Inventiones math.

5, 2 3 7 - 242 (1968)

On the Representation of Hermitian Forms

as Sums of Squares
DANIEL G. QUILLEN* (Cambridge, Mass.)

There is an extensive literature (e.g. [1], Ch.VI) concerned with

Hilbert's problem of representing a positive-valued form with real
coefficients as a sum of squares. The corresponding problem for her-
mitian forms does not seem to have been touched. In this paper we
present a theorem about hermitian forms which arose in connection
with a priori estimates for partial differential equations. This theorem is
proved by the methods of a priori estimates and one of its corollaries
is the Hilbert Nullstellensatz over C, so we obtain a direct analytic proof
of the latter theorem.
Let T and V be finite dimensional complex vector spaces and let
E(u 1, u2) be a hermitian form on Sm T | V where S,~ denotes m-th sym-
metric tensors. If ~ e T a n d v e V l e t
E^ (~, v)=im!)-2 E(~m | v, ~ | v).
Then E ^ is a real-valued smooth function on T x V such that

E ^ (s (, t v)= Isl 2m Itl 2 E ^ ((, v) (1)

for s, teC, and the map E~--~E ^ yields a 1-1 correspondence between
hermitian forms on Sm T | Vand smooth functions on T x Vsatisfying (1).
The inverse procedure of passing from a function to a form is polariza-
tion (see formulas (2) below).
Let I(I 2 be a hermitian scalar product on T a n d define the "N-th
prolongation" of E to be the hermitian form E N on S m + N T |
(EN) ^ (~,v)=(N!)- x Ir E ^ (r
Theorem. The following conditions are equivalent:
(i) E ^ (~, v)>O for ~ 4=0 and v4=O.
(ii) For sufficiently large N, E N is positive definite.
We first derive an expression for E N as an integral. By choosing an
orthonormal basis for T we may identify T with the space of C-linear
functions on C ~ in such a way that the orthonormal basis is z 1, ..., z~
where z ~ i s t h e i-th coordinate function. Then S k T may be naturally
identified with the holomorphic polynomial functions homogeneous of
* This research was supported in part by N S F GP-6959.

17 Inventionesmath.,Vol. 5
238 D.G. Quillen:

degree k and Sk Thas the basis z", tetl= k where a is a multi-index. If we

identify V with C' and let e t ..... e, be the canonical basis of C r, then
Sk T | can be naturally identified with the holomorphic polynomial
functions homogeneous of degree k with values in C'. Then S , T |
has the basis z~e~, 10~1= m, 1 < i < r, so there are complex numbers a~ i, pj,
I ~ l = [ f l l = m , l < i , j < r with aaj,~-a~i,a~-
* such that
e(ul, u 2 ) = E ~! [3! aa,,~j ba, c~j
(2)
e ^ (~, ~)=Y a,,.~ ~ v, ~* ~ ~7
if u, = E b,, z~ e,, u~= E c,, z~ e,, r E r z ', and v=(v,, ..., v,), where *
denotes complex conjugate. The matrix a~,aj may be determined from
the Taylor series of E ^ at (0, 0) which shows that E and E ^ determine
each other.
Let E~ be the differential quadratic form
e ~ (f, g ) = E a~i.aJ(D~fi)(Dag~)*
where f and g are holomorphic functions on C" with values in C" and
Di=d/Oz i. If dG=n-"exp(-Izl2)dxtdyl...dx"dy" is the Gaussian
measure on C ~, then we have the following expression for E N.
Lemma 1. If f, geSm+~T| then
EN(f, g)= ~ E ~ (f,, g) dG.
Proof. If u~, u2 are two holomorphic polynomial functions on C", let
(.~, U~)o= ~ ~, u~ ao.
By integrating by parts we obtain the basic formula
(V' ~,, ~:)o = (.,, ~',,~)0,
from which follows
(~, ~"~ - S~ ~* [3
m-[0t! a=/~. (3)
By definition
(E") ^ (r -1 Ir162 E (V!)-'r162162
I~l=tr
= E (7!) -1 a~i.,J r r+~ ~162 r+t3 v*,
hence
(EN)~ (f,g)= E (~!)-1E~tD~f,D~g).
I~l=N
By linearity it suffices to prove the l e m m a when f = z ~ ei and g = z~ei
where 16[=lel=m+N. But
E~(f, g)=(E~)" (f, g)(0)= E (~!)-"~,.,~(D'" z~)(D ~ " z~)*
 On the Representation of Hcrmitian Forms as Sums of Squares 239

and
SE~(f,g)dG = ~. ar D~z~)o.

These two coincide by (3), proving the lemma.

Introduce norms
]lullS= S ~ (7!) -1 JD~uj2 dG
I~l=q
tlult~,-- j (q!)-' Izl 2q lui 2 dG
where u ~ S , T | Vis thought of as a function with values in V a n d Ivj2=
vi v*. Let ( , )q and ( , )~, be the associated hermitian forms. Let
Ckq be the binomial coefficients.
L e m m a 2. I f u ~ S k T | V, then f]ui[2 = Cqk []u]]2.
Proof. By the binomial theorem
C~~ X v= (1 + X) *

where C~* = 6l(~!(6 - 7)!) -~ and X = (Xa . . . . . X.). Substituting X~ . . . . .

X. = t and comparing coefficients of t q we have
X C~a= C~ I'1- (4)
I~1=q
Using (3) and (4) we have
11: e,]l 2 = ~ (~'!)-~ liD ~ z'lloz = Z (7 !)- x (~ !)2 ((~_~)!)-1
l~l=q Ivl=q
=~! C i~I=C~ II:e~ll~,
Combining this with the readily verifiedfact that : e~ and z# ejin Sk T | V
are orthogonal for both ( , }~ and ( , )o unless c~= fl and i=j, we obtain
the lemma.
L e m m a 3. There is a constant C depending only on q such that
II.ll~=<,~ilL~ c Y tlull~
for every u in S . T | V. J<=q
Proof We shall need the c o m m u t a t i o n formulae
(~!fl!)-'z"DP u= Z (- ])Jr('y!(fl-~)!(cx-y)!)-ID"-~'z~-' u
(5)
(o~!ill)-'D" zO u= Z (yt(fl-~)!(~z--y)l)-'z #-~ D "-~ u
where the first may be obtained as the coefficient of s ~ t # in the power
series expansion of the Weyl c o m m u t a t i o n relations
:" e '~ u ( z ) = e - " e'D :" u(z)

i.e. e a ' u ( z + t ) = e - " e*(=+Ou(z+t)

17"
240 D.G. Quillen:

and where the second may be derived similarly (see also [2], p. 153). Thus
Ilul[2. = ~, (a!) -~ ttz~u[12= ~ (a!)-'(u,V~z~U)o
I~1=q I~l=q
= 7` 7 ` ~ ! ( ~ ! ( ~ - # ~ ) - ' ( u , : - ~ D ~ - ~ U ) o
I~1=~
= Y 7` c : ( ( ~ - ~ ) 0 -' ItD'-~utt~
I~l=q
from which the lemma immediately follows.
Proof of the Theorem. The implication (ii)=> (i) is clear, so assume (i)
holds. If ueSm+NT| then by Lemma 1, (2), and (5) we have with
indices running like 1 =<i, j=<r, la[=]fl]=m
EN (U, u)= ~" a~,i,ljj(D" ui, D# U))o
=~, a,i,#1(u,, z~ D# U~)o
= ~ Z ( - 1)J,~a~,,p: ~! ~!(~!(B-~')! (~-~)!)-~ : - ' u,(:-~ u,)* ao
= S [ E^ (z*, u)+ r u)] do
where Q(z,u) is of degree <m in ziz* and of degree 1 in uiu]'. Thus
for some constant C
[Q(z, u)i = C ~ (i!)-' Iz[ zl lul z,
i<m

while by hypothesis on E there is an e > 0 such that

i~ ^ (:,,,)i>_-~(m~)-~ Izl~" luP.

Therefore
EN(U, U)> I (~(m ~)- 'lzl 2m-- C Z (i~)-1 Izl2') lul 2 dO
i<m

=~ Ilull~,- c Z tlull~,
i<:m

->__e Ilull~- c ' Z llulr~

i<m

using Lemma 3. By Lemma 2 we have E~C(u,u)>=P(N) IlUl[o

2 where

P ( N ) = ~ C,,"+N- C' 7` Cii+N

i<rn

is a polynomial of degree m in N with positive leading coefficient. Thus

for N sufficientlylarge P(N)> 0 so E N is positive definite and the theorem
is proved.
We shall now give some applications of the theorem when V= C
and T = C n.
 On the Representation of Hermitian Forms as Sums of Squares 241

Corollary 1. Let
E^(Q = ~ a~t~ ~ ~*t~

be a hermitian form homogeneous of degree m on C" which is strictly

positive, i.e. E ^ (~) > 0 if ~ oe O. Then for sufficiently large N,
Z I (W
i
= y, 2
bl=N
where Pi (~) is a basis for the holomorphic polynomials of degree m + N on C".
Corollary 2. Let
E- (f) = fr =ma~~ D"f(Dt~ f ) *
be a differential quadratic form in n variables homogeneous of degree
(m, m), such that a ~ , - a , ~* . I f E ~ has no non-zero characteristics, i.e.
~a,p('~*#>0 for ( + 0 , then for sufficiently large N we may write
E (Y!)-' E ~ ( D V f ) = E [Pi(D)f[ 2
I~l =N i
where Pi(D) is a basis for the homogeneous polynomials of degree m+ N
in D 1..... D..
These corollaries are just reinterpretations of the implication (i)=~ (ii)
of the theorem.
Example. Let E ~ ( f ) = ]D~ u[ 2 - [ D 1 D 2 u[ 2 + ID2z u[ 2 where n = 2. T h e dif-
ferential quadratic form E - is not a sum of squares because E ~ (z 1 z 2) < 0.
But E~ (0=(l~d 2-1(212) 2 + [(x ~zl 2 > 0 if ( # : 0 so the t h e o r e m applies.
In fact E a - ( f ) = [D~ f[2 + ]D 3 f l z and E N is positive definite for N > 3.
Corollary 3. (Projective Nullstellensatz over C . ) / f Qi(z) 1 < j < l are
homogeneous polynomials in z ~..... z" without a common zero ~ 4:0, then
the ideal generated by the Qj contains all homogeneous polynomials of
sufficiently high degree.
Proof We m a y assume that the Qj have the same degree m. Let
g ^ (~)= ~,
J
IQj(Qt2;
then E ^ (~)>0 if ~4:0 so by Corollary 2 for all sufficiently high N, there
is an expression

E (~'!)-~ ID~'QJ(D)fl2=~ Ie~(o)fl 2

j IyI=N
where P/(D) is a basis for the space W of h o m o g e n e o u s polynomials
of degree m + N . It follows that the set DrQj(D) for I~,I=N, l N j < l
spans W for otherwise we could take f to be a non-zero p o l y n o m i a l in
242 D.G. Quiilen: On the Representation of Hermitian Forms as Sums of Squares

z homogeneous of degree m + N killed by the D r Qj(D) and get a contra-

diction. This proves the corollary.
Corollary 3 implies the usual Nullstellensatz by means of the Rabino-
witz trick and making polynomials homogeneous by adding an extra
variable. Hence we have given a direct analytical proof of the Null-
stellensatz over C bearing some resemblance to the Liouville theorem
proof of the fundamental theorem of algebra.

Bibliography
1. Jacobson, N.: Lectures in abstract algebra III. Princeton: van Nostrand 1964.
2. TrOves, F.: Linear partial differential equations with constant coefficients. New York:
Gordon & Breach 1966.

Dr. D. G. Quiilen
Massachusetts Institute of Technology
Department of Mathematics
Cambridge, Mass. 02139, USA

(Received February 27, 1968,)