Boletı́n de Matemáticas

Nueva Serie, Volumen XI No. 1 (2004), pp. 51–68

FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A

FUNCTION

DAVID MOND (*)

A la memoria de Jairo Charris, mi profesor

1. Introduction: a determinantal factorisation of D4

Let us begin with a simple example. Consider the 2-variable family of matrices
 
x2 0 0
M (x1 , x2 ) =  0 x1 x2 
0 x2 x1
Its determinant is the well known D4 singularity, f (x1 , x2 ) = x21 x2 − x32 . So
M det
(C2 , 0) −→ Mat3 (C) −→ C
is a factorisation of D4 . A deformation of M induces a deformation of its determi-
nant. The deformation theory of D4 is very well known: a miniversal deformation
is given by
F (x1 , x2 , u1 , u2 , u3 , u4 ) = x21 x2 − x32 + u1 x22 + u2 x2 + u3 x1 + u4 .
However, deformations of M induce only a subset of the possible deformations
of D4 ; it turns out that every deformation of D4 obtained by deforming M is
parametrised-right-equivalent to one induced from the deformation
F1 (x1 , x2 , v1 , . . . , v7 ) = x21 x2 − x32 +
v1 x21 − (v1 + v6 )x22 + v5 x1 x2 + (v1 v5 − v3 v7 − v2 v4 )x1
(1)
+(v2 v7 − v1 v6 + v3 v4 )x2 + (v2 v6 − v3 v5 )v7 .

(*) David Mond. Mathematics Institute, University of Warwick, Coventry CV4 7AL, England
E-mail: [email protected] .
51
52 DAVID MOND

Deformation of M within the (smaller) class of symmetric matrices leads to an

even more restricted class of deformations of its determinant, all of which are
parametrised-right-equivalent to deformations induced from the deformation
F2 (x1 , x2 , w1 , . . . , w4 ) = x21 x2 −x32 +
w1 (x21 − x22 ) + (w1 w4 − w22 − w32 )x1 + 2w2 w3 x2 + w4 x1 x2 − w32 w4 . (2)
These assertions are proved by introducing suitable notions of equivalence, Kdet
s
and Kdet , for matrix and symmetric matrix families like M - both of which are, in
fact, liftings, of right equivalence of determinants. These notions, due principally
to Jim Damon (though see [3] for the precise version given below), are recalled
in Section 3 below. The deformations F1 and F2 of det M are the determinants
s
of versal deformations of M — versal, that is, with respect to Kdet - and Kdet -
equivalence. As these two matrix deformations we may take, respectively,
 
x2 + v1 v4 v7
M1 (x, v) =  v2 x1 + v5 x2 
v3 x2 + v6 x1
and  
x2 + w1 w2 w3
M2 (x, w) =  w2 x1 + w4 x2 
w3 x2 x1
By definition of the notion of versal deformation for right-equivalence, F1 and
F2 are parametrised-right-equivalent to deformations induced from F . That is,
denoting by B, B1 and B2 the base-spaces (C4 , 0), (C7 , 0) and (C4 , 0) of F , M1
and M2 respectively, there are maps ρ1 : B1 → B and ρ2 : B2 → B such that F1
and F2 are parametrised-right-equivalent, respectively, to the deformations ρ∗1 (F )
and ρ∗2 (F ) given by
ρ∗1 (F )(x, v) = F (x, ρ1 (v))
and
ρ∗2 (F )(x, w) = F (x, ρ2 (w)).
By inspection of the formulae for M1 and M2 , one can see that there is a map
ρ21 : B2 → B1 such that
M2 (x, w) = M1 (x, ρ21 (w));
in fact the definition of Kdet -versality guarantees the existence of a map ρ21 such
that M2 is parametrised Kdet -equivalent to ρ∗21 (M1 ).
So our factorisations of D4 , and their asociated deformation theories, have given
rise to a triangle of base spaces and inducing maps
ρ21
B2 −→ B1
ρ1 & . ρ2 (3)
B
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 53

Question: Are such triangles commutative?

This paper is concerned with this and other related questions. Our main result,
in Section 6, is an infinitesimal condition for the commutativity of such diagrams.
We begin by placing these questions in a more general context, in the category of
factorisations of a function. The questions about commutativity of these diagrams
become questions about the functoriality of certain maps from this category to the
categories of OX /Jg -modules and algebras.

2. The category of factorisations

F f
Fix a map g : X → Z. The set of all factorisations X −→ Y −→ Z of g (i.e.,
for which f ◦ F = g) forms a category, in which a morphism
1 F f1 2 F f2
α : (X −→ Y1 −→ Z) → (X −→ Y2 −→ Z)
is a map α : Y1 → Y2 giving rise to a commutative diagram
X
F1 . & F2
α
Y1 −→ Y2
f1 & . f2
Z
Since in this case F2 = α ◦ F1 and f1 = f2 ◦ α, morphisms correspond to the
re-bracketing of triple factorisations of g: the morphism above can be written
α : (f2 ◦ α, F1 ) → (f2 , α ◦ F1 ),
where for brevity we are representing each factorisations as an ordered pair of
maps.

We will denote the category of factorisations of g by F(g).

Example 2.1. Let Symn (C) be the space of symmetric n × n matrices over C.
Then the previous section is concerned with the morphism
(C2 , 0)
M. &M
i
Sym(C) −→ Mat3 (C)
det & . det
C
where i is inclusion.
Remark 2.2. Let M be an n × n matrix whose entries are functions and let
g = det M . Let M ∗ be the adjugate matrix of signed submaximal minors of
M . Then M M ∗ = M ∗ M = gIn . Any pair of matrices (M, N ) such that M N =
N M = gIn is known in commutative algebra and representation theory as a matrix
54 DAVID MOND

factorisation of g. Matrix factorisations correspond to maximal Cohen-Macaulay

modules on the hypersurface {g = 0}, and are extensively studied, e.g. in [4],[9].
However, the morphisms of matrix factorisations referred to by algebraists in this
context are not morphisms in our sense. For this reason, here we avoid the term
‘matrix factorisation’.
Example 2.3. If there is a morphism α : (f1 , F1 ) → (f2 , F2 ) then F2 = α ◦ F1
and f1 = f2 ◦ α. It follows that F1∗ (Jf1 ) ⊂ F2∗ (Jf2 ). For
F1∗ (∂f1 /∂xj ) =
P

k ((∂f2 /∂yk ) ◦ α)∂αk /∂xj ◦ F1
P


= k (∂f2 /∂yk ) ◦ α ◦ F1 ∂αk /∂xj ◦ F1

∈ F2∗ (Jf2 )
P


= k (∂f2 /∂yk ) ◦ F2 ∂αk /∂xj ◦ F1
We will refer to the natural algebra-homomorphism O/F1∗ (Jf1 ) → O/F2∗ (Jf2 ) which
results from this inclusion as the comparison map.
Proposition 2.4. The map C : F(g) → O/Jg -algebras taking (f, F ) to
O/F ∗ (Jf ), and taking a morphism α : (f1 , F1 ) → (f2 , F2 ) to the comparison map
O/F1∗ (Jf1 ) → O/F2∗ (Jf2 ), is a covariant functor F(g) → O/Jg -algebras. 2
Here and where possible in the rest of the paper, we use O in place of OX ,
where X is the domain of the map g whose factorisations we are considering.
Example 2.5. As initial and final objects in F(g) we have the factorisations
id m g
(Cm , 0) −→
C
(Cm , 0) −→ (C, 0)
and
g id
(Cm , 0) −→ (Cn , 0) −→
C
(C, 0),
since for any other factorisation f ◦ F = g there are unique diagrams
Cm Cm
idCm . &F F . &g
α α∞
Cm 0
−→ Cn and Cm −→ C
g& .f f& . idC
C C
in which α0 = F and α∞ = f . We have C(g, idCm ) = O/Jg and C(idC , g) = 0.
Thus C takes the initial and final objects of F(g) to the initial and final objects of
the category of O/Jg -algebras.

I am grateful to Ignacio de Gregorio, Daan Krammer, Victor Goryunov and

Jan Stevens for helpful conversations on the topics of this paper.
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 55

3. KV and Kf -equivalence
In [1], Jim Damon generalised the notion of contact equivalence of map-germs
(X, x0 ) → (Y, y0 ), as follows.
Definition 3.1. Given the germ of a subvariety (V, y0 ) ⊂ (Y, y0 ) the map-germs
F, G : (X, x0 ) → (Y, y0 ), are KV -equivalent if there are diffeomorphisms Φ of
(X × Y, (x0 , y0 )) and ϕ of (X, x0 ) such that
(1) πX ◦ Φ = ϕ ◦ πX , (i.e. Φ lifts ϕ)
(2) Φ takes X × V to itself, and
(3) Φ induces a diffeomorphism graph(F ) → graph(G).
Here πX : X × Y → X is projection. In the analytic category, “diffeomor-
phism” means bianalytic automorphism. Ordinary contact equivalence becomes
KV -equivalence with V = {y0 }.
In the case that V is a hypersurface in Y with equation f , the closely related
group of equivalences Kf was introduced in [3]; the definition is that of KV except
that (2) is replaced by the condition
2 0 . f ◦ πY ◦ Φ = f ◦ πY .
Lemma 3.2. (i) If F and F 0 are Kf -equivalent then f ◦ F and f ◦ F 0 are right-
equivalent. (ii) If F and F 0 are KV -equivalent then f ◦ F and f ◦ F 0 are contact-
equivalent.
Proof By (1) and (3) of the definition,
Φ(x, F (x)) = (ϕ(x), F 0 (ϕ(x))). (7)
Now (2’) implies
f (F (x)) = f ◦ πY (Φ(x, F (x)) = f ◦ πY (ϕ(x), F 0 (ϕ(x))) = f (F 0 (ϕ(x))),
i.e. f ◦ F and f ◦ F 0 are right-equivalent, while (2) implies
F −1 (V ) = (gr(F ))−1 (X × V ) = (gr(F ))−1 (Φ−1 (X × V ))
which is equal to (F 0 ◦ ϕ)−1 (V ) by (7). Thus (f ◦ F )−1 (0) = (f ◦ F 0 ◦ ϕ)−1 (0),
from which contact-equivalence follows. 2

In 4.4 below we show that the converse to both parts of this Lemma is in general
false.

Damon showed that when X and Y are smooth, the tangent space to the KV -
orbit of a germ F is
T KV F = tF (θX,x0 ) + F ∗ (Der(− log V )),
where Der(− log V ) is the OY,y0 -module of germs of vector fields on Y which are
tangent to V at its smooth points.
56 DAVID MOND

The expression for T Kf F is the same as that for T KV F except that

Der(− log V ) is replaced by Der(− log f ), the OY -module of vector fields tangent
to all of the level sets of f .
Our concern with these equivalence relations and deformation theories is mo-
tivated by the fact that many naturally occuring maps and functions arise, and
are naturally deformed, as composites. The discriminant hypersurface Df (set of
critical values) of a map-germ f : (Cn , 0) → (Cp , 0) with n ≥ p, is a ‘non-linear
section’ of the discriminant DF of a stable unfolding F , and as equation for Df we
can take the composite of an equation h for DF with a suitable ‘base-change’ map
i. In fact Damon shows in [2] that the A-codimension of f is equal to the KDF -
codimension of i, and in [3] Damon and the author showed that the discriminant
of a stable perturbation of f has homotopy type of a wedge of (p − 1)-spheres,
whose number is equal to the Kh -codimension of i.

The T 1 ’s for the deformation theory of the germ F under Kf and KV -equivalence
have usually been denoted by TK1 f F and TK1 V F . In keeping with our change of
viewpoint in this paper, we will call them T 1 (f, F ) and Te1 (f, F ) instead. That is,
θ(F ) θ(F )
T 1 (f, F ) = , Te1 (f, F ) = .
T Kf F T KV F
Example 3.3.
T 1 (g, idCm ) = 0 = Te1 (g, idCm ).
T 1 (idC , g) = OX /Jg , Te1 (idC , g) = OX /(g + Jg ).
F f
Proposition 3.4. For any factorisation X −→ Y −→ C of g, T 1 (f, F ) and
Te1 (f, F ) are O/Jg -modules.
Proof Let x1 , . . . , xm and y1 , . . . , yn be variables on X and Y respectively. We
have to show that for each i and j,
∂g ∂
∈ T Kf F.
∂xi ∂yj
This is easy: since Der(− log f ) contains the Hamiltonian vector fields
∂f ∂ ∂f ∂
χk,` (f ) := − ,
∂yk ∂y` ∂y` ∂yk
we have
    
∂g ∂ X ∂f ∂Fk ∂ X ∂Fk ∂f ∂
= ◦F = χk,j (f )◦F +tF ◦F
∂xi ∂yj ∂yk ∂xi ∂yj ∂xi ∂yk ∂xi
k k6=j

2
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 57

In [5], the relation between Kf equivalence and right-equivalence of the compos-

ites, and KV -equivalence and contact-equivalence of the composites, was further
explored. We summarise the main result: let L• (f ) be a free OY -resolution of
OY /Jf , and let K• (f ◦ F ) be the Koszul complex on the first order partials of
f ◦ F . There is a morphism of complexes K• (f ) → L• (f ), and composing the
pull-back of this by F , with the canonical morphism K• (f ◦ F ) → F ∗ (K• (f )),
we get a morphism φf : K• (f ◦ F ) → F ∗ (L• (f )) lifting the comparison mor-
phism OX /Jf ◦F → OX /F ∗ (Jf ). Let C• (φf ) denote the cone of the morphism
of complexes φf . There is an analogous construction of a morphism of com-
plexes φV : K e • (f ) → F ∗ (L
e • ) lifting the comparison morphism OX /(f ◦ F +
Jf ◦F ) → OX /F ∗ ((f ) + Jf ), and of its cone, C• (φV ).
Theorem 3.5. ([5])
TK1 f F = H1 (C• (φf )), TK1 V F = H1 (C• (φV )).
2
If φ : A• → B• is a morphism of complexes, there is a long exact sequence
· · · → Hk (A• ) → Hk (B• ) → Hk (C• (φ)) → Hk−1 (A• ) → · · · ,
and thus we have
Corollary 3.6. There are canonical long exact sequences
OX OX
· · · → H1 (K• (g)) → H1 (F ∗ (L• )) → T 1 (f, F ) → → ∗ →0 (8)
Jg F (Jf )
and
e • (g)) → H1 (F ∗ (L
e • )) → Te1 (f, F ) →OX OX
· · · → H1 (K → ∗ →0
(g + Jg ) F ((f ) + Jf )
(9)
in which the maps T 1 (f, F ) → OX /Jg and Te1 (f, F ) → OX /(g + Jg ) are induced
by F ∗ (tf ) : θ(F ) → OX . 2
Note that H1 (F ∗ (L• )) = TorO
1 (OY /Jf , OX )
Y
and H1 (F ∗ (L
e • )) =
TorO
1 (OY /(f + Jf ), OX ).
Y

Remark 3.7. (i) Let g = f ◦F . Then OCm /Jg is the T 1 of g for right-equivalence.
The morphism TK1 f F → T 1 g in (8) is a map between deformation functors, telling
us which of the first order deformations of g we can get by deforming F alone. A
similar statement holds for TK1 V F → OCm /(g) + Jg ), with contact-equivalence in
place of right-equivalence.

(ii) In fact we can say more: if F : (Cm × B, 0) → (Cn , 0) is a Kf -miniversal

deformation of F , then f ◦ F is isomorphic to a deformation induced from an
Re -miniversal deformation G : (Cm × S, 0) → (C, 0) of g by an inducing map
58 DAVID MOND

1
Λ : B → S. Because of the freeness of TR/S G := OCm ×Cµ /Jrel G over OS , the
Kodaira-Spencer map gives an isomorphism
M
TSs ' π∗ (TR/S
1
G)s = 1
TR/S G(x,s)
x∈Cm

where TS is the tangent sheaf on S. By coherence, there is a neighbourhood of

0 ∈ B on which there is an epimorphism
M
TBb → π∗ (TK1 f /B F)b = TK1 f /B Fb .
x∈Cm

Then there is a commutative diagram

π∗ (TK1 f /B F)b 1


−→ π∗ TR/S G Λ(b)
↑ ↑' (10)
db Λ
TBb −→ TSΛ(b)

in which the vertical maps are the Kodaira-Spencer morphisms and the map at
the top is the relative version of the morphism TK1 f F → OCm /Jg in the long exact
sequence (8). It follows from commutativity that Λ is a submersion at precisely
those points b over which OCm /Fu∗ (Jf ) vanishes.

(iii) Where the left hand Kodaira Spencer map is an isomorphism also (e.g. for
symmetric matrix families in two and three variables, with f = det, as shown in
[5]), the relative version of the morphism TK1 f F → OCn /Jg can be identified with
the derivative of the inducing map Λ. As a consequence of this, Λ is an immersion
at those points b over which H 1 (L• (f )) (which is of course just TorO
1
Cn
(OCn /Jf ,
OCm )) vanishes.

(iv) In [6], V. V. Goryunov and V. Zakalyukin observe that for simple symmetric
matrix families M in two variables, and with f = det, the Kdet -codimension of M
is equal to the Re -codimension (= Milnor number) of det M , and that morover
the discriminant of the inducing map Λ is the image under Λ of the set of points
u such that for some x ∈ C2 , Mu (x) has corank ≥ 2. They conjecture that the
assumption of simplicity can be omitted in these statements. The first part of this
conjecture was proved in [5]. That the second part also holds follows from (ii)
above, since V (Jdet ) is precisely the set of matrices of corank ≥ 2.

4. Is T 1 a functor?
From now on we consider only T 1 = T 1 Kf . An analogous theory can be developed
for Te1 .
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 59

Suppose we are given a morphism

Cm
F1 . & F2
α
Cn −→ Cp . (11)
f1 & . f2
C
The map α gives rise to an O-linear map F1∗ (tα) : θ(F1 ) → θ(F2 ), sending d(F 1t
dt |t=0
d(α◦F1,t )
to dt . The spaces T 1 (fj , Fj ) and Te1 (fj , Fj ) are quotients of θ(Fj ) and
|t=0
one might hope that F1∗ (tα) would gives pass to the quotient and give rise to
morphisms of O-modules (and therefore of O/Jg -modules, in view of 3.4)
T 1 (α) : T 1 (f1 , F1 ) → T 1 (f2 , F2 )
and
Te1 (α) : Te1 (F1 , F1 ) → Te1 (f2 , F2 ).
Since the existence of the morphism α means that F2 = α ◦ F1 and f1 = f2 ◦ α, we
change our notation slightly, and write F for F1 and f for f2 . For the existence
of O-linear maps T 1 (α) and Te1 (α), what is required is that
F ∗ (tα)(T Kf ◦α F ) ⊆ T Kf (α ◦ F )
and
F ∗ (tα)(T Kα−1 (V ) F ) ⊆ T KV (α ◦ F ).
Clearly F ∗ (tα)(tF (θCm )) = t(α ◦ F )(θCm ) ⊆ T Kf (α ◦ F ); however, as we will see
in Example 4.3, F ∗ (tα)(F ∗ (Der(− log f ◦ α)) and F ∗ (tα)(F ∗ (Der(− log α−1 (V )))
are not always contained in T Kf (α ◦ F ) and T KV (α ◦ F ).
Example 4.1. For any factorisation (F, f ) of g, let α∞ : (f, F ) → (idC , g) be the
unique morphism to the final object (g, idC ). Then T 1 (α∞ ) and Te1 (α∞ ) are just
the maps T 1 (f, F ) → O/Jg and Te1 (f, F ) → O/(g + Jg ) in the exact sequences (8)
and (9), and therefore are OX -linear. Of course, it is easy to check this directly.
Example 4.2. Let f (x1 , x2 ) = x2 x2 and let α : C → C2 be the inclusion α(x1 ) =
(x1 , 0). Then f ◦ α = 0, so Der(− log(f ◦ α)) = θC . But there is no extension of
∂/∂x1 ∈ θC to a vector field in Der(− log f ).
Example 4.3. Define f : C4 → C and α : C3 → C4 by
f (w, x, y, z) = (z + w)x + (z − w)y + w2 + z 2 ,
α(x, y, z) = (0, x, y, z),
so that
f ◦ α(x, y, z) = xz + yz + z 2 = z(x + y + z).
Both Der(− log f ◦ α) and Der(− log α−1 (V )) contain the constant vector field
∂/∂x−∂/∂y, while f itself has an isolated (indeed non-degenerate) critical point at
60 DAVID MOND

0, so that Der(− log f ) and Der(− log V ) are contained in mC4 ,0 θ(f ). Since tα = α
is a linear injection, tα(Der(− log f ◦ α))⊆ / α∗ (Der(− log f )) and
−1 ∗
tα(Der(− log α (V )))⊆ / α (Der(− log V )).

Now define F : C2 → C3 by F (u, v) = (u2 , u2 , v), and let g = f ◦α◦F = 2u2 v +v 2 .

Then T 1 (f ◦ α, F ) is the quotient of θ(F ) by the OC2 -module T Kf ◦α F , which
contains the two linearly independent constant vector fields tF (∂/∂v) = ∂/∂z
and ∂/∂x − ∂/∂y. On the other hand, up to scalar multiples T Kf (α ◦ F ) and
T KV (α ◦ F ) contain only one constant vector field, namely t(α ◦ F )(∂/∂v). As
F ∗ (tα) is a linear injection with constant matrix,
F ∗ (tα)(T Kf ◦α F ) /⊆T KV f ◦ F,

and T 1 (α) : T 1 (f ◦ α, F ) → T 1 (f, α ◦ F ) and Te1 (α) : Te1 (f ◦ α, F ) → Te1 (f, α ◦ F )

cannot be O-linear.

A straightforward calculation shows that T 1 (f ◦α) has length 1 and T 1 (f, α◦F )
has length 5.
Remark 4.4. The previous examples make clear that the converse to Lemma
3.2 does not in general hold. In the case of Example 4.3, the family Ft (u, v) =
(u2 + t, u2 − t, v) is Kf ◦α - trivial, from which it follows, by 3.2, that f ◦ α ◦ Ft
is R-trivial. Nevertheless α ◦ Ft is not Kf -trivial, nor even KV -trivial (where
V = {f = 0}).
4.1. T 1 (α) as a C-linear map. Suppose that for j = 1, 2, T 1 (fj , Fj ) has finite
length, and let Fj be a Kfj -miniversal deformation of Fj , with base BFj . Since
α ◦ F1 is a deformation of F2 , it is isomorphic to a deformation induced from F2
by a base change map ρ1,2 : BF1 → BF2 . Although the map ρ1,2 is not unique, its
derivative at 0 ∈ BF1 is. Modulo the identification of T0 BFj with T 1 (fj , Fj ), it is
the map
dF1,t d(α ◦ F1,t )
7→ .
dt |t=0 dt |t=0

In other words, it is induced by F1∗ (tα) : θ(F1 ) → θ(F2 ).

Now suppose also that g has isolated singularity, and let G : X × Bg → C be
an Re -miniversal deformation of g. By versality of G there are base-change maps
ρj : BFj → Bg , such that fj ◦ Fj ∼R un ρ∗j (G). Once again, these are not unique,
but their derivatives at 0 are; after the requisite identifications, they are given by
dF1,t dfj ◦ F1,t
7→ .
dt |t=0 dt |t=0

In other words, this map is induced by Fj∗ (tfj ) : θ(F ) → θ(g), and coincides with
the map T 1 (f, F ) → OX /Jg in the long exact sequence 8).
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 61

Lemma 4.5. We can always choose ρ1 = ρ2 ◦ρ1,2 and thus ensure that the diagram
ρ1,2
B F1 −→ B F2
ρ1 & . ρ2 (12)
Bg
commutes.
Proof Given a group of equivalences G acting on a space of map-germs, let Gun
denote the corresponding group of parametrised equivalences acting on deforma-
tions.
By definition of ρ1,2 , we have
α ◦ F1 ∼Kf2 ,un ρ∗1,2 (F2 ).
From this it follows, by the parametrised version of Lemma 3.2, that
f2 ◦ α ◦ F1 ∼Run f2 ◦ ρ∗1,2 (F2 ) = ρ∗1,2 (f2 ◦ F2 ).
But f2 ◦ α = f1 , so
f1 ◦ F1 ∼Run ρ∗1,2 (f2 ◦ F2 ).
Moreover,
f2 ◦ F2 ∼Run ρ∗2 G,
so
f1 ◦ F1 ∼Run ρ∗1,2 (ρ∗ (G)) = (ρ ◦ ρ1,2 )∗ (G).
Therefore we can take, as ρ1 , the composite ρ2 ◦ ρ1,2 . 2

5. Coherent morphisms
Continuing with the notation and assumptions of the previous section, let H :
(Cm ×B) → (Cn , 0) be any deformation of F1 . Then H also induces a deformation
α◦H of F2 . The deformations H and α◦H are, respectively, Kf1 ,un -equivalent and
Kf2 ,un -equivalent to the deformations of F1 and F2 induced by base-change maps
1i 2 i
B −→ BF1 and B −→ BF2 , whose derivatives at 0 are uniquely determined.
Thus, there is a diagram
BB (13)
|| BB
i1
|| BBi2
|| BB
} || B!
ρ1,2
BF1 / BF2

Example 4.3 shows that the maps i1 and i2 cannot always be chosen so as to
make this diagram commute. For example, in 4.3 the 1-parameter deformation
H(u, v, t) = (u2 +tu, u2 , v) of F1 is infinitesimally Kf1 -trivial, so that i1 : B → BF1
has vanishing derivative at 0. On the other hand, the deformation α ◦ H of F2 is
not infinitesimally Kf2 -trivial, so that i2 : B → BF2 is an immersion.
62 DAVID MOND

Definition 5.1. The morphism α : (f1 , F1 ) → (f2 , F2 ) is coherent if whenever H1

and H2 are deformations of F1 on the same parameter space, then
H1 ∼Kf1 ,un H2 ⇒ α∗ (H1 ) ∼Kf2 ,un α∗ (H2 ) (14)
Proposition 5.2. (1) Let α be a coherent morphism. Then
(i) we can always choose the base-change map i2 in diagram (13) to be equal to
ρ1,2 ◦ i1 , and thus ensure that (13) commutes, and
(ii) T 1 (α) : T 1 (f1 , F1 ) → T 1 (f2 , F2 ) is O-linear.
(2) The composite of coherent morphisms is coherent.
Proof (1)(i) The proof is the same as the deduction of 4.5 from the parametrised
version of 3.2.
(ii) All that is needed for O-linearity is that whenever the 1-parameter defor-
mation H of F1 is infinitesimally trivial, then so should α ◦ H be infinitesimally
trivial also. This is immediate from the definition of coherence.
Statement (2) is obvious. 2

Example 4.3 shows that immersions are not necessariy coherent. Projection,
however, always defines a coherent morphism.
Proposition 5.3. The morphism π in the diagram
Cm
(F, G) . &F
π
Cn × Cp −→ Cn ,
f ◦π & .f
C
is coherent.
Proof Suppose that (Fi , Gi ) for i = 1, 2 are Kf ◦π un -equivalent deformations of
(F, G) on base B. Let Φ : B × X × Y1 × Y2 → B × X × Y1 × Y2 be a diffeo-
morphism in Kf ◦π un which transforms the graph of (F1 , G1 ) to that of (F2 , G2 ).
Write Φ(u, x, y1 , y2 ) = (u, ΦX (u, x), ΦY1 (u, x, y1 , y2 ), ΦY2 (u, x, y1 , y2 ) (using the
subindices merely as labels). Note that by definition of Kf ◦π un -equivalence, we
have
(1) Φ(0, x, y1 , y2 ) = (0, x, y1 , y2 ) for all (x, y1 , y2 ) (Φ is an unfolding of the
identity on X × Y1 × Y2 ), and
(2) f (ΦY1 (u, x, y1 , y2 )) = f (y1 ) for all (u, x, y1 , y2 ).
Define Ψ : B × X × Y1 → B × X × Y1 by
Ψ(u, x, y1 ) = (u, ΦX (u, x), ΦY1 (u, x, y1 , G1 (u, x)).
By (1) and the inverse function theorem, Ψ is a diffeomorphism, and unfolds the
identity on X × Y1 , by (2) it preserves the value of f , and evidently it transforms
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 63

the graph of F1 to that of F2 . Hence F1 and F2 are Kf un -equivalent. 2

This example of coherence, though trivial, will be useful in the proof of the
main theorem of the next section.
Proposition 5.4. The canonical morphisms α0 and α∞ associated to the initial
and terminal object,
Cm Cm
idCm . &F F . &g
α α
∞
Cm 0
−→ Cn and Cm −→ C ,
g& .f f& . idC
C C
(where α0 = F, α∞ = f ) are coherent.
Proof If F : Cm × B → Cm is any deformation of idCm , then there is an unfold-
ing of idCm , Φ : Cm ×B → Cm ×B, such that F ◦Φ = π, where π : Cm ×B → Cm
is projection. It follows that α0 ◦ F ◦ Φ = F ◦ π. Thus every deformation of F
arising as the composite with α of a deformation of idCm is trivial. Hence any two
such are equivalent to one another, and a fortiori α0 is coherent.
Coherence of α∞ , is essentially a parametrised version of 3.2(i). For KidC is
just right-equivalence R; so it is necessary to show that if F1 ∼Kf un F2 then
f ◦ F1 and f ◦ F2 are Run -equivalent deformations of g. Suppose Φ(u, x, y) =
(u, φX (u, x), ΦY (u, x, y)) is a Kf un -equivalence transforming the graph of F1 to
that of F2 . Then
ΦY (u, x, F1 (u, x)) = F2 (u, ΦX (u, x)).
But by definition of Kf un -equivalence, f (ΦY (u, x, y)) = f (y) for every u, x, y, and
in particular
f (F1 (u, x)) = f (F2 (u, ΦX (u, x)).
In other words f ◦ F1 and f ◦ F2 are Run -equivalent. 2

6. An infinitesimal condition for coherence

The simplest condition to guarantee O-linearity of the map T 1 (α) associated with
a morphism α in Cg is
tα Der(− log(f ◦ α)) ⊂ α∗ Der(− log f )).


(15)
For O-linearity is just the inclusion
F ∗ (tα)(tF (θCm ) + F ∗ (Der(− log f ◦ α)) ⊂ t(α ◦ F )(θCm ) + (α ◦ F )∗ (Der(− log f )).
(16)
Clearly F ∗ (tα)(tF (θCm )) ⊂ t(α ◦ F )(θCm ). By applying F ∗ to the inclusion (15)
we deduce the inclusion of the second summand on the left of (16) in the second
64 DAVID MOND

summand on the right.

Condition (15) is a coherent version of the linear algebraic statement

dx α(ker dx (f ◦ α)) ⊂ ker dα(x) f. (17)
However, whereas (17) always holds, (15) does not, as demonstrated by Example
4.3.
Example 6.1. (1) In fact (15) holds for any flat morphism (see Remark 6.2
below). In Section 7 we give examples of non-flat morphisms satisfying
(15).
(2) The canonical morphism α∞ to the final object in Cg described in Example
2.5 satisfies (15). The canonical morphism from the initial object does not
always satisfy (15).
Remark 6.2. Since
Der(− log(f ◦ α)) = Z1 (K• (f ◦ α))
and
α∗ (Der(− log f )) = B1 α∗ (L• (f )) ,

(15) is equivalent to the vanishing of the morphism

Z1 α∗ (L• (f ))


Z1 (K• (f ◦ α)) O p
H1 (K• (f ◦ α)) = →
= Tor1 ( C , OCn ),
B1 (K• (f ◦ α)) ∗
B1 α (L• (f )) Jf
induced by tα in the long exact sequence S(α, F ).
Theorem 6.3. If (15) holds then the morphism α is coherent.
Proof Let A denote the graph-embedding of α: A(y) = (y, α(y)). We can recover
α from A by composition with the projection π2 : Y1 × Y2 → Y2 . As projection
gives rise to a coherent morphism, and the composite of coherent morphism is
coherent, it will be enough to prove the result for the immersion A, provided, that
is, that we can show that if (15) holds for α then it holds for A. That is, we have
first to show that if
tα(Der(− log f ◦ α)) ⊂ α∗ (Der(− log f ))
then
tA(Der(− log f ◦ π2 ◦ A)) ⊂ A∗ (Der(− log f ◦ π2 )).
In fact this is straightforward: Der(− log f ◦ π) = (Der(− log f ) ⊗ OY1 ×Y2 ) ⊕
θY1 ×Y2 |Y2 , while f ◦ π2 ◦ A = f ◦ α, so
tA(Der(− log f ◦ π2 ◦ A)) = tA(Der(− log f ◦ α)) ⊆ tα(Der(− log f ◦ α)) ⊕ θ(A|Y2 )
⊆ α∗ (Der(− log f )) ⊕ θ(A|Y2 ) = A∗ (Der(− log f ◦ π2 )).
Now we prove the theorem under the assumption that α is an immersion. By
this assumption, after a diffeomorphism we can write Y2 = Y1 ×N for some smooth
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 65

manifold N , and take α : Y1 → Y2 to be the inclusion of Y1 as Y1 × {0} ⊂ Y1 × N .

To ease our notation, we rename Y1 simply Y . The operator tα converts a vector
field on Y to “the same” vector field on Y × {0} ⊂ Y × N , and can thus be
suppressed in our notation.
Let Fi : B × X → Y , for i = 1, 2, be Kf ◦α un -equivalent deformations of F :
X → Y , and let Φ be a Kf ◦α un -equivalence transforming the graph of F1 to that
of F2 . It has the form
Φ(u, x, y) = (u, φX (u, x), φY (u, x, y))
and satisfies f ◦ α(φY (u, x, y)) = f ◦ α(y), Φ(0, x, y) = (0, x, y). We have to extend
Φ to a diffeomorphism Φ̃ of B × X × Y1 × N ,
Φ̃(u, x, y, n) = (u, Φ̃X (u, x), Φ̃Y (u, x, y, n), Φ̃N (u, x, y, n)),
such that f (Φ̃Y (u, x, y, n), Φ̃N (u, x, y, n)) = f (y, u). Any such extension will trans-
form the graph of α ◦ F1 to that of α ◦ F2 , since both graphs are contained in
B × X × Y × {0}.
Suppose first that dim B = 1, and let t be a coordinate on B. Since Φ is an
unfolding of the identity on X × Y , Φ(t, x, y) is the image of (0, x, y) under the
integral flow of a time-dependent vector field
∂ X ∂ X ∂
+ ξi (t, x) + ηj (t, x, y) .
∂t i
∂x i j
∂y j

The vector field η := j ηj ∂y∂ j lies in Der(− log f ◦ α) ⊗ OB×X×Y , since the value
P

of f ◦ α is preserved along the flow lines. The hypothesis, (15), implies that η
extends to a vector field η̃ ∈ Der(− log f ) ⊗ OB×X×Y ×N . The map Φ̃ sending
(0, x, y, n) to its image under the time t integral flow of the extended vector field
∂
+ χ(u, x, y) + η̃(u, x, y, n)
∂t
is a Kf un -equivalence extending Φ.
To deal with the general case where B is smooth of dimension d, we identify B
with Cd and let B1 = C × {0}, B2 = C2 × {0}, . . . , Bd = B. Suppose inductively
that the diffeomorphism Φj := Φ|Bj−1 ×X×Y has been extended to a diffeomor-
phism Φ̃j−1 of Bj−1 × X × Y × N which preserves the value of f as required. Let
t1 , . . . , tj be coordinates on Bj . Once again, Φ((t1 , . . . , tj , 0, . . . , 0)), x, y) is the
image of Φ((t1 , . . . , tj−1 , 0, . . . , 0), x, y) under the integral flow of a tj - dependent
vector field
∂
+ ξ(t1 , . . . , tj , x, y) + η(t1 , . . . , tj , x, y)
∂tj
with η ∈ Der(− log f ◦ α) ⊗ OBj ×X×Y . By (15), η extends to η̃ ∈ Der(− log f ) ⊗
OBj ×X×Y ×N , and now we extend Φj to Φ̃j of Bj ×X ×Y ×N by the now standard
procedure. This completes the proof. 2
66 DAVID MOND

The converse to 6.3 does not hold. For example, (15) does not always hold for
the morphism α0 : (g, idCm ) → (f, F ), but α0 is always coherent.
It is useful to keep in mind the diagram
ker{α∗ (tf ) : θ(α) → θ(f ◦ α)}
% -
tα(Der(− log f ◦ α)) --→ α∗ (Der(− log f )) ,
↑ ↑
tα(Ham(f ◦ α)) → α∗ (Ham(f ))
where Ham(f ) denotes the module generated by the Hamiltonian vector fields
ηij (f ). The inclusion (15) is the broken arrow.
Two cases in which, in view of 6.2, (15) obviously holds, are
(1) where Tor1OCp (OCp /Jf , OCn ) = 0, in which case the right upper arrow is
an equality, and
(2) where f ◦ α has isolated singularity, in which case the left lower arrow is
an equality.
The first of these two occurs, of course, if α is flat. So there are two subcategories
of Cg in which all morphisms are coherent:
(1) The subcategory whose objects are those of Cg , and whose morphisms are
the flat morphisms of Cg , and
(2) The full subcategory of Cg consisting of factorisations (f, F ) where f has
isolated singularity.
Since (2) excludes, for example, determinantal factorisations, it is rather restric-
tive. More interesting than (1) is the subcategory whose objects are Cohen-
Macaulay factorisations (f, F ), that is, where OCn /Jf is Cohen-Macaulay, and
whose morphisms are those morphisms α of Cg such that the codimension of
V (α∗ (Jf )) in Cn is equal to the codimension of Jf in Cp . For such morphisms,
α∗ (L• (f )) is acyclic, condition (1) above holds, and thus α is coherent.
Given a factorisation (f, F ) of g, call codim V (Jf ) − codim V (F ∗ (Jf )) the di-
mensional defect of (f, F ) and denote it by defect(f, F ). The dimensional defect
plays an important role in the relation bewteen τ := dimC T 1 Kf F and the rank of
the vanishing homology of g = f ◦ F under perturbations of F — see [5].
We will say a morphism α : (f1 , F1 ) → (f2 , F2 ) has finite length if T 1 (f, α) has
finite length.
Lemma 6.4. Suppose that α has finite length. Then a necessary and sufficient
condition for equality of codim V (α∗ (Jf )) and codim V (Jf ) is that defect(f, α◦F ) =
defect(f ◦ α, F ).
Proof By exactness of S(f, α), the finite length of T 1 (f, α) implies that
codim V (Jf ◦α ) = codim V (α∗ (Jf )). So
defect(f, α ◦ F ) = defect(f ◦ α, F ) ⇔ codim V (Jf ) = codim V (Jf ◦α )
 FUNCTORS ON THE CATEGORY OF FACTORISATIONS OF A FUNCTION 67

⇔ codim V (Jf ) = codim V (α∗ (Jf )).

2
For each integer k, let Cg (k) denote the full subcategory of Cg consisting of fac-
torisations with defect k.
Theorem 6.5. Let (f1 , F1 ) and (f2 , F2 ) be Cohen-Macaulay factorisations in
Cg (k), and let α : (f1 , F1 ) → (f2 , F2 ) be a morphism of finite length. Then α
is coherent.
Proof Let L• (f ) and L• (f ◦α) be free resolutions of OCp /Jf and OCn /Jf ◦α , over
OCp and OCn respectively. As OCp /Jf is Cohen Macaulay, and, by the lemma,
codim V (α∗ (Jf )) = codim V (Jf ), it follows that α∗ (L• (f )) is acyclic. This means
in particular that TorO1
Cp
(OCp /Jf , OCn ) = 0. 2

7. Symmetric determinantal factorisations revisited

A symmetric matrix family S : (Cm , 0) → Symn (C) can also be viewed as a
(unrestricted) matrix family. Thus there is a morphism of factorisations
Cm
Ss . & Sg
i
Symn (C) ,→ Matn (C) (18)
dets & . detg
C
of det S (we have added subscripts to S and to det in order to indicate their target
and source). This morphism is coherent, although i∗ (L• (detg )) is not acyclic.
We take, as L• (det) and L• (dets ), the well known Gulliksen-Negård and Jozefiak
resolutions J• and GN• (see [7] and [8] respectively), which are shown below.
It turns out that there is a split morphism of complexes p• : J• → i∗ (GN• ); in
particular (15) holds, and coherence follows by Theorem 6.3. The splitting has
the rather remarkable form
p• r
0 → J• −→ i∗ (GN• ) −→
•
Hom(J4−• , O) → 0; (19)
in some sense it explains the fact that J1 = symmetric matrices over O = OSymn (C)
and J3 = skew-symmetric matrices over O are complements in matn (O). The split
short exact sequence of complexes (19) is shown vertically in the following diagram.
ds
3 ds
2 ds
1
0 → skn (O) −→ sln (O) −→ symn (O) −→ O → 0
p3 ↓ p2 ↓ p1 ↓ p0 ↓
g g g g
d4 d3 d2 d1
0 → O −→ matn (O) −→ sln (O) ⊕ sln (O) −→ matn (O) −→ O → 0
r3 ↓ r2 ↓ r1 ↓ ↓
(ds
1)
t (ds
2)
t (ds
3)
t
0 → O −→ symn (O) −→ sln (O) −→ skn (O) → 0
(20)
where
p0 is the identity p1 is inclusion p2 (Y ) = (Y, −Y t ) p3 is inclusion
r1 (X) = 21 (X − X t ) r2 (Y1 , Y2 ) = Y1 + Y2t r3 (Z) = 21 (Z + Z t ) r4 = id
68 DAVID MOND

The morphism p• exists despite the fact that i∗ (GN• ) is not acyclic. The diagram
above shows that H1 (i∗ (GN• )) ' Ext3 (O/Jdet , O). The fact that p1 = ti lifts to
a morphism J2 → i∗ (GN2 ) is equivalent to the image of p1 not meeting the ho-
mology of i∗ (GN• ).

It is of course well-known that OMatn (C) /Jdet is Gorenstein and that GN• is self-
dual. Does this, together with the Z2 -action generated by transposition, of which
Symn (C) is the fixed-point set, explain the remarkable split extension (19)?

References
[1] J. N. Damon Deformations of sections of singularities and Gorenstein surface singularities,
Amer.J.Math. 109 (1987) no. 4, 695-721.
[2] J. N. Damon A-equivalence and the equivalence of sections of images and discriminants,
Singularity Theory and its Applications, Warwick 1989, Part I, D.Mond and J.Montaldi
(Eds.), Lecture Notes in Maths. 1461, Springer Verlag, 1991.
[3] J. N. Damon and D. Mond, A-codimension and the vanishing topology of discriminants,
Invent. Math. 106 (1991), 217–242.
[4] D. Eisenbud, Homological algebra on complete intersections, with an application to group
representations, Trans. Amer. Math. Soc. 260 (1980), 35-54
[5] V. V. Goryunov and D. Mond, Tjurina and Milnor numbers of matrix singularities,
math.AG/0307025 at arXiv.org, 2003
[6] V. V. Goryunov and V. M. Zakalyukin, Simple symmetric matrix singularities and the
subgroups of Weyl groups Aµ , Dµ , Eµ , to appear in Moscow Mathematical Journal.
[7] T. H. Gulliksen and O. G. Negård, Un complexe résolvent pour certains idéaux
déterminantiels, C.R. Acad. Sci. Paris Sér. A 274 (1972), 16-18.
[8] T. Józefiak, Ideals generated by minors of a symmetric matrix , Coment. Math. Helvetici
53 (1978), 595–607.
[9] Y. Yoshino, Cohen-Macaulay modules over Cohen-Macaulay rings, London Math. Soc.
Lecture Notes 146, Cambridge University Press, 1990

Personal note: Jairo Charris was the first person I spoke to when I went to enquire
about doing a Master’s degree in the National University in 1975. He opened his
office door as I walked nervously along the corridor, dripping from a recent down-
pour, and asked me, in English, if he could help. He could, and did. Enormously.
He invited me into mathematics, and taught me that it was worth doing and
that I could do it. His combination of informality, friendliness and passion for
mathematics was incredibly attractive. To him, I owe my career.