---
title: Generalized Alpha-Beta Divergence
url: https://www.emergentmind.com/topics/generalized-alpha-beta-divergence
type: topic
---

# Generalized Alpha-Beta Divergence

Generalized Alpha-Beta divergence denotes a class of parameterized divergence constructions that unify or interpolate many standard discrepancy measures, including Kullback–Leibler, Hellinger, Itakura–Saito, squared Euclidean, chi-square-type, Rényi-related, Gamma-related, and log-det divergences. In the literature, the expression is used in several closely related senses: a Tweedie/variance-function framework that generates \(\alpha\)- and \(\beta\)-divergences from a common dual cumulant; the two-parameter Cichocki–Amari Alpha–Beta family and its scale-invariant and logarithmic variants; matrix and operator-valued log-det analogues; and a recent \(\psi\)-generated superfamily called generalized alpha-beta divergence [1209.4280] [1412.7146] [1805.01045] [2507.04637].

## 1. Terminological scope and canonical definitions

A standard discrete formulation writes the Alpha–Beta divergence between positive measures \(P\) and \(Q\) on a finite set \(\mathcal{X}\) as
\[
D_{AB}^{\alpha,\beta}(P,Q)=\sum_{x\in\mathcal{X}} d_{AB}^{(\alpha,\beta)}(P(x),Q(x)),
\]
with pointwise term
\[
d_{AB}^{(\alpha,\beta)}(p, q) =
\frac{-1}{\alpha\beta}\left(
p^{\alpha}q^{\beta}
-
\frac{\alpha\,p^{\alpha+\beta}}{\alpha+\beta}
-
\frac{\beta\,q^{\alpha+\beta}}{\alpha+\beta}
\right),
\qquad
\alpha,\beta\neq 0,\ \alpha+\beta\neq 0,
\]
together with continuous extensions for \(\alpha=0\), \(\beta=0\), \(\alpha=-\beta\), and \(\alpha=\beta=0\) [2310.09129]. This family is the basic two-parameter AB divergence used in generalized divergence theory, in robust NMF, and in exact divergence computation for decomposable graphical models [2112.04583].

A more recent superfamily replaces the raw power terms by a scalar generating function \(\psi\). With
\[
{p}_\alpha=\left(\int_A p^\alpha(a)\,d\mu(a)\right)^{1/\alpha},
\qquad
{p,q}_{\alpha,\beta}=\int_A p^\alpha(a)\,q^\beta(a)\,d\mu(a),
\]
the generalized alpha-beta divergence is
\[
d_{GAB}^{(\alpha,\beta),\psi}(P, Q)
=
\frac{1}{\beta(\alpha+\beta)} \psi\big( {p}_{\alpha+\beta}^{\alpha+\beta} \big)
+
\frac{1}{\alpha(\alpha+\beta)} \psi\big( {q}_{\alpha+\beta}^{\alpha+\beta} \big)
-
\frac{1}{\alpha\beta} \psi\big( {p,q}_{\alpha,\beta} \big),
\]
for \(\alpha,\beta,(\alpha+\beta)\neq 0\), with edge-case extensions defined by limits and derivatives [2507.04637]. This construction is explicitly presented as a superfamily of Hellinger distance, power divergence, density power divergence, logarithmic density power divergence, S-divergence, logarithmic S-divergence, Gamma divergence, and \((\phi,\gamma)\)-divergence [2507.04637].

The validity of this \(\psi\)-generated family is characterized in the log-domain by \(\Psi(x)=\psi(e^x)\). For \(\alpha,\beta\neq 0\) and \(\alpha+\beta\notin\{0,1\}\), the divergence is nonnegative for all dominated sub-probability measures if and only if \(\Psi\) is strictly increasing and convex; on the special line \(\alpha+\beta=1\), the admissibility condition weakens to local monotonicity conditions near \(1\) [2507.04637]. This establishes a general mechanism for constructing new alpha-beta-type divergences from admissible \(\psi\).

## 2. Variance-function and Tweedie construction

A distinct but foundational line derives \(\alpha\)- and \(\beta\)-divergences from exponential dispersion models (EDMs). For scalar, separable divergences \(D(\mathbf{x},\boldsymbol{\mu})=\sum_i d(x_i,\mu_i)\), an EDM has density
\[
p(x|\theta,\varphi)=h(x,\varphi)\exp\left\{\varphi^{-1}\big(\theta x-\psi(\theta)\big)\right\},
\]
mean \(\mu=\psi'(\theta)\), and variance function \(\mathrm{Var}(x)=\varphi\,v(\mu)\). In the Tweedie case, \(v(\mu)=\mu^p\), and the corresponding dual cumulant \(\phi_p\) generates both divergence families [1209.4280].

The Tweedie-generated \(\beta\)-divergence is the Bregman divergence of \(\phi_p\),
\[
d_\beta(x,\mu)
=
\phi(x)-\phi(\mu)-(x-\mu)\phi'(\mu)
=
\frac{x^{2-p}}{(1-p)(2-p)}
-\frac{x\,\mu^{1-p}}{1-p}
+\frac{\mu^{2-p}}{2-p},
\]
with special cases
\[
p=0:\ \tfrac12(x-\mu)^2,\qquad
p=1:\ x\log\frac{x}{\mu}-x+\mu,\qquad
p=2:\ \frac{x}{\mu}-\log\frac{x}{\mu}-1.
\]
The same \(\phi_p\), reparameterized by the ratio \(x/\mu\), yields the \(\alpha\)-divergence as a Csiszár \(f\)-divergence,
\[
d_\alpha(x,\mu)
=
\mu\,\phi_p\!\left(\frac{x}{\mu}\right)
=
\frac{x^{2-p}\mu^{p-1}}{(1-p)(2-p)}
-\frac{x}{1-p}
+\frac{\mu}{2-p},
\]
with notable cases including forward KL at \(p=1\), reverse KL at \(p=2\), and Hellinger at \(p=\tfrac32\) [1209.4280].

The same idea extends beyond Tweedie power laws. For a general EDM variance function \(v(\mu)\), the dual cumulant satisfies \(\phi''(\mu)=1/v(\mu)\), and the generalized beta divergence admits the compact integral form
\[
d_\beta(x,\mu)=\int_\mu^x \frac{x-t}{v(t)}\,dt.
\]
This generalization replaces the power variance function \(\mu^p\) by an arbitrary \(v(\mu)\), thereby extending beta divergence from the Tweedie family to any EDM, with explicit examples for Bernoulli, negative binomial, and hyperbolic secant variance functions [1306.3530]. A plausible implication is that, in this variance-function view, the “beta parameter” is replaced by a functional choice of \(v(\mu)\).

## 3. Special cases, identities, and geometry

Within the Tweedie framework, the same index \(p\) simultaneously specifies the variance function \(v(\mu)=\mu^p\), the corresponding Tweedie law \(Tw_p(\mu,\varphi)\), the \(\beta\)-divergence, and the \(\alpha\)-divergence [1209.4280]. The relation
\[
d_\beta(x,\mu)=\mu^{1-p}\,d_\alpha(x,\mu)
\]
implies that KL divergence at \(p=1\) is the unique divergence that is both an \(\alpha\)- and a \(\beta\)-divergence. Moreover, \(d_{p_1}(x,\mu)=d_{p_2}(\mu,x)\) if and only if \(p_1+p_2=3\), so symmetry occurs at \(p=\tfrac32\), yielding the Hellinger case [1209.4280].

The same family also exhibits scale relations. Tweedie models satisfy
\[
c\,Tw_p(\mu,\varphi)=Tw_p(c\mu,c^{2-p}\varphi),
\]
and the induced divergences obey
\[
d_\beta(cx,c\mu)=c^{2-p}d_\beta(x,\mu),\qquad
d_\alpha(cx,c\mu)=c\,d_\alpha(x,\mu).
\]
These identities explain, for example, the scale behavior associated with the Gamma/Itakura–Saito case \(p=2\) [1209.4280].

The principal scalar specializations can be summarized as follows.

| \(p\) | Distribution | Divergence specializations |
|---|---|---|
| \(0\) | Gaussian | \(\beta\): squared Euclidean; \(\alpha\): Pearson-type \((\tfrac12\chi^2)\) |
| \(1\) | Poisson | \(\alpha=\beta=\) KL |
| \(\tfrac32\) | compound Poisson | \(\alpha\): Hellinger |
| \(2\) | Gamma | \(\beta\): Itakura–Saito; \(\alpha\): reverse KL |
| \(3\) | inverse Gaussian | \(\alpha\): reverse Pearson |

Along the \(\beta\)-line generated by negative Tsallis entropies, the geometric behavior on the probability simplex has recently been quantified by a sharp Pinsker-type inequality. For \(p,q\in\mathrm{relint}(\Delta^K)\),
\[
D_\alpha(p\Vert q)\ge \frac{C_{\alpha,K}}{2}\|p-q\|_1^2,
\]
with explicit optimal constants \(C_{\alpha,K}\): dimension-free \(C_{\alpha,K}=2^{1-\alpha}\) for \(\alpha\le 1\), dimension-dependent \(C_{\alpha,K}\asymp K^{1-\alpha}\) for \(1<\alpha\le 2\), and no global positive constant for \(\alpha>2\) when \(K\ge 3\) [2602.05744]. This identifies a precise metric profile for one canonical \(\beta\)-slice inside the broader alpha-beta landscape.

## 4. Statistical interpretation, estimation, and variational objectives

The Tweedie construction gives \(\beta\)-divergence a direct probabilistic interpretation. For EDMs,
\[
d_\beta(x,\mu)=\frac12 d_\nu(x,\mu)=\varphi\big(L_x(x)-L_x(\mu)\big),
\]
where \(d_\nu\) is the unit deviance and \(L_x(\mu)\) is the log-likelihood evaluated at mean parameter \(\mu\). Hence, for fixed \(\varphi\), minimizing \(\beta\)-divergence is equivalent to maximizing the log-likelihood. The Gaussian least-squares identity at \(p=0\) is therefore generalized to the whole Tweedie class [1209.4280].

For model selection across divergence families, one approach is to place \(\beta\) under an explicit likelihood. The Exponential Divergence with Augmentation (EDA) density
\[
p_{\mathrm{EDA}}(x;\mu,\beta,\phi)
=
\frac{1}{Z(\mu,\beta,\phi)}
\exp\left\{R(x,\beta)-\frac{1}{\phi}D_\beta(x\|\mu)\right\}
\]
exists for all \(\beta\in\mathbb{R}\) and supports maximum-likelihood selection of the best \(\beta\). Alpha-divergence is then selected by the transformation
\[
y_i=\frac{x_i^\alpha}{\alpha^{2\alpha}},
\qquad
m_i=\frac{\mu_i^\alpha}{\alpha^{2\alpha}},
\qquad
\beta=\frac{1}{\alpha}-1,
\qquad
D_\beta(y_i\|m_i)=D_\alpha(x_i\|\mu_i),
\]
which reuses the same likelihood machinery. The same framework extends to \(\gamma\)- and Rényi divergences through scale-normalized reductions to \(\beta\)- or \(\alpha\)-minimization [1406.1385].

A different generalization arises in variational inference through the scale invariant Alpha–Beta divergence
\[
D_{sAB}^{\alpha,\beta}(p\|q)
=
\frac{1}{\beta(\alpha+\beta)}\log\int p^{\alpha+\beta}
+
\frac{1}{\alpha(\alpha+\beta)}\log\int q^{\alpha+\beta}
-
\frac{1}{\alpha\beta}\log\int p^\alpha q^\beta,
\]
with continuous extensions on the parameter boundaries. This objective contains \(D_{\mathrm{KL}}(q\|p)\), \(D_{\mathrm{KL}}(p\|q)\), Rényi divergence on the line \(\alpha+\beta=1\), Gamma divergence on the line \(\alpha=1\), Hellinger-type and chi-square-type cases, and a log-Euclidean limit at \((\alpha,\beta)=(0,0)\) [1805.01045]. In the variational setting, \(\lambda=\alpha+\beta\) governs robustness to outliers, while \(\beta\) controls mass-covering versus mode-seeking behavior [1805.01045].

## 5. Superfamilies and further generalizations

One route to broader alpha-type generalization replaces arithmetic and geometric means by pairs of strictly comparable quasi-arithmetic means. For strictly increasing generators \(f,g\) with \(f\circ g^{-1}\) strictly convex, the quasi-arithmetic \(\alpha\)-divergence is
\[
I_\alpha^{f,g}[p:q]
=
\frac{1}{\alpha(1-\alpha)}
\int\Bigl(M_{1-\alpha}^f(p,q)-M_{1-\alpha}^g(p,q)\Bigr)\,d\mu,
\qquad \alpha\in(0,1),
\]
with limit cases
\[
I_1^{f,g}[p:q]
=
\int\left(
\frac{f(q)-f(p)}{f'(p)}
-
\frac{g(q)-g(p)}{g'(p)}
\right)d\mu,
\qquad
I_0^{f,g}[p:q]=I_1^{f,g}[q:p].
\]
These generalized \(1\)- and \(0\)-divergences decompose as generalized cross-entropies minus entropies and admit conformal Bregman representations through monotone embeddings [2001.09660]. A plausible implication is that mean-comparison geometry supplies an alternative route to alpha-beta-type families without starting from power functions.

The \(\psi\)-generated GAB framework broadens this program further. It contains AB divergence when \(\psi(x)=x\), logarithmic AB/AC divergence when \(\psi(x)=\ln x\), and, through reparameterizations, power divergence, density power divergence, logarithmic density power divergence, S-divergence, logarithmic S-divergence, Gamma divergence, and \((\phi,\gamma)\)-divergence [2507.04637]. The characterization by strict increase and convexity of \(\Psi(x)=\psi(e^x)\) provides a necessary and sufficient criterion for the validity of the generalized divergence away from the special line \(\alpha+\beta=1\) [2507.04637].

Another branch extends AB divergence from non-negative scalars to real and complex vectors. For \(\mathbf{p},\mathbf{q}\in\mathbb{C}^m\),
\[
D^{(\alpha,\beta)}(\mathbf{p},\mathbf{q})
=
D_{AB}^{(\alpha,\beta)}(\|\mathbf{p}\|,\|\mathbf{q}\|)
+
\|\mathbf{p}\|^\alpha \|\mathbf{q}\|^\beta \bigl(1-\cos(\angle_{pq})\bigr),
\]
where the angular term uses the Euclidean angle defined through the Hermitian inner product. This construction particularizes to \(\tfrac12\|\mathbf{p}-\mathbf{q}\|^2\) when \((\alpha,\beta)=(1,1)\), and its weighted form
\[
D_{\mathbf{W}}^{(\alpha,\beta)}(\mathbf{p},\mathbf{q})
=
D^{(\alpha,\beta)}(\mathbf{W}^{1/2}\mathbf{p},\mathbf{W}^{1/2}\mathbf{q})
\]
particularizes to Mahalanobis squared distance at \((1,1)\) [2508.03272]. The associated right-sided centroid has the closed form
\[
\mathbf{c}_\star^{(\alpha,\beta)}
=
M_\alpha(\|\mathbf{x}\|)
\,\exp_{1-\alpha}\big(-\beta\,\xi_\alpha(\mathbf{x})\big)\,
\frac{E[\mathbf{x}^{(\alpha)}]}{\|E[\mathbf{x}^{(\alpha)}]\|},
\]
which makes the roles of \(\alpha\) and \(\beta\) explicit: \(\alpha\) controls weighting and mean type, while \(\beta\) modulates the effect of angular dispersion [2508.03272].

A separate symmetrization line constructs the Alpha-Beta-Symmetric divergence, a symmetric Hilbertian metric on \(\mathbb{R}_+\) and on probability measures that recovers Euclidean, Hellinger, Jeffreys, and symmetrized Itakura–Saito-type cases and induces positive definite kernels for kernel methods [1803.00001].

## 6. Matrix, operator, and structured-distribution formulations

For symmetric positive definite matrices \(A,B\), the Alpha–Beta log-det divergence is
\[
D^{(\alpha,\beta)}_{AB}(A\|B)
=
\frac{1}{\alpha\beta}
\log\det\left(
\frac{\alpha (AB^{-1})^\beta + \beta (AB^{-1})^{-\alpha}}{\alpha+\beta}
\right),
\]
equivalently
\[
D^{(\alpha,\beta)}_{AB}(A\|B)
=
\frac{1}{\alpha\beta}\sum_{i=1}^n
\log\left(
\frac{\alpha \lambda_i^\beta + \beta \lambda_i^{-\alpha}}{\alpha+\beta}
\right),
\]
where \(\lambda_i\) are the eigenvalues of \(AB^{-1}\) [1412.7146]. This family contains, through parameter choices and limits, Stein’s loss, S-divergence/Jensen-Bregman LogDet divergence, Logdet Zero (Bhattacharyya) divergence, and the affine invariant Riemannian metric, with \((\alpha,\beta)=(0,0)\) yielding
\[
D^{(0,0)}_{AB}(A\|B)=\frac12\sum_{i=1}^n \log^2(\lambda_i)
\]
and therefore the squared AIRM [1412.7146].

The same log-det idea extends to positive definite unitized trace class operators on a Hilbert space. The resulting infinite-dimensional Alpha-Beta Log-Det divergences generalize finite-dimensional Alpha-Beta Log-Det divergences, include the infinite-dimensional affine-invariant Riemannian distance and the infinite-dimensional Alpha Log-Det divergences as special cases, and admit closed-form formulas via Gram matrices for covariance operators on an RKHS [1610.08087].

For high-dimensional discrete distributions, alpha-beta divergence can be computed exactly when the distributions are represented as decomposable models, i.e. chordal Markov networks. Exact computation of the joint divergence is tractable in time exponential in the treewidth of the computation graph, and this framework extends to exact marginal and conditional alpha-beta divergence by decomposing marginals through \(N\)-partitions and conditionals through quotient constructions on chordal graphs [2112.04583] [2310.09129]. Applications in the literature include pixelwise analysis of distributional changes in QMNIST and a divergence-based quantification of error behavior in superconducting quantum computers [2310.09129].

Across these formulations, the recurrent theme is unchanged: generalized alpha-beta divergence is a parameterized bridge between Bregman-type, \(f\)-divergence-type, log-det, and power-divergence geometries. What varies is the underlying domain—nonnegative scalars, probability measures, complex vectors, SPD matrices, operators, or graphical models—and the structural object from which the divergence is generated: a Tweedie variance function, a power-law AB kernel, a scale-invariant log-integral functional, a quasi-arithmetic mean pair, or a general \(\psi\)-transform.

Source: https://www.emergentmind.com/topics/generalized-alpha-beta-divergence