1/4) - x2^(-1/4)); as
x2 --> oo, this converges. (The integral over finite regions converges
by continuity of the integrand.)
On the other hand, if a<>b, the integrand is now of the form
x^(-3/4)[sqrt(a/2)-sqrt(b/2)] plus a function as discussed in the previous
paragraph. While the integral of that function converges, the
integral of the first part is a non-zero constant multiple of
(x1^(1/4) - x2^(1/4) ), so that as x2 --> oo, the integrals diverge.
A3. Let d be the number expressed by the digits d1...d9, and likewise
e and f. Then we are given that for each i, the number
d + (e_i-d_i)*10^(9-i) is a multiple of 7. Summing, we see
9d + (e-d) ==0 mod 7, i.e., d == -e. Since likewise each
e + (f_i-e_i)*10^(9-i)==0 and e==-f, we add the congruences to see
(f_i-d_i)*10^(9-i) ==0. Since 10 is a unit mod 7, we divide and
conclude f_i == d_i, as needed.
A4. Suppose this is not possible. Let us pick any starting point on
the necklace and either of the possible directions, and read off the
labels to get a sequence of integers a_1, a_2,... Note that
a_i=a_j if i=j mod n, and that by assumption the sum of any n consecutive
terms is less than n.
If an appropriate cut in the necklace is not possible, then for any i
there is a j such that a_i+a_(i+1)+...+a_j is at least equal to the
number of terms (j-i+1); in other words, the average of this subset is
at least 1.
Very well, find the first such j which corresponds to i=1; call it j_1.
Find the first such j corresponding to i=j_1+1; call it j_2.
Proceed in this manner to find consecutive stretches within the
sequence whose average is at least 1.
Consider the collection of reductions of the j_k modulo n -- there can
only be a finite number, so that at some point we will have j_k = j_l
mod n. Then it is true that the each of these sequences
a_(j_k + 1), ..., a_(j_(k+1))
a_(j_(k+1)+1),...,a_(j_(k+2))
...
a_(j_(l-1)+1),...,a_(j_l)
consists of terms with an average at least equal to 1. Consequently
the lot of them
a_(j_k+1), ..., a_(j_l)
must also average at least 1.
On the other hand, since j_k = j_l mod n, this sequence consists of
a number of repetitions of the whole sequence of labels taken around
the circle, hence their average is the same as the average of those
n labels, which is less than one. This contradiction shows that at some
point the construction of the indices j_k must fail, i.e., that
from some starting point it is possible always to keep the averages
less than 1.
A5. Yes, the functions are necessarily linearly dependent.
The eigenvalue procedure used here should be familiar to those who
have taken a course in differential equations which covered linear
equations with constant coefficients.
We are given a matrix A of non-negative coefficients and functions
x=(x_1(t), ..., x_n(t)) with Dx=Ax and x(t) --> 0; the question
asks if there is necessarily a linear combination of the x_i which
vanishes identically.
Rephrasing in a coordinate-free way, we have a function x:R -> R^n
which tends to 0 in R^n and such that Dx=Ax for some linear map
A: R^n -> R^n; we ask if the image of x lies in some proper subspace
of R^n. If a is a real eigenvalue of A, we may choose a new basis
of R^n which begins with an eigenvector corresponding to this
eigenvalue; then Dx_1=ax_1 means x_1 is a multiple of exp(a t).
Since x(t) --> 0, either this multiple is zero (so that x does
lie in a subspace) or a<0. Likewise if a+bi is a complex eigenvalue,
then its complex conjugate a-bi is also so that in a suitable basis we
have (x_1, x_2) = exp(a t)*(cos(bt), sin(bt)) up to a constant multiple.
So again, either the image of x lies in a subspace or the real part a
of these eigenvalues is negative.
We conclude that the conditions Dx=Ax, x(t)-->0, and A real imply
that either the image of x lies in a subspace of R^n or else all
the eigenvalues have negative real part. In particular, this last
condition will mean the trace of A will be negative. This is impossible
if A has only non-negative entries.
A6. We have n vectors (x_i, y_i, z_i) consisting of permutations of
{1,2,3}. We are investigating the sum (s, t, u) of these vectors. It
seems more helpful to let x'_i=x_i-2, y'_i=y_i-2, and z'_i=z_i-2 so that
x'_i+y'_i+z'_i=0, and thus (x,t,u)=(2n, 2n, 2n)+(s', t', u'). These
new vectors can be thought of as random steps in the plane x+y+z=0,
each of length sqrt(2) but in directions 60 degrees apart (as may be
computed by noting that the cosine of the angle between any two is
a dot product (equal to +-1 or +-2) divided by the lengths). So the
issue is to compare the probabilities that a random walk of this type
returns to the origin or to one of the six vectors neighboring the origin.
(Geometrically, this motion can be seen as a random walk on an
infinite Chinese-Checkers board.)
If we let p_n(v) be the probability the walker is at position v
after n steps, then a standard treatement of conditional probability
shows p_(n+1)(v) is the average of the value of p_n(w) at the
six neighbors w of v.
Clearly, over the long run, this averaging will make p_n(v) nearly
constant over short regions; in particular, p_n(0) will roughly equal
the probability p_n(v) of landing at any one of the 6 neighbors of 0,
and is thus 1/6 of the probability of landing somewhere in that set
(for large n). This will force p_n(0) to be less than 1/4 of that
probability, of course.
Now, the function p_n is the n-fold convolution of p_1, which ought
to make it possible to evaluate p_n(v) precisely, but I'm sure there's
a way to estimate just a few values such as p_n(0) without as much
work. Unfortunately I've just had to endure a few hours of a social function
and my brain is now blocked; someone else will have to provide details.
==============================================================================
B1. Let's try to see what partitions would look like which did _not_
have pi(x)=pi(y) and pi'(x)=pi'(y) for any pair of x and y.
First note that there can not be any equivalence class under pi which
has four or more elements. If, say, {a, b, c, d} is contained in a
single equivalence class under pi, then by assumption each of these
four lie in different equivalence classes under pi'; indeed, they must
lie in classes of different cardinalities. This would force the sum of the
cardinalities of the equivalence classes under pi' to be at least
1+2+3+4=10, a contradiction.
Likewise, pi cannot have two or more equivalence classes of cardinality
three, nor of cardinality two; and it cannot have four or more classes of
cardinality one.
So the sum of the sizes of the classes under pi is at most
3 x 1 + 1 x 2 + 1 x 3 = 8, which is insufficient.
An examination of the proof shows that 9 is the maximal cardinality of
the large set; indeed consider these partitions of {1, ..., 10}
{{1, 2, 3, 4}, {5, 6, 7}, {8, 9}, {10}}
{{1, 5, 8, 10}, {2, 6, 9}, {3, 7}, {4}}
B2. Let b be the larger of the two semi-axes.
The length of the arc is integral sqrt( 1 + ((c/a) sin(x/a))^2 dx
for x in [0, 2 pi a]. Let u= x/a; this is then the integral
integral sqrt( a^2 + (c sin u)^2 ) du, u in [0, 2 pi]
The length of the ellipse is twice the length of the curve
(x/a)^2 + (y/b)^2 = 1 above the x-axis. This is then
2*integral sqrt(1 + ((b/a)^2 x/y)^2 ) dx, x in [-a, a]
using implicit differentiation. Now, a variation of the trig parameterization
of an ellipse suggests the substitution x = a sin(u), y = b cos(u)
(with u ranging over [-pi/2, pi/2]) to give us
2*integral sqrt(1 + ((b/a) sin(u)/cos(u) )^2 ) (a cos u) du
=2*integral sqrt(a^2 cos^2 u + b^2 sin^2 u) du
=2*integral sqrt( a^2 + (b^2-a^2) sin^2 u) du
Using the periodicity of the sine, this is the same as the integral
integral sqrt( a^2 + (b^2-a^2) sin^2 u) du, u in [0, 2 pi]
If the ellipse rolls without slipping so that one revolution traverses
one period of the curve, then the two lengths must be equal. Clearly
c=+-sqrt(b^2-a^2) will make that true, but it is clear geometrically
that the length of the sine curve increases with |c|, so these can
be the only solutions.
B3. We need to find the sum
Sum det(A)
where A ranges over all n x n matrices with entries in {0, 1, ..., 9}
and with A_{1, 1} <>0.
The key observation is that most of these determinants cancel. Excluding
the matrices with the bottom two rows equal, we can pair each matrix with
a unique other matrix by permuting the bottom two rows. These matrices
have determinants which are negatives of each other.
All the remaining matrices have determinant zero since they have two
equal rows!
Grand total: zero.
The argument fails when n=1 or 2. In the former case, we have simply
Sum (n, 1 <= n <=9 ) = 45. In the latter, we would have a pairwise
cancellation as before but for the lack of matrices with a 0 in
the upper left; thus the sum is the negative of Sum det [[0,a],[b,c]]
which is clearly Sum( -ab : a, b, c in {0, 1, ..., 9}). This in turn
equals -(Sum(a))*(Sum(b))*(Sum(1)) (the sums taken over a, b, and c
respectively), so that the number we seek when n=2 is 45^2*10=20250.
B4. The ellipsis is unclear, I suppose, but it presumably means that
if x is the continued fraction 2207-1/(...), then x=2207-1/x.
Therefore, x^2-2207x+1=0. The number y we seek has y^8=x, so that
y is the solution to the equation y^16-2207y^8+1 = 0.
The hint suggests that y satisfies a quadratic equation with integral
coefficients, so we try to factor the above polynomial. Starting with
a factor of the form (y^8 + a y^4 +- 1) we see y^8 +-47 y^4 +1 are
indeed factors, one of which has no real roots. In turn we likewise try
factors of the form (y^4 + a y^2 +- 1) and deduce y^4 +- 7y^2 + 1
are factors, again only one with a real root. Finally, we try to
factor this with (y^2 + a y +- 1) and see y satisfies y^2 - 3y + 1,
in other words, y=(3 + sqrt(5))/2.
B5. Looks to me like a win for player one, but I don't really have an
elegant characterization of her strategy.
If at any time she can arrange it so after her turn there are just two
heaps with the same cardinality, then it suffices thereafter to simply
mirror in the one pile the moves of player one in the other pile. Likewise
if there are two pairs of heaps of equal sizes.
Also, if a player can force there to be a single pile with 4 or 6
beans, then the game is essentially a direct win for that player.
Combining these comments, a player can win if she can leave a pair of
equal heaps, together with a heap of 4 or 6.
Here's how to get to one of those states: Player one should change
(3, 4, 5, 6) to (2, 4, 5, 6). Then here's the response to plays by
player two:
(0, 4, 5, 6)? (0, 4, 5, 5)
(2, 3, 5, 6)? (2, 0, 5, 6) (not yet in good state; see below)
(2, 4, 4, 6)? (0, 4, 4, 6)
(2, 4, 5, 5)? (0, 4, 5, 5)
In the second case, player two can respond in several ways; follow with
(0, 0, 5, 6)? (0, 0, 5, 5)
(2, 0, 4, 6)? (0, 0, 4, 6)
(2, 0, 5, 5)? (0, 0, 5, 5)
[B6: There was an error in this solution; see below.]
B6. The first observation is that the union of the sets is to be {1, 2, ...},
meaning that 0 is not to occur; in particular, [alpha] <> 0 => alpha >=1.
Now just how many elements are in S(alpha)? More precisely, for any N
let N_alpha=|S(alpha) intersect (0,N)|. The intersection includes
[n alpha] iff n alpha < N, which happens only for n less than N/alpha,
i.e., N_alpha < N/alpha (with a difference at most 1.) But
since the union of the three sets is to be all of {1, 2, ...} we must
have N_alpha + N_beta + N_gamma = N (recall the sets are to be
disjoint), i.e.
N/alpha+N/beta+N/gamma > N
(with a difference at most 3).
Now on the one hand, taking N --> oo shows 1/alpha + 1/beta + 1/gamma = 1.
On the other hand, this in turn means
N/alpha + N/beta + N/gamma = N;
which contradicts the previous inequality.
==============================================================================
III. Comments and feedback. For the most part these are posts by other
people taken from sci.math postings.
First let me mention that Dan Bernstein posted a complete set of solutions
(in TeX) shortly after the exam. The answers seem to be essentially the
same as mine except that he carried A-6 and B-6 properly to completion.
I'll include those solutions below. The attribution is
From: djb@silverton.berkeley.edu (D. J. Bernstein)
Date: 4 Dec 1995 21:54:25 GMT
Newsgroups: sci.math
Subject: 00A07 1995 Putnam problems and unofficial solutions (TeX)
Posts from everybody are segregated by problem.
A2============================================================================
From: baloglou@panix.com (George Baloglou)
Newsgroups: sci.math
Subject: Re: Putnam spoilers -- first go-round (A2)
Date: 5 Dec 1995 03:51:38 -0500
When a = b, an obvious substitution (u = x+a) shows the integral to be
equal to
- INT[a 2a]{(x^.5 - (x-a)^.5)^.5} - lim INT[M M+a]{(x^.5 - (x - a)^.5)^.5}
M-->+oo
Our integral is equal to the first (certainly finite) term: indeed, the
second term is equal to zero, as a limit of integrals of a zero-approaching
function over a constant length interval that "drifts" toward infinity.
(This solution was shown to me by a colleague of mine.)
When a =/ b, let us first observe that (x+a)^.5 + (x-b)^.5 - 2(x)^.5
is positive for a > b and negative for a < b (the proof is routine), while,
for large x, 1/(((x+a)^.5 - x^.5)^.5 + (x^.5 - (x-b)^.5)^.5) > 1 ; that
is (radical manipulation), for sufficiently large x the integrand is either
(case a > b) greater than (positive) (x+a)^.5 + (x-b)^.5 - 2(x)^.5 or
(case a < b) smaller than (negative) (x+a)^.5 + (x-b)^.5 - 2(x)^.5 .
The antiderivative of (x+a)^.5 + (x-b)^.5 - 2(x)^.5 can be written as
(2/3)*(((1+(a/x))^1.5 + (1-(b/x))^1.5 - 2)/(x^(-1.5))); l' Hopital's rule
shows that the limit of the second factor as x approaches +oo is equal to
the limit of a*((x+a)^.5) - b*((x-b)^.5) , easily seen to be +oo for a > b
and -oo for a < b. Putting everything together, we can now conclude that
our integral diverges to +oo for a > b but diverges to -oo for a < b.
I tried hard to bound -(((x+a)^.5 - x^.5)^.5 - (x^.5 - (x-a)^.5)^.5)
by a converging integrand but did not succeed; that is, and save for the
power series approach, the substitution trick *seems* to be the only way
to handle the case a = b. A very fine Calculus problem, indeed!
A2============================================================================
Newsgroups: sci.math
From: rjchapma@exeter.ac.uk (R.J.Chapman)
Subject: Re: 1995 Putnam problems and unofficial solutions (TeX)
Date: Wed, 6 Dec 1995 09:51:01 GMT
A2: As x->infinity we have
sqrt(x+a) - sqrt(x)
= sqrt(x)(-1+sqrt(1+a/x))
= 1/(2a sqrt(x)) ( 1 + O(1/x))
and so
sqrt(sqrt(x+a) - sqrt(x)) = x^(-1/4)/(sqrt(2a))(1 + O(1/x)).
Similiarly
sqrt(sqrt(x) - sqrt(x-b)) = x^(-1/4)/(sqrt(2b))(1 + O(1/x)).
and so the integrand equals c x^(-1/4) + O(x^(-5/4)), and the
integral converges iff c = 1/sqrt(2a) - 1/sqrt(2b) = 0 i.e., iff a = b.
A4=============================================================================
From: mvs@unlinfo.unl.edu (Mark V. Sapir)
Newsgroups: sci.math
Subject: Re: Putnam exam: list of questions
Date: 3 Dec 1995 18:41:06 GMT
By the way, problem A4 about necklace is the well known problem about cars
and gasoline stations which is at least 40 years old. I think that first
time this problem was suggested on one of the Moscow olympiads in the fifties.
One can find this problem in many collections of olympiad problems.
A4============================================================================
From: fardila@athena.mit.edu (Federico Ardila)
Newsgroups: sci.math
Subject: Re: Putnam spoilers -- first go-round
Date: 4 Dec 1995 18:45:07 GMT
Here's my solution to A-4, I thought it was kinda neat.
It'll probably look gross here because I'll have to write it
in equations, but it's much nicer once you draw a picture,
and it is pretty straightforward once we choose where to cut.
First, subtract one to each number on the beads. Then their
sum is now -1, and the condition that a_1+...+a_k<=k-1
now becomes a_1+...+a_k<=-1, i.e., this sum is negative.
Consider all the possible sums of the beads on substrings
of the necklace of any size, i.e., all the possible sums of
numbers in consecutive beads.
Since there is a finite number of such sums, there must be
a largest sum. Let S be the largest substring of the necklace
which gives rise to this largest sum.
Now make the cut at one end of S, so the beads are numbered
a_1,...,a_k,a_(k+1),...,a_n, where a_(k+1),...,a_n are the
beads in S.
Let S(i,j)=a_i+a_(i+1)+...+a_j, where a_(n+1)=a_1, etc...
Then for j=1,2,...,k
S(1,j)=S(k+1,j)-S(k+1,n)
but the substring k+1,k+2,...,j is longer than the substring
k+1,k+2,...,n, so it does not give rise to the maximal sum.
Then S(k+1,j)=0, so S(1,j)<=-1<0
Therefore this labelling satisfies the conditions of the problem.
A4============================================================================
That positive label is $M$, and the label immediately before it is $L$.)
We now construct a smaller necklace,
by merging $L$ and $M$ into $L+M-1$.
The new necklace has exactly $n-1$ beads.
The sum of its labels is exactly $n-2$.
Hence, by induction, there is a way to cut the smaller necklace
so that its labels $y_1,y_2,\ldots,y_{n-1}$
satisfy $\sum_{i=1}^k y_i\le k-1$ for each $k$.
Now the ``same'' cut of the original necklace will work.
In $y_1,y_2,\ldots,y_{n-1}$ we un-merge $L+M-1$ into $L$ and $M$
to obtain $x_1,x_2,\ldots,x_n$;
if the merged label $L+M-1$ appears at position $p$,
we have $x_i=y_i$ for $i
p$. We check the desired condition in several cases. For $k
To: rusin@washington.math.niu.edu
Subject: Putnam A6
Since you seem to be compiling Putnam solutions, and didn't have
a complete A6, I humbly offer my own. (It can be phrased in the same
random walk terms you used--however, instead of worrying about whether
the ratio of the probabilities approaches 1/6, we do some (slightly)
more precise calculations.)
A6. To simplify discussion, we say that the numbers in each column are
-1, 0, and 1. Let P_n be the probability that the 3xn matrix has
row sums 0, 0, 0, and let Q_n be the probability the sums are -1, 0, and 1
in some order. We need to show that, for some n >= 1995, Q_n >= 4 P_n.
Let A_n be the probability the sums are (-2, 0, 2) in some order,
B_n the probability they're (2, -1, -1), and C_n the probability they're
(-2, 1, 1). We note that we can express the probabilities for the 3x(n+1)
matrix recursively in terms of those for those for the 3xn matrix:
for the 3x(n+1) to come out to (0, 0, 0), the 3xn has to come out to
(-1, 0, 1), and the last column has to be placed correctly, so
P_{n+1} = 1/6 Q_n
To have the 3x(n+1) come out to (-1, 0, 1), the row sums for the 3xn must
be one of the five cases listed above (each row sum would have to have
abolute value at most 2). If the 3xn row sums are (0, 0, 0), the 3x(n+1)
sums have to be (-1, 0, 1), while in the other cases we end up with
(-1, 0, 1) for 1 or 2 of the possible values of the last column:
Q_{n+1} = P_n + 1/3 Q_n + 1/6 A_n + 1/3 B_n + 1/3 C_n.
Also, if the row sums for the 3xn come out to (-1, 0, 1), any of the
five cases above can arise depending on the last column, and we get
A_{n+1}, B_{n+1}, C_{n+1} >= 1/6 Q_n = P_{n+1},
so (note that A_1 = B_1 = C_1 = P_1 = 0), for all n >= 1,
Q_{n+1} >= 1/3 Q_n + 11/6 P_n.
Thus, for any n >= 1, if Q_n < 4 P_n, then P_n > 1/4 Q_n, and
Q_{n+1} > 1/3 Q_n + 11/6 (1/4 Q_n) = 19/24 Q_n = 19/4 P_{n+1} > 4 P_{n+1}.
So, either Q_1995 >= 4 P_1995, or Q_1996 >= 4 P_1996, proving the proposition.
A6============================================================================
[Bernstein:]
Let $(x,y,z)$ be the original row sums.
Note that $x+y+z=6n$.
Now $a=b-1=c-2$ if and only if $(x-2n,y-2n,z-2n)$ is in $S$,
where
$S=\setof{(0,1,-1),(0,-1,1),(1,0,-1),(-1,0,1),(1,-1,0),(-1,1,0)}$;
and $a=b=c$ if and only if $(x-2n,y-2n,z-2n)=(0,0,0)$.
For any $(a,b,c)$, write $p_n(a,b,c)$ for the probability that
$(x,y,z)$ equals $(2n+a,2n+b,2n+c)$.
We are asked to show that
$4p_n(0,0,0)\le \sum_{s\in S} p_n(s)$
for some $n\ge 1995$.
In fact, this is true for $n=2001$ or $n=2002$.
Define
$X=\setof{(1,1,-2),(-1,-1,2),(1,-2,1),(-1,2,-1),(-2,1,1),(2,-1,-1)}$,
$Y=\setof{(0,2,-2),(0,-2,2),(2,0,-2),(-2,0,2),(2,-2,0),(-2,2,0)}$,
and $C=\setof{(0,0,0)}$.
Define $S_n=\sum_{s\in S} p_n(s)$;
define $X_n$, $Y_n$, and $C_n$ similarly.
With some work one can verify that
$6C_{m+1}=S_m$;
$6S_{m+1}=6C_m+2S_m+2X_m+Y_m$;
$6X_{m+1}\ge 2S_m+2Y_m$; and
$6Y_{m+1}\ge S_m+2X_m$.
Hence $36C_{m+2}=6C_m+2S_m+2X_m+Y_m$
and $36S_{m+2}\ge 12C_m+15S_m+6X_m+6Y_m$.
If $4C_{m+1}>S_{m+1}$ then
$$4S_m>6C_m+2S_m+2X_m+Y_m$$
so $2S_m>6C_m+2X_m+Y_m$.
If also $4C_{m+2}>S_{m+2}$ then
$$24C_m+8S_m+8X_m+4Y_m>12C_m+15S_m+6X_m+6Y_m$$
so $12C_m+2X_m>7S_m+2Y_m\ge 4S_m>12C_m+4X_m+2Y_m$
so $0>2X_m+2Y_m$, contradiction.
In particular, for $m=2000$, we must have
$4C_{m+1}\le S_{m+1}$ or $4C_{m+2}\le S_{m+1}$
as claimed.
A6=======================================================================
From: rusin@washington.math.niu.edu (Dave Rusin)
Newsgroups: sci.math
Subject: Re: Putnam spoilers [problem B6] (Also a reprise on A6)
Date: 5 Dec 1995 03:08:49 GMT
While we're polishing off the Putnam problems, let me mention that
Sandy Kutin has mailed me the rest of the proof of A-6 (the random
walk). I'm still sort of bothered by the fact that rather a lot of
information about the probability distributions seems to be required.
I was trying to formulate a different completion to the proof; maybe
someone can help me out here.
The goal is to determine accurately the probability distribution
p_n(v) that the walker is at vertex v (in the lattice Z[w], where
w^2-w+1=0) after n steps. We know p_0, of course, and p_1(v)=(1/6)
for the vectors w^i (and is zero otherwise). The distribution p_n
is the n-fold convolution p_1*p_1*...p_1, where (f*g) means
(f*g)(v) = Sum f(v-w)g(w).
OK, well here's my question. Is there an obvious transform f --> T(f)
rather like the Laplace transform (say) which will give
T(f*g) = (T(f)).(T(g)) (the pointwise product)? Unlike the situation with
functions on the real line, we don't have the exp homomorphism to
change sums to products. Well, fine; then change the domain of the
transforms or something (there's no need to have T(f) also defined
on the lattice). Of course my goal is to find such a transform with
a workable inverse T^(-1) (f); for then I need only write
p_n = T^(-1) ( T (p_1*...*p_1) ) = T^(-1) ( T(p_1)^n ). If T is ever
calculable, surely T(p_1) will be, and I ought to be able to raise it
to a power. Then T^(-1) might be difficult to compute in general, but I
only need to get the values of T^(-1) (P_n) (v) for v=0 and v=w, and
Bob's your uncle.
The ideal of a random walk on this lattice is such a natural and pretty
thing that I can't imagine it hasn't been studied before. Any
suggestions?
A6============================================================================
From: jrr@atml.co.uk (John Rickard)
Newsgroups: sci.math
Subject: Re: Putnam spoilers -- first go-round
Date: 5 Dec 1995 12:44:34 GMT
The 3-fold convolution is
1 3 3 1
3 6 6 6 3
3 6 15 15 6 3
1 6 15 12 15 6 1
3 6 15 15 6 3
3 6 6 6 3
1 3 3 1
and already each number is at most 1/4 the sum of the surrounding 6
numbers; it follows that the conclusion holds for all n >= 3 (indeed,
the first n - 3 columns can be chosen however you like, and only the
last 3 need be random as specified).
(I suppose this would need fleshing out with some more explanation to
be an adequate solution.)
A6============================================================================
From: Greg Lawler (duke)
Date: Tue, 5 Dec 95 08:30:44 EST
To: rusin@washington.math.niu.edu
Subject: Re: Putnam spoilers [problem B6] (Also a reprise on A6)
Problem 6A is a direct consequence of a theorem in probability theory.
Instead of using 1,2,3 as the numbers use -1,0,1 (so the average is zero).
Consider the random walk (X_n,Y_n) where X_n denotes the sum of the
first row and Y_n denotes the sum of the second row (we have not rearranged
the rows yet).
Then the two probabilities that we have to compare are:
a) the probability that after n steps we are at (0,0)
b) the probability that after n steps we are at (1,0), (1,-1), (0,1),
(0,-1), (-1,0), (-1,-1).
By the local central limit theorem for two dimensionsal random walks,
the probabilty
of a) is asymptotic to c/n where c is some constant depending only on
the (co)variance of the random walk. The probability of b) is asymptotic to
6c/n for the same constant c. Hence
P(a)/P(b)
approaches 1/6 as n goes to infinity.
A6============================================================================
From: Greg Lawler