Seperating Free Energy into Potential and Kinetic Terms

We coarse-grain a Hamiltonian system H(q→,p→) with the function f∈ℝn→ℝN mapping microstates to macrostates Q→=f(q→). The free energy of a macrostate can be calculated from the restricted partition function.

e−F(Q→,P→)kBT=1hn∫Vdq→∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∏iδ(sci(∂F∂Pi−∑j∂fi∂qj∂H∂pj))e−H(p→,q→)kBT

Here his Planks constant, ci are the unit constant with units of Qi, and s is some the unit of time. They make the equations inside the Dirac delta functions unitless so dimensional analysis still works.

We assume a Hamiltonian in a quadratic form with respect to the momentum.

H(q→,p→)=12p→⊤𝐌(q)p→+U(q)=12∑k∑lMkl(q)pkpl+U(q)

The velocities can be caluclated as below

∂H(q→,p→)∂pj=∑kSjk(q)pk

where

𝐒(q)=12(𝐌⊤(q)+𝐌(q))

is a symmetric matrix. Substituting for the velocities we get the below.

e−F(Q→,P→)kBT=1hn∫Vdq→∏iδ(1ci(Qi−fi(q→))))∫−∞∞dp→∏iδ(sci(∂F∂Pi−∑j∂fi∂qj∑kSjk(q)pk))e−12∑k∑lp→⊤𝐌(q→)p→+U(q→)kBT

The potential energy can be split out of the exponent and since it only depends on position it can be moved out of the momentum integral. And the quadratic part can be switched to use 𝐒 because p→⊤𝐒p→=p→⊤𝐌p→

=1hn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→))))∫−∞∞dp→∏iδ(sci(∂F∂Pi−∑j∂fi∂qj∑kSjk(q)pk))e−β(12p→⊤𝐒(q→)p→)

Let's assume that the free energy is also quadratic, and try to solve for 𝐑(Q→) and V(Q→).

F(Q→,P→)=P→⊤𝐑(Q→)P→+V(Q→)

So we have the below equality.

e−β(P→⊤𝐑(Q→)P→+V(Q→))=1hn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∏iδ(sci(∑lRilPl−∑j∂fi∂qj∑kSjk(q)pk))e−β(12p→⊤𝐒(q→)p→)

Substitute in the Fourier definition for the Dirac delta.

=1hn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∏i1τ∫−∞∞dkeiksci(∑lRilPl−∑j∂fi∂qj∑kSjk(q)pk)e−β(12p→⊤𝐒(q→)p→)

Do change of variables on the ki's to factor out the sci. So now ki has units of sci

=1hn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∏i1τcis∫−∞∞dkeik(∑lRilPl−∑j∂fi∂qj∑kSjk(q)pk)e−β(12p→⊤𝐒(q→)p→)

Expand all the products out. Also let C=∏ici.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∫−∞∞⋯∫−∞∞dk1…dkN∏ieiki(∑lRilPl−∑j∂fi∂qj∑kSjk(q)pk)e−β(12p→⊤𝐒(q→)p→)

A product of exponents becomes an exponent of sums.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∫−∞∞⋯∫−∞∞dk1…dkNei∑iki(∑lRilPl−∑j∂fi∂qj∑kSjk(q)pk)e−β(12p→⊤𝐒(q→)p→)

Combine the exponents.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∫−∞∞⋯∫−∞∞dk1…dkNe−β(12p→⊤𝐒(q→)p→)+i∑iki(∑lRilPl−∑j∂fi∂qj∑kSjk(q)pk)
=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞dp→∫−∞∞⋯∫−∞∞dk1…dkNe−β12p⊤𝐒p+ik⊤(𝐑P−𝐉f𝐒p)

Swap the order of integration.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞⋯∫−∞∞dk1…dkN∫−∞∞dp→e−β12p⊤𝐒p+ik⊤(𝐑P−𝐉f𝐒p)

The part the doesn't depend on the k's can be moved out.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞⋯∫−∞∞dk1…dkNeik⊤𝐑P∫−∞∞dp→e−β12p⊤𝐒p−ik⊤𝐉f𝐒p

Do complete the squares on the exponent

−β12p⊤𝐒p−ik⊤𝐉f𝐒p
=−12p⊤(β𝐒)p+(−i𝐒𝐉f⊤k)⊤p

And assume 𝐒 is positive definite then do complete the squares

=−12(p−i(1β𝐉f⊤k))⊤(β𝐒)(p−i(1β𝐉f⊤k))−121β(𝐉f⊤k)⊤𝐒(𝐉f⊤k)

Then substitute it back in to get

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞⋯∫−∞∞dk1…dkNeik⊤𝐑P∫−∞∞dp→e−12(p−i(1β𝐉f⊤k))⊤(β𝐒)(p−i(1β𝐉f⊤k))−121β(𝐉f⊤k)⊤𝐒(𝐉f⊤k)

The part that doesn't depend on momentum can be moved out.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞⋯∫−∞∞dk1…dkNeik⊤𝐑Pe−121β(𝐉f⊤k)⊤𝐒(𝐉f⊤k)∫−∞∞dp→e−12(p−i(1β𝐉f⊤k))⊤(β𝐒)(p−i(1β𝐉f⊤k))

for an arbitrary multidimensional gaussian and symmetric positive definite matrix A.

∫dx→e12(x−a)⊤𝐌(x−a)=τndet𝐌

You can show it works even if the offset a is an complex vector.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞⋯∫−∞∞dk1…dkNeik⊤𝐑Pe−121β(𝐉f⊤k)⊤𝐒(𝐉f⊤k)1βnτndet𝐒

Now the expression is independent of the fine grain momentum. We need to do complete the squares again.

ik⊤𝐑P−121β(𝐉f⊤k)⊤𝐒(𝐉f⊤k)

Rearrange a bit.

=−121βk⊤(𝐉f𝐒𝐉f⊤)k+i(𝐑P)⊤k

Completing the squares yields the below.

=−121β(k+iβ(𝐉f𝐒𝐉f⊤)−1𝐑P)⊤(𝐉f𝐒𝐉f⊤)(k+iβ(𝐉f𝐒𝐉f⊤)−1𝐑P)−βP𝐑⊤(𝐉f𝐒𝐉f⊤)−1𝐑P

Note this requires that 𝐉f𝐒𝐉f⊤ is positive definite. Substitute it back in.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))∫−∞∞⋯∫−∞∞dk1…dkNe−121β(k+iβ(𝐉f𝐒𝐉f⊤)−1𝐑P)⊤(𝐉f𝐒𝐉f⊤)(k+iβ(𝐉f𝐒𝐉f⊤)−1𝐑P)−βP𝐑⊤(𝐉f𝐒𝐉f⊤)−1𝐑P1βnτndet𝐒
=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))e−βP𝐑⊤(𝐉f𝐒𝐉f⊤)−1𝐑P∫−∞∞⋯∫−∞∞dk1…dkNe−121β(k+iβ(𝐉f𝐒𝐉f⊤)−1𝐑P)⊤(𝐉f𝐒𝐉f⊤)(k+iβ(𝐉f𝐒𝐉f⊤)−1𝐑P)1βnτndet𝐒

Then do the same thing again with the Gaussian.

=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))e−βP⊤𝐑⊤(𝐉f𝐒𝐉f⊤)−1𝐑PβNτNdet(𝐉f𝐒𝐉f⊤)1βnτndet𝐒

This almost gives a solution for seperating the free energy, but not quite there yet.

e−β(P⊤𝐑(Q)P+V(Q))=CτNsNhn∫Vdq→e−βU(q)∏iδ(1ci(Qi−fi(q→)))e−βP⊤𝐑⊤(𝐉f𝐒𝐉f⊤)−1𝐑PβNτNdet(𝐉f𝐒𝐉f⊤)1βnτndet𝐒

This gives a final, exact, expression of (with O being whatever units make the rightmost sqrt dimensionless)

exp⁡(−β(P⊤𝐑(Q)P+V(Q)))=OτNhn∫Vdq→exp⁡(−βU(q)+ln⁡(CsNOβNτNdet(𝐉f𝐒𝐉f⊤)1βnτndet𝐒))exp⁡(−βP⊤𝐑⊤(𝐉f𝐒𝐉f⊤)−1𝐑P)∏iδ(1ci(Qi−fi(q→)))

Seperating with additional asumptions

Let 𝐋=𝐉f(q→)𝐒(q→)𝐉f⊤(q→). If 𝐋 can be written solely as a function of Q→ instead of q→, so it only depends on the coarse grain positions, then we can move it out of the integral. We can do that because it is practically constant in the integral since any time it differs the Dirac delta function removes it's contributions

exp⁡(−β(P⊤𝐑(Q)P+V(Q)))=OτNhnexp⁡(−βP⊤𝐑⊤𝐋−1(Q→)𝐑P)∫Vdq→exp⁡(−βU(q)+ln⁡(CsNOβNτNdet𝐋(Q→)1βnτndet𝐒))∏iδ(1ci(Qi−fi(q→)))

Then a simple solution as 𝐑(Q→)=𝐋(Q→), and in the exponent 𝐋−1(Q→)𝐑(Q→) (or the transpose since L symmetric.

exp⁡(−β(P⊤𝐑(Q)P+V(Q)))=OτNhnexp⁡(−βP⊤𝐑P)∫Vdq→exp⁡(−βU(q)+ln⁡(CsNOβNτNdet𝐋(Q→)1βnτndet𝐒))∏iδ(1ci(Qi−fi(q→)))

And the momentum contribution can be divided out on both sides.

exp⁡(−βV(Q))=OτNhn∫Vdq→exp⁡(−βU(q)+ln⁡(CsNOβNτNdet𝐋(Q→)1βnτndet𝐒))∏iδ(1ci(Qi−fi(q→)))

Then just rearrange and basic algebra.

exp⁡(−βV(Q))=OτNhnCsNβNτNdet𝐋(Q→)∫Vdq→exp⁡(−βU(q)+ln⁡(1O1βnτndet𝐒))∏iδ(1ci(Qi−fi(q→)))
V(Q)=−1β(ln⁡(1τN)+ln⁡(CsNβNτNdet𝐋(Q→))+ln⁡(OτNhn∫Vdq→exp⁡(−βU(q)+ln⁡(1O1βnτndet𝐒))∏iδ(1ci(Qi−fi(q→)))))