Patrick Lei

All notes

Simons Summer Math Workshop 2023

Lectures by Brett Parker, Bernd Siebert, Danilo Lewanski and Dimitri Zvonkine, Simons Center, Summer 2023.

Simons Math Summer Workshop 2023
Notes by Patrick Lei
Lectures by Various

Simons Center for Geometry and Physics

Disclaimer

These are my notes from the second and third weeks of the 2023 Simons Math Summer Workshop.

These notes were taken during the lectures using neovim. Any errors are mine and not the speakers’. In addition, my notes are picture-free (but will include commutative diagrams) and are a mix of my mathematical style and that of the lecturers. Also, notation may differe between lecturers. If you find any errors, please contact me at plei@math.columbia.edu.

Acknowledgements

I would like to thank Mark Gross, Mark McLean, Simon Donaldson, Kenji Fukaya, John Pardon, and the Simons Center for organizing the workshop.

1 Degenerations of holomorphic curves, tropical geometry, gluing theorems, and exploded manifolds (Brett Parker)

1.1 Geometry of holomorphic curves

Consider a map f:Σ→M, where Σ is a Riemann surface and M is a manifold with metrics gΣ,gM respectively.

Definition 1.1.1.

The L2 energy of f is given by

where |D⁢f| is defined using gΣ and gM on T∗⁢Σ⊗T⁢M.

This is conformally invariant, as can be seen by computing the energy after scaling gΣ. This is also analogous to the definition of geodesics, where we consider stationary points of

We really want to consider holomorphic maps, so we choose an almost complex structure J∈End⁡(T⁢M) such that

This defines a 2-form by

which is not necessarily closed. We also choose a complex structure j on Σ.

Lemma 1.1.2.

The L2 energy of f satisfies the inequality

with inequality if and only if df∘j=J∘df (meaning that f is holomorphic).

Locally, if we choose coordinates x,y such that the complex structure is x+i⁢y, then

with equality if and only if f is holomorphic.

Remark 1.1.3.

It is tempting to consider just the area in the inequality, but that would not tell us how to parameterize the curve, just like how defining geodesics by minimizing length doesn’t tell us how to parameterize the geodesic.

If ω is closed, then Eω⁢(f)=∫Σf∗⁢ω is deformation invariant, so f is holomorphic if and only if f minimizes L2-energy and L2⁢(f)=Eω⁢(f).

1.1.1 Regularity

Lemma 1.1.4 (Monotonicity).

Consider a ball of radius R such that a nonconstant holomorphic curve f passes through its center. Then

where εR depends on gM and R.

Morally, we should think that it takes some amount of energy for our curve to escape the ball.

Now note that the Cauchy-Riemann equation is elliptic, so we have elliptic regularity. Specifically, we have

which gives us bounds on all higher derivatives, so solutions are smooth. Unfortunately, L2 energy bounds do not guarantee derivative bounds.

Theorem 1.1.5 (Removable singularity).

If L2⁢(f)<∞ then f:Σ∖p extends over p.

Lemma 1.1.6.

Consider f:B1⁢(0)→M. If L2⁢(f) is small, then

1.1.2 Think-thin decomposition of the domain

For a holomorphic curve f:Σ→M with Eω⁢(f)<ω, our goal is to decompose Σ into long thin annuli A with low energy and “thick” regions Ci such that

  • •

    Eω⁢(f|A)<ε0;

  • •

    |d⁢f| is bounded on C;

  • •

    The diameters of the Ci are bounded;

  • •

    Either χ⁡(Ci)<0 or Eω⁢(f|Ci)>ε0.

Remark 1.1.7.

This requires changing the metric in the same conformal class to eliminate bubbling.

Taking the limit as the lengths of the annuli go to infinity, we obtain a degerated holomorphic curve with many components.

1.1.3 Noncompact stuff

Some examples of noncompact manifolds we want to consider are N×ℝ×S1⏟ℂ×, where N is compact, or (ℂ×)k⋊N, where the fibers are holomorphic and J,g are ℂ×-invariant. In the first case, we are using a ℂ×-invariant metric, so the manifold really does look like a cylinder. A third case we will consider are those which are asymptotically cylindrical, for example M∖D, where M is a complex manifold and D is a normal crossings divisor.

In these cases, we must use a different taming form ω, so g≠ω⁡(−,J−). For example, in the cylindrical case we can consider

αh is a connection form, t,θ are coordinates on ℝ×S1, and ρh looks like a smoothed step function. We obtain the inequality

or we consider the L2 energy of f locally on (ℂ×)k⋊N. Unfortunately, this means we only have local energy bounds.

Exercise 1.1.8.

Let f:A→ℂ× be holomorphic with |dff|<c0. Write A={e−T≤|z|≤eT}. Show that

  1. (a)

    |df⁡(z)(f⁡(z))−df⁡(1)f⁡(1)|≤e−T⁢c0⁢(|z|+1|z|);

  2. (b)

    f⁡(z)=eh⁡(z)⁢a⁢zm with h⁡(z)≤2⁢(|z|+1|z|)⁢e−T⁢c0.

Therefore, in logarithmic coordinates, a holomorphic curve in (ℂ×)2 has its thin regions travelling in straight lines and the thick regions are very small. In the tropical limit, we obtain a piecewise-linear curve.

In all of these cases, we can replace the RHS of the first part of the exercise with

where 0<δ<1.

1.2 Exploded manifolds

One of the problems that we have is dealing with what happens in our moduli spaces as things go off to infinity. Of course, the moduli space of the tropical curves is not compact, but it is relatively nice. To deal with the problem in the usual geometry, we will develop the formalism of exploded manifolds.

1.2.1 Explosion

Let R be a ring. The exploded semiring is R⁢tℝ with multiplication given by

and addition given by

If R is the ring with one element, then we get the tropical numbers. Then there is a tropical part homomorphism

where a represents the tropical number 0⁢ta∈tℝ. There is also a smooth part homomorphism

The goal is to define the functor

First, we will define the exploded point

For any log-scheme X†, we will define

Here, the structure morphisms to 𝒪X are as expected. A holomorphic morphism of these exploded objects is simply a morphism X†×p†→Y†×p† over p†.

Example 1.2.1.

We will now consider the points p†→𝖤𝗑𝗉𝗅⁡(ℂ,0). These correspond to morphisms

This is determined by the value h⁡(z⁢t0)∈ℂ×⁢t[0,∞). We may now consider a new coordinate z~ on 𝖤𝗑𝗉𝗅⁡(ℂ,0). Thus, as a set, we have

This will admit the notion of a tangent space and a metric, and geometrically we will see ℂ∖0 with a cylindrical end at 0 and then infinitely many cylindrical ends in the [0,∞) direction.

1.2.2 Exploded manifolds

Changing perspective, we will consider 𝖤𝗑𝗉𝗅⁡(ℂ,0) as an exploded manifold. We will have

  • •

    A set of points ℂ×⁢t[0,∞);

  • •

    A topology induced from the analytic topology on ℂ using the smooth part homomorphism;

  • •

    A sheaf of functions valued in ℂ×⁢tℝ. These functions will look something like f⁡(⌈z~⌉)⁢z~m⁢ta, where m∈ℤ, a∈ℝ, and f:ℂ→ℂ× which is regular in some sense.

Here, regular could mean holomorphic, smooth, Ck, or another form of regularity known as C∞,1¯. This means that for all δ<1, the function |f⁡(z)−f⁡(0)|⁢|z|−δ extends to be continuous at 0 and so do all of its derivatives |Dk⁢f⁢(z)|⁢|z|−δ, where the deriatives are defined using the real or imaginary parts of z⁢∂∂z.

Definition 1.2.2.

An exploded space B is a set B equipped with

  • •

    A topology induced from a surjective map B→⌈B⌉, where ⌈B⌉ is Hausdorff;

  • •

    A sheaf of ℂ×⁢tℝ-valued functions ℰ×⁢(B) containing constants.

Example 1.2.3.

For the exploded manifold 𝖤𝗑𝗉𝗅⁡(X†), then the set B is the set of maps p†→X†, the smooth part is ⌈𝖤𝗑𝗉𝗅⁡(X†)⌉=X with the analytic topology, and ℰ×⁢(𝖤𝗑𝗉𝗅⁡(X†))=ℳXgp⁢tℝ.

Example 1.2.4.

We will consider an important non-logarithmic example. We will consider 𝕋=ℂ×⁢tℝ with ⌈𝕋⌉=pt and ℰ×⁢(𝕋) the set of monomials c⁢ta⁢zm.

Exercise 1.2.5.

Show that for any exploded space B, maps B→𝕋 correspond to sections of ℰ×⁢(B).

Example 1.2.6.

The most general example will be 𝕋Pm, where P⊂ℝm is an integral affine polytope defined by inequalities of the form α⋅x+a≥0, where α∈ℤm and a∈ℝ. We also allow strict inequalities. We will define the set 𝕋Pm⊂(ℂ×⁢tℝ)m as the points whose tropical part lie in P.

We will now set

to be the quotient of integral affine maps on ℝm that are nonnegative on P by those which are strictly positive. As a log scheme, this is

Note that M is finitely generated. Writing an integral affine map α⋅x+a, we can write the monomial zα⁢ta, and this gives us ζ=⌈zα⁢ta⌉. Choosing generators for M, we have a map 𝕋P→ℂk given by (ζ1,…,ζk). The image of this map will be called ⌈𝕋P⌉.

Finally, we can write

where h:ℂk→ℂ× is regular.

More generally, our coordinate charts for exploded manifolds will be ℝk×𝕋Pm, where the smooth part and sheaf of exploded functions are as expected.

Example 1.2.7.

One relatively simple example is 𝕋[0,∞)mm=𝖤𝗑𝗉𝗅(ℂ,0)m.

Example 1.2.8.

Another simple example is 𝕋(0,∞)1=𝖤𝗑𝗉𝗅⁡(Spec⁡ℕ→ℂ).

Definition 1.2.9.

An exploded manifold is an exploded space which is locally isomorphic to ℝk×𝕋Pm.

Exercise 1.2.10.
  1. (a)

    Maps f:B→𝕋Qn correspond to n-tuples f1,…,fn∈ℰ×⁢(B) such that (f¯)∈Q;

  2. (b)

    Maps 𝕋Pm→𝕋Qn correspond to integral maps f:P→Q;

  3. (c)

    Maps f:B→ℝ correspond to global sections f−1⁢(ex⁢t0)∈ℰ×⁢(B) valued in ℝ>0⁢t0, which are the same as regular maps ⌈B⌉→ℝ.

Example 1.2.11.

We will now consider the example of a compact polytope. As a set, we have

Geometrically, the copies of the ℂ× will be asympotically cylindrical only at one end. Note that the M here is generated by two elements (zero at one of the two endpoints), where ζ1=⌈z⌉ and ζ2=⌈z−1⁢tℓ⌉. Next, we note that

which is the model for a node.

Exercise 1.2.12.

Show that 𝕋[0,ℓ]1={z1⁢z2=tℓ}⊆𝕋[0,∞)2=𝖤𝗑𝗉𝗅⁢(ℂ,0)2.

Note that there is a map 𝖤𝗑𝗉𝗅⁢(ℂ,0)2→z1⁢z2𝖤𝗑𝗉𝗅⁡(ℂ,0), and so we see 𝕋[0,1]1 as a fiber of this map.

1.2.3 Geometric structures

As usual, vector fields will correspond to real derivations on ℰ⁡(B), which is a sheaf of (semi)-rings. Recall that a derivation v must be linear and satisfy the Leibniz rule:

Then the condition of being real means that if f is ℝ⁢tℝ-values, so is v⁡(f). We also require that

for constant functions. For example, there is an isomorphism

generated by the vector fields

Definition 1.2.13.

A vector v∈Tpℤ⁢B⊂Tp⁢B is integral if v⁢ff∈ℤ for all f∈ℰ×⁢(B).

Definition 1.2.14.

An almost complex structure J is a regular section of T∗⁢B⊗T⁢B such that J2=−Id and (J⁢v)⁢f=i⁡(v⁢f) for any v∈Tpℤ⁢B.

Definition 1.2.15.

An exploded manifold B is complete if ⌈B⌉ is compact and locally isomorphic to ℝn×𝕋pm for a closed polytope P.

1.3 Families of exploded curves

1.3.1 Exploded curves

Definition 1.3.1.

A curve C is a complete, 2-dimensional exploded manifold with an almost complex structure. A regular curve in an almost complex exploded manifold (B,J) is a regular map f:C→B from a curve. Such an f is holomorphic if df∘j=J∘df.

The local models for exploded curves are open subsets of:

  • •

    ℂ, corresponding to regular points;

  • •

    𝕋[0,ℓ]1, corresponding to nodes;

  • •

    𝕋[0,∞)1, corresponding to ends.

The underlying topological space ⌈C⌉ is simply a nodal curve, while the tropical part will have vertices corresponding to smooth components, edges corresponding to nodes, and infinite legs corresponding to the ends. At every node, we will have an extra choice of ℂ×⁢t(0,∞) gluing information, which makes the tropical part a metric graph.

Automorphisms of holomorphic curves are holomorphic automorphisms h:C→C which commute with f. These induce automorphisms of ⌈f⌉, but not every automorphism of ⌈f⌉ respects the gluing data, so they may not lift. Thus there are exploded curves with finite automorphism group as curves, but whose underlying curves are not stable.

Definition 1.3.2.

A curve f is stable if the automorphism group Aut⁡⌈f⌉ is finite.1

1.3.2 Families

Definition 1.3.3.

A family Π:B̂→B0 is a complete map such that for all p∈B̂, the maps

are surjective.

For example, the explosion of any normal crossings degeneration is a family in our sense.

Definition 1.3.4.

A regular family of curves f̂ in a family B̂ is a regular commutative diagram

such that the left arrow is a family whose fibers are curves. A morphism α:ĝ→f̂ is a pullback diagram

Using this, we may define the moduli stack ℳ∞,1⁢(B) of C∞,1 curves in B. Then there is a sequence of inclusions

where ℳ is stable holomorphic curves and ℳst is C∞,1 stable curves. This carries a topology, where an open substack U⊂ℳ∞,1 is such that its preimage U⁡(f̂)⊂F⁡(f̂) is open and for any morphism α:f̂→ĝ, F⁢(α)−1⁢U⁢(ĝ)=U⁡(f̂).

1.3.3 Gromov compactness

Given an appropriate taming form ω on B, the open substack

of stable curves with at most a prescribed energy E is in fact compact. Equivalently, any sequence of ℳg,n,E has a subsequence fi such that there exists a regular family f̂ and maps αi:fi→f̂ such that ⌈F⁡(α)⌉⊂⌈F⁡(f̂)⌉ converges.

1.3.4 Kuranishi structure

Recall that Tf⁢ℳst give first-order deformations of f. We can also consider

which is a section of Y⁡(f)=(T∗⁢C⊗f∗⁢T⁢B)0,1 that vanishes on the edges of C. We will call the sections of this 𝒴⁡(f) and this is a vector bundle on ℳst. For any holomorphic f, we obtain a linearization

of the ∂¯-equation.

Theorem 1.3.5.

If Df⁢∂¯ is surjective, then there exists an open neighborhood U⊂ℳst of f and a regular family f̂⊂U with finite automorphism group G such that ℳ∩U=f̂/G.

Therefore, when we have transversality, the moduli space is an exploded orbifold. Unfortunately, this does not always hold.

Theorem 1.3.6.

For an arbitrary f∈ℳ, there exists an open U⊂ℳst and a nice finite rank V⊂𝒴 over U, f̂⊂U, and G as above such that ∂¯−1⁡(V)=f̂/G and and ℳ∩U=∂¯−1⁡(0).

1.3.5 Refinements

Definition 1.3.7.

A refinement of an exploded manifold B is a complete, bijective submersion B′→B.2

If we consider the fiber product diagram

of a refinement with itself, it must always look like the above.

Example 1.3.8.

Let (Xn,D) be a compact toric variety. Then there is a refinement map

whose tropical part is the toric fan of X in ℝn.

If we do a subdivision of the tropical part of an exploded manifold, then on the smooth part we obtain some kind of degeneration. For example, refining the toric fan of ℙ2 as gives a new scheme. Importantly for us, the virtual fundamental class [ℳ⁡(B′)]vir is a refinement of [ℳ⁡(B)]vir.

Finally, we would like to discuss the relationships between various moduli spaces. There is a diagram as follows:

Example 1.3.9.

Consider the example of lines in ℙ2. If we move a general line into one of the boundary divisors, the tropical picture gets shifted in the direction we move in. Rescaling (following Ionel), we can produce a new rubber, and the curve then lives in the rubber, as in Figure 1.1.

↝↝
Figure 1.1: Tropical shift

A more complicated example is shown in Figure 1.2.

↝↝
Figure 1.2: More interesting tropical shift

The moduli space is simply ℳ⁡(𝕋2)=𝕋2, and after the refinement, we obtain that ℳ⁡(ℙ2) has tropical part the toric fan of ℙ2 blown up at the three fixed points. The smooth part is simply ℙ2 blown up at 3 points, and the moduli space itself is its explosion. As in the case of relative enumerative geometry, the rubber components need to be quotiented out by symmetries in the approaches of Ionel and Tehrani. However, we should note differences in the two approaches, as seen in Figure 1.3.

Figure 1.3: Ionel’s approach can distinguish between the two situations, Tehrani’s cannot.

1.4 Computations with Calabi-Yau manifolds

We will do another example of a surface, and then we will move on to threefolds. Consider X=Bl(0,1)⁡(ℙ1×ℙ1) with the strict transform of the toric divisor. Our goal is to compute the number nd of degree d covers of the exceptional divisor E. The blowup modifies our integral affine structure by introducing a cut as in Figure 1.4.

cut
Figure 1.4: Cut

We can compute n1 by using the gluing formula for a tropical invariant as in Figure 1.5:

=
Figure 1.5: Computing n1

We then see that 1=1×n1, so n1=1.

To compute nd, we obtain the picture in Figure 1.6, which gives us

=?
Figure 1.6: Computing nd

If you are a combinatorialist, it is easy to obtain

We will now consider some threefold examples. Let B be a Calabi-Yau 3-exploded manifold. Then B¯ is a 3-dimensional integral affine manifold with codimension 2 singular locus Γ. In some sense, we generically expect that Γ has

  • •

    Edges (span a wall coming out of the edge where holomorphic curves can go);

  • •

    Framing changes (where the edge changes the direction by the wall does not);

  • •

    Positive vertices where three edges meet with the sum of their normal directions is 0;

  • •

    Negative vertices whose normal directions are all the same.

In principle, we can compute invariants for all of these local models, but this is actually very difficult.

The nonsingular case is simply 𝕋3. There are inward arrows v,w,−v−w and we obtain (via a quantum deformation)

We then obtain

where q12=ei⁢ℏ2.

We will now consider 𝖤𝗑𝗉𝗅⁡(X,D)×𝕋 (the edge). In fact the picture in Figure 1.5 is still valid, but instead we obtain 1=[1]q×n1, so n1=1[1]q. To compute nd, we have the formula

which yields us

We will now consider the case of curves lying on a wall. For each edge on the singular locus, we need to glue over 𝖤𝗑𝗉𝗅⁡(X,D). Fortunately gluing constrained to the wall factors through intersection with E, so we obtain a Fock space Hℓ associated with the edge.

In any situation as in Figure 1.7, we obtain an operator Wv,ℓ:Hℓ→Hℓ. These satisfy the commutation relation

These are the relations for the quantum torus W1,∞.

vHℓHℓ
Figure 1.7: Obtaining an operator

Performing another tropical computation as in Figure 1.8, we obtain the formula

vw=vw+v+w[v∧w]
Figure 1.8: On the wall

Moving on to the framing change, let the legs be ℓ1 and ℓ2. The framing change is a map

which is in fact an isomorphism of representations of W1,∞. To show this, we consider the tropical picture in Figure 1.9.

Fv=Fv
Figure 1.9: Framing change

Now we consider the positive vertex. We will have ℓ1,ℓ2 pointing inward and ℓ3 pointing out. The vertex will be an intertwiner

of representations of the quantum torus. The pictures are as in Figure 1.10.

v=vv⋅
Figure 1.10: Positive vertex

In the case of the negative vertex (point all edges inwards), the operator N diagonalizes as

There is a nice interpretation of this due to Bryan-Pandharipande in terms of representations of the symmetric group as usual.

2 Logarithmic and punctured Gromov-Witten invariants, tropicalization, and gluing formalism (Bernd Siebert)

2.1 Introduction to log geometry

We will be relatively liberal in the category we work in. We may consider either algebraic varieties or complex manifolds/analytic spaces. These get generalized to algebraic stacks (or analytic stacks), which we will consider later.

2.1.1 Normal crossings divisors

Let X be smooth and D⊆X be a divisor such that locally, D=V⁡(z1⁢⋯⁢zk), where z1,…,zn are local coordinates. This is really bad for algebraic geometry. Instead, we may consider

Definition 2.1.1.

A simple normal crossings divisor is such that there exists U⊆X open and a smooth π:U→𝔸r such that D∩U=π−1⁢(V⁡(z1⁢⋯⁢zr)).

All such D can be written as D=⋃Di, where the Di⊆X are smooth divisors intersecting transversely. In general, we need an étale U→X such that the preimage of D is a simple normal crossings divisor.

For a simple normal crossings divisor D=⋃Di, we obtain line bundles Li with sections si and corresponding sheaves ℒi=𝒪X⁢(Di). Every section si is a map 𝒪X→𝒪X⁢(Di), or equivalently, a map 𝒪X⁢(−Di)→𝒪X. In addition, the normal bundles of Di⊆X is

Iterating, we see that Di∩D⊆Dj is a simple normal crossings in Di with normal bundle

All of this is contained in ℳX≔𝒪X/D×∩𝒪X↪𝒪X. Unfortunately, this is a sheaf of multiplicative monoids and has no additional structure.

Example 2.1.2.

Consider 𝔸z,w2 with D=V⁡(z⁢w). If we consider the sheaf ℳX, away from D we see all h∈𝒪×, on the z-axis we have h⋅wb,h∈𝒪×, and at the origin we see h⋅za⁢wb. As a monoid, we have

We can then take the associated abelian sheaf ℳXgp, which replaces the ℕ-factors by ℤ-factors. This has a discrete part ℳ¯X=ℳX/𝒪X× with quotient map κ. In general, in the simple normal crossings case, we have

We may recover the line bundles in the simple normal crossings case. Taking the groupification, we obtain

and thus ℳXgp is a 𝒪X×-torsor. Considering 𝒯≔κ−1⁢((a1,…,ar)), we can brute force the sheaf

2.1.2 Toric geometries

Consider a finitely generated submonoid P⊆(ℤn,+). Explicitly, we can write P=ℕ⁢m1+⋯+ℕ⁢mr. The most imporant case is when we take σ∨∩ℤn, where σ⊆(ℝn)∗ is a rational polyhedral cone. It is customary to write (ℝn)∗≃N⊗ℝ and P⊆M for the two different ℤn.

Using P, we obtain a finitely-generated ℂ-algebra ℂ⁡[P]={∑m∈Pam⁢zm}, where the sums are finite. Explicitly, we can consider the map

which give equations for the ring. The relations always come from relations in P, so ker⁡φ is generated by binomial equations zm1⁢zm2=zm1′⁢zm2′ (in the saturated case). Applying the Spec functor, we obtain

In fact, we always have P⊆σ∨∩M with σ=Hom⁡(P,ℝ≥0) with equality if and only if P is saturated.

Example 2.1.3.

Consider P⊆ℕ⋅2+ℕ⋅3⊆ℤ=M. Then ℂ⁡[P]=ℂ⁡[x,y]/(x3−y2).

Examples 2.1.4.
  1. (a)

    Consider the cone generated by (0,1) and (k,1) in Nℝ. Then the dual cone is generated by (−1,k) and (1,0) over ℝ, and the monoid P=σ∨∩M is generated by (1,0),(0,1), and (−1,k). Writing the corresponding variables as z,w,t, we obtain

    which is the Ak−1 singularity. Note that this is the base change under t↦tk of the normal crossings degeneration (z⁢w−t).

  2. (b)

    Let σ∨ be the cone over the convex hull of the square {(0,0),(1,0),(1,1),(0,1)}. Then

    is a tensor product

Because we have P=σ∨∩M, we obtain ℂ⁡[P]⊆ℂ⁡[M]. This gives us an M-grading, which induces a (ℂ×)n-equivariant embedding

Definition 2.1.5.

A toric variety is a (ℂ×)n-equivariant partial compactification of (ℂ×)n.

Remark 2.1.6.

We will want P to be saturated, which corresponds to the toric variety being normal.

Definition 2.1.7.

The toric divisor is the complement xσ∖(ℂ×)n, and its components Di can be read off from the facets of P (or Pℝ=ℝ≥0⋅P).

An alternative construction of toric varieties is as follows. Consider N≃ℤn. Then let σ⁡(1) be the set of rays, labelled ρ1,…,ρr. Then we define the Cox ring to be

Then there is a map M→Map⁡(σ⁡(1),ℤ) given by m↦(ρi↦⟨m,ni⟩), where ni is the primitive generator of ρi. Grading R by Γ=Map⁡(σ⁡(1),ℤ)/M, we obtain an action of (ℂ×)r−n-action on R. Taking the categorical quotient Spec⁡RΓ, we obtain Xσ.

Exercise 2.1.8.

Show that RΓ=ℂ⁡[P] where P=σ∨∩M.

The upshot is that an affine toric variety can be written as Spec⁡ℂ⁡[P]=𝔸r⫽0(ℂ×)r−n.

In this case, if X=Xσ, then generalizing ℳ𝔸n=𝒪𝔸n∖V⁡(z1⁢⋯⁢zn)∩𝒪𝔸n↪𝒪𝔸n, we can write

As long as 0 is the only invertible element of P, we have

Now we have an embedding

The image corresponds to principal Cartier divisors.

Example 2.1.9.

If σ is the cone over the square, then D=∑ai⁢Di is Cartier if and only if a1+a3=a2+a4.

2.1.3 Abstract log structures

Consider α:ℳX→𝒪X in the étale topology such that α−1⁢(𝒪X×)→𝛼≃𝒪X×. This definition is useless in this generality, but it does automatically provide us with

as before.

We will instead provide an alternative point of view, due to Deligne-Faltings. We want the following data:

  • •

    A sheaf of finitely generated ℳ¯ inside ℳ¯Xgp constructible;

  • •

    For all U⊆X, a map

  • •

    We have compatibility ℳ¯→𝒟⁢ivX which is morally a symmetric monoidal functor.

Examples 2.1.10.
  1. (a)

    Consider log points, which are given by Spec⁡(Q→ℂ)=(Spec⁡ℂ,Q⊕ℂ×). Here, we need Q a finitely generated monoid with Q×={0}, and we have

  2. (b)

    We can consider pullback log structures for morphisms f:Y→X with a log structure ℳX on X. We obtain a log structure on Y by writing

    and send (s,h)↦f∗⁢α⁢(s)⋅h.

Note that if Q=σ∨∩M and 0↪Xσ is the inclusion of a 0-dimensional orbit, the pullback log structure is f∗⁢ℳXσ=Spec⁡(Q→ℂ).

An important class of examples are those with charts. Consider an open set U⊆X (maybe in the étale topology) and a map f:U→Spec⁡ℂ⁡[P], where P is as in the previous subsection, with isomorphisms ℳX|U≃f∗⁢ℳSpec⁡ℂ⁡[P]. The fine log structures are those with local charts and a fine saturated log structure is one with P saturated.

Definition 2.1.11.

A log morphism f:(X,ℳX)→(Y,ℳY) is a map f:X→Y (which gives us f♯:f−1𝒪Y→𝒪X) and a morphism of sheaves f♭:f−1⁢ℳY→ℳ𝒳 making the diagram

commute.

We will refer to ℳ¯ as the ghost sheaf. Others call it the characteristic sheaf, but Bernd prefers to call it the ghost sheaf.

2.1.4 Log smooth morphisms

These morphisms are locally given by toric morphisms. Suppose we have a morphism Q→P of monoids. Then we require the existence of a diagram

Example 2.1.12.

The basic example is the diagram

Examples 2.1.13.
  1. (a)

    We can allow Q=0. Then we require the existence of a divisor D⊆X such that (X,D) is a toroidal pair.

  2. (b)

    If we consider Q=ℕ and P=ℕr with the map 1↦(1,…,1), these are base changes of normal crossings degenerations.

  3. (c)

    Let 0†=Spec⁡(ℕ→ℂ) be the standard log point. Then if Y is a curve, the central fiber

    is a normal crossings degeneration.

  4. (d)

    Let f∈ℂ⁡[z0,…,z3] be homogeneous of degree 4. Define

    This is singular, where the singular locus (X0′)sing∩V⁡(f) consists of 24 A1 singularities. If we consider X0′⊆X′, this is a very bad at the singularities, meaning they are not fine. To resolve this, we can simply blow up the singularities in some order to obtain a smooth total space X which is a normal crossings degeneration and hence log smooth.

    Suppose we have a normal crossings degeneration X0⊆X. Then this X0 is d-semistable in the sense of Friedman. What this means is that

    is a trivial bundle. He proved that a d-semistable K3 surface is smoothable. In higher dimension, this is very tricky.

    This story can be reinterpreted in terms of log geometry following Kawamata and Namikawa. They showed that a normal crossings variety X0 is d-semistable if and only if there exists a log structure ℳX0 and a log smooth morphism

    All of this has a symplectic analogue due to McLean-Tehrani-Zinger, which states that symplectically d-semistable implies symplectically smoothable.

2.1.5 Kato-Nakayama spaces

For any (X,ℳX), this is a topological space (X,ℳX)KN→X which provides a “topological smoothing”, where if we have a diagram

then the KN functor will give us a map X0KN→S1 which should be thought of as a restriction of π to the locus lying over a small circle.

2.2 Tropicalization

Traditionally, we would begin with a valued ℂ-algebra, for example

If we choose an ideal I⊆K⁡[z1±,…⁣zn±], we obtain an algebraic subvariety V⁡(I)⊆(K×)n. Applying the valuation

we can define the tropicalization in various ways:

where 𝗍𝗋𝗈𝗉(I)⊆(ℝ,max,+)[x1±,…,xn±] is the tropicalization of I.

If we have convergent Puiseux series, we can consider (Xt)t≠0⊆(ℂ×)n, and then take

2.2.1 Tropicalization in log geometry

Consider a fine log space X=(X¯,𝖬X). For all x∈X, we will write Px=ℳ¯X,x. This is a finitely generated monoid with Px×={0}. In algebraic geometry, if y∈X is a non-closed point and x is a specialization of Y, then there is a generization map

Here, F⊆Px is a face {p∣α⁡(p)∈𝒪X,y×}.

In the analytic world, we can instead find an open set U around x such that ℳ¯X⁢(U)→ℳ¯X,x=Px is an isomorphism. Then, we have a diagram

defining the generization map.

Dualizing, consider the rational polyhedral cone σx=Hom⁡(Px,(ℝ≥0,+)). Because we have the surjective generization map which is projection along a face F, we have

In the case of ℙ1×ℙ1, we can recover the correspondence between orbits and the fan, and finally we set

Proposition 2.2.1.

If X is the toric variety corresponding to a fan Σ, then 𝖳𝗋𝗈𝗉⁡X=Σ with the embedding into Nℝ forgotten.

Warning 2.2.2.

It should be noted that if X=Bl1⁡(ℙ1×ℙ1) and D=⋃4ℙ1, then

In fact, if we take any toric surface with four boundary divisors (for example a Hirzebruch surface), we obtain the same tropicalization. This is in fact a feature of the theory.

Example 2.2.3 (Whitney umbrella).

Consider coordinates x,y,z on 𝔸3 and consider D~=V⁡(x⁢y). Consider the ℤ/2⁢ℤ action given by

Taking the quotient, we obtain

Removing the origin, we can set X=𝔸3∖{0} with divisor D~/(ℤ/2⁢ℤ). If we consider ℓ~ to be the z-axis, we note that

so ℳX|ℓ is locally constant (it sees the monodromy of swapping x and y). Finally, we get that 𝖳𝗋𝗈𝗉⁡(X) is the generalized cone complex (locally a limit of cone complexes) shown in Figure 2.1.

lim→id(a,b)↦(b,a)=
Figure 2.1: Tropicalization of the Whitney umbrella

2.3 Log smooth curves

These are simply normal crossings degenerations of smooth curves with marked points. These will be domains for stable log maps, which will define logarithmic Gromov-Witten theory. There are three types of local models for ℳC|C0. Let z be the fiber direction and t be the base direction.

  • •

    At the generic points η of C0, the map is given by (z,t)↦t, so

  • •

    At a marked point p of C0, we can see both the base and fiber directions, so we obtain

    Here, this receives a map from ℕ going into the first factor.

  • •

    At a node q, we have (z,w)↦z⁢w=t, and thus

    This receives a map from ℕ where 1↦(1,1).

Generally, we want to consider curves C→Spec⁡(ℕ→ℂ) which are log smooth, integral, have C¯ reduced, have fine separated log structures, and are of relative dimension 1. In this case, C is a nodal curve, and locally on C:

  • •

    At η, ℳ¯C,η=Q;

  • •

    At p, ℳ¯C,p=Q⊕ℕ;

  • •

    At q, ℳ¯C,q=Q⊕ℕℕ2. Here, the map from ℕ is given by 1↦(s,(1,1)). We should note that

In the universal case, we have Q=ℕnumber of nodes. Interpreting this as a (semi-)universal deformation, where each node gets smoothed in a different direction. The log structure is then given by

If the curve is stable, we obtain the moduli space ℳ¯g,k with universal curve 𝒞g,k. Considering the nodal locus Dg,k, we also have normal crossings divisors (in the sense of Deligne-Mumford stacks) D̂g,k∪Γk, where the Γk are the marked points. This gives us a log smooth smooth morphism

Restricting to one fiber, we obtain the universal log structure on stable curves.

2.3.1 Tropicalization of log curves

Over the standard log point, consider a curve C→𝜋0†. Tropicalization will give us a tropical curve

In fact, if Γ denotes the fiber over 1, then in fact ΣC is the cone over Γ as a polyhedral complex.

  • •

    At η, we obtain a vertex Vη of Γ. Note that Hom⁡(ℳ¯C,η,ℝ≥0)=ℝ≥0, so we get a copy of a line.

  • •

    At a node q, we have

    Therefore, we can consider the two projections x1 (to the y-axis) and x2. Dualizing, σ is the cone generated by (1,0) and (1,k), and so πtrop is the projection to the first coordinate. Therefore, we obtain an interval of length k.

  • •

    At a marked point p, we obtain a copy of ℝ≥0.

The upshot is that Γ is the dual graph of C with vertices corresponding to components, edges corresponding to nodes, and legs corresponding to marked points. We also give Γ a ℤ-affine structure, which means that each of the edges have a length k, which corresponds to the local equation of the node (an Ak−1 singularity).

Over a more general log point Spec⁡(Q→ℂ), the tropicalization ΣC lives over the cone τ=Hom⁡(Q,ℝ≥0). Note that faces of τ corespond to collapsing some edges. In the universal case, then τ=ℝ≥0number of nodes.

2.4 Logarithmic Gromov-Witten invariants

2.4.1 Stable log maps

As in the case of (non-logarithmic) stable maps, we can simply write down the definition as expected (although the automorphisms of the underlying map need to be finite). The targets will be a fine separated log scheme X=(X¯,ℳX), and later we will require our target to be log smooth and projective.

Over a log point W≔Spec⁡(Q→ℂ), we will have a map f:C→X, where C is a log curve over W. Note that we need Q≠0 to allow nodal domains. The stability condition is that C¯→X¯ is a stable map in the ordinary sense.

Problem 2.4.1.

The set of possible values for Q are not bounded. This is common in log moduli problems, so we need to use the geometry of the moduli problem to constrain the choice of Q. For example, for nodal curves, we will choose Q=ℕnumber of nodes.

The solution here is to consider basic stable log maps. If we consider the tropicalization

this is something called a tropical stable map. Now for all s∈τ∘, the map

has the same type, which is given by

  • •

    The combinatorial type of Γs. In other words, no contraction of edges is allowed;

  • •

    The smallest cells σ⁡(−) of ΣX containing h⁡(Vη),h⁡(Eq),h⁡(Lp);

  • •

    The contact orders up at Lp and uq at Eq. For example, at a marked point, hs maps the ray Lp (with endpoint Vη) to a ray inside σ⁡(p) starting somewhere on σ⁡(η). Then the unit vector along this ray is called up. For an edge, the story is similar, but it depends on an implicit orientation of Eq.

We now obtain a (local) tropical moduli space of type

where τ is also a rational polyhedral cone. The moduli space parameterizes tropical space maps of a given type.

Examples 2.4.2.
  1. 1.

    Let X=ℙ2 with the toric log structure and C be a line. We will take the rays of Σℙ2 to be the opposite of the usual ones and suppose that up1=(1,1),up2=(−1,0),up3=(−1,−1) as in Figure 2.2. We then obtain

    This forces the curve to have two components, where the one containing p3 maps to the vertical divisor and the one containing p1,p2 is collapsed to a point.

    abEqup2=(−1,0)up1=(1,1)up3=(−1,−1)1,23
    Figure 2.2: Most generic (tropically) case

    Taking the a→0 limit as in Figure 2.3, then τ=ℝ≥0 and the image of p3 moves along the bottom divisor. In the b→0 limit, shown in Figure 2.4, we unbreak the curve, and taking the total limit as (a,b)→(0,0) (see Figure 2.5), we obtain the most general (in the geometric sense) situation.

    b1,23
    Figure 2.3: a→0 limit of the previous example
    a123
    Figure 2.4: b→0 limit of the previous example
    123
    Figure 2.5: (a,b)→(0,0) limit of the previous example

    This appears from a degenerating family of general lines. If we consider a family C→𝔸t1 with maps C→ℙ2, then ΣC0 can be any of the possibilities we discussed before.

    Recall that the expanded degneration is point of view was done originally by Jun Li in the case of a smooth divisor and by Ranganathan for a general divisor. In our world, 𝖳𝗋𝗈𝗉⁡(C) defines a cone subdivision Σ of ΣX×ℝ≥0 (if dimX>2, choose a subdivision such that Γ⊂Σ⁡(1)).

    In the previous example, let X=XΣ, which is a blowup of ℙ2×𝔸1. Then over 𝔸1∖{0}, we simply obtain ℙ2×𝔸1∖{0}, but

    where Σρ is some fan. In the dual picture, we obtain a polyhedral decomposition. Therefore, we obtain a morphism from the expanded degeneration moduli space to our moduli space. An example expanded degeneration is shown in Figure 2.6.

    ℙ1×ℙ1ℙ2ℙ2
    Figure 2.6: Example of an expanded degeneration
  2. 2.

    This example will have τ≠ℝ≥0k. We will choose X=(ℙ1,0) and suppose C has five components in a chain with C2,C3,C4 each containing one marked point. Then the contact orders Lp1,Lp2 are vertical while Lp3 is finite. In the end, we obtain

Definition 2.4.3.

A stable log map

is basic if and only if ΣC→ΣX is the universal tropical stable map of some type.

2.4.2 Moduli spaces

Theorem 2.4.4 (Abramovich-Chen, Gross-Siebert).

There exists a good moduli space of basic stable log maps which is a Deligne-Mumford stack (with log structure) locally of finite type over ℂ that fulfills the valuative criterion of properness and has a perfect obstruction theory to the stack 𝔐 of log curves with any fine separated log structure on the base. Fixing the topological data

the stack is proper and we obtain a virtual fundamental class [ℳ⁡(X,β)]vir.

2.5 Artin fans

Consider P=σ∨∩M. Suppose that Xσ has dimension n, and then we will define the toric stack

For any scheme W, any map W→𝒜σ is a diagram

of a 𝔾mn-bundle Y→W and an equivariant map Y→Xσ.

This has the property that if τ⊆σ is a face, then 𝒜τ⊆𝒜σ is an open substack. Recall that

and therefore Xτ×(ℂ×)codim⊆Xσ is open. For example, we can consider

Now for any X=(X¯,ℳX), we obtain a diagram of cones ΣX.

Definition 2.5.1.

The Artin fan of X is the algebraic stack

The Artin fan algebraizes tropical geometry in the following precise sense.

Proposition 2.5.2.

Assume that 𝒳 has a Zariski covering by various 𝒜σ (for example if X is a Zariski fine separated log scheme which is log-smooth over ℂ). Then for all finite separated log-schemes T, there is a canonical bijection

As an application, we may define the algebraic stack ℳ⁡(𝒳) of log maps into 𝒳, whose data consists of a domain C→Spec⁡(Q→ℂ) and a tropical stable map

2.6 Log Gromov-Witten invariants of fixed type

Let τ be the type of tropical stable maps to Σ𝒳=ΣX.

Definition 2.6.1.

A marking of a stable log map

by τ is an identification of τ with a face of

These markings define closed substacks ℳ⁡(𝒳,τ)⊆ℳ⁡(𝒳). Pulling back to X, we obtain the Cartesian diagram

In order to have Gromov-Witten invariants, we need a virtual fundamental class. We need the following:

  • •

    The stack ℳ⁡(𝒳,τ) is pure-dimensional of dimension 3⁢g−3+k+dimB−dimτ;

  • •

    The morphism ℳ⁡(X)→ℳ⁡(𝒳) is virtually smooth, so ℳ⁡(X,τ)→ℳ⁡(𝒳,τ) is as well.

Now we can define

2.7 Punctured Gromov-Witten invariants and the gluing formalism

2.7.1 Rigid tropical curves and virtual decomposition

Example 2.7.1.

Let X′=V⁡(t⁢f3⁢(z0,z1,z2,z3)+z0⁢⋯⁢z2)⊆ℙz0,…,z33×𝔸t1 be a degeneration of cubic surfaces. We resolve X′ to obtain X, whose central fiber X0 is a union ⋃3Bl3⁡ℙ2, as displayed in Figure 2.7. Then Σ⁡(X)=Σ⁡(X0) is the cone over the unit right triangle.

p1p2
Figure 2.7: Central fiber X0

Now consider the invariant of g=0 curves of degree d=3 passing through p1,p2. There are 12=9+3 such curves. The two possibilities are shown in Figure 2.8. The first gives 3⋅3=9 and the second gives 3 possibilities.

p1p2Ep2p1p1p2p1p2
Figure 2.8: Two possibilities (above and below)

We have the decomposition result

In this example, we obtain 12=(3+3+3)+3.

2.7.2 Splitting

Now we will consider X→B where B is a point or a curve over a log point. Our goal is now to compute

(whether or not τ is rigid) by splitting τ along edges.

Example 2.7.2.

We will consider the splittings as in Figure 2.9. Split edges give a pair of “punctured points” which in turn produce punctured stable maps.

split
Figure 2.9: Splittings τ1,τ2,τ3,τ4 of τ
Theorem 2.7.3.

There is a Cartesian diagram

where the vertical arrows are virtually smooth. The “ev” makes the bottom arrow representable and finite, and is defined by

where there is one choice for each puncture from splitting or node to split. We should note that the morphism X→𝒳 is smooth (in the ordinary sense), so the fiber product is a harmless operation.

Corollary 2.7.4.

If we know δ∗ev⁢[ℳev⁢(𝒳,τ)] on ∏iℳev⁢(𝒳,τi), we obtain a gluing formula GW⁡(τ)=F⁡(GW⁡(τi)).

Theorem 2.7.5 (Gluing theorem).

There is an fs-Cartesian diagram

Here, the tilde gives a bigger log structure, and the prime changes the non-reduced structure. Finally, fs-Cartesian means that the fiber product is modeled on the fiber product of cones in the following sense for toric varieties:

For example, if Xτ=𝔸2 and Xσ1,Xσ2 are lines with different slopes, then the fs-fiber product is empty.

By work of Yixian Wu, the gluing stratum of X (contained in type τ) are toric, so we obtain the gluing formula

The ω are obtained as solutions to perturbation problems for τ via displacement along edges. There is a similar formula in a different setting due to Venugopalan-Woodward.

Example 2.7.6.

Consider a degeneration of ℙ1×ℙ1 to ⋃2ℙ2 via the cone over the skeleton in Figure 2.10.

Figure 2.10: Skeleton

If we consider curves of bidegree (1,1), then τ is simply the skeleton. We can split along the diagonal edge and then perturb by the tangent vector ξ=(1,0). The possibilities for ω1,ω2 are shown in Figure 2.11.

ξξ
Figure 2.11: Possible perturbations of the skeleton

2.7.3 Punctured curves and punctured GW invariants

Definition 2.7.7.

A punctured curve C∘=(C¯,ℳC∘)→Spec⁡(Q→ℂ) is defined almost as a log curve is before, but instead of Q⊕ℕ at marked points, we admit submonoids Q∘⊆Q⊕ℤ such that Q⊕ℕ⊆Q∘ and for any (g,k) with k<0 implies that α⁡(q,k)=0. This is a fine log structure, but is not separated.

Example 2.7.8.

Let C=C1∪C2 with node q. Let ι:C1↪C and q=ι⁡(p). Then

is a puncturing at p.

Definition 2.7.9.

A stable punctured map is a stable map C∘→X such that ℳ¯C∘,p is generated by Q⊕ℕ and f¯♭⁢(ℳ¯X,f⁡(p)) for all punctures p.

The second condition on generation of ℳ¯C∘,p is equivalent tropically to saying that the bounded leg Lp for the punctured point p extends as far as possible. After some work, we may define tropical punctured maps, their types, and moduli spaces of punctured stable maps as before.

Example 2.7.10.

Recall the example of Bl1⁡(ℙ1×ℙ1), where we blow up on the interior of one of the boundary divisors. Then we consider a line degenerating to the strict transform of the blown up boundary divisor. We will have two marked points and one puncture which moves to the exceptional divisor. The classical and tropical pictures are shown in Figure 2.12.

E
Figure 2.12: A comparison of punctured and unpunctured invariants. Punctures are blue and the exceptional divisor is orange.

The virtual dimension is 0, and in fact we have

As an application, we may consider a degeneration X→S of a Calabi-Yau or (X,D) a log Calabi-Yau. Let B⊆Σ⁡(X) be the Kontsevich-Soibelman skeleton, which is an integral affine manifold. From 1-punctured invariants, we obtain a wall structure on B (for example in the previous example, consider E) with walls swept out by images of tropical punctured maps of wall type, which are not defined here. From the wall structure, we produce an intrinsic mirror ring from 1-punctured invariants with up to 2 marked points. Then the formula of Yixian Wu tells us that enumerative and algebraic wall crossings are the same. For a reference, see The canonical wall structure and intrinsic mirror symmetry by Gross and Siebert.

3 A spin on Hurwitz theory and topological recursion (Danilo Lewanski)

The impatient reader may skip to Figure 3.4 for a flowchart which summarizes this chapter.

3.1 Hurwitz theory

We are interested in studying d:1 branched covers f:C→D from a curve C of genus g to a curve D of genus h. Then we know that f has k ramification points, above which we see ramification profiles, which are partitions μi of d. Therefore, Hurwitz theory enumerates branched coverings of Riemann surfaces with specified ramification data. There are several common specializations:

  1. 1.

    We choose one or two partitions to be special, and their parts are variables. All other ramification profiles are of the same type.

  2. 2.

    We can also set h=0, so we are looking for coverings of ℙ1.

Definition 3.1.1.

Let B be a connected and compact Riemann surface and fix x1,…,xk∈B and partitions μ1,…,μk. Then the Hurwitz number is defined to be

where the sum is over equivalence classes [f] of branched covers f. We will write Hd∘ to enforce connected domains and Hd∙ to have possibly disconnected domains.

Examples 3.1.2.
  1. 1.

    Consider the map

    Then we know that Aut⁡(f)=μd is the d-th roots of unity, and therefore

  2. 2.

    Let E be given by the equation y2=x3+a⁢x+b. There is a unique map f:E→ℙ1 of degree 2. This map has a unique involution given by y↦−y, and therefore

Exercise 3.1.3.

We list a few more invariants:

  1. 1.

    H∘⁢(S1×S1;(2,1)4)=9;

  2. 2.

    Compute H∘⁢(Σ2;(3),(2,1)6) and H∘⁢(Σ5,(3)4,(2,1)6).

It is unclear from the specification of the problem that a single cover even exists for a given g,h,μi. We of course do have a classically known linear condition.

Theorem 3.1.4 (Riemann-Hurwitz).

For any f, we have

This theorem is proved by lifting a triangulation of B passing through the xi. Any cover satisfies this formula, but not all data which satisfy the formula admit a connected covering.

3.1.1 Spin Hurwitz numbers

Definition 3.1.5.

A spin structure on B is a holomorphic line bundle θ on B such that θ⊗2≅ωB.

Examples 3.1.6.
  1. 1.

    Let B=ℙ1. Then there is a unique spin structure given by 𝒪⁡(−1).

  2. 2.

    If θ is a spin structure on B and all μji are odd, then

    is a spin structure on the domain of f. Here, xi(j) is the j-th preimage of xi. This follows from the formula

    Note that if we compute the degree of both sides, we obtain the Riemann-Hurwitz formula.

Definition 3.1.7.

The spin Hurwitz number is defined by

where Arf(f:C→B)=h0(C,θf).

It turns out that all spin Hurwitz numbers are positive, but this is somewhat unexpected. However, we can enumerate spin structures on curves.

Theorem 3.1.8.

On any curve of genus g, there are 2g−1⁢(2g+1) even spin structures and 2g−1⁢(2g−1) odd spin structures.

In Table 3.1, we can see these numbers for low genus.

Table 3.1: Enumeration of even and odd spin structures
g even spin structures odd spin structures
0 1 0
1 3 1
2 10 6
3 36 28

3.1.2 Burnside formula

Problem 3.1.9.

Computing Hurwitz or spin Hurwitz numbers directly from the geometry is extremely hard.

The solution to this is that there exists an automorphism-preserving bijection

where

is an extension

defined by Giachetto-Kramer-Lewanski.

For example, we can now compute

where

Also, this 𝔨h is defined in Cavalieri-Miks and is a correction factor. For example, C(2,1,…,1) is the sum of all transpositions and C(d) is the sum of all cycles of length d. This formula is can be generalized to the spin setting, but we will instead give a different formula. We will now restate the formula using characters.

Definition 3.1.10.

For a partition λ, define

By orthogonality of characters, we obtain Fλ⁢Fφ=Fλ⁢δλ⁢φ, so the Fλ form an idempotent basis. Inverting this formula, we obtain

and inserting us into the previous formula, we obtain

Theorem 3.1.11 (Burnside formula).

In the spin setting, there is the following theorem:

Theorem 3.1.12.

where χ~,C~ are the same as the ones without tildes but for 𝔖~d and

Finally, here

It should also be noted that strict partitions must be strictly decreasing.

We will now fix a positive integer r. Classically, we consider (r+1)-completed cycles and Hurwitz numbers with B=ℙ1. We will write

and all other μi will be

where the correction terms generate 12⁢sinh⁡(z/2). In the spin setting, we will consider B=(ℙ1,𝒪⁡(−1)), force the μi to be odd, and have the correction terms generate 12⁢coth⁡(z/2).

Remark 3.1.13.

We should note that we would like to generalize the Gromov-Witten/Hurwitz correspondence of Okounkov-Pandharipande to the spin case, where we consider spin (r+1)-completed cycles. The other motivation is the Zvonkine conjecture/theorem relating to the ELSV formula.

Definition 3.1.14.

Define the function

where

Adding a tilde to everything, we get the spin version. Here, we have

In the spin case, we define

Then the completed cycles are defined to be

In the spin case, the definition is the same.

Definition 3.1.15.

We will define the (r+1)-completed cycles Hurwitz numbers by

where

In the spin case, simply consider (ℙ1,𝒪⁡(−1)) and add a tilde to everything,

These are very hard to compute, and closed formulae are very hard, so we instead develop recursions, such as cut and join, topological recursion, and the Toda equations. The following result was conjectured by Pandharipande.

Theorem 3.1.16 (Okounkov).

The series

satisfies the equation

which is the first equation of the Toda lattice.

3.1.3 Operator interpretation of Hurwitz numbers

One of the main ingredients in this is a Fock space interpretation of Hurwitz numbers hg;μ,ν∙,r as a vacuum expectation value of the operator

where

is determined by the Riemann-Hurwitz formula.

Definition 3.1.17.

The vacuum v∅ is defined to be the Maya diagram in Figure 3.1.

−12−321232
Figure 3.1: Vacuum vector

Note that there exists a bijection between partitions and the charge zero sector of Maya diagrams (where the are the same number of black stones on the left and white stones on the right). For example, the partition (3,2,2) corresponds to the Maya diagram in Figure 3.2;

Figure 3.2: Vector corresponding to λ=(3,2,2)

Now we will define the operator

where ψk† attempts to remove a black stone from position k, ψk−m attempts to drop a black stone in position k−m, and

Note that even though αm is an infinite sum, when applied to any charge zero Maya diagram, it only has finitely many nonzero elements. Also, recall that the normal ordering is defined by

We can also build Maya diagrams from partitions by using the Russian notation for partitions as seen in Figure 3.3.

Figure 3.3: Maya diagram and Russian partition notation corresponding to λ=(3,2,2)

If we apply α2, we obtain

This is exactly the same as the Murnaghan-Nakayama rule, which implies that

The αn also satisfy the relations

It now remains to define ℱr+1 such that its eigenvalue is the completed cycle. We want

and the correct definition is

We will also define other operators

where ζ⁡(z)=2⁢sinh⁡(z2). As studied by Okounkov-Pandharipande, these satisfy the relation

Note that if we set z=0, ℰm⁢(0)=αm and if we set m=0, then

In the spin setting, we may define

where we define

to be the uncharged fermion. The commutation relations become

There is also the very nice property that

so the spin operators are either even or odd.

3.2 Topological recursion

Topological recursion was introduced by Eynard-Orantin (with work from Chekhov) from random matrix theory, where integrals of the form

over some space of N×N matrices are considered. In the N→∞ limit, we want to expand these integrals in terms of powers of N−1. There are various matrix models, including the one used to prove Witten’s conjectures and other including Hurwitz numbers. The problem is to compute all correlators in this model, so we first compute the spectral curve of the matrix model and then use topological recursion to compute the correlators, which give the coefficients. It was realized in 2007 that the matrix model is not necessary in this setup, but we can still apply topological recursion from the spectral curve. For example, we can compute both psi-integrals on ℳ¯g,n and the Weil-Petersson volumes from some spectral curves, but these do not have well-defined matrix models. In 2017, Kontsevich-Soibelman found a way to do without the spectral curve, but there must be a replacement. Also in 2017, a “categorification” was introduced by Andersen-Borot-Orantin, called Geometric Recursion.

Definition 3.2.1.

A spectral curve is the data 𝒮≔(Σ,x⁡(z),y⁡(z),B⁡(z1,z2)) and a point p∈Σ, where:

  • •

    Σ is a not necessarily compact or connected Riemann surface;

  • •

    x:Σ→ℂ is a function such that dx is meromorphic with finitely many simple zeroes 𝔠≔{c1,…,cr};

  • •

    y:Σ→ℂ is a meromorphic function not branched at 𝔠;

  • •

    B is a symmetric bidifferential on Σ×Σ with a double pole on the diagonal with biresidue equal to 1 and no other pole. In other words,

Definition 3.2.2.

Topological recursion is a process that takes a spectral curve 𝒮 and returns a system

of symmetric n-differentials on Σn with poles at most at the branch points 𝔠 of x as follows:

First set

Now we need a recursion kernel K. Recall that x has simple poles, so locally there exists a Galois involution σi such that x⁡(z)=x⁡(σi⁢(z)) locally near ci. Then we may set

We are now able to define

This should be thought of as constructing a genus g surface (with n punctures) by either gluing the kernel K (a pair of paints) to a genus g−1 surface (with n+1 punctures) or by using K to glue a disjoint union of surfaces of genus g1,g2.

Example 3.2.3.

This example comes from the paper arxiv:0709.1453 by Bouchard, Klemm, Marino, and Pasquetti, which is about mirror symmetry of toric Calabi-Yau threefolds (here the mirror object is the spectral curve). There is a spin-off, which arises from considering the framed topological vertex as the framing f→∞.

Theorem 3.2.4 (Eynard, Mulase, Safnuk).

Consider the spectral curve 𝒮H with the data x⁡(z)=log⁡(z)−z,y⁡(z)=z,B=Bcan. Topological recursion in this case produces the Hurwitz numbers via the expansion

This example can be generalized to any hg,μ∘,r by replacing x⁡(z)=log⁡(z)−zr.

Example 3.2.5.

Consider the spectral curve Sψ given by x⁡(z)=z22,y⁡(z)=z,B=Bcan. Running topological recursion, we obtain the intersection numbers of psi-classes on the moduli spaces of curves via the expansion

In practice, running topological recursion gives us integrals over ℳ¯g,n of some cohomological field theory, but this requires some technical conditions.

3.2.1 Running topological recursion in practice

We will now check some cases of Theorem 3.2.4. We will begin with the unstable cases, where we compute

and this implies that

We cannot see the Hurwitz numbers here, but we may consider the Lambert function

which has the property that W⁢eW=z. Here, we can see the Hurwitz numbers using the ELSV formula

where we make the conventions

This gives us

In the spin case, there is a similar result.

Theorem 3.2.6 (Alexandrov-Shadrin).

Consider the spectral curve given by the data

which has new poles on the antidiagonal. Usually, the rank of the CohFT is equal to the number of branch points, but in the spin case, we have the action of G=ℤ/2⁢ℤ, and so the the number of branch points is twice the rank of the CohFT.

3.3 The ELSV formula

Consider the cohomological field theory from Theorem 4.3.3

where λi=ci⁢(𝔼g). In the spin setting, we will consider the moduli space

parameterizing curves with a line bundle L such that L⊗r≅Klog⊗s⁢(−∑ai⁢xi), where we require that (2⁢g−2+n)⁢s≅∑ai (this suffices for the existnece of an L). Then we will define the cohomological field theory

where ℒa→r,s is the universal line bundle on the universal curve 𝒞. These Chern characters can be written in the form

where Bm⁢(x) are the Bernoulli polynomials defined by

and κm≔π∗⁢ψn+1m−1 is the pushforward of ψn+1 under forgetting the last marked point. In addition, ja is the gluing map, where ψ′ has label a and ψ′′ has label r−a.

This cohomological field theory can be used to obtain double ramification cycles, Masur-Veech volumes, χ⁡(ℳg,n), ELSV-type formulae for (spin) Hurwitz numbers, the Gromov-Witten theory of ℙ1, and the object Θg,n.

Theorem 3.3.1 (Shadrin, …).

The Hurwitz numbers can be computed as

where μi=[μi]⁢r+⟨μi⟩ is the Euclidean division.

The goal now is to find a spin analogue for this formula.

Theorem 3.3.2 (Lewanski-Popolitov-Shadrin-Zvonkine,15).

Topological recursion yields hg,μ∘,r if and only if the previous theorem holds.

Theorem 3.3.3 (Giachetto-Kramer-Lewanski).

Topological recursion yields hg,μ∘,r,spin if and only if

Note that in the r=s=ai=1 case, Mumford’s formula yields

We can obtain the right hand side via topological recursion, but it is hard to find a geometric interpretation of such. Fortunately it does exist in our setting.

Theorem 3.3.4.

Applying the recipe for topological recursion, Ωspin is almost given by

up to various prefactors and an oddness gluing condition.

In fact, Ωspin⁢(r,1)=(e⁢p12)∗⁢(Wr=2∩(ε22⁢R)∗⁢(r,1)). Here, we consider the diagram

In the simplest possible case, we set r=2. Then R=1, and thus we obtain

In the end, we obtain

Therefore, we can rephrase Theorem 3.3.3 as

Theorem.

Topological recursion yields hg,μ∘,r,spin if and only if

There is a correspondence as follows. With a single Λ, we obtain a 1-dimensional theory, which yields the Gromov-Witten theory of ℙ1, which is Hurwitz numbers. The Spin Hurwitz case corresponds to Kähler surfaces, which are 2-dimensional and yields Λ⁢Λ. In dimension 3, we can consider Mariño-Vafa, the Gromov-Witten theory of Calabi-Yau threefolds, and CohFTs of the form Λ⁢Λ⁢Λ.

3.4 Gromov-Witten theory

Recall that Gromov-Witten invariants of a smooth projective variety X are defined by

Also recall that the virtual dimension of ℳ¯g,n⁢(X,β) is

In the case when dimX=0, we obtain the Gromov-Witten theory of a point, which is simply computing integrals

which by the famous result of Kontsevich is a τ-function for the KdV hierarchy. When dimX=1, the Gromov-Witten theory of curves was studied in three papers by Okounkov-Pandharipande. In this section, we will consider dimX=2.

Recall that a spin curve is a curve (C,θ) where θ⊗2≅ωC. If D∈|KX| is a smooth canonical divisor, then (D,𝒩D/X) is a spin curve by the adjunction formula

Theorem 3.4.1 (Lee-Parker, Kiem-Li).

The Gromov-Witten theory of surfaces with smooth canonical divisors localizes on spin curves. In particular,

Here, note that virdim⁡ℳ¯g,n⁢(D,d)=g−1+n+d⁡(1−g⁡(D)).

Theorem 3.4.2 (Kiem-Li).

In the case when g⁡(D)=0, recall there is only one spin structure on ℙ1. Then

where we have the diagram

of the universal map to ℙ1.

3.4.1 Spin GW/H

Following the strategy of Okounkov-Pandharipande, there is a spin GW/H correspondence, but unfortunately the degeneration formula for spin curves is not understood.

Conjecture 3.4.3 (Lee, 2019).

The stationary Gromov-witten invariants are given by

Theorem 3.4.4 (Giachetto-Kramer-Lewanski, 2021).

Conjecture 3.4.3 is true for ℙ1;

The idea of the proof is to use virtual localization to obtain something like

and then transform this into the ELSV formula. We encode the spin Gromov-Witten theory of equivariant spin (ℙ1,𝒪⁡(−1)) by the generating series

and this can be expressed in terms of

Remark 3.4.5.

Exploiting the expression of Hurwitz numbers as vacuum expectations of uncharged fermions, one can write an explicit algorithm to compute the Gromov-Witten invariants of (ℙ1,𝒪⁡(−1)). In particular, applied to d=1 and d=2 obtains Conjecture 3.4.6.

Conjecture 3.4.6 (Maulik-Pandharipande).

Define 𝒰d⁢(z1,…,zn) by

Then in low degree,

Because of corrections in the completed cycles, there are corrections in the expressions for 𝒰d for d≥3. For example,

We may also consider the case of n=1 but arbitrary degree. Then

But now positive energy operators annihilate the vacuum, all operators in the vacuum expectation are cases of ℰmspin⁢(z), and the ℰmspin are closed under commutation, so we commute operators to the right until we reach

In the first step of the recursion, we obtain

The 𝒰d and 𝒱d are enough to compute everything, and in the end, we obtain

where

For example, we can write

3.4.2 Integrability

In the non-spin case, the Hurwitz number hg,μ,ν∘,r obey the 2D Toda hierarchy by a theorem of Okounkov, and then we can use the Gromov-Witten/Hurwitz correspondence of Okounkov-Pandharipande to obtain that the Gromov-Witten theory of ℙ1 satisfies the 2D Toda hierarchy also. This has a ℤ-grading, so in fact it is harder than the spin case. In the spin case, the Hurwitz numbers hg,μ,ν∘,r,spin satisfy the 2BKP hierarchy, so the Gromov-Witten theory of (ℙ1,𝒪⁡(−1)) does as well. In fact, in the spin case τ0=τ1. Here, we mean that

is a tau function of the 2BKP hierarchy.

Now there are several missing ingredients for the Gromov-Witten theory of spin curves,

Theorem 3.4.7.

The spin degeneration formula implies the spin stationary Gromov-Witten/Hurwitz correspondence.

The other missing ingredients are:

  • •

    Virasoro constraints (to deal with insertions of 1);

  • •

    How to deal with odd cohomology classes.

Signed count of Riemann surface covers by parity of spin structures Gunninghamhspin=⟨ uncharged fermions ⟩TR on {x⁡(z)=log⁡(z)−z2⁢Ry⁡(z)=zB=12⁢Bcan+12⁢Bspintry (0,1),(0,2)Ωspin Givental Teleman Ω⋅WittenΛ⁡(1)⁢Λ⁢(12)r=2R=1[ℳ¯g,n⁢(ℙ1,d)]vir∩cg−1+d virtual localization Kiem-Liℙ1 Spin GW/H Transfer 2BKP integrability to GW side GW of spin curves missing ingredients GW of surfaces with smooth KX Kiem-Li Lee-Parker
Figure 3.4: Summary of this chapter

4 Teleman’s classification of semisimple cohomological field theories and applications (Dimitri Zvonkine)

Cohomological field theories are a union of two topics:

  1. 1.

    Topological field theories (equivalently, fusion algebras or commutative Frobenius algebras);

  2. 2.

    The moduli space ℳ¯g,n of curves.

4.1 Topological field theories

Definition 4.1.1.

Let V be a finite-dimensional ℂ-vector space. A topological field theory on V is a series of maps

satisfying the following axioms:

  1. 0.

    For any g,n,m, ωg,n,m is Sn×Sm-invariant;

  2. 1.

    ω0,1,1=id:V→V;

  3. 2.

    The ωg,n,m satisfy the following gluing axioms:

    1. (a)

      For any g,n,m,

    2. (b)

      For any g1,g2,n1,n2,m1,m2, we have

If we choose a basis eμ of V, we can write a tensor (ωg,n,m)μ1,…,μnν1,…,νm. As an application of the axioms, we can see that for any g1,g2,

Remark 4.1.2.

This definition is a result of a dialogue between physicists and mathematicians and encodes (in a nice encapsulated way for mathematicians) the data of a diffeomorphism-invariant quantum field theory.

Examples 4.1.3.
  1. (a)

    Let V=ℂ. Then all ωg,n,m are simply numbers. For example, we can consider c≠0 and set ωg,n,m=c2−2⁢g−n−m. In the case where dimV=1, this is the only example.

  2. (b)

    Now suppose r≥1 and set V=ℂ⁢⟨e0,…,er−1⟩. We can then solve the combinatorial problem of labelling half-edges graphs with remainders modulo r such that each edge or vertex sums to 0, and define

    Now consider the 2-form η≔ω0,2,0:V⊗2→ℂ. Then in this example ηa⁢b=δa+bmodr. We may also define η∨≔ω0,0,2.

    Proposition 4.1.4.

    η and η∨ are inverses of each other. Here, we consider η:V→V∗ and η∨:V∗→V.

    Proof.

    Simply apply the gluing axiom and obtain ω0,1,1, which is the identity by definition. □

    With the data of η and η−1, inputs and outputs are symmetric, so the data of a topological field theory is determined by ωg,n,0.

  3. (c)

    Let G be a finite group and set V=Z⁢ℂ⁢[G]. We know that V is spanned by conjugacy classes. Now define

    We can define

    For example, if G=Sd is the symmetric group, then V is spanned by partitions of d, and ωg,n,0 is a Hurwitz number.

    Stepping back, note that a finite-dimensional commutative ℂ-algebra is either ℂ⊕⋯⊕ℂ or has at least one nilpotent element. The Burnside formula allows us to move to the idempotent basis from the usual basis

Definition 4.1.5.

A commutative Frobenius algebra is a commutative associative algebra with unit (V,⋅,𝟙) with a counit δ:V→ℂ such that η⁡(u,v)≔δ⁡(u⋅v) is nondegenerate.

Example 4.1.6.

The ring ℂ⁡[x,y]/(x2,x⁢y,y2) cannot be promoted to a Frobemius algebra.

Example 4.1.7.

The counit δ=1|G|⁢[𝟏G]⁢(−) turns V=Z⁢ℂ⁢[G] into a Frobenius algebra.

Theorem 4.1.8.

Topological field theories are equivalent to Frobenius algebras.

Proof.

Let’s begin with a topological field theory. Given V, we can construct

  • •

    The unit as ω0,0,1;

  • •

    The counit as ω0,1,0;

  • •

    The product as ω0,2,1.

We can see that the product is associative by cutting ω0,3,1 in two different ways and commutative by S2-equivariance of ω0,2,1. Clearly, η=ω0,2,0 is nondegenerate from the axioms of a TFT.

In the other direction, we can define ωg,n,m by cutting any surface into pairs of pants. We already know that pairs of pants are simply the product in the Frobenius algebra, so we get the answer. To check consistency, we can modify our decomposition and repeatedly use associativity of the Frobenius algebra. □

Now suppose we actually want to compute ωg,n,1⁢(v1,…,vn) from the Frobenius algebra structure. First, we set α=ω1,0,1∈V, and then the gluing axiom tells us that

Then we may compute

4.2 Moduli of curves

Recall that there is a moduli space ℳg,n parameterizing smooth and proper curves of arithmetic genus g over ℂ with n distinct (numbered) marked points whenever 2−2⁢g−n<0. This has a compactification ℳ¯g,n of stable curves. Recall that a stable curve has:

  • •

    The only singularities allowed are nodes;

  • •

    The marked points are all smooth;

  • •

    The condition 2−2⁢g−n<0 applies to every component if we treat all nodes and marked points as special.

Recall that ℳ¯g,n is a smooth proper Deligne-Mumford stack of dimension 3⁢g−3+n.

Example 4.2.1.

Consider the space ℳ0,4. Classically, there is an isomorphism

where

is the cross-ratio, and thus ℳ0,4=ℙ1∖{0,1,∞}. Compactifying, we obtain ℳ¯0,4=ℙ1, where the points 0,1,∞ correspond to the t→0,t→1,t→∞ limits and are geometrically represented by curves with two components. The markings on the components are (14)⁢(23),(13)⁢(24),(12)⁢(34) in the three cases.

We will conclude this section by giving a picture of ℳ¯g,n, which can be seen in Figure 4.1. A general point corresponds to a smooth curve, and then there are various boundary divisors corresponding to curves with nodes. There is a boundary divisor corresponding curves with non-separating nodes, and these have self-intersections which correspond to having more nodes, where the codimension is given by counting the number of nodes. The boundary is a normal crossings divisor. There are other boundary divisors which correspond to separating nodes, and they have a combinatorial description based on how the genus and marked points are split among the components.

××××××××××××××××××
Figure 4.1: Picture of ℳ¯2,3. The divisor with self-intersection is the image of q:ℳ¯1,5→ℳ¯2,3 while the other divisor is the image of s:ℳ¯1,2×ℳ¯1,3→ℳ¯2,3.

It is important to note that the components of the boundary are given by products of smaller moduli spaces. In the boundary components with reducible curves, we can form the gluing map

In the components with a non-separating node, there is of course the map

which is 2-to-1 onto the image (since we forgot the last two markings, we can swap the two and obtain the same glued curve).

The last interesting map we should discuss are the forgetful maps

which are defined by deleting a marked point and then stabilizing the resulting curve by collapsing rational tails (one node and either zero or one marked points) and bridges (two nodes and no marked points).

4.3 Cohomological field theories

Definition 4.3.1.

Let V be a finite-dimensional ℂ-vector space. A cohomological field theory on V is the data of a unit vector 0≠𝟙∈V, a nondegenerate symmetric bilinear form η, and a collection

satisfying the following axioms:

  1. 0.

    Each Ωg,n is Sn-equivariant, where Sn acts on ℳ¯g,n by permuting the marked points (here, we will only consider the even part of the cohomology of ℳ¯g,n);

  2. 1.

    For all u,v∈V, Ω0,3⁢(u,v,𝟙)=η⁡(u,v);

  3. 2.

    We have the following gluing axioms:

    1. (a)

      For any g,n, we have

    2. (b)

      Choose a basis (eμ) for V. Then we have

  4. 3.

    (Optional, flat unit, string equation) For any g,n, we have

We will now give several examples of cohomological field theories.1

Definition 4.3.2.

Recall that if 𝒞→𝜋ℳ¯g,n is the universal curve, then there is a sheaf

on ℳ¯g,n. Above any stable curve C, the fiber is simply H0⁢(ωC). Recall that because nodes are Gorenstein, ωC is a line bundle. In the more analytic setting, the sections are abelian differentials, which are simply meromorphic 1-forms on each component with poles allowed only at the nodes and opposite residues on the branches. By Serre duality, connectedness, and flatness, 𝔼 is a vector bundle of rank g, called the Hodge bundle.

Theorem 4.3.3.

If we set V=ℂ, 𝟙=1, η=(1), and

we obtain a cohomological field theory.

Proof.

The first two axioms are clearly satisfied, so we check the gluing axioms. We will first show that

Because poles are not allowed at the node (by the residue formula), we have q∗⁢𝔼g=𝔼g1⊕𝔼g2. By the splitting axiom for the total Chern class, we are done. We now turn to showing that

There is an exact sequence

given by α↦Resn+1⁡α, so by the trivial bundle axiom for the total Chern class, we are done. In fact, the string equation is also satisfied because ℰ does not depend on the marked points. □

Remark 4.3.4.

For any CohFT, if we restrict Ωg,n to the H0⁢(ℳ¯g,n)=ℂ, we obtain a topological field theory.

Remark 4.3.5.

For any ai, we may consider any characteristic class of the form

and obtain a CohFT.

Definition 4.3.6.

Let (C,x1,…,xn) be a smooth curve with marked points and KC be the cotangent bundle. Then an r-spin structure on C is a line bundle ℒ with an isomorphism ℒ⊗r≃KC⁢(−∑ai⁢xi), where a1,…,an∈{0,…,r−1}. Now note that if we fix the ai, then there are r2⁢g possible choices of ℒ corresponding to the possible choices of r-torsion in Pic⁡(C).

We now obtain a moduli space ℳg;a1,…,an1/r with an étale map

We can compactify this to ℳ¯g;a1,…,an1/r, but now the map to ℳ¯g,n is ramified at the boundary. Then we may define Witten’s class on ℳ¯g;a1,…,an1/r and push it forward to ℳ¯g,n.

We may now set V=ℂ⁢⟨e0,…,er−2⟩, write 𝟙=e0, and set ηa⁢b=δa+b,r−2 to be the antidiagonal form. Then we will set Ωg,n⁢(ea1,…,ean) to be Witten’s class. This is a cohomological field theory.

Definition 4.3.7.

Suppose that h0⁢(C,ℒ)=0 (this only happens in genus 0). Then if π:𝒞→ℳ¯g;a1,…,ar1/r is the universal curve, R1⁢π∗⁢ℒ is a vector bundle. Then we may define the Witten class

In the complicated case when h0 doesn’t vanish, the Witten class may be defined using cosection localization. In our situation, let X be the space of sections of ℒ. Then if s∈Γ⁡(C,ℒ), we have

which gives us sr−1:V→𝒪.

Remark 4.3.8.

Suppose σ⁢L→𝒪 is a (meromorphic) cosection. Then we can consider c1⁢(L)=P−Z, so in the case of an obstruction bundle V and a cosection L→𝒪, we may consider [X]∩c1⁢(L)⋅e⁡(V). For example, if we consider

over ℂx,y2, σ is surjective except over (x,y)=(0,0). We can then blow up to obtain the situation of V⊕L.

Example 4.3.9.

These Witten classes can be hard to compute, but there are some relatively simple examples. First, there is

and a second relatively accessible example is

For the final example, let 𝔤 be a simple Lie algebra, for example 𝔰⁢𝔩N. Then let G be the compact real form of the Lie group associated to 𝔤, for instance S⁢UN. Then let T⊆G be the maximal torus with Lie algebra 𝔱. Then we may consider positive weights, which in the case of 𝔰⁢𝔩N are parameterized by Young diagrams of length at most N−1. We will define the level ℓ to be all partitions with width at most ℓ. We will then set the data

where μ∗ is the complement of μ in an N×N square (corresponding to the adjoint representation).

Example 4.3.10.

In the simplest example, irreducible representations of S⁢U2 are given by integers (or really weights on the half-line). They are really just Symn⁡ℂ2, where ℂ2 is the defining representation.

We will now construct the Verlinde bundle 𝕍 on ℳ¯g,n. We will do the construction for U⁡(1)=S1 (although this will break down) and for a simple group G.

Table 4.1: Construction of Verlinde bundle
U⁡(1)=S1 Simple group G
Curve C Curve (C,x1,…,xn)
Jac⁡(C) parameterizing flat connections on C×U⁡(1) Moduli space ℳ of flat connections on C×G with monodromies exp⁡(μi/ℓ) at xi.
ω=[Θ] Canonical symplectic form ω
Line bundle ℒ=𝒪⁡(ℓ⁢Θ) on Jac⁡(C) Line bundle ℒ on ℳ such that c1⁢(ℒ)=ℓ⁢ω
Vector bundle H0⁢(Jac⁡(C),ℓ) of θ-functions of weight ℓ Vector bundle 𝕍C=H0⁢(ℳ,ℒ)

We have defined a vector bundle 𝕍 on ℳg,n, which by a theorem of Faltings extends to ℳ¯g,n as a vector bundle. Then there are the formulae

These imply that taking ch⁡(𝕍g,n) is a cohomological field theory.

Remark 4.3.11.

It is in extending to ℳ¯g,n where the construction for U⁡(1) breaks down because line bundles have degrees, and when we degenerate a bundle of degree 0, the two components could have nonzero degrees.

Remark 4.3.12.

The topological part of this cohomological field theory is simply ch0⁡(𝕍)=rk⁡(𝕍) and is given by the Verlinde formula (using the idempotents of the Frobenius algebra).

4.4 Givental group action

4.4.1 Stable graphs

Recall our picture Figure 4.1 of ℳ¯g,n. If π:𝒞→ℳ¯g,n is the univeral curve and si:ℳ¯g,n→𝒞 is the section corresponding to the i-th marked point, we may construct the line bundle

corresponding to the i-th cotangent line. Then we define the psi-classes

The boundary strata correspond to stable graphs which have vertices corresponding to components, edges corresponding to nodes, and half edges (legs) corresponding to marked points. Each vertex is labelled with a nonnegative integer corresponding to a genus, and each leg is labelled with a number. An example is displayed in Figure 4.2.

230αβ3γ1ψ2ψ22ψψ3ψψ2ψ
Figure 4.2: Example of a stable graph in ℳ¯7,3 and associated tautological class. Red half edges are forgotten from larger moduli spaces. This stable graph describes the image of a map ℳ¯2,4+1×ℳ¯3,2+2×ℳ¯0,5→ℳ¯2,4×ℳ¯3,2×ℳ¯0,5→ℳ¯7,3.

For the forgetful morphism p:ℳ¯g,n+1→ℳ¯g,n, we may define the kappa-class

Then the tautological classes are those which can be obtained from 1 using the maps p,q,s.

4.4.2 Group action on cohomological field theories

Define the formal variable

where ti∈V. Then we define

Composing these translations corresponds to addition of power series. If we consider the action Ta+Tb, we see that adding ma copies of Ta and mb copies of Tb gives us a leading coefficient of

We may check that T preserves the gluing axiom, but nodes are always glued using η−1, so we are fine. The only thing that is not preserved is the string equation (or flat unit).

We will also consider operators

These are required to satisfy the symplectic condition

where the adjoint is taken with respect to η. Now define the cohomological field theory (R⁢Ω)g,n⁢(v1,…,vn) as a sum over stable graphs Γ of genus g with n legs of Ω modified by the following procedure in Figure 4.3:

  • •

    At every vertex, place a copy of Ω;

  • •

    For any leg corresponding to the ith marked points, insert R−1⁢(ψi)⁢vi;

  • •

    At any edge, insert a copy of

    where ψ′,ψ′′ are the ψ-classes at the two components of the node.

  • •

    Divide by |Aut⁡Γ|.

230ΔΔΔΔR−1⁢(ψ3)⁢v3R−1⁢(ψ3)⁢v3R−1⁢(ψ3)⁢v3
Figure 4.3: Contribution of a stable graph to R⁢Ω

We need to prove that:

  1. 1.

    This gives a group action: Ra⁢(Rb⁢Ω)=(Ra⁢Rb)⁢Ω;

  2. 2.

    If Tb=R⁢Ta, then R⁡(Ta⁢Ω)=Tb⁢(R⁢Ω);

  3. 3.

    This R⁢Ω is a cohomological field theory.

This is complicated, so we will restrict ourselves to computing q∗⁢(R⁢Ω). Recall that q corresponds to creating a new edge, so if Γ does not have an edge separating the curve into curves which can be glued by q, we create a new edge and insert η−1. If the edge is already present, then the normal line bundle to the boundary divisor corresponding to q has c1=−(ψ′+ψ′′), so we replace Δ with

We therefore sum over all contributions to obtain

This is precisely what is needed to satisfy the gluing axiom.

4.5 Teleman’s classification

We are now ready to formulate Teleman’s theorem.

Theorem 4.5.1 (Teleman).

Let Ω be a cohomological field theory, ω be its topological part, and assume that the Frobenius algebra corresponding to ω is semisimple. Then there exists a unique T∈ψ2⁢V⁢[[ψ]],R∈End⁡(V)⁢[[ψ]] such that

Moreover, Ω satisfies the string equation if and only if

which is equivalent to the condition that R⁢T preserves the vector −𝟙⁢ψ.

Remark 4.5.2.

The condition of preserving −𝟙⁢ψ is the same as preserving the vertex of Givental’s Lagrangian cone, which is the dilaton shift −𝟙⁢ψ.

We are now ready to return to our examples.

  1. (a)

    In the case of the Hodge bundle 𝔼, the R-matrix is given by

    where Bm+1 is the Bernoulli number. This is computed using Mumford’s formula, which computes the Chern characters of 𝔼. For example,

    where δ is the sum of all boundary divisors.

  2. (b)

    We will now consider Witten’s 3-spin class. In this case

    The product is given by

    and therefore the only nonzero element is W0,3⁢(0,0,1)=1. Therefore, the product is given by:

    Because e1 is nilpotent (in fact the algebra is ℂ⁡[x]/x2), we are not semisimple. The solution to this is to shift the Witten class by 3⁢e1 to obtain

    This brings us to a semisimple point in the Frobenius manifold, where we also have

    and thus the product is modified to have

    It is easy to see that the algebra is ℂ⁡[x]/(x2−1) and the idempotents are

    The shifted Witten class is no longer of pure degree, but the correction terms have lower degrees of the same parity.

    We may now apply the classification theorem. From the product and the quadratic form, the topological part is

    We can now write the R-matrix. First, define the power series (asymptotic expansions of the Airy function)

    Now set

    If we extract the part of degree g−1+∑ai3, we obtain a formula for Witten’s class, and if we consider the higher degree parts, we obtain tautological relations on the moduli space of curves. The simplest tautological relation is the WDVV equation on ℳ¯0,4. A new relation, due to Getzler, lives on ℳ¯1,4. If we consider the simpler case of ℳg, then there are no relations between the κi up to degree g3. There is a conjecture that all tautological relations come from Witten’s class, and for example, if we consider the degree 2 part of the g=1,n=4,a1=⋯⁢a4 case, we obtain Getzler’s relation.

  3. (c)

    We now consider the Verlinde bundle for 𝔰⁢𝔩2⁢(ℂ) at level ℓ=1. There are only two representations e∅=𝟙,e□, and η is diagomal. Then the Verlinde bundle is a pullback except in the case

    where Γ runs over stable graphs with vertices of even degree. The contribution of Γ is given by inserting e−ψi4 at marked points and inserting

    at the edges.

4.6 Relation to topological recursion

Consider a disk near xi with simple ramification, so we have the coordinate zi=x−xi. Recall from Definition 3.2.2 that topological recursion requires the specification of the data yi⁢(z) for all i, the Bergman kernel B⁡(zi,zj)⁢dzi⁢dzj for all i,j, and the series ξi⁢(x).

Applying the Laplace transform, we obtain T⁡(ψ) and Δ⁡(ψ′,ψ′′) in place of y and B (which yield Ω), where we drop even powers of z and send

Then topological recursion produces forms given by