Cascade of transitions in molecular information theory
Extracted source text · arXiv 1804.03827
\documentclass[aps,pre,twocolumn,showpacs,superscriptaddress,groupedaddress]{revtex4-1}
% for review and submission
%\documentclass[aps,preprint,showpacs,superscriptaddress,groupedaddress]{revtex4}
%% for double-spaced preprint
%\usepackage{subfiles}
%\usepackage{lipsum}
%\usepackage{multicol}
\usepackage{graphicx} % needed for figures
\usepackage{dcolumn} % needed for some tables
\usepackage{bm} % for math
\usepackage{amssymb} % for math#
\usepackage[utf8]{inputenc}
\usepackage{amsmath,mathtools,amssymb,latexsym}
\usepackage{xspace,colortbl,color}
\usepackage{fullpage,enumerate}
%\usepackage{subfigure}
%\usepackage{bm}
\usepackage{epsfig,amsopn,graphicx}
\hyphenation{ALPGEn}
\hyphenation{EVTGEN}
\hyphenation{PYTHIA}
\def\hM{{\bm{M}}}
\def\hK{{\bm{K}}}
\def\hg{{\boldsymbol{\gamma}}}
\def\hG{{\boldsymbol{\Gamma}}}
\def\bS{{\boldsymbol{\Sigma}}}
\def\bG{{\bm{G}}}
\def\bg{{\bm{g}}}
\def\bea{\begin{eqnarray}}
\def\eea{\end{eqnarray}}
\def\etal{et al.}
\def\nn{\nonumber}
\def\om{\omega}
\def\f{\frac}
\def\bn{{\bm{n}}}
\def\be{{\bm{\hat{e}}}}
\def\bell{\bm \ell}
\def\br{\bm r}
\def\bZ{\bm Z}
\def\p{\partial}
\def\red#1{\textcolor{red}{#1}}
\def\blue#1{\textcolor{blue}{#1}}
\def\mP{\mathbb{P}}
\def\mX{\mathbb{X}}
\def\mY{\mathbb{Y}}
\def\cH{\mathcal{H}}
\DeclareMathOperator*{\argmax}{\arg\!\max}
\def\texitem#1{\vskip 5pt \par \noindent\hangindent 12pt
\hbox to 16pt {\hss #1 ~}\ignorespaces}
\newtheorem{lemma}{Lemma}
\newtheorem{proof}{Proof}
\begin{document}
%\lipsum
%opening
\title{Cascade of transitions in molecular information theory}
\author{Suman G. Das}
\affiliation{Simons Centre for the Study of Living Machines, National Centre for Biological Sciences (TIFR),
GKVK Campus, Bellary Road, Bangalore 560065, India}
\author{Madan Rao}
\affiliation{Simons Centre for the Study of Living Machines, National Centre for Biological Sciences (TIFR),
GKVK Campus, Bellary Road, Bangalore 560065, India}
\author{Garud Iyengar}
\affiliation{Industrial Engineering and Operations Research, Columbia University, New York, NY 10027}
\begin{abstract}
Biological organisms are open, adaptve systems that can respond to changes in environment in specific ways.
Adaptation and response can be posed as an optimization problem, with a tradeoff between the benefit
obtained from a response and the cost of producing environment-specific responses. Using recent results in stochastic thermodynamics,
we formulate the cost as the mutual information between the environment and the stochastic response. The problem of
designing an optimally performing network now reduces to a problem in rate distortion theory -- a branch of information theory
that deals with lossy data compression.
We find that as the cost of unit information goes down, the system undergoes a sequence of transitions, corresponding
to the recruitment of an increasing number of responses, thus improving response specificity as well as the net payoff. We derive formal equations for the transition points and exactly solve them for special cases.
The first transition point, also called the {\it coding transition}, demarcates the boundary between a passive response and
an active decision-making by the system. We study this transition point in detail, and derive three classes of asymptotic
behavior, corresponding to the three limiting distributions of the statistics of extreme values. Our work points
to the necessity of a union between information theory and the theory of adaptive biomolecular networks, in particular
metabolic networks.
\end{abstract}
\pacs{05.40.-a,65.40.gd,64.70.qd,87.10.Vg,87.10.Ca,87.10.Mn}
\maketitle
\section{Introduction}
Biological systems are distinctive in their ability to adaptively respond
to different environments. This applies not
only to organisms, but even % single
individual
cells in multicellular organisms\cite{alberts}. Adaptive behavior contains the notion of
specificity, i.e. the
% where a
response % must be
is tailored towards the particular
stimulus~\cite{specificity1,specificity2,bialek}. % that produces it
The response results in a benefit
to the cell. For example, % in the context of energy metabolism,
the
synthesis of enzymes specific to the breakdown of
a particular nutrient in the environment leads to a free-energy gain for
the cell. The net free-energy payoff should however include, apart from this gain, the free-energy {\it cost} of response
specificity. An optimal adaptive network % that
% has evolved to perform optimally will
evolves to maximize % the net payoff, which is
the difference between the benefit and
the specificity cost.
The degree of specificity in this biological context has been hard to quantify; % part of the
% reason is that
part of the reason is that specificity is not the
property of particular pathways, but of the entire repertoire of
possible cellular responses and their regulatory mechanisms. Recent
developments in information thermodynamics \cite{parrondo} allow
us to build a coarse-grained model of specificity cost in a system.
We model the adaptive response as a problem in
% communication
coding or data compression. Consider the schematic in \figurename{~\ref{schematic}}. The environment comprises of
a diversity of states indexed by $\alpha$, and they are presented to the cell with frequencies $f_\alpha$. % presented with
% a particular
The cell produces a response $i$ with
probability $p(i\vert \alpha)$. In information theoretic terms, the
environment state $\alpha$ is the signal and the response $i$ is the \emph{code}
or the representation, and the mapping $p(i \vert \alpha)$
is a random code.
The stochastic response to the environment
models the fact that responses
at the cellular level are
inherently noisy. % , and therefore the response must be characterized by a
% response probability distribution $p(i|\alpha)$.
% In terms of information theory, $p(i|\alpha)$ represents a channel, with
% $\alpha$ as an input and $i$ as the (probabilistic)
% output \cite{cover}.
From a bio-chemical perspective, the adaptive response essentially involves a transfer of
information from one set of molecules to another.
For example, many bacteria such as {\it E. coli} can grow on a wide
variety of sugars \cite{alon}. The concentration of various
enzymes in a bacterial cell contains information about the nutrients in
its growth medium~\cite{bialek}.
\begin{figure}
\includegraphics[height=2.8in,width=3.1in]{schematic}
\caption{A cell (organism) must respond adaptively to different environmental
states. The environment comprises a diversity of nutrients that
occur with frequency $f_\alpha$ ($\alpha=1,\dots,n$). The cell responds to the environment
with response probabilities
$p(i|\alpha)$, where $i$ labels the cellular response. Adaptive behavior is seen as a communication channel
between environment and response.}
\label{schematic}
\end{figure}
Since the environmental states occur with frequency $f_\alpha$, the
% nput and output pair satisfy a joint
% distribution
the joint distribution of the environment state $\alpha$ and the response
$i$ is given by $P(i,\alpha)=p(i|\alpha) f_\alpha$.
The mutual information between the environment $\alpha$
and reponse $i$ is
given by
\begin{eqnarray*}
I
& = & \sum_\alpha f_\alpha \sum_i p(i|\alpha) \log
\frac{P(i,\alpha)}{p(i)f(\alpha)},\\
& = &
\sum_\alpha f_\alpha \sum_i p(i|\alpha) \log \frac{p(i|\alpha)}{p(i)},
\end{eqnarray*}
where $p(i)=\sum_\alpha P(i,\alpha)$ is the marginal distribution of the
response. The mutual information $I$ denotes the average number of nats about the
environment state $\alpha$ learned from observing the response
$i$. Thus, the mutual information is high when the residual uncertainty in
the environment state $\alpha$ is low after observing the response $i$,
i.e. the response is very specific to each environment state.
% This is a measure of the specificity of response, since it is the
% number of bits
% about the input that can be recovered from the output. In perfectly
% general terms, the result of a classical measurement of
% a system with discrete states indexed by $\alpha$ is given by the conditional distribution $p(i|\alpha)$, and $I$ tells us the
% amount of information gained from the measurement.
We assume here that when the environment states change to a new randomly
chosen state,
the adaptive response of the system is instantaneous, and the memory of
the previous response is erased. This dynamics occurs on a scale much faster than adaptive evolution
over fitness landscapes.
Recent results in information thermodynamics have established that
the minimum free energy consumed in such measurement and erasure processes
is proportional
to the mutual information $I$
\cite{parrondo,tlusty_07,sgm}.
Thus, the cost of specificity % , viewed as a measurement process,
is proportional to $I$.
This cost is of fundamental thermodynamic origin, connected to problems involving the Maxwell's demon \cite{parrondo},
and it is also irreducible, so that it cannot be circumvented by better system design.
Here we ignore the cost of producing the response machinery, but focus only on the operational cost of regulating the machinery
to produce the appropriate response.
% It must be stated that
Note that mutual information is
a lower bound for this cost; the precise value of the free energy cost
depends on the specifics of the system, and may, in general, be
significantly
greater than this lower bound. Nevertheless, what we lose in sacrificing some of the
biological realism, we gain in the generality and
tractability of the model. This allows us to identify qualitative regimes of behavior even in the absence of detailed data.
Having identified the cost of specificity, we can understand the performance of an adaptive system
as involving a tradeoff between benefit and cost. We ask what an optimal response distribution, $p(i|\alpha)$, looks like,
given the abundances of various environmental states. We investigate this question by deriving the general features of
these distributions. Our main focus will be on biological systems that perform energy catabolism, because in these systems
it is the free energy cost that is of primary relevance in analyzing system performance. In other systems, such as those
involving the synthesis of proteins through transcription and translation, it is not obvious how to construct a
simple qunatitative indicator of performance,
and one should investigate
whether the free energy cost is indeed an important limiting factor. Nonetheless, the information theoretic
approach has been suggested as relevant
for the origin of molecular codes in a variety of contexts \cite{tlusty_07,tlusty_iop}.
In the next section, we state our model for energy catabolism formallly as an optimization problem.
\section{Model}
Consider an organism that is exposed to $n$ different nutrients % free energy
% sources (nutrients)
indexed by $\alpha$, that occur
in the environment with probability $f_\alpha$. % , and confer a maximum free
% energy benefit of $E_\alpha$.
The organism % ,
% when presented with the source
on sensing environment
$\alpha$ produces a response (i.e. the synthesis of the appropriate
transporters, enzymes, etc.) indexed by $i$
with a probability $p(i|\alpha)$.
Let $E_{i\alpha} \geq 0$ denote the free-energy benefit to the organism from
response $i$ in % the
% presence of free-energy source
environment
$\alpha$.
%Note that we are \emph{not} assuming
%the space of responses $i$ is the same as that of the free energy source
%$\alpha$.} % is
% The response needs to be highly
% specific, so that the free energy $E_\alpha$ is
% extracted only if $i=\alpha$.
The average free energy benefit is given by
\begin{equation}
G=\sum_i \sum_\alpha f_\alpha p(i|\alpha) E_{i\alpha}.
\end{equation}
The free energy cost of sensing the environment and producing a % specific
particular response, i.e. coding for the environment, is
proportional to the mutual information~\cite{parrondo,tlusty_07,sgm,tlusty_iop}.
%\begin{equation}
% I=\sum_\alpha f_\alpha \sum_i p(i|\alpha) \log \big(\frac{p(i|\alpha)}{p(i)}\big),
%\end{equation}
%where $p(i) = \sum_{\alpha} p(i\vert \alpha) f_{\alpha}$.
The net free energy payoff is therefore
\begin{equation}
F=G(p)-\kappa^{-1} I(p),
\end{equation}
where the constant $\kappa$ is determined by general thermodynamic and
chemical parameters that are independent of the channel. A higher
$\kappa$ is associated with a lower
cost of specificity.
An optimal metabolic network would tune the channel $p(\cdot|\cdot)$
to maximize
the payoff $F$, i.e. solve the optimization problem
\begin{equation}
\mbox{max}_{p(\cdot,\cdot)}\ G - \kappa^{-1}\ I(p).
\label{RD}
\end{equation}
Note that this problem is equivalent to the problem
\begin{equation}
\label{eq:metabolic-2}
\mbox{min}_{p(\cdot,\cdot)}\ -\kappa \sum_i\sum_{\alpha} f_{\alpha}
p(i\vert \alpha) E_{i\alpha} + I(p),
\end{equation}
i.e. a rate distortion problem with the distortion function $d(i,\alpha) =
- E_{i\alpha}$ \cite{cover}. The distortion function is a loss term
that measures the performance of the code, wheareas $I$ is the quantity of information
with which to minimize the loss.
%information about the source contained
%in the code.
%Rate distortion theory seeks to minimize the loss given a constraint on $I$.
In our interpretation, the first term in \eqref{RD} is a benefit term that depends on the specificity of
the molecular code, wheareas the fidelity incurs a cost given by the second term.
% The optimal channel is a function of the parameter set
%$\{\{f_\alpha,E_\alpha\},\kappa\}$.
This formulation is not limited to metabolic responses. It should be
applicable wherever the free energy cost is relevant,
and the benefit is linear in the conditional probabilities. For example,
rate distortion theory has been used to investigate
the origins of the genetic code in \cite{tlusty_07}.
%Our goal in this paper is to ivestigate the behavior
%of the optimal channel and the mutual information as a function of $\kappa$ and the nutrient distribution.
As a generic example of how the optimal solution behaves, we numerically compute in \figurename{~\ref{bif}} the
mutual information and conditional probabilities
for a particular parameter set when $n=6$, as a function of $\kappa$.
The data are generated by calculating the optimal channel by an alternating
minimization procedure called the Blahut-Arimoto algorithm \cite{cover}. The procedure
is described in Appendix A.
% The alogrithm used
% to generate the solution is described in
% Appendix A.
The notable feature of the plot is that the mutual information $I$ is $0$ at
$\kappa=0$, and $I$ undergoes
a sequence of bifurcations as $\kappa$
increases, analogous to continuous transitions in thermodynamic
systems. The first bifurcation point
has been called the coding transition in a related context
\cite{tlusty_iop} -- this is the value of $\kappa$ subsequent to which the mutual
information becomes non-zero.
Each transition point is characterized by the
fact that a new response is ``switched on''. For example, in \figurename{~\ref{bif}},
the response probability $p(2)$ becomes non-zero at the coding
transition point. As $\kappa$
increases, we see a sequence of such transitions, where typically only one
response is activated at a time.
We call this sequence of transitions a cascade. Similar cascades have been
observed previously at the interface of statistical mechanics
and computer science, such as in the application of deterministic
annealing to clustering, classification and other
computational problems~\cite{rose}.
The following section finds analytical solutions for the cascades.
\begin{figure}
\hspace{.3cm} \includegraphics[height=1.8in,width=2.75in]{cascade}
\includegraphics[height=1.7in,width=2.4in]{pcasc}
\caption{Cascade of transitions with $n=6$. Parameters: $E_{\alpha,\alpha}=\alpha+0.2$,
$f_\alpha \sim 6/(\alpha-0.4)$,
$E(i,j)=0.2 E(i,i)$, for $j \neq i$. (Upper panel) The benefit and mutual
information as a function of $\kappa$ are plotted.
The inset 2
zooms in on the first three transition points, marked by arrows. (Lower panel) The first three responses to be turned
on. The response $p(1)$ is the optimal non-coding response; $p(2)$ is turned on at the first transition point, also called
the coding transition.
}
\label{bif}
\end{figure}
\section{Analysis of cascades}
When $\kappa=0$, the cost of unit information is infinite, therefore the
mutual information $I=0$. Thus, the response $p(i\vert \alpha)$ is, in
fact, independent of the
environment, i.e. $p(i\vert \alpha) = p(i)$. % i.e the response is
% independent of the
% environment state.
Since $I(p) = 0$, the optimization problem reduces to
\[
\max_{p}\ \sum_ip(i) \Big(\sum_{\alpha} f_{\alpha} E_{i\alpha}\Big).
\]
It is clear that the optimal $p^\ast(i) = \delta_{ir}$, where $r =
\text{argmax} \{\sum_{\alpha} f_{\alpha}
E_{i\alpha}\}$. Thus, the organism uses the response $r$
irrespective of the environment state $\alpha$.
We will refer to $r$ as the optimal non-coding response.
As $\kappa$ increases, at some point the mutual information becomes
non-zero, i.e. the response $p(i\vert \alpha)$ \emph{does} depend on
$\alpha$. {This is called a
coding transition in a related context~\cite{tlusty_iop} because the
organism senses and codes the environment in terms of the response.}
As $\kappa$ continues to increase, we see more and more responses being
activated, each accompanied by a kink
in the mutual information and the benefit $G$ (see \figurename{~\ref{bif}}). In the simplest interpretation,
each response represents the expression of a particular gene (or operon).
The task of regulating the expression is carried out by transcription
factors in the cells, and the free energy cost
associated with the specificity of the responses arises out of the
activity of the regulatory machinery. The existence
of cascades tells us that in an optimal metabolic network, a particular
metabolic response may not be expressed
even when the appropriate enzyme-coding genes
and the corresponding nutrients are present in the system.
%When the distribution of nutrients in the environment changes,
%the same system may produce a response that it previously did not.
The number of responses available from an
optimal metabolic network is limited by the unit cost of information.
The distinctive feature of the cascades is that the responses are
activated one at a time. It is not clear {\it a priori}
that it should be so. One may think naively that for any $\kappa>0$, all responses would be active;
or that, in general, multiple responses should be switched on at a transition point.
We demonstrate in the following subsection that cascades are indeed generic. % is what one should generically expect.
\subsection{Cascades are generic}
% A closed expression for the solution of the general rate distortion problem is not
% known. But the criteria for
% a channel to be optimal are known.
We show below that under generic choice of parameters, two
different responses cannot transition at the same $\kappa$.
We start by assuming that the vector of free-energy $E_i =
\{E_{i\alpha}\}_{\alpha=1}^n$ corresponding to response $i$ are
distributed according to a distribution that has an $n$-dimensional
density. Thus, $\mathbb{P}(E_i \in A) = 0$ for any set $A$ that has
dimension at most $(n-1)$.
Let $p(i) = \sum_{\alpha} p(i\vert \alpha) f_{\alpha}$ denote the marginal
probability for response $i$. For $\kappa > 0$, results in Chapter 10 of~\cite{cover}
establish that $p$ is optimal, if and only if,
\begin{equation}
\label{eq:metabolic-3}
\sum_{\alpha} \frac{f_{\alpha} e^{\kappa E_{i\alpha}}}{\sum_j p(j)
e^{\kappa E_{j\alpha}}} \left\{
\begin{aligned}
\leq 1 && p(i) = 0,\\
= 1 && p(i) > 0.
\end{aligned}
\right.
\end{equation}
Strong convexity of mutual information implies that the optimal
solution $p_{\kappa}$ is unique and a continuous function of $\kappa$.
The optimal solution for $\kappa=0$ can be obtained by taking the limit of
$\kappa \searrow 0$. % , the above condition has to be evalued in the
% limit of small $\kappa$.
Now, fix a response index $s$. We call $\kappa_s$ a transition point
for the response $s$ if $p_{\kappa_s}(s) = 0$ and there exists
$\epsilon_s>0$ such that $p_{\kappa}(s) > 0$ for all $\kappa \in (\kappa_s,
\kappa_s + \epsilon_s]$. By the continuity property it follows that
\[
\sum_{\alpha} \frac{f_{\alpha} e^{\kappa_s E_{s\alpha}}}{J_{\kappa_s}(\alpha)}=1
\]
where $J_{\kappa}(\alpha) = \sum_j p_{\kappa}(j)
e^{\kappa E_{j\alpha}}$.
Suppose $\kappa_s$ is a transition point for another index $t \neq
s$, i.e. $p_{\kappa_s}(t) = 0$ and there exists an $\epsilon_t > 0$ such
that $p_{\kappa}(t) > 0$ for all $\kappa \in (\kappa_s, \kappa_s +
\epsilon_t]$. Thus, a necessary condition is that
\[
\sum_{\alpha} \frac{f_{\alpha} e^{\kappa_s E_{t\alpha}}}{J_{\kappa_s}(\alpha)}=1.
\]
Note that $p_{\kappa_s}(t) = 0$ implies that $E_t= \{E_{t\alpha}\}_{\alpha= 1}^n$ is
not included in the sum defining $\{J_{\kappa_s}(\alpha)\}_{\alpha = 1}^n$.
%Thus, the vector ${E_t}$ lies in an $(n-1)$ dimensional hyperplane.
Therefore,
\begin{eqnarray*}
\lefteqn{\mathbb{P}\big(\text{$\kappa_s$ is a transition point for
$t$}\big)}\\
& \leq & \mathbb{P}\left( \sum_{\alpha} \frac{f_{\alpha} e^{\kappa_s
E_{t\alpha}}}{J_{\kappa_s}(\alpha)}= 1 \right) = 0,
\end{eqnarray*}
where the last equality follows from the fact that the probability that
$E_t$ lies in an $(n-1)$ dimensional hyperplane is zero, which follows from our starting assumption.
(We reiterate that the case of $\kappa=0$ has to be analysed as a limit, in which case the same conclusions will follow.)
% The transition point for a response, say $s$, is the point where the
% equality above is satisfied, but with $p(s)=0$.
% Thus the condition for the transition point $\kappa^\ast$ is
% \[
% \sum_{\alpha} \frac{f_{\alpha} e^{\kappa^\ast E_{s\alpha}}}{J(\alpha)}=1
% \]
% where $J(\alpha)=\sum_j p(j) e^{\kappa E_{j\alpha}}$. If the responses
% $s$ and $t$ transition at the same $\kappa^\ast$,
% then we have
% \[
% \sum_{\alpha} \frac{f_{\alpha} e^{\kappa_s E_{t\alpha}}}{J(\alpha)}=
% \sum_{\alpha} \frac{f_{\alpha} e^{\kappa_s E_{s\alpha}}}{J(\alpha)}=1
% \]
% The parameters $f_\alpha$ and $E_{i \alpha}$ form an $N$-dimensional space,
% and together with $\kappa^\ast$ they inhabit an $(N+1)$-dimensional space.
% The equations above represent the intersection of two $N$-dimensional curves, which is an $(N-1)$-dimensional
% curve. Thus, for a solution to the above equation to exist, the parameters $f_\alpha, E_{i,\alpha}$ must be restricted
% to an $(N-1)$-dimensional space. But a {\it generic} choice of parameters must be from a finite density on the full
% $N$-dimensional space, which will therefore not satisfy the above equations.
% \iffalse
% Define the set of indices $A(\kappa)$ such that for all
% $i \in A(\kappa)$, either $p(i)>0$ or $\kappa$ is a transition point for $i$, ie
% $p(i)>0$ everywhere in the interval $(\kappa,\kappa+\epsilon]$ for some non-zero $\epsilon$.
% We call this set the {\it active set} at $\kappa$.
% \iffalse
% In this case, the equivalent of Eq. \ref{casc0} is
% \begin{equation}
% p(i|\alpha) = p(i) \exp[\kappa (\frac{\lambda_\alpha}{f_\alpha}+E_{i,\alpha})],
% \end{equation}
% Further, define the set of indices $A_\kappa$ such that $p(\beta)>0$ at $\kappa$, ie they represent the set
% of responses that are on at $\kappa$. Let us call this the active set.
% Using $\sum_i p(i|\alpha)=1$, we can derive the relation
% \begin{eqnarray}
% \lambda_\alpha &=& \frac{f_\alpha}{k} \ln \frac{1}{\sum\limits_{\beta} p(\beta) e^{k E_{\beta \alpha}}}\\
% &=& \frac{f_\alpha}{k} \ln \frac{1}{\sum\limits_{\beta \in A_\kappa} p(\beta) e^{k E_{\beta \alpha}}}.
% \label{casclambda}
% \end{eqnarray}
% Since $\sum_\alpha p(i|\alpha)f_\alpha=1$, the equations corresponding to \ref{casc2} are
% \begin{equation}
% \sum_{\alpha}f_\alpha \exp[\kappa (\frac{\lambda_\alpha}{f_\alpha}+E_{i,\alpha})]=1
% \label{casca}
% \end{equation}
% Notice that at the transition point for $i$, we have
% \[
% \lambda_i = \frac{f_i}{\kappa} \ln \frac{1}{\sum\limits_{A_\kappa} p(\beta) e^{\kappa E_{\beta i}}},
% \]
% which does not depend explicitly depend on the set $\{E_{i \alpha}\}$, since $i\notin\{\beta_\kappa\}$.
% At the value of $\kappa$ which induces the transition for $i$, we have
% \[
% \lambda_i= \frac{f_i}{k} \ln \frac{1}{\sum_\beta p(\beta) e^{k E_{i \beta}}}
% \]
% where $\beta$ belongs to the index set such that $p(\beta) > 0$.
% If $i$ and $j$ transition simultaneously, then we also have
% \begin{equation}
% \sum_{\alpha}f_\alpha \exp[\kappa (\frac{\lambda_\alpha}{f_\alpha}+E_{j,\alpha})]=1.
% \label{cascb}
% \end{equation}
% Let the equations \ref{casca},\ref{cascb} be satisfied simultaneously by $\kappa^\ast$.
% %Now notice that the lhs of both equations are strictly convex functions of $\kappa$, and each equation has
% %at most one solution for $\kappa$. Further the solution of the first equation is independent of the set $\{E_{j\alpha}\}$,
% %while that of the second is independent of the set $\{E_{i \alpha}\}$ (it can be seen from Eq. \ref{casclambda} that all the
% %$\lambda_\alpha$ are independent of these sets, since
% %$p(i)=p(j)=0$ at the joint transition point). Thus, generically, the solutions will not coincide, and we cannot expect joint
% %transitions.
% \fi
% Consider the scenario where two responses $i$ and $j$ make a simultaneous transition at $\kappa^*$. We want to find if this situation can
% be generic, ie if the property of joint transitions is stable under small perturbation to the parameters.
% Imagine a small
% perturbation to the set $\{E_{j,\alpha}\}$. Let the new transition points for $i$ and $j$ be $\kappa_i^\ast$ and
% $\kappa_j^\ast$. Generically, we should have $\kappa_j^\ast \ne \kappa^\ast$.
% Let us consider the case (if it exists) where the the property of joint transitions is robust to small perturbations, i.e
% $\kappa^\ast_j=\kappa^\ast_i=\kappa_{new}^\ast$. Further, let the active set for this perturbed problem be denoted by
% $A^p(\kappa)$.
% Now consider the following facts:\\
% {\it 1)} The active set $A^p({\kappa_{new}^\ast})$ is identical to the set $A({\kappa^\ast})$, since, by continuity,
% a sufficiently small perturbation will not alter which responses -have been turned on before $i$ and $j$ make
% the transition.
% \blue{{\it 2)}{\it Given} the active set at any $\kappa$, the optimal channel can be determined
% by optimizing over the active set; so, if the active set is given at the
% transition point of any $j$, we can solve for the
% corresponding $\kappa$, and the solution will not depend on the set
% $\{E_{j \alpha}\}$ explicitly,
% since $j$ does not belong to the
% active set. }
% \red{I don't understand the comment that ``one can solve for the
% corresponding $\kappa$''. Also, by your definition of the active set $j$
% \emph{does} belong to the active set.}
% {\it 3)} From the arguments above, we see that
% $\kappa_{new}^\ast=\kappa^\ast$, since the active sets are identical,
% and the only source of difference in parameters, namely the set $\{E_{j
% \alpha}\}$, does not influence the optimization
% over the active sets at the transition point of $j$ in either case.
% But this violates the condition that for
% generic perturbations,
% $\kappa_j^\ast \ne k^\ast$, indicating that a joint transition requires fine-tuning,
% and therefore is not generic.
% \fi
\subsection{Structure of bifurcation points for diagonal response}
Starting in this section we will focus on the special case where the
benefits are diagonal, i.e $E_{i \alpha}= E_\alpha \delta_{i \alpha}$.
This is a reasonable approximation in the context of metabolic responses,
since enzymes are generally highly specific to
substrates. In this section, we focus on computing the bifurcation
points~$\kappa$. % in the system.
% We introduce the simplification that
Define $g_\alpha=f_\alpha E_\alpha$.
%Assume that a bifurcation happens at some $\kappa^\ast$, with two
%alternative scenarios: \\
Then, the object to be maximized is
\[
\mathcal{L} = G- \kappa^{-1} I +\sum_\alpha \lambda_\alpha \sum_i p(i|\alpha),
\]
where the last term, involving the set of Lagrange multipliers $\lambda_\alpha$, ensure the normalization
of the response probabilities.
Equating the derivative of this to zero, we have
\begin{equation}
p(i|\alpha) = p(i) \exp[\kappa (\frac{\lambda_\alpha}{f_\alpha}+\frac{g_i}{f_i} \delta_{i,\alpha})].
\label{casc0}
\end{equation}
Solving for $\lambda_\alpha$ and substituting the result, we obtain the relations:
\begin{eqnarray}
p(i|\alpha) &=& \frac{p(i) e^{\kappa E_i \delta_{i \alpha}}}{1+p(\alpha) (e^{\kappa E_\alpha}-1)}
% p(i|i) &=& \frac{p(i) e^{k E_i}}{1+p(i) (e^{\kappa E_i}-1)}.
\end{eqnarray}
We use the relation $p(i)= \sum_\alpha p(i|\alpha) f_\alpha$ in
conjunction with the above equations to obtain
\begin{eqnarray}
\sum\limits_{\alpha \neq i} \frac{f_\alpha}{1+p(\alpha) (e^{\kappa E_\alpha}-1)}
+\frac{f_i e^{\kappa E_i}}{1+p(i)(e^{\kappa E_i}-1)} = 1 \nn,
\end{eqnarray}
for $p(i)>0$.
The sum can be rewritten as
\begin{eqnarray}
\sum\limits_{\alpha} \frac{f_\alpha}{1+p(\alpha) (e^{\kappa E_\alpha}-1)}
+\frac{f_i (e^{\kappa E_i}-1)}{1+p(i)(e^{\kappa E_i}-1)} = 1 \nn
\end{eqnarray}
Define $C=\sum\limits_{\alpha} \frac{f_\alpha}{1+p(\alpha) (e^{\kappa
E_\alpha}-1)}$. Then, from the above
equation,
\begin{equation}
p(i) = \frac{f_i}{1-C} - \frac{1}{e^{\kappa E_i} -1}.
\label{ex1}
\end{equation}
Thus, it follows that
\begin{equation}
\label{eq:ex-2}
p(i) > 0 \quad \Leftrightarrow \quad h_i(\kappa) := f_i(e^{\kappa E_i} - 1) > 1-C.
\end{equation}
Let us define the active set as $A = \{i: p(i) > 0\}$. Then substituting \eqref{ex1} in the definition
of $C$, we get
that
\[
C=(1-C) \sum\limits_{i \in A} \frac{1}{e^{\kappa E_i}-1} + \sum\limits_{j \notin A} f_j,
\]
Substituting the solution for $C$ in \eqref{ex1}
we obtain that for all $i \in A$,
\begin{equation}
p(i) = \frac{f_i}{\sum\limits_{m \in A} f_m}\big(1+ \sum\limits_{j
\in A} \frac{1}{e^{\kappa E_j}-1}\big)
-\frac{1}{e^{\kappa {E_i}} - 1}.
\label{ex2}
\end{equation}
This does not specify the solution for the marginals of the response
probability completely,
since we still need to know the active set $A$. We need to understand how the active set changes as $\kappa$ increases.
At $\kappa=0$, the only member of the active set is the optimal non-coding response. As $\kappa$ increases, we expect
the size of the active set to increase. However, it is also possible for responses to exit the active set;
see \figurename{~\ref{exitgraph}} for such an example. Such cases make the composition of the active set difficult to track and
the the bifurcation points harder to determine. Below we find conditions under which the behavior of the active set is more regular,
and use it to find the bifurcation points analytically.
First, we fix $\kappa$ and reorder the indices such that $h_i(\kappa) \geq h_j(\kappa)$
for $i \leq j$. Then \eqref{eq:ex-2} implies that the active set $A = \{i: i
\leq \ell\}$, i.e. we only need to search over $\ell$. The index $\ell$ is
optimal for $\kappa$ if $p_i > 0$ for all $i \leq \ell$, and $p_i=0$ for $i>l$.
Next, we characterize special cases where there exist thresholds $\{\kappa_i: 1
\leq i \leq n\}$ such that $p(i) > 0$ for all $\kappa > \kappa_i$. In other words, no response exits
the active set as $\kappa$ increases, and the cascade consists entirely of transitions into the active set.
\begin{figure}
\includegraphics[height=2in,width=2.5in]{exitgraph}
\caption{Here we show a simulation of the optimal marginal probabilities $p(i)$, for $n=40$; the diagonal benefit values
$E_{ii}$ were
chosen randomly from a uniform distribution, whereas the off-diagonal $E_{i\alpha}$ were chosen uniformly from
$[0,\sigma_{od} E_{i,i}]$, with $\sigma_{od}=0.9$. The red curve shows a response that is switched on at a particular
value of $\kappa$
and exits the active set at a higher value of $\kappa$.
Some of the other responses that do not exit the active set are shown
in dashed lines for comparison. As $\sigma_{od}$ decreases, instances of exits from the active set become rare.
}
\label{exitgraph}
\end{figure}
% The transition point $\kappa$
% for the response $i$ is given
% by
% \begin{equation}
% \frac{f_i}{\sum\limits_{l \in A_{\ell}} f_l} \big(1+ \sum\limits_{j \in
% A_{\ell}} \frac{1}{e^{\kappa E_j}-1}\big)
% -\frac{1}{e^{\kappa {E_i}} - 1}=0,
% \label{eq:trans}
% \end{equation}
% There are instances where we can find
% the active set independently, and solve Eq.(\ref{ex2}) for the bifurcation points.
% From the approach given here, we can prove some further properties of
% active sets. For example, once a response enters the active set, can it
% exit the active set as $\kappa$ increases? We can prove the following
% proposition.
\begin{lemma}
Suppose either (a) $E_i \equiv E$ or (b) $f_i \equiv f = \frac{1}{n}$. Then
there exists thresholds $\kappa_m$ such that $p(m) > 0$ for all $\kappa >
\kappa_m$.
\end{lemma}
From \eqref{eq:ex-2} it follows that in both the special cases (a) and
(b), the ordering $\{h_i(\kappa): 1\leq i \leq n\}$ is independent of
$\kappa$. Order the indices $i$ such that $h_i(\kappa) \geq h_j(\kappa)$
for all $i < j$. From \eqref{ex2}, it follows that $p(m) > 0$ if, and only if,
\[
\sum_{j \leq m} f_j\left( \frac{1}{h_i(\kappa)} -
\frac{1}{h_j(\kappa)} \right) < 1
\]
Define $\Delta_i(\kappa) = \frac{1}{h_{i+1}(\kappa)} -
\frac{1}{h_i(\kappa)}$ for $i < n$. We show that in both cases (a) and
(b), $\Delta_i(\kappa) = \frac{1}{h_{i+1}(\kappa)} -
\frac{1}{h_i(\kappa)}$ which is clearly monotonically decreases with $\kappa$. Define
$\kappa_m$ to be the solution of the equation
\[
\sum_{j \leq m} f_j\left( \frac{1}{h_i(\kappa)} -
\frac{1}{h_j(\kappa)} \right)\\
= \sum_{j < m} \big(\sum_{k \leq j} f_k\big) \Delta_j(\kappa) = 1.
\]
Then the monotonicity of $\Delta_j(\kappa)$ implies that $p(m)~=~0$ for
all $\kappa \leq \kappa_m$ and $p(m) > 0$ for all $\kappa > \kappa_m$.
For case (a)
\[
\Delta_i(\kappa) = \left(\frac{1}{f_{i+1}} - \frac{1}{f_i} \right)
(e^{\kappa E}-1)^{-1},
\]
and clearly monotonically decreasing in $\kappa$.
Define $x={(e^{\kappa_m E}-1)}^{-1}$, and $F_m=\sum_{j\leq m} f_j$.
From (\ref{eq:ex-2}) we have
\[
\frac{f_m}{F_m} (1+ mx)-x=0.
\]
Thus,
\[
\kappa_m = \frac{1}{E} \ln \left(\frac{F_m}{f_{m}}-(m-1)\right).
\]
The organism transitions to a coding mode when at least two responses have
a positive probability of occurring. This transition occurs at
\[
\kappa_2 =\ln(\frac{f_1}{f_2}).
\]
Thus, the coding transition point depends rather weakly on the ratio
of highest to second highest frequencies.
% The mutual information $I_m$ at $\kappa_m$ is given by
% \begin{eqnarray*}
% \lefteqn{I_m = } \\
% && H(f)- \Big[ \sum_{\beta >m } f_\beta \ln \frac{1}{f_\beta}+(m-1)f_m \ln \frac{1}{f_m}\nn \\
% && \mbox{} + (F_m-(m-1)f_m)
% \ln \big(\frac{1}{F_m -(m-1)f_m}\big) \Big], \nn
% \end{eqnarray*}
% where $H(f)= - \sum f_\alpha \ln f_\alpha$ is the entropy of the
% distribution of nutrient frequencies.
% The second term is positive and gives the conditional entropy of the nutrients.
% To solve a special case of this, let us consider exponentially decaying
% frequencies,
Let us treat some particular instances of this case.
Consider exponentially decreasing frequencies of the form $f_m \sim 2^{-c_n m}$, where $c_n$
is some constant that, in general, depends on the total number of nutrients $n$.
%First, consider the case where $c_n$
%is independent of $n$; we see that the ratio of any two frequencies is now independent of $n$, and therefore
%$F_m/f_m$ is independent of $n$. In that case, $\kappa_m$ is independent of $m$.
The solution for the transition points is:
\[
\kappa_m=\ln \big[\frac{2^{c_n}(2^{c_n m}-1)}{2^{c_n}-1}-(m-1)\big]
\]
Consider the case where $c_n=1$.
Then,
% and $m=1,..,n$. Then
\[
\kappa_m=\ln (2^m-m).
\]
At large $m$, the transition point $k_m$ is linear in $m$, although the frequencies
decay exponentially in $m$. Thus, % it does not require a very
the information cost need not be very low for the responses to rare
nutrients to be active. This feature can be generically expected because
of the logarithmic dependence of the transition point on
$\frac{F_m}{f_m}$. A numerical verification of the result above is given in \figurename{\ref{analytical}a}.% As an illustration of this point,
The fact that $\kappa_m$ is independent of $n$ is somewhat counterintuitive,
since we would expect the transitions to occur earlier when the diversity is larger.
This would occur if we choose a $c_n$ that goes to zero asymptotically, in which case, for large $n$,
\[
\kappa_m \approx \frac{c_n}{2}(m-1).
\]
In particular, if $c_n=\frac{1}{n}$, then $\kappa_m \approx \frac{m-1}{2n}$, so the transition points are inversely
proportional to the diversity $n$. The crucial diffetrence between the two cases is that in the former, the ratio of the
frequencies are independent of $n$, whereas in the latter case, the different frequencies start to
crowd together, i.e the ratio of any two frequencies $f_l$ and $f_m$
approach $1$ in the limit of large $n$.
%\[
% \kappa_m=\ln \big[\frac{2^{\frac{1}{n}}}{2^{\frac{1}{n}}}(2^{\frac{m}{n}}-1)-(m-1) \big]
%\]
\iffalse
Now consider
a different case, where $l_n=2^\frac{1}{n}+1$;
In this case,
\[
\kappa_m =
\]
\fi
For case~(b), the derivative
\[
\Delta_i'(\kappa) = \frac{E_i e^{\kappa E_i}}{f(e^{\kappa E_i}-1)^2} -
\frac{E_{i+1} e^{\kappa E_{i+1}}}{f(e^{\kappa E_{i+1}}-1)^2}
\]
Define $\alpha = \frac{E_{i}}{E_{i+1}} > 1$ and $x = e^{\kappa
E_{i+1}}>1$. Then
\[
\Delta_i'(\kappa) = \frac{\alpha x^{\alpha}}{f(x^{\alpha}-1)^2} -
\frac{x}{f(x-1)^2}.
\]
Fix $x > 1$, and define $g(\alpha) = \frac{\alpha
x^{\alpha}}{(x^{\alpha}-1)^2}$.
Then
\[
g^\prime(\alpha)= \frac{x^\alpha W(\alpha \ln(x))}{(x^\alpha-1)^3}
\]
where $W(y)=e^y(1-y)-(1+y)$ for $y \geq 0$. Since $x>1$ and $\alpha >
1$, the sign of
$g^\prime(\alpha)$ is the same as that of
$W$. We see that $W(0)=0$, and
$\frac{dW}{dy}=-(y e^y+1) < 0$ for all $y$. Therefore $W(y) < 0$ for
$y>0$; consequently, $g^\prime(\alpha) < 0$ and $\Delta_i(\kappa)' =
\frac{1}{f}\big(g(\alpha) - g(1)\big) < 0$,
i.e. $\Delta_i(\kappa)$ is monotonically decreasing.
% \iffalse
% \begin{eqnarray*}
% \ln(g(\alpha))'
% & = & \frac{1}{\alpha} + \ln(x) -
% \frac{2x^{\alpha}\ln(x)}{x^{\alpha}-1}\\
% & = & \frac{1}{\alpha}\Big( 1 -
% \big(\frac{x^{\alpha}-1}{x^{\alpha}-1}\big)\ln(x^{\alpha})\Big)
% \end{eqnarray*}
% Let $w(y) = \big(\frac{1+y}{y-1}\big)\ln(y)$ for $y > 1$. Then $\lim_{y
% \downarrow 1} w(y) = 1$. We will next establish that $w'(y) \geq 0$.
% The derivative
% \[
% w'(y) = \frac{1}{y(y-1)^1}\big( y^2 - 1 - 2y\ln(y) \big).
% \]
% Let $u(y) = 2y^2 - 1 - 2y\ln(y)$. Then $u(1) = 0$, and $u'(1) = 0$, and
% $u''(y) = 2 - 2/y \geq 0$ for all $y \geq 1$. Thus, it follows that $u(y)
% \geq 0$ for all $y \geq 0$. Thus, it follows that $w(y) \geq 2$ for all $y
% \geq 1$. Consequently,
% \[
% \ln g(\alpha)' < -\frac{1}{\alpha} \leq 0.
% \]
% \fi
\begin{figure}
\includegraphics[width=2.8in,height=2.6in]{hm}
\caption{The fraction of cases where at least one exit from the active set happens as $\kappa$ increases from $0$.
The diagonal benefit values $E_{ii}$ and the unnormalized frequencies are chosen uniformly from $[0.5,1.5]$. The
off-diagonal benefit values $E_{ij}$ are chosen uniformly from $[0,\sigma_{od} E_{ii}]$. When $\sigma_{od}$ is less than
$0.5$, exits become extremely rare, and typically no exits are found in a set of $10^3$ different realizations.
}
\label{exit}
\end{figure}
In this case, one cannot solve for the bifurcation points $\kappa_m$
exactly. However, one can approximate $\kappa_m$ when
$\kappa_m E_i \ll 1$ for all $i \leq m$.
%The last condition would hold, for example, when the benefit values are chosen independently from a distribution that has
%a sufficiently fast tail, such as a distribution with a finite support (as will be clearer in the next section).
We can then
expand \eqref{ex2} up to linear order in $\kappa E_i$ and obtain the solution
\[
\kappa_m \approx \frac{m}{E_{m}} - \sum_{j=1}^m \frac{1}{E_j}.
\]
See \figurename{\ref{analytical}b} for comparison of the prediction with data.
Unlike in Case~(a), the bifurcation points depend more sensitively on the
benefit values $E_i$. Once again, consider
the coding transition point, $\kappa_2$. If $E_2 \ll E_1$, it can be shown that $\kappa_2 \approx \ln(2)(\frac{E_1}{E_2})$,
i.e the coding transition depends linearly rather than logarithmically on
the ratio of highest to second-highest benefits.
In an approximate sense, a low-frequency nutrient is more likely to be
coded for than a low-benefit nutrient.
In general, the bifurcation points can be solved for when the ordering of the $h_i$ are independent of $\kappa$. This
happens when $f_i \le f_j$ if and only if $E_i \le E_j$ for all $i,j$. When this condition is violated, the ordering depends on
$\kappa$, and the
active set becomes harder to identify. When one also allows for the off-diagonal elememts $E_{i \alpha}$ to be non-zero,
we find numerically that exits from the active set do happen. It is shown in \figurename{~\ref{exit}} that
exits are common when the off-diagonal terms are large, but as they become smaller, exits become very rare and
eventually cannot be numerically detected at all.
% To prove this, order the responses $i$ in decreasing order of $h_i$, i.e. $h^{\kappa}_1 \geq h_2^\kappa
% \geq \ldots \geq h_n^\kappa$. From \eqref{eq:ex-2} it follows that the active sets
% are ordered in decreasing order of $h$, i.e. for an active set of size $l$, we only need to consider sets
% of the form $A_{\ell} = \{i: 1 \leq i \leq \ell\}$. When either the frequencies or the benefits are equal,
% we can see from \eqref{eq:ex-2} that the ordering of the $h_i$ is independent of $\kappa$. So, if a response $i$
% is not in the active set, any response $j>i$ cannot be in the set either. Thus, the only way in which a response can enter
% the active set at $\kappa_1$ and exit it at a higher value $\kappa_2$ is if the active set
% remained unchanged between $\kappa_1$ and $\kappa_2$, and $h_i < h_j$ for all $j\neq i$ in the active set.
% In that case, from \eqref{ex2}, points $\kappa_1$ and $\kappa_2$
% must be the solution of \eqref{eq:trans} for a given active set. When $E_j=E$ for all $j$, the equation reduces to
%Since, generically, $h_i \neq h_j$ for $i \neq j$, it follows that the
%responses $i$ are turned on sequentially rather than simultaneously (consistent with the argument of the previous section).
% \[
% \frac{f_i}{\sum\limits_{l \in A_{\ell}} f_l} \big(1+ \sum\limits_{j \in A_{\ell}} \frac{1}{e^{\kappa E}-1}\big)
% -\frac{1}{e^{\kappa {E}} - 1}=0,
% \]
% and it is easily shown that this equation has at most one solution in $\kappa$, and therefore cannot be satisfied by
% both $\kappa_1$ and $\kappa_2$. Similarly, when $f_j=\frac{1}{n}$ for all $j$, the equation reduces to
% \[
% \sum_{j \in A_{\ell}}[ \frac{1}{e^{\kappa E_i}-1}-\frac{1}{e^{\kappa E_j}-1}]=1,
% \]
% where $E_i<E_j$ for all $j\neq i$ in the active set, and each term in the square bracket is monotonic decreasing.
% Thus, the equation cannot be
% satisfied by more than one value of $\kappa$. This completes the proof.
% \subsection{Bifurcation with off-diagonal response}
% \noindent
% Before moving on, we mention a few properties of the off-diagonal
% responses, i.e of $p(i|\alpha)$, where $i \neq \alpha$.
% As we have seen,
The optimal $p(i\vert \alpha)$ for $\alpha \ne i$ satisfies
\begin{equation}
\frac{p(i \vert \alpha)}{p(i)} = \frac{1}{1+p(\alpha)(e^{\kappa E_\alpha}-1)} <
p(\alpha \vert \alpha)
\label{casc0}
\end{equation}
% A number of properties are worth noting. First, every off-digaonal
% response probablity is less than or equal to the
% corresponding marginal
% probability, and therefore less than or equal to the diagonal response probability.
This makes intuitive sense, since the
off-diagonal responses do not contribute to the benefit term. Note that
the off-diagonal $p(i \vert \alpha)$ is positive
whenever a response is active, i.e. $p(i)>0$. In an experimental context, such
non-cognate responses may show up merely as
noise, but our analysis here indicates that it may be the result of
an optimization process. Although the off-diagonal
terms do not increase the benefit, they % can reduce the
contribute to decreasing the
mutual information,
and therefore increase the net payoff.
Third, when the response $\alpha$ is not active, we have $p(i|\alpha)=p(i)$, i.e the metabolic response contains no
information about the nutrients in the non-active set.
\begin{figure}
\includegraphics[height=2in,width=3in]{analytical}
\caption{ {\bf a.} Mutual information and net payoff for an exactly solvable case, with $E_m=1$, and
$f_m= \frac{2^{-m}}{1-2^{-n}}$,
for $n=7$. The dashed
lines are from the numerically obatined solutions, whereas the symbols are the exact solutions at the transition point.
{\bf b.} The approximate transition points for the case $f_i=1/n$, $E_\alpha=1-(\alpha-1)^2/n^2$, for $n=8$, are marked as
triangles. They are compared with the numerically obtained transition points for the second to fifth responses. The approximation
fits the data well up to the third transition point.}
\label{analytical}
\end{figure}
% \iffalse
% \subsection{Exits from the active set}
% When either the frequencies are equal or the benefits are equal, the cascades have the property that a response
% in the active set never exits the set as $\kappa$ is increased. We prove it as follows.
% First, consider the case where the benefits are all equal to $E_i=E$ , and the frequencies are arbitrary.
% We start by assuming that a member of
% the active set exits, and subsequently derive a contradiction. Consider the first response to exit an active set,
% and let its index be $l$. Two cases must be considered. (a) After $l$ was activated and before it exited, at least one other response was activated, say $m$.
% Then at some value of $\kappa$, the optimal channel has $p(l)>0$ and $p(m)=0$. This must mean that $f_l > f_m$,
% for otherwise, at that $\kappa$, one could switch
% the probabilities $p(l|\cdot)$ and $p(m|\cdot)$ and improve the benefit while keeping the mutual information the same,
% indicating that the channel is not optimum. When $l$ exits, we have $p(l)=0$ but $p(m)>0$ (since $l$ is the first
% response to exit), and therefore by the same argument as above $f_l < f_m$, which contradicts the above condition.
% (b) After $l$ was activated, no other response is activated till it exits, ie the active set remains unchanged between
% $\kappa_1$ (where $l$ enters the active set) and $\kappa_2$ (where $l$ exits the active set).
% In this case, we get from Eq. \ref{ex2} the condition that $\kappa_1$ and $\kappa_2$ must satisfy the equation
% \[
% \frac{f_i}{\sum\limits_{l \in A_{\ell}} f_l} \big(1+ \sum\limits_{j \in A_{\ell}} \frac{1}{e^{\kappa E}-1}\big)
% -\frac{1}{e^{\kappa {E}} - 1}=0.
% \]
% It is easily shown that this equation has at most one solution in $\kappa$, and therefore cannot be solved
% by both $\kappa_1$ and $\kappa_2$.
% Second, consider the case where the frequencies are equal, and the benefits are arbitrary.
% We proceed as before. The proof of case (a) is essentially identical. The proof for case (b) is
% as follows:
% The equation above, in this case, can be written as
% \[
% \sum_{j \in A_{\ell}}[ \frac{1}{e^{\kappa E_i}-1}-\frac{1}{e^{\kappa E_j}-1}]=1.
% \]
% Here $E_i < E_j$ for all $j \neq i$, and each term in the square bracket is monotnic decreasing. Thus, the equation cannot be
% satisfied by more than one value of $\kappa$.
% \iffalse
% {Let $p(i) = \sum_{\alpha} p(i\vert \alpha) f_{\alpha}$ denote the marginal
% distribution of the organism's response. Then results in Chapter 10
% in~\cite{cover} establish that $p$ is optimal, if and only if,
% \begin{equation}
% \label{eq:metabolic-3}
% \sum_{\alpha} \frac{f_{\alpha} e^{\kappa E_{i\alpha}}}{\sum_j p(j)
% e^{\kappa E_{j\alpha}}} \left\{
% \begin{aligned}
% \leq 1 && p(i) = 0,\\
% = 1 && p(i) > 0.
% \end{aligned}
% \right.
% \end{equation}
% }
% \fi
% \fi
% Now consider a different case, where $f_i=\frac{1}{n}$. We cannot solve for the bifurcation points exactly, but we
% can make an approximation when $n$ is large, and $m \ll n$, and in
% the regime $\kappa E_\alpha \ll 1$.
% %The last condition would hold, for example, when the benefit values are chosen independently from a distribution that has
% %a sufficiently fast tail, such as a distribution with a finite support (as will be clearer in the next section).
% We can then
% expand \eqref{eq:trans} up to linear order in $\kappa E_i$ and obtain the solution
% \[
% \kappa_m \approx \frac{m}{E_{m}} - \sum_{j=1}^m \frac{1}{E_j}.
% \]
% Unlike in the case of frequencies, the transition point depends more sensitively on the benefit values. Once again, consider
% the coding transition point, $\kappa_2$. If $E_2 \ll E_1$, it can be shown that $\kappa_2 \approx \ln(2)(\frac{E_1}{E_2})$,
% ie the coding transition depends linearly rather than logarithmically on the ratio of highest to second-highest benefits.
% In an approximate sense, a low-frequency nutrient is more likely to be coded for than a low-benefit nutrient.
Among
all the transition points, the coding transition
is special because it separates the regime of constant response from a
regime where the organism must sense the environment.
We devote the next section to the analysis of the coding transition,
i.e. the value $\kappa^\ast$ when the conditional distribution $p(i\vert
\alpha)$ is no longer uniform for all environment states $\alpha$.
\section{Asymptotics of the Coding Transition and the theory of extreme
values}
% As we have seen before, when
% dependent decisions are taken
% by a cell.
The coding transition has been suggested as the origin of
molecular codes in biology \cite{tlusty_iop}.
Here we ask the question: how generic is the phenomenon of metabolic
coding? When the number of nutrients is
small, $\kappa^\ast$ has a finite value, but as the number of
nutrients increases, there is an increasing
benefit to be had from distinguishing the nutrients, and we could expect
$\kappa^\ast$ to be small. But does it go to
$0$ asymptotically, and if so, at what rate?
We will see in the following that the answer depends on
the distribution of frequencies and benefits of the nutrients, and certain
universality classes of the coding behavior
can be identified from the theory of extreme value statistics.
Recall that when
$\kappa=0$, the optimal response is non-coding, and the
optimization problem becomes:
\[
\max_{p} \sum_i p(i) g_i,
\]
where $g_i =f_i E_i$.
The optimal solution is $p_0(i) = \delta_{ir}$, where $r =
\text{argmax} \{g_i\}$. This solution remains optimal provided
\[
\sum_{\alpha} \frac{f_{\alpha}e^{\kappa E_{i\alpha}}}{\sum_j p_0(j)
e^{\kappa E_{j\alpha}}} \leq 1
\]
for all $i \neq r$. For the special case where $E_{i\alpha} = E_{\alpha}
\delta_{i\alpha}$, this condition reduces to
% The coding transition occurs at $\kappa = \kappa^\ast$
\[
f_i e^{\kappa E_i} + f_r e^{-\kappa E_r} \leq f_i + f_r,
\]
for all $i \neq r$. Thus, the coding transition point $\kappa^\ast$ is
smallest positive value of $\kappa$ that solves one of the
$n-1$ equations
\[
\frac{f_i}{f_r} (e^{\kappa E_i}-1)= 1- e^{-\kappa E_r}
\]
for $i \neq r$. Let $x = e^{\kappa E_r}$, $s_i = \frac{E_i}{E_r}$, and
$w_i = \frac{f_i}{f_r}$. % Then
% without loss of generality, we can put $E_r$ =1, and $\kappa$ is taken
% to be expressed in the units of ${E_r}^{-1}$,
% and $E_i$ is expressed in units of $E_r$.
Then the above equation reduces to
\begin{equation}
w_i x^{1+s_i}=(1+w_i)x-1.
\label{eq:wi}
\end{equation}
% where $w_i=\frac{f_i}{f_r}$, and $x=e^\kappa$ and $\rho_i=E_i$. Note that $\kappa$ and $E_i$ are now dimensionless.
One solution is $x^\ast=1$ corresponding to $\kappa=0$. Our goal is to compute
the non-trivial solution $x^\ast > 1$.
% In real scenarios, the
In typical scenarios, the precise distribution of nutrient benefits
$E_{\alpha}$ are unlikely to be known, and hence the precise value of
$x^\ast$ cannot be computed. However, one may be able to identify certain
universal features of the coding transition $x^\ast$ as a function of the
distribution of benefits $\{E_{\alpha}\}_{\alpha = 1}^n$ and the
frequencies $\{f_{\alpha}\}_{\alpha = 1}^n$.
We assume that $E_\alpha$ are independent identically distributed samples
from a fixed distribution $\pi_E$, and
$f_{\alpha} = \frac{h_{\alpha}}{\sum_{\alpha^\prime} h(\alpha')}$ where
each term $h_{\alpha}$ is an independent sample from the distribution
$\pi_h$.
% But under some reasonable
% assumptions, we can extract a few
% universal features of the transition point. In fact, the asymptotoic
% scaling of $\rho$ with $n$
% can be showed to belong to some universality classes according to
% the theory of extreme value statistics.
% In general, let the $n$ benefit values be drawn independently from the distribution $Q(x)$.
% For distributions with faster than power law tails, $\kappa \to 0$ as $n \to \infty$, and we can make a useful
% approximation as follows.
We will focus on cases where the coding transition occurs at
$\kappa^\ast \approx 0$ as $n \rightarrow \infty$, equivalently, at
$x^\ast = 1 + y$, for $y \ll
1$. Substituting $y_i=x_i -1$ in
\eqref{eq:wi}, and expanding up to second order in
$y_i$, we get
\begin{equation}
y_i= \frac{2(1-s_i w_i)}{1-\frac{1}{2} s_i w_i (1-s_i)},
\label{eq:smally}
\end{equation}
which holds for all $y_i \ll 1$. Since $s_i w_i = \frac{f_i E_i}{f_rE_r} = \frac{g_i}{g_r}<1$, we have that
$y_i>0$, as required.
% The solution $y$ can be small in the following two different scenarios:
% \begin{eqnarray}
% s w & \lesssim & 1 \nn\\
% s & \gg & 1/s w \label{eq:s_gg_1}
% \end{eqnarray}
% %The second option implies that $E_r \gg E_i$.
We show in the Appendix
that % under fairly general
% sufficient conditions
when the distribution of $E$ is bounded, the above equation further reduces to
% , the second condition does not hold; and further,
% when the first condition holds, we have $s_i \to 1$.
% in the Appendix
% However, under fairly general conditions,
%the second condition cannot
%hold. % gives sufficient conditions that rule
% this out, in which case we end up with
% We end up with
\begin{equation}
\label{eq:s-cond}
y_i \approx {2(1-s_i w_i)},
\end{equation}
and therefore the transition corresponds to the largest value of $s_i w_i$.
Under this approximation, the coding transition is determined by the
distribution of the product $g_i = f_i E_i$, % i.e the
% product of $f_i$ and $E_i$ that matters,
and not the distributions of the individual terms $f_i$ and $E_i$.
% and the individual distributions are of no consequence. So let us define a
% new variable
Define $\rho_i=g_i/g_r$, where $g_r$
is the benefit associated with the non-coding response. Clearly, $\rho_i < 1$, and
\begin{equation}
\kappa^\ast=2(1-\rho),
\end{equation}
where $\rho = \max_{i \neq r} \{\rho_i\}$. % is the value of $\rho_i$ maximized
% over all $i \neq r$
Thus, the distribution of $\kappa^\ast$ is determined by the distribution
of $\rho$, i.e.
% Further,
% we want to find the mean value of $\kappa^\ast$, where the average is
% taken over the distribution from which the
% $g_i$ are chosen. Thus we need to find
% the mean of the quantity $\rho$. The problem now reduces to one in the
% statistics of extreme values, where we need to
% find the mean of
the ratio of the second highest to highest value out of
$n$ values chosen independently from an identical
distribution $Q(x)$.
%and for the rest of this section we will denote the mean itself
%simply as $\rho$, without using any additional notation to indicate averages.
%We are interested in the distribution
%\[
%P_r (\rho)=\int P(x_{max}=x,x_{2max}=\rho x) dx
%\]
%To evaluate this, we first ask for the joint ditribution $P_2(x_1,x)$ of
%the two highest values,
%$x$ and $x_1$, $(x \ge x_1)$.
Along the lines of Eq 13 in \cite{satya}, we can write
\[
P_2(x,x_1) = n Q(x) P_{max} (x_1;n-1) \Theta(x-x_1)
\]
where the term $n Q(x_1)$ is the probability that {\it any} one of the $n$
chosen numbers has the value $x_1$, and
$P_{max} ( x;n-1)$ is the probability that among the remaining $n-1$
numbers (which were
chosen independently from $Q(x)$), $x$ is the highest value; the Heaviside
step function $\Theta$ ensures that
the probability is zero when $x>x_1$. We need to characterize the ratio
$\rho=x/x_1$
using this distribution. It is natural to split
the problem into three different classes of distribution $Q$, which are
known to exhibit different
extreme value behavior \cite{book}.
\subsection{Asymptotic classes}
Suppose $n$ values $X_i$ are sampled independently from a distribution
$Q$. Let $x_n = \max_{1\leq i \leq n} \{X_i\}$ denote the maximum of these
$n$ samples. Then
% and the
% largest value is $x$,
there exist % scaling
coefficients $a_n$, and $b_n$ such that the distribution of
$x_n$ for large $n$ is given by $b_n^{-1} f(\frac{x-a_n}{b_n})$, where the
function $f$ is one of three universal
limiting forms of extreme value statistics. The precise form the limiting
distribution $f$ is determined by the tail of the
parent distribution $Q$. In this section, we compute the asymptotic behavior
of the value of the coding transition $\kappa^\ast$ for particular distrubutions belonging to each of the three
canonical classes \cite{satya}. We consider the classes in decreasing order of how fast the tail
of the parent distribution decays.
% Below we consider standard examples from the three classes, and find the
% asymptotic behavior of $\kappa^\ast$.
\vskip 5 pt
\noindent{\bf (i) Bounded distributions}\
Consider a bounded distribution of the form $Q(x) \sim (a-x)^{\beta-1}$,
$\beta>0$ with the support $0 \le x\le a$. This covers a broad range of
asymptotic behaviors that commonly occur as $x \to a$
\cite{satya}. In this case, the asymptotic distribution $f(z)$ of the
scaled maximum $z = b_n^{-1}(x_n -
a_n)$ is of the Weibull form
\[
f(z)\sim{(-z)}^{\beta-1} \exp\big(-(-z)^\beta\big)\
\]
for $a_n = a$ and $b_n = n^{-1/\beta}$.
% and $Q(x)=0$
% for $x > a$. Also, $\beta>0$.
Let $x$ denote second highest value among the $n$ terms, and let $x +
\epsilon$, $\epsilon > 0$, denote the highest value among the $n$
terms. Then the joint distribution of $(x,\epsilon)$ is given by
% The joint distribution of the second highest value $x$, and the difference
% between the highest and second highest values
% $\epsilon$ is given by
$n Q(x+\epsilon) P_{max}(x;n-1)$. In Appendix C, we show that this form of
the joint distribution implies that
%\red{What do you mean by the ``scaling form''?
%Need a formal definition here}.
the mean value of $\kappa^\ast = 2(1-\rho) =
2\langle\frac{\epsilon}{x_n+\epsilon}\rangle$, % using
% the above distribution we show in the Appendix C that,
and in the
large $n$ limit it is given by
\[
\kappa^\ast \sim n^{-1/\beta}.
\]
For % an exponential
a {uniform}
distribution $\beta=1$, and $\kappa^\ast$ is inversely proportional to
$n$.
The asymptotic decay
of the transition point is fastest for this class of distributions.
Comparison of the result with numerical data in \figurename{~\ref{scaling}a} reveals an excellent match.
\begin{figure}
\includegraphics[width=2.5in,height=1.8in]{scaling1}
\hskip 10 pt \includegraphics[width=2.1in,height=2.7in,angle=-90]{scaling2}
\caption{The scaling of the coding transition point as a function of total number of nutrients.
(a) The result for bounded distribution, with the predicted scaling shown in dashed lines. The symbols represent data points
obtained from numerical simulation. (b) The inverse of $\kappa^\ast$ plotted on log-linear scale. The curves have been
normalized by the value at $n=10^4$ for clear visibility in a single frame.
The theory predicts
a linear curve; but the convergence to the asymptotic logarithmic regime is extremely slow, and only the approach to the
regime can be detected in the plots.
}
\label{scaling}
\end{figure}
\vskip 5 pt
\noindent{\bf (ii) Unbounded distributions with light tails}~
Next, we consider the case where the support of $Q$ is unbounded; however, the
tail of $Q$ is faster than power law. We restrict ourselves to distributions of
the form $Q(x)\sim\exp(-x^\delta)$,with $\delta>0$~\cite{satya}, which
cover a broad class of distributions that
occur in natural settings, including exponential, Gaussian and stretched
exponential form~\cite{stretched}. In this case, the limiting
distribution of the scaled maximum $z = b_n^{-1}(x_n-a_n)$ is the Gumbel
distribution of the form
$f(z)=e^{-z-e^{-z}}$, and
\begin{eqnarray}
a_n&=& (\ln
n)^{\frac{1}{\delta}}+\frac{1}{\delta}(\frac{1}{\delta}-1)(\ln
n)^{\frac{1}{\delta}-1} \ln(\ln n) \\
b_n &=& \frac{1}{\delta}(\ln n)^{\frac{1}{\delta}-1}.
\label{ab}
\end{eqnarray}
%{Clearly write an expression for $f(z)$ and the scaling constants
% $a_n$ and $b_n$ for all the other cases if there are any. Done.}
% The maximum value is localized around $a_{n}$, and $b_n \to \infty$ if
% the distribution is slower
% than an exponential and $b_n \to 0$ if it is faster than an
% exponential. In this case,
% \[
% \bar{z} \approx {n}{}\int dx \int d\epsilon~\frac{\epsilon}{x}
% Q(x+\epsilon) P_{max}(x;n-1),
% \]
% and the joint distribution
% \[
% P(x,\epsilon) = n Q(x+\epsilon) P_{max} (x,N-1),
% \]
% where $\epsilon \ge 0$.
Using this limiting form, and a calculation similar to the bounded case (see
Appendix C for details), we get
\[
\kappa^\ast\sim {(\ln n)}^{-1}.
\]
In this case, the coding transition point goes to zero very slowly. Consequently,
even in a highly nutrient diverse environment, % we cannot be sure that
the organism % system is in the
may not sense and encode the environment.
% coding regime.
Comparison with numerical simulation data is shown in \figurename{~\ref{scaling}b}.
\vskip 5 pt
\noindent{\bf (iii) Power law distributions}~~
When $Q$ has a power law tail, we can show that $\kappa^\ast$ does not go
to zero in the large $n$ limit.
Consider $Q(x)$ of the form $Q(x) = \alpha e^{-x^{-\alpha}} x^{-(1+\alpha)}$,
$\alpha > 0$. In this case, the asymptotic distribution
$f(z) \sim \frac{\alpha}{z^{1+\alpha}} e^{-z^{-\alpha}}$ is the Fr{\'e}chet distribution, with $a_n=0$,
and
$b_n=n^\frac{1}{\alpha}$.
% \iffalse
% So, we have
% \begin{eqnarray}
% \bar{z} &=& \frac{n}{b_n} b_n^{1+\alpha}\int dx \int d\epsilon ~
% \frac{\epsilon}{x+\epsilon}
% \frac{e^{-(x+\epsilon)^{-\alpha}}}{(x+\epsilon)^{1+\alpha}}
% x^{-(1+\alpha)}
% e^{-({\frac{x}{b_n}})^{-\alpha}} \nn \\
% &=& {n} b_n^{\alpha} \int_0^\infty dx~x^{-(1+\alpha)}
% e^{-({\frac{x}{b_n}})^{-\alpha}} \int_0^\infty d\epsilon ~
% \frac{\epsilon e^{-(x+\epsilon)^{-\alpha}}}{(x+\epsilon)^{2+\alpha}} \nn \\
% &\approx& {n} b_n^{\alpha} \int_0^\infty dx~x^{-(1+\alpha)}
% e^{-({\frac{x}{b_n}})^{-\alpha}} \int_0^\infty d\epsilon ~
% \frac{\epsilon}{(x+\epsilon)^{2+\alpha}} \nn \\
% &=& {n} b_n^{\alpha} \int_0^\infty dx~x^{-(1+\alpha)}
% e^{-({\frac{x}{b_n}})^{-\alpha}} x^{-\alpha}
% \int_0^\infty dy ~
% \frac{y}{(1+y)^{2+\alpha}} \nn \\
% &\sim& {n} b_n^{\alpha} \int_0^\infty dx~x^{-(1+\alpha)}
% e^{-({\frac{x}{b_n}})^{-\alpha}} x^{-\alpha}\nn \\
% &=& {n} b_n^{\alpha} \int_0^\infty dx~x^{-(1+\alpha)}
% e^{-({\frac{x}{b_n}})^{-\alpha}} x^{-\alpha}\nn \\
% &=& {n} b_n^{-\alpha}\int_0^\infty dw~w^{-(1+\alpha)}
% e^{-({{w}})^{-\alpha}} w^{-\alpha}\\
% &=& \int_0^\infty dw~w^{-(1+\alpha)} e^{-({{w}})^{-\alpha}} w^{-\alpha}\\
% &\sim& O(1),
% \end{eqnarray}
% which is a contradiction.
% \fi
%We % can calculate the ratio of highest to second highest values, and
%show that in this case the ratio of the largest to the second
%largest value does
% that it does
%not go to $1$ as $n\rightarrow \infty$.
As before, let $x$ denote the
second largest value and $x+\epsilon$ denote the highest value. Then
% If the highest value is
% $x+\epsilon$, and the second highest is $x$, then we can show
% that in the limit of large $n$,
% \begin{eqnarray}
% \langle \frac{\epsilon}{x} \rangle &\sim& \frac{{n}^{\frac{2}{a}}}{a-1}
% ~\mbox{for $a > 1$} \nn \\
% &=& \infty ~\mbox{for $a \leq 1$.}
% \end{eqnarray}
% We show it as follows:
\begin{eqnarray}
{\langle \frac{\epsilon}{x} \rangle} &\sim& \frac{n^{\frac{2}{1+\alpha}}}{\alpha-1}, \mbox{~ $\alpha > 1$}\nn\\
&\sim& \infty, \mbox{~~~~~~~$\alpha \le 1$.}
\end{eqnarray}
The details of the calculation are in Appendix C.
For $n \rightarrow \infty$, we see that
% We see that in the large n limit,
the mean ratio of the highest to second highest values diverges, even for
$\alpha >1$ (where the mean of the
parent distribution is finite). We show in Appendix C that this implies $\kappa^\ast$ also diverges
asymptotically. The heavy tails of
the distributions cause the largest selected value to dominate over
all other values, and thus the organism is not in the coding regime.
\subsection{Discussion}
When % An organism in the regime
$\kappa>\kappa^\ast$, the organism senses the environment, % is engaged
% in a decision making process, and its
% response
encodes information about the environment, and takes an appropriate
response. Note that the response $i$ is a potentially a random map from
the environment state $\alpha$. This stochastic response is, often, critical for the
survival of the population. The value of $\kappa^\ast$ depends on the
distribution of environmental states and the free benefit associated the
particular response; therefore, the precise value of $\kappa^\ast$ cannot
be precisely calculated. However, universal
behavior emerges in the limit of a large number of environmental
states. The behavior falls into three different classes % s, which are
% consistent with the hierarchy of the behavior
as a function of the tail of the
distributions of the environment frequency $f_{\alpha}$ and the free
energy benefit $E_{\alpha}$. We show that only limiting behaviour of
$\kappa^\ast$ is a function of the product $g_{\alpha} = f_{\alpha}
E_{\alpha}$.
When $g_{\alpha}$ are chosen from a bounded distribution, the mean $\kappa^\ast$
goes to $0$ as $n^{-1/\beta}$ as a function of the number of environment
states $n$.
%In the ecology literature, s
%such organisms are often dubbed {\it generalists} \cite{generalists}, which can use a number of different nutrient sources for its metabolic needs.
In the case of unbounded distributions with light tails, the mean $\kappa^\ast$
decays to zero as $(\ln n)^{-1}$, and the mean $\kappa^\ast$ can diverge
for
% the parent distribution has finite moments,
% whereas it diverges for
power law % parent
distributions. The divergence is an artifact of the power law
distribution, but what it does imply is that there
is no typical $\kappa$ above which we can expect an organism to be in the
coding regime.
The slow convergence in the
case of exponential or Gaussian tails is more surprising,
since one would have expected
the organism to code for the environment % regime for a large number of
% nutrients
when the product is drawn from such distributions.
Which of these cases, if any, occur in natural environments is a difficult
question to answer. We may expect that in highly
biodiverse environments, such as in soil microbial communities, the number
of available nutrients is high.
Such communities harbor hundreds of different species of microbes, and
their interactions
often promote chemical diversity \cite{metabolites}.
A thorough quantitative study of nutrient distributions in such
communities would
be helpful in making predictions about the coding behavior of metabolic systems.
\section{Conclusion}
We have formulated the problem of adaptive response in biological systems
as an optimization problem where the tradeoff is
between the benefit obtained from specific responses and the cost of
producing them. Our main result
is that there is a cascade of responses as a function of the cost of
information, and the number of
responses that are active in an optimal metabolic system depends on the
cost and the distribution of nutrients.
Even when the system is capable of producing $n$ different responses, the
number of responses it actually uses can be significantly smaller.
The decision to express certain genes is executed by the regulatory
machinery, and evolutionary processes may lead
to machinery that produces the optimal response when supplied with a
certain environment.
The traditional
view has been that in order to metabolize new nutrients, genotypic changes
that produce new or altered enzymes are required.
However, our model points to the
existence of
epigenetic mechanisms that carry out a decision theoretic task -- that of
constructing the optimal response channel
from within the repertoire of responses already available.
%Recent years have seen an explosion in the field of epigenetics, and indeed regulatory mechanisms that cause switches
%between metabolic phenotypes are now known in some detail
%For example, a prion protein like element converts wild fungi from metabolic specialists
%to generalists \cite{fungi}, pathogenic fungi can switch from biotrophic to necrotrophic lifetyles \cite{biotrophic}, etc.
Our model is based on a particular way of looking at such regulatory
mechanisms -- namely that they % evolved to solve a class
optimal solutions of a particular % problems
in rate distortion problem.
In this % present
paper, we only characterize the structure of the optimal channe. We % have
% made no hypotheses
do not describe how these regulatory
mechanisms evolve, or indeed about the kind of molecular processes
necessary to establish the optimal channel.
More detailed studies are required to address these questions. Further
investigations may bring out new surprises,
but we believe that some of the central features of our results will
continue to hold.
% \iffalse
% We have seen that as the number of nutrients in the system grows, the
% coding transition point shows three different kinds of
% behavior. Only in one case -- where the expected benefit values are
% drawn from a bounded distribution -- the transition
% point goes to zero quickly as the number of nutrients grows. It is
% reasonable to ask what this implies about molecular coding
% in the metabolic world.
% Strictly speaking, the results here apply if the number of nutrients
% available to a cell is sufficiently large.
% However, in the case where the free energy benefits are drawn from a
% bounded distribution, we see that the asymptotic behavior
% already holds for $n \sim 10$ (see \gurename{ \ref{scaling}}a). When the
% benefits are drawn from an unbounded
% distribution, the approach to zero either does not happen or is
% logarithmically slow, and we cannot conclude that
% the cell is generically in the coding regime.
% It is likely that environments corresponding to all the different
% regimes exist. Extreme metabolic specialists may inhabit
% environments where the beneifts follow the power law distribution, such
% that a single nutrient dominates the scenarion.
% \fi
\section{Appendix}
% \iffalse
% \section{Why Cascades?}
% \subsection{Diagonal Benefits Only}
% Why are the responses turned on one at a time? First, define $g_\alpha=f_\alpha E_\alpha$.
% %Assume that a bifurcation happens at some $\kappa^\ast$, with two alternative scenarios: \\
% Then, the Lagrangian to be maximized is
% \[
% \mathcal{L} = G- \kappa^{-1} I +\sum_\alpha \lambda_\alpha \sum_i p(i|\alpha).
% \]
% Equating the derivative of this to zero, we have
% \begin{equation}
% p(i|\alpha) = p(i) \exp[\kappa (\frac{\lambda_\alpha}{f_\alpha}+\frac{g_i}{f_i} \delta_{i,\alpha})].
% \label{casc0}
% \end{equation}
% Firstly, we will notice that $\sum_i p(i|\alpha)=1$, which, from the above equation, implies that
% \begin{equation}
% \lambda_\alpha=\frac{f_\alpha}{\kappa} \log \frac{1}{1+p_\alpha(e^{\kappa \frac{g_\alpha}{f_\alpha}}-1)}.
% \end{equation}
% Clearly, $\lambda_\alpha \leq 0$. As the transition that switches on $\alpha$ is approached from above, we have $p(\alpha) \to 0^+$, and consequently
% $\lambda_\alpha \to 0^-$.
% From the normalization condition $\sum p(i|\alpha) f_\alpha = p(i)$, and for $p(i)>0$, the above equation implies,
% \begin{equation}
% \sum_{\alpha\ne i}f_\alpha \exp[\kappa (\frac{\lambda_\alpha}{f_\alpha})]+
% f_i \exp[\kappa (\frac{\lambda_i}{f_i}+\frac{g_i}{f_i})]=1
% \label{casc2}
% \end{equation}
% Precisely at the transition point for the response $i$, we have, for $i$,
% \begin{equation}
% \sum_{\alpha\ne i}f_\alpha \exp[\kappa (\frac{\lambda_\alpha}{f_\alpha})]+
% f_i \exp[\kappa \frac{g_i}{f_i}]=1. \label{casc1}
% \end{equation}
% If the responses $i$ and $j$ are turned on simultaneously, i.e at the same value of $\kappa$, then $j$ will also satisfy
% a similar equation simultaneously. This will imply,
% \begin{equation}
% f_i e^{\kappa \frac{g_i}{f_i}}+f_j=f_i+f_j e^{\kappa \frac{g_j}{f_j}}.
% \end{equation}
% It is easily shown that the above equation has at most one solution for
% $\kappa$. Moreover, the solution is determined
% solely by $f_i,g_i,f_j,g_j$, wheareas the true value of $\kappa$
% will be determined by the entire set of frequencies and benefits. This
% clearly shows that simultaneous transitions cannot be
% generically expected. However, in some special cases they might
% hold. The equation above gives us a necessary condition for
% simultaneous transition to two responses. A sufficient condition is
% $f_i=f_j$ and $g_i=g_j$; sufficiency is apparent from the
% fact that {\it any} $\kappa$ that satisfies Eq. \ref{casc1} also
% satisfies the corresponding equation for $j$.
% %These two facts imply that the new optimal channel for any
% %$\kappa \le \kappa_{new}^\ast$ is identical to that of the unperturbed problem.
% %Based on this, we see that $\kappa_{new}^\ast$ should satisfy Eq. \ref{casca} with the exact same parameters,
% %and therefore $\kappa_j^\ast=\kappa_{new}^\ast=\kappa^\ast$. This violates the condition that for generic perturbations,
% %$\kappa_j^\ast \ne k^\ast$, indicating that a joint transition is not immune to small perturbations,
% %and requires fine-tuning.
% \iffalse
% Now, consider the case $\kappa_{new}^\ast \le \kappa$.
% In this case, from the argument above, we see that equation \ref{casca} for $\kappa_i^\ast$ is identical to the unperturbed case,
% and therefore we have $\kappa_i^\ast=\kappa_{new}^\ast=\kappa^\ast$. On the other hand, consider the case
% $\kappa_{new}^\ast \ge \kappa^\ast$; here, once again, the transition point for $i$ must still be $\kappa^\ast$,
% since the
% solution over the index set $\{A,i\}$ at $\kappa^\ast$ does not depend on $\{E_{j,\alpha}\}$. Therefore,
% $\kappa_i^\ast =\kappa^\ast = \kappa_{new}^\ast$. Thus, in either case, a joint transition is possible only if
% $\kappa_{new}^\ast=\kappa^\ast$, which will not generically happen, since an arbitrary small perturbation to the parameters
% will change the transition points for $j$.
% \fi
% \iffalse
% 1) The $i-th$ and $j-th$ responses are turned on, i.e $p_i,p_j\gtrsim 0$
% when $\kappa \gtrsim \kappa^\ast$.
% Without loss of generality, $g_i \geq g_j$. Further, define
% $\epsilon_1(\alpha)=p(i|\alpha),\epsilon_2(\alpha)=p(j|\alpha)$.
% \\
% 2) Only the $i-th$ response is turned on, ie $p_i\gtrsim 0$ and $p_j=0$
% for $\kappa \gtrsim \kappa^\ast$. However, we now
% assign $p(i|\alpha)=\epsilon_1(\alpha)+\epsilon_2(\alpha)$, while the
% conditional probabilities for all $l \neq i,j$ are
% left the same as in scenario 1 above.
% The difference in benefit between scenario 2 and scenario 1 is
% \[
% g_i(\epsilon_1(i)+\epsilon_2(i))-(g_i \epsilon_1 (i) + g_j \epsilon_2
% (j))=(g_i-g_j) \epsilon_2(j) \geq 0.
% \]
% Consider a distribution where $f_\alpha=1/n$. Let the benefits in
% decreasing order be $E_1>E_2>\ldots>E_{n-1}>E_n$.
% The transitions will occur precisely in this order, because the
% probabilities are identical and the cost
% does not distinguish between alternatives.
% The $m$-th transition, which activates the $m+1$-th response, satisfies
% the condition:
% \begin{equation}
% e^{\kappa E_{m+1}} = 1+ \sum_{\alpha=1}^m \frac{p_\alpha(e^{\kappa
% E_\alpha}-1)}{1+p_\alpha(e^{\kappa E_\alpha}-1)}
% \end{equation}
% The equation is difficult to solve exactly since the $p_\alpha$'s have
% to be known at the transition point. But in some
% special cases we can predict the transitions. Let us say that our
% benefits are $E_\alpha=1-\frac{\alpha-1}{n}$, and $n$
% is large. For $m\sim O(1)$, we can show that the $m+1$-th transition
% happens approximately at $\kappa_m=m/n$, whereas for
% for large $m$ the transitions happen at $\kappa_m = \frac{\log
% (m+1)}{1-\frac{m}{n}}$.
% \fi
% \fi
\subsection{Algorithm for finding optimal channel}
Here we consider the problem of computing the optimal $p(i\vert \alpha)$
for $\kappa \neq 0$. Let
\begin{eqnarray*}
J(p,q) & = & \sum_{i\alpha} f_{\alpha} p(i\vert \alpha) \log \left(
\frac{p(i\vert \alpha)}{q(i)} \right)\\
& = & I(p) + D(p\|q),
\end{eqnarray*}
where $D(p,q)$ denotes the relative entropy of $p$ and $q$.
Then, $J(p,q)$ is jointly convex in $(p,q)$ and moreover, $\min_{q} J(p,q)
= I(p) + \min_q D(p,q) = I(p)$.
Since $G(p) - \kappa^{-1} J(p,q)$ is jointly concave in $(p,q)$, and
$\max_{(p,q)} G(p) - \kappa^{-1} J(p,q) = \max_p G(p) - \kappa^{-1} I(p)$,
it follows
that the coordinate descent will converge to the optimal solution. Thus,
the following algorithm will compute the optimal solution for
$\kappa>0$:
\begin{enumerate}[1.]
\item Set $q^0 = \frac{1}{m}$ where $m$ is number of responses.
\item % Compute the solution of the optimization problem
For $k \geq 0$, set
\[
p^{k} \leftarrow \argmax_{p} \{ G(p) - \kappa^{-1} J(p,q^k)\}
\]
It is easy to check that $p^k(i\vert \alpha) \propto q^k(i) e^{\kappa
E_{i\alpha}}$. Thus, $p^k(i\vert \alpha) = \frac{q^k(i) e^{\kappa
E_{i\alpha}}}{\sum_j q^k(j) e^{\kappa E_{j\alpha}}}$
\item % Compute the solution of the optimization problem
For $k \geq 0$, set
\[
q^{k+1} \leftarrow \max_{q} \{ G(p^k) - \kappa^{-1} J(p^k,q)\}.
\]
It is easy to check $q^{k+1}(i) = \sum_{\alpha} f_{\alpha} p^k(i\vert \alpha)$.
% \item Take $q(i)=q^\ast(i)$, $p(i \vert \alpha)= p^\ast(i\vert \alpha) $
% and go to step 1.
\end{enumerate}
This algorithm converges to the optimum solution.
\subsection{Product of two random variables}
Suppose $f_i$, $i = 1, \ldots, n$ are IID samples from $P_f$ and $E_i$, $i
= 1, \ldots, n$, are IID samples from $P_e$. We assume that $P_e$ is
supported over $[0,a]$ with $a < \infty$.
Let $g_i = f_i E_i$, $i = 1,
\ldots, n$, and let its distribution be denoted by $P_g$. Permute indices such that $g_1 = f_1E_1$ is the largest value
among $\{g_i\}$ and $g_2 = f_2E_2$ is the second larges value among
$\{g_i\}$. We will restrict ourselves to distributions such that $g_2/g_1
\rightarrow 1$ as $n \rightarrow \infty$.
Recall that $s_i = \frac{E_i}{E_1}$ and $s_iw_i = \frac{g_i}{g_1} =
\frac{f_iE_i}{f_1E_1} \leq 1$. Note that $s_2w_2 \rightarrow 1$ as $n
\rightarrow \infty$.
% There are $n$ such equations.
Let $x_i>1$ denote the non-trivial solution of \eqref{eq:wi} corresponding
to $w_i$, $i = 2, \ldots, n$. Let $y_i = x_i -1$.
% Let the smallest value in the set
% $\{y_i\}$ be
% $y_t$. We restrict ourselves to parent distributions
% such that $y_t \ll 1$ for large $n$. \red{Is it clear, apriori, that such
% distributions exist?} Our goal is to find $y_t$ for large
% $n$. % Consider any $y_i \ll 1$ (this includes the case
% $y_t$). \red{Why does such a $y_i$ exist?}
Substituting $1+y_i$ in \eqref{eq:wi}, and expanding up to second order in
$y_i$, we obtain the solution
\begin{equation}
y_i= \frac{2(1-s_i w_i)}{1-\frac{1}{2} s_i w_i (1-s_i)}.
\label{eq:smally}
\end{equation}
This solution is valid provided $y_i$ is small in the limit of large
$n$. Since $s_i \geq 0$ and $s_iw_i \leq 1$, it follows that
\[
y_i \leq 4(1-s_iw_i).
\]
Since $s_2w_2 \rightarrow 1$, it follows that $y_2 \rightarrow 0$. Since
the coding transition is determined by the smallest value of $x_i$, $i
\geq 2$, in the limit of large $n$, one can limit to indices $i$ with $y_i
\ll 1$.
% Since $y_i$ is small (which was used to do the expansion above), the above
% expression must be small.
Let $A_i = 2(1-s_iw_i)$ denote the numerator and $B_i = 1 -
\frac{1}{2}s_iw_i(1-s_i)$ denote the denominator in the expression of
$y_i$. Then we have that $0 \leq A_i \leq 2$ and $\frac{1}{2} \leq B_i \leq 1 +
\frac{1}{2}s_i$. We show below that there exists a bound $S$ such that, in
the limit of large $n$, $s_i \leq S$ for all $i \geq 2$. Thus, $B_i \leq
1 + S/2$ in the limit of large
$n$. Thus, $y_i \ll 1$ only if $s_iw_i \lesssim 1$. We also show that $s_i
\lesssim 1$ for all $i$ such that $s_iw_i \lesssim 1$. Thus, it follows
that $y_i \approx 2(1-s_iw_i)$.
% Notice that the numerator N is bounded above by
% $N=2$ (since $s_i w_i \le 1$), and the denominator $D$ is bounded below
% by $1/2$ (since $s_i w_i \le 1$ and $s_i > 0$). Thus,
% $y_i$ is small in one of three ways: either $N$ is small but $D$ is not
% large, or $D$ is large but $N$ is not small,
% or both $N$ is small and $D$
% is large. (The case where both $N$ and $D$ vanish or both diverge but
% their ratio is small is ruled out, since $N$ is bounded
% above and $D$ is bounded below).
% We notice that
% \[
% D=1+\frac{1}{2}(s_i w_i) s_i-\frac{1}{2}s_i w_i \leq 1+\frac{1}{2}(s_i
% w_i) s_i \leq
% 1+\frac{1}{2}s_i.\]
% At this point, we introduce a condition: $E$ is chosen from a bounded distribution.
% We show below that when this is the case, $s_i \not \gg 1$, and
% therefore $D$ is not large. We are then left with the only
% option that $N$ must be small, which means that $s_i w_i \lesssim 1$. This condition holds for all $y_i \ll 1$.
%From this, we
%see that $y_t$ corresponds to the highest value of $s_i w_i$.
% We show one further thing below, which is that when $E$ is bounded, then
% for all $i$ such that
% $s_i w_i \lesssim 1$, we must have $s_i \lesssim 1$. Then we have
% $y_i = (1-s_i w_i)$ (which holds for all $y_i \ll 1$). From this, it becomes clear that
% $y_t$, which is the smallest $y_i$, corresponds to the largest value of $s_i w_i$.
% We need to complete the proof by showing that when $E_i$ is bounded, the
% following two statements hold:\\
\begin{lemma} Suppose the distribution $P_e$ is supported over the
bounded set $[0,a]$. Then the following holds:
\begin{enumerate}
\item there exists a bound $S$ such that $s_i =
\frac{f_i}{f_1} \leq S$ for all $i \geq 2$ and sufficiently large $n$.
\item Suppose $s_i w_i \lesssim 1$. Then we must have $s_i \lesssim 1$.\\
\end{enumerate}
\end{lemma}
% \iffalse
% Consider the two random variables $f$ and $E$ with distributions $P_f(f)$
% and $P_e(E)$. The product variable is
% $g=fE$ with distribution
% \[
% P_g(g)=\int f^{-1} P_f(f) P_e(f^{-1}g) ~df.
% \]
% Let $n$ values of $f$ and $E$ be drawn from their distributions, giving us
% $n$ values of $g$, which follow the above distribution.
% Let $g_1$ denote the highest value of $g$, and $g_2$ denote the second
% highest value. Let $g_i=f_i E_i$, $i = 1, 2$. % where $f_a$ and $E_a$ are
% % the choices of $f$ and $E$ thatgive us $g_1$. Likewise, let $g_2=f_b E_b$.
% We limit ourselves to distributions where $g_2/g_1 \lesssim 1$ in the
% limit of large $n$.
% We show that when the distribution of $E$ is bounded, i.e when $P_e$ is
% supported on $[0,a]$ for some $a>0$ that
% % Then \eqref{eq:s-cond} holds provided
% \begin{eqnarray*}
% s_i & = & \frac{E_i}{E_1} \not \gg 1, \qquad \forall i = 1, \ldots, n,\\
% s_2 & \approx & 1
% \end{eqnarray*}
% % $ % does not hold for any
% % for all $i$, and (2) $s_2 = \frac{E_2}{E_1}
% % \approx 1$,
% in the limit of large $n$.
% % The statement in the main text, $s \approx 1$, is equivalent to the
% % statement that $E_b/E_a \approx 1$ in the limit of large $n$ .
% % We show that these codnitions holds when the distribution of $E$ is bounded,
% % % i.e. is chosen from a bounded distribution, i.e.
% % .
% % Since $sw < 1$, $s \gg 1/sw$ cannot since this would violate the condition
% % that $s \not \gg 1$. Thus, \eqref{eq:s_gg_1} cannot hold.
% \fi
% First, recall that $g_i = f_i E_i$.
% Let the higest value of $g$ be $g_1=f_1E_1$.
% The values $E_i$ are chosen from $P_e(E)$ which has a finite support,
% i.e $P_e(E)$ is $0$ unless $0 \leq E \leq a$.
% The $f_i$ are chosen from $P_f(f)$.
\noindent We break our proof down
into two cases.
\texitem{(1)} $P_f$ is bounded with support on $[0,c]$. \vskip 5pt
% \noindent{\it Proof of statement (1)} \vskip 5 pt
\noindent In this case,
$P_g$ is supported on $[0,ac]$. Since $P_g$ has finite support,
it belongs to the Weibull domain, by definition. In this case, the
distribution of $g_1$ is $b_n^{-1} h(\frac{g_1-ac}{b_n})$, where
\[
h(z)\sim{(-z)}^{\beta-1} \exp\big(-(-z)^\beta\big)\
\]
and $b_n = n^{-1/\beta} \ll a$. This distribution is concentrated at
$g_1 \lesssim ac$.
%Furthermore, for values drawn from a
%Weibull domain distribution, $g_2 \lesssim g_1$.
% Consider the particular
% form $P_e(E) \sim (a-E)^{\beta-1}$,
% $\beta>0$ with the support $0 \le f\le a$. (\red {This form is probably
% not essential. But we can discuss that if you find the
% proof of this special case acceptable.})
% and $Q(x)=0$
% for $x > a$. Also, $\beta>0$.
Since $f_1 \leq c$ and $E_1 \leq a$, % we see that their
the product $g_1 = f_1 E_1
\lesssim ac$ if, and only if,
$f_1 \lesssim c$ and $E_1 \lesssim a$. The latter condition means that
$E_i/E_1 \leq a/E_1 \rightarrow 1$ as $n \to \infty$. Thus, we are
guaranteed that the first statement in the Lemma holds for any $S>1$.
%does not hold for any $i$, i.e $s_i \not \gg 1$. \\ \vskip 2 pt
% \noindent {\it Proof of statement (2)} \\ \vskip 2 pt
Since the denominator in \eqref{eq:smally} is bounded, $y_i$ is
small only if $s_i w_i \lesssim 1$, or equivalently, $g_i \lesssim g_1$.
Since $g_1 \lesssim ac$, it follows that
% Furthermore, for values drawn from a
% Weibull domain distribution, $g_2 \lesssim g_1$,
% implying that
$g_i \lesssim ac$. This holds only if $f_i \lesssim c$ and $E_i
\lesssim a$, i.e. $s_i = E_i/E_1 \approx 1$. This establishes the second
statement in the Lemma.
\vskip 5 pt
\texitem{(2)} $P_f$ % is an
is unbounded % distribution
with a light tail. \vskip 5pt
\noindent We consider $P_f(f) \sim \exp(-f^\delta)$
with $\delta>0$. Thus,
\begin{eqnarray*}
\lefteqn{P(g)}\\
& = & \int_{a^{-1}g}^\infty f^{-1} P_f(f) P_e(f^{-1} g)\ dx \\
& = & \int_0^\infty
(a^{-1}g+\epsilon)^{-1}P_f(a^{-1}g+\epsilon)
P_e\big(\frac{g}{a^{-1}g+\epsilon}\big)~d\epsilon
\end{eqnarray*}
We are interested in the tail of the distribution, i.e. $P(g)$ for $g \gg
1$.
%\red{What about other unbounded distributions? What happens in those
% cases?}
For large
$g$, % we can make the following approximation in the integrand:
$P_f(a^{-1}z+\epsilon) \sim \exp(-(a^{-1}g)^\delta)\exp(-\frac{\delta
\epsilon}{a^{-1}g})$. Then,
\begin{eqnarray*}
&&P(g) \nn \\
&\approx& e^{-(a^{-1}g)^\delta} \int_0^\infty \frac{\exp(-\frac{\delta
\epsilon}{a^{-1}g})}{1+\frac{\epsilon}{a^{-1}g}}P_e\big(\frac{a}{1+\frac{\epsilon}{a^{-1}g}}\big)
~\frac{d\epsilon}{a^{-1}g}\\
&\approx& P_f(a^{-1}g) \int_0^\infty \frac{\exp{(-\delta
\epsilon^\prime)}}{1+\epsilon^\prime}P_e\big(\frac{a}{1+\epsilon^\prime}\big) ~d\epsilon^\prime \nn
\\
&\sim& P_f(a^{-1}g).\nn \\
\end{eqnarray*}
Thus, the tail of $P_g$ follows the same form as $P_f$, except
that $g$ is now scaled by $a^{-1}$.
The distribution $P_f(f)$ belongs to
the Gumbel domain. The distribution of the extreme value $f_{max}$ is $d_n^{-1}
h(\frac{f_{max}-c_n}{d_n})$, where $h(z)\sim \exp(-z-e^{-z})$,
$c_n \sim (\ln(n))^{1/\delta}$, and $d_n \sim (\ln(n))^{1/\delta-1} \ll
c_n$. Therefore, $f_{max}$ is localized around $c_n$, i.e $f_{max}/c_n \approx
1$.
% For the parent distribution
The distribution $P(g) \sim P_f(a^{-1}f)$, so we have $g_1/(a c_n) \approx 1$.
Thus, $E_1 \approx (a c_n)/f_1$; however, since % the distribution of the
% maximum value of $f$ is localized around
$f_1 \leq f_{max}$, and the maximum value of $E_1$ is $a$, this equality can be true
only if $c_n/f_1 \approx 1$ and $E_1 \approx a$. The latter condition
implies that $E_i/E_1 \leq a/E_1 \to 1$ as $n \to \infty$. Thus, any $S>1$
satisfies the first statement in the Lemma.
% \\ \vskip 5 pt
% %does not hold for any $i$, ie $s_i \not \gg 1$. \\ \vskip 2 pt
% %For the Gumbel domain, we have
% %$g_2/g_1 \to 1$.
% \noindent {\it Proof of statement (2)} \\
% \noindent
Since the denominator in \eqref{eq:smally} is bounded, $y_i$ is
small only if $s_i w_i \lesssim 1$, or equivalently, $g_i \lesssim g_1$.
% For small $y_i$ we must have $s_i w_i \lesssim 1$, which implies $g_i
% \lesssim g_1$, and therefore
Thus,
$g_i/(a c _n) \approx 1$; consequently, $E_i
\approx (a c_n)/f_1$, and because $f_1/c_n \approx 1$, we have
$E_i \approx a$. % Comparing with $E_a$, we see that
Thus, it follows that $s_i = E_i/E_1 \approx 1$.
% In either of the cases above
%In the two cases above, we did not prove the statement that
%$g_2/g_1 \lesssim 1$, but this follows from
%the % calculation
%argument similar to those used to establish the asymptotic scaling of
%$\kappa^\ast$ in the main text.
% \subsection{Product of two random variables}
% Consider two random variables $x,y$ with joint distribution $P_1(x)
% P_2(y)$. Define $z=xy$. Then,
% \begin{eqnarray*}
% P(z,x)&=& x^{-1} P_1(x) P_2(x^{-1}z) \nn \\
% P(z,y)&=& y^{-1}P_2(y) P_1(y^{-1} z)\nn \\
% P(z) &=&\int x^{-1} P_1(x) P_2(x^{-1}z) \\
% & = & \int y^{-1}P_2(y) P_1(y^{-1}
% z)
% \end{eqnarray*}
% Without loss of generality, we consider the tail of $P_1(x)$ to be as slow
% or slower than $P_2(y)$. Let us now consider the various scenarios for
% $P_1$ and $P_2$.
% \vskip 5pt
% \texitem{(1)} $x$ is supported on $[0,c_1)$, and $y$ on $[0,c_2]$, and
% so $P(z)$ is supported on $[0,c_1 c_2)$.
% \noindent From this it is clear
% that $P(z)$ belongs to the Weibull domain, so at large $n$, $a_n^z \to c_1
% c_2$. \red{Why is it obvious that $P(z)$ belongs to the Weibull domain?
% What is $a^z_n$?}
% So, the extreme value of $z$
% gets contributions jointly
% from the extreme value of $x$ (i.e, $c_1$) and that of $y$ (i.e $c_2$).
% Thus, clearly, the option (2) above is
% ruled out.
% \red{What is ``option (2) above''? You should write this explicitly
% here. Also, why does it follow from $a_n^z \to c_1
% c_2$ that one must have contributions from both extreme values?}
% Consider the special case where $P_1$ and $P_2$ are both uniform. % In
% % that case,
% % we find
% Then
% \[
% P(z,x)= \frac{(c_1 c_2)^{-1}}{x} \Theta (x-\frac{z}{c_2}),
% \]
% from which we obtain
% \[
% P(z)\sim \ln(\frac{c_1 c_2}{z})
% \]
% with $z\leq c_1 c_2$. The maximum value of $z$, $a_n^z \to c_1 c_2$. Also,
% \[
% P(x|z) \sim \frac{(c_1 c_2)^{-1}}{x \ln (\frac{c_1 c_2}{z})}
% \Theta(x-\frac{z}{c_2}).
% \]
% Note that when $z=c_1 c_2 -\epsilon$, the Heaviside function puts the
% restriction that $x> c_1-\frac{\epsilon}{c_2}$, and thus
% as $z \to c_1 c_2$, we must have $x \to c_1$. By the same arguments, as $z
% \to c_1 c_2$, $y \to c_2$.
% \red{How can we conclude from here that $s_i = \frac{f_i E_i}{f_r W_r}
% \rightarrow 1$? Not clear at all!}
% \vskip 5pt
% \texitem{2.} $P_2$ is supported on $[0,c]$, and $P_1$ is unbounded.
% \noindent In this case, $P(z)$ is unbounded. The joint distribution $
% P(z,x) = x^{-1} P_1(x) P_2(x^{-1} z)$.
% Thus,
% \begin{eqnarray*}
% \lefteqn{P(z)}\\
% & = & \int_{c^{-1}z}^\infty x^{-1} P_1(x) P_2(x^{-1} z)\ dx \\
% & = & \int_0^\infty
% (c^{-1}z+\epsilon)^{-1}P_1(c^{-1}z+\epsilon)
% P_2\big((c^{-1}z+\epsilon)^{-1}z\big)~d\epsilon
% \end{eqnarray*}
% We are interested in the tail of the distribution, i.e. $P(z)$ for $z \gg
% 1$. Consider
% the form $P_1(x)\sim \exp(-x^\delta)$.
% \red{What about other unbounded distributions? What happens in those
% cases?}
% For large
% $z$, % we can make the following approximation in the integrand:
% $P_1(c^{-1}z+\epsilon) \sim \exp(-(c^{-1}z)^\delta)\exp(-\frac{\delta
% \epsilon}{c^{-1}z})$. Then
% \begin{eqnarray*}
% P(z)&\sim& e^{-(c^{-1}z)^\delta} \int_0^\infty \frac{\exp(-\frac{\delta
% \epsilon}{c^{-1}z})}{1+\frac{\epsilon}{c^{-1}z}}
% ~\frac{d\epsilon}{c^{-1}z}\\
% &\sim& P_1(c^{-1}z) \int_0^\infty \frac{\exp{(-\delta \epsilon^\prime)}}{1+\epsilon^\prime} ~d\epsilon^\prime \nn \\
% &\sim& P_1(c^{-1}z).\nn \\
% \end{eqnarray*}
% Thus, the tail of the distribution follows the same form as $P_1$, except
% that $z$ is now scaled as $c$, and therefore
% $a_n \to c a_n$, i.e the largest value of $z$ gets contribution from
% $y\lesssim c$.
% \red{How can we conclude from here that $s_i = \frac{f_i E_i}{f_r W_r}
% \rightarrow 1$? Not clear at all!}
% \iffalse
% {\bf 2.} $P_2$ is supported on $[0,c]$, and $P_1$ is unbounded.
% In this case, $P(z)$ is unbounded. The joint distribution is
% \[
% P(z,x) \sim \frac{1}{x} P_1(x) P_2(\frac{z}{x})
% \]
% Clearly, the conditional distribution of $z$ is $P(z|x)=\frac{1}{x}
% P_2(\frac{z}{x})$. If $n$ pairs of $(z,x)$ are generated
% from the joint distribution, let the largest $x$ be $x_{max} \sim a_n$,
% and the largest $z$ be $z_{max}$. Let the former belong
% to the pair $(z_m,x_{max})$, and the latter to $(z_{max},x_m)$. We will
% show that $x_m,z_m,z_{max} \sim O(x_{max})$.
% We have,
% \[
% P(z|x_{max})\sim P(z|a_n)=\frac{1}{a_n} P_2(\frac{z}{a_n}).
% \]
% Now consider the probability that $z_m$ is less than $d_n$.%, where
% $d_n/a_n \to 0$ as $n \to \infty$.
% \begin{eqnarray}
% P(z_m< d_n) = \int_0^{d_n}\frac{dz}{a_n} P_2(\frac{z}{a_n})\nn \\
% = \int_0^{\frac{d_n}{a_n}} P_2(w)~ dw,
% \end{eqnarray}
% At large $n$, the integral vanishes unless $d_n$ is at least
% $O(a_n)$. Further, we must have
% $z_m<c a_n$, and so $d_n$ is also at most $O(a_n)$.
% Thus, $z_m \sim O(a_n)$. Further, $z_{max} < c x_{max}$, so
% $z_{max}$ can be at most $O(a_n)$. Thus, we have $z_{max} \sim z_m$,
% and therefore $x_m \sim x_{max}$. All four quantities
% are $O(a_n)$.
% This also tells us that $y_m=\frac{z_{max}}{x_m} \sim O(1)$. Since
% $y_{max}=c$, we have $y_m \sim y_{max}$.
% Then our arguments imply that $E_r \sim E_{max}$, and so it is not possible to have $E_i \gg E_r$ for any $i$.
% \fi
\subsection{$\kappa^\ast$ in different asymptotic regimes}
Suppose $X_i$, $i = 1, \ldots, n$, denote $n$ IID draws from the
density $Q$. Let $x$ denote the \emph{second} largest of the $n$ terms, and
let $x+\epsilon$ denote the largest term of the $n$ terms. Then the joint
density $P(x,\epsilon)$ of $(x,\epsilon)$ is given by $n
P(x,\epsilon) = Q(x+\epsilon)P_{\max}(x;n-1)$, where $P_{\max}(x;n-1)$ denotes the density
of the largest of $n-1$ draws from the distribution $Q$. We analyze the
behavior of the mean $\langle \kappa^\ast \rangle$ for three different
asymptotic regimes.
\noindent {\bf Bounded distribution:}\\
\begin{eqnarray}
\lefteqn{ \langle \kappa^\ast\rangle = 2 \Big\langle
\frac{\epsilon}{x+\epsilon} \Big\rangle} \nn \\
&\sim& {n}{}\int dx \int d\epsilon~\frac{\epsilon}{x+\epsilon} Q(x+\epsilon) P_{max}(x;n-1)\nn \\
&\approx& \frac{n}{a}\int dx \int d\epsilon~{\epsilon} Q(x+\epsilon) P_{max}(x;n-1)\label{case2:1} \\
&=& \frac{n}{a b_n}\int_0^a dx \int_0^{a-x} d\epsilon~{\epsilon}
(a-x-\epsilon)^{\beta-1} e^{-{(\frac{a-x}{b_n})}^\beta} \nn \\
&=& \frac{n}{a b_n} b_n^\beta \int_0^{a/b_n} dz \int_0^{b_n z} d\epsilon~ \epsilon {\big(z-\frac{\epsilon}{b_n}\big)}^{\beta-1}
z^{{\beta}-1} e^{-z^\beta} \nn \\
&& \mbox{(with the transformation $z=\frac{a-x}{b_n}$)} \nn \\
%&& \text{\red{ Did you also
% take a limit in $n$?}} \nn\\
&=& \frac{n}{a}~ {b_n}^{\beta+1} \int_0^{a/b_n} dz \int_0^z dy ~ y (z-y)^{\beta-1} z^{\beta-1} e^{-z^\beta} \nn \\
&& \mbox{(with the transformation $y=\frac{\epsilon}{b_n}$)} \nn \\
&\approx& \frac{n}{a}~ {b_n}^{\beta+1} \int_0^\infty dz \int_0^z dy ~y(z-y)^{\beta-1} z^{\beta-1} e^{-z^\beta} \label{case2:2} \\
&\sim& n^{-\frac{1}{\beta}}, \label{case2:3}
\end{eqnarray}
where \eqref{case2:1} follows from the fact that the integrand is non-zero only if $x\lesssim z$, \eqref{case2:2}
follows from the fact that $b_n \to 0$ and the integrand vanishes at large
$z$, and \eqref{case2:3} follows
from $b_n \sim n^{-\frac{1}{\beta}}$.
\\
\noindent {\bf Unbounded distribution with light tails:}\\
\noindent With the scaling transformation $z=\frac{x-a_n}{b_n}$, we have
\[
P(z,\epsilon)=n Q(a_n + b_n z +\epsilon) f(z)
\]
\begin{eqnarray*}
\lefteqn{\langle \kappa^\ast \rangle} \nn\\
& = & \int dz~\int ~ d\epsilon \frac{2\epsilon}{a_n+b_n
z+\epsilon} e^{-{(a_n
+ b_n z + \epsilon)}^\delta-z-e^{-z}}\nn \\
\end{eqnarray*}
Since $a_n \gg 1$, and $a_n \gg b_n$, the integrand has
finite values only when $z \sim O(1)$, and
$\epsilon \ll a_n + b_nz$. Expanding the first term in the exponent as
${(a_n+\epsilon)}^\delta +\delta(a_n+\epsilon)^{\delta-1}b_n z$, and using
the explicit forms of the scaling constants above, and the fact that
$\epsilon \ll a_n$, it follows that the second term is $~c z$, where $c$
is independent of $n$. Thus, the integral over $z$ yields
\[
\langle \kappa^\ast \rangle \sim \frac{1}{n} \int d\epsilon~ \frac{\epsilon
\exp[-(a_n+\epsilon)^\delta]}{a_n}
\]
Using $\epsilon \ll a_n$,
\begin{eqnarray}
\langle \kappa^\ast \rangle &\sim& \frac{1}{n} {\exp(-a_n^\delta)}\int d\epsilon ~\epsilon
\frac{\exp(-\delta a_n^{\delta-1}\epsilon)}{a_n} \nn \\
&\sim& \frac{1}{n} \frac{\exp{(-\ln n -\ln ((\ln
n)^{\frac{1}{\delta}-1}}))}{ a_n^{2\delta-1}}\nn \\
&\sim& \frac{1}{b_n a_n^{2\delta-1}}\nn \\
&\sim& \frac{b_n}{a_n}\nn \\
&\sim& \frac{1}{\delta} {(\ln n)}^{-1}
\end{eqnarray}
where the last two steps use the forms of the scaling constants given
in \eqref{ab}.\\
%Therefore,
%\[
% \kappa^\ast\sim {(\ln n)}^{-1}.
% \]
\noindent {\bf Power Law:}
% \red{$\langle \kappa^\ast \rangle \sim \Big\langle \frac{\epsilon}{x+\epsilon}
% \Big\rangle$. In the power law case $\epsilon \gg 1$, and so the behavior
% of $\Big\langle \frac{\epsilon}{x+\epsilon}
% \Big\rangle$ could potentially be very different from $\Big\langle \frac{\epsilon}{x}\Big\langle$.
% In fact, since $\epsilon \gg 1$, it might be the case that $\Big\langle \frac{\epsilon}{x+\epsilon}
% \Big\rangle \sim 1$. Can you comment?}\\
%\iffalse
% \iffalse
% Since we have $s_i w_i \ll 1$, it must be true that $s_i \ll 1$ or $w_i \ll 1$.
% First, consider the case $w_i \ll 1$.
% Rearrange \eqref{eq:wi} to get
% \begin{equation}
% w_i = \frac{x-1}{x(x^{s_i}-1)}%=\frac{x-1}{x(x-1)}=\frac{1}{x}.
% \label{eq:wi2}
% \end{equation}
% Then \eqref{eq:wi2} can be satisfied only if $x \gg 1$, and therefore $\kappa$ is large.
% Consider the alternate possibility, $s_i \ll 1$. Consider the case that
% $x$ is {\it not} large. Then, $x^{s_i} \approx 1$,
% and then $w_i \approx 1+w_i -\frac{1}{x}$, which is solved by $x \approx
% 1$, which we ruled out above. Thus, $x$ must be large.
% \fi
% \iffalse
% Consider the other case, $s_i \ll 1$. Once again, the only way in which
% $w_i \ll 1$ is if $x \gg 1$ and so $\kappa \gg 1$.}
% If $E_i$ is chosen from a bounded distribution,
% then the only way $g_1 \gg g_2$ is if $f_1 \gg f_2$. Then $s_i$ is bounded above whereas
% $w_i \approx 0$; using this in \eqref{eq:wi}, we see that for finite
% $x$, the lhs goes to $0$ and the rhs goes to $-1$. The
% only way to avoid this is if $x$ diverges, and therefore $\kappa^\ast$ diverges.
% \fi
The ratio of largest to second-largest value is $1+\langle \frac{\epsilon}{x}\rangle$. Now,
\begin{eqnarray*}
\lefteqn{\langle \frac{\epsilon}{x} \rangle} \nn \\
&=& n^{} \int dx~ P_{max}(x,n-1) \int d\epsilon \frac{\epsilon}{x}
Q(x+\epsilon)\nn \\
&\sim& n \int dx~ P_{max}(x,n-1) \frac{x^{1-\alpha}}{\alpha-1} \label{pl1}\\
&=& \frac{n}{\alpha-1} b_{n-1}^\alpha\int dx
~ {x^{-2\alpha}}e^{-\big(\frac{b_{n-1}}{x}\big)^\alpha}\\
&\approx& \frac{n^{\frac{2}{\alpha}}}{\alpha} \int dy ~ y^{-2\alpha}
e^{-\big(\frac{1}{y}\big)^{\alpha}} \nn \\
&\sim&\frac{n^{\frac{2}{1+\alpha}}}{\alpha-1}.\nn
\end{eqnarray*}
% \noindent {\bf Response }
Therefore the ratio of largest to second largest value diverges, i.e $g_1 \gg g_2$.
Since $\kappa^\ast$ is not small, the formula
$\langle \kappa^\ast \rangle \sim \Big\langle \frac{\epsilon}{x+\epsilon}
\Big\rangle$ does not work. We will demonstrate now that $\kappa^\ast$, in
fact, diverges.
%Assuming it does, the ratio is $\frac{x}{x+\epsilon}$, not
%$\frac{\epsilon}{x+\epsilon}$.
%The former
%goes to $0$, not $1$.
The equation \eqref{eq:wi} is exact.
Rearrange \eqref{eq:wi} to get
\begin{equation}
s_i w_i= \frac{s_i(x-1)}{x(x^{s_i}-1)}.%=\frac{x-1}{x(x-1)}=\frac{1}{x}.
\label{eq:wi2}
\end{equation}
We have seen that for power law, the largest term $g_1 \gg g_2$, the
second largest term, and therefore $s_i w_i =
\frac{g_i}{g_1} \rightarrow 0$ for all $i$. We show that this is
possible only if $x \gg 1$, and therefore $\kappa \gg 1$.
First, it is easy to show that for $x > 1$ (which is our region of interest)
\[
s_i w_i =\frac{s_i(x-1)}{x(x^{s_i}-1)} \ge \frac{1}{x^{1+s_i}}.
\]
Since $s_i w_i \rightarrow 0$, we must have $\frac{1}{x^{1+s_i}}
\rightarrow 0$.
%Consider the case where $E_i=E$ for all $i$, and therefore $s_i=1$. Then
%we require $w_i \ll 1$.
% \iffalse
% First, consider the possibility that $x=1+a$, where $a \ll 1$, which
% corresponds to $\kappa \approx 0$. Putting this in
% \eqref{eq:wi2}, and using the expansion $(x^{s_i}-1)={(1+a)}^{s_i}
% -1\approx s_i a$, we have
% \begin{equation}
% s_i w_i\approx \frac{1}{x}.
% \end{equation}
% Since $g_1 \gg g_2$, we must have $s_i w_i \ll 1$. Therefore, the above
% solution cannot hold, since $x \approx 1$
% by assumption. Thus we have stablished that $x$ cannot be close to $1$,
% and equivalently $\kappa$ cannot be small.
% Since $ s_i w_i \ll 1$, the rhs of \eqref{eq:wi2} must be small. We show
% in the following that this would require
% $x \gg 1$, and therefore $\kappa \gg 1$.
% Suppose $s_i$ is finite. In that case, since $x$ is not close to $1$,
% the rhs small only if $x\gg 1$. Suppose $s_i$ is small; then it is
% easily shown that the rhs goes to the value
% $(x-1)/(x \ln x)$. Once again, this can be small only if $x \gg 1$,
% since $x$ is not close to $1$.
We show below that $s_i$ is bounded with high probability, and therefore,
we must have $x \rightarrow \infty$ with high probability.
% in order to have $s_i w_i \ll 1$.
%Finally, the rhs can be small if $s_i \gg 1$, but we rule out this possibility below.
As before, consider the case that $E$ is drawn from a bounded
distribution. Further, let $f$ be drawn from
a power law distribution, $P(f) \sim \frac{\exp(-f^{-\alpha})}{f^{1+\alpha}}$. Then
\begin{eqnarray*}
\lefteqn{P(g)}\\
& = & \int_{a^{-1}g}^\infty f^{-1} P_f(f) P_e(f^{-1} g)\ dx \\
& = & \int_0^\infty
(a^{-1}g+\epsilon)^{-1}P_f(a^{-1}g+\epsilon)
P_e\big(\frac{g}{a^{-1}g+\epsilon}\big)~d\epsilon\nn \\
&\approx&\int_0^\infty \frac{1}{(a^{-1}g+\epsilon)^{2+\alpha}} P_e\big(\frac{a}{1+\frac{\epsilon}{a^{-1}g}}\big)
~\frac{d\epsilon}{a^{-1}g}\\
&\approx& \frac{1}{(a^{-1}g)^{2+\alpha}}\int_0^\infty \frac{1}{(1+\frac{\epsilon}{a^{-1}g})^{2+\alpha}} P_e\big(\frac{a}{1+\frac{\epsilon}{a^{-1}g}}\big)
~{d\epsilon}\\
&\approx& \frac{1}{(a^{-1}g)^{1+\alpha}}\int_0^\infty \frac{1}{(1+\frac{\epsilon}{a^{-1}g})^{2+\alpha}} P_e\big(\frac{a}{1+\frac{\epsilon}{a^{-1}g}}\big)
~\frac{d\epsilon}{a^{-1}g}\\
&\approx& P_f(a^{-1}g) \int_0^\infty \frac{1}{(1+\epsilon^\prime)^{2+\alpha}}
P_e\big(\frac{a}{1+\epsilon^\prime}\big) ~d\epsilon^\prime \nn
\\
&\sim& P_f(a^{-1}g).\nn
\end{eqnarray*}
Since the tail of $P_g$ behaves similar to that of $P_f$, it follows that
$P(g) \sim {g}^{-(1+\alpha)}$, for $g \gg 1$. Next,
\begin{eqnarray}
{P}(E|g) &=& \frac{P(g|E) P_e(E)}{P(g)}\nn \\
&=& \frac{P_f(\frac{g}{E})P_e(E)}{P(g)}\frac{df}{dg} \label{eq:eg}\\
&\sim& E^{\alpha} P_e(E) \nn \\ \nn
% &\sim& \frac{E^{1+\alpha}}{{g^{1+\alpha}}}\nn \\.
\end{eqnarray}
where we have used that $g \gg E$ and used the forms of the tails of $P_f$ and $P_g$ in the last
step. Note that the $P(E\vert g)$ is independent of $g$ in the limit of
% We see that the rhs is independent of $g$ (note that this happens only for
$g \gg 1$ and $g \gg E$.
%and so we simply denote $P_c(E)=P(E|g)\sim E^{1+\alpha} P_e(E)$, for large $g$.
%So $P(E_1|g_1)\sim \frac{E_1^{1+\alpha}}{{g_1^{1+\alpha}}}$.
Thus,
\begin{eqnarray*}
\lim_{\epsilon \rightarrow 0} P(E < \epsilon| g) & = & \lim_{\epsilon
\rightarrow 0}
\int_0^\epsilon ~dE ~ {E^{\alpha}}P_e(E)\\
& \leq & \lim_{\epsilon \rightarrow 0} \Big(\epsilon^{\alpha} \int_0^\epsilon ~dE~ P_e(E)\Big) = 0.
\end{eqnarray*}
% &\ll& \int_0^\epsilon ~dE~ P_e(E),
% %\sim \epsilon^{2+\alpha} \to 0
% \end{eqnarray}
% where the last relation holds for $\epsilon \ll 1$.
Fix $\delta \ll 1$. Then, there exists $\epsilon>0$, independent of $n$, such that $P(E <
\epsilon| g) < \delta$ when $g \gg 1$.
% Since the last term above is $\le 1$, we see that
% $P(E< \epsilon|g ) \ll 1$, and therefore with very very high probability,
% $E \not \ll 1$ for sufficiently large
% $g$.
Recall that in the power law case, the largest term $g_1 \gg 1$, it
follows that, with probability $1-\delta$,
% we must have $E_1 \not \ll 1$, or
% equivalently $\frac{1}{E_1} \not \gg 1$. Further, $E_i \le a$; therefore,
$s_i = E_i/E_1 \le \frac{a}{\epsilon}$ for all $i$ and $n \geq 1$.
% , implying that $s_i \not \gg 1$.
% \iffalse
% \begin{eqnarray}
% P(E_1=\epsilon_1 | g_1) &=& P(g_1 | E_1=\epsilon_1) \frac{P(E_1=\epsilon_1)}{P(g_1)}\nn \\
% &=&P(g_1 | E_1=\epsilon_1) P(E_1=\epsilon_1) g_1^{1+\alpha}\nn \\
% &=&P(g_1,\epsilon_1) g_1^{1+\alpha}\nn
% % &=& P(f_1=
% \end{eqnarray}
% \fi
%Following arguments
%similar to those in the previous subsection, we see that %$g_i \sim f_i a$ for large $g_i$, and therefore
%$s_i \not \gg 1$.
\begin{thebibliography}{}
\bibitem{alberts} Alberts, Bruce, et al. Essential cell biology. Garland Science, 2013.
\bibitem{specificity1}
Gasch, Audrey P., et al. ``Genomic expression programs in the response of yeast cells to environmental changes."
Molecular biology of the cell 11.12 (2000): 4241-4257.
\bibitem{specificity2}
Fajardo, Alicia, and José L. Martínez.
``Antibiotics as signals that trigger specific bacterial responses." Current opinion in microbiology 11.2 (2008): 161-167.
\bibitem{bialek} Taylor, Samuel F., Naftali Tishby,
and William Bialek. ``Information and fitness." arXiv preprint arXiv:0712.4382 (2007).
\bibitem{parrondo} Parrondo, Juan MR, Jordan M. Horowitz, and Takahiro Sagawa.
``Thermodynamics of information." Nature physics 11.2 (2015): 131.
\bibitem{alon} Aidelberg, Guy, et al. "Hierarchy of non-glucose sugars in Escherichia coli." BMC systems biology 8.1 (2014): 133.
\bibitem{cover} Cover, Thomas M., and Joy A. Thomas. Elements of information theory. John Wiley and Sons, 2012.
\bibitem{tlusty_07} Tlusty, Tsvi. ``Rate-distortion scenario for the emergence and evolution of noisy molecular codes."
Physical review letters 100.4 (2008): 048101.
\bibitem{sgm} Das, Suman G., Madan Rao, and Garud Iyengar.
``Universal lower bound on the free-energy cost of molecular measurements." Physical Review E 95.6 (2017): 062410.
\bibitem{tlusty_iop}Tlusty, Tsvi. ``A simple model for the evolution of molecular codes driven by the interplay of accuracy,
diversity and cost." Physical Biology 5.1 (2008): 016001.
\bibitem{rose} Rose, Kenneth, Eitan Gurewitz, and Geoffrey C. Fox.
"Statistical mechanics and phase transitions in clustering." Physical review letters 65.8 (1990): 945;
Rose, Kenneth.
``Deterministic annealing for clustering, compression, classification, regression, and related optimization problems."
Proceedings of the IEEE 86.11 (1998): 2210-2239.
\bibitem{satya} Sabhapandit, Sanjib, and Satya n. Majumdar.
``Density of near-extreme events." Physical review letters 98.14 (2007): 140201.
\bibitem{book} Leadbetter, Malcolm R., Georg Lindgren, and Holger Rootzén.
Extremes and related properties of random sequences and processes. Springer Science and Business Media, 2012.
\bibitem{stretched} Laherrere, Jean, and Didier Sornette. ``Stretched exponential distributions in nature and economy:“fat tails”
with characteristic scales." The European Physical Journal B-Condensed Matter and Complex Systems 2.4 (1998): 525-539.
\bibitem{generalists} Morgante, J. S., et al. ``Evolutionary patterns in specialist and generalist species of Anastrepha."
Fruit Flies. Springer, New York, NY, 1993. 15-20.
\bibitem{metabolites} Karlovsky, Petr. ``Secondary metabolites in soil ecology."
Secondary metabolites in soil ecology. Springer, Berlin, Heidelberg, 2008. 1-19.
\bibitem{fungi} Jarosz, Daniel F., et al. ''An evolutionarily
conserved prion-like element converts wild fungi from metabolic specialists to generalists." Cell 158.5 (2014): 1072-1082.
\bibitem{biotrophic} Kabbage, Mehdi, Oded Yarden, and Martin B. Dickman. ``Pathogenic attributes of Sclerotinia sclerotiorum:
switching from a biotrophic to necrotrophic lifestyle." Plant Science 233 (2015): 53-60.
\end{thebibliography}
\end{document}
Let us look at the $\delta$-function approximation carefully. If we do the transformation $x=a_n + b_n z$, the expression
for the joint distribution becomes
\[
P(z,\epsilon) = n b_n Q(a_n+b_n z+\epsilon) \frac{1}{b_n} f(z)=Q(a_n+b_nz+\epsilon) f(z)
\]
where $f(z)$ is the Gumbel form $f(z)=e^{-z-e^{-z}}$.
Then
\[
P(\epsilon)=n \int_{-a_n/b_n}^\infty Q(a_n+b_n z + \epsilon) f(z).
\]
It helps to look at the variable $\epsilon=1-\rho$, so that $\kappa^\ast=2z$. The mean is
\[
\bar{\epsilon}= n {a_{n-1}} \int_0^1 \frac{\epsilon}{(1-\epsilon)^2} Q(\frac{a_{n-1}}{1-\epsilon})dz.
\]
Upon the transformation $w=\frac{z}{1-z}$, the integral above reduces to
\[
\bar{z}=n {a_{n-1}}\int_0^\infty dw \frac{w}{(1+w)^3} Q\big(a_{n-1} (1+w)\big).
\]
Consider the special case where the tail of the parent distribution is of the form $exp(-x^\delta)$, $\delta\ge 1$.
\begin{eqnarray}
\bar{z} &\sim& n {a_{n-1}}\int_0^\infty dw \frac{w}{(1+w)^3} \exp [(a_{n-1})^\delta (1+w)^\delta]\nn \\
&\approx& n {a_{n-1}}\int_0^\infty dw \frac{w}{(1+w)^3} \exp [-(a_{n-1})^\delta (1+\delta w)]\nn \\
&=& n {a_{n-1}} \exp (-(a_{n-1})^\delta) \nn \\
&& \int_0^\infty dw \frac{w}{(1+w)^3} \exp [-\delta (a_{n-1})^\delta w)]\nn \\
&\approx& n {a_{n-1}} \exp (-(a_{n-1})^\delta) \nn \\
&& \int_0^\infty dw~ {w} \exp [-\delta (a_{n-1})^\delta w)]\nn \\
&\sim& n \frac{e^{-a_{n-1}^\delta}}{a_{n-1}^{2\delta-1}}\nn \\
&\sim& \big(\ln (n-1)\big)^{1/\delta-2}.
\end{eqnarray}
\pagebreak
\iffalse
Since $a_{n-1}$ is asymptotically large, we may approximate the integral as
\begin{equation}
\bar{z}= \int_0^1 {z} Q\big({a_{n-1}}{(1+z})\big)dz
\end{equation}
Consider the special case where the tail of the parent distribution is of the form $exp(-z^\delta)$.
Then the integral above reduces approximately to
\begin{eqnarray}
\bar{z} &\sim& n\int_0^1 {z} \exp{[-{a_{n-1}}^\delta {(1+z)}^\delta}] dz \nn\\
&\approx& ne^{{-(a_{n-1})}^\delta} \int_0^1 e^{-[\delta {(a_{n-1})}^\delta] z} z ~dz \nn\\
&\sim& \frac{1}{\delta^2 {(\ln (n))^2}}, \\
\nn
\end{eqnarray}
where the last step follows from the fact that $a_n \sim (\ln (n))^{1/\delta}$. Thus $\kappa^\ast \sim 1/ {(\ln n)}^2$.
It is interesting to note that the scaling is independent of $\delta$, so it might be good to check if this scaling is
generic for all this distributions of this type.
{\bf Case 2}
Consider the tail of $Q$ to be light and upper bounded by $a$, i.e of the form $(a-x)^{b-1}$.
\vskip 36pt
let the mean density of states near the maximum be given by the function $\bar{c(r;n)}$, where
\[
{c(r;n)}=\langle \frac{1}{n-1} \sum_{E_i \ne E_{max}} \delta\big(r-(E_{max}-E_i)\big) \rangle.
\]
The location of the second highest value $r_s$ can be approximately found through the relation
\[
\int_0^{r_s} {c(r;n)} dr = \frac{1}{n-1}.
\]
A calculation of the exact statistics of the second highest value is required if one is interested in its distribution,
but for the purpose of finding the scaling property of $\rho$, $r_s$ as defined above should suffice.
From the integral above, we see that $r_s$ must be an asymptotically vanishing function
of $n$, and so up to leading order, $r_s$ is given by
\begin{equation}
r_s \approx \frac{{{c(0;n)}}^{-1}}{n-1} \label{s}
\end{equation}
where we define $$c(0;n)=\lim_{r \to 0} c(r;n)$$, which is the correct
expression in the case of a discontinuous density at
$r=0$ \cite{satya}. Then using the above equations,
\begin{equation}
\kappa^\ast = \frac{2{{c(0;n)}}^{-1}}{a_n (n-1)}.
\end{equation}
According to the finsings about $c(r;n)$ given in \cite{satya}, we arrive at the following cases:
{\bf Case A} ~~~ When the tail of $Q(E)$ is faster than exponential, the maximum value is localized near
$a_n$, where $a_n$ depens on $Q(E)$. The denisty of states is given by
\[
c(r;n)=Q(a_n-r).
\]
now consider some special cases of the distribution $Q$.
{\it 1. Uniform distribution} ~The energy values are chosen uniformly from $(0,E_u)$. In this case $a_n=E_u(1-\frac{1}{n})$,
and $c(0;n)=Q\big(E_u (1-1/n)\big)=\frac{1}{E_u}$. So,
\[
\kappa^\ast \approx \frac{2 E_u^{-1}}{n-2}.
\]
{\it 2. }
% \iffalse
% What is important for us is that the extreme value in the superexponential case is localised
% around $a_n$, and the typical difference between the highest and second-highest value is $b_n$, where
% these two quantities depend on the details of the distribution.
% For a distribution with a super-exponential
% tail, $P_E(E) \sim exp(-E^{1+\delta})$, we have $a_n={(\ln n)}^{1/(1+\delta)}$, and
% $b_n={(1+\delta)}^{-1} {(\ln n)}^{-\frac{\delta}{1+\delta}}$. Then
% \[
% \rho=1-b_n/a_n=1- \frac{(1+\delta)^{-1}}{\ln n},
% \]
% and so
% \[
% \kappa^\ast= \frac{2(1+\delta)}{\ln n}.
% \]
% \fi
\vskip 6pt
\noindent {\it Case 3}:
Quite generically, we can expect $\kappa^\ast$ to approach to zero for asymptotically large $n$.
Consider, as before, the case of diagonal $E_{i\alpha} = \delta_{i\alpha}
E_{\alpha}$, but now we allow arbitrary distributions $f_{\alpha}$
Small $\kappa$ allows us to expand \eqref{cond} up to quadratic order
in $\kappa$, and the transition
happens when
\begin{eqnarray}
\kappa^\ast= \min_{\{i \ne r\}} \frac{2(f_r E_r -f_i E_i)}{f_r {E_r}^2+f_i {E_i}^2} \nn \\
\approx \min_{\{i \ne r\}} \frac{2(f_r E_r -f_i E_i)}{f_r {E_r}(E_r+{E_i})}.
\end{eqnarray}
\fi
\vskip 72 pt
\red {next, we consider the problem of computing the optimal $p(i\vert \alpha)$
for $\kappa \neq 0$. Let
\begin{eqnarray*}
J(p,q) & = & \sum_{i\alpha} f_{\alpha} p(i\vert \alpha) \log \left(
\frac{p(i\vert \alpha)}{q(i)} \right)\\
& = & I(p) + D(p\|q),
\end{eqnarray*}
where $D(p,q)$ denotes the relative entropy of $p$ and $q$.
Then, $J(p,q)$ is jointly convex in $(p,q)$ and moreover, $min_{q} J(p,q)
= I(p) + \min_q D(p,q) = I(p)$.
Since $G(p) - \kappa^{-1} J(p,q)$ is jointly concave in $(p,q)$, and
$\max_{(p,q)} G(p) - \kappa^{-1} J(p,q) = \max_p G(p) - \kappa^{-1} I(p)$,
it follows
that the coordinate descent will converge to the optimal solution. Thus,
we have the following algorithm will compute the optimal solution for
$\kappa > 0$:
\begin{enumerate}[1.]
\item Compute the solution of the optimization problem
$
\max_{p} \{ G(p) - \kappa^{-1} J(p,q^k)\}
$
It is easy to check that $p^\ast(i\mid \alpha) \propto q^k(i) e^{\kappa
E_{i\alpha}}$.
\item Compute the solution of the optimization problem
$
\max_{q} \{ G(p^k) - \kappa^{-1} J(p^k,q)\}.
$
From the above analysis, it follows that $q^\ast(i) = \sum_{\alpha}
f_{\alpha} p^k(i\vert \alpha)$.
\end{enumerate}
next, consider the case where
\[
G_{\epsilon} (p) = \sum_{\alpha} f_{\alpha} \sum_{i} p(i \vert \alpha)
\left(1-\frac{\epsilon}{2} p(i \vert \alpha)\right)E_{i\alpha}
\]
with $\epsilon > 0$. When $\kappa = 0$, the optimization problem reduces
to
\[
\max_p \sum_i p_i g_i(1- \frac{\epsilon}{2}p_i),
\]
where $g_i = \sum_{\alpha} f_{\alpha} E_{i\alpha}$. Using Lagrange
multipliers, this problem reduces to searching for a multiplier $\lambda$
that satisfies
\[
\sum_i \Big( 1- \frac{\lambda}{g_i}\Big)^+ = \epsilon.
\]
The function $f(y) = y (1-\frac{\epsilon}{2} y)$ is a concave
function of $y$; therefore, using the theory of convex conjugates we can write
\[
h(y) = \min_z \left\{ zy - \frac{1}{2\epsilon} (1-z)^2\right\},
\]
where the minimum is achieved at $z^\ast = 1 - \epsilon y$. We can write
the convex optimization problem
\begin{eqnarray*}
\lefteqn{\max_{p} \kappa^{-1} G(p) - I(p) = }\\
&& \max_p \sum_{i,\alpha} f_{\alpha} E_{i\alpha} h(p(i\vert \alpha)) -
\kappa^{-1} I(p)\\
&& \max_p \min_{z(i\vert \alpha)} \Big\{
\sum_{i,\alpha} f_{\alpha} E_{i\alpha} p(i\vert \alpha) z(i \vert \alpha) -
\kappa^{-1} I(p)\\
&& \hspace*{0.75in}
+ \frac{1}{2\epsilon} \sum_{i,\alpha} f_{\alpha} E_{i\alpha} \big(1-z(i\vert
\alpha)\big)^2 \Big\}
\end{eqnarray*}
This is saddle-point problem where the payoff is concave in $p$ and convex
in $z$. The pair $(p^\ast, z^\ast)$ is saddle-point provided
\begin{eqnarray*}
p^\ast & = & \text{argmax}_p \Big\{
\sum_{i,\alpha} f_{\alpha} z^\ast(i \vert \alpha) E_{i\alpha} p(i\vert \alpha) -
\kappa^{-1} I(p) \Big\}\\
z^\ast (i\vert \alpha) & = & \text{argmin}_{z(i\vert \alpha)}
\big\{p^\ast(i\vert \alpha) z(i \vert
\alpha)
+ \frac{1}{2\epsilon} \big(1-z(i\vert
\alpha)\big)^2\big\}\\
& = & 1 - \epsilon p^\ast(i\vert \alpha)
\end{eqnarray*}
Thus, checking whether the non-coding solution $p_0^\ast$ is optimal for a
non-zero $\kappa$ amounts to checking whether \eqref{eq:metabolic-3} holds
for $E_1_{i\alpha} = E_{i\alpha}(1-\epsilon p^\ast(i\vert\alpha))$.
A simple best-response strategy should converge to the optimal
solution. Thus, we have the following solution algorithm:
\begin{enumerate}[1.]
\item Solve
\[
\max_p \Big\{
\sum_{i,\alpha} f_{\alpha} z^k(i \vert \alpha) E_{i\alpha} p(i\vert \alpha) -
\kappa^{-1} I(p) \Big\}
\]
This is equivalent to the earlier problem with $E_{i\alpha} = z^k(i\vert
\alpha)$.
\item Solve
\[
\min_{z(i\vert \alpha)} \Big\{
\sum_{i,\alpha}f_{\alpha} E_{i\alpha} \Big(p^k(i\vert \alpha) z(i \vert \alpha)
+ \frac{1}{2\epsilon} \big(1-z(i\vert
\alpha)\big)^2\Big) \Big\}
\]
It is easy to check that the optimal solution here is $z^\ast(i\vert
\alpha) = 1 - \epsilon p^k(i\vert \alpha)$.
\end{enumerate}
}
\red{
next, consider the case where $G_{\epsilon}(p) = \sum_{\alpha} f_{\alpha}
\sum_i p(i\vert \alpha) \big( 1 + \epsilon f(i \vert \alpha)\big)$,
i.e. $G_{\epsilon}(p)$ is a \emph{convex} function of $p$. In this case,
$\max_{p} \kappa G_{\epsilon}(p) - I(p)$ is not a convex optimization
problem. However, one can still write
\begin{eqnarray*}
\lefteqn{\max_{p} \Big\{ G_{\epsilon}(p) -\kappa^{-1} I(p) \Big\} =}\\
&& \max_{p} \max_{z} \Big\{\sum_{i,\alpha}f_{\alpha} E_{i\alpha}p(i\vert \alpha)
z(i\vert \alpha) - \kappa^{-1} I(p) \\
&& \hspace*{0.75in} -\frac{1}{2\epsilon}\sum_{i,\alpha} f_{\alpha} E_{i\alpha}(z-1)^2 \Big\}
\end{eqnarray*}
The co-ordinate ascent algorithm for this case is given by
\begin{enumerate}[1.]
\item Solve
\[
\max_p \Big\{
\sum_{i,\alpha} f_{\alpha} E_{i\alpha} z^{k}(i\alpha) p(i\vert \alpha) -
\kappa^{-1} I(p) \Big\}
\]
This is equivalent to the earlier problem with $E_{i\alpha} = z^k(i\vert
\alpha)$.
\item Solve
\[
\max_{z(i\vert \alpha)} \big\{ p^k(i\vert \alpha) z(i \vert \alpha)
- \frac{1}{2\epsilon} \big(z(i\vert
\alpha)-1\big)^2 \big\}
\]
It is easy to check that the optimal solution here is $z^\ast(i\vert
\alpha) = 1 + \epsilon p^k(i\vert \alpha)$.
\end{enumerate}
Since the objective is \emph{not} jointly concave so co-ordinate wise
optimization is not guaranteed to reach a global optimum. A good heuristic
would be to start with $p^0(i \vert\alpha)$ equal to the solution
corresponding to $\kappa = 0$.
}
To find the extremum of $F$, we take the derivatives of the cost and benefit functions:
\begin{eqnarray}
\frac{\delta I}{\delta p(i|\alpha)}=f_\alpha \log \frac{p(i|\alpha)}{p_i}\\
\frac{\delta G}{\delta p(i|\alpha)}=\delta_{i,\alpha} f_\alpha E_\alpha.
\end{eqnarray}
The probabilities $p(i|\alpha)$ satisfy the normalization conditions $\sum_i p(i|\alpha)=1$, which can be
taken care of by introducing the Lagrange multipliers $\mu_\alpha$. The extremum of $F$ then satisfies the equations
\cite{tlusty_07,tlusty_iop},
\begin{equation}
\delta_{i,\alpha} f_\alpha E_\alpha- k^{-1} f_\alpha \log \frac{p(i|\alpha)}{p_i}+\mu_\alpha=0.
\end{equation}
The solution is given by:
\begin{equation}
\frac{p(i|\alpha)}{p_i} = \frac{1+\delta{i,\alpha}(e^{\kappa
E_\alpha}-1)}{1+p_\alpha(e^{k E_\alpha}-1)}
\end{equation}
When $k=0$, the solution is $p(i|\alpha)=p_i$, which is precisely equivalent to $I=0$; it is hardly surprising, since the
sensing cost diverges as $k$ vanishes. But as we see below, the network is non-sensing over a finite interval of $k$, i.e till
$k$ reaches a critical value $k_c$ that is determined by the nutrient distribution $\{f_\alpha,E_\alpha\}$. In this non-sensing
region, the payoff is $F=\sum_\alpha p_\alpha f_\alpha E_\alpha$, and the optimal solution is of course
$p(i|\alpha)=p_i=\delta_{i,\beta}$, where $\beta$ is the nutrient with the highest expected free energy benefit,
i.e $f_\beta E_\beta \ge f_\alpha E_\alpha$ for all $\alpha$. The network thus produces just one response all the time.
The transition boundary can be found by examining the stability of this solution. The curvature of $F$ is given by the
Hessian
\cite{tlusty_07}:
\begin{equation}
\cH_{\alpha i,\omega j}= \delta_{i,j}\frac{f_\alpha}{p_i} (\delta_{\alpha,\omega} \frac{p_i}{p(i|\alpha)}-f_{\omega}).
\end{equation}
The variation in $F$ around the extremum is given by $\sum \cH_{\alpha i,\omega j} \delta p(i|\alpha) \delta p(j|\omega)$.
We now introduce the very useful scaling given in \cite{tlusty_07},
$\delta \tilde{p}_{(i|\alpha)}=\delta{p}_{(i|\alpha)}\sqrt{\frac{f_\alpha}{p_i}}$. In this new basis,
the Hessian becomes
\begin{equation}
{\tilde{\cH}}_{\alpha i,\omega j}= \delta_{i,j}(\delta_{\alpha,\omega}g_{i,\alpha}-\sqrt{f_\alpha f_{\omega}}).
\end{equation}
The vectors $\delta\tilde{p}(i|\alpha)$ have to satisfy the normalization relations
$\sum_i \sqrt{\frac{p_i}{f_\alpha}} \delta\tilde{p}(i|\alpha)=0$; i.e the eigenvectors relevant to the problem are
orthogonal to the vector $p_i$. In the non-coding region, the vector lies along the $\beta$ direction, and the
normalization relation amounts to $\delta\tilde{p}(\beta|\alpha)=0$ for all $\alpha$.
At the optimal solution, $g_{i,\alpha}=\frac{1+p_i(e^{\kappa E_\alpha}-1)}{1+\delta_{i,\alpha}(e^{\kappa E_\alpha}-1)}$.
It is interesting to note that in the non-coding region and for $i=\alpha \neq \beta$, we have
$\frac{p_i}{p(i|\alpha)}=e^{-k E_\alpha}\neq 1$, even though the mutual information is zero. This is not a fallacy
because as $p(i)$ and $p(i|\alpha)$ both go to zero while maintaining a fixed ratio, the mutual information vanishes
as well. This is crucial to our theory and a point of departure from usual molecular coding theory \cite{tlusty_07}, where
it is assumed that $\frac{p(i|\alpha)}{p_i}=1$ in the non-coding region, allowing a convenient separation
between the input and output variables.
Our non-coding region is the region of parameter space where all the eigenvalues of $\tilde{\cH}$ are positive.
The sensing transition
happens at the bounday of this region, where the lowest eigenvalue becomes zero. In general it is difficult to find the eigenvalues
exactly, but the matrix $\tilde{\cH}$ can be nuerically diagonalized through standard algorithms. We will also see that
in some special cases of interest we can find the phase boundary exactly.
\begin{figure}
\includegraphics[width=3in,height=3in]{f1b.pdf}
\end{figure}
To find the phase boundary, we consider the existence of an eigenvector
$x_{i,\alpha}$ with eigenvalue zero, and satisfying the corresponding
linear equations:
\begin{eqnarray}
&&\sqrt{f_\beta} \sum_\omega \sqrt{f_\omega} x_{\omega,i} = x_{\beta,i}e^{\kappa E_\beta}, \nn \\
&&\sqrt{f_\alpha} \sum_\omega \sqrt{f_\omega} x_{\omega,i} = x_{\alpha,i}[1+\delta_{i,\alpha} (e^{-kE_\alpha}-1)],
~~~~ \alpha\neq \beta \nn \\
&&x_{\omega,\beta}=0.\nn\\
\end{eqnarray}
Consider the simplest case of two nutrients with $E_1=E_2=E$. now $\tilde{\cH}$ is a four dimensional matrix, with the
three independent parameters, $f_1$,$k$ and $E$. The imposition of the normalization
condition means that the only non-zero components of the vector are $x_{1,2}$, $x_{2,2}$, which now satisfy just two equations:
\begin{eqnarray}
\sqrt{f_1} (\sqrt{f_1} x_{1,2}+\sqrt{f_2} x_{2,2})=e^{k E} x_{1,2}\\
\sqrt{f_2} (\sqrt{f_1} x_{1,2}+\sqrt{f_2} x_{2,2})=e^{-k E} x_{2,2}.
\end{eqnarray}
This can be solved to obtain the phase boundary,
\begin{equation}
k=\frac{1}{E} \log \big[ \frac{1+\sqrt{1-4 f_1 f_2}}{2 f_2}\big].
\end{equation}
We see that when $f_1=f_2=1/2$, the transition happens at $k=0$.
%This property is more general. Whenever the distribution is
%\section{Introduction}
% Thus, we have that
% \begin{equation}
% I_{ss} \leq \min\left\{\frac{1}{p_{\min}},
% \frac{1}{2}\right\} \frac{\Delta \dot{\sigma}_y}{w} = \frac{\Delta
%\pagebreak
%{onecolumn}
%\lipsum[1]
%\onecolumn
%\begin{multicols}{1}
\end{document}
%%% Local Variables:
%%% mode: latex
%%% TeX-master: t
%%% End: