-
Iterated Entropy Derivatives and Binary Entropy Inequalities
Authors:
Tanay Wakhare
Abstract:
We embark on a systematic study of the $(k+1)$-th derivative of $x^{k-r}H(x^r)$, where $H(x):=-x\log x-(1-x)\log(1-x)$ is the binary entropy and $k>r\geq 1$ are integers. Our motivation is the conjectural entropy inequality $α_k H(x^k)\geq x^{k-1}H(x)$, where $0<α_k<1$ is given by a functional equation. The $k=2$ case was the key technical tool driving recent breakthroughs on the union-closed sets…
▽ More
We embark on a systematic study of the $(k+1)$-th derivative of $x^{k-r}H(x^r)$, where $H(x):=-x\log x-(1-x)\log(1-x)$ is the binary entropy and $k>r\geq 1$ are integers. Our motivation is the conjectural entropy inequality $α_k H(x^k)\geq x^{k-1}H(x)$, where $0<α_k<1$ is given by a functional equation. The $k=2$ case was the key technical tool driving recent breakthroughs on the union-closed sets conjecture. We express $ \frac{d^{k+1}}{dx^{k+1}}x^{k-r}H(x^r)$ as a rational function, an infinite series, and a sum over generalized Stirling numbers. This allows us to reduce the proof of the entropy inequality for real $k$ to showing that an associated polynomial has only two real roots in the interval $(0,1)$, which also allows us to prove the inequality for fractional exponents such as $k=3/2$. The proof suggests a new framework for proving tight inequalities for the sum of polynomials times the logarithms of polynomials, which converts the inequality into a statement about the real roots of a simpler associated polynomial.
△ Less
Submitted 22 December, 2023;
originally announced December 2023.
-
Prompt have evil twins
Authors:
Rimon Melamed,
Lucas H. McCabe,
Tanay Wakhare,
Ye** Kim,
H. Howie Huang,
Enric Boix-Adsera
Abstract:
We discover that many natural-language prompts can be replaced by corresponding prompts that are unintelligible to humans but that provably elicit similar behavior in language models. We call these prompts "evil twins" because they are obfuscated and uninterpretable (evil), but at the same time mimic the functionality of the original natural-language prompts (twins). Remarkably, evil twins transfe…
▽ More
We discover that many natural-language prompts can be replaced by corresponding prompts that are unintelligible to humans but that provably elicit similar behavior in language models. We call these prompts "evil twins" because they are obfuscated and uninterpretable (evil), but at the same time mimic the functionality of the original natural-language prompts (twins). Remarkably, evil twins transfer between models. We find these prompts by solving a maximum-likelihood problem which has applications of independent interest.
△ Less
Submitted 29 April, 2024; v1 submitted 12 November, 2023;
originally announced November 2023.
-
Extremal Graphs for a Spectral Inequality on Edge-Disjoint Spanning Trees
Authors:
Sebastian M. Cioabă,
Anthony Ostuni,
Davin Park,
Sriya Potluri,
Tanay Wakhare,
Wiseley Wong
Abstract:
Liu, Hong, Gu, and Lai proved if the second largest eigenvalue of the adjacency matrix of graph $G$ with minimum degree $δ\ge 2m+2 \ge 4$ satisfies $λ_2(G) < δ- \frac{2m+1}{δ+1}$, then $G$ contains at least $m+1$ edge-disjoint spanning trees, which verified a generalization of a conjecture by Cioabă and Wong. We show this bound is essentially the best possible by constructing $d$-regular graphs…
▽ More
Liu, Hong, Gu, and Lai proved if the second largest eigenvalue of the adjacency matrix of graph $G$ with minimum degree $δ\ge 2m+2 \ge 4$ satisfies $λ_2(G) < δ- \frac{2m+1}{δ+1}$, then $G$ contains at least $m+1$ edge-disjoint spanning trees, which verified a generalization of a conjecture by Cioabă and Wong. We show this bound is essentially the best possible by constructing $d$-regular graphs $\mathcal{G}_{m,d}$ for all $d \ge 2m+2 \ge 4$ with at most $m$ edge-disjoint spanning trees and $λ_2(\mathcal{G}_{m,d}) < d-\frac{2m+1}{d+3}$. As a corollary, we show that a spectral inequality on graph rigidity by Cioabă, Dewar, and Gu is essentially tight.
△ Less
Submitted 4 April, 2021;
originally announced April 2021.
