# Effective action

In quantum field theory, the effective action is a modified expression for the action, which takes into account quantum-mechanical corrections, in the following sense:

In classical mechanics, the equations of motion can be derived from the action by the principle of stationary action. This is not the case in quantum mechanics, where the amplitudes of all possible motions are added up in a path integral. However, if the action is replaced by the effective action, the equations of motion for the vacuum expectation values of the fields can be derived from the requirement that the effective action be stationary. For example, a field $\phi$ with a potential $V(\phi)$, at a low temperature, will not settle in a local minimum of $V(\phi)$, but in a local minimum of the effective potential which can be read off from the effective action.

Furthermore, the effective action can be used instead of the action in the calculation of correlation functions, and then only tree diagrams should be taken into account.

## Mathematical details

Everything in the following article also applies to statistical mechanics. However, the signs and factors of i are different in that case.

Given the partition function Z[J] in terms of the source field J, the energy functional is its logarithm.

$E[J] = i\ln Z[J]$

Some physicists use W instead where W = −E. See sign conventions

For systems with two-particle interactions, the above Feynman diagrams arise at first order in the perturbation expansion of both Z and E. The perturbation expansion for Z consists of all diagrams which are closed, while the perturbation expansion for E consists of all diagrams which are both closed and connected.

In multiple areas of mathematics and information theory, including statistical mechanics, one writes the partition function as

$E[J]=-\ln Z[J] \,$

Just as Z is interpreted as the generating functional (aka characteristic function(al)/moment-generating function(al) of the probability distribution function(al) eS[φ]/Z) of the time ordered VEVs/Schwinger function (aka moments) (see path integral formulation), E (a.k.a. the second characteristic function(al)/cumulant-generating function(al)) is the generator of "connected" time ordered VEVs/connected Schwinger functions (i.e. the cumulants) where connected here is interpreted in the sense of the cluster decomposition theorem which means that these functions approach zero at large spacelike separations, or in approximations using Feynman diagrams, connected components of the graph.

$\langle\phi(x_1)\cdots\phi(x_n)\rangle_{con}=(-i)^{n+1}\left.\frac{\delta^n E}{\delta J(x_1)\cdots \delta J(x_n)}\right|_{J=0}$

or

$\langle\phi^{i_1}\cdots\phi^{i_n}\rangle_\text{con}=(-i)^{n+1}E^{,i_1\dots i_n}|_{J=0}$

in the deWitt notation

Then the n-point correlation function is the sum over all the possible partitions of the fields involved in the product into products of connected correlation functions. To clarify with an example,

\begin{align} & {} \quad \langle\phi(x_1)\phi(x_2)\phi(x_3)\rangle\\ &=\langle\phi(x_1)\phi(x_2)\phi(x_3)\rangle_\text{con} +\langle\phi(x_1)\phi(x_2)\rangle_\text{con}\langle\phi(x_3)\rangle_\text{con} +\langle\phi(x_1)\phi(x_3)\rangle_\text{con}\langle\phi(x_2)\rangle_\text{con} \\ &+\langle\phi(x_1)\rangle_\text{con}\langle\phi(x_2)\phi(x_3)\rangle_\text{con} +\langle\phi(x_1)\rangle_{con}\langle\phi(x_2)\rangle_\text{con}\langle\phi(x_3)\rangle_\text{con} \end{align}

Assuming E is a convex functional (which is debatable), the Legendre transformation gives a one-to-one correspondence between the configuration space of all source fields and its dual vector space, the configuration space of all φ fields. If E is not convex, we take the Fenchel conjugate instead. φ here is a classical field and not a quantum field operator.

Slightly out of the usual sign conventions for Legendre transforms, the value

$\phi=-{\delta\over\delta J}E[J]$

or

$\phi^i=-E^{,i} \,$

is associated to J. This agrees with the time ordered VEV <φ>J. The Legendre transform of E is the effective action (this corresponds to the rate function, which is the Fenchel conjugate of the cumulant-generating function, a common construction in statistics; e.g. the Chernoff bound)

$\Gamma[\phi]=-\langle J,\phi\rangle-E[J] \,$

or

$\Gamma[\phi]=-J_i \phi^i - E[J] \,$

where

$\phi=-{\delta\over\delta J}E[J]$

and

$J=-{\delta\over\delta \phi}\Gamma[\phi]$

or

$J_i=-\Gamma_{,i}. \,$

There are some caveats, though, the major one being we don't have a true one-to-one correspondence between the dual configuration spaces.

The Dyson equation relating the full propagator, bare propagator and the 1PI self-energy in the absence of tadpoles

Let us first consider the case without tadpoles, i.e. $\langle \phi \rangle =0$ for J=0. In that case, Γ[0] gives the zero-point energy, the first functional derivative of Γ at φ=0 is zero, the second functional derivative gives the inverse of the full propagator, and the nth functional derivative for n ≥ 3 gives the one particle irreducible correlation functions or 1PI correlation functions. The Dyson equation relates the full propagator, the bare propagator and the 1PI self-energy. The n-point connected functions are given as the sum over all trees with n ≥ 3 1PI's as nodes and full propagators as edges.

But what if we have tadpoles? We can always adjust the source J so that there are no tadpoles, i.e. $\langle \phi \rangle =0$. This corresponds to adding a Feynman rule corresponding to a coupling to the source. For any Feynman diagram, a subtadpole is a subgraph corresponding to a component not connected to any of the external legs which arises after cutting of an edge. Any Feynman diagram with a subtadpole can be evaluated as nonzero, but we can group these diagrams into equivalence classes (two connected diagrams are equivalent if they only vary in their subtadpoles). Therefore, we only need to consider the sum of all connected graphs without subtadpoles. The sum over all graphs in an equivalence class with subtadpoles is zero, since J is adjusted so that $\langle \phi \rangle =0$. Any graph without subtadpoles do not contain any couplings to the source. A Taylor expansion of the effective action about φ = 0 gives the 1PI's corresponding to these value of the source according to the rules of the previous paragraph. So, we compute the 1PI's to get the Taylor series about $\langle \phi \rangle =0$. Then, from the effective action that we get from the Taylor series, we find the value of φ which minimizes the effective action. This gives us the VEV of φ when J = 0. Then, we now perform a Taylor series expansion about this VEV after shifting the field φ to a new field redefinition $\phi'=\phi - \langle \phi \rangle$ (this is the background field method). Now we can compute the n-point correlations about the J = 0 vacuum.

## One loop approximation

The one-loop approximation to the effective action is

$\Gamma[\phi]=S[\phi]+\frac{1}{2}Tr\left[\ln {S^{(2)}[\phi]}\right]+\cdots.$

## References

• J.Goldstone, A.Salam, and S.Weinberg, Phys.Rev.127, 965 (1962)
• G.Jona-Lasinio, Nuovo Cimento 34, 1790 (1964)
• S.Weinberg: The Quantum Theory of Fields, Vol.II, Cambridge University Press 1996
• D.J.Toms: The Schwinger Action Principle and Effective Action, Cambridge University Press 2007