Moment closure
Moment closure is needed when equations for the mean, variance and other low-order moments depend on higher moments that we do not want—or cannot afford—to track indefinitely.
Why does the problem appear?
For a nonlinear stochastic model, the equation for the first moment may contain the second moment:
\[\frac{d}{dt}E[X_t]\quad\text{depends on}\quad E[X_t^2].\]We then derive an equation for \(E[X_t^2]\), but that may contain \(E[X_t^3]\). The third-moment equation may require the fourth moment, and so on.
This is the moment hierarchy.
A concrete example: stochastic logistic growth
Consider
\[dX_t=rX_t\left(1-\frac{X_t}{K}\right)dt+\sigma X_tdW_t.\]Let
\[m_1=E[X_t],\qquad m_2=E[X_t^2],\qquad m_3=E[X_t^3].\]The mean equation is
\[\boxed{\frac{dm_1}{dt}=rm_1-\frac{r}{K}m_2}.\]So knowing only the mean is not enough.
Applying Itô's lemma to \(X^2\) gives
\[\boxed{\frac{dm_2}{dt}=(2r+\sigma^2)m_2-\frac{2r}{K}m_3}.\]Now the second-moment equation requires \(m_3\). If we derive the third-moment equation, a fourth moment appears.
What does “closing” the system mean?
Suppose we decide to keep only \(m_1\) and \(m_2\). We need some approximation
\[m_3\approx F(m_1,m_2).\]After inserting that approximation, the equations for \(m_1\) and \(m_2\) contain no new unknown moments. The system is then closed.
The simplest closure: ignore variability
If we assume the distribution is concentrated very tightly around its mean, then approximately
\[X\approx E[X].\]This suggests
\[E[X^2]\approx E[X]^2,\qquad E[X^3]\approx E[X]^3.\]Equivalently, this assumes the variance is negligible.
A useful identity before choosing a closure
Write the mean as \(\mu=E[X]\), the variance as \(V=\operatorname{Var}(X)\), and the third central moment as
\[\mu_3=E[(X-\mu)^3].\]Expanding gives
\[\boxed{E[X^3]=\mu^3+3\mu V+\mu_3}.\]This identity shows exactly what information is missing when we know only the mean and variance: the third central moment.
Normal closure
A normal distribution is symmetric, so its third central moment is zero:
\[\mu_3=0.\]Under a normal or Gaussian closure,
\[\boxed{E[X^3]\approx \mu^3+3\mu V}.\]Since
\[V=m_2-m_1^2,\]this can also be written entirely in raw moments:
\[\boxed{m_3\approx3m_1m_2-2m_1^3}.\]What assumption did we actually make?
We did not prove that the biological state is normally distributed. We assumed that its third central moment can be approximated by the Gaussian value, zero.
This distinction is important because biological distributions can be skewed, bounded or concentrated near extinction.
Why closure can change the result
Other closure ideas
There is no single universally correct closure. The approximation should reflect the model and the expected shape of its distribution.
| Closure idea | Main assumption or purpose |
|---|---|
| mean-field / zero-variance | fluctuations around the mean are neglected |
| normal / Gaussian | higher central moments are approximated using Gaussian properties |
| lognormal | useful when positive variables have right-skewed, multiplicative variability |
| distribution-specific | assume a family appropriate to the biological process and express higher moments through its parameters |
Why positivity matters
Population counts and concentrations are non-negative. A normal approximation has support over the whole real line, including negative values.
If the distribution is far from zero and relatively narrow, that may be a harmless approximation. Near extinction or for strongly skewed populations, it may be inappropriate.
Moment closure in epidemic models
Epidemic models contain nonlinear interactions. For example, infection frequently involves a product such as
\[SI.\]Taking expectations gives
\[E[SI],\]not simply
\[E[S]E[I].\]The difference is covariance:
\[\boxed{E[SI]=E[S]E[I]+\operatorname{Cov}(S,I)}.\]A mean-field approximation replaces
\[E[SI]\approx E[S]E[I],\]which is equivalent to neglecting the covariance between susceptible and infectious populations.
Why covariance naturally appears
If an epidemic happens to have more infectious individuals than average, it may simultaneously have fewer susceptible individuals than average. Thus \(S\) and \(I\) need not fluctuate independently.
Ignoring their covariance can therefore alter the expected infection term.
Closure order
A first-order closure keeps means and approximates second or higher moments. A second-order closure keeps means, variances and covariances, then approximates moments of order three and above.
Keeping more moments can retain more information, but it produces a larger system and still requires an assumption at the next level.
How should a closure be chosen?
A useful closure should be guided by the stochastic mechanism, expected distributional shape, population size, proximity to boundaries, and the biological question.
It should also be tested where possible.
Validation against stochastic simulation
One practical approach is to simulate the original stochastic model many times, estimate the mean and variance from those trajectories, and compare them with the closed moment equations.
If the closure agrees over the parameter range of interest, it may provide a fast and interpretable approximation. If it fails, a different closure or direct simulation may be needed.
When closure is especially difficult
Moment closure can become unreliable when distributions are strongly skewed, multimodal, close to absorbing boundaries, or dominated by rare events. Extinction is an important example: a distribution may contain a large probability mass at zero together with a separate distribution of surviving populations.
Moment closure is an approximation, not a new biological law
The original SDE or event-based stochastic model contains the modelling assumptions about the biological process. Closure is a mathematical approximation used to make equations for selected statistics tractable.
Its usefulness depends on whether the discarded higher-order information matters for the question being studied.