Solutions to the 74th William Lowell Putnam Mathematical Competition

Saturday, December 7, 2013

Kiran Kedlaya and Lenny Ng

A1 Suppose otherwise. Then each vertex v is a vertex for A4 Let w01 , . . . , w0k be arcs such that: w0j has the same length
five faces, all of which have different labels, and so the as w j ; w01 is the same as w1 ; and w0j+1 is adjacent to
sum of the labels of the five faces incident to v is at least w0j (i.e., the last digit of w0j comes right before the first
0 +1 +2 +3 +4 = 10. Adding this sum over all vertices digit of w0j+1 ). Since w j has length Z(w j ) + N(w j ), the
v gives 3 × 39 = 117, since each face’s label is counted
sum of the lengths of w1 , . . . , wk is k(Z + N), and so the
three times. Since there are 12 vertices, we conclude
concatenation of w01 , . . . , w0k is a string of k(Z + N) con-
that 10 × 12 ≤ 117, contradiction.
secutive digits around the circle. (This string may wrap
Remark: One can also obtain the desired result by around the circle, in which case some of these digits
showing that any collection of five faces must contain may appear more than once in the string.) Break this
two faces that share a vertex; it then follows that each string into k arcs w001 , . . . , w00k each of length Z + N, each
label can appear at most 4 times, and so the sum of all adjacent to the previous one. (Note that if the num-
labels is at least 4(0 + 1 + 2 + 3 + 4) = 40 > 39, contra- ber of digits around the circle is m, then Z + N ≤ m
diction. since Z(w j ) + N(w j ) ≤ m for all j, and thus each of
w001 , . . . , w00k is indeed an arc.)
A2 Suppose to the contrary that f (n) = f (m) with n < m,
and let n · a1 · · · ar , m · b1 · · · bs be perfect squares where We claim that for some j = 1, . . . , k, Z(w00j ) = Z and
n < a1 < · · · < ar , m < b1 < · · · < bs , ar , bs are minimal N(w00j ) = N (where the second equation follows from
and ar = bs . Then (n · a1 · · · ar ) · (m · b1 · · · bs ) is also a the first since Z(w00j ) + N(w00j ) = Z + N). Otherwise,
perfect square. Now eliminate any factor in this product since all of the Z(w00j ) differ by at most 1, either
that appears twice (i.e., if ai = b j for some i, j, then Z(w00j ) ≤ Z − 1 for all j or Z(w00j ) ≥ Z + 1 for all j.
delete ai and b j from this product). The product of what
In either case, |kZ − ∑ j Z(w0j )| = |kZ − ∑ j Z(w00j )| ≥
remains must also be a perfect square, but this is now a
product of distinct integers, the smallest of which is n k. But since w1 = w01 , we have |kZ − ∑ j Z(w0j )| =
and the largest of which is strictly smaller than ar = bs . | ∑kj=1 (Z(w j ) − Z(w0j ))| = | ∑kj=2 (Z(w j ) − Z(w0j ))| ≤
This contradicts the minimality of ar . ∑kj=2 |Z(w j ) − Z(w0j )| ≤ k − 1, contradiction.
Remark: Sequences whose product is a perfect square
occur naturally in the quadratic sieve algorithm for fac- A5 Let A1 , . . . , Am be points in R3 , and let n̂i jk denote a
toring large integers. However, the behavior of the func- unit vector normal to ∆Ai A j Ak (unless Ai , A j , Ak are
tion f (n) seems to be somewhat erratic. Karl Mahlburg collinear, there are two possible choices for n̂i jk ). If n̂ is
points out the upper bound f (n) ≤ 2n for n ≥ 5, which a unit vector in R3 , and Πn̂ is a plane perpendicular to
holds because the interval (n, 2n) contains an integer of n̂, then the area of the orthogonal projection of ∆Ai A j Ak
the form 2m2 . A trivial lower bound is f (n) ≥ n + p onto Πn̂ is Area(∆Ai A j Ak )|n̂i jk · n̂|. Thus if {ai jk } is area
where p is the least prime factor of n. For n = p prime, definite for R2 , then for any n̂,
the bounds agree and we have f (p) = 2p. For more
discussion, see https://oeis.org/A006255. ∑ ai jk Area(∆Ai A j Ak )|n̂i jk · n̂| ≥ 0.
A3 Suppose on the contrary that a0 + a1 y + · · · + an yn is Note that integrating |n̂i jk · n̂| over n̂ ∈ S2 , the unit
nonzero for 0 < y < 1. By the intermediate value theo- sphere in R3 , with respect to the natural measure on
rem, this is only possible if a0 + a1 y + · · · + an yn has the S2 gives a positive number c, which is independent of
same sign for 0 < y < 1; without loss of generality, we n̂i jk since the measure on S2 is rotation-independent.
may assume that a0 + a1 y + · · · + an yn > 0 for 0 < y < 1. Thus integrating the above inequality over n̂ gives
For the given value of x, we then have c ∑ ai jk Area(∆Ai A j Ak ) ≥ 0. It follows that {ai jk } is area
definite for R3 , as desired.
a0 xm + a1 x2m + · · · + an x(n+1)m ≥ 0
Remark: It is not hard to check (e.g., by integration
for m = 0, 1, . . . , with strict inequality for m > 0. Taking in spherical coordinates) that the constant c occurring
the sum over all m is absolutely convergent and hence above is equal to 2π. It follows that for any convex body
valid; this yields C in R3 , the average over n̂ of the area of the projection
of C onto Πn̂ equals 1/4 of the surface area of C.
a0 a1 an
+ 2
+···+ > 0, More generally, let C be a convex body in Rn . For n̂
1−x 1−x 1 − xn+1
a unit vector, let Πn̂ denote the hyperplane through the
a contradiction. origin perpendicular to n̂. Then the average over n̂ of the
 2

volume of the projection of C onto Πn̂ equals a constant Now suppose that f = 1 + ∑Nn=1 an cos(2πnx) ∈ C.
(depending only on n) times the (n − 1)-dimensional When n is an integer, cos(2πn/3) equals 0 if 3|n and
surface area of C. −1/2 otherwise. Thus an cos(2πn/3) = −an /2 for all
Statements of this form inhabit the field of inverse prob- n, and f (1/3) = 1 − ∑Nn=1 (an /2). Since f (1/3) ≥ 0,
lems, in which one attempts to reconstruct information ∑Nn=1 an ≤ 2, whence f (0) = 1 + ∑Nn=1 an ≤ 3.
about a geometric object from low-dimensional sam-
B3 Yes, such numbers must exist. To define them, we make
ples. This field has important applications in imaging
the following observations.
and tomography.
Lemma 1. For any i ∈ {1, . . . , n}, if there exists any S ∈ P
A6 (by Harm Derksen) Consider the generating functions
containing i, then there exist S, T ∈ P such that S is the disjoint
f (x, y) = xa yb , union of T with {i}.
∑
(a,b)∈S
Proof. Let S be an element of P containing i of minimum car-
g(x, y) = ∑ w(a, b)xa yb . dinality. By (ii), there must be a subset T ⊂ S containing P
(a,b)∈Z2 with exactly one fewer element than S. These sets have the
desired form.
Then A(S) is the constant coefficient of the Laurent
polynomial h(x, y) = f (x, y) f (x−1 , y−1 )g(x, y). We may Lemma 2. Suppose S1 , S2 , T1 , T2 ∈ P have the property that
compute this coefficient by averaging over unit circles: for some i ∈ {1, . . . , n}, S1 is the disjoint union of T1 with {i}
Z 2π Z 2π and S2 is the disjoint union of T2 with {i}. Then
(2π)2 A(S) = h(eis , eit ) dt ds
0 0 f (S1 ) − f (T1 ) = f (S2 ) − f (T2 ).
Z 2π Z 2π
2
= f (eis , eit ) g(eis , eit ) dt ds. Proof. By (i) we have
0 0

Consequently, it is enough to check that g(eis , eit ) is a f (T1 ∪ T2 ∪ {i}) = f (S1 ) + f (T2 ) − f (T1 ∩ T2 )
nonnegative real number for all s,t ∈ R. But g(eis , eit ) = f (T1 ∪ T2 ∪ {i}) = f (T1 ) + f (S2 ) − f (T1 ∩ T2 ),
16G(cos s, cost) for
from which the claim follows immediately.
G(z, w) = zw + z2 + w2 − z2 w − zw2 − z2 w2 .
We now define f1 , . . . , fn as follows. If i does not ap-
If z, w ∈ [−1, 1] and zw ≥ 0, then pear in any element of P, we put fi = 0. Otherwise, by
Lemma 1, we can find S, T ∈ P such that S is the disjoint
G(z, w) = zw(1 − zw) + z2 (1 − w) + w2 (1 − z) ≥ 0. union of T with {i}. We then set fi = f (S) − f (T ); by
Lemma 2, this does not depend on the choice of S, T .
If z, w ∈ [−1, 1] and zw ≤ 0, then
To check that f (S) = ∑i∈S fi for S ∈ P, note first that 0/ ∈
2
G(z, w) = (z + w) − zw(1 + z)(1 + w) ≥ 0. P by repeated application of (ii) and that f (0) / = 0 by
hypothesis. This provides the base case for an induction
Hence g(eis , eit ) ≥ 0 as desired. on the cardinality of S; for any nonempty S ∈ P, we may
apply (ii) to find T ⊂ S such that S is the disjoint union
B1 Note that of T and some singleton set { j}. By construction and
the induction hypothesis, we have f (S) = f (T ) + f j =
c(2k + 1)c(2k + 3) = (−1)k c(k)(−1)k+1 c(k + 1) j + ∑i∈T fi = ∑i∈S fi as desired.
= −c(k)c(k + 1)
B4 Write f0 (x) = f (x) − µ( f ) and g0 (x) = g(x) − µ(g),
= −c(2k)c(2k + 2).
so that 01 f0 (x)2 dx = Var( f ), 01 g0 (x)2 dx = Var(g),
R R
R1 R1
It follows that ∑2013 1006 and 0 f0 (x) dx = 0 g0 (x) dx = 0. Now since |g(x)| ≤
n=2 c(n)c(n + 2) = ∑k=1 (c(2k)c(2k +
M(g) for all x, 0 ≤ 01 f0 (x)2 (M(g)2 − g(x)2 ) dx =
R
2) + c(2k + 1)c(2k + 3)) = 0, and so the desired sum is
c(1)c(3) = −1. Var( f )M(g)2 − 01 f0 (x)2 g(x)2 dx, and similarly 0 ≤
R

Var(g)M( f )2 − 01 f (x)2 g0 (x)2 dx. Summing gives

R
Remark: Karl Mahlburg points out the general formula
c(n) = (−1)b0 b1 +b1 b2 +···+bk−1 bk for n having binary rep- Z 1
resentation bk · · · b0 . 2
Var( f )M(g) +Var(g)M( f ) ≥ 2
( f0 (x)2 g(x)2 + f (x)2 g0 (x)2 ) dx.
0
B2 We claim that the maximum value of f (0) is 3. This (1)
is attained for N = 2, a1 = 34 , a2 = 32 : in this case Now
f (x) = 1+ 43 cos(2πx)+ 23 cos(4πx) = 1+ 34 cos(2πx)+ Z 1
2 2 1 2
3 (2 cos (2πx)−1) = 3 (2 cos(2πx)+1) is always non-
( f0 (x)2 g(x)2 + f (x)2 g0 (x)2 ) dx − Var( f g)
0
negative. Z 1 Z 1
= ( f0 (x)2 g(x)2 + f (x)2 g0 (x)2 − ( f (x)g(x) − f (y)g(y) dy)2 ) dx;
0 0
 3

substituting f0 (x) + µ( f ) for f (x) everywhere and the induction hypothesis, we compute the number of
g0 (x) + µ(g) for g(x) everywhere, and using the fact functions which iterate X into {1, . . . , k + 1} but not into
that 01 f0 (x) dx = 01 g0 (x) dx = 0, we can expand and {1, . . . , k} to be
R R

simplify the right hand side of this equation to obtain

n−k
Z 1
2 2 2 2
∑ (n − k − 1) · · · (n − k − ` + 1)(k + `)nn−`−1
( f0 (x) g(x) + f (x) g0 (x) ) dx − Var( f g) `=1
0
Z 1
By rewriting this as a telescoping sum, we get
= f0 (x)2 g0 (x)2 dx
0 n−k
Z 1 Z 1
− 2µ( f )µ(g) f0 (x)g0 (x) dx + ( 2
f0 (x)g0 (x) dx) ∑ (n − k − 1) · · · (n − k − ` + 1)(n)nn−`−1
0 0 `=1
Z 1 n−k
≥ −2µ( f )µ(g) f0 (x)g0 (x) dx. − ∑ (n − k − 1) · · · (n − k − ` + 1)(n − k − `)nn−`−1
0
`=1
Because of (1), it thus suffices to show that n−k−1
= ∑ (n − k − 1) · · · (n − k − `)nn−`−1
Z 1 `=0
2µ( f )µ(g) f0 (x)g0 (x) dx ≤ Var( f )M(g)2 + Var(g)M( f )2 . n−k
0
(2) − ∑ (n − k − 1) · · · (n − k − `)nn−`−1
`=1
Now since (µ(g) f0 (x) − µ( f )g0 (x))2 ≥ 0 for all x, we
have = nn−1 .
Z 1 Z 1
as desired.
2µ( f )µ(g) f0 (x)g0 (x) dx ≤ (µ(g)2 f0 (x)2 + µ( f )2 g0 (x)2 )dx
0 0 Second solution: For T a set, f : T → T a function, and
= Var( f )µ(g)2 + Var(g)µ( f )2 S a subset of T , we define the contraction of f at S as
the function g : {∗} ∪ (T − S) → {∗} ∪ (T − S) given by
≤ Var( f )M(g)2 + Var(g)M( f )2 ,

establishing (2) and completing the proof. ∗
 x=∗
g(x) = ∗ x 6= ∗, f (x) ∈ S
B5 First solution: We assume n ≥ 1 unless otherwise spec- 
 f (x) x 6= ∗, f (x) ∈/ S.
ified. For T a set and S1 , S2 two subsets of T , we say
that a function f : T → T iterates S1 into S2 if for each
Lemma 2. For S ⊆ X of cardinality ` ≥ 0, there are `nn−`−1
x ∈ S1 , there is a j ≥ 0 such that f ( j) (x) ∈ S2 .
functions f : {∗} ∪ X → {∗} ∪ X with f −1 (∗) = {∗} ∪ S which
Lemma 1. Fix k ∈ X. Let f , g : X → X be two functions such iterate X into {∗}.
that f iterates X into {1, . . . , k} and f (x) = g(x) for x ∈ {k +
1, . . . , n}. Then g also iterates X into {1, . . . , k}. Proof. We induct on n. If ` = n then there is nothing to check.
Otherwise, put T = f −1 (S), which must be nonempty. The
Proof. For x ∈ X, by hypothesis there exists a nonnegative in- contraction g of f at {∗}∪S is then a function on {∗}∪(X −S)
teger j such that f ( j) (x) ∈ {1, . . . , k}. Choose the integer j as with f −1 (∗) = {∗} ∪ T which iterates X − S into {∗}. More-
small as possible; then f (i) (x) ∈ {k + 1, . . . , n} for 0 ≤ i < j. over, for given T , each such g arises from `#T functions of the
By induction on i, we have f (i) (x) = g(i) (x) for i = 0, . . . , j, so desired form. Summing over T and invoking the induction
hypothesis, we see that the number of functions f is
in particular g( j) (x) ∈ {1, . . . , k}. This proves the claim.
n−`  
n−` k
We proceed by induction on n−k, the case n−k = 0 be- ∑ k ` · k(n − `)n−`−k−1
ing trivial. For the induction step, we need only confirm k=1
that the number x of functions f : X → X which iter- n−` 
n−`−1 k

ate X into {1, . . . , k + 1} but not into {1, . . . , k} is equal =∑ ` (n − `)n−`−k = `nn−`−1
k − 1
to nn−1 . These are precisely the functions for which k=1
there is a unique cycle C containing only numbers in
as claimed.
{k + 1, . . . , n} and said cycle contains k + 1. Suppose C
has length ` ∈ {1, . . . , n − k}. For a fixed choice
of `, We now count functions f : X → X which iterate X
we may choose the underlying set of C in n−k−1 `−1 ways into {1, . . . , k} as follows. By Lemma 1 of the first
and the cycle structure in (` − 1)! ways. Given C, the solution, this count equals nk times the number of
functions f we want are the ones that act on C as speci- functions with f (1) = · · · = f (k) = 1 which iterate X
fied and iterate X into {1, . . . , k} ∪C. By Lemma 1, the into {1, . . . , k}. For such a function f , put S = {k +
number of such functions is n−` times the total num- 1, . . . , n} ∩ f −1 ({1, . . . , k}) and let g be the contraction
ber of functions that iterate X into {1, . . . , k} ∪ C. By of f at {1, . . . , k}; then g−1 (∗) = ∗ ∪ {S} and g iterates
 4

its domain into ∗. By Lemma 2, for ` = #S, there are zone which have not previously occurred. Con-
`(n − k)n−k−`−1 such functions g. For given S, each sequently, after the first move into the end zone,
such g gives rise to k` functions f with f (1) = · · · = the rest of the game consists of enumerating all
f (k) = 1 which iterate X into {1, . . . , k}. Thus the num- positions in the end zone in some order.
ber of such functions f is – At this point, we may change the rules without
n−k   affecting the outcome by eliminating the rule on
n−k ` repetitions and declaring that the first player to
∑ ` k `(n − k)n−k−`−1 move into the end zone loses (because #Z = n + 1
`=0
n−k   is even).
n−k−1 `
=∑ k (n − k)n−k−`
`=0 ` − 1 To determine who wins in each position, number the
n−k−1 spaces of the board 1, . . . , n from left to right. Define the
= kn .
weight of a position to be the sum of the labels of the
occupied spaces, reduced modulo n + 1. For any given
The desired count is this times nk , or knn−1 as desired.
position outside of the end zone, for each s = 1, . . . , n
Remark: Functions of the sort counted in Lemma 2 can there is a unique move that adds s to the weight: if s is
be identified with rooted trees on the vertex set {∗} ∪ X empty that a move of type 1 there does the job. Other-
with root ∗. Such trees can be counted using Cayley’s wise, s inhabits a block running from i + 1 to j − 1 with
formula, a special case of Kirchoff’s matrix tree theo- i and j empty (or equal to 0 or n+1), so the type 2 move
rem. The matrix tree theorem can also be used to show at i + j − s (which belongs to the same block) does the
directly that the number of rooted forests on n vertices job.
with k fixed roots is knn−k−1 ; the desired count follows
We now verify that a position of weight s outside of
immediately from this formula plus Lemma 1. (One
the end zone is a win for the player to move if and
can also use Prüfer sequences for a more combinatorial
only if s 6= (n + 1)/2. We check this for positions in
interpretation.)
Pi for i = n − 2, . . . , 0 by descending induction. For
B6 We show that the only winning first move for Alice is positions in Pn−2 , the only safe moves are in the ex-
to place a stone in the central space. We start with some tremal blocks; we may thus analyze these positions
terminology. as two-pile Nim with pile sizes equal to the lengths
of the extremal blocks. In particular, a position is a
By a block of stones, we mean a (possibly empty) se- win for the player to move if and only if the extremal
quence of stones occupying consecutive spaces. By the blocks are unequal, in which case the winning move
extremal blocks, we mean the (possibly empty) max- is to equalize the blocks. In other words, a position is
imal blocks adjacent to the left and right ends of the a win for the player to move unless the empty spaces
playing area. are at s and n + 1 − s for some s ∈ {1, . . . , (n − 1)/2},
We refer to a legal move consisting of placing a stone in and indeed these are precisely the positions for which
an empty space as a move of type 1, and any other legal the weight equals (1 + · · · + n) − (n + 1) ≡ (n + 1)/2
move as being of type 2. For i = 0, . . . , n, let Pi be the (mod n + 1). Given the analysis of positions in Pi+1
collection of positions containing i stones. Define the for some i, it is clear that if a position in Pi has weight
end zone as the union Z = Pn−1 ∪ Pn . In this language, s 6= (n+1)/2, there is a winning move of weight t where
we make the following observations. s + t ≡ (n + 1)/2 (mod n), whereas if s = (n + 1)/2
then no move leads to a winning position.
– Any move of type 1 from Pi ends in Pi+1 .
It thus follows that the unique winning move for Al-
– Any move of type 2 from Pn ends in Pn−1 . ice at her first turn is to move at the central space, as
– For i < n, any move of type 2 from Pi ends in Pi ∪ claimed.
Pi+1 . Remark: Despite the existence of a simple description
– At this point, we see that the number of stones of the winning positions, it is nonetheless necessary to
cannot decrease until we reach the end zone. go through the preliminary analysis in order to establish
– For i < n − 1, if we start at a position in Pi where the nature of the end zone and to ensure that the repe-
the extremal blocks have length a, b, then the only tition clause does not affect the availability of moves
possible moves to Pi decrease one of a, b while outside of the end zone. However, it is not strictly nec-
leaving the other unchanged (because they are essary to study Pn−2 separately: none of the positions
separated by at least two empty spaces). In par- in Pn−1 has weight (n + 1)/2, so following the strategy
ticular, no repetition is possible within Pi , so the of forcing the weight to equal (n + 1)/2 cannot force a
number of stones must eventually increase to i+1. first move into the end zone.
– From any position in the end zone, the legal moves Remark: It is easy to see that Alice’s winning strategy
are precisely to the other positions in the end is to ensure that after each of her moves, the stones are
placed symmetrically and the central space is occupied.
 5

However, it is somewhat more complicated to describe – The property of being balanced is invariant under
Bob’s winning strategy without the modular interpreta- left-right symmetry. This will permit some sim-
tion. plification in the following arguments.
Remark: To resolve a mild ambiguity in the problem – Every position in Pn−2 is unbalanced, because
statement, it should be clarified that the initial position a0 < a0 + a1 < a1 + (n + 1).
(with no stones placed) should be treated as having oc- – For a position x ∈ P1 to be balanced, in order
curred previously once the first move has been made. to have f (x,t) ≡ 0 (mod n + 1) for some t, the
This only affects the case n = 1. unique occupied space must be n + 1 − t. We
Remark: For the analogous problem with n even, must then have A(x) − t = 1 + · · · + n − (n + 1) =
David Savitt has conjectured (based on the cases n = 2 (n/2 − 1)(n + 1), so x is balanced if and only if
and n = 4) that Alice has a winning strategy, and her f (x, n/2 − 1) = 0. This occurs if and only if the
possible winning moves at her first turn are to place a occupied space is n/2 or n/2 + 1.
stone in one of the two central spaces. We give a par- – From every balanced position x ∈ Pn−k for k ≥ 3,
tial analysis based on an argument from Art of Problem every move leads to an unbalanced position. To
Solving user gnayijoag, with some clarification from check this, we need only consider moves at or to
Savitt. the left of the balance point at of x. Let y be the
We first revise the endgame analysis from the original result of a move from x. If the move is at at , then
solution. Define the sets Pi and the end zone Z as be-
f (y,t 0 ) ≡ f (x,t) − at 0 (y) = −at 0 (y) (mod n + 1)
fore. The first six observations from the previous solu-
tion remain correct; however, now the number of posi- and the latter is not a nonzero residue because
tions in Z is odd, so the first player to move into Z wins. at 0 ∈ {1, . . . , n}. For a move to the left of at , the va-
That is, every position in Pn−2 is a winning position for cant spaces to the right of at remain at a0 , . . . , at−1
the player to move. Consequently, the positions in Pn−3 and 0 < A(x) − A(y) < at ; consequently,
can be identified with two-player Nim on the extremal
blocks (the subdivision between the two internal blocks f (y,t − 1) = f (x,t − 1) − (A(x) − A(y))
being immaterial).
≥ ( f (x,t) + at − at−1 + (n + 1)) − (at − 1)
This suggests that if we want to introduce a numeri- = n + 2 − at−1 > 0.
cal invariant that detects the difference between win-
ning and losing positions for the player to move, we Meanwhile, either at remains vacant, or at and
must consider a formula that selectively discards some at+1 are filled while some space b in between be-
information about some of the stones. To this end, for comes vacant; in either case, we have f (y,t) <
a position x ∈ Pn−k for k ≥ 2 with vacant spaces at f (x,t) = 0. Since f (y,t) < 0 < f (y,t − 1), y is
a0 > · · · > ak−1 (or a0 (x) > · · · > ak−1 (x) if this needs unbalanced.
to be clarified), define
To complete the analysis, one would need to show that
A(x) = a0 + · · · + ak−1 from every unbalanced position in Pn−k for k ≥ 3, there
f (x,t) = A − at − t(n + 1) (t = 0, . . . , k − 1); is a move to some balanced position; this would then
show that a position in the game is a win for the player
note that f (x, 0) > · · · > f (x, k − 1). We say that x is to move if and only if it is unbalanced, from which the
balanced if f (x,t) = 0 for some (necessarily unique) conjecture of Savitt would follow.
choice of t, in which case we refer to at as the balance
point of x; otherwise, we say that x is unbalanced. This
definition then has the following properties.