Patrick Lei

All notes

Quantum D-modules of projective bundles and blowups

SCMS, Oct 2026.

Abstract.

In this lecture series, we review the basics of quantum D-modules and Givental’s formalism. We then explain the projective bundle formula and the blowup formula, which are proved using Fourier transforms associated with rank-one torus actions. If time permits, we will also discuss a nonabelian version of the Fourier transform.

1 GW theory

One motivation for us comes from mirror symmetry. Let X be a Fano variety (or more generally, log Calabi-Yau). Then we expect there to be a mirror Xˇ, which is an LG model. Via some version of Saito theory (which may not actually be well-defined), we obtain a Gauss-Manin conjecture. This will correspond to the quantum D-module of X, which is defined using Gromov-Witten theory.

Another motivation comes from the study of other areas of algebraic geometry1. Katzarkov-Kontsevich-Pantev-Yu define Hodge atoms which establish an irrationality criterion for varieties and was used to prove the irrationality of the cubic fourfold. Recently, an unreleased internal model at OpenAI proved the generalized Mukai conjecture using Gromov-Witten theory.

1.1 Quantum D-modules.

Let X be a smooth projective variety and (ϕi)i be a homogeneous basis with corresponding coordinates (τi)i. Then we define the (big) quantum cohomology

where ℂ⁡[Q]≔ℂ⁡[NEℕ⁢(X)] keeps track of effective curve classes, and here we are taking a graded completion of ℂ⁡[Q,τ] where

Example 1.1.

Let X=ℙn and τ=0. Then the curve classes are just non-negative integers, so Q has degree 2⁢(n+1). In this case, ℂ⁢⟦Q⟧=ℂ⁡[Q].

Example 1.2.

Let X have ℂ⁡[Q]=ℂ⁡[Q1,Q2] where deg⁡Q1=2 and deg⁡Q2=−2. In this case, we actually have

which lives in all even degrees.

Now let α,β,γ∈H∗⁢(X). The quantum product is defined by the formula

where (−,−)X is the Poincaré pairing and τ=∑iτi⁢ϕi.

Remark 1.3.

We can define the quantum cohomology as a formal Frobenius manifold, but this is equivalent to the approach taken by the speaker, so we will skip it.

Definition 1.4.

We will now define the quantum D-module

The connection ∇X is known as the quantum connection and is defined by

Here, the first definition makes sense only for ξ∈H2⁢(X) and the Euler vector field is defined by

Additionally, the grading operator is defined by μX⁢(ϕi)=deg⁡ϕi−dimX2⁢ϕi. Finally, the pairing

is given by the formula

Example 1.5.

Let X=ℙ1. Then we have

and

with the relevant connections.

Remarks 1.6.
  1. (1)

    The object QDM⁡(X) is not a D-module in the usual sense of D-modules.2

  2. (2)

    The objects QH and QDM contain the same information, namely that one can be recovered from the other.

  3. (3)

    The ring QH∗⁡(X) is a quantization of the usual cohomology ring H∗⁢(X) as a ring and QDM⁡(X) is a quantization of H∗⁢(X) as a graded H∗⁢(X)-module.

We will now produce the fundamental solution MX⁢(τ) to the quantum connection. This is given by the formula

for all α,β∈H∗⁢(X).

Proposition 1.7.

We have

Moreover, MX⁢(τ)⁢1=JX⁢(τ) and MX⁢(τ)⁢ϕi=z⁢∂τiJX⁢(τ). We also have the asymptotics

and the property

1.2 Givental formalism.

Define

and the genus zero descendent GW potential

Now define the symplectic vector space

This is actually functorial, but we will skip the discussion of this. Then define the form

by the formula

This makes ℋX an infinite-dimensional symplectic vector space over ℂ⁢⟦Q⟧.

There is a decomposition

where

Finally, we can define ℒX to be the graph of dℱX and refer to it as the Givental cone.

Theorem 1.8.

ℒX⊂[H]X is an overruled Lagrangian cone; namely:

  1. (1)

    ℒX⊂(ℋX,ΩX) is Lagrangian;

  2. (2)

    It is a cone, namely that if I⁡(τ)∈ℒX and F⁡(τ,z)⁢I⁡(τ)∈ℒX if F⁡(τ,z)=1+O⁡(τ);

  3. (3)

    For any f∈ℒX, we have Tf⁢ℒX∩ℒX=z⁢Tf⁢ℒX.

ℋ+ℋ−ℒXTf⁢ℒXz⁢Tf⁢ℒXf
Figure 1. Schematic of the Givental cone
Example 1.9.

When X=ℙ1, we see that the basis 1,p,z,p⁢z,z2,… is dual to the basis z−1⁢p,z−1,z−2⁢p,z−2,z−3⁢p,…. When we multiply by z, we see that ℋ− is not preserved.

Remark 1.10.

Let F be a formal function on ℋ+. Then the graph ℒ of dF is an overruled Lagrangian cone if and only if F satisfies the string equation, dilaton equation, and topological recursion relations.

There is an informal description of the Lagrangian cone as

In particular, we see that

is an isomorphism intertwining ∇ and d. This essentially explains all of genus zero Gromov-Witten theory.

Table 1. Comparison
QDM⁡(X) Tτ⁢ℒX
differential operator hard easy
description as ℂ[z]⟦Q,τ⟧-module easy hard

For example, if we consider

we prove that the bottom arrow is an isomorphism, then go around the diagram to obtain the top arrow.

2 Equivariant GW theory

Now let X be a smooth semiprojective (projective over affine) T-variety, where T=(ℂ×)r is a torus. We can run the same formalism as before to obtain

This still intertwines ∇ and d, but there are additional operators, called shift operators, which we will denote by 𝕊 and 𝒮.

In this setting, we have the genus-zero potential

and therefore ⋆τ, ∇z⁢∂τi, ∇z⁢ξ⁢Q⁢∂Q, and ∇z2⁢∂z. Here, we will denote by τ=(τi)i∈I coordinates corresponding to a homogeneous basis ϕi)i∈I of H∗⁢(X) and τ=(τi,ℓ→)i∈I,ℓ→∈ℤ≥0r coordinates corresponding to a basis (ϕi⁢λℓ→)i,ℓ→ of HT∗⁢(X)≅H∗⁢(X)⊗HT∗⁢(pt) (here this is a non-canonical isomorphism of HT∗⁢(pt)-modules due to Goresky-Kottwitz-MacPherson).

Note here that we need to make the modification

and define

and the operator

We also can define

and the Givental cone ℒX,T⊂ℋX,T≔HT∗⁢(X)⁢((z−1))∩HT⁢t⁢i⁢m⁢e⁢s⁢ℂz×∗⁢(X)loc. Note here that the localization occurs in the category of graded rings, so expressions of the form ⟨ϕiz+ψ,…⟩ live in HT×ℂ×∗⁢(pt)loc.

2.1 Shift operators.

Example 2.1.

Let X=ℂn carry the scaling action of ℂ×. Note that τ0 has degree 2, so

and

Since the classical and quantum products coincide, ∇ and P are easy to define. The operator 𝕊 is an endomorphism of ℂ⁡[z,λ]⁢⟦τ⟧ given by the formula

Note that Mℂn,ℂ×⁢(τ)=exp⁡(τ/z) gives an isomorphism QDMℂ×⁡(ℂn)≃Tτ⁢ℒℂn,ℂ×, and 𝕊=M−1∘𝒮∘M. In this setting, 𝒮 is given by

so the terms containing τ come from M.

Remark 2.2.

From the example we can see that QH and ∇z⁢∂τ do not depend on n, but 𝕊 does.

Remark 2.3.

Later, we will use shift operators to lift Kirwan surjectivity

to obtain an isomorphism

More generally, we expect that the “completion”

is a quantization of the usual Kirwan map κL and that QDMT⁡(X) is a Kähler moduli space. with various limit points being GIT quotients X⫽LT or X⫽L′T.

⋯X⫽LTX⫽L′TQDMT⁡(X)LQDMT⁡(X)L′Kähler moduliQDMT⁡(X)
Figure 2. Equivariant Kähler moduli with GIT quotients as limit points

In the example, if we consider small quantum cohomology, we see that

Dependence on the dimension can be recovered by considering the structure as a ℂ⁡[Q]-module.

Remark 2.4.

In symplectic topology, we can study quantum cohomology using the loop space

From the torus action on X, we obtain a torus action on L⁢X given by

There is also an action of π1⁢(T) on L⁢X, where if g:S1→T is a group homomorphism, then

This action of π1⁢(T) is where the shift operators come from.

We will now actually introduce shift operators, which defines a ℂ⁡⟦NEℕT⁡(X)⟧-module structure on QDMT⁡(X), where NEℕT⁡(X) is an equivariant analogue of the monoid NEℕ⁡(X) of effective curve classes.

Definition 2.5.

Define

to be the subgroup generated by vertical curve classes with respect to the fibration

where XT≔(X×E⁢T)/T and by sx⁣∗⁢σ for any x∈XT and σ∈H2⁢(B⁢T,ℤ).

Lemma 2.6.

The complex

is exact.

Sketch of proof.

By definition, we see that N1T↠H2T⁢(pt,ℤ). Then we consider the diagram

where the vertical arrows are injective by definition and the bottom arrow is injective by the spectral sequence for X→XT→B⁢T. This implies the top arrow is injective, so it remains to show that the relevant image and kernel coincide. It is enough to show that for all x,x′∈XT lying on different fixed components and σ∈H2⁢(B⁢T,ℤ), then sx⁣∗⁢σ−sx′⁢σ∈N1⁢(X). It is clear that they are linked by a chain of T-fixed rational curves. Note that if x,x′ lie on the same fixed component, then they define the same homology class.

Now let σ:ℙ1→B⁢T. Then between sx⁣∗⁢σ and sx′⁣∗⁢σ is a chain of Hirzebruch surfaces. For simplicity, suppose there are two surfaces 𝔽k and 𝔽k′, both attaching to sx′′⁣∗⁢σ. If P and P′ denote fiber classes, we see that

□

Now let β∈N1T⁢(X) and set k=−β¯∈H2T⁢(pt,ℤ). Define

where the action is defined by (k11). This can be viewed as some form of a “twisted” graph space and is the fiber product

There is an action of T^≔T×ℂ× on Ek given by the formula

Denote by X0 and X∞ be the fibers at 0=[1,0] and ∞=[0,1]. Then the action of T^ on X0 is given by

whereas the action on X∞ is given by

The identity map X0≃X∞ induces a nontrivial map

Here, we have to identify T^ with itself by (λ,z)↦(λ⁢zk,z).

Now for any

there exists some τ^∈HT∗⁢(Ek) such that τ^|X0=τ and τ^|X∞=Φk⁢(τ). Moreover, there exists a unique connected component F0⁢(k)⊆Xk⁡(ℂ×) whose normal bundle has only negative weights (using Morse theory for the real moment map, this is the top connected component of the fixed locus, which is the critical locus). Let σ0⁢(k)∈NEℕ⁡(Ek) be the section class associated to F0⁢(k).

Definition 2.7.

For β∈N1T⁢(X), define the shift operator

as follows: for any α0∈HT^∗⁢(X0) and α∞∈HT^∗⁢(X∞), we have

Note that Φ−β¯∘𝕊β is HT^∗⁢(pt)-linear, but because Φ−β¯ is not linear, neither is 𝕊β, and in fact we have

Equivalently, for all λ∈HT∗⁢(pt), we have the commutator identity

Remark 2.8.

The map τ↦τ^ is not HT^∗⁢(pt) is not HT∗⁢(pt)-linear, so we need to work over ℂ⟦Q,τ⟧.

Definition 2.9.

Define

Proposition 2.10.

The shift operators give QDMT⁡(X) the structure of a ℂ⁡⟦NEℕT⁢(X)⟧-module.

2.2 Shift operators for Givental cones.

Define

where the additional factor of ℂ× acts trivially. By the localization formula, we know

Definition 2.11.

For β∈N1T⁢(X), define

as follows: for any fixed component F⊂XT and f∈ℋXrat, we have

where we decompose NF=⨁α∈Hom⁡(T,ℂ×)NF,α and ρF,α are the Chern roots of NF,α.

Proposition 2.12.
  1. (1)

    We have 𝒮β1∘𝒮β2=𝒮β1+β2;

  2. (2)

    𝒮β commutes with commutators via

  3. (3)

    The shift operators intertwine the fundamental solution:

Corollary 2.13.
  1. (1)

    We have 𝕊β1∘𝕊β2=𝕊β1+β2;

  2. (2)

    𝕊β commutes with commutators via

3 Fourier transforms

There are two kinds of Fourier transforms, which were introduced by Iritani: continuous and discrete. Let W be a smooth semi-projective T-variety, where T=ℂ×.

3.1 Discrete Fourier transform.

This is a quantization of the Kirwan map. Let L be a T-linearized ample line bundle on W such that ∅≠Ws=Wss is acted on freely by T and W⫽LT is smooth and projective. Then there is the Kirwan map

Conjecturally, this should quantize to

It will intertwine the shift operator 𝕊 with the Novikov variable q and the equivariant parameter with the quantum connection ∇z⁢κ⁢(λ)⁢q⁢∂q.

Example 3.1.

Let W=ℂn with the diagonal action. Then the Kirwan map is

At the level of small quantum D-modules, we have

but we have a nontrivial shift operator

In the quotient, we have

where SQH∗⁡(ℙn−1)=ℂ⁡[p,q]/(pn−q). Note here that SQDM⁡(ℙn−1)=⨁i=0n−1ℂ⁡[z,q]⋅pi whereas SQDMℂ×⁡(ℂn)=⨁i=0n−1ℂ⁡[z,𝕊]⋅λi. Note here that

We may now define the map

More generally, we have

In this case, FT is an isomorphism. It is clear that it intertwines 𝕊 and q. It is a bit less obvious that it intertwines λ and ∇z⁢q⁢∂q, but this is also not hard to check.

Note that QDMT(W)=HT∗(W)[q]⟦𝒬,𝕌⟧ has the action of 𝕊β for all β∈NEℕT⁡(W), whereas QDMT(W)loc=HT∗(W)loc[z]⟦𝒬,𝕌⟧ has the action of β∈N1T⁢(W).

Definition 3.2.

Let cL⊂cT⁢(W)⊆NT1⁢(W)ℝ be the GIT chamber. Then we define

In progress work of Iritani-Sanda will show that cT⁢(W)∨=NEℕT⁡(W).

Now we can define QDMT⁡(W)L to be the graded completion of

Here, note that deg⁡𝕊β≔2⁢(c1T⁢(W),β).

Conjecture 3.3 (Reduction conjecture, Iritani, Iritani-Sanda).

There exists τ(𝕌)∈H∗(W⫽LT)⟦cL,ℕ∨,𝕌⟧, or equivalently, τ:HT∗⁢(W)→H∗⁢(W⫽LT) over ℂ⟦cL,ℕ∨,𝕌⟧, such that there exists FTL and (⋅)L∧ fitting into the diagram

such that λ on the left side is intertwined with ∇z⁢κL⁢(λ)⁢q⁢∂q and z⁢κL⁢(λ)⁢q⁢∂q+κL⁢(λ)∪− on the right side, and 𝕊β on the left side is intertwined with qβ on the right side. Moreover, the bottom arrow is given by

Example 3.4.

Let W=ℂn with diagonal torus action. Then Jℂn,ℂ×⁢(τ=0)=1∈T0⁢ℒℂn,ℂ×. Its Fourier trasnform is given by

Note that 𝒮:f⁡(λ,z)↦λn⁢f⁢(λ−z,z), so 𝒮−1:f⁡(λ,z)↦f⁡(λ+z,z)(λ+z)n, and therefore

Applying again, we see that

On the other hand, we see that 𝒮⁢1=λn is killed by the Kirwan map, so in fact we see that

To recover the big I-function, just replace 1 by e𝕌/z.

Theorem 3.5.

The reduction conjecture is true for:

  • •

    Projective bundles ℙ⁡(V)=V⫽ℂ×, due to Iritani-Koto;

  • •

    (Possibly non-split) toric bundles ⨁iVi⫽T, due to Koto;

  • •

    When T=ℂ× and there exists a ℂ×-fixed divisor such that T acts on ND/W with weight ±1 and W⫽Lℂ×=D, due to Iritani. This plays a crucial role in the blowup formula.

3.2 Continuous Fourier transform.

This is a quantization of equivariant localization. Let WT⊇F be a connected component of the fixed locus. Then there exists

where cF∈ℤ≥0 is determined the action of T on NF/W. Here, it is defined as a phase asymptotic approximation of the integral

Informally, define

Now consider the map

This intertwines 𝕊β with Qβ+σ0⁢(−β¯)⋅e−z⁢β¯⁢∂λ. Applying the classical idea of the Mellin transform, we can consider the integral

and turn the shift operator into a variable Sβ. Unfortunately, this integral is ill-defined because the integrand contains log⁡z and log⁡sF.

Instead, we will consider a stationary phase approximation. Here, SF≔SσF⁢(1), where σF:B⁢G→XT is a section. Then, we can rewrite the integral as

where

Here, if we write

then rα≔rk⁡(Nα) and wα≔αλ. Finally, we define

Setting cF≔|∑αrα⁢wα|∈ℤ≥0 and λ0≔SF1/cF (up to scaling by a rational number), then we have the stationary phase approximation

Note here that η′⁢(λ0)=0.

Definition 3.6.

We define

Here, we write η⁡(λ0+λ~)=η⁡(λ0)−cF2⁢λ⁢λ~2+η≥3⁢(λ~), where η≥3⁢(λ~)∈O⁡(λ~3). For all 0≥j<cF, there is ℱF,j⁢(f) corresponding to all of the critical points of η.

Theorem 3.7 (Iritani).

For all 0≤j≤cF−1 there exists σj:HT∗⁢(W)→H∗⁢(F) and a Fourier transform FTF,j fitting into a commutative diagram

which intertwines 𝕊β and 𝒮β on the left with formal variables Sβ on the right and intertwines λ on the left with ∇z⁢λ⁢S⁢∂S+λj and z⁢λ⁢S⁢∂S+λj on the right. Here, the λj are the critical points of η⁡(λ).

Remark 3.8.

Iritani actually introduced

Example 3.9.

Again, consider W=ℂn with the diagonal action of ℂ×. Then set F={0}=Wℂ×. Note here that λ0=S1/n and η⁡(λ)=λ⁢log⁡S−n⁡(λ⁢log⁡λ−λ). Then we see that

where n′=n when n is odd and n′=2⁢n when n is even. Note that this sends

Also, this intertwines λ and 𝕊 with ∇z⁢S⁢∂S=z⁢S⁢∂S+S1/n and S. Moreover, ∇z2⁢∂z is intertwined with ∇z2⁢∂z=z2⁢∂z−n⁢S1/n.

4 Projective bundle formula

Let W≔V be a vector bundle of rank r over a smooth projective base B. Assume that there exists a surjection 𝒪B⊕N↠V∨, which ensures that V is semi-projective. If this is not true, we can replace V by V⊗𝒪B⁢(−M) for some M. Then we see that

for some L and VT=B.

The reduction conjecture predicts the existence of a Fourier transform

There is also a continuous Fourier transform

We will write ℂ⁡⟦Q⟧≔ℂ⁡⟦NEℕ⁡(B)⟧ and note that

where σ∈H∗⁢(B). We note that

Here, β∈NEℕ⁡(ℙ⁡(V)) is associated to q(β,𝒪ℙ⁡(V)⁢(1))⁢Qπ∗⁢β, where π:ℙ⁡(V)→B is the structure morphism.

Theorem 4.1 (Iritani-Koto).

There exists

for i=0,…,r−1 such that there exists

Here, QDM(ℙ(V))loc=QDM(ℙ(V))⊗ℂ[z]((q−1/r′))⟦Q,τ⟧. Moreover, Φ intertwines the quantum connections and is linear over ℂ[z]((q−1/r′))⟦Q,τ⟧.

Sketch of proof.

The first step is to prove the reduction conjecture for ℙ⁡(V)=V⫽ℂ×. The second step is to prove that FTL is an isomorphism. Once we have done this, we define a ℂ[z,q]⟦Q,τ⟧-module homomorphism

as

Unfortunately, this is linear only over ℂ[z,q]⟦Q,𝕌⟧, so we need a lift u:H∗⁢(ℙ⁡(V))→Hℂ×∗⁢(V) (here, u⁡(τ)∈Hℂ×∗⁢(V)⁢⟦τ⟧ and τ⁢(u⁡(τ))=τ) of τ:Hℂ×∗⁢(V)→H∗⁢(ℙ⁡(V)). This is not hard to find, so we obtain a morphism

where σ¯i=σi∘u. Now if we set p=c1⁢(𝒪ℙ⁡(V)⁢(1)) and consider ϕ∈H∗⁢(B), we see that

where λi=ζri⁢q1/r. Once we invert q1/r, we see that Φ is an isomorphism. □

Remark 4.2.

For example, when B is a point, then z𝕀(𝕌)=M{r−1(τ(𝕌))⋅R (here the big I-function is defined using light markings) via Birkhoff factorization. This defines τ⁡(𝕌) as a mirror map.

Corollary 4.3.

As rings, we have an isomorphism

over ℂ⁡((q−1/r))⁢⟦Q⟧. Moreover, the genus-zero Gromov-Witten invariants of ℙ⁡(V) can be recovered from the genus-zero invariants of B and the total Chern class c⁡(V).

Theorem 4.4.

We have

where JV is the equivariant J-function of V.

Sketch of proof.

Consider the short exact sequence

This gives us an inclusion V↪𝒪⊕N, which descends to

which is cut out by a regular section of 𝒬⊠𝒪ℙN−1⁢(1). We can now use quantum Lefschetz to relate ℒℙ⁡(V) and ℒB×ℙN−1,(𝒬⊠𝒪⁡(1),eℂ×). Quantum Riemann-Roch relates this to ℒB×ℙN−1, the work of Brown relates ℒB×ℙN−1 and ℒB, and finally quantum Riemann-Roch relates ℒB and ℒV. □

Remark 4.5.

To prove a mirror theorem for general non-split toric bundles, this strategy fails because 𝒬 is not convex in general. However, we can use virtual localization to reduce to the case of projective bundles.

5 Blowup formula

Let W≔BlZ×0⁡(X×ℙ1).

Theorem 5.1.

There exists a map

and a map

Here, c is the codimension of Z and the maps are defined over ℂ((q−1/(c−2)))⟦Q,τ~⟧.

Corollary 5.2.

We obtain a corresponding decomposition

over ℂ⁡((q−1/(c−2)))⁢⟦Q⟧.

Theorem 5.3.

The genus-zero Gromov-Witten invariants of BlZ⁡X can be recovered from those of X and Z, and from c⁡(NZ/X).

Proof.

Consider the master space W with the action of ℂ×. It has two GIT quotients W⫽Lℂ×=X and W⫽L~=BlZ⁡X. There are three fixed components

so we have three continuous Fourier transforms and two conjectural discrete Fourier transforms

The first step of the proof is to show that FTX contains essentially the same information as FTL and that FTBlZ⁡X contains essentially the same information as FTL~. The next step is to prove that

is an isomorphism. Then we combine the Fourier transforms to construct Ψ and finally show that Ψ is an isomorphism. □

Example 5.4.

Let X=ℙm+n−1⊇Z=ℙn−2. Assume that m≥1 and n≥2 and set X~=BlZ⁡X≅ℙ⁡(𝒪ℙm⊕n+1⊕𝒪ℙm⁢(−1)). Note that

where p=c1⁢(𝒪X⁢(1)) and h=c1⁢(𝒪ℙm⁢(1)). Using the J-function, we can compute the small quantum cohomology

where deg⁡Q=2⁢n and deg⁡q=2⁢m. Doing some algebra, we can obtain an ad hoc blowup formula and deduce

On the other hand, an ad hoc version of the projective bundle formula viewing X~ as an ℙn−1-bundle over B=ℙm gives us

Therefore, we have the following picture of SQH∗⁡(X~):

qQq=0Q≠0Q=0,q≠0q,Q≠0H∗⁢(ℙm)⊕nℂn⁡(m+1)H∗⁢(X~)H∗⁢(X)⊕H∗⁢(Z)⊕m