A parametric probabilistic model $\mathcal{P} = \{ P_{\boldsymbol \theta}\mid\boldsymbol \theta \in \Theta \}$ is a family of probability distributions that can be specified using a finite number of parameters $\theta$. The parameter space $\Theta \subseteq \mathbb{R}^k$ is the set of all allowable values for the parameter.

Categories of Parametric Models

Exponential Family

A parametric model is called an exponential family, if it has the form $f( \mathbf{x} \mid \boldsymbol \theta ) = c(\boldsymbol \theta) h(\mathbf{x}) \exp \{ \sum_{i=1}^l w_i(\boldsymbol \theta) t_i(\mathbf{x}) \}$, where $c$ and $h$ are nonnegative, and all functions are real-valued. Additionally, if $k < l$, the parametric model is called a curved exponential family; if $k = l$, it is called a full exponential family.

Examples of exponential families:

  1. Discrete distributions: Binomial, Poisson, negative binomial;
  2. Continuous distributions: Normal, Gamma, Beta.

Theorem (Natural Parameterization): For an exponential family, denote natural parameters $\eta_i = w_i(\boldsymbol \theta)$, then the parametric model can be rewritten as $f(\mathbf{x} \mid \boldsymbol \eta) = c^∗(\boldsymbol \eta) h(\mathbf{x}) \exp \{ \sum_{i=1}^l \eta_i t_i(\mathbf{x}) \}$. And the natural parameter space $\{ \boldsymbol \eta \mid \int f(\mathbf{x} \mid \boldsymbol \eta)~\mathrm{d}\mathbf{x} < \infty \}$ is convex.

Location-Scale Family

The location-scale family with standard PDF $f(x)$ is the parametric model $\mathcal{P} = \{ f_{(\mu,\sigma)} = \frac{1}{\sigma} f(\frac{x-\mu}{\sigma}) \mid (\mu,\sigma) \in \mathbb{R}\times\mathbb{R}^+ \}$ with location parameter $\mu$ and scale parameter $\sigma$.

Theorem: $X \sim f_{(\mu,\sigma)} \Leftrightarrow \exists Z \sim f(z)$, s.t. $X = \sigma Z + \mu$.

Wrapped Distributions

A wrapped probability distribution $p_w(\theta) = \sum_{k \in \mathbb{Z}} p(\theta + 2\pi k)$ on a unit n-sphere is the distribution of random variable $Z = \arg e^{i \phi}$, where $\phi \sim p$ with support $\mathbb{R}$.

Parametric Families by Support

  1. Discrete:
    1. finite: Bernoulli, binomial, categorical, multinomial, hypergeometric, Zipf;
    2. semi-infinite: Poisson, geometric, negative binomial;
  2. Continuous (Euclidean):
    1. bounded: uniform, triangular, semicircle, Beta, logit-normal;
    2. semi-infinite: exponential, Erlang, Gamma, inverse Gamma, generalized inverse Gaussian; Pareto, Fréchet, Weibull; Chi-squared, Chi, F; log-normal, log-Cauchy, log-Laplace, log-logistic;
    3. infinite: Gaussian, Cauchy, Student's t, Fisher's z; Laplace, logistic, Gumbel;
  3. Manifold:
    • simplex: Dirichlet;
    • circle: von Mises;
    • sphere: von Mises–Fisher [@Fisher1953], Bingham [@Bingham1974], Kent [@Kent1982];
    • torus: bivariate von Mises [@Mardia1975];
    • Stiefel and Grassmann manifolds: matrix Langevin @Chikuse2003;

Matrix Manifolds

The set $M_{m, n}(\mathbb{R})$ of m-by-n real matrices is a smooth $(m n)$-manifold, because it is an $(m n)$-dimensional real vector space: $M_{m, n}(\mathbb{R}) \cong \mathbb{R}^{mn}$.

The set $M^∗_{m, n}(\mathbb{R})$, or $\mathbb{R}_∗^{mn}$, of m-by-n full-rank real matrices is a smooth $(m n)$-manifold, because it is an open subset of the smooth manifold of m-by-n real matrices: $M^∗_{m, n}(\mathbb{R}) = \{X \in M_{m, n}(\mathbb{R}) : \text{rank}~X = \min\{m, n\}\}$, $M^∗_{m, n}(\mathbb{R}) \in \mathcal{T}_{M^∗_{m, n}(\mathbb{R})}$. In particular, the general linear group $\text{GL}(n, \mathbb{R})$ of nonsingular order-n real matrices is a smooth $n^2$-manifold.

Stiefel manifold $V_{k, n}$ is the space of orthonormal k-frames in the Euclidean n-space: $V_{k, n} = \{X \in M_{n,k}(\mathbb{R}) : X^T X = I\}$, $k \le n$. The Stiefel manifold is a smooth submanifold of the Euclidean space of n-by-k real matrices: $V_{k, n} \subset M_{n,k}(\mathbb{R})$, with dimension $k (2n - k - 1) / 2$. When $k = 1$, the Stiefel manifold coincides with the $(n-1)$-sphere: $V_{1,n} = \mathbb{S}^{n-1}$. When $k = n$, the Stiefel manifold coincides with the orthogonal group: $V_{n,n} = O(n)$.

Grassmann manifold or Grassmannian $G_{k, n}$ is the space of rank-k projection matrices, or equivalently, k-subspaces in the Euclidean n-space: $G_{k, n} = \{\text{Span}(X) : X \in V_{k, n}\}$, or $G_{k, n} = \{X X^T : X \in V_{k, n}\}$. The Grassmann manifold is a compact smooth manifold of dimension $k (n - k)$. When $k = 1$, the Grassmann manifold coincides with the $(n-1)$-dimensional projective space, i.e. the quotient manifold of lines in the Euclidean n-space: $G_{1,n} = \mathbb{RP}^{n-1}$.

The set $\mathcal{M}(k, m \times n)$ of m-by-n real matrices of rank k is an embedded $k (m + n - k)$-submanifold of $M_{m,n}(\mathbb{R})$. Other Riemannian metrics can be induced on the fixed-rank matrix manifold via Riemannian submersion $\pi(M, N) = M N^T$, given Riemannian metrics on the total space [@Absil2014].

The set $S_+(k, n)$ of rank-k order-n positive-semidefinite matrices: $S_+(k, n) = \{V \Lambda V^T : V \in V_{k, n}, \lambda \in \mathbb{R}_+^k, \Lambda = \text{diag}(\lambda)\}$.

Oblique manifold $\text{OB}_{m,n}$ is the set of m-tuples of points on the $(n-1)$-sphere that span the Euclidean n-space: $\text{OB}_{m,n} = \{X \in M^∗_{m, n}: \text{diag}(X X^T) = I\}$, or equivalently $\text{OB}_{m,n} = \{X \in \prod_{i=1}^m \mathbb{S}^{n-1} : \text{Span}~X = \mathbb{R}^n\}$.

Flag manifold $F_K(\mathbb{R}^n)$ is the set of all flags of type $K$ in the Euclidean n-space, i.e. a nested sequence of linear subspaces: given $K = (k_i)_{i=1}^m$, $F_K(\mathbb{R}^n) = \{(S_i)_{i=1}^m : S_i \in G_{k_i,n}, \forall i < j, S_i \subset S_j \}$.

Essential manifold is the set of essential matrices, i.e. the product of a skew-symmetric matrix and a rotation matrix: $E_n = \{\Omega Q : \Omega \in M_n(\mathbb{R}), \Omega = - \Omega^T, Q \in O(n)\}$.

Euclidean group $\text{SE}(3)$ is the Cartesian product of the order-3 special orthogonal group and the Euclidean 3-space: $\text{SE}(3) = \text{SO}(3) \times \mathbb{R}^3$.


🏷 Category=Statistics