At the heart of modern information geometry and functional analysis lies a powerful symmetry: the ability to describe a system from two different, yet mathematically equivalent, perspectives. This symmetry is rooted in the Legendre transform, a mathematical operation that maps a convex function to its dual. While traditionally used in thermodynamics to switch between energy and entropy, recent research shows that this duality is the "Rosetta Stone" connecting disparate fields like quantum mechanics, communication theory, and statistical inference.
The Geometry of Duality
In classical geometry, we usually think about distances and angles on a single surface. However, Frank Nielsen explains in An Elementary Introduction to Information Geometry that statistical models are best understood through dualistic structures. Instead of one connection (a rule for moving along a surface), these manifolds possess a pair of conjugate connections that are dual with respect to a metric tensor.
This dualism creates "dually flat" manifolds. In these spaces, even if the manifold itself is curved, we can find two different coordinate systems where the geometry behaves linearly. As Ciaglia et al. argue, this framework is essential for understanding how information flows and transforms across different coordinate systems in both classical and quantum regimes. It allows researchers to treat statistical inference not just as a series of calculations, but as a trajectory through a geometric space.
Bridging Mechanics and Information
The power of this duality is most evident when it bridges the gap between information and physical motion. Leok et al. demonstrate in Connecting Information Geometry and Geometric Mechanics that the divergence functions used to measure "distance" between probability distributions are mathematically equivalent to the discrete Lagrangians used to describe the dynamics of physical systems.
This connection suggests that the way a system processes information is fundamentally linked to its physical constraints. For instance:
- Thermodynamics: Frederic BARBARESCO proposes a framework where thermodynamic properties emerge as geometric features of Lie groups, using Legendre transforms to link characteristic functions to their duals.
- Symplectic Geometry: Frank Nielsen shows in Symplectic Bregman divergences that dually flat manifolds can be analyzed using the tools of Hamiltonian mechanics, providing a new way to study the thermodynamics of machine learning.
- Complex Geometry: Darvas et al. utilize the transform to map geometric quantities in Kähler geometry to simpler problems involving convex functions on a polytope.
Functional Inequalities as Information Limits
Perhaps the most striking application of this duality is in the realm of functional inequalities. The Brascamp-Lieb inequality is a fundamental result in analysis that bounds the integral of a product of functions. While it appears to be a purely analytic tool, Liu et al. have reframed it through the lens of information theory.
In A Forward-Reverse Brascamp-Lieb Inequality, the authors use Legendre-Fenchel duality to establish an "entropic formulation" of these inequalities. They show that these analytic bounds are actually dual representations of relative entropy properties within communication channels. Specifically, forward inequalities correspond to broadcast channels, while reverse inequalities correspond to multiple access channels.
This duality provides a unified framework: instead of proving dozens of specialized inequalities, researchers can derive them as specific instances of a broader functional analytic principle. As Liu et al. further demonstrate, this relationship allows tools from functional analysis to be used to prove "strong converses"—mathematical proofs that communication is impossible beyond certain rates—in common randomness generation.
Deforming the Duality
While the standard Legendre transform is powerful, it is limited to systems that follow classical Shannon entropy. To handle complex systems, researchers have begun "deforming" the transform. Fernando Rosas introduces generalized Legendre transforms by replacing standard convexity with a flexible "link function," allowing for the exploration of non-Shannon entropies and curved statistical manifolds.
However, these generalizations are not as disconnected as they first appear. In Generalized Legendre Transforms Have Roots in Information Geometry, Frank Nielsen reveals that these generalized transforms are equivalent to the standard Legendre transform applied to affine-deformed functions. The deformation operator acts as an involution—applying it twice returns the original function—grounding these abstract mathematical tools in the familiar geometric structures of dually flat manifolds.
By treating functional analysis and information geometry as two facets of a shared dualistic structure, this research direction provides a unified language for understanding how information, geometry, and physical laws interact across scales.
Key Foundational Work
Koszul Information Geometry and Souriau Geometric Temperature/Capacity of Lie Group Thermodynamics (2014) — contributed 1 concept elements to this direction.
Topological-antitopological fusion and the model (1994) — contributed 1 concept elements to this direction.