Patrick Lei

All notes

The Count of Instantons

Lectures by Nikita Nekrasov, Columbia, 2023-24.

The Count of Instantons
2023-24
Notes by Davis Lazowski and Patrick Lei
Lectures by Nikita Nekrasov

Columbia University

Disclaimer

These notes were taken during the lectures using neovim (Patrick) and VSCode (Davis). 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 differ between lectures. If you find any errors, please contact me at plei@math.columbia.edu.

Unfortunately, I cannot guarantee that the notation in these notes is consistent. Please do not judge this too harshly.

Description

Graduate level introduction to modern mathematical physics with the emphasis on the geometry and physics of quantum gauge theory and its connections to string theory. We shall zoom in on a corner of the theory especially suitable for exploring non-perturbative aspects of gauge and string theory: the instanton contributions. Using a combination of methods from algebraic geometry, topology, representation theory and probability theory we shall derive a series of identities obeyed by generating functions of integrals over instanton moduli spaces, and discuss their symplectic, quantum, isomonodromic, and, more generally, representation-theoretic significance.

Quantum and classical integrable systems, both finite and infinite-dimensional ones, will be a recurring cast of characters, along with the other (q⁢q-) characters.

Acknowledgements

Notes were taken by Patrick Lei for all lectures until March 22, 2024 except for the lecture on December 8, 2023. Davis Lazowski provided notes for the lecture on December 8, 2023 and for all lectures from March 29, 2024 onwards. Slight edits (in form, not in content) were made to the notes taken by Davis to make them fit with Patrick’s stylistic conventions.

1 Classical mechanics

We will be discussing three types of physics in an attempt to create something mathematically interesting:

  • •

    Classical physics;

  • •

    Statistical physics;

  • •

    Quantum physics;

1.1 Classical physics

1.1.1 Hamiltonian dynamics

We will begin with a space of classical states, which is most commonly known as a phase space. This is a symplectic manifold (M,ω), where dimM=2⁢m and ω∈Ω2⁢(M) satisfies

This carries a function

called a Hamiltonian. Then there is a vector field VH described by

which generates a 1-parameter group gt of symplectomorphisms of M. The evolution law of the physical system is given by

Because gt acts by symplectomorphisms, the graph

is a Lagrangian submanifold. Recall that a submanifold L⊂M of a symplectic manifold is called Lagrangian if dimL=m and ω|L=0.

Exercise 1.1.1.

Locally any symplectic manifold is given by M=ℝ2⁢m with the symplectic form

This coordinate system on ℝ2⁢m is unique up to Sp⁡(2⁢m).

Now if V is a vector field such that 𝖫𝗂𝖾V⁡ω=0, then dιV⁢ω=0. Thus there locally exists a Hamiltonian hV. We also need the Hamiltonian vector field to be linear in the local coordinates, so the Hamiltonian itself must be quadratic.

Example 1.1.2.

An important example of a symplectic manifold is T∗⁢B for any smooth manifold B. There is a 1-form θ on T∗⁢B given by the following formula. If v is a tangent vector at the point (p,b), then

Then ω=dθ is a symplectic form.

Example 1.1.3.

Another large class of examples are obtained by symplectic reduction. Here, we suppose that a symplectic manifold M carries the action of a compact Lie group G by exact symplectomorphisms. This defines a moment map M→𝜇𝔤∗ by the formula

There is some ambiguity in the choice of constants, but in the end we obtain a new space

In practice, we want the moment map μ to be equivariant with respect to the coadjoint action on 𝔤∗. Then the manifold M⫽G has a canonical symplectic form, but this requires a lot of work.

Now consider M=ℝ2⁢m and G=U⁡(1), where we write M=(ℝ2)m and U⁡(1) acts by rotations. Then the moment map is actually

so μ−1⁢(0) is a sphere. We then obtain

We made no use of the complex numbers, so the fact that we obtain a complex manifold will be viewed as a bonus. The reduced symplectic form is simply r times the Fubini-Study form.

Note that ℙm−1 is compact, so the interpretation that phase space records position and momentum breaks down. In this case, our phase space is called the classical spin phase space, where the motion is by rotations rather than by translation.

Remark 1.1.4.

We will often consider time-dependent Hamiltonians, where x˙=VH⁡(t)⁢(x).

1.1.2 Lagrangian mechanics

There is another point of view, where dynamics on M are given by an optimization problem on the space

of paths in M. We will consider an action

where we assume that ω=dθ. If we further require that γ⁡(0)∈L0 and γ⁡(1)∈L1, our dynamics are well-defined if L0,L1 are Lagrangian submanifolds. Of course, we are looking for paths were δ⁢𝕊=0.

1.1.3 Classical field theory

Classical field theory should be thought of as an infinite-dimensional version of classical mechanics, where we study loopified versions of finite-dimensinoal manifolds. We want to consider integrals

where Σ is the spacetime, h is a metric, and ϕ are the fields. Fields could be one of several options:

  • •

    Scalars f:Σ→X, where X is a Riemannian manifold;

  • •

    Connections ∇ on principal bundles

    called gauge fields;

  • •

    The metric h itself, called gravity.

We can introduce more complexity into the problem by varying the action and looking for solutions of PDEs, introducing boundary to Σ, and other operations. Also note that Lagrangian mechanics can be interpreted as a 1-dimensional classical field theory.

1.2 Statistical physics

In statistical physics, a point is replaced by a cloud of points, or a probability measure. For example, the measure could contain the term e−β⁢H, where H was the classical Hamiltonian and β is a parameter of our distribution (inverse temperature). The system often flows to a stationary distribution, which is determined by the outside world. In reality, the distribution will have the form

This factor Z is called the partition function, and most of our energy is spent on computing this partition function.

1.3 Classical physics with multiple times

Suppose we have Hamiltonians H1,…,Hk such that [VHi,…,VHj]=0. Then we obtain dynamics

and an action

defined on paths inside the cube [0,ε]p. The extreme case of this is an integrable system where the Hi are functionally independent, and the maximum possible value of p is m.

Theorem 1.3.1 (Liouville-Arnold).

If the motion is finite (fits in a compact set), then locally M is a fibration

such that the Hi factor through B and the VHk span the rotations on each S1 factor of Tm.

A typical trajectory is a winding of the torus, where if θi are the angle coordinates on Tm, there is the formula

Generically, these paths will have dense image.

In the special case of an integrable system, there are action-angle variables, where the symplectic form is

The θi are defined up to S⁢L⁢(m,ℤ) affine transformations. If Ci∈H1⁢(Tm,ℤ) form a basis, then the Ii are defined by

Here, Ci is transported to other fibers via the Gauss-Manin connection. We should note that the Ii are not well-defined, but the quantities Ii⁢(b′)−Ii⁢(b) are well-defined.

Because the Hk are defined on the base B, we can write Hk⁢(I1,…,Im). Fixing τ1,…,τm∈ℝ, we can flow along

we obtain

In “reality,” which is non-integrable, consider an approximation

Then we can understand the approximate evolution with respect to H by averaging H1 over Tm.

1.4 Gauge symmetry

We would now like to discuss the idea of gauge symmetries and gauging in the Lagrangian formalism. Recall that the phase space (P,ω) carries an action

and we want to consider the effect of an action of a group G on P. The moment map

is equivariant with respect to the coadjoint representation. Note that P⫽G has a symplectic form ω~, and locally P looks like

Thus ω restricted to a tubular neighborhood of μ=0 has the form

Recall that we are looking for extrema of 𝕊, and we need to find a path on the quotient space. We need to enlarge the space of variables to include A∈ΩI1⁢(𝔤). Now we will define

The space of possible A has an infinite-dimensional symmetry generated by

The action is given by

Note that A transforms as a connection, not as a 1-form.

Remark 1.4.1.

The translation by g−1⁢dg compensates for the change of d−1ω=∑pi⁢dqi under the action of G.

1.4.1 Rational Calogero-Moser model

The first example is called the Calogero-Moser-Sutherland model. The phase spaces are

with the standard form

For any ν∈ℝ+, the Hamiltonian is given by

These are systems of n particles where they start out very far away from each other, are brought closer together, and then repel themselves apart again. These turn out to be integrable systems, and in fact they can be obtained by reduction of a system on a higher-dimensional symplectic manifold.

Define the unreduced phase space

This is a pair of N×N Hermitian matrices (P,Q) with a vector z∈ℂN. The Liouville form will be written as

Then we may define the Hamiltonians

The flows look like

so they clearly commute. This system carries a U⁡(N)×U⁡(1) symmetry, where

Because this preserves the symplectic form, we may perform the symplectic reduction. Becuase U⁡(n) is not simple, there is a free parameter ν, so the moment map is given by

We only need to solve μ=0 up to U⁡(N), so we choose a diagonal representative of {u−1⁢Q⁢u}. Thus, assume that Q=diag⁡(x1,…,xN) is diagonal with x1≥⋯≥xN. Generically, the inequalities are strict. Then

If i=j, then |zi|2=ν, so the remaining U⁢(1)N-action can be used to set zi=zi∗=ν. We can now compute

for the non-diagonal elements. We cannot compute the diagonal elements of P, so we obtain

In this form, the Hamiltonians become

which is what we wanted. We need to show that ω=∑dpi∧dxi, so we will compute the Poisson brackets of various functions. Recall that functions on (P,ω) form a Lie algebra with the Poisson bracket

We note that

The functions we will consider are the resolvents

Because the trace is cyclic, we obtain

Therefore

On the reduced space, the functions become

This is equivalent to {xi,xj}=0={pi,pj}, so {pi,xj}=δi⁢j.

Note that this system has an alternative presentation where we assume that P=diag⁡(p~1,…,p~N) and

Then the Hamiltonians reduce to

and the flows are given by

This system is not very interesting, but we could instead consider

and obtain a system with position and momentum exchanged.

1.4.2 Trigonometric Calogero-Moser (Sutherland)

Now consider P=T∗⁢U⁢(N)×ℂN. Then we have a triple (P,g;z) where P is Hermitian and g(=exp⁡(−1⁢Q)) is unitary. The moment map is given by

We may choose to either diagonalize P as diag⁡(p~1,…,p~N) or diagonalize g as diag⁢(e−1⁢x1,…,e−1⁢xN). Making the latter choice, we obtain

The Hamiltonians in this case are

Making the former choice, we obtain another integrable system called the rational relativistic Calogero-Moser system or the rational Ruijsenaars model. In this model, the Hamiltonians look like

Here, a relativistic particle in 1+1 dimensions has energy and momentum given by

so E2−p2=m2.

1.5 Infinite-dimensional symmetries

We will now replace 𝔤=𝖫𝗂𝖾 with 𝔤̂=𝖬𝖺𝗉𝗌⁡(S1,𝔤)̂, which is a central extension of the space of maps with commutator given by

Then 𝔤̂∗ is not 𝔤 but instead

with the pairing

We will also consider the Lie algebra 𝖬𝖺𝗉𝗌⁡(ℝ,𝔤), but this requires us to specify some kind of boundary conditions at ∞. We may also consider L2⁢(ℝ)⊗𝔤. In the case of S1, note that

is a 2-cocycle and that

is 1-dimensional, so this is the only nontrivial cocycle.

The corresponding group is given by the following construction. Define

Then L⁢Ĝ is a nontrivial U⁡(1)-bundle

Note that H2⁢(L⁢G,ℝ)≃ℝ. The cohomology H3⁢(G,ℤ) is nontrivial with a nontrivial class given by

Then there is an evaluation map

and then

represents c1⁢(L⁢Ĝ→L⁢G). Therefore, we have an identification

where 𝖬𝖺𝗉𝗌⁡(D2,G)̂=𝖬𝖺𝗉𝗌⁡(D2,G)×U⁡(1) with multiplication

To embed 𝖬𝖺𝗉𝗌⁡((D2,S1),(G,1)) as a normal subgroup, we make use of the fact that π2⁢(G)=0, so any map g can be extended to g~:B3→G. Then we define

The fact that this construction is well-defined is the Polyakov-Wiegmann formula.

We will now discuss the adjoint and coadjoint actions of L⁢Ĝ on ĝ,ĝ∗ respectively. Infinitesimally, we have

The action on the dual space is given by

Therefore, we obtain

Note that Ak is a g-connection 1-form on S1.

There is now a natural candidate for a symplectic form, which is

Here, δ is the differential in the space of fields. The moment map μ:T∗⁢ĝ→ĝ∗ is given by

where G⁡(k,c,A,ξ)=−k⁢dξ+[A,ξ]. We will now compute

We will now solve the moment map equation with the assumption that k≠0. We will scale k=1, so the equation becomes

This is a first order matrix differential equation with periodic coefficients which can be studied using Floquet-Lyapunov theory. This says that there exists g such that

is constant and lies in a maximal Cartan of 𝔤. What this means is that we can write

These satisfy the equations G˙⁢G−1=A and G⁡(0)=1. The monodromy is

This must commute with ξ0, so we can bring

Recall that GA can be brought to T⊂G and A can be brought to α∈𝔱 and gA=exp⁡(2⁢π⁢α). There is still some remaining symmetry by

where λ∈Λ∨ is in the lattice of coroots and u∈S1=ℝ/2⁢π⁢ℤ is the coordinate. This shifts α↦α+λ while preserving monodromy. The second kind of remaining symmetry is the Weyl group W≔N⁡(T)/T. Taking their semidirect product, we obtain the affine Weyl group.

If we consider the weight space decomposition of the moment map equation and β is a root of 𝔤, then the equation for this component is

Because ξβ⁢(u)=e−u⁢⟨β,α⟩, this is generically not 1, so ξβ=0. Therefore ξ∈𝔱, and we obtain

Note that T parameterizes conjugacy classes of P⁢exp⁡∮A and that T=𝔱/Λ∨. Unfortunately, the reduced space is an orbifold, not a manifold.

We will now attempt to remedy this situation by modifying the quotient. Instead of setting the moment map to be 0, we want to consider an orbit. We want 𝒪=ℙN−1, and if G=S⁢U⁢(N), L⁢G acts on ℙN−1 by evaluation at some 0∈S1. We choose z∈ℂN such that z†⁢z=N up to z∼z⁢ei⁢α, and the modified equation is

Remark 1.5.1.

While most of the orbits are infinite dimensional, we are taking some limit where A becomes a distribution on S1 supported on finitely many points.

We first apply Floquet-Lyupanov to make A=diag⁡(a1,…,aN) diagonal with ∑ai=0. Then on each coordinate we obtain

Because ξi⁢i⁢(+0)=ξi⁢i⁢(2⁢π−0)=ξi⁢i⁢(−0) for any 0∈S1, |zi|2=1. Using the maximal torus, we may force zi=1. Then we obtain

Finally, the initial value is

Note that this appeared in our study of the Sutherland system.

2 Quantum mechanics

2.1 Quantization

The motivation for complexification of our systems is quantization. Recall that if S is the action of a Lagrangian system, many systems are not described by solving the variational equation

but by meditating on formal path integrals (due to Feynman)

The classical system is obtained via stationary phase approximation. Mathematically, this is ill-defined, but if X is a finite-dimensional manifold, we can consider oscillating integrals

Here, X is one of many possible cycles in the complexification Xℂ and μ is viewed as a holomorphic top-degree form, so this is just a period. If dimX=n, then we may consider other Γ such that

converges. These will satisfy

where

is set to force the integral to converge. These Γ are chosen to flow from critical points of S (equivalently, of W=i⁢S) in Xℂ into X≪ℂ. We can construct cycles using Lefschetz thimbles. We can choose a critical point p which satsifies dW⁢(p)=0 and then consider the steepest descent flow for Re(Wℏ), or in other words

where the gradient is taken with respect to some metric on Xℂ, and finally take the union Γp of descending trajectories. If we choose a Hermitian metric, then Im(Wℏ) is actually constant. For a generic choice of ℏ, the critical points have different imaginary parts, but in the special settings we may have Stokes phenomena and wall-crossing behavior of solutions.

Exercise 2.1.1.

What happens for W=∑i=1nzi2?

2.2 Holomorphic symplectic dynamics

Now let Mℂ be a holomorphic symplectic manifold and consider 𝔛ℂ=𝖬𝖺𝗉𝗌⁡(S1,Mℂ) and

We will take S1=ℝ/ℤ and we want to find points with dW=0.

Remark 2.2.1.

When we quantize everything, we will obtain

We obtain the equations

If β is real, this is the usual Hamiltonian dynamics, but there may not be real solutions if β is not real. On the other hand, if β=i⁢βE is purely imaginary (also known as going to Euclidaen time), then there may be solutions where q is real and p is purely imaginary.

Example 2.2.2.

Let M=ℝ2 and

where λ,a are parameters. This is usually called the Higgs potential. Then the energy is

We can see that there is a q↦−q symmetry, so near ±a there are two copies of the same physics. After complexifying, we obtain Mℂ=ℂ2, while the zero set CE of E is a Riemann surface. If we compactify, we will obtain an elliptic curve with the hyperelliptic form

Note that H1⁢(CE,ℤ)=ℤ⊕ℤ, so there are two independent cycles corresponding to classically allowed physics. Also, it is clear that

Therefore, this family degenerates to a union of two copies of ℙ1 when E=0 (and one of our distinguished cycles is the vanishing cycle) and has another critical point when E=λ⁢a44. Now ω=dqp is a holomorphic differential on CE, and the Hamilton equation tells us that any

must satisfies γ∗⁢ω=β⁢dt. Now in the low-energy region, [γ]∈H1⁢(CE,ℤ) can be specified by two integers:

where [A] is the vanishing cycle and [B] satisfies A∩B=1. Note B is described up to multiples of A, so n is well-defined, but m is defined only up to multiples of 2⁢n. We can then find β by the period integral

which are functions of E. Therefore, the E=Em,n⁢(β) for n∈ℤ,m∈ℤ/2⁢n⁢ℤ satisfy a transcendental equation

We can solve this equation approximately in the β→∞ limit, where

where Ω2=2⁢λ⁢a2 is the classical period of very small oscillations around the q=a critical point. This is all obtained by Picard-Lefschetz theory using the fact that

where the +⋯ can be computed using knowledge of elliptic integrals. The solution where n=0 corresponds to classical physics, while the solution where m=0 corresponds to tunneling between the two critical points. Therefore, the general (m,n) solution is some superposition of classical motion and tunneling.

2.3 Algebraic integrable systems

We will now consider the class of systems which generalize the following feature of the previous example: our manifold (Mℂ2⁢n,ωℂ) has a Lagrangian fibration M2⁢n→Bn to some open subset of ℂn by polarized abelian varieties. These are called algebraic integrable systems. This structure gives the structure of special Kähler geometry on Bn.

For us, a polarization is simply an integral class t∈H2⁢(F,ℤ), where F=π−1⁢(b). We will now define action variables (ai,aD,i) on B. If we have γ∈H1⁢(Fb,ℤ), this path can be transported canonically over paths connecting b to b′ that avoid the discriminant locus Ξ. We now obtain a 2-chain in Mℂ covering the path, and we can obtain a number

for a distinguished choice of b∗. Therefore, we obtain a map

where Bn∖Ξ~ is the space of choices (b,γ), where b∈Bn∖Ξ and γ is a path connecting b∗ to b in Bn∖Ξ up to homotopy. The image of a is a Lagrangian with respect to t, where if Ai,Bi is a basis of 1-cycles in H1⁢(Fb∗,ℤ), then t⁡(Ai∩Bj)=δij. Now because

locally there exists ℱ⁡(a) such that

Definition 2.3.1.

This ℱ is called a prepotential.

Example 2.3.2.

One example is the elliptic Calogero-Moser system. Here, let E be an elliptic curve parameterized by τ∈ℋ and consider

Then let pi be the coordinates in the fiber directions and zi be the coordinates on the copies of E and define

This was proven to be an algebraic integrable system by Krichever (before algebraic integrable systems were defined). Set

where the theta function is defined by

Here, we make the usual substitution q=e2⁢π⁢i⁢τ. This operator satisfies the equations

and has an expansion of the form

Therefore, the spectrum of L⁡(z) forms an n-sheeted cover over E, and in T∗⁢E, if g⁡(C)=n, then

Then we will obtain a Lagrangian fibration where the fiber is the Jacobian Jac⁡(C).

2.4 How to compute some path integrals exactly

Recall our setting of a symplectic manifold with Hamiltonian (M,ω,H). We will assume that there is an action of T≅U⁢(1)r with moment map μ:M→𝔱∗. We will also assume that H is linear in the moment map, so

Example 2.4.1.

Consider M=ℝ2⁢n with Darboux coordinates pi,qi, the standard action of U⁢(1)n, and moment map given by

Then the Hamiltonian

generates an action ℝ↪T which is dense for generic ξ.

If we now assume that M is compact and the fixed points of the T-action are isolated, then to obtain the statistical-mechanical partition function for some inverse temperature β, we have the Duistermaat-Heckman formula

where mi⁢(p) are the weights of the T-action on Tp⁢M. With these weights, H near p behaves like

Example 2.4.2.

If M=S2 and H=cos⁡θ is the cosine of the azimuthal angle, we embed M⊂Mℂ≅T∗⁢S2, and the contour corresponding to the south pole goes into the cotangent directions.

The Duistermaat-Heckman formula can be stated in the setting of T-equivariant cohomology. If G is a Lie group acting on M, we will consider the Cartan model of equivariant cohomology.

where we require that

The grading on differential forms is deformed to be

Then the equivariant differential is defined by

where V:𝔤→𝖵𝖾𝖼𝗍⁡(M) is a linear map. We can now compute

so we set HG∗⁢(M) to be the cohomology of this complex. We can now compute equivariant integrals of f∈ΩG∗⁢(M). Note that ∫Mf∈𝖥𝗎𝗇⁢(𝔤)G and

for all g∈G. Also note that

We will now prove the Duistermaat-Heckman formula. If μ:M→𝔤∗ is the moment map, we claim that

This follows from the definition of the moment map. We then see that

and this is actually a Duistermaat-Heckman integral. If the G-action is free (and G is compact), then every closed form is exact. Because G is compact, there exists a G-invariant metric. Assuming that G=T for now, choose some generic ξ¯∈𝖫𝗂𝖾⁡T. Then define

We then obtain

where the function part is nonzero, and finally if

we obtain f=D⁢ψ whenever D⁢f=0. Now we can replace the integral in the Duistermaat-Heckmann formula with

where Mε=M∖⋃pBε⁢(p), and finally use Stokes’ theorem to obtain the desired result.

2.5 Infinite-dimensional generalizations and supersymmetry

Now suppose that M=L⁢N, where (N,g) is a Riemannian manifold. Then define for ξ,η∈Tρ⁢M=Γ⁡(S1,ρ∗⁢T⁢N), we can define a 2-form by

where ∇t is the Levi-Cevita connection. The Hamiltonian that generates the standard S1-action is

Then the partition function of supersymmetric quantum mechanics on N (index of the Dirac operator on N) is given by

where  is the Â-genus. The term on the right should be viewed as the contribution of non-isolated components of the fixed locus and can be thought of as Fourier modes of infinitesimal loops. Of course, there is the question of the role of β in this formula.

2.5.1 The Dirac operator

First, associated to the metric g|Tx⁢N we have the Clifford algebra, which is generated by symbols v for v∈Tx⁢N=V with the relation

This is the analogue of the the Heisenberg-Weyl algebra over a symplectic vector space (W,ω), which is generated by elements ŵ for w∈W with the relation

The Clifford algebra Cl has irreducible representations, which are called spinors. If we choose an orthonormal basis {ei} of V, we can write any element as

In addition, we see that the Clifford algebra relation is a deformation of the exterior algebra, so Cl≅Λ∗⁢V as ℝ-vector spaces. The space of quadratic elements of the Clifford algebra is in fact 𝖫𝗂𝖾⁡O⁡(V). Because O⁡(V) is not simply connected, representations of its Lie algebra integrate to Spin⁡(V), which is the universal cover of S⁢O⁢(V) (at least if dimV>2). The group-like elements take the form

Now we impose a complex structure compatible with the metric and define

for a=1,…,⌊dimV2⌋. These satisfy the relation

Finally, we define the vector space

We impose that c∗⁢|vac⟩=0, and in general that c∗ moves to the left. We then define Γ=edimV if dimV is odd. It satisfies

From now on, we will assume that dimV=2⁢k.

Note that S is an irreducible representation of 𝒞⁢𝓁, but recall that 𝔬⁡(V) is much smaller. It is spanned by elements of the form ca⁢cb,ca⁢cb∗,ca∗⁢cb∗. There is a ℤ/2-valued conserved charge called (−)F which is preserved by 𝔬⁡(V). For S, it is defined by

Therefore, we can split

by the parity of the F-charge. Then for any u∈V, we have

These are called Dirac matrices.

Unfortunately, there is not always a global spinor bundle on N. The obstruction is the Stiefel-Whitney class

If this is nonzero, then we cannot glue spinors into a vector bundle and thus the manifold N does not have a spin structure. Therefore, we will now assume that w2⁢(T⁢N)=0. We will now define the Dirac operator

by the local formula

Here, ∇i is a spin cover of the Levi-Civita connection and γi is the physicist notation for ei. We also define D∗ to be the same operator applied to S−. The symbol of D∗⁢D or of D⁢D∗ is given by the well-known formula

Because of the analogy with the Laplacian, the spaces HS+=ker⁡D and HSi=ker⁡D∗ are known as harmonic spinors. There is a lucky coincidence that if Ĥ=D⁢D∗+D∗⁢D, then

for any β>0. Note here that working on ℝn,

Now suppose that ψ(k)∈Γ⁡(S+) satisfies

Then D⁢ψ(k)=χ(k)∈Γ⁡(S−) satisfies

This tells us that

Note that this cancellation is our first example of supersymmetry. Finally, we take the β→0 limit.

2.5.2 The heat kernel

If we were considering on flat space the ordinary heat kernel

then Tr⁡eβ⁢Δ is computed by

Note that in flat space, the heat kernel must satisfy the PDE

As β→0,

There must also be the integral formula

so in fact, it is given by

In our curved situation, we obtain the limit

which can be well-approximated by the expression in flat space. Considering instead the loop space L⁢T∗⁢N, we now obtain

Note that the [D⁢x⁢(s)⁢D⁢p⁢(s)] carries the term ∏t1(2⁢π⁢β)dimM2. Also, if we integrate out the p, we obtain the previous term

2.5.3 Supersymmetric Duistermaat-Heckmann

We will now rewrite the integral we wanted to compute at the beginning of this section in new notation as

Here, we rewrite

where (ψi⁢(s))∈Γ⁡(S1,x∗⁢T⁢N). The ψi satisfy the formula

Finally, here ε=1β is the equivariant parameter for the action of U⁡(1) on the cotangent directions.

Now, the Duistermaat-Heckmann normalized integral (which is not the Atiyah-Singer normalized integral) is

where xi⁢(t) splits into Fourier modes as

and thus splits

and α are the Chern roots of T⁢N. This exactly reproduces the A^-genus. The Atiyah-Singer normalized integral has an extra factor of 1∏sβdimM2.

2.6 Duality

There are two main examples: T-duality in 2d sigma-models and S-duality in 4d gauge theories.

2.6.1 p-form generalized gauge theory

We will begin with p-form generalized gauge theory in D spacetime dimensions. For some A∈Ωp⁢(MD), define

This is invariant under A→A+dB. The space of fields in the ℝ-type theory is given by

In the U⁡(1)-type setting, the space of fields is smaller and is given by

where Ωℤp is the space of all p-forms alpha such that

for all integral cycles Zp.

Example 2.6.1.

When p=0 and D=2, then the U⁡(1)-type theory describes maps M2→U⁡(1). Locally, these look like A:M2→ℝ such that A∼A+n. Of course, not every manifold is simply connected. Therefore, the true space of fields is

This has a decomposition by the topological type as follows. Let t∈ℝ be a coordinate with

Then the topological type is given by

In general, we will assume that the curvature form dA is not exact, but instead integral, so

This can be achieved on an open cover Uα by choosing

such that

Then the space of fields is the space of connections on all U⁡(1)-bundles and its connected components are parameterized by

The partition function is now given by

where

is a closed form.

We may also replace f:M2→U⁡(1) by functions

We introduce tensors

In this form, the partition function becomes

where dBi⁢j=0 and Bi⁢j∈ΩD−2⁢(p+1)⁢(MD). We also have

We now want to find the critical points of ℒ, where δ⁢ℒ=0. The equations for A become

for all i. If we set Fi≔dAi, the equation becomes

which are the generalized Maxwell equations. Therefore, we obtain

where 𝒩⁡(G,h) is a regularized version of

which is defined as follows. Define

This is defined for Res≫0, and taking the analytic continuation, we define

We may need to assume that D=2⁢(p+1). For example, if D=2, then the ⟨ci,∗cj⟩ term knows only about the conformal structure of M2. Finally, Hp+1⁢(MD,Γ) is taken modulo torsion.

2.6.2 Duality

The duality is given by inverting

when D=4 and inverting

when D=2. We may consider transformations of the form

or more generally

where

Applying Poisson resummation, we obtain

Here, we have the identity

in dimension 2. Note that when D=2, ∗2=−1, while when D=4, ∗2=+1. In the 4-dimensional setting, if we define

the relation is

Therefore,

The duality must give some reassignment of the degrees of freedom. Locally, if

which is noncompact, we must discuss the Hamiltonian system. The phase space is given by the stack

The symplectic form is given by

Here,

is the electric field and dA is the magnetic field. Now we can write the Hamiltonian as

We return to the space of spatial fields 𝒜, which has a decomposition into connected components indexed by

The component corresponding to some c is

which is the space of flat p-connections.

All of this can be understood via the quantum mechanics of a particle on S1. If we define

then θ is the analogue of (Bi⁢j). Then if the standard symplectic form on T∗⁢S1 is given by

E2 quantizes to −∂t2=Ĥ, and the spectrum of this operator is given by

For example, if En=0, then n and −n have the same eigenvalues, if θ=12, then the spectrum is doubly degenerate, and for any other value of θ the spectrum is simple. Therefore, the duality exchanges the electric and magnetic fields whenever D=2⁢(p+1).

3 Instantons

There are three main approaches:

  • •

    Morse theory;

  • •

    Twisted supersymmetric σ-models;

  • •

    Twisted supersymmetric gauge theory.

These have a Hamiltonian and Lagrangian approach, and the latter will motivate the study of integrals on moduli spaces of instantons.

3.1 Morse theory

Note that this is not the same as what topologists call Morse theory. Let (M,g) be a compact Riemannian manifold and f:M→ℝ. Suppose all x such that dfx=0 have nondegenerate Hessian, or in other words,

We will consider Ω∗⁢(M) with the differential

Using the standard scalar form

we define the adjoint

Physicists refer to the study of this package as supersymmetric quantum mechanics. Here, the Hamiltonian is

Note that ker⁡H≅H∗⁢(M), and if H⁢ψ=0, then

so we see that H⁢ψ=0 if and only if D⁢Ψ=D∗⁢ψ=0.

There are also excited states where H⁢ψi=Ei⁢ψi for Ei>0. Then we compute

so the states D⁢ψi, D∗⁢ψi, and D⁢D∗⁢ψi are also eigenstates with eigenvalue Ei.

There is then the following trick. If f↦t⁢f and g is fixed for t≫0, then

where ‖∇f‖2 is a very large potential outside of the critical locus of f. We also have

on the ground states, which are isomorphic to H∗⁢(M). More generally, we note that

Ignoring the t−1⁢d∗ term, the Hamiltonian becomes

In the world where f↦t⁢f and g↦t⁢g, we obtain states

As t→∞, we obtain Hin and Hout, which are distributions.

Example 3.1.1.

Now consider M=ℝ and f=ω⁢x22. Note in this case we need to use L2 differential forms. Then we have basic operators

Finally, we set α=ψ⁡(x)⁢dx.

The spectrum is given in the following way. If H⁢ψ=0, then the equality (∂x+ω⁢x)⁢ψ=0 implies that ψ=e−f, which is fine if ω>0 and bad if ω<0. Then suppose there is ψn such that H⁢ψn=n⁢|ω|⁢ψn. Then we obtain

Then if P is a polynomial,

as t→∞ (here t scales ω). This tells us that Hin0≅Hout1 are regular functions, while eigenstates of H∞ are monomials. Similarly, Hin1≅Hout0 are distributions supported at x=0, while the eigenstates of H∞ are those of the form ∂xnδ⁡(x). Observe that

Then, note that δ⁡(x)⁢dx is an invariant distribution-valued 1-form. The pairing between the in and out states must be the usual pairing between functions and distributions.

In higher dimensions, there will be both attracting and repelling states, so we can consider something like

and i the end ψin is a distribution supported on the attracting manifold. This concludes the study of the local picture.

However, we must consider the compact picture. For example, consider S1=ℝ∪∞ where the Morse function f is attracting at x=0 and repelling at x=∞. Then near ∞, the states corresponding to monomials will look like P.V.⁡xn. Applying x⁢ddx and expanding

then the scaling x→t⁢x,ε⁢t⁢o⁢t⁢ε gives us correction terms of log⁡ε and log⁡t to ψ∞(n+1). Here, the principal value

is the finite part in the ε-expansion of

Therefore, we see that critical points can talk to each other via excited states.

3.2 σ-models

We will consider maps Σ→(X,ω) from a Riemann surface to a symplectic manifold. We will also choose an almost complex structure J which is tame with respect to ω. If we consider the functional

its critical points are simply harmonic maps.

In the Hammiltonian formalism, consider M=L⁢X. If ξ∈Tℓ⁢L⁢X, then

Note that

and

Then 𝕊 can be rewritten as

where B is the B-field, satisfying dB=0. Therefore, the absolute minima are solutions to the PDE

Example 3.2.1.

Consider X=ℝ2 with the standard symplectic form and complex structure. Let

Let t be the loop variable and s be the noncompact direction. Then we obtain the equations

which are the Cauchy-Riemann equations. If z=s+i⁢t and w=exp⁡(z), the Cauchy-Riemann equations give

3.3 4d supersymmetric Yang-Mills

The word symmetry here is actually a confusing one. There are two types of symmetries one encounters in physics, and one is not a symmetry. Sometimes what we mean by symmetry is a redundancy; mathematically what that means is that we’re studying a quotient of some space X by some symmetry G, X/G. This is a local symmetry. The other type of symmetry is a global symmetry where we really have G acting on the space we are studying X. In the case of a local symmetry, platonically there is a space X with action of G, but that space is not accessible to us. Rather we only can see X/G. Occasionally, it’s the case that we find there exists Y/H=X/G. We call this a duality.

Because the groups we deal with in physics are typically infinite dimensional groups, for example 𝖬𝖺𝗉𝗌⁡(M,G) of a manifold into a finite dimensional group, these groups typically have lots of normal subgroups. Sometimes rather than the full 𝖬𝖺𝗉𝗌⁡(M,G), our local symmetry is some sort of normal subgroup that preserves additional structure: for example, possibly the normal subgroup that sends certain marked points {xi} of M into certain subgroups {Hi} of G. The global symmetry will then be whatever is left from 𝖬𝖺𝗉𝗌⁡(M,G), for example in this case ×iG/Hi.

For example, say we are studying (M,g) a smooth Riemannian 4-manifold, with G a compact simple Lie group, and a principal G-bundle 𝒫 over M. Let 𝒜𝒫 the affine space of G-connections on 𝒫. There is a group g𝒫:=Γ⁡(M4,G×A⁢d𝒫) acting on 𝒜𝒫; what we really want to study is the quotient

The functional we seek to integrate over B𝒫 is

The measure here is induced from the L2 metric, and T⁢r refers to the killing form. Our integral consists of two parts.

  • •

    The first is the Yang-Mills action;

  • •

    The second is a *topological term* which does not depend on the metric and picks out some topological equivalence class of principal bundles. We normalise it so that the whole thing is an integral class.

Since the second term is topological, we know eventually that our integral is equivalent to

where Zn is the integral taken over bundles of Pontryagin class n. We could try to write this on a lattice, viewing 𝒜P as the edges of a graph and g𝒫 as the vertices. But if you did this naively you would find that the coupling has to go to zero. Instead, we can try a trick which we started to discuss last time. If we write

the Yang-mills term in the functional is equal to

Then we can rewrite the whole functional as

where

is some function in the upper half-space. We could try to study the limit c2→0, while keeping τ finite. In that limit the functional goes to

Studying the moduli space of ℳ𝒫:={A|FA+=0}/g𝒫, which we roughly expect to be finite dimensional because Λ2,+⁢ℝ4≃ℝ3, we can at the very least compute its virtual dimension. To do so, assume

  • •

    A0 is so that FA0+=0;

  • •

    A=A0+δ⁢A so that FA0+δA+=dA0+δ⁢A;

  • •

    wherein dA0+:Ω1⁢(M)⊗ad⁡𝒫→Ω2,+⁢(M)⊗ad⁡𝒫.

To the linear level, it suffices to computer ker⁢d+. Since

the virtual tangent bundle is

which is the first cohomology of the Atiyah-Hitchin-Singer complex

wherein the maps are dA0,dA0+. If H0,H2 of this complex don’t vanish, we could have singularities or obstructions and therefore trouble counting dimension.

3.3.1 “theory of differential forms on ℳ𝒫”

We could try to overcome the infinite dimensionality of the problem by seeking to develop a theory of differential forms on ℳ𝒫, the space of anti-self dual connections, i.e. those A such that FA+=0. To do so, we would need some notion of local coordinates on ℳ𝒫. Let’s first try to do this on the space of connections without the ASD condition.

Define

on the space of connections 𝒜𝒫. Then

This is a one-form, hence an ’odd’ object, i.e. a fermion. So Q meaps even to odd and vice versa.

But remember that A:=Aμ⁢𝑑xμ is a redundant description, because connections related by gauge transformation are the same. To deal with this, rather than dD⁢R we will deal with *equivariant differential forms*, and write

on the Cartan model for equivariant cohomology, so that this differential is an operator on

Since H∙⁢((M×E⁢G)/G)≃H∙⁢(M/G), if we want this relation not only at the level of chomology but also on the actual space of differential forms, it’s enough to study invariant forms, hence why we’re taking invariants by g𝒫. This is like passing from M/G→M×𝔤/G, which slightly improves the situation of stabilisers and therefore the space of equivariant differential forms. Then,

The extra factor of the Lie algebra also defines an object ϕ¯, so that Q⁢ϕ¯=η where Q⁢η=[ϕ,ϕ¯]. Here ϕ¯ lives in the 𝖫𝗂𝖾⁡(g𝒫) inside our forms itself, whilst ϕ lives in the space F⁢u⁢n⁢(g𝒫) we tensor by. Then

on the space of g𝒫-invariant forms, Q2=0 since Q is just the Lie derivative associated to transformation along ϕ.

3.3.2 Finite-dimensional model

Suppose we are in a similar finite dimensional situation.

  • •

    Suppose G acts on a Riemannian manifold (N,g);

  • •

    Suppose f∈(Ω∙⁢(N)⊗𝖥𝗎𝗇⁡(𝖫𝗂𝖾⁡(G)))G;

  • •

    If G acts freely, then there is a well-defined quotient by G. If f is equivariantly closed,it should correspond go to ϕ∈Ω∙⁢(N,G).

Remark 3.3.1.

Let π:N→N/G. If ϕ is closed so that ιV⁢π⋆⁢ϕ=0 for all ϕ∈𝔤, with ϕ∈Ω∙⁢(N/G).

Further, characteristic classes of our bundle N→N/G are represented by invariant polynomials on our Lie algebra, (S∙⁢𝔤∨)G. The functions ϕ should somehow correespond to these, ‘it looks too good to be a coincidence’. This consideration hits on the fact that we likely want to use a metric when we make this map.

Let’s do what we already did a previously. Using the metric, we can build a one-form:

It has the following useful property: if D⁢f=0, we can multiply

This will not change the cohomology class, but it will change the representative. For example,

is given by

We can write this as

In order to not worry about contours, we treat ϕ,ϕ¯ as complex conjugates. There is a canonical measure on 𝖫𝗂𝖾⁡(G),

which is just the Haar measure which we have for free. Insert this into the integral ∫f⁢eD⁢θ as an additional term by which we integrate out ϕ.

A procedure like this is ‘universal’ in the sense that it produces a differential form on N. Because all the integrations I’ve done commute with D, this kind of averaging results in a form which is still equivariantly closed. The secret of Yang-Mills theory is that it does exactly that: it takes something simple on the space of connections and does some sort of integration like that to produce a form on the space of connections.

But there’s a second piece, the anti self-dual condition, which we did not consider. If we have M=s−1⁢(0)/G for some section s of a vector bundle E→N, how do we restrict differential forms to the vanishing locus of this section? Naively the idea is to just send

where δ is just a δ function or δ form. To do this properly we need to add on a Koszul complex into our complex of forms which effectively has s as a differential.

At the level of mysterious formulae, this is achieved in the following way. Extra complexes imply we have extra fields; introduce the field χ which for our purposes will be the self dual form χ∈Ω2,+⁢(M)⊗∏ad⁡𝒫. This χ represents the (−1) term in the Koszul complex, so that the Koszul differential of χ is δ⁢χ:=FA+=s

Then we have a vector bundle EP over 𝒜𝒫 with fibers

Naively Q2 is dA+, so not zero on the nose. This is fixed in the following way: we define another field, χ∈Ω2,+⁢(M)⊗∏ad⁡𝒫, and a further field H∈Ω2,+⁢(M)⊗ad⁡𝒫, so that

Remember also that we had fields

  • •

    A a connection on 𝒫;

  • •

    ϕ¯∈Ω0⊗ad⁡𝒫;

  • •

    ψ∈Ω4⁢(M)⊗ad⁡𝒫;

  • •

    η∈Ω0⊗ad⁡𝒫;

  • •

    ϕ∈Ω0⊗ad⁡𝒫.

which is a minimal list of fields for Yang-Mills theory. Q acted on these fields by

  • •

    Q⁢χ=H;

  • •

    Q⁢H=[χ,ϕ];

  • •

    Q⁢A=ψ;

  • •

    Q⁢ψ=dA⁢ϕ;

  • •

    Q⁢ϕ¯=η;

  • •

    Q⁢η=[ϕ,ϕ¯];

  • •

    Q⁢ϕ=0.

Note that for all but ϕ there is a canonical pairing between fields just by their type, called Berezin measure. This gives us a measure:

where we have only ‘cheated’ in the last term, whose definition leaves something to be said in the fully infinite dimensional setting.

Consider the functional

We got these terms from looking at our metric g=∫Tr⁡(δA∧⋆δA) and applying Q to create the most general supersymmetric or Q-closed functional. This is the same as

Since H is quadratic, we could integrate it out.

Also integrating χ,η out represents the Atiyah-Hitchin-Singer complex, because it imposes the conditions dA+ψ=0,dA⋆ψ=0, equivalent to ψ∈ker⁢dA+, ψ is orthogonal to Im(dA), which gives us the first cohomology of the AHS complex.

The zero modes of χ correspond to the second cohomology of the AHS complex, and the zero modes of η correspond to the first cohomology. It’s something to keep in mind because zero modes are significant because they drop out of the exponential and so observables ‘have to provide the missing zero modes’ for the fermionic integrations to be nonzero. Also, we could determine that the functional implies that ϕ is ‘the curvature of the universal bundle evaluated at the point x’, but we would need to talk more about this later on.

3.3.3 When M is not compact

There is one important change to the whole story when M is not compact; in other words, if the metric is such that there is some ‘end’ which is infinitely far away, the story has to be modified. We don’t need to integrate over all fields, rather we want to restrict to those where the curvature goes to zero at ∞. Naively that’s because the integral would be divergent if the curvature was not zero.

Since we want ϕ to approach a constant a, ϕ¯→a¯, such that [a,a¯]=0, we only use those gauge transformations which approach unity sufficiently fast at infinity. The rate at which they go to 1 is fixed by the requirement that ‖dϕ‖2<∞. The result of that is that a finite dimensional group of those gauge transformations preserved now acts as a global symmetry! Instead of integrals over the moduli space of instantons of closed differential forms, one passes to equivariant differential forms for the moduli space of framed instantons, where framed means the vanishing condition above.

4 Gauge-theoretic instanton counting

The difference between mathematicians and physicists is that mathematicians like closed 4-manifolds while physicists like ℝ4. For mathematical precision, we will let M4 be a compact Riemannian manifold and 𝒫→M be a principal G-bundle for some compact Lie group G. These are classsified by the second Chern class

which physicists call the instanton charge. We are looking for connections ∇=d+A whose curvature satisfies

which are called instantons. This equation in fact only depends on the conformal class of the metric.

Being anti-self-dual implies that the connection gives a minimum of the Yang-Mills action

If M4 is a complex surface and the metric g is Hermitian, then F∇0,2=0, which is equivalent to requiring that that ∂¯A2=0. Therefore, any representation E of G gives rise to a holomorphic vector bundle ℰ on M4. This will come into play later when we cheat by replacing instantons with torsion-free sheaves.

The moduli space of such instantons is

where 𝒢𝒫 is the group of sections of the associated Ad𝒢-bundle acting by

For generic metrics g, this is a manifold of dimension

where h∨ is the dual Coxeter number. We should note that even though ℳk as a space depends only on the conformal class of the metric, the metric depends on g itself. It also lives inside the space 𝒜𝒫/𝒢𝒫 of all connections on 𝒫. Our aim is to understand the infinite-dimensional integral

Here, ϑ is valued on S1. In the g2→0 limit, the asymptotics of the integral should become

where q=e−8⁢π2g2+i⁢ϑ and the measure is obtained by regularity and determinants of δ⁢SYM.

Instead of trying to understand this very complicated measure μ, we can try to study simpler but still interesting integrals. For examples, Donaldson invariants are defined as integrals over ℳk of cohomology classes associated to 2-cycles Σi and 0-cycles p on M4. The problem is that ℳk is non-compact because of the conformal invariance of the equations leading to delta-function solutions. Uhlenbeck discovered that via a complicated system of gauge transformations, these points can be filled in by ideal solutions, so there is the Uhlenbeck compactification

4.1 The case of ℝ4

This is the case that physicists are worried about. The problem is of course that ℝ4 is non-compact, so we can view either ℝ4=S4∖∞ or ℝ4=ℂ2=ℙ2∖ℙ∞1. There are already interesting solutions in this case, but because our metric is singular at infinity, we will consider the moduli space

where 𝒢𝒫∞⊂𝒢𝒫 is the set of those elements satisfying g⁡(x)→1 as x→∞. This is a hyperkähler manifold of dimension 4⁢k⁢N when G=S⁢U⁢(N).

It is well-known that this (or a slight modification ℳ~kframed) is a Nakajima quiver variety corresponding to the quiver data

Here, we need to modify the instanton equations to

which in algebraic geometry corresponds to changing the value of the moment map.

This moduli space has an instanton interpretation if we replace ℝ4 by a noncommutative ℝζ4, which has coordinates z1,z2,z¯1,z¯2 with commutators

In the N=2 case, we have ℳ2framed=ℝ4×(ℝ4∖0)/ℤ2 while ℳ¯kframed=ℝ4×ℝ4/ℤ2 and ℳ~kframed=ℝ4×T∗⁢S2.

There is also an action of U⁡(2)×S⁢U⁢(N), where for elements

and

the corresponding vector field is denoted by V⁡(a,ε). Given a choice of

the partition function is

This is not the most general formulation, but it is a distilled version of Yang-Mills with some supersymmetry. It turns out that Zk is actually a rational function of degree −2⁢k⁢N in a and ε.

One generalization of this is

where

Here, Γ2 has the asymptotics

and solves the equation

In the limit as ε1,ε2→0, we expect the asymptotics

so the question now is to evaluate ℱ⁡(a,Λ). In the case when G=S⁢U⁢(N), we can do it by localization.

A standard computation tells us that

where f and g are defined using the (relative) arms and legs of the two Young diagrams. For example, when i=j, we end up with f=ε1⁢(arm□+1)−ε2⁢leg□ and g=−ε1⁢arm□+ε2⁢(leg□+1).

Treating Zk as a probability measure on the set of Young diagrams, we obtain the observables

This function knows essentially everything about the shape of the diagrams, and its vacuum expectation is

It satisfies the property that

has no poles in x and is in fact a polynomial of degree N. Physically, this is interpreted as an interaction between two neighboring instanton sectors.

If we send ε1,ε2→0, then ⟨Y⁡(x)⟩=𝒴⁡(x) we then obtain the algebraic equation

where the coefficients of T⁡(x) are defined by

This is a hyperelliptic curve.

4.2 Mathematical realization

From now on, we will set G=U⁡(N) and consider the moduli space

Recall that this is a Nakajima quiver variety corresponding to the Jordan quiver. The ADHM equations are

Here, ζ→=(ζℝ,ζℂ,ζ¯ℂ)∈ℝ3=ℝ⊕ℂ. It can in fact be rotated to (ζ,0,0) by S⁢O⁢(3), so we can assume ζ>0. Here, the action of g∈U⁡(k) is given by

Then ℳ¯kframed⁢(N) has an action of U⁡(N) given by

and these are symmetries. There is also a U⁡(2)-action given by

In fact, if we only consider μℂ, it in fact has G⁢L⁢(2)-symmetry.

We are interested in the fixed points of our U⁡(2)×U⁡(N) action on ℳ¯kframed⁢(N). To be a fixed point of (u,h)∈U⁡(2)×U⁡(N), the tuple β=(B1,B2,I,J) must obey (u,h)⁢β=gu,h⁢β for some gu,h∈U⁡(k) depending on u,h. It is in fact enough to consider the maximal torus of U⁡(2)×U⁡(N), so we may assume that

Now we obtain

To make this compatible with the gauge transformations, the assignment

is a group homomorphism, giving a decomposition

where λ⁡(ε,a) is a ℤ-linear function.

We will now use the Kempf-Ness theorem and solve μℂ=0 and then divide by G⁢L⁢(k). Given a solution, we want g∈G⁢L⁢(k) such that (g−1⁢B1⁢g,g−1⁢B2⁢g,g−1⁢I,J⁢g) solves μℝ=0. This is not always possible, so we need stable solutions in the sense of geometric invariant theory.

Definition 4.2.1.

The tuple (B1,B2,I,J) is stable if and only if for all subspaces K′⊂ℂk such that B1⁢(K′)⊂K′, B2⁢(K′)⊂K′, and I⁡(ℂN)⊂K′, then K′=K. Equivalently, ℂ⁢⟨B1,B2⟩⁢I⁢(ℂN)=ℂk.

Theorem 4.2.2.

We have an isomorphism

Using this, write ℂN=⨁Nα as the standard decomposition into eigenspaces of the standard maximal torus of U⁡(N). Then let Iα=I⁡(Nα). Our equations then become

Then ℂ⁢⟨B1,B2⟩⁢Iα=Kα is spanned by eigenvectors of gε,a−1 with eigenvalues ei⁢aα⁢ei⁡(p⁢ε1+q⁢ε2) for integers p,q≥0. Note if ψ∈ℂk satisfies gε,a−1⁢ψ=λ⁢ψ, we have

We are still interested in the fate of J, and in fact we will see that J=0, which will imply that [B1,B2]=0.

It is sufficient to replace ℂk by Kα and ℂN by ℂ, so we reduce to the case of N=1. This gives us

so taking the trace gives us Tr⁡I⁢J=0. Therefore, J⁡(I)=0, viewing I∈ℂk. Next, we see that

Similarly, J⁡(B2⁢I)=0. This is also true for any powers of B1,B2, and in fact for any polynomial of B1,B2, so using the fact that I is a cyclic vector for B1,B2, we see that J=0.

We now see that to each Kα, we obtain an ideal ℐα⊂ℂ⁡[x1,x2] to be the annihilator of Iα under the assignment xi↦Bi. Thus we can identify

By a standard argument, the action of ℂ××ℂ× tells us that ℐα must be a monomial ideal, and this corresponds to a partition of kα≔dimKα, or equivalently a Young diagram of size kα. The only thing we know is that

so some of the kα could be zero. We will call the α-th partition λ(α) and its components

We have therefore proven the following theorem:

Theorem 4.2.3.

There is a bijection

Example 4.2.4.

When k=1, we have

and the fixed points are given by {0}×(0,…,1,…,0).

4.3 Explicit localization computation

In order to compute our integrals using equivariant localization, we need to compute

as a representation of the maximal torus 𝕋. Its equivariant Chern character (also the character of the representation) is given by

where hα=e−1⁢aα and qi=e−1⁢εi. Then any

must satisfy

up to

Therefore, Tλ¯⁢ℳ¯kframed⁢(N) is the H1 of a complex

where the two arrows are given by

Lemma 4.3.1.

We have H0⁢(𝒯)=H2⁢(𝒯)=0.

Proof.

Suppose γ∈ker⁡d1. Then

But this space is all of ℂk, so γ=0.

Now suppose there exists γ~∈End⁡ℂk such that

for all (β1,β2,i,j). But this implies that

which is exactly d1(γ~)=0, which implies γ~=0. □

To compute the character, we note that

We also know that

where

This also gives us

We are now able to compute

It is also easy to see that

and

so

This is not obviously a pure character, but taking its dual gives us

Rewriting

and noting that

we can compute

We then see that

Taking only the nonpositive part, we obtain

Putting it all together, we obtain

We are now able to compute the integrals we wanted as

We now consider the tautological complex

which should be thought of as a holomorphic square root of the tangent complex. This is given by the formulae

In fact, the ADHM equation is equivalent to this being a complex for all z∈ℂ2. We see what happens when z=0. In this case, we see that if κ~∈coker⁡δ2, then

for all (κ1,κ2,ν), so it annihilates all of B1,B2,I, and thus κ~=0. Therefore

However, in general, this complex has both H0 and H1. We will compute these over λ¯. We can explicitly see that

is the space of corners. We will compute the character of the H1 using the equation

If we note that

we see that

is the character of the ideal corresponding to λ(α). Therefore, the coefficient of wα is simply

and the first term yields Ch⁡H1 while the second term yields Ch⁡H0.

To see how this works in practice, fix N=1. Then ℐ is a codimension k ideal in ℂ⁡[z1,z2]. We are looking for

and this is given by

4.4 q⁢q-characters

Notation 4.4.1.

From now on, we will denote K=ℂk and N=ℂn.

The main tool for dealing with instanton integrals (equivalently, sums over ensembles of Young diagrams) is the q⁢q-character. Here, we will have a discrete set S, and the densities μs∈ℂ for s∈S must satisfy

For example, we can consider s=(λ(1),…,λ(N)). It is important to note that μs depends on

but we can make choices such that all μs are positive real nnumbers, giving us a probability measure.

Definition 4.4.2.

A function 𝒪:S→ℂ is called an observable if

Before we continue, we will consider a tautological bundle over instanton moduli spaces, which we will call K. In fact, it literally corresponds to the ℂk in the construction of the Nakajima variety. Then we are interested in quantities

where

We can see that

where

normalizes the sum ∑sμs=1 and nΞ⁢(k) si the degree of Ξ|s.

Examples 4.4.3.
  1. 1.

    In pure super-Yang-Mills theory, Ξ=1 and Z is homogeneous in (a1,…,aN,ε1,ε2,Λ);

  2. 2.

    In super-QCD, we set

    where c denotes the homogeneized Chern polynomial. Here, the input is given by

    where the mf are the “flavors” and q takes the role of Λ previously.1

  3. 3.

    The 𝒩=2∗ theory, or adjoint super-QCD, is defined by

Theorem 4.4.4.

The Chern polynomials at the fixed points are given by

Examples 4.4.5.

For example, in super-QCD, we obtain

This series has a finite radius of convergence and is well-defined for |q|<1. It also admits a non-trivial analytic continuation.

In the 𝒩=2∗ theory, we obtain

Remark 4.4.6.

We should view q as lying on ℳ¯0,4, and the points 0,1,∞ become places where the curve degenerates.

Instead of the tautological bundle, we may also consider the tautological complex

which defines some K-theory class. Here, we set K to live in degree −1. For any K-theory class

we may define its total Chern class by

Definition 4.4.7.

The Y⁡(x)-observable is defined by

We may evaluate it at a general fixed point to obtain

where c□=ε1⁢(i−1)+ε2⁢(j−1). We can then compute the expected value

which is a meromorphic function of x with poles at

where (i,j) lie in a finite set with cardinality at most k at order qk and satisfy 1≤i,j≤k.

Recall the gamma function

which is defined for Rez>0 and has an analytic continuation which is meromorphic with first order poles at the nonpositive integers. We will need the generalization

Before this, recall that z⁢Γ⁢(z)=Γ⁡(z+1) and Γ⁡(1)=1. Similarly, if we set

this satisfies the functional equation

The Barnes 2-gamma function will be defined the functional equation

Definition 4.4.8.

The fundamental q⁢q-character for A1-theory is defined by

where

Definition 4.4.9.

The A1 q⁢q-character is defined by the formula

where

Remark 4.4.10.

This is the beginning of quantum field theory, where we should think about this as an operator product corresponding to

when we make the ωi collide.

We would also like to discuss the expected value more precisely.

Definition 4.4.11.

The expectation value of the Y⁡(x)-observable is

Using equivariant localization, we see that

To speed up our calculations, we will express Y⁡(x)|λ¯ as a plethystic exponential

where

In addition,

where

is the character of K at λ¯. Using the plethystic exponential, we can write

where

Theorem 4.4.12.

The expected value

is an entire function of x.

Sketch of proof for fundamental q⁢q-character.

We need to check pole cancellation for the expression

Note that x is a pole if there exists α∈{1,…,N} and p,q∈ℤ>0 such that

and (p,q)∈∂−λ(α). Therefore, we have

Considering the other term, we need to compute the residue

The poles of this expression require (p,q)∈∂+λ(α). All of the combinatorics works out, so it remains to check that the measures work out, or more precisely that

Using the plethystic exponential and setting ξ=eaα+c□, we can write

We are now able to evaluate the residue. We see that

The desired result follows from more computation, namely that

and that

In the Â0-theory, the partition function is defined to be

where we magically have new parameters ε3 and ε4=−(ε1+ε2+ε3). In this case, the measure is given by

This is invariant under swapping ε3,ε4 and ε1,ε2, which corresponds to transposing λ¯. In this case, the fundamental q⁢q-character is given by

Previously, we were integrating over generalisations of H⁢i⁢l⁢bn⁢ℂ2. Today we will generalise even further, and integrate over products of such:

Definition 4.4.13.

We will call spaces like the above moduli spaces of quiver instantons.

4.5 The theory for general quivers

Definition 4.5.1.

Let Γ=(VΓ,EΓ,s,t) be a quiver, wherein VΓ is its set of vertices, EΓ its set of edges, and s,t the source, target maps EΓ→VΓ. sending an edge to where it starts or ends.

Definition 4.5.2.

Let Γ a quiver. Let m:VΓ→ℤ≥0 be its flavours, and n:VΓ→ℤ>0 its colors. Then the moduli space of quiver instantons is

Remark 4.5.3.

Each ℂ2 is acted on by a torus ℂ××ℂ×. Further, we can work equivariantly with respect to each G⁢L⁢(n⁡(v)). Define

Remark 4.5.4.

Each ℳ¯k⁡(v)⁢(n⁡(v)) has a tautological bundle, Kv. We could choose a tuple of complex numbers μ(v)→, one for each vertex, and a tuple of complex numbers μ(e)→, one for each edge, and evaluate

Definition 4.5.5.

Let ℋ⁢om⁢(ℱ1,ℱ2) denote the sheaf on

parametrising the data of two points in our moduli space corresponding to two torsion-free sheaves. Then define

wherein π is the projection map forgetting the ℂ⁢ℙ2 factor.

In principle, we could imagine integrating the above integral. However, not every Γ leads to a nice story. One possible restriction would be to hope that our integrals converge. A basic condition for that would be:

that is, the dimension of our moduli space is at least the degree of the chern class of the bundle we’re integrating over. (The tautological bundle term is clear, and

gives a term like

Also note

Remark 4.5.6.

The above restriction is linear in k on both sides. We could for the restriction to be true for every k. That leads to the condition that

Theorem 4.5.7.

Fix Γ. There exists n→,m→ satisfying the above condition if either

  1. 1.

    Γ is an ADE Dynkin diagram as an unoriented graph;

  2. 2.

    Γ is an affine ADE Dynkin diagram as an unoriented graph, m=0, and n is related to the Dynkin indices.

Example 4.5.8.

Choose the quiver with one vertex and ℓ loops. Then the inequality reads

This is satisfied if either

  1. 1.

    ℓ=0, m≥2⁢n (the A1 case);

  2. 2.

    ℓ=1,m=0 (the Â0 case).

In more detail, the solutions are (described here but not drawn due to the note-taker’s inadequacy)

  1. 1.

    Âr: a cyclic graph with r nodes, all with the mark N;

  2. 2.

    D̂r: a linear graph of r−3 nodes, with two additional nodes attached to the first and last nodes. The marks of the linear nodes are all 2⁢N; the marks of the additional nodes are all N.

  3. 3.

    Êr: an exceptional series of graphs.

    1. (a)

      Ê6, the graph with 7 vertices, a linear graph with three nodes with two chains of length two added at the end. Dynkin marks are N,2⁢N,3⁢N on the linear part and then 2⁢N,N on each chain.

    2. (b)

      Ê8, the graph with 9 vertices, namely a linear graph of length 7 with one additional node attached to the sixth node. The marks on the linear component are N,2⁢N,3⁢N,4⁢N,5⁢N,6⁢N,4⁢N,2⁢N and the added node has mark 3⁢N.

    3. (c)

      Ê7, left as an exercise.

Remark 4.5.9.

In physics, this inequality is related to the condition of asymptotic freedom for the gauge theory associated to the quiver. It’s hard to explain this precisely. These integrals approximate the partition function of the gauge theory, which is a path integral over the space of all connections. You could try to define the path integral rigorously by discretising. Asymptotic freedom, in that picture, is an equation relating to when the discretised path integral is a good approximation to our original path integral.

Example 4.5.10.

The standard model could be related to the A3 quiver with m=0 and n=1,2,3 (though we are of course modeling supersymmetric versions of the theory with our integrals.)

Remark 4.5.11.

If you’ve found a solution of this inequality, you could take a limit:

This limit is a finite dimensional shadow of renormalisation group flow. Taking this limit removes one factor from our chern class; essentially it just removes one flavour. This means that, replacing m→m−1, we could find a solution to the inequality for the new m. So it suffices to solve for solutions where the inequality is an equality: together with the trick of ’freezing’ one of the nodes, this gives all solutions. For example, we can get Ar by freezing a node on Âr+1, sending qr+1→0,q0→0.

Remark 4.5.12.

We have parameters

and

Example 4.5.13.

When we seek to get Ar by freezing Âr+1 as above, after a little bit of computation we find that

and

Let’s reiterate the idea of how we study integrals like the above, namely the use of q⁢q-characters. Since there is some ADE structure to this story, it might not be so surprising that what we called q⁢q-characters have something to do with the characters of algebras related to the ADE series. Again ∫ℳ¯Γ⁢(k)n will be related to a combinatorial sum, namely a sum over a doubly-indexed set of Young diagrams λv,α, where v∈VΓ and α∈1,…,n⁡(v).

Definition 4.5.14.

Let

Then start to define

Theorem 4.5.15.

There is a completion of this sum such that <Xv⁢(x)> has no poles in x. This has the same size as the sum defining characters of affine Kac-Moody groups.

Example 4.5.16.

Let’s study the A2 case with marks N,N and two framing nodes N,N. Redefine so μe=0 by rescaling. Then,

Here, the term we needed to add was q1⁢q2⁢P1⁢(x)⁢P2⁢(x)Y2⁢(x), and this exactly cancels all remaining pools. So it stops after 3 terms. Why 3 terms? A2 is related to s⁢l3. This looks like a character: a character of what? The large x asymptotics are like

There’s something called a conformal extension of a simple Lie group G. It has to do with the fact that there’s a centre. For any group with a centre, we can add a torus factor for every cyclic factor in the centre. This relates, for example, SL to GL.

Remark 4.5.17.

The “fundamental” in q⁢q-characters denotes that they’re related to highest weight.

4.6 General q⁢q-characters

Definition 4.6.1.

A general q⁢q-character depends on

For w:VΓ→ℤ≥0. We could try to define a q⁢q-character Xω→⁢(x). It would look like

The point is that it’s not just the product of fundamental characters; there are corrections. You’re trying to encode all orders of expansion of this function.

Once we define a general qq-character, we’ll then be able to set the stage for the limit shape story, that is, Seiberg-Witten geometry.

Remark 4.6.2.

Remember that a quiver is an oriented connected finite graph, possibly with more than one arrow between nodes. Then quiver gauge theory depends on the data of Ni,Mi, vector spaces for each i∈V⁢e⁢r⁢t⁢i⁢c⁢e⁢s⁢(Q).

Recall there is an ADE classification given by imposing the condition that

The N,M are discrete parameters of our theory, but there are also continuous parameters a(i)∈𝖫𝗂𝖾⁡(T)⊂L⁢i⁢e⁢(GL⁡(Ni)), and m(i)∈𝖫𝗂𝖾⁡(T)⊂𝖫𝗂𝖾⁡(GL⁡(Mi)). Further there are qi∈ℂ, complex numbers of small absolute value.

Finally there are parameters me∈ℂ, for e∈E⁡(Γ), and there are two parameters

Definition 4.6.3.

We may then define a partition function

wherein α is the color parameter, and runs from 1 to ni, and

Remark 4.6.4.

There is some redundancy in this description. For μi∈ℂ, i∈VΓ, we may shift

Remark 4.6.5.

There are physics expectations that as ε1,ε2→0, or even just ε2→0, the asymptotics might be computable.

To study all of this, we’ll need another auxiliary object.

Definition 4.6.6.

A Nakajima quiver variety depends on the data of a quiver Q and data defined by tuples of integers v→i∈VΓ, w→i∈VΓ. We want at least one wi>0 and the remaining wi to be positive. To this data, we assign complex vector spaces Vi=ℂvi,Wi=ℂwi. Then the Nakajima quiver variety is defined as

where ⫽⁣⫽ denotes the symplectic quotient. We may estimate its dimension to be 2⁢∑ivi⁢(wi−vi)+2⁢∑i→jvi⁢vj.

Remark 4.6.7.

These spaces carry obvious tautological bundles

and hence a tautological complex Ci:

wherein all the maps are the obvious ones given by the quiver representation. The complex moment map condition is precisely the condition that this is a complex.

Remark 4.6.8.

ℳQ⁢(v→,w→) has a G⁢L⁢(w→)-global symmetry. There is another symmetry factor, namely by permuting edges, and finally a ℂq⋆ which acts by scaling the moment map. There is some redundancy in this description.

Definition 4.6.9.

Let Yi⁢(x) the observable defined by

We will now give a general formula for qq-characters. Let’s fix vectors

which we think of as equivariant parameters for G⁢L⁢(W→), and an equivariant parameter −ε1+ε22 for ℂq×, and define

where everything is taken equivariantly with respect to all symmetries listed above and

Definition 4.6.10.

𝒳u→⁢(x) is a fundamental q⁢q-character.

Proposition 4.6.11.

⟨𝒳u→⁢(x)⟩ has no poles in x.

Example 4.6.12.

Let’s recover the A→1-case we studied earlier. Here, ℳA1⁢(v,w)=T⋆⁢G⁢r⁢(v,w). The symmetry group is PGL⁡(w)×ℂ×. The maximal torus T⊂PGL⁡(w)×ℂ× acts with isolated fixed points, labeled by the v-dimensional planes spanned by the coordinate axes in the basis chosen by our choice of maximal torus. In other words, fixed points are labelled by partitions of {1,…,w}=ℐ⊔𝒥 into two disjoint sets, so that |ℐ|=v.

In terms of the Nakajima data, let I:W→V,J:V→W, and impose the stability condition that J is of maximal rank. Hence it’s the space of operators

So the tangent space is

At the fixed point I=0,J⁢eγ=eiγ, we can identify the tangent space as

Therefore,

where the first summand comes from T⁢G⁢r, and the second from T⋆⁢G⁢r.

So the qq-character takes the form

Of course the canonical complex takes the form

Since I=0 and J is injective, up to equivariant weight, C+=W.V, and Ch⁡(C+)=∑j∈𝒥euj, Ch⁡(C−)=∑i∈ℐeui−ε1−ε2.

Hence the final factor gives an overall answer of

This is the qq character for A1-quiver theory, namely SQCD. It has 2w terms. This formula is already indicative that not all such formulas come as rational expressions in the Y whose arguments are just shifted. If we were to set for example ui=uj, we would have a smaller torus. The fixed point locus would not be isolated but it would still be compact. So this integral is finite. So even though there’s terms like ui−uj, we should expect the limit as these two things approach to exist, but we might get some derivatives of Y appearing.

4.7 Limit shapes

We’ve defined many partition functions

dependent on many parameters. The parameters εi were special, recording the weights of G⁢L⁢(2) symmetry for 𝔤⁡(ℂ⁢ℙ2,ℂ⁢ℙ∞1). These partition functions were essentially sums over infinite sets whose elements are finite subsets of ℂ. So these subsets had the form

Each such set carries a certain complex weight. There is a way to specify all the parameters of the problem such that this weight is real and positive. In that case we could talk about the induced probability distribution. Then we could be interested in the limiting behaviour as ε1,ε2→0. In this probablistic case that would require us to find the limiting measure. We can draw a curve which describes the boundary of these finite sets, and seek its limiting shape as ε1,ε2→0.

Physically, ε1,ε2 effectively confine instantons into a polydisc of volume ≈(ε1⁢ε2)−1. Then taking the limit ε1,ε2→0 is a thermoydynamic limit. We view Z=e⁢x⁢p⁢(ε1−1⁢ε2−1⁢ℱ⁢(a→,m→,q→)+…), and our goal is to extract the function ℱ. The typical size of a partition will be

Remark 4.7.1.

A probabilistic interpretation of Z exists at all because the moduli space of instantons, ℳ¯k⁢(n), is a complex symplectic manifold. So each tangent space is a symplectic vector space, and so the weights of the symmetry group therefore split in pairs. So if w appears in T⁢ℳ¯k⁢(n), so does ε1+ε2−w. That defines a symplectic form on ℳk⁢(n)¯, which roughly descends from that on the space of all connections:

If we specialize ε1+ε2=0, then

Each w is a linear function of the a→,ε→ terms. Choosing ε1∈i⁢ℝ,aα(i)∈i⁢ℝ, then (−w2)∈ℝ+. The numerator is trickier, but can also be made positive real by judicious specialisation of our parameters.

If we stay away from the locus wherein

which is the locus wherein some of the weights could become 0 in the denominator, we won’t have any convergence issues.

Remark 4.7.2.

Remember our

has ρi with compact support near aα(i), and we remember that

Then

All these δ-functions have signs. We should think that the density between these ± will concentrate, so we might thank that in the limit we might get a continuous function which equals zero almost everywhere except at a finite number of intervals where the ± signs concentrate. The question is: can we express μ in terms of this limiting function?

We have

wherein

The advantage of writing it this way is that it’s a quadratic functional, so extremising it is straightforward.

Now let’s extremise μλ subject to the constraints that ∫Ii,αρi⁢(x)⁢𝑑x=1, ∫Ii,αx⁢ρi⁢(x)⁢𝑑x=aα(i). Doing so, we will get linear equations on the ρi. The most concise way to write those linear equations is via

wherein Pi⁢(x)=∏f=1mi(x−μf(i)).

After analytic continuation, the only thing we’ll be able to control about the Yi are their moments

Studying analytic continuation is related to the action of the Weyl group. So to solve these equations, it suffices to find the invariants of the Weyl group action. So the strategy now is to find Weyl invariant functions of the Yi.

4.7.1 Review of Lie theory and Weyl groups

Let’s review some Lie theory so that in the above formula for Y+⁢Y−. You will recognise something of classical mechanics.

Let Q∨=coroot lattice=H⁢o⁢m⁢(ℂ⋆,T) wherein T is our maximal torus. The weights form the dual lattice Λ⊂𝔥⋆ (they are subsets of the dual to the Cartan). It’s formed by all elements λ so that λ⁡(α∨)∈ℤ for every coroot. We let

for t∈T.

Remark 4.7.3.

In the lattice Q∨ of coroots, there is a basis of simple coroots (α1∨,…,αr∨) such that the adjoint action decomposes a⁢dt⁢(eα)=tα⁢eα, wherein eα is a root (it’s a lattice of weights, namely the weights of the adjoint representation). This has a basis of roots, ℤ⁢α1,…,ℤ⁢αr. Finally there is a lattice of coweights,

wherein Z is the centre of G. Also, Z=Weights/Roots.

Definition 4.7.4.

For ξ∈η, define simple reflection by ri⁢(ξ)=ξ−αi⁢(x⁢i)⁢αi∨. It is an interesting theorem that W⁢e⁢y⁢l=𝕎=N⁡(T)/T is the action generated by the ri. This can of course also be defined in a multiplicative form. For t∈T, an element of the maximal torus, we may parametrise t=∏ziαi∨. Applying rj to such an element, the coordinates zi of this torus will transform in the following way:

wherein Ci⁢j is the Cartan matrix. My point is that if you look closely at this formula, you will see that analytic continuation of the Yi across the cuts is essentially the action of the Weyl group by simple reflections.

Remark 4.7.5.

The more precise statement is the following. Let g⁡(x)∈ℂ⁢GΓ, the group associated to the quiver somewhat extended. Specifically ℂ⁢G denotes the conformal extension of our group, which is given by

Essentially, it’s where we replace all cyclic factors in the centre by ℂ× factors. Write g⁡(x)=∏I∈Vertices~ΓYi⁢(x+μi)αi∨⁢Pi⁢(x+μi)−λi∨, where Vertices~ denotes that we need to take the universal cover. The claim is that under analytic continuation

Remark 4.7.6.

I don’t know how well this is known in the case of matrix models, but at lesat for some matrix models, there is a very similar property where this is some combination of resolvents of densities of matrices which, continuing across the cuts, they transform in the vector representation of the Weyl group. Therefore what I’m about to say here also applies there.

Claim 4.7.7.

In any linear representation V, χV⁢(g⁡(x)) is an analytic function of x. With proper normalisation, it is just a polynomial in x.

Claim 4.7.8.

Choose a fundamental representation of G. The equations χi⁢(g)=Ti⁢(x) define curves in (ℂ×)r×ℂ. The tori give coordinates Y1,…,Yr; the last coordinate is x. There is a Weyl group which acts on this vector space, under which the curve is Weyl invariant. This curve is called a Cameral curve. Of course you have heard about spectral curves, which are curves you can assign to a matrix; its eigenvalues form a curve. That’s what you would do in the case of G⁢Ln. But in general these guys are subject to more intricate Weyl symmetries, so it’s not really clear what sort of spectral object to consider. What one can do, for any W′⊂W, one can try to form a smaller curve by taking the quotient.

Example 4.7.9.

Let’s take the Ar quiver, and take the case wherein we have fundamental matter fields at the first and last node and all labels are N. Then g=diag⁡(z1⁢Y1P1,…,zr⁢YrYr−1′,zr+1⁢PrYr)∈G⁢L⁢(r+1). Then invariants are given by

the claim is this is a monic polynomial in x of degree n.

This is an example of a Cameral curve. It maps Cameral→ℂx with degree (r+1)! since any permutation gives another solution of our equations. Now Weyl=S⁡(r+1) acts on this curve, and inside we for example have a subgroup S⁡(r+1)⊃S⁡(r). Now Cameral/S⁡(r)=Spectral is the spectral curve. It keeps track of one eigenvalue. However, for example, we could also consider Cameral/S⁡(r−1) to remember two eigenvalues. This special branched structure of course has to do with Gelfand-Tsetlin patterns and is special to A-type.

Spaces of curves like this describe either monopoles in ℝ2×S1 for ADE type or instantons on ℝ2×T2 for affine ADE type. This curve in question is actually a quasimap of degree n (in affine type at least) from ℙ1 into the moduli spaces of G-bundles on an elliptic curve.

4.8 Use of q⁢q-characters

The main idea is to use the analyticity of ⟨Xw→⁢(x)⟩, together with the fact that for large x the observables behave like polynomials, to conclude ⟨Xw→⁢(x)⟩ is a polynomial in x. Therefore, the moments

are zero for k>0. This will give us a set of equations on the partition function.

Remark 4.8.1.

One can amputate legs with ni=1. Recall that we have a harmonicity condition; if there exists i∈V⁢e⁢c⁢tΓ for which ni=1, then it imposes the stringent requirement that 2=mi+∑e∈s−1⁢(i)nt⁡(e)+∑e∈t−1⁢(i)ns⁡(e). That is, it cannot be connected to ‘too many’ other nodes. There were limited possibilities:

  1. 1.

    A0̂ theory: one node and one loop.

  2. 2.

    Two more complicated graphs separated by a node with marking 1.

Since ni=1, there is only one partition λ(i) at this node. Hence,

Now, as x→∞, to O⁡(x−1), we find

The good thing about the size of the partition is that its expectation value

is simply the logarithmic derivative with respect to fugacity of the partition function.

Example 4.8.2.

Consider A1-theory wherein the framing node has rank 2 and the other node has rank 1. Then

We have another formula

wherein a□=λi−j is arm length, ℓ□=λjt−i is leg length, c□=ε1⁢(i−1)+ε2⁢(j−1) is content. For this theory, we have a unique qq-character, so

since it has no poles.

Now use the computation in the above remark to expand ⟨⋯⟩. Specifically, let’s try to interact the coefficient of x−1. From the first term, this is going to be

From this, we derive a differential equoation for Z:

gotten from

Define

formally. However, we can do better. Take the Gaussian

The main point is that you can use tricks like this to get rid of all degree 1 nodes.

Example 4.8.3.

Consider the example of D4-theory. Write the shape symmetrically, so that there’s one central node with ni=2, three rank one gauge nodes and a framing node also of rank 1 attached. Also give the framing node mass m and give gauge node i the fugacity qi, where 0 is the label of the central node. Then this can be amputated in the sense that

where

(think of q=q0+small correction).

Remark 4.8.4.

This is related to the phenomenon known as bubbling. In an ideal world we would be integrating over actual moduli spaces of instantons, but here we have compactified these moduli spaces by relaxing the condition that they correspond to actual vector bundles. The price we pay is that the rank 1 story is already nontrivial, even though Maxwell theory doesn’t admit instantons on manifolds without holes.

Nonetheless, say that you performed a blowup, luing in a small ℙ1. Then this manifold has nontrivial b2, i.e. it has holes. Then these solve Maxwe’ll’s equations, but from differential geometeric point of view you have a family of metrics that diverges as the exceptional divisor goes to 0. Unfortunately it’s not defined at a point unless you blow up this point.

Remark 4.8.5.

This is somehow related to mirror maps.

Example 4.8.6.

Finally, let’s consider the example of Â0 theory. It’s related to earlier work of Nekrasov-Okounkov. In full generality,

Of course,

Setting μ2=ε32/ε12 and ε1=−ε2, the sum becomes something like

which vanishes if μ=h□. We will prove some formula like

The fundamental qq-character of this theory is

This is a complicated formula, but we don’t want to study the full story; we just want to extract its x−1 coefficient. The result is that

This is already a differential equation for Z.

But the partitions λ(1) describe certain ideals in ℂ⁡[z1,z2], i.e. they parametrise 0-dimensional subschemes in a plane with coordinates ε1,ε2. However, somehow in this story a second plane emerged with equivariant parameters ε3,ε4 whose points were also related to λ. This relation says there is some sort of correspondence between instantons on either plane. This is a transverse intersection of two planes in 4d, and there is no flat deformation of this cross into a smooth surface. So there is something useful coming out of this. This whole structure is called crossed instantons.

We don’t know what crossed instantons are, even though we do know what their moduli space is. Potentially, Z could be singular when ε2/ε1 is a positive rational number or 0 or ∞. However no poles appear in ε3. Moreover, if ε3=−ε1,ε3=−ε2, then this whole thing collapses because if the partition is nontrivial you have boxes for which the armlength is zero. So we may write

where Zregular is regular. We can more naturally write this as an equation of logarithms as

which means that if we define

and note that

the denominator is symmetric under exchanging ε3,ε4. Further in the small q expansion we have no singularities. Further, the differential equation we wrote above implies we can exchange the εi to obtain

This is a meromorphic function on ℙ2 with no singularities, therefore it is constant in the ε. It suffices to evaluate it at any point. We find that it’s =−log⁡ϕ⁡(q), where ϕ is the Euler function. Many years ago, Nekrasov-Okounkov proved it by representing this formula as a trace. That proof was much more complicated than this qq-character method.

Example 4.8.7.

Consider the case of an Ar quiver with all gauge nodes rank 1 and framing nodes of rank 1 at either end. Add Coulomb parameters μ1,μr,a(i) and sometimes write μ1=a0,μr=ar+1 for more symmetric notation. Then,

There are r fundamental q⁢q-characters in this case. The result is that

where pi+=ε1+ε2+ai−ai+1, pi−=ai−ai+1.

Here, we pass to parameters zi so that zi/zi−1=qi. What this formula should make us think about is that a chiral block of a partition function of a Gaussian free field living on a sphere with vertex operators inserted at points zi. This would, for example, be a relevant computation in string theory with strings coming from infinity somehow. The pieces that go off to infinity would be represented by a vertex operator.

Remark 4.8.8.

Studying a Gaussian free field

the Green’s function of the Laplacian, for Σ a plane, is

There’s a small issue since this quantity isn’t dimensionless, which is related to the fact that the constant/zero mode does not enter into the action and so we must make a choice. In principle, we should add

Then we’d like to define vertex operators

which give us

where G is the Green’s function. The analytic piece Ψ will look very much like our expression for Z except Z has two types of momenta, unless ε1=−ε2.

One can slightly modify the Gaussian free field action to include a term

depending on the scalar curvature. Having this scalar term with some coefficient modifies our computations, and can be fine-tuned to exactly reproduce our formula for ZAr.

Remark 4.8.9.

Let us list the topics which we have not covered yet:

  1. 1.

    Spiked (and the special cases of crossed and folded) instanton moduli spaces. Their study goes under the name of gauge origami. It’s a kind of model-building device which builds statistical models of random partitions which can also be interpreted as tools to produce natural observables, e.g. q⁢q-characters, defects, etc.

  2. 2.

    Compactness theorem (hence, regularity of expectation values of observables).

  3. 3.

    N>1 cases where one can write closed (system) of PDEs on the partition function Z, or on some expectation values ⟨⋯⟩.

    It turns out these equations, which often are derived more-or-less combinatorially, are actually the same as the BPZ equations (Belavin-Polyakov-Zamolodchikov) in 2d CFT, or the KZ equations (Knizhnik-Zamolodchikov). This is one of the best incarnations of the somewhat mysterious, but nowadays better understood, correspondence known as the BPS/CFT correspondence.

    Since the world of CFT includes representation of things like Kac-Moody algebras, and the BPS world includes things like Donaldson invariants, this is a very powerful relation.

4.9 Gauge origami

Previously, we explained that the basic example of the q⁢q-character is an observable in the statistical mechanical model wherein we sum over n-tuples of partitions

which represent a subset of ℂ of the form {aα+ϵ1⁢(i−1)+ϵ2⁢(j−1)∣α∈1,…,N;(i,j)∈λ(α)}.

We discovered that a q⁢q-character χ⁡(x), which is an observable for this model, is described by a formula like

wherein the sum labelled a subset of ℂ of the form

However, if we tried to trace the algebrogeometric origin of these two subsets, it was different: the first subset came by doing gauge theory on the plane described by z1⁢z2 whilst the second subset was in fact derived from a partition function on a transversal plane z3⁢z4.

The only connection between these two planes is that ϵ1+ϵ2+ϵ3+ϵ4=0. So there’s a 3d torus acting on a 4d space relating them. We now aim to present a more unified story where we show that the qq-character is the result of projecting a moduli space involving both planes onto one of the planes. Let us proceed to define that. Unfortunately, the story is local, although there are many indications it can be globalised.

4.9.1 Local story

Think about the union of ℂA2=∪A∈σnA⁢ℂA2 wherein A∈{12,13,14,23,24,34}, |A|=6 labels the 6 coordinate hyperplanes, and nA≥0 are the possible integer multiplicities of the hyperplanes.

We view

and act on it with a 3d torus. We seek to understand the nA in terms of certain sheaves. As before, fix K=ℂk,Na=ℂnA. Pictorially we can draw a tetrahedron with vertices 1,2,3,4, so that the edges label pairs of vertices, hence 2d planes. So attach to an edge A the number nA. Within the tetrahedron draw an interior ball labelled by K with arrows IA,JA from and to edges to the ball. We also label endomorphisms of K, Bi.

This is some homogenised version of the ADHM construction, however in the ADHM construction we only have one multiplit=cty so in that setting one doesn’t see the point of all the operators we have introduced above. To start, let’s define an operator for A=a⁢b with a<b:

Then the equations look like

These equations are valued in End⁡(K)⊗ℂ3, but we may also represent that as taking values in six copies of 𝖫𝗂𝖾⁡(U⁡(k))⊗ℝ6 by endowing K with Hermition structure. The ℝ6 here is actually an irreducible representation of S⁢U⁢(4).

Furthermore we will impose a unifying equation:

Why is seven important? It’s 8−1. There are 8 hermition operators acting on K. Modding out by the action of U⁡(K) by

we seek to introduce as many equations as variables.

The intuition is the following: if we draw a ’spiked’ picture where we connect the z12,z23 planes by another plane which bounds them, the I somehow ’injects’ this plane. Since the Ba are roughly operators of multiplicationby the za. You can ask: why are we writing the μi in a sort of strange way, m⁢u12+μ34†=0? There is a vaninishing theorem which says

Theorem 4.9.1 (Vanishing theorem).

The μA=0 all vanish individually, the Ba⁢IA=0 if a∉A, and JA⁢ba=0.

The way this is proven is by squaring the equations. To do everything algebraically, we would have to introduce derived geometry at this point. But the differential geometry is a little easier.

Remark 4.9.2.

This story is a precursor to generalising Donaldson-Thomas theory to CY4s. It has to do with how we really want to integrate the Pfaffian, and we somehow do this by imposing half the equations.

Ultimately, we find that

This is to be conmpared with the ordinary ADHM construction

Remark 4.9.3.

The subspaces Ka⁢b=ℂ⁡[Ba,Bb]⁢Ia⁢b⁢(Na⁢b) can overlap. For example, B3⁢K12=0, and B1⁢K12=0, so if K34∩K12≠∅, the overlap (call it K~12,34) must satisfy B3⁢K~12,34=0, etc.

In a box picture, the corner boxes correspond to ker⁡(B1)∩ker⁡(B2). So elements in K~12,34 must come from corner boxes. The situation relevant for crossed instantons is when n12⁢n34≠0, and rest =0. To take a shortcut, just look at the fixed points of the torus action

which is associated to the box i⁢j∈λ(α). The vectors which conceivably could belong to K12∩K34 here are those annihilated by B1,B2, again the corner boxes.

On the other hand, look at the corner boxes of K34. Generically these don’t intersect, but nongenerically, these cornes can become 1 or become linearly dependent. They will form a ’butterfly configuration’ where some corner boxes overlap. This will happen when we choose the equivariant torus to be a smaller subtorus, and hence have some relation between the equivariant parameters. That is, we have some equation like

which is, notably, one of the conditions for a qq-character to have a pole.

Example 4.9.4.

Set k=1,n12=n,n34=m. Then the Ba∈ℂ; since they are just numbers, they drop out of our equations. Therefore we just have equations

where I12:ℂ2→ℂ,J12:ℂ→ℂn,I34:ℂm→ℂ,J34:ℂ→ℂm. Further there are lots of equations like B4⁢J12=0,…. There is also a stability condition stating that there exist λ,μ∈ℂ so that λ12+μ⁢I34≠0.

Let’s start by assuming that at least one of B1,B2≠0. On this branch, J34,I34=0 by our assumption that one of the Bi doesn’t vanish. Hence the stability condition implies I12≠0. Working this out, we get a copy of T⋆⁢ℙn−1. So one branch of the moduli space is ℂ(B1,B2)2×T⋆⁢ℙn−1. The other branch of the moduli space comes when one of (B3,B4)≠0. On this branch,

Further, there is another branch wherein all the Bi are zero, and we get a copy of T⋆⁢ℙn+m−1 which parametrises relations between the I12,I34.

Remark 4.9.5.

In general, let ℳ¯k⁢(n→) denote the moduli space of spiked instantons which depends on n→=(n12,n13,…).

There is then a function

wherein k→⁢(Bi,IA,JA)=(ka⁢b=dim(Ka⁢b)).

We can also map

Then,

Further,

Then what is the variable x? It has to do with the framing symmetry. We have groups ∏AU⁡(nA) which act only on the (I,J). The diagonal subgroups act trivially except x acts nontrivially. Here h in equivariant parameter.

Remark 4.9.6.

Consider U⁢(1)3⊂U⁡(K) which acts by multiplying the Ba→ei⁢ϕa⁢Ba. Let

Consider the orbifold classification by Γ⊂H. We get answers of type ADE. An interesting case is Γ⊂S⁢U⁢(2)34. To be specific, we need to embed Γ into H by specifying all the decompositions

and similarly for N34, where d⁢i⁢m⁢N⁢12,ω=n→12, for example. Now ℳ¯k is connected, but its fixed point locus might not be. It splits,

which gives us more parameters for our basic counting problem. That is, the generating function will now depend on many additional fugacities in addition to our other parameters:

One can reverse the roles of n12,n34. Then, the same calculation will be computing some integrals over the moduli space of instantons related to the work of Kronheimer-Nakajima.