Free Energy-Entropy Duality

• Free Energy–Entropy Duality is a framework that links thermodynamic free energy and entropy-like quantities via exact variational identities and dissipation laws.
• It employs canonical formulations, Rényi entropy as a finite-difference slope, and convex-analytic tools to reinterpret equilibrium and nonequilibrium behaviors.
• The theory extends to quantum channels and stochastic systems, offering insights into extractable work, relative entropy, and irreversible dissipation.

Searching arXiv for the cited papers to ground the article in current arXiv records. arXiv search query: (Baez, 2011) Free energy–entropy duality denotes a family of exact correspondences, variational identities, and dissipation laws that relate entropy-like quantities to free energy, or free-energy excess to entropy deficits, across equilibrium statistical mechanics, information theory, stochastic thermodynamics, and quantum resource theories. The phrase is not used uniformly in the literature. In the canonical Gibbs setting, the baseline relation is $F=U-TS$, with ordinary entropy obtained from the temperature derivative of $-F$. Baez showed that Rényi entropy is obtained by replacing this derivative with a finite temperature difference quotient, so that entropy of order $q$ is the secant slope of $-F$ under the rescaling $T=T_0/q$ (Baez, 2011). Other formulations identify excess free energy with relative entropy to equilibrium (Gopalkrishnan, 2013, Wilming et al., 2017), interpret entropy and free energy as convex-conjugate potentials linked by the Fenchel–Young equality (Miao et al., 2023), extend the relation to quantum channels (Badhani et al., 14 Oct 2025), or assign a thermodynamic meaning even to reverse relative entropy by passing to a dual ensemble in which the roles of energy and probability are exchanged (Crooks, 17 Feb 2026).

1. Canonical thermodynamic backbone

In the canonical ensemble, the relevant state is the Gibbs distribution or Gibbs density matrix. For a quantum Hamiltonian $H$, the equilibrium state at temperature $T$ is

$$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$

and the Helmholtz free energy is

$F(T)=-kT\ln Z(T).$

In the classical finite-state formulation, one likewise has

$F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$

with $-F$0 the Shannon entropy in nats. In both settings, equilibrium is the free-energy minimizer, and the standard entropy–free-energy relation is recovered as

$-F$1

This derivative identity is the ordinary Shannon/von Neumann case of the broader duality family (Baez, 2011, Gopalkrishnan, 2013).

A recurrent structural point in the literature is that free energy can be read in two equivalent ways. First, it is an energetic balance, $-F$2. Second, once the reference equilibrium state is fixed, the excess free energy above equilibrium becomes an information-theoretic distance. This shifts the emphasis from absolute thermodynamic potentials to nonequilibrium deviation from the Gibbs state. A plausible implication is that much of the modern literature uses “free energy–entropy duality” to name this second viewpoint rather than the textbook identity alone (Gopalkrishnan, 2013, Wilming et al., 2017).

2. Rényi entropy as a free-energy secant slope

A particularly sharp equilibrium formulation is given by the Rényi family. For a probability distribution $-F$3,

$-F$4

and for a density matrix $-F$5,

$-F$6

As $-F$7, these reduce to Shannon and von Neumann entropy. Baez proves that if the system starts in thermal equilibrium at temperature $-F$8, and $-F$9, then

$q$0

This is an exact identity, not a heuristic analogy: Rényi entropy of the initial equilibrium state is the finite-difference quotient of $q$1 between two temperatures (Baez, 2011).

The same result can be rewritten in $q$2-calculus form. If the $q$3-derivative is

$q$4

then

$q$5

Ordinary entropy is recovered in the limit $q$6, when the secant slope becomes the tangent slope. Baez emphasizes this geometrically: ordinary entropy is the slope of the tangent line, whereas Rényi entropy is the slope of the secant line (Baez, 2011).

The physical interpretation is equally explicit. If the bath temperature is suddenly divided by $q$7, so that $q$8, then during re-equilibration the maximum extractable work is

$q$9

and therefore

$-F$0

Thus Rényi entropy measures extractable work per temperature change in a quench protocol. The paper stresses that this applies to both classical and quantum systems and does not require modifying statistical mechanics itself; the deformation is in the entropy–free-energy relation, not in the Gibbs formalism (Baez, 2011).

3. Relative entropy, excess free energy, and informational distinguishability

A second major formulation identifies excess free energy with relative entropy to equilibrium. For a classical finite-state system with Gibbs distribution $-F$1,

$-F$2

where

$-F$3

This is the precise form of the Szilard–Landauer correspondence: information relative to equilibrium is available free energy, up to the factor $-F$4 (Gopalkrishnan, 2013).

The quantum analogue is

$-F$5

with

$-F$6

In the axiomatic resource-theoretic formulation, this identity is decisive because once quantum relative entropy is singled out by continuity, data processing, additivity, and super-additivity, the free-energy difference is likewise singled out as the unique continuous extensive monotone of athermality under the corresponding Gibbs-preserving catalytic operations (Wilming et al., 2017).

The literature has also pushed the asymmetry of relative entropy further. The familiar interpretation applies to $-F$7, but not directly to the reverse divergence. The 2026 dual-ensemble construction shows that

$-F$8

for a nonequilibrium ensemble $-F$9 and its equilibrium reference $T=T_0/q$0, while

$T=T_0/q$1

after constructing a dual pair $T=T_0/q$2 in which the original probabilities define an effective energy spectrum and the original equilibrium probabilities are retained. The reverse divergence is therefore not excess free energy of the original ensemble, but of a thermodynamically dual one (Crooks, 17 Feb 2026).

Before the table below, one caution is essential: these identities all involve a Gibbs reference or a dual Gibbsized construction. They are not generic statements about arbitrary pairs of distributions.

Setting	Entropy-like quantity	Free-energy relation
Canonical nonequilibrium state	$T=T_0/q$3 or $T=T_0/q$4	$T=T_0/q$5 or $T=T_0/q$6
Rényi equilibrium state	$T=T_0/q$7	$T=T_0/q$8
Dual ensemble construction	$T=T_0/q$9	$H$0
Quantum channels	$H$1	$H$2

Taken together, these results suggest that “duality” often means complementarity between a reference-relative entropy and a free-energy excess, with equilibrium acting as the zero of both descriptions (Gopalkrishnan, 2013, Wilming et al., 2017, Badhani et al., 14 Oct 2025, Crooks, 17 Feb 2026).

4. Convex-analytic and variational formulations

Another line of work treats free energy–entropy duality as a theorem of convex analysis. In that framework, one starts with a convex function $H$3 and its Legendre–Fenchel conjugate

$H$4

The pair obeys the Fenchel–Young inequality

$H$5

with equality when $H$6 and $H$7. The thermodynamic interpretation given in “On Thermodynamic Information” is that $H$8 is entropy-like, $H$9 is free-energy-like, the bilinear term is energy-like, and $T$0 is entropy production; equilibrium is precisely the Fenchel–Young equality on a “thermo-doubled” state space (Miao et al., 2023).

The same paper makes the maximum-entropy/minimum-free-energy equivalence explicit. Fixing the dual variable and maximizing $T$1 is the entropy-side problem; fixing the primal variable and minimizing $T$2 is the free-energy-side problem. In the large-deviation example,

$T$3

so relative entropy and log-partition function appear as a convex-conjugate pair (Miao et al., 2023).

A related but distinct formulation appears in “Duality Symmetry, Two Entropy Functions, and an Eigenvalue Problem in Gibbs' Theory”. There the Massieu free entropy $T$4 is Legendre–Fenchel dual to $T$5, and the paper argues that Gibbs entropy and Shannon/Sanov entropy correspond to different laws of large numbers: arithmetic sample means in one case, empirical counting frequencies in the other. This yields a second free-energy-like object

$T$6

with $T$7 interpreted as energy parameters conjugate to empirical frequencies (Commons et al., 2021).

Approximate local versions also exist. In Bethe–Kikuchi theory, generalized belief propagation converges to critical points of local approximations of the free energy $T$8, the Shannon entropy $T$9, and the variational free energy $$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$0. The paper stresses that this local Legendre duality can be degenerate on loopy hypergraphs, so the $$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$1-$$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$2 relation may become many-valued rather than one-to-one (Peltre, 2022).

5. Dynamical dissipation, entropy production, and de Bruijn-type identities

In nonequilibrium dynamics, the duality frequently appears not as a static conjugacy but as a dissipation law. For Tsallis-type stochastic thermodynamics of Markov processes, one has a generalized entropy $$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$3 and generalized free energy $$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$4 linked by

$$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$5

The same irreversible quantity $$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$6 is therefore both the free-energy dissipation rate and the entropy-balance remainder after subtraction of generalized excess heat (Peng et al., 2010).

For exactly solvable Brownian motors and overdamped Brownian particles, the literature develops analogous current-based balances. In multi-bath systems, the isothermal identity $$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$7 is replaced by generalized rate equations such as

$$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$8

or

$$\rho_T=\frac{1}{Z(T)}e^{-H/kT}, \qquad Z(T)=\operatorname{tr}\!\left(e^{-H/kT}\right),$$9

together with

$F(T)=-kT\ln Z(T).$0

These formulas make the same structural point: entropy production measures irreversibility, and free-energy decay is its energetic counterpart once heat flow or mechanical work is accounted for. The papers are careful that, in nonisothermal settings, no single global temperature need exist, so the relevant objects are temperature-weighted entropy rates and generalized free-energy dissipation rates rather than a naive Helmholtz functional (Taye, 2015, Taye, 2016).

A stronger infinite-volume result is the de Bruijn-type identity for reversible continuum birth-and-death dynamics. There the specific relative entropy with respect to a Gibbs measure,

$F(T)=-kT\ln Z(T).$1

acts as the nonequilibrium free energy, and the paper proves

$F(T)=-kT\ln Z(T).$2

Here $F(T)=-kT\ln Z(T).$3 is the specific Fisher information and $F(T)=-kT\ln Z(T).$4 is a local entropy-production density written as a symmetrized relative entropy between a birth-weighted law and a reduced Palm law. In this setting, free-energy loss, Fisher information, and entropy production become exactly equivalent descriptions of reversible relaxation toward Gibbs equilibrium (Jahnel et al., 13 Feb 2026).

6. Process-level, generalized, and domain-specific extensions

The channel-level theory of quantum processes defines free energy itself as distinguishability from thermalization. With the absolutely thermal channel $F(T)=-kT\ln Z(T).$5, channel free energy is

$F(T)=-kT\ln Z(T).$6

The paper derives several channel analogues of free energy–entropy duality. One is the exact complementarity identity

$F(T)=-kT\ln Z(T).$7

Another is the Helmholtz-type bound

$F(T)=-kT\ln Z(T).$8

which becomes equality for replacer channels. Operationally, channel free energy equals maximal extractable work from partial thermalization,

$F(T)=-kT\ln Z(T).$9

and one-shot distillation and formation rates are governed by hypothesis-testing and max-relative entropy. A common misconception is corrected explicitly in that work: the channel setting does not support a universal exact identity $F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$0 for all processes; the exact equality survives only for state-preparation channels, while the general statement is an inequality plus complementary entropy identities (Badhani et al., 14 Oct 2025).

Rényi-generalized quantum thermodynamics preserves the formal shape of the Helmholtz relation. With

$F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$1

the nonequilibrium free energy can be written as equilibrium free energy plus a correction involving traditional or sandwiched Rényi relative entropy, together with $F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$2-terms that vanish as $F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$3. The claim there is “form invariance”: the thermodynamic laws retain their structure when entropy, internal energy, and relative entropy are all deformed consistently (Misra et al., 2015).

The same general vocabulary also appears outside standard thermodynamics. In benchmarked risk-sensitive portfolio control, free energy–entropy duality rewrites the exponential criterion

$F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$4

as a stochastic differential game under an equivalent measure with entropic regularization, producing explicit affine saddle-point controls and a quadratic value function. This suggests that, in that domain, “free energy” is the log-moment-generating functional and “entropy” is the relative-entropy penalty for adverse measure distortion (Lleo et al., 16 Apr 2026, Lleo et al., 18 Jun 2026).

A final caution comes from concrete materials modeling. In skyrmion thermodynamics, the low-temperature picture is indeed one of entropic stabilization: skyrmions can be energetically unfavorable but entropically favored, so $F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$5 becomes negative. Yet the same study shows that at elevated temperature both $F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$6 and $F_{E,T}(p)=\langle E\rangle_p-k_BT\,H(p),$7 can change sign. This suggests that free energy–entropy duality is often structurally exact while its phenomenology remains strongly regime-dependent (Schick et al., 2020).

The literature therefore supports no single universal theorem under the heading “free energy–entropy duality.” Instead, it exhibits a recurrent structure with several exact realizations: entropy as derivative or finite-difference quotient of free energy; excess free energy as relative entropy to equilibrium; free energy and entropy as convex-conjugate potentials; dissipation of free energy as entropy production or Fisher information; and complementarity identities in generalized or process-level theories. The unifying theme is that entropy and free energy are not independent descriptors but paired coordinates of equilibrium, nonequilibrium deviation, and irreversible relaxation (Baez, 2011, Miao et al., 2023, Badhani et al., 14 Oct 2025, Crooks, 17 Feb 2026).