Coupled Cluster Theory Through the Lens of Cumulants

Posted on April 14, 2026 in Science

One of the most satisfying connections I've found recently is between coupled cluster (CC) theory and cumulant expansions from probability theory.

The exponential ansatz¶

The CC wave function is written as:

$$|\Psi\rangle = e^{\hat{T}} |\Phi_0\rangle$$

where $\hat{T} = \hat{T}_1 + \hat{T}_2 + \cdots$ is the cluster operator. This exponential structure is not just a convenient parameterization — it's a cumulant-generating construction.

Why this matters¶

In probability theory, cumulants $\kappa_n$ relate to moments through:

$$\ln \mathbb{E}[e^{tX}] = \sum_{n=1}^{\infty} \kappa_n \frac{t^n}{n!}$$

A Gaussian distribution has $\kappa_1 = \mu$, $\kappa_2 = \sigma^2$, and $\kappa_{n \geq 3} = 0$. CCSD — which truncates at $\hat{T}_2$ — is making the same approximation: it assumes the many-body correlation is well described by its first two cumulants.

A diagnostic for reliability¶

This gives us a clear criterion for when CC should work: the CI coefficient distribution should be approximately Gaussian. We can check this with the configuration entropy:

$$S = -\sum_I |c_I|^2 \ln |c_I|^2$$

When $S/S_{\max} \lesssim 0.3$, the distribution is dominated by a few configurations (approximately Gaussian in the cumulant sense), and CC converges rapidly. When it approaches 1, higher cumulants matter and CC methods struggle — the onset of strong correlation.

More on this in a future post, where I'll trace the connection through to Jaynes' maximum entropy framework.