THE IMAGES OF MULTILINEAR POLYNOMIALS EVALUATED

ON 3 × 3 MATRICES.
arXiv:1306.4389v2 [math.AG] 30 Jun 2013

ALEXEY KANEL-BELOV, SERGEY MALEV, LOUIS ROWEN

Abstract. Let p be a multilinear polynomial in several noncommuting vari-

ables, with coefficients in a algebraically closed field K of arbitrary charac-
teristic. In this paper we classify the possible images of p evaluated on 3 × 3
matrices. The image is one of the following:
• {0},
• the set of scalar matrices,
• a (Zariski) dense subset of sl3 (K), the matrices of trace 0,
• a dense subset of M3 (K),
• the set of 3−scalar matrices (i.e., matrices having eigenvalues (β, βε, βε2 )
where ε is a cube root of 1), or
• the set of scalars plus 3−scalar matrices.

1. Introduction
This paper is the continuation of [BeMR1], in which we considered the question,
reputedly raised by Kaplansky, of the possible image set Im p of a polynomial p on
matrices.
Conjecture 1. If p is a multilinear polynomial evaluated on the matrix ring Mn (K),
then Im p is either {0}, K (viewed as K the set of scalar matrices), sln (K), or
Mn (K).
Here sln (K) is the set of matrices of trace zero.
This subject was investigated by many authors (see [AlM], [BrK], [Ch], [Ku1],
[Ku2], [LeZh]). For review and basic terminology we refer to our previous pa-
per [BeMR1]. (Connections between images of polynomials on algebras and word
equations are discussed in [BKP]; also see [La], [LaS], [S].)
Recall that a polynomial p (written as a sum of monomials) is called semi-
homogeneous of weighted degree d with (integer) weights (w1 , . . . , wm ) if for each
monomial h of p, taking dj,h to be the degree of xj in h, we have
d1,h w1 + · · · + dn,h wn = d.
A semi-homogeneous polynomial with weights (1, 1, . . . , 1) is called homogeneous of
degree d.
In [BeMR1] we settled Conjecture 1 for n = 2 and classified the possible images
for semi-homogeneous polynomials:

Date: July 22, 2013.

2010 Mathematics Subject Classification. Primary 16R99, 15A24, 17B60; Secondary 16R30.
Key words and phrases. Noncommutative polynomial, image, multilinear, matrices.
This work was supported by the Israel Science Foundation (grant no. 1207/12).
The second named author was supported by an Israeli Ministry of Immigrant Absorbtion
scholarship.
1
2 ALEXEY KANEL-BELOV, SERGEY MALEV, LOUIS ROWEN

Theorem 1. Let p(x1 , . . . , xm ) be a semi-homogeneous polynomial evaluated on

the algebra M2 (K) of 2 × 2 matrices over a quadratically closed field. Then Im p is
either {0}, K, sl2 (K), the set of all non-nilpotent matrices in sl2 (K), or a dense
subset of M2 (K) (with respect to Zariski topology).
A homogeneous polynomial p is called multilinear if dj,h = 1 for each 1 ≤ j ≤ n
and each monomial h of p (and thus d = n).
Examples were given in [BeMR1] of homogeneous (but not multilinear) polyno-
mials whose images do not belong to the classification of Theorem 1.
Our research in this paper continues for the 3 × 3 case, yielding the following:
Theorem 2. If p is a multilinear polynomial evaluated on 3 × 3 matrices then Im p
is one of the following:
• {0},
• the set of scalar matrices,
• sl3 (K), (perhaps lacking the diagonalizable matrices of discriminant 0),
cf. Remark 7.
• a dense subset of M3 (K),
• the set of 3−scalar matrices, or
• the set of scalars plus 3−scalar matrices.

2. Images of Polynomials
For any polynomial p ∈ Khx1 , . . . , xm i, the image of p (in R) is defined as
Im p = {r ∈ R : there exist a1 , . . . , am ∈ R such that p(a1 , . . . , am ) = r}.
Remark 1. Im p is invariant under conjugation, since
ap(x1 , . . . , xm )a−1 = p(ax1 a−1 , ax2 a−1 , . . . , axm a−1 ) ∈ Im p,
for any nonsingular a ∈ Mn (K).
We recall the following lemmas (for arbitrary n) proved in [BeMR1]:
Lemma 1 ([BeMR1, Lemma 4]). If ai are matrix units, then p(a1 , . . . , am ) is
either 0, or c · eij for some i 6= j, or a diagonal matrix.
Lemma 2 ([BeMR1, Lemma 5]). The linear span of Im p is either {0}, K, sln ,
or Mn (K). If Im p is not {0} or the set of scalar matrices, then for any i 6= j the
matrix unit eij belongs to Im p.
Another major tool is Amitsur’s Theorem [Row3, Theorem 3.2.6, p. 176], that
(k)
the algebra of generic n × n matrices (generated by matrices Yk = (ξi,j ) whose
(k)
entries {ξi,j , 1 ≤ i, j ≤ n} are commuting indeterminates) is a non-commutative
domain UD whose ring of fractions with respect to the center is a division algebra
which we denote as UD g of dimension n2 over its center F1 := Cent(UD)).
g

Remark 2. Assume Char (K) = 0. Suppose t is a commuting indeterminate, and

f (x1 , . . . , xm ; t) is a polynomial taking values under matrix substitutions for the xi
and scalars for t. If there exists unique t0 such that f (x1 , . . . , xm ; t0 ) = 0, then
t0 is a rational function with respect to the entries of xi . If this t0 is fixed under
simultaneous conjugation of the matrices x1 , . . . , xm , then t0 is in the center of
Amitsur’s division algebra UD, g implying f ∈ UD. g If Char (K) = k 6= 0, then tkl is
0
a rational function for some l ∈ N0 .
 IMAGES OF MULTILINEAR POLYNOMIALS 3

Remark 3. In Remark 2 we could take a system of polynomial equations and

l
polynomial inequalities. If t0 is unique, then it is a rational function (or tk0 if
Char (K) = k).
In fact, we need a slight modification of Amitsur’s theorem, which is well known.
Viewing  
g ⊆ Mn F (ξ (k) ) : 1 ≤ i, j ≤ n, k ≥ 1
UD i,j

g which by
we can define the reduced characteristic coefficients of elements of UD,
[Row2, Remark 24.67] lie in F1 .
Lemma 3. Assume Char (K) = 0. If an element a of UD g has a unique eigenvalue
α (i.e., of multiplicity n), then a is scalar. If Char (K) = k 6= 0 then a is k l −scalar
for some l.
g and a − αI is nilpotent, and
Proof. If Char (K) = 0, then α is the element of UD
thus 0.
l
g therefore akl − αkl I is nilpotent,
If Char (K) = k then αk is the element of UD,
l
and thus 0. Thus a is k −scalar. This is impossible if k is not the devisor of the
size of the matrices n. 
g must divide n.
Lemma 4. The multiplicity of any eigenvalue of an element a of UD
In particular, when n is odd, a cannot have an eigenvalue of multiplicity 2.
Proof. Recall [Row1, Remark 4.106] that for any element a in a division algebra,
represented as a matrix, the eigenvalues of a occur with the same multiplicity, which
thus must divide n. 
Proposition 1. Suppose we have a homomorphism ϕ : UD → A given by the spe-
g specializes
cialization ϕ(Yk ) = ak . Then any characteristic coefficient of Yk in UD
to the corresponding characteristic coefficient of ak .
Proof. Any characteristic coefficient can be expressed as the ratio of two central
polynomials, in view of [Row3, Theorem 1.4.12]; also see [BeR, Theorem J, p. 27]
whichPsays that for any characteristic coefficient αℓ of the characteristic polynomial
t
λt + ℓ=1 (−1)ℓ αℓ λt−ℓ that we have
t
X
αℓ f (a1 , . . . , at , r1 , . . . , rm ) = f (T ℓ1 a1 , . . . , T ℓt at , r1 , . . . , rm ), (1)
ℓ=1
P
summed over all vectors (ℓ1 , . . . , ℓt ) where each ℓi ∈ {0, 1} and ℓi = t, where f is
any t−alternating polynomial nonidentity (and t = n2 ), so any characteristic coef-
ficient of a polynomial specializes according to the specialization from UD g induced
from ϕ, where we take f (a1 , . . . , at , r1 , . . . , rm ) 6= 0. 
For the proof of our results we need the following consequence of Proposition 1:
Proposition 2. All of the functions in Donkin’s theorem [D] can be embedded
g
in UD.
For n > 2, we also have an easy consequence of the theory of division algebras.
Lemma 5. Suppose for some polynomial p and some number q < n, that pq takes
on only scalar values in Mn (K), over an infinite field K, for n prime. Then p takes
on only scalar values in Mn (K).
4 ALEXEY KANEL-BELOV, SERGEY MALEV, LOUIS ROWEN

g of degree n,
Proof. We can view p as an element of the generic division algebra UD
and we adjoin a q-root of 1 to K if necessary. Then p generates a subfield of
g The latter is impossible, so the dimension is 1; i.e., p is
dimension 1 or n of UD.
already central. 
2.1. The case M3 (K). Now we turn specifically to the case n = 3.
Lemma 6. We define functions ωk : M3 (K) → K as follows: Given a matrix a,
let λ1 , λ2 , λ3 be the eigenvalues of a, and denote
X
ωk := ωk (a) = λi1 . . . λik .
1≤i1 <i2 <···<ik ≤3

Let p(x1 , . . . , xm ) be a semi-homogeneous, trace vanishing polynomial.

3
2 (p(x1 ,...,xm ))
Consider the rational function H(x1 , . . . , xm ) = ω
ω3 (p(x1 ,...,xm ))2 (taking values in
K ∪ {∞}). If Im H is dense in K, then Im p is dense in sl3 .
Proof. Note that ω2 (p)3 and ω3 (p)2 are semi-homogeneous. Thus, Im H is dense
in K iff the image of the pair (ω2 (p)3 , ω3 (p)2 ) is dense in K 2 . But since ω2 and ω3
are algebraically independent, so are ω2 (p)3 and ω3 (p)2 , so we conclude that the
image of the pair (ω2 (p)3 , ω3 (p)2 ) is dense in K 2 . Thus, the set of characteristic
polynomials of evaluations of p is dense in the space of all possible characteristic
polynomials of trace zero matrices. Therefore, the set of all triples (λ1 , λ2 , −λ1 −λ2 )
of eigenvalues of matrices from Im p is dense in the plane x + y + z = 0 defined in
K 3 , implying that Im p is dense in sl3 . 
Let K be an algebraically closed field. We say that a polynomial p is trace-
vanishing if each of its evaluations have trace 0; i.e., tr(p) is a trace identity of p.
Also, for char(K) 6= 3 we fix a primitive cube root ε 6= 1 of 1; when char(K) = 3
we take ε = 1.
Theorem 3. Let p(x1 , . . . , xm ) be a semi-homogeneous polynomial which is trace-
vanishing on 3 × 3 matrices. Then Im p is one of the following:
• {0},
• the set of scalar matrices (which can occur only if Char K = 3),
• a dense subset of sl3 (K), or
• the set of 3−scalar matrices, i.e., the set of matrices with eigenvalues
(γ, γε, γε2 ), where ε is our cube root of 1.
Proof of Theorem 3. We define the functions ωk : Mn (K) → K as in Lemma 6,
ω2 (p(x1 ,...,xm ))3
and consider the rational function H = ω 3 (p(x1 ,...,xm ))
2 (taking values in K ∪ {∞}).

If ω2 (p) = ω3 (p) = 0, then each evaluation of p is a nilpotent matrix, contradict-

ing Amitsur’s Theorem. Thus, either Im H is dense in K, or H must be constant.
If Im H is dense in K, then Im p is dense in sl3 by Lemma 6.
So we may assume that H is a constant, i.e., αω23 (p) + βω32 (p) = 0 for some
α, β ∈ K not both 0. Fix generic matrices Y1 , . . . , Ym . We claim that the eigenvalues
λ1 , λ2 , −λ1 − λ2 of q := p(Y1 , . . . , Ym ) are pairwise distinct. Otherwise either they
are all equal, or two of them are equal and the third is not, each of which is
impossible by Lemmas 3 and 4 since q ∈ UD. g
′ ′ ′ ′
Let λ1 , λ2 , −λ1 − λ2 be the eigenvalues of another matrix r ∈ Im p. Thus we
have the following:
αω23 (r) + βω32 (r) = αω23 (q) + βω32 (q) = 0.
 IMAGES OF MULTILINEAR POLYNOMIALS 5

Therefore we have homogeneous equations on the eigenvalues. Dividing by λ62

and λ′6
2 respectively, we have the same two polynomial equations of degree 6 on
λ1 λ′1 λ′1
λ2 and λ′ , yielding six possibilities for λ′ . The six permutations of λ1 , λ2 , and
2 2
λ′
λ3 = −λ1 − λ2 define six pairwise different λ′1 unless (λ1 , λ2 , λ3 ) is a permutation
2
(multiplied by a scalar) of one of the following triples: (1, 1, −2), (1, −1, 0), (1, ε, ε2).
The first case is impossible since the eigenvalues must be pairwise distinct. The
second case give us an element of Amitsur’s algebra UD g with eigenvalue 0 and thus
determinant 0, contradicting Amitsur’s Theorem. In the third case the polynomial
p is 3−scalar. Thus, either p is 3−scalar polynomial, or each matrix from Im p will
have the same eigenvalues up to permutation and scalar multiple. Note for p being
3−scalar this is true also.
Assume that for some i ∈ {2, 3} that tr pi is not identically zero. Then λi1 , λi2 ,
and λi3 are three linear functions on tr pi . Hence we have the PI (polynomial
identity) (pi − λi1 )(pi − λi2 )(pi − λi3 ). Thus by Amitsur’s Theorem, one of the factors
is a PI. Hence pi is a scalar matrix. However i 6= 2 by Lemma 4. Hence i = 3. In this
case the image of p is the set of matrices with eigenvalues {(γ, γε, γε2) : γ ∈ K}.
Thus, we may assume that p satisfies tr(pi ) = 0 for i = 1, 2 and 3. Now
ω1 (p) = tr(p) = 0 and 2ω2 (p) = (tr(p))2 − tr(p2 ) = 0.
Hence ω1 = ω2 = 0 if char(K) 6= 2; in this case ω3 is either 0 (and hence p is PI)
or not 0 (and hence p is 3−scalar).
So assume that char(K) = 2. Recall that
0 = tr(p3 ) = λ31 + λ32 + λ33 = λ31 + λ32 + λ33 − 3λ1 λ2 λ3 + 3λ1 λ2 λ3 .
But λ31 + λ32 + λ33 − 3λ1 λ2 λ3 is a multiple of λ1 + λ2 + λ3 (seen by substituting
−(λ1 + λ2 ) for λ3 ) and thus equals 0. Thus, 0 = 3λ1 λ2 λ3 = λ1 λ2 λ3 = ω3 (p), and
the Hamilton-Cayley equation yields p3 + ω2 p = 0. Therefore, p(p2 + ω2 ) = 0 and
by Amitsur’s Theorem either p is PI, or p2 = −ω2 (which is central), implying by
Lemma 5 that p is central. 

g
Example 1. The element [x, [y, x]x[y, x]−1 ] of U D takes on only 3−scalar values
(see [Row3, Theorem 3.2.21, p. 180]) and thus gives rise to a homogeneous polyno-
mial taking on only 3−scalar values.
Now we consider the possible image sets of multilinear trace-vanishing polyno-
mials.
Lemma 7. If p is a multilinear polynomial, not PI nor central, then there ex-
ist a collection of matrix units (E1 , E2 , . . . , Em ) such that p(E1 , E2 , . . . , Em ) is a
diagonal but not scalar matrix.
Proof. By Lemmas 1 and 2, the linear span of all p(E1 , E2 , . . . , Em ) for any matrix
units Ei such that p(E1 , E2 , . . . , Em ) is diagonal includes all Diag{x, y, −x − y}.
In particular there exist a collection of matrix units (E1 , E2 , . . . , Em ) such that
p(E1 , E2 , . . . , Em ) is a diagonal but not scalar matrix. 
Theorem 4. Let p be a multilinear polynomial which is trace-vanishing on 3 ×
3 matrices over a field K of arbitrary characteristic. Then Im p is one of the
following:
• {0},
6 ALEXEY KANEL-BELOV, SERGEY MALEV, LOUIS ROWEN

• the set of scalar matrices,

• the set of 3−scalar matrices, or
• for each triple λ1 + λ2 + λ3 = 0 there exist a matrix M ∈ Im p with
eigenvalues λ1 , λ2 and λ3 .

Proof. If the polynomial ω2 (p) (defined in the proof of Theorem 3) is identically

zero, then the characteristic polynomial is p3 −ω3 (p) = 0, implying p is either scalar
(which can happen only if Char (K) = 3) or 3−scalar. Therefore we may assume
that the polynomial ω2 (p) is not identically zero. Let

fα,β (M ) = αω2 (M )3 + βω3 (M )2 .

It is enough to show that for any α, β ∈ K there exists a non-nilpotent matrix

M = p(a1 , . . . , am ) such that fα,β (p(a1 , . . . , am )) = 0, since this will imply that the
β
image of H (defined in Lemma 6) contains all − α and thus K ∪{∞}. (For example,
if α = 0 and β 6= 0, then ω3 (M ) = 0, implying ω2 (M ) 6= 0 since ω1 (M ) = 0
and M is non-nilpotent, and thus H = ∞.) Therefore, for any trace vanishing
polynomial (i.e., a polynomial x3 + γ1 x + γ0 ) there is a matrix in Im p for which
this is the characteristic polynomial. Hence whenever λ1 + λ2 + λ3 = 0 there is a
matrix with eigenvalues λi .
Without loss of generality we may assume that a = p(Y1 , . . . , Ym ) and b =
p(Ỹ1 , Y2 . . . , Ym ) are not proportional for generic matrices Ỹ1 , Y1 , . . . , Ym , cf. [BeMR2,
Lemma 4]. Consider the polynomial ϕα,β (t) = fα,β (a + tb). There are three cases
to consider:
CASE I. ϕα,β = 0 identically. Then fα,β (a) = 0, and a is not nilpotent by
Proposition 2.
CASE II. ϕα,β is a constant β̃ 6= 0. Then fα,β (b + ta) = t6 ϕα,β (t−1 ) = β̃t6 ; thus
fα,β (b) = 0, and b is not nilpotent by Proposition 2.
CASE III. ϕα,β is not constant. Then it has finitely many roots. Assume that
for each substitution t the matrix a + tb is nilpotent; in particular, ω2 (a + tb) = 0.
Note that ω2 (a+tb) equals the sum of principal 2×2 minors and thus is a quadratic
polynomial (for otherwise ω2 (b) = 0 which means that ω2 (p) is identically zero, a
contradiction). Hence ω2 (a + tb) has two roots, which we denote as t1 and t2 . If
t1 = t2 , then t1 is uniquely defined and thus, in view of Remark 2, is a rational
function in the entries of a and b, and a+t1 b is a nilpotent rational function (because
we assumed that one of a + t1 b and a + t2 b is nilpotent, but here they are equal.)
At least one of t1 and t2 is a root of ϕα,β .
If only t1 is a root, then t1 is uniquely defined and thus, by Remark 2, is a rational
function; hence, a+t1 b is a nilpotent polynomial, contradicting Proposition 2. Thus,
we may assume that both t1 and t2 are roots of ϕα,β . But ϕα,β (ti ) is nilpotent,
and in particular ω3 (a + ti b) = 0. Thus there exists exactly one more root t3 of
ω3 (a + tb), which is uniquely defined and thus, by Remark 2, is rational. Hence we
may consider the polynomial q(x1 , . . . , xm , x̃1 ) = a + t3 b, which must satisfy the
condition tr(q) = det(q) = 0. This is impossible for homogeneous q by Theorem 3,
and also impossible for nonhomogeneous q since the leading homogenous component
qd would satisfy tr(qd ) = det(qd ) = 0, a contradiction. 

Remark 4. Assume that char(K) = 3 and p is a multilinear polynomial, which

is neither PI nor central. Then, according to Lemma 7 there exists a collection of
 IMAGES OF MULTILINEAR POLYNOMIALS 7

matrix units Ei such that

p(E1 , . . . , Em ) = Diag{α, β, γ}
is diagonal but not scalar. Without loss of generality, α 6= β. Hence p3 (E1 , . . . , Em ) =
Diag{a3 , β 3 , γ 3 } and α3 6= β 3 because char(K) = 3. Therefore p is not 3−scalar.
Theorem 5. If there exist α, β, and γ in K such that α + β + γ, α + βε + γε2
and α + βε2 + γε are nonzero, together with matrix units E1 , E2 , . . . , Em such that
p(E1 , E2 , . . . , Em ) has eigenvalues α, β and γ, then Im p is dense in M3 .
Proof. Define χ to be the permutation of the set of matrix units, sending the indices
1 → 2, 2 → 3, and 3 → 1. For example, χ(e12 ) = e23 . For triples T1 , . . . , Tm (each
Ti = (ti,1 , ti,2 , ti,3 )) consider the function

f (T1 , . . . , Tm ) = p(t1,1 x1 + t1,2 χ(x1 ) + t1,3 χ−1 (x1 ), t2,1 x2 + t2,2 χ(x2 ) + t2,3 χ−1 (x2 ),
. . . , tm,1 xm + tm,2 χ(xm ) + tm,3 χ−1 (xm )). (2)
Opening the brackets, we have 3m terms, each of which we claim is a diagonal
matrix. Each term is a monomial with coefficient of the type
χkπ(1) χkπ(2) · · · χkπ(m) xπ(1) xπ(2) · · · xπ(m) ,
where ki is −1, 0 or 1, and π is a permutation. Since we substitute only matrix
units in p, by Lemma 1 the image is either diagonal or a matrix unit with some
coefficient. For each of the three vertices v1 , v2 , v3 in our graph define the index
ιℓ , for 1 ≤ ℓ ≤ 3 to be the number of incoming edges to vℓ minus the number of
outgoing edges from vℓ . Thus, at the outset, when the image is diagonal, we have
ι1 = ι2 = ι3 = 0.
We claim that after applying χ to any matrix unit the new ι′ℓ will all still be
congruent modulo 3. Indeed, if the edge 12 ~ is changed to 23,
~ then ι′1 = ι + 1 and
′ ′
ι3 = ι3 + 1, whereas ι2 = ι2 − 2 ≡ ι2 + 1. The same with changing 23 ~ to 31~ and
~ ~ ~ ~
31 to 12. If we make the opposite change 21 to 13, then (modulo 3) we subtract 1
throughout. If we make a change of the type ii ~ 7→ jj,
~ then ι′ = ιℓ for each ℓ.
ℓ
k1 k2 km
If p(χ x1 , χ x2 , . . . , χ xm ) = eij , this means that the number of incoming
edges minus the number of outgoing edges of the vertex i is −1 (mod 3) and
the number of incoming edges minus the number of outgoing edges of j is 1
(mod 3), which are not congruent modulo 3. Thus the values of the mapping f
defined in (2) are diagonal matrices. Now fix 3m algebraically independent triples
T1 , . . . , Tm , Θ1 , . . . , Θm , Υ1 , . . . , Υm . Assume that Im f is 2−dimensional. Then
Im df must also be 2−dimensional at any point. Consider the differential df at the
point (Θ1 , T2 , . . . , Tm ). Thus,
f (Θ1 , T2 , . . . , Tm ), f (T1 , T2 , . . . , Tm ), f (Θ1 , Θ2 , . . . , Tm )
belong to Im df . Thus these three matrices must span a linear space of dimension
not more than 2. Hence they lie in some plane P . Now take
f (Θ1 , Θ2 , T3 , . . . , Tm ), f (Θ1 , T2 , T3 , . . . , Tm ), f (Θ1 , Θ2 , Θ3 , T4 , . . . , Tm ).
For the same reason they lie in a plane, which is the plane P because it has two
vectors from P . By the same argument, we conclude that all the matrices of the
type f (Θ1 , . . . , Θk , Tk+1 , . . . , Tm ) lie in P. Now we see that
f (Θ1 , . . . , Θm−1 , Tm ), f (Θ1 , . . . , Θm ), f (Υ1 , Θ2 , . . . , Θm )
8 ALEXEY KANEL-BELOV, SERGEY MALEV, LOUIS ROWEN

also lie in P . Analogously we obtain that also

f (Υ1 , . . . , Υk , Θk+1 , . . . , Θm ) ∈ P
for any k.
Hence for 3m algebraically independent triples
T1 , . . . , Tm ; Θ1 , . . . , Θm ; Υ1 , . . . , Υm ,
we have obtained that f (T1 , . . . , TM ), f (Θ1 , . . . , Θm ) and f (Υ1 , . . . , Υm ) lie in one
plane. Thus any three values of f , in particular Diag{α, β, γ}, Diag{β, γ, α} and
Diag{γ, α, β}, must lie in one plane. We claim that this can happen only if
α + β + γ = 0, α + βε + γε2 = 0, or α + βε2 + γε = 0.
Indeed, Diag{α, β, γ}, Diag{β, γ, α} and Diag{γ, α, β}, are dependent if and
only if the matrix  
αβγ
β γ α
γαβ
is singular, i.e., its determinant 3αβγ − (α3 + β 3 + γ 3 ) = 0. But this has the desired
three roots when viewed as a cubic equation in γ.
We have a contradiction to our hypothesis. 

Remark 5. If there exist α, β, and γ such that α + β + γ = 0 but (α, β, γ) is

not proportional to (1, ε, ε2 ) or (1, ε2 , ε), with matrices E1 , E2 , . . . , Em such that
p(E1 , E2 , . . . , Em ) has eigenvalues α, β and γ, then either all diagonalizable trace
zero matrices lie in Im p, or Im p is dense in M3 (K). If α + βε + γε2 = 0 but
(α, β, γ) is not proportional to (1, ε, ε2 ) or (1, 1, 1), then all diagonalizable matrices
with eigenvalues α + β, α + βε and α + βε2 lie in Im p or Im p is dense in M3 (K).
Remark 6. The proof of Theorem 5 works also for any field K of characteristic 3.
In this case ε = 1. Hence, if there are α, β, and γ in K such that
α + β + γ 6= 0,
together with matrix units E1 , E2 , . . . , Em such that p(E1 , E2 , . . . , Em ) has eigen-
values α, β and γ, then Im p is dense in M3 . Therefore, for Char K = 3, any
multilinear polynomial p is either trace-vanishing or Im p is dense in M3 (K).
Theorem 6. If p is a multilinear polynomial such that Im p does not satisfy the
equation γω1 (p)2 = ω2 (p) for γ = 0 or γ = 41 , then Im p contains a matrix with
two equal eigenvalues that is not diagonalizable and of determinant not zero. If
Im p does not satisfy any equation of the form γω1 (p)2 = ω2 (p) for any γ, then
the set of non-diagonalizable matrices of Im p is Zariski dense in the set of all
non-diagonalizable matrices, and Im p is dense.
Proof. If not, then by [BeMR2, Lemma 4] there is at least one variable (say, x1 ) such
that a = p(x1 , x2 , . . . , xm ) does not commute with b = p(x̃1 , x2 , . . . , xm ). Consider
the matrix a + tb = p(x1 + tx̃1 , x2 , . . . , xm ), viewed as a polynomial in t.
Recall
Q that the discriminant of a 3×3 matrix with eigenvalues λ1 , λ2 , λ3 is defined
as 1≤i<j≤3 (λi − λj )2 . Thus, the discriminant of a + tb is a polynomial f (t) of
degree 6. If f (t) has only one root t0 , then this root is defined in terms of the entries
of x̃1 , x1 , x2 , . . . , xm , and invariant under the action of the symmetric group, and
 IMAGES OF MULTILINEAR POLYNOMIALS 9

thus is in Amitsur’s division algebra UD. g By Lemma 3, a + t0 b is scalar, and the

uniqueness of t0 implies that a and b are scalar, contrary to assumption.
Thus, f (t) has at least two roots - say, t1 6= t2 , and the matrices a+t1 b and a+t2 b
each must have multiple eigenvalues. If both of these matrices are diagonalizable,
then each of a + ti b have a 2−dimensional plane of eigenvectors. Therefore we have
two 2−dimensional planes in 3−dimensional linear space, which must intersect.
Hence there is a common eigenvector of both a+ti b and this is a common eigenvector
of a and b. If a and b have a common eigenspace of dimension 1 or 2, then there is at
least one eigenvector (and thus eigenvalue) of a that is uniquely defined, implying
a∈U gD by Remark 2, contradicting Lemma 3. If a and b have a common eigenspace
of dimension 3, then a and b commute, a contradiction.
We claim that there cannot be a diagonalizable matrix with equal eigenvalues on
the line a + tb. Indeed, if there were such a matrix, then either it would be unique
(and thus an element of U g D, which cannot happen), or there would be at least two
such diagonalizable matrices, which also cannot happen, as shown above.
Assume that all matrices on the line a+ tb of discriminant zero have determinant
zero. Then either all of them are of the type Diag{λ, λ, 0} + e12 or all of them are
of the type Diag{0, 0, µ} + e12 . (Indeed, there are three roots of the determinant
equation det(a + tb) = 0, which are pairwise distinct, and all of them give a matrix
with two equal eigenvalues, all belonging to one of these types, since otherwise one
eigenvalue is uniquely defined and thus yields an element of U g D, which cannot
happen.
In the first case, all three roots of the determinant equation det(a+tb) = 0 satisfy
the equation (ω1 (a+tb))2 = 4ω2 (a+tb). Hence, we have three pairwise distinct roots
of the polynomial of maximal degree 2, which can occur only if the polynomial is
identically zero. It follows that also (ω1 (a))2 − 4ω2 (a) = 0, so (ω1 (p))2 − 4ω2 (p) = 0
is identically zero, which by hypothesis cannot happen.
In the second case we have the analogous situation, but ω2 (p) will be identically
zero, a contradiction.
Thus on the line a+ tb we have at least one matrix of the type Diag{λ, λ, µ} + e12
and λµ 6= 0. Consider the algebraic expression µλ−1 . If not constant, then it takes
on almost all values, so assume that it is a constant δ. Then δ 6= −2, since otherwise
this matrix will be the unique matrix of trace 0 on the line a+tb and thus an element
of Ug D, contrary to Lemmas 3 and 4. Consider the polynomial q = p − δ+2 trp
. At the
same point t it takes on the value Diag{0, 0, (δ −1)λ}+e12. Hence all three pairwise
distinct roots of the equation det q(x1 + tx̃1 , x2 , . . . , xm ) = 0 will give us a matrix of
the form Diag{0, 0, ∗} + e12 (otherwise we have uniqueness and thus an element of
g
U D), contradicting Lemma 4. Therefore q satisfies an equation ω2 (q) = 0. Hence,
p satisfies an equation ω1 (p)2 − cω2 (p) = 0, for some constant c, a contradiction.
Hence almost all non-diagonalizable matrices belong to the image of p, and they
are almost all matrices of discriminant 0 (a subvariety of M3 (K) of codimension 1).
By Amitsur’s Theorem, Im p cannot be a subset of the discriminant surface. Thus,
Im p is dense in M3 (K). 

Remark 7. Note that if ω1 (p) is identically zero, and ω2 (p) is not identically
zero, then Im p contains a matrix similar to Diag{1, 1, −2} + e12 . Hence Im p
contains all diagonalizable trace zero matrices (perhaps with the exception of the
diagonalizable matrices of discriminant 0, i.e. matrices similar to Diag{c, c, −2c}),
10 ALEXEY KANEL-BELOV, SERGEY MALEV, LOUIS ROWEN

all non-diagonalizable non-nilpotent trace zero matrices, and all matrices N for
which N 2 = 0. Nilpotent matrices of order 3 also belong to the image of p, as we
shall see in Lemma 9.

3. Proof of the main Theorem

Lemma 8. A matrix is 3-scalar iff its eigenvalues are in {γ, γε, γε2 : γ ∈ K}, where
γ 3 ∈ K is its determinant. The variety V3 of 3-scalar matrices has dimension 7.
Proof. The first assertion is immediate since the characteristic polynomial is x3 −γ 3 .
Hence V3 is a variety. The second assertion follows since the invertible elements
of V3 are defined by two equations: tr(x) = 0 and tr(x−1 ) = 0 and thus a V3 is a
variety of codimension 2. 

Lemma 9. Assume Char K 6= 3. If p is neither PI nor central, then the variety

V3 is contained in Im p.
Proof. According to Lemma 2 there exist matrix units E1 , E2 , . . . , Em such that
p(E1 , E2 , . . . , Em ) = e1,2 . Consider the mapping χ described in the proof of Theo-
rem 5. For any triples Ti = (t1,i , t2,i , t3,i ), let
f (T1 , T2 , . . . , Tm ) = p(. . . , t1,i Ei + t2,i χ(Ei ) + t3,i χ2 (Ei ), . . . ).
Im f (a subset of Im p) is a subset of the 3−dimensional linear space
L = {αe12 + βe23 + γe31 , α, β, γ ∈ K}.
Since e12 , e23 and e31 belong to Im f , we see that Im f is dense in L, and hence at
least one matrix a = αe12 + βe23 + γe31 for αβγ 6= 0 belongs to Im p. Note that
this matrix is 3−central. Thus the variety V3 , excluding the nilpotent matrices, is
contained in Im p. The nilpotent matrices of order 2 also belong to the image of p
since they are similar to e12 .
Let us show that all nilpotent matrices of order 3 (i.e., matrices similar to e12 +
e23 ), also belong to Im p. We have the multilinear polynomial
f (T1 , T2 , . . . , Tm ) = q(T1 , T2 , . . . , Tm )e12 +r(T1 , T2 , . . . , Tm )e23 +s(T1 , T2 , . . . , Tm )e31 ,
therefore q, r and s are three scalar multilinear polynomials. Assume there is no
nilpotent matrix of order 3 in Im p. Then we have the following: if q = 0 then
either rs = 0, if r = 0 then sq = 0, and if s = 0 then qr = 0. Assume q1 is the
greatest common divisor of q and r and q2 = qq1 . Note both qi are multilinear
polynomials defined on disjoint sets of variables. If q1 = 0 then r = 0 and if
q2 = 0 then s = 0. Note there are no double efficients, and thus r = q1 r′ is a
multiple of q1 and s = q2 s′ is a multiple of q2 . The polynomial r′ cannot have
common devisors with q2 , therefore if we consider any generic point (T1 , . . . , Tm )
on the surface r′ = 0 then r(T1 , . . . , Tm ) = 0 and q(T1 , . . . , Tm ) 6= 0. Hence
s(T1 , . . . , Tm ) = 0 for any generic (T1 , . . . , Tm ) from the surface r′ = 0. Therefore r′
is the divisor of s. Remind both q1 and q2 are multilinear polynomials defined on
disjoint subsets of {T1 , T2 , . . . , Tm }. Without loss of generality q1 = q1 (T1 , . . . , Tk ),
and q2 = q2 (Tk+1 , . . . , Tm ). Therefore r′ = r′ (Tk+1 , . . . , Tm ) and it is divisor of s.
Also remind s = s′ q2 so q2 (Tk+1 , . . . , Tm ) is also divisor of s. Hence r′ = cq2 where
c is constant. Thus r = q1 r′ = cq1 q2 = cq. However there exist (Tk+1 , . . . , Tm ) such
that q = 0 and r = 1 (i.e. such that f (Tk+1 , . . . , Tm ) = e23 ). A contradiction. 
 IMAGES OF MULTILINEAR POLYNOMIALS 11

Remark 8. When Char K = 3, then V3 is the space of the matrices with equal
eigenvalues (including also scalar matrices). The same proof shows that all nilpotent
matrices belong to the image of p, as well as all matrices similar to cI + e12 + e23 .
But we do not know how to show that scalar matrices and matrices similar to
cI + e12 belong to the image of p.
Proof of Theorem 2. First assume that Char K 6= 3. According to Lemma 9
the variety V3 is contained in Im p. Therefore Im p is either the set of 3-scalar matri-
ces, or some 8−dimensional variety (with 3-scalar subvariety), or is 9−dimensional
(and thus dense).
It remains to classify the possible 8−dimensional images. Let us consider all
matrices p(E1 , . . . , Em ) where Ei are matrix units. If all such matrices have trace 0,
then Im p is dense in sl3 (K), by Theorem 4. Therefore we may assume that at least
one such matrix a has eigenvalues α, β and γ such that α + β + γ 6= 0. By Theorem
5 we cannot have α + β + γ, α + βε + γε2 and α + βε2 + γε all nonzero. Hence
a either is scalar, or a linear combination (with nonzero coefficients) of a scalar
matrix and Diag{1, ε, ǫ2 } (or with Diag{1, ε2 , ǫ}, without loss of generality - with
Diag{1, ε, ε2 }). By Theorem 6, if Im p is not dense, then p satisfies an equation of
the type (tr(p))2 = γtr(p2 ) for some γ ∈ K. Therefore, if a scalar matrix belongs to
Im p, then γ = 31 and Im p is the set of 3−scalar plus scalar matrices. If the matrix
a is not scalar, then it is a linear combination of a scalar matrix and Diag{1, ε, ε2 }.
Hence, by Remark 5, Im p is also the set of 3−scalar plus scalar matrices. In any
case, we have shown that Im p is either {0}, K, the set of 3−scalar matrices, the set
of 3−scalar plus scalar matrices (matrices with eigenvalues (α+ β, α+ βε, α+ βε2 )),
sl3 (K) (perhaps lacking nilpotent matrices of order 3), or is dense in M3 (K).
If Char K = 3, then by Remark 6 the multilinear polynomial p is either trace-
vanishing or Im p is dense in M3 (K). If p is trace-vanishing, then by Theorem 4,
Im p is one of the following: {0}, the set of scalar matrices, the set of 3−scalar
matrices, or for each triple λ1 + λ2 + λ3 = 0 there exists a matrix M ∈ Im p with
eigenvalues λ1 , λ2 and λ3 . 

4. Open problems
Problem 1. Does there actually exist a multilinear polynomial whose image eval-
uated on 3 × 3 matrices consists of 3−scalar matrices?
Problem 2. Does there actually exist a multilinear polynomial whose image eval-
uated on 3 × 3 matrices is the set of scalars plus 3−scalar matrices?
Remark 9. Problems 1 and 2 both have the same answer. If they both have
affirmative answers, such a polynomial would a counter-example to Kaplansky’s
problem.
Problem 3. Is it possible that the image of a multilinear polynomial evaluated on
3 × 3 matrices is dense but not all of M3 (K)?
Problem 4. Is it possible that the image of a multilinear polynomial evaluated
on 3 × 3 matrices is the set of all trace vanishing matrices without discriminant
vanishing diagonalazable matrices?
12 ALEXEY KANEL-BELOV, SERGEY MALEV, LOUIS ROWEN

References
[AlM] Albert, A.A.; Muckenhoupt. On matrices of trace 0, Michigan Math. J. 1 (1957), 1–3.
[BKP] Belov, A.; Kunavskii, B; Plotkin,E; Word equations in simple groups and poloynomial
equations in simple algebras, Vestnik St. Petersburg University, 46 (2013), 3–13
[BeMR1] Belov, A.; Malev, S.; Rowen, L. The images of non-commutative polynomials evaluated
on 2 × 2 matrices, Proc. Amer. Math. Soc 140 (2012), 465–478.
[BeMR2] Belov, A.; Malev, S.; Rowen, L. Power-central polynomials on matrices, preprint,
(2013).
[BeR] Belov, A.; Rowen, L.H. Computational Aspects of Polynomial Identities, A. K. Peters
Ltd., Wellesley, MA. (2005).
[BrK] Bresar, M. and Klep, I., Values of Noncommutative Polynomials, Lie Skew-Ideals and
the Tracial Nullstellensatz, arXiv:0810.1774
[Ch] Chuang, C.-L. On ranges of polynomials in finite matrix rings, Proceeding of the Amer-
ican Mathematical Society 110 (1990), no. 2, 293–302.
[D] Donkin, S. Invariants of several matrices, Invent. Math. 110 (1992), no. 2, 389–401.
[Ku1] Kulyamin, V.V. Images of graded polynomials in matrix rings over finite group algebras
Russ. Math. Surv.55 (2000), 345–346.
[Ku2] Kulyamin, V.V. On images of polynomials in finite matrix rings, Thes. Cand. Phys.-
Math. Sci., Moscow Lomonosov state University Moscow (2000).
[La] Larsen M. Word maps have large image, Israel J. Math. 139 (2004) 149–156.
[LaS] Larsen M.; Shalev, A. Word maps and Waring type problems, J. Amer. Math. Soc. 22
(2009), 437-466.
[LeZh] Lee, Ts.-Kw.; Zhou, Y. Right Ideals generated by an idempotent of finite rank, Linear
Algebra and its Applications 431 (2009), 2118–2126.
[Row1] Rowen, L.H., Graduate algebra: Commutative view, AMS Graduate Studies in Mathe-
matics 73 (2006).
[Row2] Rowen, L.H., Graduate algebra: Noncommutative view, AMS Graduate Studies in Math-
ematics 91 (2008).
[Row3] Rowen, L. Polynomial identities in ring theory, Academic press, New York (1980).
[S] Shalev, A. Word maps, conjugacy classes, and a non-commutative Waring-type theorem,
Annals of Math., to appear.

Department of mathematics, Bar Ilan University, Ramat Gan, Israel

E-mail address: beloval@math.biu.ac.il
E-mail address: malevs@math.biu.ac.il
E-mail address: rowen@math.biu.ac.il