NONSINGULAR BILINEAR MAPS, SPACES OF

MATRICES, IMMERSIONS AND EMBEDDINGS ∗

ZORAN PETROVIĆ †
Faculty of Mathematics, Studentski trg 16,
11000 Belgrade, Serbia
E-mail: [email protected]

We present in this short paper relations between vector spaces of matrices which
satisfy some rank conditions, problem of the existence of nonsingular bilinear maps
and the problem of the existence of immersions and embeddings of real projective
spaces into Euclidean spaces.

1. Introduction
In every differential topology textbook one finds Whitney’s immersion and
embedding theorems. But, the real story of immersions started with the
results of Hirsch ([15]) and Smale ([30]). We do not intend to follow closely
all these developments. Rather, we restrict ourselves mainly to the problem
of the existence of immersions of real projective spaces into Euclidean spaces
and the problems related to this one.
We show in the next section why the real projective spaces are especially
interesting as we discuss immersions of manifolds and how one attempts to
prove the existence of particular immersions. These methods turn out to
be related to the question of the dimension of the vector spaces of matrices
on which some condition on the rank is imposed and which we discuss in
the section dedicated to the spaces of matrices. These problems are also
interesting per se, since these are such basic objects and the questions we
ask are so natural, that one would certainly like to know the answer to
them.

∗ MSC 2000: 57R42, 57R40, 15A03.

Keywords: bilinear maps, immersions, embeddings, spaces of matrices.
† Work supported by the Ministry of Science under contract No 1854.

427
428

Finally, we point out to some recent works which show how the question
of nonsingular bilinear maps is related to special kinds of embeddings into
Euclidean spaces.

2. Immersions
One of the long-standing problems of differential topology had certainly
been the resolution of the so-called immersion conjecture: Every compact
n-dimensional manifold M n admits an immersion in the Euclidean space
R2n−α(n) , where α(n) stands for the number of 1’s in the binary expansion
of n. Certainly, the most natural question one can ask after hearing about
this conjecture is: How this strange number α(n) comes into play? But,
surprisingly, this number is rather natural for this problem. Namely, one
can show, using only Stiefel–Whithey characteristic classes, that the fol-
t1 t2 t
lowing product of real projective spaces RP 2 × RP 2 × · · · × RP 2 k does
not immerse into R2n−k−1 , if n = 2t1 + · · · + 2tk is the representation of n
as the sum od different powers of 2. From this example we can conclude
two things. First, one cannot do better than what the immersion conjec-
ture claims. Second, the real projective spaces are rather important for the
immersion problem.
Actually, one can say that the work on the immersion conjecture has started
with the paper [22] where the author had shown, using the Steenrod squares,
that w̄i (M n ) = 0 for i > n − α(n) for any n-dimensional compact mani-
fold M n . Of course, w̄i (M n ) stand for the Stiefel-Whitney classes of the
normal bundle of M n . In the papers [6],[7] and [8], the authors first ex-
panded this algebraic result by computing the ideal of all relations among
the Stiefel-Whitney classes of normal bundles of all compact, n-dimensional
manifolds and then proceeded to find the geometric realization of the alge-
braic structures which appeared in this investigation, namely to construct
certain commutative diagram of spaces and maps realizing corresponding
diagram od algebras. This program was finally completed by Ralph Cohen
in 1985 (see [10]). Since this paper is rather long and involved and the
result is so important, some topologists think that it would be nice to find
another proof, perhaps using some other methods.
The following question still remains unresolved:
Q: What is the smallest k such that RP n # Rn+k ?
We use # to denote an immersion.
One can approach this problem from two directions. First, one tries to
 429

actually construct immersions and second, to apply various methods from

algebraic topology in order to show that some immersions cannot exist.
We will be mainly interested in the first approach. As far as proving the
nonexistence of immersions, let us just mention that various methods have
been used, mainly the obstruction theory, but also some methods specially
suited for this problem.
Let us go back to the question of constructing immersions. Suppose
t ≥ n. An axial mapg : RP n × RP n −→ RP t is a map g such that
g(x, ∗) = x = g(∗, x) for x ∈ RP n , where ∗ stands for the base point and
we regard real projective space RP n as embedded into RP t . The existence
of such a map is equivalent to the existence of an immersion of RP n into
RP t (see [19],[29]). It is interesting to note that Hopf in 1940 (see [17]) used
cohomology to prove that certain axial maps do not exist. Much later, some
generalized cohomology theories were used to prove nonimmersion results
(see [2], [9], [5]).
One can construct axial maps using nonsingular bilinear maps.

Definition 2.1 We define a bilinear map

f : Rn+1 × Rn+1 −→ Rn+k+1

to be nonsingular if f (x, y) = 0 occurs only when x = 0 or y = 0.

We see that this notion is a kind of generalization of the product without

zero divisors. For example, one can take k = 0 only if n = 0, 1, 3 or 7 in
which case we get the usual products for real numbers, complex numbers,
quaternions and octonions. The relevance of this notion to our immersion
problem stems from the following theorem (see e. g. [11]).

Theorem 2.1 Existence of a bilinear map f : Rn+1 × Rn+1 −→ Rn+k+1

which is nonsingular, implies RP n # Rn+k .

Actually, one can show that the existence of nonsingular bilinear map im-
plies existence of a map which is homotopic to the corresponding axial
map. But, the previous theorem can be proved without the help of axial
maps. This is related to the spaces of matrices and will be dealt with in
the following section.
Various methods were used to construct such maps. For example, we have
the following general result:

RP n # R2n−α(n)−k(n) ,
430

where k(n) = 0, 1 or 4, depending on n (see [23], [20]). As one can see,

this result is pretty close to the general result which holds for all manifolds.
For concrete n one has better results, but we do not intend to discuss these
results here.
Let us look at this problem from a slightly different perspective. Any
bilinear map f : Rn+1 × Rn+1 −→ Rn+k+1 naturally induces a linear map
L : Rn+1 ⊗ Rn+1 −→ Rn+k+1 . Since Rn+1 ⊗ Rn+1 ∼ = M (n + 1, n + 1, R)
(we denote by M (m, n, R) the space of all m × n matrices over R), we can
reformulate our question as follows:

Q: What is the highest dimension of the vector subspace of M (n, n, R)

which does not contain matrices of rank 1?

Namely, if V is any subspace not containing matrices of rank 1, the pro-

jection L : Rn+1 ⊗ Rn+1 −→ M (n + 1, n + 1, R)/V gives us a nonsingular
bilinear map. This brings us to the discussion of spaces of matrices.

3. Spaces of matrices
Let us first introduce some notation. We denote by LF (m, n; k) the maximal
dimension of a subspace of M (m, n, F), seen as a vector space over the real
numbers, consisting, in addition to the zero matrix, only of matrices of rank
at least equal to k (F is either R, C or H, where H denotes the skew-field
of quaternions). Similarly, lF (m, n; k) stands for the maximal dimension of
the subspace of the above mentioned space consisting only of matrices of
rank equal to k (and the zero matrix of course).
If we look at the problem of finding subspaces consisting of matrices of rank
at least equal to k we mention two results. Handel ([14]) showed that if
k = 2, one can find such a subspace which has a basis consisting only of
rank 2 matrices. Petrović ([25]) has generalized this result and has proved
that it is true for all k.
How about matrices of fixed rank? Let us first look at matrices of high
rank. In the paper [3] the authors have shown that the following is true:

lR (n, n; n) = ρ(n);
lC (n, n; n) = 2ν2 (n) + 2;
n
lH (n, n; n) = ρ( ) + 4.
2
Before we proceed, let us explain this notation. The Radon-Hurwitz number
 431

ρ(n) (see [3], [27], [18]) is defined as follows:

ρ((2m + 1)24a+b ) = 8a + 2b , a ≥ 0, 0 ≤ b ≤ 3.
Of course, if n is odd, we take ρ(n/2) = 0. The notation ν2 (n) stands for
the highest power of 2 dividing n. The paper [3] refers to the famous Adams
paper [1]. In this paper Adams has proved that the maximum number of
linearly independent vector fields on S n−1 is ρ(n) − 1. In this paper Adams
has used the full machinery of multiplicative structure as well as operations
in K-theory to prove this result. At the end of the correction to their
paper [3](see [4]) the authors pose a question to relate their results to some
topological phenomena.
In my thesis [25] and the paper [26], I have used only the additive structure
of real K-theory as well as symplectic K-theory to prove these results. Since
the Bott periodicity is inherent in the relation between real and symplectic
K-theory, this shows that it is also responsible for this algebraic phenomena.
Various people have worked on the problem of determining the values of
lF (m, n; k) and LF (m, n; k). Let us present a general method for deal-
ing with this kind of problem. Suppose we have matrices A1 , . . . , Ar in
M (m, n, R). We define a bundle map
f
nξr−1 −→ εm ,
where by ξr−1 we denote the canonical line bundle over RP r−1 and by εm
the trivial m-bundle over the same space, as follows
 
λ1
 .. 
f ([x]; λ1 x, . . . , λn x) = ([x]; (x1 A1 + · · · + xr Ar )  . ),
λn
where [x] is the class of x = (x1 , . . . , xr ) ∈ S r−1 in RP r−1 . This construc-
tion was introduced for the first time in [31] for the complex case, and later
used in [32], [24], [21], [25] for the complex and the real case (we concentrate
in our presentation to the real case).
If our matrices A1 , . . . , Ar all belong to a vector subspace of matrices of
constant rank k and are linearly independent, then Im(f ) becomes a vector
bundle of dimension k we have the following splittings
Im(f ) ⊕ ν m−k = εm ; Im(f ) ⊕ η m−k = nξr−1 .
To extract further information from these splittings, Stiefel-Whitney classes
and real K-theory have been used in the previously mentioned papers. Of
432

course, in order to find subspaces of matrices of constant rank, one does not
use algebraic topology or algebraic geometry — one constructs such spaces
directly. These methods can only show which subspaces may not exist.
Meshulam (see [24]) has shown that lR (n, n; k) ≤ n. In [26] one may find
that lR (n, n; 2) = 2[n/2] if n 6= 3 while lR (3, 3; 2) = 3. Lam and Yiu have
used topological K-theory to determine lR (n, n; n − 1), lR (n, n − 1; n − 2)
and lR (n, n; n − 2) (for details, see [21]).
Sylvester has shown that LC (n, n; n − 1) = 4 (see [31]), while in [28] one
finds that LR (n, n; n − 1) = 4 if n = 4k + 2. It is shown in [26] that for all
n one has LR (n, n; n − 1) ≤ 2ν2 (n) + 1. For the results concerning L(n, n; k)
the authors have used some rudimentary algebraic geometry. While the
previously mentioned inequality can certainly be improved, one cannot hope
for the nice result for the real case as the one we have for the complex one.
Namely, it is clear that LR (4k +1, 4k +1; 4k) ≥ lR (4k +1, 4k +1; 4k) = ρ(4k)
(see [21]). Therefore, LR (n, n; n − 1) can be arbitrarily large for n of the
form 4k + 1. The reader should consult papers mentioned in the references
below in order to get further information.
Let us, at the end of this section, show how one can prove the existence of
an immersion of RP n into Rn+k from the existence of a nonsingular bilinear
map f : Rn+1 × Rn+1 −→ Rn+k+1 . We follow [25].
Any nonsingular bilinear map f : Rn+1 ×Rn+1 −→ Rn+k+1 induces a linear
map L : Rn+1 −→ Hom(Rn+1 , Rn+k+1 ) defined by L(x)(y) := f (x, y). This
map is mono and for any x 6= 0 L(x) is also mono (these facts follow from
the definition of a nonsingular bilinear map). We conclude from this the
following inequality lR (n + 1, n + k + 1; n + 1) ≥ n + 1. Using the previous
results, one sees that there exist the following splitting
εn+k = (n + 1)ξn ⊕ ν k .
Since (n + 1)ξn ∼
= τ (RP n ) ⊕ ε1 . it follows that τ (RP n ) has a stable normal
k-bundle and so by a theorem of Hirsch (see [15]), RP n can be immersed
into Rn+k .

4. Recent developments
Suppose that we would like to be more stringent and instead of looking
for arbitrary nonsingular bilinear maps, we look for the symmetric ones.
Of course, by a symmetric bilinear map f we understand a map f such
that f (x, y) = f (y, x). Actually, people have used these maps (see [16])
to construct embeddings of real projective spaces into Euclidean ones. But
 433

some recent developments shows that they can be used to get some more
interesting results.
In [12] and [13], the notion of totally skew embeddings have been introduced.
Namely, if we have an embedding M ,→ RN , then this embedding is totally
skew if for all x 6= y in M and lines l in Tx M and l0 in Ty M one has: l
and l0 are skew (as lines in RN ). For example, the embedding of R into R3
given by the map x 7→ (x, x2 , x3 ) is totally skew, while one has the result
that there are no totally skew embeddings of S 1 into R3 . In [12] one finds
the following
Theorem 4.1 If we denote by N (M n ) the smallest dimension N of the
Euclidean space for which there exists a totally skew embedding of M n into
RN , then 2n+1 ≤ N (M n ) ≤ 4n+1. If M n is closed, then N (M n ) ≥ 2n+2.
The reader may ask for the relevance of these results to our general theme
concerning nonsingular bilinear maps. Here it is. In the same paper, the
authors have proved that if B : Rn+1 × Rn+1 −→ Rm is a symmetric
nonsingular bilinear map, then one totally skew embedding S n ,→ Rn+1 ×
Rm is given by x 7→ (x, B(x, x)).
So, it is interesting to ask what do we know about the existence of sym-
metric nonsingular bilinear maps. Let us give a simple example. We know
that if we multiply two polynomials with real coefficients we get a zero
polynomial only if one of the factors is zero. But this multiplication gives a
symmetric bilinear map (look at the coefficients of these polynomials) and
this result shows that this map is nonsingular. So, we have a symmetric
nonsingular bilinear map m : Rn+1 × Rn+1 −→ R2n+1 . If we use polyno-
mials with complex coefficients and see C as R2 we can show that there
exist a symmetric nonsingular bilinear map m : R2k × R2k −→ R4k−2 . This
example looks rather simple. The reader might wonder why we discuss such
a simple example at all. Well, for the very simple reason: it is not known
whether this result can be improved — it is not known whether there exist
a symmetric nonsingular bilinear map g : R2k × R2k −→ R4k−3 . So, here is
certainly something to think about.

References

1. J. F. Adams, Vector fields on spheres, Ann. of Math. 75 (1962), 630–632.

2. L. Astey and D. Davis Nonimmersions of real projective spaces implied byBP,
Bol. Soc. Mat. Mex. 24 (1979), 49–55.
3. J. F. Adams P. Lax and R. Phillips, On matrices whose real linear combina-
tions are nonsingular, Proc. Amer. Math. Soc. 16 (1965), 318–322.
434

4. J. F. Adams, P. Lax and R. Phillips, Correction to “On matrices whose real

linear combinations are nonsingular”, Proc. Amer. Math. Soc 17 (1966),
945–947.
5. R. Bruner, D. Davis, and M. Mahowald, Nonimmersions of real projective
spaces implied by tmf, Contemp. Math. AMS 293 (2002), 45-68.
6. E. H. Brown Jr. and F. P. Peterson, Relations among characteristic classes
I, Topology 3 (1964), 39–52.
7. E. H. Brown Jr. and F. P. Peterson, On immersions of n-manifolds, Adv. in
Math. 24 (1977), 74–77.
8. E. H. Brown Jr. and F. P. Peterson, A universal space for normal bundles of
n-manifolds, Comment. Math. Helv. 54 (1979), 405–430.
9. D. Davis, A strong nonimmersion theorem for real projective spaces, Ann. of
Math. 120 (1984), 517–528.
10. R. Cohen, The immersion conjecture for differentiable manifolds, Ann. of
Math. 122 (1985), 237–328.
11. S. Gitler, Two problems studied by Heinz Hopf, Lecture Notes in Mathematics
279 (1972), Springer-Verlag, 134–174.
12. M. Ghomi and S. Tabachnikov, Totally skew embeddings of manifolds,
preprint.
13. M. Ghomi, Tangent bundle embeddings in Euclidean space , to appear in
Comment. Math. Helv.
14. D. Handel, On subspaces of tensor products containing no elements of rank
one, J. of Algebra 14 (1970), 523–527.
15. M. W. Hirsch, Immersions of manifolds, Trans. Amer. Math. Soc. 93 (1959),
242–276.
16. H. Hopf, Systeme symmetrischer Bilinearformen und euklidische Modelle der
projektiven Raüme, Vjschr. naturf. Ges. Zurich 85 (1940), 165–177.
17. H. Hopf, Ein topologischer Beitrag zur reellen Algebra, Comment. Math. Helv.
13 (1940–41), 219–239.
18. A. Hurwitz, Über der Komposition der quadratischer Formen, Math. Ann.
88 (1923), 1–25.
19. I. M. James, On the immersion problem for real projective spaces, Bull. Amer.
Math. Soc. 69 (1963), 231–238.
20. K. Y. Lam, Construction of some nonsingular bilinear maps, Bol. Soc. Mat.
Mex. 13 (1968), 88–94.
21. K. Y. Lam and P. Yiu, Linear spaces of real matrices of constant rank, Linear
Algebra Appl. 195 (1993), 69–79.
22. W. S. Massey, On the Stiefel-Whitney classes of a manifold, Amer. J. Math.
82 (1960), 92–102.
23. R. J. Milgram, Immersing projective spaces, Annals of Math. 85 (1967), 473–
482.
24. R. Meshulam, On k-spaces of real matrices, Linear and Multilinear Algebra
26 (1990), 801–803.
25. Z. Z. Petrović, On spaces of matrices satisfying some rank conditions, Ph. D.
Thesis, The Johns Hopkins University, Baltimore, 1996.
26. Z. Z. Petrović, On nonsingular matrices and Bott periodicity, Publ. Inst.
 435

Math. 65 (79) (1999), 97–102.

27. J. Radon, Lineare Scharen orthogonalen Matrizen, Abh. Math. Sem. Univ.
Hamburg 1 (1922), 1–14.
28. E. G. Rees, Linear spaces of real matrices of given rank, Contemporary Math-
ematics 188 (1995), 219–229.
29. B. J. Sanderson, A nonimmersion theorem for real projective spaces, Topology
2 (1963), 209–211.
30. S. Smale, The classification of immersions of spheres in Euclidean spaces,
Ann. Math. 69 (1959), 281–290.
31. J. Sylvester, On the dimension of spaces of linear transformations satisfying
rank conditions, Linear Algebra Appl. 78 (1986), 1–10.
32. R. Westwick, Spaces of matrices of fixed rank, Linear and Multilinear Algebra
20 (1987), 171–174.