Mutual Power-Divergence Information Strictly Fails Subadditivity for Every Exponent Below the Golden Ratio
medium tier2026-10-31OP-726reviewed by Claude Code中文版
Abstract
Does subadditivity of mutual $f$-information across conditionally independent observations hold for power-divergence generators, and what characterizes the exact range of exponents where it is valid? For exponents $\beta \in (0, 1)$, mutual power-divergence information $I_\beta(X; Y)$ generalizes mutual information to the power-divergence family generated by $f_\beta(t) = (t^\beta - \beta t + \beta - 1)/(\beta(\beta - 1))$. In the conditionally independent observation model $Y \leftarrow X \rightarrow Z$, subadditivity requires $I_\beta(X; (Y, Z)) \le I_\beta(X; Y) + I_\beta(X; Z)$. Sufficient conditions for general $f$-divergences are known, but they do not identify the exact range of exponents for the power-divergence family; numerical evidence points to a boundary at the golden ratio conjugate. This post demonstrates that subadditivity strictly fails for every exponent $\beta$ strictly below the golden ratio conjugate $\beta_* = (\sqrt{5} - 1)/2 \approx 0.618034$. Specifically, for every $\beta \in (0, \beta_*)$, there exists an explicit, strictly positive binary Markov chain $Y \leftarrow X \rightarrow Z$ exhibiting strict superadditivity, $I_\beta(X; (Y, Z)) > I_\beta(X; Y) + I_\beta(X; Z)$. The counterexample is constructed via an asymmetric binary channel with a rare input symbol of probability $p$ and crossover probability of order $\kappa p$. The subadditivity slack vanishes at $p = 0$, and its derivative there can be made positive by a choice of the channel parameters exactly when $1 - \beta - \beta^2 > 0$, that is, when $\beta < \beta_*$. These results have been formalized and verified in the Lean 4 proof assistant.
1. Introduction
In information theory, mutual $f$-information $I_f(X; Y)$ generalizes classical mutual information by measuring the divergence between a joint probability distribution $P_{XY}$ and the product of its marginal distributions $P_X \otimes P_Y$ through a convex generator function $f$. A fundamental structural property of classical mutual information is subadditivity under conditional independence: if two observations $Y$ and $Z$ are conditionally independent given an underlying source $X$—forming the Markov chain $Y \leftarrow X \rightarrow Z$—the information provided jointly by $(Y, Z)$ cannot exceed the sum of the informations provided individually, so that $I(X; (Y, Z)) \le I(X; Y) + I(X; Z)$. For general $f$-divergences, however, this subadditivity inequality does not hold universally. Determining the exact class of convex generators $f$ and parameter families for which subadditivity holds across all conditionally independent observations is a question studied in arXiv:2610.21231, which gives sufficient conditions on the generator.
A canonical one-parameter family of generators is the power-divergence family, indexed by an exponent $\beta \in (0, 1)$, given by $$f_\beta(t) = \frac{t^\beta - \beta t + \beta - 1}{\beta(\beta - 1)}.$$ This family encompasses several standard divergence measures, including the divergence generated by $f_{1/2}(t) = 2(\sqrt{t} - 1)^2$ at exponent $\beta = 1/2$. For this family, the mutual power-divergence information between finite random variables $X$ and $Y$ is defined as $I_\beta(X; Y) = D_{f_\beta}(P_{XY} \parallel P_X \otimes P_Y)$. The question posed here, beyond what the paper asks, is whether subadditivity, $$I_\beta(X; (Y, Z)) \le I_\beta(X; Y) + I_\beta(X; Z),$$ holds for all conditionally independent finite random variables $Y \leftarrow X \rightarrow Z$ if and only if $\beta \ge \beta_* = \frac{\sqrt{5}-1}{2} \approx 0.618034$, where $\beta_*$ is the unique positive root of the algebraic equation $\beta^2 + \beta - 1 = 0$.
The work in arXiv:2610.21231 develops differential criteria that establish subadditivity for broad classes of $f$-divergences; it does not treat the power-divergence family exponent by exponent. Before this result, verified refutations of subadditivity for power divergences were available only at isolated exponents, $\beta = 1/4$ and $\beta = 1/2$, and an earlier conjecture recorded here proposed the sufficient range $\beta \ge 2/3$. Whether subadditivity fails everywhere below $\beta_*$ was open.
Detailed numerical searches provided evidence for the critical role of the golden ratio exponent $\beta_*$. In parameterized families of binary channels with rare input symbols of probability $10^{-L}$ and conditionally independent observations with crossover probabilities $a$ and $10^{-L}$, the boundary of subadditivity violations was observed to approach $0.615286$ for $a = 10^{-2}$, $0.618006$ for $a = 10^{-4}$, and $0.618033986$ for $a = 10^{-8}$. Furthermore, high-precision evaluations at $L = 200$ and $a = 10^{-12}$ revealed negative subadditivity slack of $-4.6999 \times 10^{-234}$ at $\beta = 0.6180339887$ and positive slack of $4.7722 \times 10^{-234}$ at $\beta = 0.6180339888$. Nevertheless, numerical evaluations cannot establish the failure of subadditivity across an entire continuum of exponents, nor do they elucidate the mathematical mechanism connecting the golden ratio equation $\beta^2 + \beta - 1 = 0$ to the subadditivity slack.
This post presents the formal verification in Lean 4 of the failure half of the golden ratio classification. We show that for every real exponent $\beta$ in the open interval $(0, (\sqrt{5}-1)/2)$, there exists an explicit, strictly positive binary probability distribution on $X$ and identical transition kernels from $X$ to $Y$ and $Z$ such that $Y \leftarrow X \rightarrow Z$ forms a Markov chain on which mutual power-divergence information strictly fails subadditivity: $$I_\beta(X; (Y, Z)) > I_\beta(X; Y) + I_\beta(X; Z).$$ This establishes that $\beta \ge \beta_*$ is strictly necessary for subadditivity on arbitrary finite alphabets. The proof is constructive: it analyzes an asymmetric binary channel with a rare input symbol of probability $p$ and crossover probability of order $\kappa p$, differentiates the subadditivity slack at $p = 0$, and shows that the derivative is strictly positive for suitable channel parameters whenever $\beta^2 + \beta < 1$.
2. Preliminaries
We formalize probability distributions and mutual power-divergence information on finite alphabets. Throughout, random variables take values in finite types, and probability distributions are represented by their probability mass functions.
Definition 1 (power_divergence_generator): For any parameter $\beta \in \mathbb{R}$ with $\beta \ne 0$ and $\beta \ne 1$, the power-divergence generator $f_\beta : \mathbb{R} \to \mathbb{R}$ is defined by $$f_\beta(x) = \frac{x^\beta - \beta x + (\beta - 1)}{\beta(\beta - 1)},$$ where $x^\beta$ denotes the real power operation. For $x > 0$, $f_\beta(x)$ is twice continuously differentiable with $f_\beta'(x) = \frac{\beta x^{\beta-1} - \beta}{\beta(\beta - 1)} = \frac{x^{\beta-1} - 1}{\beta - 1}$ and $f_\beta''(x) = x^{\beta-2} > 0$, ensuring strict convexity on $(0, \infty)$. At $x = 1$, $f_\beta(1) = 0$ and $f_\beta'(1) = 0$, so $f_\beta$ is normalized. For $x = 0$ and $\beta > 0$, the continuous extension satisfies $f_\beta(0) = \frac{\beta - 1}{\beta(\beta - 1)} = \frac{1}{\beta}$. When $\beta = 1/2$, the generator simplifies to $f_{1/2}(x) = 2(\sqrt{x} - 1)^2$.
Definition 2 (mutual_f_information): Let $X$ and $Y$ be finite types. For any generator function $f : \mathbb{R} \to \mathbb{R}$ and any joint probability mass function $p : X \to Y \to \mathbb{R}$, the mutual $f$-information $\tilde{I}_f(X; Y)$ is defined by $$\tilde{I}_f(X; Y) = \sum_{x \in X} \sum_{y \in Y} p_X(x) p_Y(y) f\left( \frac{p(x, y)}{p_X(x) p_Y(y)} \right),$$ where $p_X(x) = \sum_{y' \in Y} p(x, y')$ and $p_Y(y) = \sum_{x' \in X} p(x', y)$ are the marginal distributions of $X$ and $Y$, respectively.
Remark 1 (Conventions for Marginals and Support): In the Lean definition of mutual_f_information, the summand is formulated as $p_X(x) p_Y(y) f(p(x, y) / (p_X(x) p_Y(y)))$. Under standard real division in Lean, division by zero yields zero. When $p_X(x) p_Y(y) = 0$, the factor $p_X(x) p_Y(y)$ vanishes, which ensures that zero-marginal terms contribute zero to the total sum. In our construction, every distribution and channel transition probability is strictly positive, so $p_X(x) > 0$ and $p_Y(y) > 0$ for all symbols, and division by zero never occurs.
Definition 3 (Markov Chain and Subadditivity): Let $X, Y, Z$ be finite alphabets. A Markov chain $Y \leftarrow X \rightarrow Z$ is specified by an input probability mass function $p_X : X \to \mathbb{R}$ satisfying $p_X(x) \ge 0$ and $\sum_{x \in X} p_X(x) = 1$, and transition kernels $P_Y : X \to Y \to \mathbb{R}$ and $P_Z : X \to Z \to \mathbb{R}$ satisfying $P_Y(x, y) \ge 0$, $\sum_{y \in Y} P_Y(x, y) = 1$, $P_Z(x, z) \ge 0$, and $\sum_{z \in Z} P_Z(x, z) = 1$ for all $x \in X$. The conditional independence of $Y$ and $Z$ given $X$ means that the joint distribution on $X \times (Y \times Z)$ is given by $$p_{XYZ}(x, (y, z)) = p_X(x) P_Y(x, y) P_Z(x, z).$$ The marginal pair distributions are $p_{XY}(x, y) = p_X(x) P_Y(x, y)$ and $p_{XZ}(x, z) = p_X(x) P_Z(x, z)$. Mutual $f$-information is subadditive on this Markov chain if $$\tilde{I}_f(X; (Y, Z)) \le \tilde{I}_f(X; Y) + \tilde{I}_f(X; Z).$$ If the reverse strict inequality holds, the mutual information is strictly superadditive.
Definition 4 (The Golden Ratio Exponent): The golden ratio conjugate exponent $\beta_*$ is defined by $$\beta_* = \frac{\sqrt{5} - 1}{2} \approx 0.618033988749895.$$ It is the unique root of the quadratic equation $\beta^2 + \beta - 1 = 0$ in the interval $(0, 1)$. For any $\beta \in (0, 1)$, the condition $\beta < \beta_*$ is equivalent to $\beta^2 + \beta < 1$, which can also be written as $1 - \beta - \beta^2 > 0$.
3. Main result
Theorem 1 (Failure of Subadditivity Below the Golden Ratio): Let $\beta \in \mathbb{R}$ satisfy $0 < \beta < \frac{\sqrt{5}-1}{2}$. Then there exist strictly positive binary probability mass functions $p_X : \{0, 1\} \to \mathbb{R}$ and transition kernels $P_Y, P_Z : \{0, 1\} \to \{0, 1\} \to \mathbb{R}$ forming a conditionally independent Markov chain $Y \leftarrow X \rightarrow Z$ such that $$\tilde{I}_{f_\beta}(X; Y) + \tilde{I}_{f_\beta}(X; Z) < \tilde{I}_{f_\beta}(X; (Y, Z)),$$ where $f_\beta$ is power_divergence_generator $\beta$, and the joint distributions are defined by $p_{XY}(x, y) = p_X(x) P_Y(x, y)$, $p_{XZ}(x, z) = p_X(x) P_Z(x, z)$, and $p_{XYZ}(x, (y, z)) = p_X(x) P_Y(x, y) P_Z(x, z)$.
In the Lean formalization, this statement is given by power_divergence_exists_binary_counterexample_of_lt_golden_ratio: the theorem takes hypotheses $\beta : \mathbb{R}$, $h\beta_0 : 0 < \beta$, and $h\beta : \beta < (\text{Real.sqrt } 5 - 1) / 2$, and concludes with the existential assertion of $p_X : \text{Fin } 2 \to \mathbb{R}$ and $P_Y, P_Z : \text{Fin } 2 \to \text{Fin } 2 \to \mathbb{R}$ satisfying strict positivity on all symbols, exact normalization $\sum_x p_X(x) = 1$, $\sum_y P_Y(x, y) = 1$, $\sum_z P_Z(x, z) = 1$, and strict superadditivity of mutual_f_information with generator power_divergence_generator $\beta$.
This result adds the following structural insights:
- Universal failure on the open interval: The failure of subadditivity is not restricted to isolated rational points such as $\beta = 1/4$ or $\beta = 1/2$. Rather, strict superadditivity occurs for every exponent in the entire continuous range $(0, \beta_*)$.
- Binary alphabets suffice: The obstruction to subadditivity does not require large alphabets, asymmetric observations, or boundary points of the probability simplex. The counterexample operates entirely within binary alphabets ($|X| = |Y| = |Z| = 2$), uses identical observation channels ($P_Y = P_Z$), and places strictly positive mass on every symbol and transition.
- Necessity of the golden ratio threshold: Because a strictly positive binary counterexample exists for every $\beta < \beta_*$, the condition $\beta \ge \beta_*$ is strictly necessary for subadditivity to hold across all finite alphabets.
theorem power_divergence_exists_binary_counterexample_of_lt_golden_ratio (β : ℝ) (hβ0 : 0 < β) (hβ : β < (Real.sqrt 5 - 1) / 2) : ∃ (pX : Fin 2 → ℝ) (PY PZ : Fin 2 → Fin 2 → ℝ), (∀ x, 0 < pX x) ∧ (∑ x, pX x = 1) ∧ (∀ x y, 0 < PY x y) ∧ (∀ x, ∑ y, PY x y = 1) ∧ (∀ x z, 0 < PZ x z) ∧ (∀ x, ∑ z, PZ x z = 1) ∧ mutual_f_information (power_divergence_generator β) (fun x y => pX x * PY x y) + mutual_f_information (power_divergence_generator β) (fun x z => pX x * PZ x z) < mutual_f_information (power_divergence_generator β) (fun x (yz : Fin 2 × Fin 2) => pX x * PY x yz.1 * PZ x yz.2)
4. Proof
Proof Idea: To test subadditivity, we consider binary random variables $X, Y, Z \in \{0, 1\}$ and set $P_Y = P_Z = W$, where $W : \{0, 1\} \to \{0, 1\} \to \mathbb{R}$ is a transition kernel. By symmetry between $Y$ and $Z$, the marginal mutual informations coincide: $\tilde{I}_{f_\beta}(X; Y) = \tilde{I}_{f_\beta}(X; Z)$. Using the power-sum representation of mutual power-divergence information, the mutual information on a single observation channel can be expressed as $$\tilde{I}_{f_\beta}(X; Y) = \frac{S_Y - 1}{\beta(\beta - 1)},$$ where $S_Y = \sum_{x, y} p_X(x) W(x, y)^\beta p_Y(y)^{1-\beta}$. Similarly, for the pair of conditionally independent observations $(Y, Z)$, the joint transition kernel is $W_2(x, (y, z)) = W(x, y) W(x, z)$, and the mutual information is $$\tilde{I}_{f_\beta}(X; (Y, Z)) = \frac{S_{YZ} - 1}{\beta(\beta - 1)},$$ where $S_{YZ} = \sum_{x, y, z} p_X(x) (W(x, y) W(x, z))^\beta p_{YZ}(y, z)^{1-\beta}$.
The difference between joint and marginal informations is therefore $$\tilde{I}_{f_\beta}(X; (Y, Z)) - 2 \tilde{I}_{f_\beta}(X; Y) = \frac{S_{YZ} - 1}{\beta(\beta - 1)} - 2 \frac{S_Y - 1}{\beta(\beta - 1)} = \frac{2 S_Y - 1 - S_{YZ}}{-\beta(\beta - 1)}.$$ Because $\beta \in (0, 1)$, the denominator factor $-\beta(\beta - 1)$ is strictly positive. Consequently, strict superadditivity $\tilde{I}_{f_\beta}(X; (Y, Z)) > 2 \tilde{I}_{f_\beta}(X; Y)$ is strictly equivalent to the positivity of the slack function: $$G(p) = 2 S_Y - 1 - S_{YZ} > 0.$$
We parameterize the input distribution and the channel as follows: - The input distribution has a rare symbol $x = 1$ of probability $p \in (0, 1/2)$: $p_X = (1 - p, p)$. - The transition kernel $W$ is given by $$W = \begin{pmatrix} 1 - \kappa p & \kappa p \\ 1 - a & a \end{pmatrix},$$ where $a \in (0, 1)$ is a fixed parameter and $\kappa > 0$ is a scaling parameter. At $p = 0$, the input distribution degenerates to $p_X = (1, 0)$, so $X = 0$ almost surely, and the channel output from $x = 0$ is deterministically $0$. Thus $S_Y(0) = 1$ and $S_{YZ}(0) = 1$, giving $G(0) = 2(1) - 1 - 1 = 0$. If we can show that the right derivative $G'(0)$ at $p = 0$ is strictly positive for appropriately chosen $a$ and $\kappa$, then by differentiability $G(p) > 0$ for all sufficiently small $p > 0$.
We state the supporting lemmas:
Lemma 1 (Power-Sum Representation of Mutual Power-Divergence Information): Let $\beta \in (0, 1)$. For any probability mass function $p : X \to Y \to \mathbb{R}$ on finite types with $p(x, y) \ge 0$ and $\sum_{x, y} p(x, y) = 1$, $$\tilde{I}_{f_\beta}(X; Y) = \frac{\sum_{x \in X} \sum_{y \in Y} p(x, y)^\beta (p_X(x) p_Y(y))^{1-\beta} - 1}{\beta(\beta - 1)}.$$ In Lean, this identity is established by mutual_power_divergence_information_eq_sum_rpow. When $p(x, y) = p_X(x) W(x, y)$ for a transition kernel $W$, the identity reduces to $$\tilde{I}_{f_\beta}(X; Y) = \frac{\sum_{x \in X} \sum_{y \in Y} p_X(x) W(x, y)^\beta p_Y(y)^{1-\beta} - 1}{\beta(\beta - 1)}.$$ Proof: Expanding the definition of mutual_f_information with power_divergence_generator, the summand is $$p_X(x) p_Y(y) \frac{(p(x, y)/(p_X(x) p_Y(y)))^\beta - \beta (p(x, y)/(p_X(x) p_Y(y))) + (\beta - 1)}{\beta(\beta - 1)}.$$ Multiplying through by $p_X(x) p_Y(y)$ splits the sum into three terms: 1. $\sum_{x, y} p_X(x) p_Y(y) (p(x, y) / (p_X(x) p_Y(y)))^\beta = \sum_{x, y} p(x, y)^\beta (p_X(x) p_Y(y))^{1-\beta}$. 2. $-\beta \sum_{x, y} p(x, y) = -\beta \cdot 1 = -\beta$. 3. $(\beta - 1) \sum_{x, y} p_X(x) p_Y(y) = (\beta - 1) \cdot 1 = \beta - 1$. The sum of the constants is $-\beta + \beta - 1 = -1$, yielding the formula in mutual_power_divergence_information_eq_sum_rpow.
Lemma 2 (Slack Derivative at $p = 0$): Let $p_X = (1 - p, p)$ and $W = \begin{pmatrix} 1 - \kappa p & \kappa p \\ 1 - a & a \end{pmatrix}$. The slack function $G(p) = 2 S_Y(p) - 1 - S_{YZ}(p)$ satisfies $G(0) = 0$, and its derivative at $p = 0$ is $$G'(0) = L(a, \kappa),$$ where $$L(a, \kappa) = -(1 - \beta) a^2 - (1 - (1 - a)^\beta)^2 + 2 \kappa^\beta \left( (a + \kappa)^{1-\beta} - (\kappa + a - a^2)^{1-\beta} \right).$$ Proof: We evaluate the marginal output probabilities for $Y$ and $(Y, Z)$: For single observation $Y$: $p_Y(1) = (1 - p) \kappa p + p a = p (a + (1 - p) \kappa)$, $p_Y(0) = (1 - p)(1 - \kappa p) + p(1 - a) = 1 - p(a + (1 - p)\kappa)$. For joint observations $(Y, Z) \in \{0, 1\}^2$: $p_{YZ}(1, 1) = (1 - p)(\kappa p)^2 + p a^2 = p (a^2 + (1 - p)\kappa^2 p)$, $p_{YZ}(0, 1) = p_{YZ}(1, 0) = (1 - p)(1 - \kappa p)\kappa p + p(1 - a)a = p ((1 - a)a + (1 - p)(1 - \kappa p)\kappa)$, $p_{YZ}(0, 0) = (1 - p)(1 - \kappa p)^2 + p(1 - a)^2$.
The entries $\kappa p$ of $W$ and the probabilities $p_Y(1)$, $p_{YZ}(0,1)$, $p_{YZ}(1,0)$, $p_{YZ}(1,1)$ all carry a factor $p$. Splitting off this factor with $p^\beta p^{1-\beta} = p$ (for $p > 0$) writes the slack as $$G(p) = 2c_0(p) - 1 - c_{00}(p) + p\,K(p),$$ where $c_0(p) = (1-p)(1-\kappa p)^\beta p_Y(0)^{1-\beta}$ and $c_{00}(p) = (1-p)(1-\kappa p)^{2\beta} p_{YZ}(0,0)^{1-\beta}$ are the terms from the common input symbol and the all-zero output, smooth near $p = 0$ with value $1$ there, and $K$ collects the remaining terms, each a product of powers of $p$ with nonnegative exponents ($p^0$, $p^\beta$, $p^{1-\beta}$) and of powers of quantities that stay positive at $p = 0$; so $K$ is continuous. Hence $G(0) = 0$ and $G'(0) = 2c_0'(0) - c_{00}'(0) + K(0)$, with $c_0'(0) = -1 - \beta\kappa - (1-\beta)(\kappa + a)$, $c_{00}'(0) = -1 - 2\beta\kappa + (1-\beta)(-1 - 2\kappa + (1-a)^2)$ and $$K(0) = 2\left((1-a)^\beta + \kappa^\beta (a+\kappa)^{1-\beta}\right) - (1-a)^{2\beta} - 2\kappa^\beta(\kappa + a - a^2)^{1-\beta}.$$ Collecting terms gives $G'(0) = L(a, \kappa)$.
Lemma 3 (Positivity of the Derivative Below the Golden Ratio): Let $0 < \beta < \frac{\sqrt{5}-1}{2}$. Then there exist parameters $a \in (0, 1)$ and $\kappa > 0$ such that $L(a, \kappa) > 0$. Proof: Two elementary bounds turn $L$ into a rational expression. First, Bernoulli's inequality $(1+s)^\beta \le 1 + \beta s$ with $s = a/(1-a)$ gives $(1-a)^{-\beta} \le 1 + \beta a/(1-a)$, hence $0 \le 1 - (1-a)^\beta \le \beta a/(1-a)$ and $$(1 - (1-a)^\beta)^2 \le \frac{\beta^2 a^2}{(1-a)^2}.$$ Second, with $u = a + \kappa$, Bernoulli's inequality for the exponent $1 - \beta$ gives $(u - a^2)^{1-\beta} \le u^{1-\beta}\left(1 - (1-\beta)a^2/u\right)$, so $$2\kappa^\beta\left(u^{1-\beta} - (u - a^2)^{1-\beta}\right) \ge 2(1-\beta)a^2\left(\frac{\kappa}{a+\kappa}\right)^\beta \ge 2(1-\beta)a^2 \frac{\kappa}{a+\kappa},$$ the last step because $\kappa/(a+\kappa) \le 1$ and $\beta \le 1$. Together, $$L(a, \kappa) \ge a^2\left((1-\beta)\frac{\kappa - a}{\kappa + a} - \frac{\beta^2}{(1-a)^2}\right).$$ Now take $a = (1 - \beta - \beta^2)/4$, which lies in $(0, 1/4)$ because $\beta^2 + \beta < 1$. Then $D = (1-\beta)(1-a)^2 - \beta^2 > 0$: indeed $(1-a)^2 \ge 1 - 2a$ and $(1-\beta)(1-2a) - \beta^2 = (1+\beta)(1-\beta-\beta^2)/2 > 0$. Take $\kappa = a\left((1-\beta)(1-a)^2 + \beta^2\right)/D + 1$. Then $\kappa D = a\left((1-\beta)(1-a)^2 + \beta^2\right) + D$, which says $(1-\beta)(1-a)^2(\kappa - a) - \beta^2(\kappa + a) = D > 0$, and the right-hand side of the bound is positive. So $L(a, \kappa) > 0$.
Lemma 4 (Extraction of Positive Perturbation): Let $G : \mathbb{R} \to \mathbb{R}$ satisfy $G(0) = 0$ and have derivative $G'(0) > 0$. Then for any $\epsilon > 0$, there exists $p \in (0, \epsilon)$ such that $G(p) > 0$. Proof: By definition of the derivative, $G(p)/p \to G'(0) > 0$ as $p \to 0$, so $G(p) > 0$ on a neighborhood $(0, \delta)$. Choosing $p \in (0, \min(\epsilon, \delta))$ gives the result.
Proof of the Main Theorem: We now follow the steps of power_divergence_exists_binary_counterexample_of_lt_golden_ratio:
1. Parameter Selection: Given $0 < \beta < (\sqrt{5}-1)/2$, we have $\beta < 1$. By Lemma 3, there exist parameters $a \in (0, 1)$ and $\kappa > 0$ such that $L(a, \kappa) > 0$.
2. Derivative Evaluation: The derivative of $G(p) = 2 S_Y(p) - 1 - S_{YZ}(p)$ at $p = 0$ equals $L(a, \kappa) > 0$ by Lemma 2. Setting $\epsilon = \frac{1}{2(\kappa + 1)} > 0$, Lemma 4 provides $p > 0$ with $p < \frac{1}{2(\kappa + 1)}$ such that $G(p) > 0$.
3. Verification of Strictly Positive Probability Distributions: We verify that $p$ and $W$ define valid, strictly positive distributions on $\{0, 1\}$:
- Since $\kappa > 0$, $2(\kappa + 1) > 2$, so $p < 1 / (2(\kappa + 1)) < 1/2$. Thus $p_X(0) = 1 - p > 1/2 > 0$ and $p_X(1) = p > 0$, with sum $p_X(0) + p_X(1) = 1$.
- Since $p < \frac{1}{2(\kappa + 1)}$, we have $2(\kappa + 1) p < 1$, so $2\kappa p < 1$, which gives $\kappa p < 1/2$. Thus $W(0, 0) = 1 - \kappa p > 1/2 > 0$ and $W(0, 1) = \kappa p > 0$, with row sum $W(0, 0) + W(0, 1) = 1$.
- For the second row, $a \in (0, 1)$ implies $W(1, 0) = 1 - a > 0$ and $W(1, 1) = a > 0$, with row sum $W(1, 0) + W(1, 1) = 1$.
We set $P_Y = W$ and $P_Z = W$. The joint transition distribution for $(Y, Z)$ given $X$ is $W_2(x, (y, z)) = W(x, y) W(x, z)$, which is non-negative and satisfies $\sum_{y, z} W(x, y) W(x, z) = 1$.
4. Superadditivity: By Lemma 1, the mutual informations satisfy $$\tilde{I}_{f_\beta}(X; Y) = \frac{S_Y - 1}{\beta(\beta - 1)}, \quad \tilde{I}_{f_\beta}(X; Z) = \frac{S_Y - 1}{\beta(\beta - 1)},$$ $$\tilde{I}_{f_\beta}(X; (Y, Z)) = \frac{S_{YZ} - 1}{\beta(\beta - 1)}.$$ The subadditivity difference is $$\tilde{I}_{f_\beta}(X; (Y, Z)) - (\tilde{I}_{f_\beta}(X; Y) + \tilde{I}_{f_\beta}(X; Z)) = \frac{2 S_Y - 1 - S_{YZ}}{-\beta(\beta - 1)}.$$ Since $0 < \beta < 1$, $-\beta(\beta - 1) > 0$. Because $G(p) = 2 S_Y - 1 - S_{YZ} > 0$, dividing by $-\beta(\beta - 1)$ gives $$\tilde{I}_{f_\beta}(X; Y) + \tilde{I}_{f_\beta}(X; Z) < \tilde{I}_{f_\beta}(X; (Y, Z)).$$ This completes the proof of power_divergence_exists_binary_counterexample_of_lt_golden_ratio.
5. Discussion
The result presented here resolves the necessity direction of the golden ratio classification for power-divergence mutual information. By establishing that a strictly positive binary counterexample exists for every exponent $\beta \in (0, (\sqrt{5}-1)/2)$, the theorem proves that subadditivity across all conditionally independent observations can hold only if $\beta \ge \beta_* = (\sqrt{5}-1)/2$.
This result moves beyond what was previously available in the literature:
- Replacement of isolated counterexamples: Previous refutations were confined to individual points such as $\beta = 1/4$ and $\beta = 1/2$. The analytical construction parameterized by $a, \kappa, p$ covers the entire continuum of exponents below $\beta_*$, demonstrating that the failure is structural and governed universally by the polynomial $1 - \beta - \beta^2$.
- Small counterexamples: The counterexample achieves strict superadditivity on a small configuration: all alphabets are binary ($|X| = |Y| = |Z| = 2$), the observation channels are identical, and all prior and transition probabilities are strictly positive. Thus, subadditivity fails without invoking multi-letter alphabets, channel asymmetries, or support boundaries.
- Elucidation of the golden ratio boundary: The asymptotic expansion of the derivative $L(a, \kappa)$ clarifies why the golden ratio conjugate appears: in the rare-symbol limit $p \to 0$ with crossover scaling $\kappa \gg a$, the fractional power difference $(a + \kappa)^{1-\beta} - (\kappa + a - a^2)^{1-\beta}$ contributes $2(1-\beta)a^2$, against which the marginal terms contribute $-(1-\beta)a^2 - \beta^2 a^2$. The resulting balance equation is precisely $(1-\beta) - \beta^2 = 1 - \beta - \beta^2 = 0$, whose unique root in $(0, 1)$ is $\beta_*$.
What remains open is the positive half of the classification: does subadditivity hold for all finite alphabets and all conditionally independent Markov chains throughout the interval $[\beta_*, 1)$, including at the golden ratio boundary $\beta_*$?
Numerical searches across asymmetric binary grids and thousands of skewed configurations on larger alphabets (such as $2 \times 2 \times 2$, $3 \times 2 \times 2$, $3 \times 3 \times 3$, $4 \times 3 \times 2$, and $6 \times 3 \times 3$) have failed to detect any subadditivity violation at or above $\beta_*$. However, establishing universal subadditivity on $[\beta_*, 1)$ remains an open problem. Addressing this question will require a new generator-specific concavity or operator-inequality framework capable of establishing the dimension-free inequality across arbitrary finite alphabets.
Why it is research-level
- Proves the failure half of the golden-ratio conjecture OP-726 for every exponent, not just at sample points: for every beta in (0, (sqrt 5 - 1)/2) there is a strictly positive binary Markov chain Y <- X -> Z on which mutual power-divergence information is strictly superadditive. The construction is a rare input symbol of probability p and a channel entry of order p, and the slack has derivative -(1-beta)a^2 - (1-(1-a)^beta)^2 + 2 kappa^beta((a+kappa)^(1-beta) - (kappa+a-a^2)^(1-beta)) at p = 0, positive for suitable a, kappa exactly when beta^2 + beta < 1. The source paper (arXiv:2610.21231) gives only sufficient conditions for subadditivity, and the knowledgebase had refutations only at beta = 1/4 and 1/2 plus numerical evidence for the threshold
Lean declarations
- power_divergence_exists_binary_counterexample_of_lt_golden_ratioFor every exponent beta in (0, (sqrt 5 - 1)/2), there exist strictly positive binary probability mass functions pX, PY, PZ forming a…
- power_divergence_generatorThe power-divergence generator f_β(x) = (x^β − βx + β − 1)/(β(β − 1)), for β ∉ {0, 1} and x ≥ 0 (β = 1/2 gives 2(√x − 1)², the squared…
- mutual_f_informationFor a joint probability mass function p_{XY} on finite sets, the mutual f-information is Ĩ_f(X; Y) = D_f(p_{XY} ‖ p_X p_Y) = Σ_{x,y}…
- mutual_power_divergence_information_eq_sum_rpowFor β ∈ (0, 1) and a joint probability mass function p_{XY}, the mutual f_β-information equals (Σ_{x,y} p(x, y)^β (p_X(x) p_Y(y))^{1−β} −…
Each one elaborated against Mathlib, and #print axioms returned
nothing outside Classical.choice · Quot.sound · propext.
References
- arXiv:2610.21231 — On the Subadditivity of Mutual f-Information
Drafted by gemini-3.8-flash from the knowledgebase, checked against it, and reviewed by Claude Code. Only the Lean is verified; the prose is a reading of it.