The value at the mode in multivariate $t$ distributions: a curiosity or not?

Christophe Ley, Anouk Neven

Abstract

It is a well-known fact that multivariate Student $t$ distributions converge to multivariate Gaussian distributions as the number of degrees of freedom $\nu$ tends to infinity, irrespective of the dimension $k \ge 1$. In particular, the Student's value at the mode (that is, the normalizing constant obtained by evaluating the density at the center) $c_{\nu,k} = \frac{\Gamma(\frac{\nu+k}{2})}{(\pi\nu)^{k/2} \Gamma(\frac{\nu}{2})}$ converges towards the Gaussian value at the mode $c_k = \frac{1}{(2\pi)^{k/2}}$. In this note, we prove a curious fact: $c_{\nu,k}$ tends monotonically to $c_k$ for each $k$, but the monotonicity changes from increasing in dimension $k=1$ to decreasing in dimensions $k \ge 3$ whilst being constant in dimension $k=2$. A brief discussion raises the question whether this a priori curious finding is a curiosity, in fine.

Introduction

In introductory statistics, the Student $t$ distribution is often introduced as a "heavier-tailed" version of the standard Gaussian distribution. A common intuition, built primarily on one-dimensional visualizations, suggests that because the $t$ distribution has more mass in its tails, it must necessarily be "flatter" or shorter at its center (the mode) to ensure the total probability integrates to one. As the degrees of freedom $\nu$ increase, the distribution converges to the Gaussian, and we expect the peak to rise accordingly.

However, Christophe Ley and Anouk Neven demonstrate that this intuition fails in higher dimensions. While the convergence to the Gaussian is universal, the path the peak takes depends critically on the dimensionality $k$.

The Dimensional Pivot

The core contribution of this work is the formal proof of a dimension-dependent "flip" in the monotonicity of the normalizing constant $c_{\nu,k}$, which represents the density value at the mode.

Figure 2: Density functions of the multivariate $t$ distribution for $k=1, 2, 3, 4$ and varying $\nu$. Note how the peak height relative to the Gaussian (green line) changes behavior as $k$ increases.

As shown in the figure above:

  • $k = 1$ (Univariate): Our standard intuition holds. The $t$ distribution is shorter than the Gaussian and grows taller as $\nu \to \infty$.
  • $k = 2$ (Bivariate): A mathematical "sweet spot." The value at the mode is perfectly constant regardless of the degrees of freedom $\nu$.
  • $k \ge 3$ (Multivariate): The intuition flips. The $t$ distribution starts out taller than the Gaussian at the mode and shrinks toward the Gaussian value as $\nu \to \infty$.

Methods

The authors provide a rigorous analytical treatment of the normalizing constant $c_{\nu,k}$. By examining the derivative of $\log(c_{\nu,k})$ with respect to $\nu$, they utilize the properties of the digamma function $\psi(z) = \frac{d}{dz} \log \Gamma(z)$. Through induction and concavity arguments, they establish that the derivative is positive for $k=1$, zero for $k=2$, and negative for $k \ge 3$. This confirms that the behavior observed in low dimensions is not representative of the general multivariate case.

Key Results and Implications

This finding challenges the simplistic "heavy tails mean flat peaks" heuristic. In high dimensions, the "tails" occupy a vast volume of space. The extra mass required for heavy tails in a $t$ distribution does not necessarily come from lowering the peak; instead, the distribution can be both peakier at the center and heavier in the tails compared to a Gaussian.

Figure 3: Probability mass $P(X \in B(0, 0.1))$ for different dimensions $k$ and degrees of freedom $\nu$.

As seen in the table above, for $k=4$, a $t$ distribution with $\nu=1$ (Cauchy-like) actually concentrates more probability mass in a small ball around the origin than the Gaussian ($\nu=\infty$) does. This confirms that the "peakiness" is not just a point-wise curiosity at the mode but reflects a local concentration of mass.

Conclusion

This "curiosity" highlights the dangers of relying on 1D intuition for multivariate statistics. While the result is mathematically elegant, its practical implications remain an open question. It suggests that robust statistical methods using $t$ distributions might behave unexpectedly in high dimensions, potentially being more sensitive to central data than their Gaussian counterparts. Future research could explore how this phenomenon affects Bayesian priors or the performance of robust estimators in high-dimensional settings.

Summary updated

Abstract

It is a well-known fact that multivariate Student $t$ distributions converge to multivariate Gaussian distributions as the number of degrees of freedom $ν$ tends to infinity, irrespective of the dimension $k\geq1$. In particular, the Student's value at the mode (that is, the normalizing constant obtained by evaluating the density at the center) $c_{ν,k}=\frac{Γ(\frac{ν+k}{2})}{(πν)^{k/2} Γ( \fracν{2})}$ converges towards the Gaussian value at the mode $c_k=\frac{1}{(2π)^{k/2}}$. In this note, we prove a curious fact: $c_{ν,k}$ tends monotonically to $c_k$ for each $k$, but the monotonicity changes from increasing in dimension $k=1$ to decreasing in dimensions $k\geq3$ whilst being constant in dimension $k=2$. A brief discussion raises the question whether this \emph{a priori} curious finding is a curiosity, \emph{in fine}.
BibTeX
@misc{ley2012value,
  author = {Christophe Ley and Anouk Neven},
  title = {{The value at the mode in multivariate $t$ distributions: a curiosity or not?}},
  year = {2012},
  eprint = {1211.1174},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1211.1174},
}
Figure