Solutions to the Fifty-Sixth William Lowell Putnam Mathematical Competition Saturday, December 2, 1995

Kiran Kedlaya

Choose a vector orthogonal to but not to . Since as , the same is true of ; . In other words, if but that is simply , then must also go to 0. However, it is easy to exhibit a solution which does not go to 0. The sum of the eigenvalues of the matrix , also known as the trace of , being the sum of the diagonal entries of , is nonnegative, so has an eigenvalue with nonnegative real part, and a corresponding eigenvector . Then is a solution that does not go to 0. (If is not real, add this solution to its complex conjugate to get a real solution, which still doesnt go to 0.) Hence one of the , say for all .
( } s } 3  }3 s } w  } ~ s }   } 3  ( }~

Hence the integral were looking at is

h 3 t s 3  B 2 3 a U CATSgb7 ) P rS'   U 3GAB ` 5 3 q ` 5 3 p Ti

, is zero, in which case

Let and be the probability that we are at the origin, or at a particular point adjacent to the origin, respectively. Then . (In fact, is times the sum of the probabilities of being at each neighbor of the origin at time , but these are all .)
w @w w v ( cw w w R 7 @b1

( ) ( "'

, so that A4 Let . These form a cyclic sequence that doesnt change when you rotate the necklace, except that the entire sequence gets translated by a constant. In particular, it makes sense to choose for which is maximum and make that one ; this way for all , which gives , but the right side may be replaced by since the left side is an integer.
e !q z h!q v b7 1 3 P v G P pe 1 z e|& 3 2 s s s {@%42 3 w 3

What wed like to argue is that for large enough , the ratio of the probabilities of being in any two particular states goes to 1. Then in fact, well see that eventually, about six times as many matrices have than . This is a pain to prove, though, and in fact is way more than we actually need.
( F) 1 P ) ( f' 1

e P ft

 e2

t P

e2 

P te

t e gn 

A3 Let and be the numbers and , respectively. We are given that and for . Sum the rst relation over and we , or . get Now add the rst and second relations for any particular value of and we get . But we know is divisible by 7, and 10 is coprime to 7, so .

 1

e gn 

The term is bounded by a constant times , whose integral converges. Thus we only have to decide whether converges. But has divergent integral, so we get convergence if and only if (in which case the integral telescopes anyway).
` 5 ( x q ( w y%yA!q P t P ur ( r bq v 7 1 v G 3 2 s s s x%2 3 p P p s  dAGk AG o P 2 e s 2 2 U n !1 l j P i g 2 dAGm  2  dACk ( he g h h h W%% 1 g h h h 6W% 1 (fe  P bt  dACd 2 U  AG !1 P b s 2 U !1  s h h %h s h h %h w 8U 3 a 7 ) P dx' w  5 vU 3 ` ) ( yI' ` 5 u vU 3

A6 View this as a random walk/Markov process with states the triples of integers with sum 0, corresponding to the difference between the rst, second and third rows with their average (twice the number of columns). Adding a new column adds on a random permutation of the vector . I prefer to identify the triple with the point in the plane, where is a cube root of unity. Then adding a new column corresponds to moving to one of the six neighbors of the current position in a triangular lattice.

}~

}~

and similarly
h $ 3 B 2 3 a U GAdFgb7 ) 2 1 & ` 5 3 ( ) feP 3 H P 3 H X

'

3 B 2 3 a U GAdScb7

'

2 1 R

` 5

P Y'

2 I3

hence
X

}~ h h h c%%@ } ~

h} 3 

}~

}~ 

V w

. The easiest proof A2 The integral converges iff uses big-O notation and the fact that for . (Here means bounded by a constant times .) So
( 5 6 W 3 GAB 3 2 1 4" U 3  B 2 3 9 VCATS@@7 P ' 1 3 7 ) 0' ( 2 R1 1 3 FD E 2 1 R Q  5 5 3 D 3 3 ( 3 3  B 2 9 7 3 2 CAA@81 H P F' 2 I3 H

}~ 

A1 Suppose on the contrary that there exist with and with . Then while , contradiction.
  &%  '    $     (  #"   !   

   

A5 Everyone (presumably) knows that the set of solutions of a system of linear rst-order differential equations with constant coefcients is -dimensional, with basis vectors of the form (i.e. a function times a constant vector), where the are linearly independent. In particular, our solution can be written as .

Now how to compute the eighth root of ? Notice that if satises the quadratic , then we have
1 2 |I3 ( 1 2 3 P ' P h 1|2 3 9  3 ` ' ' P 1 2 |I3 3 G 2 ' dP 3 G 3 ( 3

Clearly, then, the positive square roots of the quadratic satisfy the quadratic . Thus we compute that is the greater root of , is the greater root of , and is the greater root of , otherwise known as . B5 This problem is dumb if you know the SpragueGrundy theory of normal impartial games (see Conway, Berlekamp and Guy, Winning Ways, for details). Ill describe how it applies here. To each position you assign a nim-value as follows. A position with no moves (in which case the person to move has just lost) takes value 0. Any other position is assigned the smallest number not assigned to a valid move from that position. For a single pile, one sees that an empty pile has value 0, a pile of 2 has value 1, a pile of 3 has value 2, a pile of 4 has value 0, a pile of 5 has value 1, and a pile of 6 has value 0. You add piles just like in standard Nim: the nim-value of the composite of two games (where at every turn you pick a game and make a move there) is the base 2 addition without carries (i.e. exclusive OR) of the nimvalues of the constituents. So our starting position, with . piles of 3, 4, 5, 6, has nim-value A position is a win for the player to move if and only if it has a nonzero value, in which case the winning strategy is to always move to a 0 position. (This is always possible from a nonzero position and never from a zero position, which is precisely the condition that denes the set of winning positions.) In this case, the winning move is to reduce the pile of 3 down to 2, and you can easily describe the entire strategy if you so desire. B6 Obviously have to be greater than 1, and no two can both be rational, so without loss of generality assume that and are irrational. Let denote the fractional part of . Then if and only if . In particular, this means that contains elements, and similarly. Hence for every integer ,
1 t3 P P hV v h 7 3 1 2 v G  e ( V3 P 1 2 |uv t 2 h h h hv W%% 1 P 1 7 b1 1  1 2 |uv 3 2 1 2 |v g  & 7  dGh ( v ( 1 TRT9 2 A3 6 5 9 c7 ` 5 ) 6P 5 9 k2  3 H 2 yC 5 ( ( P 12Sc 3 ( P 3 12d3b 3 P 12#3ba 3 ( 1 )eP 3 1 2 4I3

B2 For those who havent taken enough physics, rolling without slipping means that the perimeter of the ellipse and the curve pass at the same rate, so all were saying is that the perimeter of the ellipse equals the length of one period of the sine curve. So set up the integrals:
s  42  Q 6 i

Let

in the second integral and write 1 as and you get

s

Since the left side is increasing as a function of , we have equality if and only if . B3 For we obviously get 45, while for the answer is 0 because it both changes sign (because determinants are alternating) and remains unchanged (by symmetry) when you switch any two rows other than the rst one. So only is left. By the multilinearity of the determinant, the answer is the determinant of the matrix whose rst (resp. second) row is the sum of all possible rst (resp. second) rows. There are 90 rst rows whose sum is the vector , and 100 second rows whose sum is . Thus the answer is
h @ggc9 9 ( gga cG a ( ( a @Sa a gI@a @a cG a 9 ( v P a cSca ( ) v 2 ' |( ) 1 v

B1 For a given , no more than three different values of are possible (four would require one part each of size at least 1,2,3,4, and thats already more than 9 elements). If no such exist, each pair occurs for at most 1 element of , and since there are only possible pairs, each must occur exactly once. In particular, each value of must occur 3 times. However, clearly any given value of occurs times, where is the number of distinct partitions of that size. Thus can occur 3 times only if it equals 1 or 3, but we have three distinct values for which it occurs, contradiction.
G 3 3 G t V 3 $ G  3 C 3 3 G 3 3 G t S 3 G

1 y2

9 @c9

7 1

9 cc9

( y

Suppose on the contrary that this is not the case; then for some constant . However, if , the probability that we chose each of the six types of moves times is already , which by Stirlings approximation is asymptotic to a . This term alone is bigger than constant times , so we must have for some . (In fact, we must have for any .)
mP 7 $ 7 9 1 @ mw cw 7 mw @w 7 v 65 8U w 7 9 @g w @ E 7 9  ( w f v v

w @

( x T

7 b1

9 @g9

cc9 9

h 3 s

7 9 mw ' 7 83

&'

7 w R 2

So the desired result, which is that some large , is equivalent to

@w v

7 @g9

for

B4 The innite continued fraction is dened as the limit of the sequence . Notice that the sequence is strictly decreasing (by induction) and thus indeed has a limit , which satises , or rewriting, . Moreover, we want the greater of the two roots.

7 |R1  2 '  42

8) Q

c ) x ' i

i x X ' 2 Q i 6 ' l ( 2 7 3

' fRP ( x

C 3

t G 3

'

2 7 1

2 A3

'

7b1 @1

v 7

2 A3 Y 1

7 v

Our desired contradiction is equivalent to showing that the left side actually takes the value 1 for some . Since the left side is an integer, it sufces to show that for some . A result in ergodic theory (the two-dimensional version of the Weil equidistribution theorem) states that if are linearly independent over the rationals, then the set
q v 1 v 1 E 7 1 2 v |uC P g 2 & 7 1 2 v uG P g

On the other hand, suppose that such a relation does hold. Since and are irrational, by the one-dimensional Weil theorem, the set of points is dense in the set of in the unit square such that is an integer. It is simple enough to show that this set meets the region unless is an integer, and that would imply that , a quantity between 0 and 1, is an integer. We have our desired contradiction.

'

'

h 9

1 2 |xv

b|2 7 1 1 2 |Av P

b|2 7 1 2

Dividing through by shows that that for all ,

1 2 xv 7 b1 v P

and taking the limit as . That in turn implies

of points is dense (and in fact equidistributed) in the unit square. In particular, our claim deffor some integers initely holds unless .