5
th of April, 2012Large spaes of matries with
bounded rank
D r
Clément de Seguins Pazzis
Lyée Sainte-Geneviève, Versailles, Frane
Université de Versailles-Saint-Quentin-en-Yvelines
dsp.profgmail.om
http://dsp.prod.free.fr/reherhe.html
I. Historial results.
II. Flanders's method.
III. Proof of the generalized Atkinson-Lloyd
theorem.
IV. Appliation to an invertibility preserver
problem.
K
denotes an arbitrary (ommutative) eld (possibly nite),n
andp
two positive integerswith
n ≥ p
,M n,p ( K )
the vetor spae of matrieswith
n
rows,p
olumns and entries inK
, andM n ( K ) := M n,n ( K )
.Two linear subspaes
V
andW
ofM n,p ( K )
areequivalent when
∃(P, Q) ∈ GL n ( K ) × GL p ( K )
suh that
W = P V Q
.The upper rank of a subset
V
ofM n,p ( K )
isrk(V ) := max
rk M | M ∈ V .
Problem : given an integer
r > 0
withr < n
andr < p
, lassify the linear subspaes ofM n,p ( K )
with upper rank
r
.For small values of
r
, the solution is known(Shur:
r = 1
; Atkinson:2 ≤ r ≤ 3
). For generalvalues, no fully enompassing answer is known.
with upper rank
r
.(ii) Classify those subspaes with a dimension
lose to the maximal one.
A. Maximal dimension
Theorem 1 (Flanders, 1962). Let
r ∈ [[1, p − 1]]
and
V
be a linear subspae ofM n ( K )
withrk(V ) = r
. Then:(a)
dim V ≤ rn
;(b) If
dim V = rn
andn > p
, thenV
is equivalentto
h
C 1 · · · C r [0] n × ( p − r )
i | C 1 , . . . , C r ∈ M n, 1 ( K )
() If
dim V = rn
andn = p
, thenV
orV T
isequivalent to
h
C 1 · · · C r [0] n × (n − r)
i | C 1 , . . . , C r ∈ M n, 1 ( K )
The ase
# K > r
:H. Flanders, On spaes of linear transformations with
bounded rank, J. Lond. Math. So. 37 (1962) 10-16.
The ase
n = p
andr = n − 1
for an arbitraryeld (extended to ane subspaes):
J. Dieudonné, Sur une généralisation du groupe orthogonal
à quatre variables, Arh. Math. 1 (1949) 282-287.
The more general ase:
R. Meshulam, On the maximal rank in a subspae of
matries, Quart. J. Math. Oxford (2) 36 (1985) 225-229.
The more general ase extended to ane
subspaes:
C. de Seguins Pazzis, The ane preservers of non-singular
matries, Arh. Math. 95 (2010) 333-342.
The naive onjeture: a subspae
V
with upperrank
r
may be extended to one with the maximaldimension
nr
. Equivalently, ifn > p
, then thematries of
V
vanish on some ommon(p − r)
-dimensional subspae ofK p
(ifn = p
, thenthis should hold for
V
or forV T
).True for
r = 1
(Shur)! False in general!Notation 1. Given
(s, t) ∈ [[0, n]] × [[0, p]]
, onedenes
R(s, t)
as the set of alln × p
matries ofthe form
[?] s × t [?] s × ( p − t ) [?] (n − s) × t [0] (n − s) × (p − t)
.
Note that
rk R(s, t)
≤ s + t
, and equality holdswhenever
s + t ≤ p
.In partiular, if
r ≥ 1
, thenR(1, r − 1)
is aounter-example to the above onjeture, with
dimension
n(r − 1) + (p − r + 1) = nr − (n − p + r − 1)
(inpartiular:
nr − (r − 1)
ifn = p
).Atkinson and LLoyd ask: an we lassify spaes
V
with upper rankr
andnr − (n − p + r − 1) ≤ dim V ≤ nr
?Theorem 2 (Atkinson and Lloyd, 1980).
Let
r ∈ [[1, n − 1]]
etV
be a linear subspae ofM n ( K )
withrk(V ) = r
.Assume that
# K > r
anddim V ≥ nr − r + 1
.Then:
(a) Either
V
is the subspae of a linear subspaeW
ofM n ( K )
withrk(W ) = r
anddim W = nr
;(b) Or
V
is equivalent toR(1, r − 1)
orR(r − 1, 1)
.bounded rank, Quart. J. Math. Oxford (2), 31 (1980)
253-262.
The ase
n > p
:L.B. Beasley, Null spaes of spaes of matries of bounded
rank, in Current Trends in Matrix Theory 45-50, Elsevier,
1987.
What about small nite elds?
An extension of Atkinson-Lloyd to an arbitrary
eld was deemed problemati: ounter-example
for
F 2
: the spae of all upper-triangular matries with zero trae:(
a c d
0 b e
0 0 a + b
| a, b, c, d, e ∈ F 2
)
ts into neither of the above ategories.
Theorem 3 (de Seguins Pazzis, 2010). Atkinson
and Lloyd's theorem holds for an arbitrary eld,
exept when
n = 3
,r = 2
,# K = 2
anddim V = 5
,in whih ase the above ounter-example is the
only exeption up to equivalene.
C. de Seguins Pazzis, The lassiation of large spaes of
matries with bounded rank, arXiv:
http://arxiv.org/abs/1004.0298
Beasley's results are also suessfully extended to
an arbitrary eld (no exeption!).
To simplify things, we take
n = p
.A. The original method.
Let
V
a linear subspae ofM n ( K )
withrk(V ) = r ∈ [[1, n − 1]]
. One may assume thatV
ontains
J r :=
I r [0] r × (n − r) [0] ( n − r ) × r [0] ( n − r ) × ( n − r )
.
In that ase, we split every matrix of
V
with thesame blok sizes:
M =
A(M ) C (M ) B (M ) D(M )
.
Assume that
# K > r
and letM ∈ V
. Now, givenan indeterminate
t
, ompute the(n − r) × (n − r)
matrix
P (t)
of all(r + 1) × (r + 1)
minors ofJ r + tM
where one takes all the rstr
rows andolumns:
→ P (t)
is a polynomial oft
with oeients inM n − r ( K )
and degree≤ r
;→ P (t)
vanishes everywhere onK
, thereforeP (t) = 0
;→
Its oeient beforet r
isD(M )
; thusD(M ) = 0 ;
→
Consequently, its oeient beforet r − 1
is−B (M )C (M )
; thusB (M )C (M ) = 0.
From
∀M ∈ V, B(M )C (M ) = 0
, one easily provesthat
dim B (V ) + dim C (V ) ≤ r(n − r)
(e.g.B (V ) × C (V )
is totally isotropi for thenon-degenerate quadrati form
(B, C ) 7→ tr(BC )
).Therefore:
dim V ≤ dim A(V ) + dim B (V ) + dim C (V ) ≤ nr.
Assume now
dim V = nr
. Then (rank theorem)V
ontains, for every invertible
P ∈ M n ( K )
,M P :=
P [0] r × ( n − r ) [0] ( n − r ) × r [0] ( n − r ) × ( n − r )
.
With the above method, one replaes
J r
withM P
and nds:
∀M ∈ V, B (M )P − 1 C (M ) = 0.
Then varying
P
yields:∀M ∈ V, B(M ) = 0
orC (M ) = 0.
One easily onludes: either
B (V ) = {0}
orC (V ) = {0}
. This yields the Flanders theorem.Same starting point, assumptions (exept on the
ardinality of
K
) and notations. Also, instead ofassuming that
V
ontainsJ r
, we simply assumethat it ontains
Q [0] r × ( n − r ) [0] ( n − r ) × r [0] ( n − r ) × ( n − r )
for some invertible
Q ∈ M r ( K )
.We take an arbitrary matrix of
M n ( K )
M 0 =
P C (M 0 ) B (M 0 ) D(M 0 )
.
We atually do not ompute the matrix of minors
of a general sum
M 0 + M
withM ∈ V
(tooompliated!). Rather, we onsider a tower of
linear subspaes of
V
:V 3 ⊂ V 2 ⊂ V 1 ⊂ V.
V 1 := Ker A
V 2 := Ker A ∩ Ker B
V 3 = Ker A ∩ Ker B ∩ Ker C.
The rank theorem shows:
dim V = dim A(V )+dim B (V 1 )+dim C (V 2 )+dim D(V 3 ).
One then omputes the matrix of all the
(r + 1) × (r + 1)
minors ofM 0 + M
whih use ther
rst rows and olumns, forM ∈ V 1
: this isdet(P )D(M 0 + M )
− B (M 0 + M ) com(P ) T C (M 0 + M ).
This matrix is
0
, and so is the one forM = 0
.Subtrating both equalities yields the
fundamental formula:
B (M ) com(P ) T c(M ) = det(P )D(M )
− B (M ) com(P ) T c(M 0 )
− B (M 0 ) com(P ) T c(M )
As
P
may be hosen invertible, this yields:D(V 3 ) = 0.
Taking the polar form of the left-hand side yields,
for every
(M, N ) ∈ V 1 2
:B (M ) com(P ) T c(N ) + B (N ) com(P ) T c(M ) = 0.
In partiular, for every
(M, N ) ∈ V 1 × V 2
:B (M ) com(P ) T c(N ) = 0.
In partiular
∀(M, N ) ∈ V 1 × V 2 , B(M )Q − 1 C (N ) = 0
, andtherefore
dim B (V 1 ) + dim C (V 2 ) ≤ r(n − r).
This yields
dim V ≤ nr
.For the ase
dim V = nr
, the proof goes on usingthe above formulas ...
generalized Atkinson-Lloyd
theorem.
We ontinue with the assumptions of the latest
setion. We now assume
dim V > nr − r + 1
.A. Step 1: showing that
B ( V 1 ) = 0
orC ( V 2 ) = {0}
.We show that either there is no non-zero matrix
in
V
of the form
[0] r × n
[?] (n − r) × n
or there is no non-zero matrix in
V
of the formh
[0] n × r [?] n × ( n − r ) i
.
Using the above inequalies, we nd
dim A(V ) > r 2 − (r − 1).
∀P ∈ A(V ), ∀(M, N ) ∈ V 1 × V 2 ,
B (M ) com(P ) T C (N ) = 0.
Lemma 4. One has
B (V 1 ) = {0}
orC (V 2 ) = {0}
.Proof. Assume the ontrary holds. Then
span
{com(P ) T | P ∈ A(V )}
does not attransitively on
K r
. Without loss of generality, onemay then assume that
∀P ∈ A(V ), com(P ) r,r = 0.
This ontradits the Flanders theorem!!
If
B (V 1 ) = {0}
, then forW := V T
, one hasC (W 2 ) ⊂ B (V 1 ) T = {0}
. No generality is thenlost by assuming that
C (V 2 ) = {0}
. ThenV 2 = V 3 = {0}
.We now assume
V 2 = {0}
. In that ase, there is alinear subspae
W ⊂ M n,r ( K )
and a linear mapϕ : W → M n,n − r ( K )
s.t.V = h
M ϕ(M )
i | M ∈ W
.
Note that
codim W < r − 1
.From there, we prove two lemmas.
Lemma 5. One has
∀N ∈ W, Im ϕ(N ) ⊂ Im(N ).
Lemma 6. There exists
C ∈ M r,n − r ( K )
s.t.∀N ∈ W, ϕ(N ) = N C
.With the last one, take
P :=
I r −C [0] ( n − r ) × r I n − r
and hek that any matrix of
V P
has all lastn − r
olumns equal to zero.Proof of Lemma 5: Denote by
U
the spae ofmatries of
W
with0
as last olumn. ForM ∈ U
,write
h
M ϕ(M ) i =
K (M ) [?] (n − 1) × (n − r) [0] 1 × r ϕ n (K (M )).
Note that:
→ codim K (U ) < r − 1 < n − 2
;→
ifrk K (M ) = r
, thenϕ n (K (M )) = 0
.→ ϕ n : K (U ) → M 1,n − r ( K )
is linear.We dedue that
ϕ n = 0
by using the followingorollary of our extension of Flander's theorem to
ane subspaes:
Corollary 7. Let
V ′
be a linear subspae ofM n,p ( K )
, withn ≥ p ≥ r
, and assume thatdim V ′ > rn
. ThenV ′
is spanned by its matriesof rang
≥ r
, unlessn = p = 2
,r = 1
and# K = 2
.a linear hyperplane
H
ofV ′
whih ontains everyrank
≥ r
matries ofV ′
, and then a paralleldisjoint ane hyperplane
H ′
, whih then hasdimension
≥ rn
, has upper rank< r
and is not alinear subspae of
M n,r ( K )
.Bak to the proof of Lemma 5: we now have
ϕ n = 0
. This means that forH = K n − 1 × {0}
,∀M ∈ W, Im M ⊂ H ⇒ Im ϕ(M ) ⊂ H
.However,
H
may be replaed by any linearhyperplane of
K n
. The onlusion follows.Lemma 6 may be extended as some kind of linear
preserver theorem:
Theorem 8 (Representation lemma). Let
m, n, p
be positive integers,
W
be a linear subspae ofM m,n ( K )
withcodim W ≤ m − 2
, andϕ : W → M m,p ( K )
be a linear map s.t.∀M ∈ W, Im ϕ(M ) ⊂ Im M.
Then there exists
C ∈ M n,p ( K )
s.t.ϕ : M 7→ M C.
Remarks 1.
•
The aseW = M m,p ( K )
is an easy exerise.•
One need only assumep = 1
.•
The onditioncodim W ≤ m − 2
is notoptimal. When
# K > 2
andn ≥ 2
, the bestupper bound is
2m − 3
(work in progress ...).The proof is a bit too tehnial for a seminar talk
...
invertibility preserver problem.
The following theorem is the rst appliation of
the Atkinson-Lloyd theorem.
Theorem 9 (de Seguins Pazzis, 2012). Let
V
bea linear subspae of
M n ( K )
with odimension≤ n − 2
, andf : V ֒ → M n ( K )
be a linearembedding whih is a strong invertibility
preserver. Then, unless
n = 3
,# K = 2
andcodim V = 1
, there exists(P, Q) ∈ GL n ( K ) 2
s.t.∀M, f (M ) = P M Q
or∀M, f (M ) = P M T Q
If
K
is innite orf (V ) = V
, then one need onlyassume that
f
is a weak invertibility preserver.This is a grand generalization of the well-known
Dieudonné theorem (ibid). The method is similar,
with the Atkinson-Lloyd theorem as a key
starting point. The above representation theorem
is also a ruial point.
non-singularity in a large spae of matries, Linear
Algebra Appl. 436-9 (2012), 3507-3530.
Remark 2. In the ase
n > p
, invertibility is naturally replaed with the assumptionrk M = p
i.e.M
has full rank: similar resultshold (urrently being written).