LXXIX.3 (1997)
Small zeros of quadratic forms over algebraic function fields
by
Albrecht Pfister (Mainz)
Dedicated to J. W. S. Cassels on his 75th birthday
1. Introduction. About 40 years ago J. W. S. Cassels [1] proved the following theorem:
Let q(x) = P
1≤i≤j≤n
q
ijx
ix
jbe a quadratic form with integer co- efficients q
ij∈ Z. Assume that q is isotropic over Q. Then there is an 0 6= a = (a
1, . . . , a
n) ∈ Z
nwith q(a) = 0 such that
(0) |a| = max |a
i| ≤ c
nQ
(n−1)/2where c
nis a constant which depends only on n and where Q = P
i,j
|q
ij|.
In the addendum to [1] he showed that the exponent (n − 1)/2 is best possible by giving an example which was found by M. Kneser. In his book [3, 6.8] he gave a better proof and estimate, in particular one can take c
n= 3
(n−1)/2.
For a generalization to totally isotropic subspaces of higher dimension see [8].
Several years later Cassels [2] published his theorem on the representa- tion of a polynomial f (x) in one variable as a sum of n squares: If such a representation is possible over the rational function field k(x) then it is already possible over the polynomial ring k[x]. He remarked that the under- lying geometrical idea for the proof was essentially the same as in his first paper [1]. This idea is as follows:
Given a 6= 0 with q(a) = 0, intersect the “quadric” q = 0 with a “line”
l = 0 passing through a. If l is chosen carefully then the second intersection point of q and l may be “smaller” than the original point a. For the choice of l one has to use the fact that Z resp. k[x] are euclidean domains.
These results have been generalized in the following directions:
[221]
(1) In 1965 I generalized Cassels’ representation theorem (= Darstel- lungssatz) to an arbitrary quadratic form q over k instead of the “unit form”
q = h1, . . . , 1i. See [4, Ch. 1] for a proof which includes the case char k = 2.
(2) In 1975 Raghavan [6] generalized Cassels’ zero theorem (= Nullstel- lensatz) to the ring of integers in an algebraic number field K. His estimate for |a| is of the same shape as in (0) but the constant c
nnow depends on n = dim q and the degree and the discriminant of K. Except for the precise value of the constant c
nthis paper essentially finishes the number theoretic case.
(3) In 1987 Prestel [5] stated and proved the zero theorem for a rational function field k(x). It reads as follows: There exists 0 6= a = (a
1, . . . , a
n) with a
i∈ k[x] and q(a) = 0 such that
deg a ≤ n − 1 2 deg Q where deg a = max
i(deg a
i), deg Q = max
i,j(deg q
ij).
This may be considered as an additive version of (0) with c
n= 1 for all n. The strengthening of the estimate is due to the fact that the valuation on k(x) which is induced by the degree is non-archimedean (= ultrametric).
In the same paper Prestel constructs an example of (a sequence of) isotropic quadratic forms over R(x, y) in n = 4 variables with coefficients q
ijof degree 2 such that the minimal degree of a non-trivial solution a is unbounded. This proves that one cannot expect results about “small” zeros for function fields in more than one variable.
The aim of the present paper is to prove the Nullstellensatz (Theorem 1) and the Darstellungssatz (Theorem 4) in the remaining open case where K is an algebraic function field in one variable over an arbitrary field k. The main difficulty is to find an argument which replaces the euclidean algorithm for k[x]. In my proofs this will be the theorem of Riemann–Roch.
I found the main breakthrough three years ago when I spent my sab- batical in Cambridge and enjoyed the privilege of being a Visiting Fellow Commoner of Trinity College.
I have been informed that Dorothea Diers (M¨ unster) has obtained very similar results in her thesis under W. Scharlau but I have not seen any details of her work.
Notation. k is an arbitrary field, K/k is an algebraic function field in one variable. As usual we assume that K/k is finitely generated and that k is algebraically closed in K. Divisors of K/k are denoted by latin capitals A, B, . . . , prime divisors by P . Ω denotes the set of all places P of K/k, S = {P
1, . . . , P
s} is an arbitrary non-empty finite subset of Ω. We write A = P
P ∈Ω
v
P(A)P . For P ∈ Ω the valuation ring of P , residue field of P
and residue degree of P are denoted by R
P, k
Pand f
Prespectively. v
P: K → Z∪∞ denotes the normalized discrete valuation of K/k corresponding to P, π = π
Pis a prime element for P . It is well known that R := R(S) :=
T
P 6∈S
R
Pis a Dedekind ring in K with quot(R) = K. For abbreviation we write R
σ, k
σ, f
σetc. instead of R
Pσ, k
Pσ, f
Pσ(σ = 1, . . . , s). Finally, let f = f (S) = max{f
1, . . . , f
s} and let g be the genus of K/k.
We consider quadratic forms q = q(x) =
X
n i,j=1i≤j
q
ijx
ix
j∈ K[x
1, . . . , x
n].
Since char k = 2 is not excluded q is not supposed to be in diagonal form nor to be non-degenerate. The only standard assumptions about q are:
q 6= 0, dim q = n ≥ 1.
In Section 2, q is an isotropic form over K. This implies n ≥ 2 since q = q
11x
21with q
116= 0 cannot be isotropic. The symmetric bilinear form corresponding to q is given by
q(x, y) := q(x + y) − q(x) − q(y) = X
i≤j
q
ij(x
iy
j+ x
jy
i).
We look for isotropic vectors 0 6= a = (a
1, . . . , a
n) ∈ R
nof q. The pole divisor A of a is the smallest non-negative divisor A ≥ 0 such that (a
i) + A ≥ 0 for i = 1, . . . , n. For a
i∈ R we see that A is a linear combination of the prime divisors P
1, . . . , P
s∈ S with coefficients from N
0: A = P
sσ=1
v
σ(A)P
σ. Similarly Q denotes the pole divisor of q, i.e. the smallest non-negative divisor such that
(q
ij) + Q ≥ 0 (1 ≤ i ≤ j ≤ n).
S and q are called compatible if all q
ij∈ R = R(S).
In Section 3 we use a slightly different notation. Here q is a form over k such that the extended form q ⊗ K represents a given element t ∈ K
∗(over K). We then work with the (n + 1)-dimensional isotropic form q ⊗ K ⊥ h−ti over K.
2. The Nullstellensatz. With the terminology introduced in Section 1 let q be an isotropic quadratic form over K.
Definition. A vector a = (a
1, . . . , a
n) ∈ R
nwith pole divisor A is called a minimal vector of q over R if
(i) 0 6= a ∈ R
n, (ii) q(a) = 0,
(iii) deg A ∈ N
0is minimal (under conditions (i), (ii)).
Clearly every isotropic form q has at least one minimal vector a over R = R(S) and then deg A depends only on K/k, q and S but not on a.
The first main result of this article is the following:
Theorem 1 (Nullstellensatz). For every isotropic quadratic form 0 6= q = q(x) = X
1≤i≤j≤n
q
ijx
ix
j∈ R[x
1, . . . , x
n]
and every minimal vector 0 6= a = (a
1, . . . , a
n) ∈ R
nof q with pole divisor A we have
deg A ≤ n(f + g − 1) + n − 1 2 deg Q.
P r o o f. 1. First we consider the case n = 2. Here we have q(x) = q
11x
21+ q
12x
1x
2+ q
22x
22.
Without loss of generality we assume q
11q
226= 0. [Otherwise (1, 0) or (0, 1) is an isotropic vector with pole divisor 0.] Then
q(x) = q
11(x
1− cx
2)(x
1− c
0x
2) with c, c
0∈ K since q is isotropic. We have
q
11c
2+ q
12c + q
22= 0, c + c
0= −q
12/q
11, cc
0= q
22/q
11. By the definition of Q we have
(q
ij) = Q
ij− Q
with non-negative divisors Q
ij≥ 0 (provided q
ij6= 0) (1 ≤ i ≤ j ≤ n). This implies v
P(q
22) − v
P(q
11) = v
P(Q
22) − v
P(Q
11) for every prime divisor P . Consider the principal divisor
(c
2) = 2(c) = 2 X
P
v
P(c)P and define γ
P∈ Z by
2v
P(c) = v
P(Q
22) − v
P(Q
11) + γ
P. Then (cc
0)
2= (q
22/q
11)
2implies
2v
P(c
0) = v
P(Q
22) − v
P(Q
11) − γ
Pfor all P .
Let C, C
0be the pole divisors of c, c
0. We have
v
P(C) + v
P(C
0) =
12max{v
P(Q
11) − v
P(Q
22) − γ
P, 0}
+
12max{v
P(Q
11) − v
P(Q
22) + γ
P, 0}.
We have to distinguish three cases for P :
(i) v
P(Q
11) − v
P(Q
22) + |γ
P| ≤ 0. Then v
P(C) = v
P(C
0) = 0.
(ii) v
P(Q
11) − v
P(Q
22) − |γ
P| ≤ 0 < v
P(Q
11) − v
P(Q
22) + |γ
P|. Then v
P(c) 6= v
P(c
0), hence
1
2
(v
P(Q
22) − v
P(Q
11) − |γ
P|) = min{v
P(c), v
P(c
0)} = v
P(c + c
0)
= v
P(q
12) − v
P(q
11) = v
P(Q
12) − v
P(Q
11) and
v
P(C) + v
P(C
0) = v
P(Q
11) − v
P(Q
12) > 0.
[Note that this case can only occur if q
126= 0.]
(iii) 0 < v
P(Q
11) − v
P(Q
22) − |γ
P|. Then
v
P(C) + v
P(C
0) =
12(v
P(Q
11) − v
P(Q
22) − γ
P) +
12(v
P(Q
11) − v
P(Q
22) + γ
P)
= v
P(Q
11) − v
P(Q
22) > 0.
With an obvious notation this implies deg C + deg C
0= X
2
f
P(v
P(Q
11) − v
P(Q
12))
+ X
3
f
P(v
P(Q
11) − v
P(Q
22))
≤ X
2+3
f
Pv
P(Q
11) ≤ deg Q
11= deg Q.
Hence we may assume without loss of generality that deg C ≤
12deg Q.
It is now easy to find a “small” isotropic vector a = (a
1, a
2) ∈ R
2by solving the linear equation a
1= ca
2with a
1, a
2∈ R \ {0}, (a
i) + A ≥ 0, A ≥ 0, deg A minimal. We need (a
2) + A − C ≥ 0. Then (a
1) + A = (c) + (a
2) + A = (c) + C + (a
2) + A − C ≥ 0.
For any divisor D let L(D) = {d ∈ K : (d) + D ≥ 0}. By the theorem of Riemann–Roch this is a finite-dimensional k-vector space of dimension
l(D) = deg D + 1 − g + i(D) where i(D) ≥ 0 is the index of speciality of D.
We want: supp A ⊂ S, l(A − C) ≥ 1. Then there is some 0 6= a
2∈ L(A − C) with a
2∈ R, a
1= ca
2∈ R. It is sufficient to find A ≥ 0 with supp A ⊂ S and deg(A − C) ≥ g. Since all multiples of f occur as degrees of divisors which are supported by S we can find such an A with
g ≤ deg(A − C) ≤ g + f − 1.
Then deg A ≤ f + g − 1 + deg C ≤ f + g − 1 +
12deg Q.
N o t e. In the case n = 2 we have shown that
deg A ≤ f + g − 1 +
12deg Q.
This estimate is better than the estimate of the theorem unless f = 1 and g = 0. Another easy estimate is obtained by taking a
2= q
11and a
1= ca
2. Then
a
21+ q
12a
1+ q
11q
22= 0,
v
P(a
1) ≥ min{v
P(q
11), v
P(q
12), v
P(q
22)} ≥ −v
P(Q) for all P ∈ Ω, hence
(a
1) + Q ≥ 0, (a
2) + Q ≥ 0, A ≤ Q, deg A ≤ deg Q.
This estimate is better than the above estimate if
12deg Q < f + g − 1.
2. From now on we assume n ≥ 3. Furthermore, we can assume that either s = 1 or that S is the exact set of poles of q, i.e. v
P(Q) > 0 for all P ∈ S. To see this let S
0= {P ∈ S : v
P(Q) > 0} ⊂ S. Then f
0= max{f
P: P ∈ S
0} ≤ f . If |S
0| ≥ 1 then application of Theorem 1 for S
0instead of S yields an isotropic vector 0 6= a = (a
1, . . . , a
n) ∈ R(S
0)
nwith
deg A ≤ n(f
0+ g − 1) + n − 1
2 deg Q ≤ n(f + g − 1) + n − 1 2 deg Q.
If S
0= ∅, i.e. Q = 0, i.e. q
ij∈ k for all i, j, then S can be replaced by any one-point subset S
00⊂ S.
The proof will be by contradiction. Hence we start with the Hypothesis. 0 6= a ∈ R
nis a minimal vector of q with
deg A > n(f + g − 1) + n − 1 2 deg Q.
Let A = P
sσ=1
v
σ(A)P
σ. Since deg A > 0 we can fix a σ ∈ {1, . . . , s}
with v
σ(A) > 0. Then B := A − P
σ≥ 0 and B < A, deg B = deg A − f
σ≥ deg A − f . By our hypothesis we have
deg A > 3(f + g − 1), deg A ≥ 3f + 3g − 2, deg B ≥ 2f + 3g − 2 > 2g − 2.
Therefore
dim L(B) = deg B + 1 − g ≥ 2f + 2g − 1 ≥ 1.
Put
V := {b = (b
1, . . . , b
n) ∈ K
n: (b
i) + B ≥ 0 for i = 1, . . . , n}.
Then V is a k-vector space, dim V = n(deg B+1−g) ≥ n > 0. For 0 6= b ∈ V we clearly have q(b) 6= 0 since the pole divisor of b is ≤ B < A.
3. The main idea of the proof is as follows: Join the point a ∈ R
non the quadric q = 0 to the point b ∈ R
noff the quadric q = 0 by a line l and intersect l with this quadric. This yields a second point of intersection
a
∗:= q(b)a − q(a, b)b ∈ R
n.
For a “good choice” of b ∈ V we can show that the pole divisor A
∗of a
∗satisfies deg A
∗< deg A, which contradicts the minimality of a.
Let us check that a
∗6= 0 and q(a
∗) = 0 for any 0 6= b ∈ V . Since q(b) 6= 0 and q(a) = 0 we know that a, b are linearly independent (over R), hence a
∗6= 0. Further,
q(a
∗) = q(b)
2q(a) + q(a, b)
2q(b) − q(b)q(a, b)q(a, b) = 0.
4. In order to compute v
σ(a
∗j) for σ ∈ {1, . . . , s} choose i(σ) ∈ {1, . . . , n}
such that
v
σ(a
j) ≥ v
σ(a
i(σ)) = −v
σ(A), j = 1, . . . , n.
Put
c
(σ):= b − b
i(σ)a
i(σ)a ∈ K
n. Then
a
∗= q
b
i(σ)a
i(σ)a + c
(σ)a − q(a, c
(σ))
b
i(σ)a
i(σ)a + c
(σ)= b
i(σ)a
i(σ)q(a, c
(σ))a + q(c
(σ))a − b
i(σ)a
i(σ)q(a, c
(σ))a − q(a, c
(σ))c
σ= q(c
(σ))a − q(a, c
(σ))c
(σ). Hence
v
σ(a
∗h) ≥ min
i,j
v
σ(q
ij) + min
j
v
σ(a
j) + 2 min
j
v
σ(c
(σ)j)
≥ −v
σ(Q) − v
σ(A) + 2 min
j
v
σ(c
(σ)j)
for all h = 1, . . . , n. We want to make v
σ(c
(σ)h) as large as possible. A priori we have
v
σ(c
(σ)h) = v
σ(b
ha
i(σ)− b
i(σ)a
h) − v
σ(a
i(σ))
≥ min
j
v
σ(b
j) ≥ −v
σ(B) for all h.
By suitable choice of b ∈ V we want to arrange that (∗) v
σ(c
(σ)j) ≥ −v
σ(B) + γ
σfor all j 6= i(σ)
where the numbers γ
σ∈ N
0(σ = 1, . . . , s) are chosen later. [Note that c
(σ)i(σ)= 0, v
σ(c
(σ)i(σ)) = ∞.]
Fix for a moment σ, j 6= i(σ) and π = π
σ. In the completion K
σ∼ = k
σ((π)) of K with respect to P
σthe element c
(σ)jhas a Laurent series
c
(σ)j= π
−vσ(B)(c
0+ c
1π + . . .) with c
0, c
1, . . . ∈ k
σ.
(∗) is fulfilled for c
(σ)jiff c
0= c
1= . . . = c
γσ−1= 0. Let W :=
M
s σ=1M
j6=i(σ)
W
σ,jwith W
σ,j:= k
γσσ. Then W is a k-vector space of dimension dim W = (n − 1) P
sσ=1
f
σγ
σ. The map
α : V → W, b 7→ M
σ
M
j
(c
0, . . . , c
γσ−1)
σ,j, is clearly k-linear.
(∗) is fulfilled for all σ and j iff 0 6= b ∈ ker α. Therefore we impose the condition dim V > dim W. This gives some upper bound for P
f
σγ
σ. 5. The above estimate for v
σ(a
∗h) leads to a useful estimate for v
σ(A
∗) only in the case
−v
σ(Q) − v
σ(A) + 2(−v
σ(B) + γ
σ) ≤ 0.
Therefore we impose the following conditions on the numbers γ
σ∈ N
0: (1) γ
σ≤ γ
σ∗:= v
σ(B) +
v
σ(Q) + v
σ(A) 2
for each σ = 1, . . . , s, (2) P
sσ=1
f
σγ
σ< n
n − 1 (deg B + 1 − g), i.e. dim W < dim V.
We compute P f
σγ
σ∗:
(i) For s ≥ 2 we have v
σ(Q) ≥ 1 for all σ by our a priori assumption.
This gives
γ
σ∗≥ v
σ(B) +
12v
σ(A),
X f
σγ
σ∗≥ deg B +
12deg A ≥ deg B +
12(deg B + 1)
=
32(deg B + 1) − 1 ≥ n
n − 1 (deg B + 1 − g) − 1
= 1
n − 1 dim V − 1 since n ≥ 3.
(ii) For s = 1 we have A = mP with m > 0, B = (m − 1)P . Then γ
1∗≥ m − 1 +
m 2
≥ m − 1 + m − 1
2 =
32(m − 1), f
1γ
1∗= f γ
1∗≥
32f (m − 1) =
32deg B.
Hence again (for n ≥ 3) 1
n − 1 dim V − 1 ≤
12dim V − 1 =
12(deg B + 1 − g) − 1 ≤
32deg B ≤ f
1γ
∗1since −
12(1 + g) ≤ 0 ≤ deg B.
Since P
f
σγ
σ∗∈ N
0this shows that P
f
σγ
σ∗is greater than or equal to the largest integer below dim V /(n − 1) in both cases.
Let now (γ
1, . . . , γ
s) be any s-tuple with 0 ≤ γ
σ≤ γ
σ∗and P
sσ=1
f
σγ
σ> 0. Replacing a fixed γ
τ> 0 by γ
τ−1 reduces this sum to P
f
σγ
σ−f
τwhere 1 ≤ f
τ≤ f . This shows that for any closed interval I ⊂
0,
n−11dim V of length ≥ f there exists a system (γ
1, . . . , γ
s) ∈ N
s0with γ
σ≤ γ
σ∗for all σ and P
f
σγ
σ∈ I.
6. We are now able to derive an estimate for deg A
∗. From our Hypothesis we have
deg B ≥ deg A − f > n(g − 1) + (n − 1)f + n − 1 2 deg Q,
1
2
deg Q + f < 1
n − 1 (deg B + n(1 − g)),
1
2
deg Q + deg B + f < n
n − 1 (deg B + 1 − g) = 1
n − 1 dim V, say
1
2
deg Q + deg B + f + ε = 1
n − 1 dim V with ε > 0.
Choose I =
12
deg Q + deg B + ε/2,
12deg Q + deg B + f + ε/2 . Then we find a system (γ
1, . . . , γ
s) as above such that P
f
σγ
σ∈ I, i.e.
(∗∗)
12deg Q + deg B < X
f
σγ
σ< 1
n − 1 dim V.
Put C := P
sσ=1
γ
σP
σ. Let W be the k-vector space corresponding to (γ
1, . . . , γ
s). Then dim W = (n − 1) P
f
σγ
σ< dim V.
Choose
0 6= b ∈ ker{α : V → W }, a
∗= q(b)a − q(a, b)b ∈ R
nand let A
∗denote the pole divisor of a
∗. The estimates of part 4 imply:
v
σ(a
∗h) ≥ −v
σ(Q) − v
σ(A) + 2 min
j
v
σ(c
(σ)j)
≥ −v
σ(Q) − v
σ(A) − 2v
σ(B) + 2γ
σ, v
σ(A
∗) ≤ v
σ(Q) + v
σ(A) + 2v
σ(B) − 2γ
σ∈ N
0,
deg A
∗= X
s σ=1f
σv
σ(A
∗) ≤ deg Q + deg A + 2 deg B − 2 deg C < deg A, since deg Q + 2 deg B − 2 deg C < 0 by (∗∗). This contradicts the minimality of the isotropic vector a and proves the theorem.
From Theorem 1 we can easily derive the following more general but
slightly weaker
Theorem 2. Let q 6= 0 be an isotropic quadratic form over K of di- mension n ≥ 2. Let Q be the pole divisor of q. For Q 6= 0 let S be the exact support of Q, and f = f (S). For Q = 0 choose S = {P } and f = f (S) = f
Pwith an arbitrary prime divisor P ∈ Ω, e.g. such that f
P= f
0:= min{f
P: P ∈ Ω}. In addition, let D be any divisor of K such that
deg D ≥ n(f + g − 1) + n − 1
2 deg Q + g.
Then there exists a non-trivial vector b = (b
1, . . . , b
n) with q(b) = 0 and (b
i) + D ≥ 0 (i = 1, . . . , n).
P r o o f. Define R = R(S) and choose a minimal vector a ∈ R
nof q with pole divisor A ≥ 0. By Theorem 1 we have
deg A ≤ n(f + g − 1) + n − 1 2 deg Q.
We put b = t · a with t ∈ K
∗and try to choose t such that (b
i) + D ≥ 0 (i = 1, . . . , n). Clearly q(b) = 0. Since (a
i) + A ≥ 0, b
i= ta
iwe need
(a
i) + (t) + D ≥ 0.
This is true if (t) + D ≥ A, i.e. t ∈ L(D − A). So we need dim L(D − A) > 0.
This is certainly the case if
deg D − deg A + 1 − g > 0, deg D ≥ deg A + g.
By our assumption on D this inequality is true.
N o t e. If we write D = D
1− D
2with (disjoint) non-negative divisors D
1, D
2then (b
i) + D
1≥ 0. This means b
i∈ R
1:= R(S
1) where S
1= supp(D
1), 0 6= b ∈ R
1n, q(b) = 0. Here S
1and q are not compatible in general.
Instead of scaling the vector a we can also scale the quadratic form q without changing the equation q(a) = 0. Thereby we can derive
Theorem 1
0. Every isotropic form q admits an isotropic vector a 6= 0 such that
deg A ≤ 3n − 1
2 (f
0+ g − 1) + n − 1 2 deg Q where f
0= min{f
P: P ∈ Ω}.
P r o o f. Let P
0∈ Ω be such that f
P0= f
0. We try to find q
0∈ K
∗such that (q
0q
ij) + mP
0≥ 0 for suitable m ∈ N
0. We need (q
0) + mP
0− Q ≥ 0.
For this it is enough that
g ≤ mf
0− deg Q ≤ g + f
0− 1, or m =
deg Q + f
0+ g − 1 f
0.
We apply Theorem 1 to the isotropic form q
0= q
0q with pole divisor Q
0≤ mP
0. This leads to an isotropic vector a ∈ R
n0(where R
0= R({P
0})) such that
deg A ≤ n(f
0+ g − 1) + n − 1
2 deg Q
0≤ n(f
0+ g − 1) + n − 1 2 mf
0≤ 3n − 1
2 (f
0+ g − 1) + n − 1 2 deg Q.
N o t e. Depending on the special values of f
0, f, g and n, Theorem 1
0or Theorem 1 may give a “smaller” isotropic vector a for q.
Example 1. For the rational function field K = k(x) and the set S = {∞} with v
∞(u) = − deg u for all u ∈ K we have R = k[x], f = 1, g = 0.
If then q
ij∈ R (1 ≤ i ≤ j ≤ n), Theorem 1 coincides with the theorem of Prestel [5]. The estimate is then best possible for all n.
Example 2. Let char k 6= 2 and let q = h1, q
2, q
3i be an anisotropic ternary quadratic form over k. Consider for g ∈ N
0the function field
K = k(t, u) with − u
2= q
2t
2g+2+ q
3.
It has genus g. All prime divisors P of K have even degree, since otherwise q ⊗k
Pwould be isotropic over the odd-degree extension k
P/k, which contra- dicts Springer’s theorem. There is one place P = ∞ over the infinite place of k(t), it has k
∞= k( √
−q
2), f
∞= 2. For S = {∞} we get R = R(S) = k[t, u]
and Q = 0 since q is a “constant” form. The vector a = (u, t
g+1, 1) is a min- imal vector of q over R with deg A = 2(g + 1). The minimality follows since u must occur in at least one component of every isotropic vector a ∈ R
3and since u has pole divisor U = (g + 1)∞ with deg U = 2(g + 1). The estimate of Theorem 1 gives the weaker estimate
deg A ≤ n(f + g − 1) = 3(g + 1).
[For g = 0 Theorem 1 is essentially sharp since deg A ≤ 3 automatically implies deg A ≤ 2.]
Example 3. Let k = Q
pythbe the pythagorean closure of Q. Then k is real with Pythagoras number p(k) = 1. It is known (see e.g. [4, Ch. 7]) that p(k(t)) ≥ 3 and that there exists a polynomial h = h(t) of degree 4 which is a sum of 3 but not of 2 squares in k[t]. Put K = k(t, u) with u
2= −h. Then K is non-real. In particular, all prime divisors of K/k have even degree, the place ∞ has degree f
∞= 2, t has pole divisor ∞, u has pole divisor 2∞.
Let S = {∞} and R = [t, u] as above. Then f = 2 and g = 1 (since h has degree 4).
We take the constant ternary quadratic form q = h1, 1, 1i with pole
divisor Q = 0. From the equation u
2+ h = 0 it follows that the form
h1i ⊕ q = 4 × h1i is isotropic over K. As is well known this implies that
q ⊗ K is isotropic as well. Let a = (a
1, a
2, a
3) be a minimal vector of q over R. Then a
i= b
i+ c
iu (i = 1, 2, 3) with b
i, c
i∈ k[t], and a
ihas pole divisor m
i· ∞ with
m
i= max{deg b
i, deg c
i+ 2}
where deg is the ordinary degree on k[t]. Theorem 1 gives the estimate deg A ≤ 3(f + g − 1) = 6 for the pole divisor A of a. We want to show that this estimate is sharp (recall that deg A is even):
Assume deg A ≤ 4. Then c
1, c
2, c
3must be constants from k (not all zero). The equation
0 = X
3 i=1(b
i+ c
iu)
2= X
b
2i− h X
c
2i+ 2u X b
ic
iimplies:
1) P
b
ic
i= 0, i.e. b
1, b
2, b
3are linearly dependent over k, say b
3= λb
1+ µb
2with λ, µ ∈ k,
2) P
3i=1
b
2i= h P
3i=1
c
2i= h · c
2with 0 6= c ∈ k.
By normalizing we can arrange c = 1. Then
h = b
21+ b
22+ (λb
1+ µb
2)
2in k[t].
Put
τ := p
λ
2+ µ
2+ 1 ∈ k, ν = λ
2µ + τ
λ
2+ 1 , ω = λ(µ − τ ) λ
2+ 1 . An easy computation yields
h = (b
1+ ωb
2)
2+ (λb
1+ νb
2)
2, i.e. h is a sum of 2 squares in k[t]: Contradiction.
F i n a l r e m a r k s t o T h e o r e m s 1, 1
0, 2
1. Since the estimates for deg A grow with the number n = dim q they can only be sharp if all proper subforms of q are anisotropic over K. For certain fields this ensures an a priori upper bound for n. If e.g. K is a C
i-field then every quadratic form of dimension > 2
iover K is isotropic, which means that we can suppose n ≤ 2
i+ 1.
2. If g 6= 1 then there is a well-known upper bound for the greatest common divisor δ of all possible divisor degrees (δ is called the index of K/k) and for the minimal prime divisor degree f
0, namely δ ≤ f
0≤ |2g − 2|.
Then f
0or f = f (S) could be eliminated from the estimates in many cases.
3. If g = 1 then f
0= δ and δ can be any natural number. For concrete examples where k is a p-adic field see e.g. [7, Cor. 13a].
3. The Darstellungssatz. Let now q 6= 0 be a quadratic form over k
and suppose that q ⊗ K represents a given element t ∈ K
∗. We look for a
representation
(1) q(u) = q(u
1, . . . , u
n) = t
for which the pole divisor U of u has small degree. The result will be slightly different depending on whether q ⊗ K is isotropic or not. For the anisotropic case we first need a variant of Theorem 1. Here a non-trivial solution in K of the equation
(2) q(a
1, . . . , a
n) − ta
20= 0
automatically satisfies a
06= 0, q(a
1, . . . , a
n) 6= 0, and leads to a solution of (1) by putting u
i= a
i/a
0. Since then a
0is (up to sign) completely deter- mined by q, t and the vector a = (a
1, . . . , a
n) we measure the “smallness” of a solution of (2) be the degree of the pole divisor A of a (as in Section 2).
Let N ≥ 0 be the divisor of zeros of t, let T ≥ 0 be the pole divisor of t, i.e. (t) = N −T . We have unique decompositions N = 2N
0+N
1, T = 2T
0+T
1with v
P(N
1), v
P(T
1) ∈ {0, 1} for all prime divisors P . Since deg N = deg T we have
2 deg(T
0+ T
1− N
0) = deg T + deg T
1− 2 deg N
0= deg T
1+ deg N − 2 deg N
0= deg T
1+ deg N
1≥ 0.
If supp T 6= ∅ let S = supp T , otherwise let S = {P } be any one-point set from Ω. Put R = R(S) and f = f (S). Call a non-trivial solution (a
1, . . . , a
n, a
0) of (2) resp. its vector a = (a
1, . . . , a
n) minimal if deg A is as small as possible.
Theorem 3. Under the above assumptions there exists a non-trivial so- lution of (2) with a ∈ R
n. If a is minimal then
deg A ≤ (n + 1)(f + g − 1) + deg(T
0+ T
1− N
0).
P r o o f. The first statement is clear since q ⊗ K represents t and K = quot(R). Suppose now that a is minimal but deg A does not satisfy the estimate of Theorem 3. Then A = P
sσ=1
v
σ(A) P
σwith P
σ∈ S, s = |S|, and v
σ0(A) > 0 for at least one σ
0since deg A > 0 by our assumption. Fix σ
0and put B := A−P
σ0. Then 0 ≤ B < A and deg B = deg A−f
0≥ deg A−f . As in the proof of Theorem 1 we shall construct a vector 0 6= a
∗∈ R
nand an element a
∗06= 0 such that q(a
∗) = ta
∗20but deg A
∗< deg A. This contradicts the minimality of a and proves the theorem.
From (2) we have
(a
i) + A ≥ 0 for i = 1, . . . , n,
hence
(t) + 2(a
0) + 2A ≥ 0, N
1+ 2N
0− 2T
0+ 2(a
0) + 2A ≥ 0, 2(N
0− T
0+ (a
0) + A) ≥ 0, and finally,
(3) (a
0) + A + N
0− T
0≥ 0.
We look for vectors b = (b
1, . . . , b
n) ∈ K
nand elements b
0∈ K such that (4) (b
i) + B ≥ 0 for i = 1, . . . , n,
(b
0) + B + N
0− T
0− T
1≥ 0.
The k-vector space V of all (n + 1)-tuples (b
1, . . . , b
n, b
0) satisfying (4) has dimension
(5) dim V ≥ (n + 1)(deg B + 1 − g) + deg(N
0− T
0− T
1).
From (2) and (4) we conclude
(ta
20) + 2A ≥ 0, (tb
20) + T − N + 2B + 2(N
0− T
0− T
1) ≥ 0, i.e. (tb
20) + 2B − T
1− N
1≥ 0, hence
(6) (tb
20) + 2B ≥ 0 and (ta
0b
0) + A + B ≥ 0.
For all i = 0, . . . , n define
a
∗i: = (q(b) − tb
20)a
i− (q(a, b) − 2ta
0b
0)b
i, a
∗: = (a
∗1, . . . , a
∗n), A
∗:= pole divisor of a
∗. It follows that
q(a
∗) = ta
∗20, (7)
(a
∗i) + A + 2B ≥ 0 for i = 1, . . . , n, (a
∗0) + A + 2B + N
0− T
0≥ 0.
(8)
In particular, we have a
∗∈ R
n.
Since B < A we cannot have q(b) − tb
20= 0 unless b = 0 and b
0= 0.
If 0 6= (b
1, . . . , b
n, b
0) ∈ V this implies that (b, b
0) is independent of (a, a
0) over K, hence a
∗6= 0. We have to choose (b, b
0) in such a way that deg A
∗<
deg A. As in the proof of Theorem 1 we choose for each σ ∈ {1, . . . , s} an index i(σ) ∈ {1, . . . , n} such that v
σ(A) = −v
σ(a
i(σ)) and put
c
(σ)j:= b
j− b
i(σ)a
i(σ)a
j(j = 0, . . . , n).
Then we get
a
∗j= (q(c
(σ)) − tc
(σ)20) a
j− (q(a, c
(σ)) − 2ta
0c
(σ)0) c
(σ)jwith c
(σ)i(σ)= 0,
v
σ(c
(σ)j) + v
σ(B) ≥ 0 (j = 1, . . . , n), v
σ(t
0c
(σ)20) + 2v
σ(B) ≥ 0.
By a suitable choice of (b, b
0) ∈ V we want to arrange for stronger inequal- ities
(9) v
σ(c
(σ)j) + v
σ(B) − γ
σ≥ 0 for i(σ) 6= j ∈ {1, . . . , n}, v
σ(tc
(σ)20) + 2v
σ(B) − 2γ
σ≥ 0
where the γ
σ∈ N
0are chosen later. If (9) holds then a
∗satisfies (10) v
σ(a
∗j) + v
σ(A) + 2v
σ(B) − 2γ
σ≥ 0 (j = 1, . . . , n).
For fixed σ condition (9) puts nf
σγ
σk-linear equations on the vector (b, b
0) ∈ V , that is, n P
sσ=1
f
σγ
σequations altogether. We take γ
σ= v
σ(B) +
0 for σ 6= σ
0, 1 for σ = σ
0. Then n P
f
σγ
σ= n(deg B + f
0) < dim V , since dim V − n(deg B + f
0)
≥ (n + 1) deg B + (n + 1)(1 − g) + deg(N
0− T
0− T
1) − n(deg B + f
0)
= deg B − nf
0− (n + 1)(g − 1) − deg(T
0+ T
1− N
0)
≥ deg A − (n + 1)(f + g − 1) − deg(T
0+ T
1− N
0) > 0
by our assumption. Therefore there exists 0 6= (b, b
0) ∈ V with (9). By (10) the pole divisor A
∗of a
∗satisfies v
σ(A
∗) ≤ v
σ(A) for σ 6= σ
0and
v
σ0(A
∗) ≤
v
σ0(A) − 2 for v
σ0(A) ≥ 2, 0 for v
σ0(A) = 1.
In any case we get v
σ0(A
∗) < v
σ0(A), hence deg A
∗< deg A: Contradic- tion.
N o t e s. 1. For a prime divisor P ∈ supp(N
1+T
1) the validity of equation (1), when read over the completion K
Pof K, implies that q ⊗ K
P, hence q ⊗ k
Pmust be isotropic. By Springer’s theorem this implies that deg P = [k
P: k] is even.
2. For t ∈ k
∗, i.e. T = N = 0, Theorem 3 coincides with Theorem 1 for the constant quadratic form q ⊥ h−ti.
Example 4. Let K = k(x) be the rational function field, and let t =
f (x) ∈ k[x] be a squarefree polynomial. If the anisotropic form q (which
remains anisotropic over K) represents t over K then deg f = d = 2m is
even and the pole divisor T of t is given by T = 2T
0with T
0= m · ∞, f =
f
∞= 1, g = 0. Theorem 3 gives us a representation q(a
1, . . . , a
n) = ta
20with
a
i∈ k[x] = R(∞), deg a
i≤ 0 + deg T
0= m (i = 1, . . . , n).
Then 0 6= q(a) ∈ k[x] with deg q(a) ≤ 2m. Since f is squarefree we get 0 6= a
0∈ k[x] with deg a
0≤ 0, i.e. a
0∈ k
∗. This proves the original repre- sentation theorem (for anisotropic form q).
We are now ready to state and prove the representation theorem for an arbitrary algebraic function field K/k with genus g.
Theorem 4 (Darstellungssatz). Let q be a quadratic form over k, dim q = n. Suppose that t ∈ K
∗is represented by the form q ⊗K. Let N = 2N
0+N
1, T = 2T
0+ T
1and S ⊂ Ω, f = f (S) be as before. Then there exists a vector u = (u
1, . . . , u
n) ∈ K
nwith pole divisor U such that q(u) = t and
deg U ≤ (n + 1)(f + g − 1) + deg(T
0+ T
1) +
12deg N
1or
deg U ≤ n(f + g − 1) + deg T.
[In the “reduced” case the first (second) estimate holds if q⊗K is anisotropic (isotropic).]
P r o o f. Since the estimates for deg U grow with n we can assume that t is not represented by e q ⊗ K for any proper k-subform e q of q (otherwise replace q by e q with dim e q < dim q). We then call q reduced with respect to t. In particular, q is non-defective, i.e. q = q
0⊥ q
00with q
0= rad q = hq
1, . . . , q
ri anisotropic over k and q
00regular over k. (See e.g. [4, Ch. 1]; for char k 6= 2 one has q
0= 0, q regular.)
F i r s t c a s e: q ⊗ K anisotropic. Here we can start with a non-trivial solution of q(a
1, . . . , a
n) − ta
20= 0 according to Theorem 3 and put u
i= a
i/a
0(i = 1, . . . , n), u = a/a
0∈ K
n.
A pole P of u
ieither comes from a pole of a
ior from a zero of a
0. In the first case we have P ∈ S. Let U = U
1+ U
2with U
1≥ 0, U
2≥ 0, supp U
1⊂ S, S ∩ supp U
2= ∅. The equation q(a) = ta
20shows:
1) For P 6∈ S we have v
P(a
i) ≥ 0, v
P(q(a)) ≥ 0, v
P(t) ≥ 0 and v
P(q(a)) = v
P(t) + 2v
P(a
0). Hence
v
Pa
ia
0+
12v
P(q(a)) ≥ −v
P(a
0) +
12v
P(q(a)) ≥ 0.
This implies v
P(U
2) ≤
12v
P(q(a)) for every P 6∈ S, i.e.
deg U
2≤
12X
P 6∈S
f
Pv
P(q(a)).
2) For P ∈ S we have v
P(a
i) + v
P(A) ≥ 0, v
P(t) + v
P(T ) ≥ 0 and again 2v
P(a
0) = v
P(q(a)) − v
P(t). Hence
v
Pa
ia
0+ v
P(A) +
12v
P(q(a)) −
12v
P(t) ≥ 0,
where v
P(A) +
12v
P(q(a)) ≥ 0. This implies v
P(U
1) ≤ v
P(A) +
12v
P(q(a)) +
1
2
v
P(T ), hence
deg U
1≤ deg A +
12deg T +
12X
P ∈S
f
Pv
P(q(a)).
Since P
all P
f
Pv
P(q(a)) = deg(q(a)) = 0 the two estimates imply
deg U = deg U
1+ deg U
2≤ deg A +
12deg T = deg A + deg N
0+
12deg N
1. With the estimate of Theorem 3 for deg A this gives
deg U ≤ (n + 1)(f + g − 1) + deg(T
0+ T
1) +
12deg N
1.
S e c o n d c a s e: q ⊗ K isotropic. Here we first show that q
0⊗ K must be anisotropic (if q is reduced). Suppose that q
0⊗ K is isotropic, say q
r= P
r−1i=1
q
ic
2iwith c
i∈ K. Let t = q(u
1, . . . , u
n) = q
0(u
1, . . . , u
r) + q
00(u
r+1, . . . . . . , u
n) be any representation with u
i∈ K. Then
q
0(u
1, . . . , u
r) = X
r i=1q
iu
2i=
r−1
X
i=1
q
i(u
2i+ c
2iu
2r) =
r−1
X
i=1
q
i(u
i+ c
iu
r)
2. Hence t is represented by e q ⊗ K where e q = hq
1, . . . , q
r−1i ⊥ q
00is a proper subform of q (over k): Contradiction.
Let now 0 6= a ∈ R(S)
nbe a solution of q(a) = 0 which satisfies the condition of Theorem 1, i.e. a has pole divisor A with deg A ≤ n(f + g − 1).
[Note that q is constant, i.e. Q = 0.] Assume for a moment that q(a, b) = 0 for all vectors b ∈ k
n. By linearity this would imply q(a, b) = 0 for all b ∈ K
n, i.e. 0 6= a ∈ rad(q ⊗ K) = q
0⊗ K. In other words, q
0(a) = 0, which is a contradiction. Hence we can choose a vector 0 6= b ∈ k
nsuch that a
0:= q(a, b) ∈ K
∗. Clearly (a
0) + A ≥ 0, in particular a
0∈ R = R(S).
We can now find a good representation of t. Put u := b + λa with some λ ∈ K. Then
q(u) = q(b) + λq(a, b) = q(b) + λa
0, i.e.
q(u) = t ⇔ λ = t − q(b)
a
0⇔ u = b + (t − q(b)) · a a
0.
Let C be the pole divisor of c := a/a
0. Since b ∈ k
nand q(b) ∈ k this shows
that the pole divisor U of u satisfies U ≤ T + C. It remains to estimate
deg C. For every P ∈ Ω we have v
P(a
0) + v
P(A) ≥ 0 and v
Pa
ia
0+ v
P(a
0) + v
P(A) ≥ 0 (i = 1, . . . , n), hence v
P(c
i) + (v
P(a
0) + v
P(A)) ≥ 0 for c
i= a
i/a
0, i.e. v
P(C) ≤ v
P(a
0) + v
P(A). This gives
deg C = X
P
f
Pv
P(C) ≤ X
P
f
Pv
P(a
0) + X
P