Symplectic geometry of homological algebra

Maxim Kontsevich

June 10, 2009

Derived non-commutative algebraic geometry

With any scheme X over ground field k we can associate a k-linear trian-
gulated category Perf(X) of perfect complexes, i.e. the full subcategory of
the unbounded derived category of quasi-coherent sheaves on X, consisting
of objects which are locally (in Zariski topology) quasi-isomorphic to finite
complexes of free sheaves of finite rank.
The category Perf(X) is essentially small, admits a natural enhancement
to a differential graded (dg in short) category up to a homotopy equivalence,
and is Karoubi (e.g. idempotent) closed. The main idea of derived non-
commutative algebraic geometry is to treat any Karoubi closed small dg
category as the category of perfect complexes on a “space”.
By a foundamental result of A. Bondal and M. Van den Bergh, any
separated scheme of finite type is affine in the derived sense, i.e. Perf(X) is
generated by just one object. Equivalently,

Perf(X) ∼ Perf(A)

for some dg algebra A, where prefect A-modules are direct summands in the
homotopy sense of modules M which are free finitely generated Z-graded A-
modules, with generators m1 , . . . , mN of certain degree deg(mi ) ∈ Z, such
that dmi ∈ ⊕j<i A · mj for all i. Algebra A associated with X is not unique,
it is defined up to a derived Morita equivalence.
Some basic properties of schemes one can formulate purely in derived
terms.

Definition 1. Dg algebra A is called smooth if A ∈ Perf(A ⊗ Aop ). It

is compact if dim H • (A, d) < ∞. This properties are preserved under the
derived Morita equivalence.

1
 For a separated scheme X of finite type the properties of smoothness and
properness are equivalent to the corresponding properties of a dg algebra A
with Perf(A) ∼ Perf(X). Smooth and compact dg algebras are expected to
be the “ideal” objects of derived geometry, similar to smooth projective vari-
eties in the usual algebraic geometry. For a smooth algebra A the homotopy
category Fin(A) of dg-modules with finite-dimensional total cohomology is
contained in Perf(A), and for compact A the category Perf(A) is contained
in Fin(A). One can define two notions of a Calabi-Yau algebra of dimension
D ∈ Z. In the smooth case it says that A! := HomA⊗Aop −mod (A, A ⊗ Aop ) is
quasi-isomorphic to A[−D] as A ⊗ Aop -module (it corresponds to the trivial-
ity of the canonical bundle for smooth schemes). Similarly, in the compact
case we demand that A∗ = Homk−mod (A, k) is quasi-ismorphic to A[D], as a
bimodule (it corresponds for schemes to the condition that X has Gorenstein
singularities and the dualizing sheaf is trivial).
The notion of smoothness for dg algebras is itself not perfect, as e.g. it
includes somewhat pathological example k[x, (1/(x − xi )i∈S ] where S ⊂ k is
an infinite subset. It seems that the right analog of smooth shemes (of finite
type) is encoded in the following notion of dg algebra of finite type due to
B. Toën and M. Vaquie:
Definition 2. A dg algebra A is called of finite type if it is a homotopy
retract in the homotopy category of dg algebras of the free finitely generated
algbera khx1 , . . . , xN i, deg(xi ) ∈ Z with the differential of the form

dxi ∈ khx1 , . . . , xi−1 i, i = 1, . . . , N .

Any dg algebra of finite type is smooth, and any smooth compact dg

algebra is of finite type. It is also convenient to replace a free graded algebra
in the definition of finite type by the algebra of paths in a finite Z-graded
quiver.
A large class of small triangulated categories (including many exam-
ples from representation theory) can be interpreted as the categories of
perfect complexes on a space of finite type with a given “support”. In
terms of dg algebras, in order to specify the support one should pick a
perfect complex M ∈ Perf(A). The corresponding category is the full sub-
category of Perf(A) generated by M , and is equivalent to Perf(B) where
B = EndA−mod (M, M )op . One can say in non-commutative terms what is
the “complement” X − Supp(M ) and the “formal completion” X bSupp M of
X at Supp(M ). The complement is given by the localization of Perf(X) =
Perf(A) at M , and is again of finite type. By Drinfeld’s construction, in
terms of dg quivers it means that we add a new free generator hM ∈

2
Hom−1 (M, M ) with dhM = idM . The formal completion is given by al-
gebra C = EndB−mod (M, M )op . E.g. when A = k[x] and M = k with x
acting trivially, we have B = H • (S 1 , k) (the exterior algebra in one variable
in degree +1), and C = k[[x]].

Examples of categories of finite type

Algebraic geometry: For any smooth scheme X the category Perf(X) '
Db (Coh(X)) is of finite type.
Topology: Let X be now a space homotopy equivalent to a finite con-
nected CW complex. Define AX := Chains(Ω(X, x0 )), the dg algebra of
chains (graded in non-positive degrees) of the monoid of based loops in X,
with the product induced from the composition of loops. This algebra is
of finite type as can be seen directly from the following description of a
quasi-isomorphic algebra.
Let us assume for simplicity that X is simplicial subcomplex in a stan-
dard simplex ∆K for some K ∈ Z≥0 . We associate with such X a finite
dg quiver QX . Its vertices are vi , i = 0, . . . , K for i ∈ X. The arrows are
ai0 ,...,ik for k > 0, where (i0 , . . . , ik ) is a face of X, and i0 < i1 < · · · < ik .
The arrow ai0 ,...,ik has degree (1 − k) and goes from vi0 to vik . We define
the differential in QX by
k−1
X
dai0 ,...,ik = (−1)j ai0 ,...,ij · aij ,...,ik
j=1

Then we have to “invert” all arrows of degree 0, i.e. add inverse arrows ai0 ,i1
for all egdes (i0 , i1 ) in X. It can be done either directly (but then we obtain
a non-free quiver), or in a more pedantical way which gives a free quiver.
In general, if want to invert a arrow aEF in a dg quiver connecting verices
E and F , with deg aEF = 0 and daEF = 0, one can proceed as follows.
To say that aEF is an isomorphism is the same as to say that the cone
C := Cone(aEF : E → F ) is zero. Hence we should add an endmorphism
hC of the cone of degree −1 whose differential is the identity morphism.
Describing hC as 2 × 2 matrix one obtains the following. One has to add 4
arrows
h0F E , h−1 −1 −2
EE , hF F , hEF
with degrees indicated by the upper index, with differentials
dh0F E = 0, dh−1 0
EE = idE −aEF · hF E ,

dh−1 0 −2 −1 −1
F F = idF −hF E · aEF , dhEF = aEF · hF F − hEE · aEF .

3
Theorem 1. The quiver QX localized in either way, is dg equivalent to AX .

In particular, if X is space of type K(Γ, 1) then AX is homotopy equiv-

alent to an ordinary algebra in degree 0, the group ring k[Γ]. In particular,
such an algebra is of finite type. In the case char(k) = 0 one can also al-
low torsion, i.e. consider orbispaces, hence Γ can be an arithmetic group, a
mapping class group, etc.
The full subcategory of finite-dimensional dg modules Fin(AX ) ⊂ Perf(AX )
is the triangulated category of sheaves whose cohomology are finite rank lo-
cal systems on X. If we invert not all arrows of degree 0 in QX for simplicial
X ⊂ ∆K , we can obtain categories of complexes of sheaves with cohomol-
ogy constructible with respect to a given CW-stratification, and even more
general categories.
Algebraic geometry II: The last example of a category of finite type
is somewhat paradoxical.

Theorem 2. (V.Lunts) For any separated scheme X of finite type the cat-
egory Db (Coh(X)) (with its natural dg enhancement) is of finite type.

Morally one should interpret Perf(X) as the category of perfect com-

plexes on a smooth derived noncommutative space Y with support on a
closed subset Z. Then the category Db (Coh(X)) can be thought as the
category of perfect complexes on the formal neighborhood X bZ . It turns out
that for the case of usual schemes this neighbourhood coincides with Y it-
self. The informal reason is that the “transversal coordinates” to Z in Y
are of strictly negative degrees, hence the formal power series coincide with
polynomials in Z-graded sense.

Fukaya categories

Let (X, ω) be a compact symplectic C ∞ manifold with c1 (TX ) = 0

The idea of K. Fukaya is that one should associate with (X, ω) a compact
A∞ Calabi-Yau category over a non-archimedean field (Novikov ring)
X
N ov := ai T Ei , ai ∈ Q, Ei ∈ R, Ei → +∞ ,
i

where numbers Ei have the meaning of areas of pseudo-holomorphic discs.

The objects of F(X) in the classical limit T → 0 should be oriented La-
grangain spin manifolds L ⊂ X (maybe endowed with a local system). There
are several modifications of the original definition:

4
 • one can allow manifolds with c1 6= 0 (in this case one get only a Z/2Z-
graded category),

• on can allow X to have a pseudo-convex boundary (see the discussion

of the Stein case below),

• (Landau-Ginzburg model), X is endowed with a potential W : X → C

satisfying some conditions at infinity (then the corresponding Fukaya-
Seidel category is not a Calabi-Yau one),

• allow X to have holes inside, then one get so called “wrapped” Fukaya
category with infinite-dimensional Hom-spaces.

Fukaya categories of Stein manifolds

The simplest and the most important case is when X is compact complex
manifold with real boundary such that there exists a strictly plurisubhar-
monic function f : X → R≤0 with f|∂X = 0 and no critical points on ∂X.
Seidel in his book gave a complete definition of the Fukaya category of
Stein manifold in terms of Lefschetz fibrations. The additional data neces-
sary for Z-grading is a trivialization of the square of the canonical bundle.
One can analyze his construction and associate certain algebra A of finite
type (over Z) such that the Fukaya category constructed by Seidel is a full
subcategory of Fin(A). We propose to consider A (or category Perf(A) and
not Fin(A)) to be a more foundamental object, and to formulate all the the-
ory in such terms. For example, for X = T ∗ Y where Y is a compact oriented
manifold, the algebra A is Chains(Ω(Y, y0 )) contains information about the
foundamental group of Y , whereas the category of finite-dimensional repre-
sentations could be very poor for non-residually finite group π1 (Y ).
Also we propose a slightly different viewpoint on AX . Namely, one can
make X smaller and smaller without changing A, and eventually contract
X to a singular Lagrangian submanifold L ⊂ X. Hence we can say that
A = AL depends only on L (up to derived Morita equivalence). One can
think for example about L being a 3-valent graph embedded in an open
complex curve X as a homotopy retract. If X is endowed with a potential,
we should contract X to a noncompact L such that Re(W )|L : L → R
is a proper map to [c,P +∞), c ∈ R, e.g. L = Rn for X = Cn with the
holomorphic potential ni=1 zi2 .
We expect that Fin(AX ) is the global category associated with a con-
structible sheaf (in homotopy sense) EL of smooth compact dg categories on

5
L depending only on the local geometry. In terms of dg algebras, AX is a
homotopy colimit of a finite diagram of local algebras. For example, if L is
smooth and oriented and spin, the sheaf EL is the constant sheaf of Perf(Z),
and the global algebra is the algebra Chains(Ω(L, x0 )) considered before.
In terms of topological field theory, the stalks of EL are possible boundary
terms for the theory of pseudo-holomorphic discs in X with boundary on L.
In codimension 1 singular Lagrangian L looks generically as the product
of a smooth manifold with the union of three rays {z ∈ C | z 3 ∈ R≥0 },
endowed with a natural cyclic order. The stalk of the sheaf EL at such a
point is Perf(A2 ), the category of representations of quiver A2 (two vertices
and one directed egde). The symmetry group of Perf(A2 ) after factoring by
the central subgroup of shifts by 2Z is equal to Z/3Z. Explicitly it can be
done by the following modification of the quiver at triple points. Namely,
consider the quiver with three vertices (corresponding to 3 objects E, F, G),
a closed arrow F → G of degree 0, two arrows E → F, E → G of degreess
−1 and 0 respectively (with differential saying that we have a morphism
E → Cone(F → G). We say that E is quasi-isomorohic to Cone(F → G),
i.e.
Cone(E → Cone(F → G)) = 0 .
This can be done explicitly by constructing a homotopy to the identity of the
above object, which is a 3 × 3-matrix. Combining all equations together we
get a quiver with 3 vertices and 12 arrows which gives a heavy but explicit
finite type model for exact triangles.
A natural example of a Lagrangian submanifold with triple point sin-
gularities comes from any union of transversally intersecting Lagarngain
submanifolds Li ⊂ X, i = 1, . . . , k. For any point x of intersection (or self-
intersection) we should remove small discs in two branches of Lagrangian
manifolds intersecting at x, and glue a small ball with two collars. The set
of triple points forms a sphere.
Global algebra AL of finite type is Calabi-Yau if L is compact, and not
Calabi-Yau in general for non-compact L. There are many examples of
(compact and non-compact) sungular Lagrangian manifolds such that

Perf(AL ) ' Db (Coh(X))

for some scheme X of finite type over Z (maybe singular and/or non-
compact). In the pictures at the end we collected several examples of this
“limiting mirror symmetry”. Categories of type AL one can consider as
“non-commutative spaces of finite type” defined combinatorially, without

6
parameters. Among other examples one can list toric varieties, maximally
degenerate stable curves, etc.

Deforming degenerate Fukaya categories

Let us assume that X is compact, in fact a complex projective manifold,

and take a complement X o to an ample divisor. New manifold X o is Stein,
and can be contracted to a singular Lagrangian L ⊂ X o . The advantage of
X o is that is has no continuous parameters as a symplectic manifold. As
was advocated by P. Seidel several years ago, one can think of F(X) as a
deformation of F(X o ). For example, if X is a two-dimensional torus (elliptic
curve) and X o is the complement to a finite set, then F(X o ) is equivalent
to Perf(Y0 ) where Y0 is a degenerate elliptic curve, a chain of copies of P1 .
In algebraic terms, holomorphic discs in X give a solution of the Maurer-
Cartan equation
dγ + [γ, γ]/2 = 0, γ ∈ C • (AL , AL )⊗m
b N ov
where C • (AL , AL ) = Cone(AL → Der(AL )) is the cohomological Hochschild
complex of smooth algebra AL , and mN ov is the maximal ideal in the ring
of integers in the Novikov field N ov.
Analogy with algebraic geometry suggests that different choices of open
o
X ⊂ X should lead to dg algebras of finite type endowed with deforma-
tions over mN ov such that algebras became (in certain sense) derived Morita
equivalent after the localization to N ov. We expect that such a formulation
will handle the cases when the deformed Fukaya category is too small, e.g.
when the mirror family consists of non-algebraic varieties (e.g. non-algebraic
K3 surfaces or complex tori).

Question about automorphisms

The group of connected components of X (with appropriate modications

for the potential/Landau-Ginzburg/wrapped cases) acts by automorphisms
of dg category F(X) over the local field N ov. One can ask whetehr this
group coincides with the whole group of automorphisms. To our knowledge,
there is no counterexamples to it! In principle, one can extend the
P group
by taking the product of X with the Landau-Ginzburg model (Cn , ni=1 zi2 )
which is undistinguishable categorically from a point. So, a more realistic
conjecture is that the automorphism group of Fukaya category coincides with
the stabilized symplectomorphism group. Why anything like this should be
true?

7
 There is an analogous statement in the (commutative) algebraic geome-
try. The group of automorphisms of a maximally degenerating Calabi-Yau
variety Y over a local non-archimedean field K maps naturally to the group
of integral piece-wise linear homeomorphisms of certain polytope (called the
skeleton, and usally homeomorphic to a sphere). The skeleton lies intrisically
in the Berkovich spectrum Y an where the latter is defined as the colimit of
sets of points X(K 0 ) over all non-archimedean field extensions K 0 ⊃ K. The
Berkovich spectrum is a very hairy but Hausdorff topological space, and the
skeleton is a naturally defined homotopy retract of Y an .
We expect that one can define some notion of analytic spectrum for a dg
algebra over a non-archimedean field, and its skeleton should be probably
a piecewise symplectic manifold (maybe infinite-dimensional). For Fukaya
type categories this skeleton should be the original symplectic manifold.