I'm confused whether the Jacobian is needed in a path integral representation of a dynamical system, as I've seen multiple conventions in the existing literature.

For simplicity, let's just consider the following ODE: $$\dot{x} = F(x) + \eta(t),$$ where $F$ is some polynomial of $x$ of degree greater than $2$, and $\eta$ is some Gaussian white noise. Normally, the MSR (Martin-Siggia-Rose) path integral can be written as $$ \int DxD\eta \,\, P(\eta)\delta(\dot{x}-F(x)-\eta) \det\Big( \frac{\delta(\dot{x}-F(x))}{\delta x} \Big) \\ = \int DxDp \, \exp\big( -p^2 + ip(\dot{x}-F(x)) \big) \det\Big( \frac{\delta(\dot{x}-F(x))}{\delta x} \Big), $$ where $p$ is the "momentum" corresponding to the noise, and I'm interpreting the last term (the determinant) as some sort of Jacobian from the measure of $x$ to the measure of $F(x)$ (is this interpretation correct?), which in general does not vanish.

However, in the spin-glass notes by Castellani, the determinant seems to be excluded in his dynamical treatment of the PSM ($p$-spin spherical model) (see equation 93 on page 26): https://arxiv.org/abs/cond-mat/0505032: $$ \int DxD\eta P(\eta) \delta(\dot{x}-F(x)-\eta) ,$$ in which case the Jacobian would have result in the Hessian of the Ising energy function, which is clearly non-trivial if $p\geq 2$. My questions is, why is the Jacobian not considered in the dynamical treatment of the PSM?

  • $\begingroup$ Apparently, the presence/absence of the determinant depends on whether one is doing the Stratonovich/Ito calculus. However, I do not understand mathematically why this is the case, so I'm leaving it as a comment for now. $\endgroup$
    – PeaBrane
    Dec 14 '20 at 23:41


To compute the Jacobian we need to discretise time, $t=k\Delta$ with some small time-step $\Delta$, and integer $k\in [0, M]$, and work with a finite number of random variables $$x_k=x(k \Delta), \eta_k = \int_{k\Delta}^{(k+1)\Delta}\eta(t)\mathrm{d}t \tag{1}\label{discr}.$$ If $\eta(t)$ is Gaussian white noise at temperature T, the $\eta_n$'s are i.i.d. zero-mean Gaussian random variables with $\langle{\eta_k\eta_m\rangle}=2T\Delta\delta_{km}$.

In discretising the Langevin's equation $$ \dot{x} = F(x) + \eta(t) \tag{2}\label{langevin}$$ we have some freedom on when to evaluate $F$. For any $\alpha\in[0,1]$, the process $$ \frac{x_{k+1}-x_k}{\Delta} = (1-\alpha) F(x_{k})+\alpha F(x_{k+1})+\frac{1}{\Delta}\eta_k \tag{3}\label{discrLang}$$ converges to (\ref{langevin}) as $\Delta\to 0$. This is a generalisation of the Itô ($\alpha=0$) and Stratonovich ($\alpha=1/2$) conventions. This corresponds to specifying the same-time correlation $\langle x(t) \eta(t)\rangle$. In general, the same SDE can define different stochastic processes depending on $\alpha$. For example, that is the case in the presence of multiplicative noise. However, for the simpler case of (\ref{langevin}), the resulting process is the same for any $\alpha$. See this paper for a nice discussion of the difference between the two prescriptions, and this for further discussion and some recent developments.

Computing the Jacobian

The probability distribution of the discretised noise ($M$ iid Gaussians) is $$ P(\boldsymbol{\eta}) = \frac{1}{(4\pi T\Delta)^{M/2}} \mathrm{e}^{\frac{1}{4T\Delta}\sum_k \eta_k^2} .$$ We can now use (\ref{discrLang}) to change variables from $\eta$'s to $x$'s,

$$ P(\mathbf{x}) = P(\boldsymbol{\eta})\det \mathbf{J}(\mathbf{x}), \qquad \text{where}\ J_{km}(\mathbf{x})=\frac{\partial \eta_{k-1}}{\partial x_m} .$$ Note that because of how we defined (\ref{discr}) the index of $\eta_k$ runs from 0 to $M-1$, while that of $x_k$ from 1 to $M$ (the initial condition $x_0$ needs to be specified separately). From (\ref{discrLang}), $$ J_{km} = \delta_{k,m}\left[1-\alpha \Delta F'(x_k)\right]+\delta_{k,m+1} \left[-1-(1-\alpha) \Delta F'(x_k) \right].$$

That is, $\mathbf{J}$ is an $M\times M$ matrix with non-zero elements only on the diagonal, and just below it. Its determinant is the product of the elements on the diagonal,

$$ \det\mathbf{J} = \prod_{k=1}^m \left[1-\alpha \Delta F'(x_k)\right] = \mathrm{e}^{\sum_k\log[1-\alpha\Delta F'(x_k)]} = \mathrm{e}^{- \alpha\Delta\sum_k F'(x_k)+\mathcal{O}(\Delta)} .$$

Therefore in the $\Delta\to0$ limit, the MSR action gets, from the Jacobian, a term $$ -\alpha \int_0^t F(x(t'))\mathrm{d}t' .$$

Note that in the Itô convention ($\alpha=0$) $\mathbf{J}$ is the identity matrix, and therefore the term above vanishes, which is why it is usually chosen in this context (as mentioned in the paper by Castellani and Cavagna).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.