Path Integral from Lattice Limit

Classical Lagrangian

Start with the definition of action S across some path x(t) where (x,x˙) is the Lagrangian density.

S(x)=tatb(x(t),dx(t)dt)dt

From the principle of least action the path should be a minimum, so the variational derivative should be zero given an perturbation δx(t) that preserves the same start and end points so δx(ta)=0,δx(tb)=0

limϵ0S(x+ϵδx)S(x)ϵ=0
limϵ01ϵtatb(x(t)+ϵδx,dx(t)dt+ϵdδxdt)(x(t),dx(t)dt)dt=0

Assume we can swap the derivative and integral.

tatblimϵ0(x(t)+ϵδx(t),dx(t)dt+ϵdδx(t)dt)(x(t),dx(t)dt)ϵdt=0

Create a dummy value.

tatblimϵ0(x(t)+ϵδx(t),dx(t)dt+ϵdδx(t)dt)(x(t)+ϵδx(t),dx(t)dt)+(x(t)+ϵδx(t),dx(t)dt)(x(t),dx(t)dt)ϵdt=0
tatblimϵ0(x(t)+ϵδx(t),dx(t)dt+ϵdδx(t)dt)(x(t)+ϵδx(t),dx(t)dt)ϵ+(x(t)+ϵδx(t),dx(t)dt)(x(t),dx(t)dt)ϵdt=0

See these are just definitions for partial derivaitves (gradients).

tatb(x(t),x(t)dt)x˙dδx(t)dt+(x(t),x(t)dt)xδx(t)dt=0

Integrate by parts on left side.

(x(tb),x(tb)dt)x˙δx(tb)(x(ta),x(ta)dt)dx˙δx(ta)+tatb(ddt(x(t),x(t)dt)x˙)δx(t)+(x(t),x(t)dt)xδx(t)dt=0

Remember δx(ta),δx(tb)=0.

tatb((x(t),x(t)dt)x(ddt(x(t),x(t)dt)dx˙))δx(t)dt=0

Since this must be true for arbitrary δx

(x(t),x(t)dt)x(ddt(x(t),x(t)dt)dx˙)=0

which is the Euler Lagrange equation. Next we define a Hamiltonian (x,q)=x˙q(x,x˙) where q=(x,x˙)x˙, usually interpreted as momentum.

(x,q)x=(x,x˙)x=ddt(x,x˙)x˙=q˙

So we get Hamiltons equations.

x˙=(x,q)q
q˙=(x,q)x

Phase Space Path Integral

We postulate that there exists an operator H^ that

x2|eΔtiH^|x1dpτeΔti(px2x1ΔtH(x1+x22,p))+O(Δt2)

So that it becomes exact in the limit that Δt approaches zero. Start with a Taylor series expansion on both sides

x2|(1ΔtiH^+O(Δt2))|x1dpτeip(x2x1)(1ΔtiH(x1+x22,p)+O(Δt2))

And multiply out

δ(x2x1)x2|ΔtiH^|x1+O(Δt2)dpτeip(x2x1)(1ΔtiH(x1+x22,p)+O(Δt2))

split the integral

δ(x2x1)x2|ΔtiH^|x1+O(Δt2)dpτeip(x2x1)dpτeip(x2x1)ΔtiH(x1+x22,p)+O(Δt2)

And simplify

δ(x2x1)x2|ΔtiH^|x1+O(Δt2)δ(x2x1)dpτeip(x2x1)ΔtiH(x1+x22,p)+O(Δt2)

Then we can see that

x2|ΔtiH^|x1dpτeip(x2x1)ΔtiH(x1+x22,p)

And canceling common terms

x2|H^|x1dpτeip(x2x1)H(x1+x22,p)

So in position basis the Hamiltonian operator is defined as

x2|H^|x1dpτeip(x2x1)H(x1+x22,p)

Defining Lattice

In D dimensional spacetime we define each point of a N××N lattice to be indexed by elements of the set

n[N]D

where [N] is the set of natural numbers smaller than N

[N]={x|x<N}

Create a field configuration assigning a real number to each point on the lattice (although this can be trivialy generalized to spinors, vectors, tensors etc). It can be represented by a ND dimensional vector space over the real numbers

([N]D)(ND)

A field configuration gives a real number for each point in the lattice

ϕ(ND)

And the field configuration can be indexed by a point on the lattice

ϕn

Defining Path Integral

So given a classical action defined by it's Lagrangian density

S[ϕ]=Ddx(ϕ(x),0ϕ(x),1ϕ(x),,D1ϕ(x))

We can write it like

S[ϕ]=Ddx(ϕ(x),ϕ(x))

Where (ϕ(x))k=kϕ(x).

To discreetize it to a lattice we define (k[D]).Lk to be the length of the lattice in each direction. Make a discrete version of the action. Defined a discreete derivative where the spacing of the lattice points in each direction are Δk=LkN and a basis vector 𝟏k in each direction so

(ϕn)k=ϕn+𝟏kϕnΔk

Then

S[ϕ]=n[N]D(k[D]Δk)(ϕn+ϕn+𝟏t2,ϕn)

When the indices go off the edge of the lattice it just loops around, called periodic boundary conditions. Here t is a chosen direction (usually time) The path integral is now defined as below. (You can also take the limit as the L's go to infinity so the lattice covers the whole universe).

𝒟ϕeiS[ϕ]:=limN(n[N]Ddϕn)exp(iS[ϕ])

Transfer matrix

So the path integral is defined as the limit of

(n[N]Ddϕn)exp(in[N]D(k[D]Δk)(ϕn+ϕn+𝟏t2,ϕn))

Split the "time" direction off from the others, so n=t,w.

(n[N]Ddϕn)exp(it[N]w[N]D1(k[D]Δk)(ϕn+ϕn+𝟏t2,ϕn))

An exponent of sums becomes a product of exponent

(n[N]Ddϕn)t[N]exp(iw[N]D1(k[D]Δk)(ϕn+ϕn+𝟏t2,ϕn))

Define a transfer matrix 𝐓((ND1))×((ND1)). Define each element of the matrix to be

𝐓(ωa,ωb)=exp(iw[N]D1(k[D]Δk)(ωa,w+ωb,w2,ωb,wωa,wΔt,ωa,w+𝟏kωa,wΔk(k[D]/t)))

for ω(ND1) (intuitively 3d time slices of 4d spacetime). So the whole matrix is defined by one-hot vectors (more generally any orthonormal basis). So the time derivative is now the difference between the two, while the spacial derivatives are the same. Back to the integral its now just equal to

(n[N]Ddϕn)t[N]𝐓(ϕt,ϕt+1)

The elements of the matrix can also be extracted with one-hot vectors where |ω(ND1) is a one-hot vector with a single 1 at ω

𝐓(ωa,ωb)=ωb|𝐓|ωa

so the path integral can now be written as

(n[N]Ddϕn)t[N]ϕt+1|𝐓|ϕt

Move the integrals around

=(wdϕN1,w)(wdϕ0,w)ϕN1,w|𝐓(wdϕN2,w)|ϕN2,wϕN2,w|𝐓(wdϕN3,w)|ϕN3,wϕ2,w|𝐓(wdϕ1,w)|ϕ1,wϕ1,w|𝐓|ϕ0,w

Because it forms an orthonormal basis so all the middle terms disappear.

(wdϕt,w)|ϕt,wϕt,w|=I

the path integral simplifies to

𝒟ϕeiS[ϕ]=limN(wdϕN,w)(wdϕ0,w)ϕN,w|𝐓N|ϕ0,w

But remember we assumed that it was periodic as if it was in a repeating L0××LN1 box so ϕ0,w=ϕN1,w and the integral is actually just over one of them.

=limN(wdϕ0,w)ϕ0,w|𝐓N|ϕ0,w=Tr(𝐓N)

Instead we can fix the endpoints in time to some field configuration. Here Lt=t1t0 and the time direction would be non periodic.

ϕt0ϕt1𝒟ϕeiS[ϕ]=limNϕt1,nw|𝐓N|ϕt0,nw

Or more concicely below.

ω0ω1𝒟ϕeiS[ϕ]=limτlimNω1|𝐓N|ω0

Most of quantum field theory can be derived from this.


Legendre transform

Momentum Eigenstates

We now define another orthogonal basis. Using the previous orthonormal basis of elements like |ω. Each element of the new basis |π is defined

ω|π=exp(i(k[D]/tΔk)w[N]D1ωwπw)

And you can prove they're orthogonal by

πa|πb=𝒟ωπa|ωω|πb
=𝒟ωexp(i(k[D]/tΔk)w[N]D1ωwπa,w)exp(i(k[D]/tΔk)w[N]D1ωwπb,w)
=𝒟ωexp(i(k[D]/tΔk)w[N]D1ωw(πb,wπa,w))

If we expand the 𝒟ω notation

=(w[N]D1dωw)exp(i(k[D]/tΔk)w[N]D1ωw(πb,wπa,w))

We see that these are really just a bunch of (unnormalized) Dirac delta functions.

=(τk[D]/tδk)(ND1)w[N]D1δ(πb,wπa,w)

Because they're unnormalized we define the measure 𝒟π to be

𝒟π=w[N]D1(dπwk[D]/tΔkτ)

So that

𝒟π|ππ|=I

We will mess around with this

ωb|T|ωa

Insert the momentum basis identity

=𝒟πωb|ππ|T|ωa

And from construction of the momentum basis

=𝒟πexp(i(k[D]/tΔk)w[N]D1ωb,wπw)π|T|ωa

and then insert another identity, using position basis this time

=𝒟π𝒟ωexp(i(k[D]/tΔk)w[N]D1ωb,wπw)π|ωω|T|ωa

And again from the construction of the momentum basis, and the exponent is negatived since its conjugated.

=𝒟π𝒟ωexp(i(k[D]/tΔk)w[N]D1ωb,wπw)exp(i(k[D]/tΔk)w[N]D1ωwπw)ω|T|ωa

And then from the definition of the transfer matrix

=𝒟π𝒟ωexp(i(k[D]/tΔk)w[N]D1ωb,wπw)exp(i(k[D]/tΔk)w[N]D1ωwπw)exp(iw[N]D1(k[D]Δk)(ωa,w+ωw2,ωwωa,wΔt,))

And

k[D]Δk=Δtk[D]/tΔk
=𝒟π𝒟ωexp(i(k[D]/tΔk)w[N]D1ωb,wπw)exp(i(k[D]/tΔk)w[N]D1ωwπw)exp(iw[N]D1(Δtk[D]/tΔk)(ωa,w+ωw2,ωwωa,wΔt,))

Rewrite as ωw=ωa,w+(ωwωa,w)

=𝒟π𝒟ωexp(i(k[D]/tΔk)w[N]D1ωb,wπw)exp(i(k[D]/tΔk)w[N]D1(ωa,w+(ωwωa,w))πw)exp(iw[N]D1(Δtk[D]/tΔk)(ωa,w+ωw2,ωwωa,wΔt,))

Then just move it over

=𝒟π𝒟ωexp(i(k[D]/tΔk)w[N]D1ωb,wπw)exp(i(k[D]/tΔk)w[N]D1ωa,wπw)exp(iw[N]D1(Δtk[D]/tΔk)((ωa,w+ωw2,ωwωa,wΔt,)ωwωa,wΔtπw))

Move the terms that don't depend in front

=𝒟πexp(i(k[D]/tΔk)w[N]D1ωb,wπw)exp(i(k[D]/tΔk)w[N]D1ωa,wπw)𝒟ωexp(iw[N]D1(Δtk[D]/tΔk)((ωa,w+ωw2,ωwωa,wΔt,)ωwωa,wΔtπw))

So now we have

=𝒟πωb|ππ|ωa𝒟ωexp(iw[N]D1(Δtk[D]/tΔk)((ωa,w+ωw2,ωwωa,wΔt,)ωwωa,wΔtπw))

(TODO make this more rigorous) In the limit that N so that Δt0 the integral on the right will converge to just the stationary points where the derivative with respect to the components of ω are zero. If the derivative isn't zero then the "swirling" of the exponent gets infinitely fast and cancells out. So the derivative of the exponent is

ωww[N]D1(Δtk[D]/tΔk)((ωa,w+ωw2,ωwωa,wΔt,)ωwωa,wΔtπw)
=w[N]D1(Δtk[D]/tΔk)(ωw12+ω˙w1Δt1Δtπw)
=w[N]D1(k[D]/tΔk)(Δtωw12+ω˙wπw)

And in the limit the term goes to zero

=w[N]D1(k[D]/tΔk)(δδω˙wπw)

So for the derivative to be zero, this restricts ω to the field configurations where

δ(ωw+ωa,w2,ωwωa,wΔt,)δω˙wπw=0

Let ωw=ωwωa,w2 and ω˙w=ωwωa,wΔt in the limit as delta t is small. So then

πw=δ(ωw,ω˙w,)δω˙w

If an inverse function exists (this assumption is adding an extra requirement to the Lagrangian) so that

ω˙w=f(ωw,πw)

And now the "velocity" can be written as a function of the "position" and "momentum". The whole equality now simplifies to

ωb|T|ωa=𝒩𝒟πωb|ππ|ωa𝒟ωexp(iw[N]D1(Δtk[D]/tΔk)((ωa,w+ωw2,f(ωa,w+ωw2,πw),)f(ωa,w+ωw2,πw)πw))

(TODO get an expression for it) Here 𝒩 is some constant depending on how flat the stationary point is.

Now we define a Hamiltonian density defined by the Legedre transform of the Lagrangian density

(ωw,πw)=(ωw,f(ωw,πw),)f(ωw,πw)πw

So now its

𝒩𝒟πωb|ππ|ωa𝒟ωexp(iw[N]D1(Δtk[D]/tΔk)((ωa,w+ωw2,πw)))