Active information requirements for fixation on the Wright-Fisher model of population genetics
Extracted source text · arXiv 2111.06909
%%
%% Copyright 2007, 2008, 2009 Elsevier Ltd
%%
%% This file is part of the 'Elsarticle Bundle'.
%% ---------------------------------------------
%%
%% It may be distributed under the conditions of the LaTeX Project Public
%% License, either version 1.2 of this license or (at your option) any
%% later version. The latest version of this license is in
%% http://www.latex-project.org/lppl.txt
%% and version 1.2 or later is part of all distributions of LaTeX
%% version 1999/12/01 or later.
%%
%% The list of all files belonging to the 'Elsarticle Bundle' is
%% given in the file `manifest.txt'.
%%
%% Template article for Elsevier's document class `elsarticle'
%% with numbered style bibliographic references
%% SP 2008/03/01
\documentclass[12pt]{article}
%\input{MarksFront8-.tex}
\usepackage{enumerate}
\usepackage{amsmath,amsthm}
\usepackage{amssymb, latexsym}
\usepackage[mathscr]{euscript}
\usepackage{hyperref}
\usepackage{color}
\usepackage{array}
\usepackage{flafter}
\usepackage{layout}
\usepackage{graphicx}
\usepackage{epstopdf}
\usepackage[ansinew]{inputenc}
\usepackage{xfrac}
\usepackage{algorithm}
\usepackage{authblk}
\providecommand{\keywords}[1]{\textbf{\textit{Index terms---}} #1}
\theoremstyle{plain}
\newtheorem{teo}{Theorem}
\newtheorem{coro}{Corollary}
\newtheorem{lema}{Lemma}
\newtheorem{prop}{Proposition}
\newtheorem{algor}{Algorithm}
\theoremstyle{definition}
\newtheorem{defi}{Definition}
\newtheorem{ej}{Example}
\theoremstyle{remark}
\newtheorem{remark}{Remark}
\author[1]{Daniel Andr\'es D\'{\i}az--Pach\'on \thanks{\texttt{Ddiaz3@miami.edu}}}
\author[2]{Robert J. Marks II \thanks{\texttt{ Robert Marks@Baylor.edu}}}
\affil[1]{Division of Biostatistics - University of Miami, Miami, FL}
\affil[2]{Department of Engineering - Baylor University, Waco, TX}
\title{Active information requirements for fixation on the Wright-Fisher model of population genetics}
\date{\today}
\begin{document}
\maketitle
\begin{abstract}
In the context of population genetics, active information can be extended to measure the change of information of a given event (e.g., fixation of an allele) from a neutral model in which only genetic drift is taken into account to a non-neutral model that includes other sources of frequency variation (e.g., selection and mutation). In this paper we illustrate active information in population genetics through the Wright-Fisher model.
\end{abstract}
\keywords{Active information, coalescence, maximum entropy, Wright-Fisher model.}
\section{Introduction}
Stochastic processes have historically been closely related to the study of biological populations. In fact, biological applications have been central to the development of concepts that are today at the core of probability and statistics. At the beginning of the twentieth century, in a quest to mathematize Darwin's model, Ronald Fisher developed fundamental concepts of statistical inference, and also the basic stochastic process in the analysis of population genetics \cite{Fisher1922}. Fisher has been crowned ``the greatest of Darwin's successors'' \cite{Edwards2011}. His theory, together with Sewall Wright's contributions, served to develop what was subsequently dubbed the Wright-Fisher model and is the focus of this paper.
Interestingly, as central as information has been to genetics since the discovery of DNA and protein synthesis, information theory has been almost universally neglected in the stochastic theory of population genetics. We introduce it here from the perspective of active information under maximum entropy \cite{DiazMarks2020}. Our main goal is to measure the information change in an event when we jump from the \textit{neutral} Wright-Fisher model, in which the only ``force'' in operation is genetic drift, to the \textit{non-neutral} Wright-Fisher model with mutations and selection. In some instances this active information is positive and large, making selection and mutation not free lunches in the sense of Wolpert \& MacReady \cite{Wolpert1997}.
In general, active information is a measure of the degree to which a process deviates from equilibrium. In particular, when it is positive, it measures the amount of guiding information needed to achieve success in certain stochastic searches \cite{Dembski2009,EwertDembskiMarks2012,MarksDembskiEwert2017}. Active information has been applied to analysis of software and modeling attempts to simulate evolution including programs named AVIDA \cite{LenskiEtAl2003}, EV \cite{Schneider2000} and Metabiology \cite{Chaitin2013}. Each model works only because external information has been applied to guide the program to success. In subsequent studies critiquing of these models, significant active information has been shown to be required for the success of AVIDA \cite{EwertDembskiMarks2009}, EV \cite{MontanezEtAl2010}, and Metabiology \cite{EwertDembskiMarks2013}.
To our knowledge, this paper is the first application of active information to population genetics and the first application of generalized active information \cite{DiazMarks2020} anywhere. Our analysis opens the door to further research on more sophisticated models of population genetics.
\section{The Wright-Fisher model}
In the simplest scenario, the neutral {\em Wright-Fisher model} (see, e.g., \cite{Durrett2008, Etheridge2011}), there are $N$ individuals of a haploid population (i.e., each individual has a single copy of each gene), and each gene has two types of alleles: $A$ and $a$. All individuals of generation $n$ are replaced in the following generation $n+1$ according to a sampling with replacement among all individuals in generation $n$. In other words, if there are $i$ individuals with the $A$ allele in generation $n$, then the probability of having $j$ individuals with that same allele in generation $n+1$ is binomial with parameters $N$ and $i/N$:
\begin{align}\label{WF}
p_{n, n+1} (i, j) = \binom{N}{j}\left(\frac{i}{N}\right)^j\left(1-\frac{i}{N}\right)^{N-j}.
\end{align}
This type of model is sometimes called the {\em forward model} because it is focuses on future offspring.
The behavior of future generations depends on the initial distribution. In the absence of any additional knowledge, we assume the $N$ elements are chosen to be either of type $A$ or type $a$ with equal probability, so that the null probability of $j$ alleles being of type $A$ in the first generation is, according to the principle of maximum entropy and assuming that both alleles $a$ and $A$ are present at time 0,
\begin{align}\label{equiWF}
p_0(i)=\frac{1}{N-1},
\end{align}
for $i= \{1, \ldots, N-1\}$. In fact, it can be shown that the number of $A$ alleles has this distribution, conditionally on the event that both alleles $a$ and $A$ have coexisted for a long time, so that none of them has been lost (see, e.g., \cite[Chapter~8.4]{CrowKimura1970}).
The only ``force'' in the neutral Wright-Fisher model with offspring probability given by (\ref{WF}) is genetic drift; i.e., the changes in alleles proportion from generation to generation are due to genetic drift. The incorporation of selection and mutation (i.e., a purely Darwinian process) is not imposed in (\ref{WF}). Following Etheridge \cite{Etheridge2011}, we will introduce these two by steps. Notice from (\ref{WF}) that the neutral probability of sampling $A$ is $i/N$, and the neutral probability of sampling $a$ is $1 - i/N$. When we introduce a selection coefficient $s$ these probabilties are modified as follows:
\begin{align}\label{WFSel}
\begin{array}{ccc}
\textbf P[A \text{ is sampled}] = \frac{i(1+s)}{i(1+s) + N - i} & \text{ and } & \textbf P[a \text{ is sampled}] = \frac{N-i}{i(1+s) + N - i}.
\end{array}
\end{align}
Thus, with respect to (\ref{WF}), when $s$ is positive, $A$ is favored in the sampling, in whose case $A$ is said to be \textit{beneficial}; on the other hand, when $s$ is negative, $A$ is not favored in the sampling, in whose case $A$ is said to be \textit{deleterious}. When $s=0$, we go back to the original drift-only scenario.
Now, suppose that every individual of type $A$ mutates to $a$ with probability $\mu_1$, and every individual of type $a$ mutates to $A$ with probability $\mu_2$. In this case, after adding both selection and mutation, the expected proportion of type $A$ offspring in the next generation is given by
\begin{align}\label{WFSelMut}
\theta_i=\frac{i(1+s)(1-\mu_1)}{i(1+s)+N-i} + \frac{(N-i)\mu_2}{i(1+s) + N-i}.
\end{align}
Notice that the first term at the RHS of (\ref{WFSelMut}) corresponds to the probability of sampling $A$ in (\ref{WFSel}) multiplied by $(1-\mu_1)$; that is, it considers the type-$A$ individuals that did not mutate to $a$. Analogously, the second term corresponds to the probability of sampling $a$ multiplied by $\mu_2$; that is, it considers the proportion of type-$a$ individuals that mutated to $A$. The two terms added in (\ref{WFSelMut}) give the proportion of type-$A$ individuals in the next generation.
We know from (\ref{WF}) that, given $i$ individuals of type $A$ at time $n$, the conditional distribution of type $A$ individuals at time $n+1$ is Bin$(N,i/N)$. The addition of selection and mutation changes the conditional distribution to Bin$(N,\theta_i)$, where $\theta_i$ is as in (\ref{WFSelMut}). This is the so-called {\em Wright-Fisher model with selection and mutation}.
\subsection{Active information in the Wright-Fisher model}
For the Wright-Fisher model, let $\psi \sim$Bin$(N, i/N)$ and $\varphi \sim$Bin$(N, \theta_i)$ be the r.v.'s obtained for the neutral and the non-neutral models described above, respectively.\footnote{We here have reversed the order of notation with respect to \cite{DiazMarks2020}.} Let the target $T$ be a given event under observation in the space $\{0,\ldots, N\}$. Conditioned on having $i$ type-$A$ alleles in the present, the {\em endogenous information}, $I_\Omega$, is the Shannon information associated with the probability $\psi$ of reaching a target $T$ under an assumption of neutrality.
When neutrality is disregarded, the probability of reaching $T$ changes and the Shannon information it generates is called {\em exogenous information}, $I_1$. The amount by which this information has changed from neutrality to non-neutrality,
\begin{equation} \label{ChristIsRizen0421}
I_+ = I_\Omega - I_1,
\end{equation}
is the \textit{active information}. It can be positive, negative or zero. Active information will be negative when the non-neutral model performs worse than the neutral one, in terms of finding the target $T$. %For domains bounded on both sides, maximum entropy is achieved by a uniform distribution. Single sided domains achieve maximum entropy with exponential and geometric distributions while unbounded domains require a normal or gaussian distribution \cite{DiazMarks2020}.
\begin{lema}\label{AIWF}
\begin{itemize}
\item[\textit{i.)}] The active information of the success event of drawing one single $A$ allele (with probability $i/N$) of the Wright-Fisher model with selection and mutation, referenced to the neutral model, is given by $I_+ = \log \frac{\theta_i}{i/N}$.
\item[\textit{ii.)}] In particular, when we add only selection but no mutations, the active information becomes
\begin{align}\label{aiwf}
I_+= \log \frac{N + Ns}{N + is}.
\end{align}
\end{itemize}
\end{lema}
\begin{remark} This computation of active information in Lemma \ref{AIWF} refers to one child drawing an allele from a pool of $i$ $A$-alleles and $N - i$ $a$-alleles. This corresponds to $\Omega = \{a, A\}$ and $T = \{A\}$. In this case $\psi(A) = i/N$ is not uniform, but it can still be motivated as a `maxent procedure', since the child draws the parent according to a uniform distribution from a pool of $N$ parents.
\end{remark}
\begin{proof}
\textit{i.} Note that the endogenous information of the success event is given by $I_\Omega = -\log(i/N)$, and the exogenous information of that same event is $I_1 = -\log \theta_i$. To obtain the active information just apply \eqref{ChristIsRizen0421}.
\textit{ii.} Replace the actual values of (\ref{WFSel}) in part \textit{i}:
\begin{align*}
I_+& = \log \frac{\frac{i(1 + s)}{i(1 + s) + N - i}}{\frac{i}{N}}\\
&= \log \frac{\frac{i(1 + s)}{N + is}}{\frac{i}{N}}\\
&= \log \frac{N + Ns}{N + is}.
\end{align*}
\end{proof}
For $N > i$, from part \textit{ii.} of Lemma \ref{AIWF}, we see that the active information of success is positive whenever $s$ is positive, and it is negative whenever $s$ is negative. It is zero when $s = 0$.
\begin{lema}
Let $\varphi\sim$Bin$(N, \theta_i)$ and $\psi\sim$Bin$(N, i/N)$. Let also $T$ be the event that, conditioned on having $i$ type $A$ alleles in the present, we obtain $j$ type $A$ alleles in the next generation. Then
\begin{align*}
I_+ (\varphi | \psi)(T) = j \log \left(\frac{\theta_i}{\frac{i}{N}}\right) + (N-j) \log \left(\frac{1-\theta_i}{1-\frac{i}{N}}\right).
\end{align*}
In particular, when $\mu_1 = \mu_2 = 0$, the active information becomes
\begin{align*}
I_+ (\varphi | \psi)(T) = j \log(1 + s) + N \log \left(\frac{N}{N + is}\right).
\end{align*}
\end{lema}
\begin{proof}
\begin{align*}
I_+ (\varphi | \psi)(T) &= \log\left(\frac{\theta_i^j\left(1-\theta_i\right)^{N-j}}{\left(\frac{i}{N}\right)^j\left(1-\frac{i}{N}\right)^{N-j}}\right) \\
&= \log \left(\left(\frac{\theta_i}{\frac{i}{N}}\right)^j \left(\frac{1-\theta_i}{1-\frac{i}{N}}\right)^{N-j}\right) \\
&= j \log \left( \frac{\theta_i}{\frac{i}{N}} \right) + (N-j) \log \left( \frac{1-\theta_i}{1-\frac{i}{N}} \right).
\end{align*}
To see the particular case, we replace $\theta_i$ by the probability of sampling $A$ in (\ref{WFSel}) to obtain:
\begin{align*}
I_+ (\varphi | \psi)(T) &= j \log \left( \frac{\frac{i(1 + s)}{N + is}}{\frac{i}{N}}\right) + (N-j) \log \left( \frac{\frac{N - i}{N + is}}{\frac{N-i}{N}}\right)\\
&= j\log \left( \frac{N(1 + s)}{N + is} \right) + (N - j) \log \left( \frac{N}{N + is}\right)\\
&= j \log(1 + s) + N \log \left(\frac{N}{N + is}\right).
\end{align*}
\end{proof}
Conveniently, it is easier to calculate the active information of the event $T$ than the individual probabilities $\varphi(T)$ and $\phi(T)$, since the combinations $\binom{N}{j}$ cancel out in the argument of the logarithm.
\begin{coro}
When mutations are absent and selection is present, the active information of fixation of the type $A$-allele in generation $n+1$, given that there are $i$ type $A$-alleles in generation $n$ (i.e., that $j = N$ in generation $n+1$, given that the proportion of $A$ alleles in generation $n$ is $i/N$), is
\begin{align}\label{fixa}
I_+ = N \log \left( \frac{N + Ns}{N + is} \right).
\end{align}
\end{coro}
\begin{remark}
Notice that the active information in (\ref{fixa}) is $N$ times the active information in (\ref{aiwf}).
\end{remark}
This is a fixation probability of $A$ in a single step, which is highly unlikely unless $i \approx N$ in generation $n$. A more interesting result has to do with the \textit{eventual} fixation of $A$, but the calculation of this probability, particularly for the model with selection and mutation, is extremely complicated. In this case, the best strategy is to approximate the process with a diffusion, as we do in the next subsection.
\subsection{Active information for the Wright-Fisher model in the limit}
The neutral Wright-Fisher model is a Markov chain with state space given by $\{0, 1, \ldots, N\}$ and transition probabilities as in (\ref{WF}). The states 0 and $N$ are \textit{absorbing}, meaning that once the chain enters in one of these two states, it cannot leave. When the chain enters the state $N$, it means that $A$ has become fixed. When the chain enters the state 0, $A$ goes extinct and $a$ is fixed. The conditional probability of fixation of $A$, given that there are $i$ alleles of type $A$ at the present, is $i/N$ \cite{Durrett2008}. However, this fixation probability and other calculations are extremely complicated when selection and mutations are present and the distribution is as in $\varphi$. For this reason, it is customary to look for diffusion approximations.
A one-dimensional diffusion is a strong Markov process on $\mathbb R$ with continuous paths (see e.g. \cite{Klebaner2012}). A diffusion $\{X_t\}_{t \geq 0}$ can be expressed as the solution of a stochastic differential equation driven by a Brownian motion with appropriate boundary conditions:
\begin{align}\label{diff}
dX_t = \mu(X_t)dt + \sigma(X_t)dB_t,
\end{align}
where $B_t$ is a Brownian motion, and $\mu(x)$ and $\sigma^2(x)$ are called the infinitesimal drift and variance of the diffusion. The limiting diffusion of the Wright-Fisher model with selection and mutation is given by (see \cite[Lemma~5.5]{Etheridge2011}):
\begin{align}
\mu(p) &= \alpha p(1-p) - v_1p + v_2(1-p)\\
\sigma^2(p) &= p(1-p).
\end{align}
For this limiting diffusion, time has been rescaled to units of $N$ generations: $\alpha = Ns$, $v_1 = N\mu_1$, $v_2 = N\mu_2$, and $p$ is the proportion of type $A$ individuals in the population. In the absence of mutations ($v_1 = v_2 = 0$), the conditional probability of fixation is given by Etheridge (\cite{Etheridge2011}, p. 68):
\begin{lema}\label{fix}
Suppose that there is no mutation ($v_1 = v_2 = 0$). If the initial proportion of $A$-alleles is $p_0$, the probability $p_{fix}(p_0)$ that the $A$-allele eventually fixes in the population (that is the diffusion is absorbed in $p=1$) is
\begin{align}
p_{fix}(p_0) =
\begin{cases}
\frac{1 - \exp(-2\alpha p_0)}{1 - \exp(-2\alpha)} & \text{ if } \alpha \neq 0,\\
p_0 & \text{ if } \alpha = 0.
\end{cases}
\end{align}
\end{lema}
%
From Lemma \ref{fix}, the next Corollary follows directly:
\begin{coro}\label{AIfix}
Under the same conditions as before, let $T$ be the event that, given that there is an initial proportion $p_0$ of $A$-alleles, the conditional event that the $A$-allele gets fixed is
\begin{align}
I_+(T) = \log \frac{\frac{1 - \exp(-2\alpha p_0)}{1 - \exp(-2\alpha)}}{p_0}.
\end{align}
\end{coro}
Notice that at fixation the search space becomes $\Omega = \{0,1\}$, and the target is $T=\{1\}$.
Of course, once we add mutation, fixation loses all meaning. With mutations, there are no absorbing states in the Markov chain. If a selected allele arises through mutation in an otherwise neutral population, then its actual frequency is $1/N$, so with a little abuse of notation, still calling $p_{fix}(\cdot)$ the first time that the mutated allele has frequency 1, we obtain
\begin{align*}
p_{fix}\left(\frac{1}{N}\right) = \frac{1 - e^{-2s}}{1 - e^{-2Ns}}.
\end{align*}
%
In this scenario, Etheridge \cite{Etheridge2011} considers three interesting cases:
\begin{enumerate}
\item Deleterious alleles: $s < 0$. If $|s| \ll 1$, and $N|s| \gg 1$, then $p_{fix}(1/N) \approx 2|s|e^{-2N|s|}$. The fixation probability of a deleterious allele is exponentially small and it decreases with increasing population size.
\item Beneficial alleles: $s > 0$. If $s \ll 1$, $Ns \gg 1$, then $p_{fix}(1/N) \approx 2s$, almost independent of population size.
\item Nearly neutral alleles: If $N|s| \ll 1$, then $A$ is nearly neutral and $p_{fix}(1/N) \approx 1/N$.
\end{enumerate}
%
For these three cases, the active information with respect to the neutral model is given respectively by
\begin{enumerate}
\item Deleterious alleles: $I_+ \approx \log \left(2N|s| e^{-2N|s|}\right)$. If we measure the information in nats, it becomes $I_+ = -2N|s| +\ln(2N|s|) < 0$.
\item Beneficial alleles: $I_+ \approx \log 2Ns > 0$. In spite of the probability of fixation being almost independent of population size, active information \textit{is} dependent on population size.
\item Nearly neutral alleles: $I_+ \approx 0$.
\end{enumerate}
%
Thus, although most alleles (beneficial or deleterious) are lost, ``fitness differences that are too small to be measured in the laboratory ($|s| < 1$) can still play an important role in evolution (if $N|s| \gg 1$)'' \cite{Etheridge2011}. This important role is made explicit by the active information measure. In fact, except on the case of nearly neutral alleles, the population size plays a very important role: for the deleterious case, it makes the active information to decrease linearly in $N$; and for the beneficial allele, it makes the active information to increase logarithmically in $N$.
\section{Active information over nonnegative reals: Coalescence}
The geometric distribution with mean $\mu$ possesses maximum entropy among all distributions over the nonnegative integers with specified mean $\mu$. Letting $\tau\sim\text{Geom}(1/\mu)$ and calling $\psi_\mu$ its distribution, this means that, if we have a search space $\Omega = \{1,2,\cdots \}$ and our only knowledge is that we are looking at a target $T \subset \Omega$ according to a distribution with mean $\mu$, every representation of the search of $T$ must start with the null (endogenous) information in terms of $\tau$, it is the equilibrium from which the active information will be measured (see Table 1 of \cite{DiazMarks2020}).
This sets the stage to think {\em backwards} of the Wright-Fisher model as the genealogy of a population instead of its offspring. Assume that our prior knowledge is that the population has size $N$. Then, since each new generation is obtained after a sampling with replacement, the probability of two given individuals at the present generation sharing the same father is $1/N$.
Defining $T := \{\tau = k\}$, for $k \in \mathbb Z^+$, we obtain that the probability of a common ancestor $k$ generations before is given by a geometric distribution with mean $N$:
\begin{align}\label{Geo}
\psi(T) = \left(1-\frac{1}{N}\right)^{k-1}\frac{1}{N},
\end{align}
and the endogenous information is $I_\Omega = -\log \psi(T)$. Suppose that after further analysis we obtain knowledge that the population is actually of size $\nu \neq N$. This defines an alternative search that uses $-\log \varphi(T)$ bits of information, where $\varphi$ is the mass density of a geometrically distributed r.v. with mean $\nu$. The active information measured in nats becomes
\begin{align*}
I_+(\varphi | \psi)(T) &= \log \frac{\left(1-\frac{1}{\nu}\right)^{k-1}\frac{1}{\nu}}{\left(1-\frac{1}{N}\right)^{k-1}\frac{1}{N}}\\
&= \log\left\{\left(\frac{1-\frac{1}{\nu}}{1-\frac{1}{N}}\right)^{k-1}\frac{N}{\nu}\right\}\\
&= \log\left\{\left(\frac{1-\frac{1}{\nu}}{1-\frac{1}{N}}\right)^k \frac{N-1}{\nu -1}\right\}
\end{align*}
%
If we rescale $k = dN$ and take $\nu$ to be $\mathcal O (N)$, then there is $0 < c < \infty$ such that $\nu \approx cN$ when $N$ is large, and the active information becomes
\begin{align}\label{ComparingGeoms}
I_+(\varphi | \psi) \approx \left(1- \frac{1}{c}\right)d-\log c,
\end{align}
which is a linear function of $d$. From this we notice that the limiting active information is positive for large $d$ when $\nu > N$, i.e. $c > 1$. On the contrary, the limiting active information is negative (positive) for small $d$ when $\nu > N$ ($\nu < N$), i.e. $c>1$ ($c<1$). This corresponds to our intuition. For instance, if tilting increases the size of the population $(c > 1)$, then we expect longer coalescence times, and this make it easier (more difficult) to find targets that correspond to unusually large (small) coalescence times.
A natural extension of this model is to consider a geometric distribution with success probability $\lambda = 1/N$, which, as $N$ grows, approaches an exponential distribution with intensity $\lambda$. The exponential distribution is also the maxent distribution over the set of all distributions with support on the nonnegative reals and intensity $\lambda$ (see Table 1 of \cite{DiazMarks2020}). Then, when our only knowledge of a search is that it is done on $\{0\} \cup \mathbb R^+$ and that it has a finite mean $\mu = \lambda^{-1} = N$, \cite{DiazMarks2020} tells us that our endogenous search must be guided by an exponential r.v. with mean $\mu$.
Consider the Wright-Fisher model and assume we are taking $N$ generations as our unit of time (i.e., just like 1 minute has 60 seconds, one unit of time here has $N$ generations), then, going backwards, the time to go from $i$ lineages to $i-1$ is exponential with intensity $i(i-1)/2$ and mean $2/[i(i-1)]$. (J. F. C. Kingman was the first to develop the coalescent in \cite{Kingman1982}; Berestycki has an excellent material explaining where the current research is \cite{Berestycki2009}.)
Any change in the distribution, as usual, is contributing information for faster or slower coalescence. To see this, define the event $T=$ ``The time to coalescence from $i$ to $i-1$ lineages is more than $t$.'' Then any other distribution altering the target, say another exponential with different mean $\mu$, is adding
\begin{align*}
I_+(T) &= \log\frac{\exp\left\{-\frac{1}{\mu} t\right\}}{\exp\left\{-\frac{i(i-1)t}{2}\right\}}\\
& = \frac{i(i-1)t}{2} - \frac{1}{\mu} t\\
& = t \left( \frac{i(i-1)}{2} -\frac{1}{\mu} \right)
\end{align*}
nats of information. Thus, in the same lines of the discrete situation, when $\mu = \frac{2}{i(i-1)}c$, which corresponds to the rescaled coalescence time in units of $N$ for a population of size $cN$, the active information becomes
\begin{align*}
I_+(T) = \frac{i(i-1)t}{2}\left(1 - \frac{1}{c}\right).
\end{align*}
Thus, the active information is positive when $\mu > \frac{2}{i(i-1)}$ and negative when $\mu < \frac{2}{i(i-1)}$. As expected, as long as we are considering another exponential distribution, shrinking the mean of the exogeneous search will lessen the probability that the coalescence time is large, whereas augmenting the former will increase the probability of the latter.
\section{Conclusion}
Generalized active information is well-suited to measure the amount of information introduced in a population genetics model when neutral models are replaced by non-neutral ones. In this paper, we focused on the Wright-Fisher model, studying how much information is added by the Darwinian paradigm (considering selection and mutations) with respect to a neutral model that only takes into account genetic drift. Other variation sources (e.g., recombination) can be considered in order to determine the amount of information they introduce to the analysis.
The Wright-Fisher model is the basic introductory model of population genetics. As such, this article opens up a research path in which the neutral and non-neutral versions of more sophisticated models can be compared through active information.
In the pre-limiting model, as it was made explicit in Remark 2, when comparing the Wright-Fisher model with mutation to the neutral Wright-Fisher model, the active information of fixation is $N$ times the active information of the success event of the binomial distributions under consideration.
When we go to the limit, as we saw at the end of Section 2, active information makes explicit fitness differences that cannot be observed in the laboratory but that still are significant asymptotically. In fact, even though the probability of fixation of a beneficial allele in a selection-only model is independent of the population size, active information shows that the information that the selection coefficient introduces plays a very important role as the population size increases. When a deleterious allele is introduced, the population size makes the active information negative; and when a beneficial allele is introduced, it makes the active information positive.
In summary, we see at least two things: First, selection does not act in the model as an innocuous force. Selection \emph{adds} information. Second, since selection is an information source (and mutations when they are present), it is not a free lunch, and the information it adds is compounded by the population size. Our research stands alongside the work of Basener and Sanford \cite{Basener2018} who through alternative analysis have demonstrated the ineffectiveness of the Wright-Fisher model to create information ex-nihilo.
When active information was originally introduced, its purpose was to measure the amount of information added by a programmer in an alternative search, referenced to a blind one. In this kind of problem the (search) space is always compact. However, non-compact spaces require a generalization of the baseline distribution beyond uniformity in order to be able to account for equilibrium in these spaces, particularly when the space is at least countably infinite. This generalization was developed in \cite{DiazMarks2020}. As such, to our knowledge, this application to population genetics is the first to use generalized active information.
\section{Acknowledgements}
The authors thank the reviewers for the careful reading of the text and their comments. This paper is better because of their suggestions and corrections.
%\subsection{Active information in thermodynamics}
%
% Let $S$ be a \textit{state space} or \textit{single-spin space} such that $S=\{s_1,\ldots,s_n\}$ and define a function $E:S \rightarrow \mathbb R$. In thermodynamics, $E(s_j)$ represents the energy of the particle in the state $s_j$ (Hamiltonians). If the $E(s_j)$'s are not all equal, then for each $\overline{E} \in [\min E(s_j), \max E(s_j)]$, there is a unique probability distribution $q$ on $S$ satisfying that $\sum q_jE(s_j) = \overline{E}$ that has maximum entropy. It is given by
% \begin{align}\label{Gibbs}
% q_j = \frac{e^{-\beta E(s_j)}}{\sum_i e^{-\beta E(s_i)}},
% \end{align}
%for a unique $\beta =\beta(\overline E)$ such that $\beta \in [\infty, \infty]$. The distribution (\ref{Gibbs}) is known as the \textit{Boltzmann-Gibbs distribution} %(For a more thorough description, see
%\cite{GeorgiiHaggstromMaes2001}. Its derivation follows the %same lines that the derivation
%that of (\ref{DiffMeanFinite}) \cite{Conrad}.
%
% A case of interest in this context is to measure the addition of information in open thermodynamical systems. For instance, Sewell has explained in different places that not even in an open system anything is possible (see \cite{Sewell2001} and Appendix D of \cite{Sewell2005}.) He shows that increasing order in a non-isolated system is still extremely improbable (as would be the case in a closed system), unless the system is receiving more order through the boundary. This order is, of course, usually represented as the negative of the entropy. Sewell is showing a more general result than the usual second law of thermodynamics assumed confined to a closed systems. Thus, if (\ref{Gibbs}) describes the distribution from which we are taking our equilibrium (endogenous) information, $I_0$, this negative entropy coming into the system through the boundaries operates as alternative (exogenous) information $I_1$. %And, as usual, we can think of
%The difference $I_+ = I_0 - I_1$ is the amount of information imparted to the system in order to get the desired increasing order.
%\section{Active information over the reals}
%
%%{\bf I'm not sure this section is mature enough yet to publish}
%
% The anthropic cosmological principle, or fine tuning, as it is usually called, has puzzled most of physicists and cosmologists for decades. It is a concept popularized at the end of the 1980's by cosmologists John Barrow and Frank Tipler \cite{BarrowTipler1988}, although others have call it into attention almost since the beginning of the twentieth century. Two things are observed in fine tuning: the constants of nature and the boundary conditions of certain physical and chemical quantities. The idea is simple: the value observed, either a constant or a boundary condition, has to be in a very small interval of the real line $\mathbb R$ in order for our universe to be able to host complex, intelligent, carbon-based life. Moreover, all the constants and boundary conditions happen to live in such permitting-life intervals with low Lebesgue measure $\mathcal L$. As an example, let's consider the gravitational law
% \begin{align*}
% F =\frac{Gm_1m_2}{d^2},
% \end{align*}
%where $F$ is the gravitational force, $m_1$ and $m_2$ are the masses of two bodies, and $d$ is the distance between them. In this case, $G$ is the constant under analysis. Its value is $6.67408 \times 10^{-11} \text N \text m^2 \text{kg}^{-2}$. But gravitational force is the weakest of the four fundamental forces (the other three being electromagnetism, and the strong and weak nuclear forces). In fact, it is $10^{-40}$ times the value of the nuclear force. However, an increment of $10^9$ in the gravitational force, so that the relation to the strong nuclear force would be $10^{-31}$, would have made life impossible. In fact, Paul Davies argues in \cite{Davies1983} that a change of $10^{-100}$ in either the gravitational force or the weak nuclear force would render intelligent life impossible. More examples have emerged (famously known are Einstein's cosmological constant, fine-tuned around $10^{-120}$, and the initial entropy of the universe, adjusted at the staggering value of $10^{-10^{123}}$) \cite{Penrose1999,Penrose2009}), but our interest here is not so much on a defense of fine tuning as it is in a proposed way to measure if fine tuning is really there, as we will see bellow.
%
%\subsection{Fine tuning}
%
% Critics have argued that there is no such a thing as fine tuning of universal constants to allow life (see, for instance, \cite{MacGrew2Vestrup2001}). Their argument goes like this: since the only thing we know is that these constants live in the real line, defenders of fine tuning are not epistemically allowed to say anything about it because when we are thinking on those values as realization of a random variable, we are assuming also a non-uniform distribution function, and such a thing is not warranted by the PrOIR. Since there are some distributions assigning high probability to the interval $\Im$, as well as others with low probability, it is argued that the selection of any such distribution has to be arbitrary.
%
%Such is not the case under the admitably speculative assumption that
% \bn
% \iii The observed value is the distribution mean, and
% \iii The underlying distribution is at maximum entropy
% \en
%
% The extension of Bernoulli's PrOIR to non-finite sets based on maximal entropy helps to partially solve the problem: Let's assume for the sake of the argument that we also know that the distribution must have finite ---but possibly unknown--- variance $\sigma^2$ (the argument can be developed under less than this condition, as it will be seen later, but for now it makes things more familiar to work with). Then the maximum entropy distributions over the real line satisfying this constraint are all those in the set of normal distributions with variance $\sigma^2$. Any other distribution will constitute an unwarranted addition of active information. The normal distributions with variance $\sigma^2$ represent those with \textit{minimal} information, as correspond to the current knowledge of the underlying setting. This is then the most conservative assumption to do. {\bf gaussian has domain less than zero}.
%
% What is left is the selection of the mean. Since the only data we have is the actual value of the constant of nature, $C$ ($G$ in the example of gravitation), then the natural thing to do is to assign to it the highest density value ---the mean. {\bf isn't it the mode?}. We will also assume that the mean is centered in the life-permitting interval $\Im:=(\mu-\Delta x, \mu+\Delta x)$, for $\Delta x \in \mathbb R^+$. Let $X\sim \mathcal N(\mu,\sigma^2)$, then the null information is given by
% \begin{align*}
% I_0 = \log\{\textbf P_{\sigma^2}[X \in \Im]\}.
% \end{align*}
% This line of thought shows that the probability depends directly on the choice of the variance, something we emphasized with the subindex in the probability above. Let $I_1 = \log\{\textbf P_{\tilde\sigma^2}[X \in I]\}$. It is then clear that if $\tilde\sigma^2 > \sigma^2$, then $I_+$ is negative. On the other hand, $I_+$ is going to be positive when $I_0 < I_1$, implying that there was an addition of information.
%
% Notice, however, that if the variance keeps on decreasing, $X$ converges weakly to a Dirac's concentration measure $\delta_C$; this implies that as $\sigma^2$ goes to zero, we are making the value of $C$ more and more deterministic, which is quite the opposite to the way the fine tuning argument is constructed from the beginning: to assume that the value is the realization of a \textit{non-trivial} random variable, not the realization of a \textit{degenerate} one ---a case that would be completely uninteresting. In fact, it is completely natural to assume that the constant might have a different value, so there is no reason to assume that it is deterministic or possesses low variance. In a philosophical sense, the criticism is epistemological (how to know what distribution to assign as the null one), not ontological (that there exist a random variable whose realization outside certain range would have avoided the appearance of intelligent life). On the other hand, if $\sigma^2$ keeps increasing, the random variable is converging to something that is not even a random variable assigning 0 probability to all the real intervals.
%
%
%
%\subsection{Exponential}
%For mean $\mu$, the exponential distribution has a density of
%$$ f_X (x) = \frac{1}{\mu} \exp \llll( \frac{x}{\mu} \rrrr) H(x) $$
%where the Heaviside step function is zero for $x<0$ and is otherwise one. %and the cumulative distribution is
%%$$ F_X (x) = \llll( 1- \exp (- x/\mu) \rrrr) H(x) $$
% The standard deviation for the exponential is also equal to the mean. Then, set
%\bea \wp \aaa \Pr [ \mu - \Delta x \leq X < \mu + \Delta x] \nnn
%%\aaa \exp \left(- (\mu - \Delta x) /\mu \right) - \exp \left(- (\mu + \Delta x) /\mu \right)\nnn
%\aaa 2 \exp (-1) \sinh\left( \frac{\Delta x}{\mu} \right ) \nn
%\eea
%For $z \ll 1$, we have $\sinh (z) \approx z$ and
%\be \wp \approx \frac{2\Delta x}{e \mu}
%\bbb{Ate2Much3}
%\ee
%
%%\begin{enumerate}
%%\item Nuclear Efficiency.
%%\begin{quote}
%%From Wikipedia:
%%Epsilon ($\varepsilon$), a measure of the nuclear efficiency of fusion from hydrogen to helium, is 0.007: when four nucleons fuse into hydrogen, 0.007 (0.7\%) of their mass is converted to energy. The value of $\varepsilon$ is in part determined by the strength of the strong nuclear force. If $\varepsilon$ were 0.006, only hydrogen could exist, and complex chemistry would be impossible. According to Rees, if it were above 0.008, no hydrogen would exist, as all the hydrogen would have been fused shortly after the big bang. Other physicists disagree, calculating that substantial hydrogen remains as long as the strong force coupling constant increases by less than about 50\%.\
%%\end{quote}
%%
%%So $\mu = 0.007$ and $\Delta x = 0.001$ so $\wp = 0.105 = 10.5\% %466288
%%$
%%
%%\item
%%Hoyle State.
%%\begin{quote}
%%From Wikipedia:
%%An older example is the Hoyle state, the third-lowest energy state of the carbon-12 nucleus, with an energy of 7.656 MeV above the ground level. According to one calculation, if the state's energy were lower than 7.3 or greater than 7.9 MeV, insufficient carbon would exist to support life; furthermore, to explain the universe's abundance of carbon, the Hoyle state must be further tuned to a value between 7.596 and 7.716 MeV.
%%\end{quote}
%%
%%So $\mu = 7.656
%%$ and $\Delta x = 0.060$ so $\wp = 0.005766194
%%% = 10.5\% %466288
%%$
%%
%%\item Gravity.\footnote{http://www.reasonablefaith.org/transcript-fine-tuning-argument}
%%\begin{quote}
%%Consider gravity, for example. The force of gravity is determined by the gravitational constant. If this constant varied by just one in $10^{60}$ parts, none of us would exist.
%%\end{quote}
%%
%%
%%
%%\end{enumerate}
%
%\subsubsection{Primordial fluctuations}
%
%The {\em amplitude of primordial fluctuations},
%
%;;;
%
%$$ Q \approx 2 \times 10^{-5} $$
%requires fine tuning to 1 part in 100 000
%of the mean density. Rees \cite{Rees2000} writes:
%
%\bc \bq
%``If $Q$ were smaller than $10^{-6}$, gas would never condense
%into gravitationally bound structures at all, and
%such a universe would remain forever dark and featureless,
%even if its initial `mix' of atoms, dark energy
%and radiation were the same as our own. On the other
%hand, a universe where $Q$ were substantially larger
%than $10^{-5}$ -- were the initial `ripples' were replaced by
%large-amplitude waves -- would be a turbulent and
%violent place. Regions far bigger than galaxies would
%condense early in its history. They wouldn't fragment
%into stars but would instead collapse into vast black
%holes, each much heavier than an entire cluster of
%galaxies in our universe $\cdots$ Stars would be packed
%too close together and buffeted too frequently to retain
%stable planetary systems.''
%\eq \ec
%
%However, with $\mu = 2 \times 10^{-5}$,
%$$ \wp = \frac{1}{\mu} \int_{10^{-6}}^{10^{-5}} \exp \llll( \frac{x}{\mu} \rrrr) dx = 0.388 $$%0.388481819
%
% {\huge Wow}
%
%%$ f_X (x) = \frac{1}{\mu} \exp \llll( \frac{x}{\mu} \rrrr) H(x) $$
%\subsubsection{Gravity}
%Consider gravity, for example. The force of gravity is determined by the gravitational constant. If this constant varied by just one in $10^{60}$ parts, none of us would exist.\footnote{http://www.reasonablefaith.org/transcript-fine-tuning-argument} Thus
%$$\frac{\Delta x}{\mu} = 10^{-60}$$
%and \eee{Ate2Much3} becomes
%$$\wp \approx 10^{-60} $$
%
%
%\subsection{Gaussian}
%
%Since the choice of the variance determines the null information, a possible way to determine what the variance should be is to use a universal information bound such as Lloyd's computational capacity of the universe \cite{Lloyd2002} or Dembski's universal probability bound \cite{Dembski1996}. Again, the actual value does not matter as much as the fact that such value can be posed. We call probability bound $\wp$. Then we can say that there is an unwarranted addition of information if $\textbf P_{\sigma^2}(\Im) \geq \wp$, and we have to account for this addition of information. Let $\sigma^2_\wp$ be the maximal variance such that the probability of $\Im$ is greater than $\wp$, {\em i.e.}
% \begin{align*}
% \sigma^2_\wp = \sup \{\sigma^2\in\mathbb R^+ | \ \textbf P_{\sigma^2}(\Im) \geq \wp \}.
% \end{align*}
% Then
% \begin{align*}
% I_0 = \log\{\textbf P_{\sigma^2_\wp}[X \in \Im]\}.
% \end{align*}
%Since $P_{\sigma^2}(\Im) $ is strictly decreasing function of $\sigma^2$, the equality version where $ \textbf P_{\sigma^2_\wp}(\Im) = \wp$ can be considered.
%
%Under a Gaussian assumption
% $$ \wp = 2 \; \mbox{erf} \left( \frac{\Delta x}{\sigma_\wp } \right)$$
%where
% $$ \mbox{erf} (z) =\frac{1}{\sqrt{2\pi}} \int_0^z e^{-x^2 /2} \; dx .$$
%Thus the largest standard deviation for available probability resource, $\wp$, is
% \begin{equation}
% \sigma_\wp= \frac{\Delta x}{ \mbox{erf}^{-1} (\wp /2)}
% \label{Ate2Much}
% \end{equation}
%and the corresponding distribution is exactly
%\begin{equation} f_X (x) = \frac{\mbox{erf}^{-1} (\wp /2) }{\sqrt{2\pi} \Delta x}
% \exp \left( \frac{-\left((x-\mu)\; \mbox{erf}^{-1} (\wp /2) \right)^2}{
% 2 \Delta x^2 }
% \right)
%\label{Ate2Much2}
%\end{equation}
%
%For $z \ll 1$ we have $\mbox{erf} (z) \approx z/\sqrt{2\pi} $. Thus, for $\wp \ll 1$, \eqref{Ate2Much} becomes
% $$ \sigma_\wp \approx \sqrt{ \frac{2}{\pi} } \; \frac{\Delta x}{ \wp} $$
%and the distribution in \eqref{Ate2Much2} is
% $$f_X (x) \approx \frac{\wp}{2\Delta x } \; \exp\left( -\pi \left( \frac{x-\mu }{2 \Delta x } \right)^2
% \right) $$
%;;;
%
%%The variance $\sigma^2_\wp$ can be upper bounded using Chebyshev's inequality
%% \begin{eqnarray}
%% P_{\sigma^2}(\Im) &=& \Pr \left[ | X-\mu | \leq \Delta x \right] \nonumber \\
%% &\geq& 1- \left( \frac{\sigma}{\Delta x} \right)^2 \nonumber \\
%% &\geq& \wp \nonumber
%% \end{eqnarray}
%%
%%
%%
%%
%%
%%for all $\sigma$. Thus
%% $$ \sigma \geq \Delta x \sqrt{1-P_{\sigma^2}(\Im)} $$
%%;;;
% The problem with this selection of $\wp$ is that the two bounds mentioned are valid for events \textit{inside our universe}, but do not work \textit{for our universe}, so the election of $\sigma_\wp^2$ seems \textit{ad-hoc}. We see this approach as a step forward to the solution of how to measure the null information in the fine tuning scenario. However, it is partial because there is still the need to figure out what $\sigma^2$ to use. Some solutions can be proposed but those go beyond the scope of this paper, whose only interest is to show how active information can be useful in more general settings than finite random variables.
%
%\subsection{Keynes' criticism}
%
%It is key in this whole section on fine tuning to understand that it is not mandatory to work with the normal distribution over the reals. We are using it because, among all the distributions with given second moment $m$, the familiy of normal distributions with second moment $m$ has maximum entropy. However, the restriction could have been different. If at some point, for instance, it looks like assuming finiteness of the second moment is not warranted, then something less is possible. Again, Table 1 of \cite{ParkBera2009} gives several examples of other distributions of maximum entropy over the real line under different moment constraints. For example, the Cauchy distributions has maximum entropy over all distributions with the restriction that $\textbf E[\ln(1 + x^2)]=2\ln 2$. It is important to keep this in mind as we proceed in this subsection.
%
%As referenced in the Introduction, Keynes attacked in his \textit{Treatise On Probability} the PrOIR, or as he called it, \textit{the principle of indifference} \cite{Keynes1921}. It seems to us too obvious that his \textit{Treatise} was written previous to 1933, the year when Kolmogorov famously published his \textit{Grundbegriffe der Wahrscheinlichkeitsrechnung} in which he axiomatized probability answering Hilbert's sixth problem, and previous to the development of entropy as. We believe that with such axiomatization many of the questionings Keynes had against the PrOIR can be solved, and this will be the matter of a future paper. But for now we want to focus in one of his criticisms that can be thought in terms of a possible criticism to fine tuning too. In \cite{Keynes1921} pp. 45--47 Keynes considers ``the specific volume of a given substance'' assuming that our only knowledge about it is that it ranges between 1 and 3. Then he considers the specific density of the substance, which is defined as the reciprocal of the volume ranging from 1/3 to 1. For definiteness, let's denote the volume by $V$ and let the density be $D:=1/V$. Keynes criticism in a nutshell is that, given the PrOIR, if one considers that $V$ is uniformly distributed on the interval $(1,3)$ then, since we do not know nothing about the the density neither, we would have to assume too that $D$ is uniformly distributed in $(1/3,1)$.
%
%This criticism commits a fundamental flaw: if we \textit{assume} that the volume of the substance is uniformly distributed in the interval $(1,3)$, then we \textit{know} by some trivial mathematics that $D$ is not uniform in $(1/3,1)$:
%\begin{align*}
% F_D(y) = \textbf P(D \leq y) = \textbf P(V \geq 1/y) = \frac{y-1}{2y}
%\end{align*}
%where $1/3 < y < 1$ . Therefore, its density is given by $f_D(y) = \frac{1}{2}y^{-2}$, for $y\in (1/3,1)$. Such knowledge should then be inputted as exogenous information. If we assume the volume is uniform, then we \textit{know} that the density is not uniform, and viceversa. So the question posed by Keynes is then reduced to this: which of the two are we going to take as the uniform one ---the volume or density?
%
%Let us assume for the sake of the argument that these two are values that need to be fine tuned in order to permit life. Then, there is an interval $I_V=(a,b)$ in which life would be possible. There is also an interval $I_D = (c,d)$ in which life would be possible in terms of the density. Now, since we are talking about fine tuning it is then required for $C := (a,b) \cap (1/d,1/c)$, that its probability is bigger than 0 or life would not exist (which entails that $C\neq \emptyset$). So calling $U_V$ and $U_D$ the uniform distributions of $V$ and $D$, respectively with induced probability distributions $\textbf Q$ and $\textbf Q'$, respectively. Let's assume without loss of generality that $\textbf Q(C) > \textbf Q'(C)$. Then we take $V$ as the random variable with the uniform distribution and, in the worst case scenario (if in reality $\textbf Q'(C) > \textbf Q(C)$) what we are getting is an upper bound of the real probability under insufficient reason.
%
%Such reasoning is not invalidating the PrOIR. If anything, it is making it stronger because it is allowing us to be conservative in what we are not allowing nature by itself, giving it a power beyond maximum entropy itself to bound our knowledge of the probability of the life permitting set $C$.
%\section{References}
\begin{thebibliography}{00}
\bibitem{Basener2018} Basener, W. F., Sanford J. C. (2018)
The fundamental theorem of natural selection with mutations.
\textit{Journal of Mathematical Biology} \textbf{767}, 1589-1622.
\bibitem{Berestycki2009}
Berstycki N. (2009)
Recent progress in coalescent theory.
\textit{Ensaios Matem\'aticos} \textbf{16}(1), 1-193.
\bibitem{Chaitin2013}
Chaitin G. J. (2013)
\textit{Proving Darwin: Making Biology Mathematical}.
Vintage (New York).
\bibitem{CrowKimura1970}
Crow J. F., Kimura M. (1970)
\textit{An Introduction to Population Genetics Theory}.
Blackburn Press.
\bibitem{Dembski2009}
Dembski W. A., Marks II R. J. (2009)
Conservation of Information in Search: Measuring the Cost of Success.
\textit{IEEE Transactions on Systems, Man and Cybernetics A, Systems \& Humans} \textbf{5}(5), 1051-1061.
\bibitem{DiazMarks2020}
D\'{\i}az-Pach\'on D. A., Marks II R. J. (2020)
Generalized active information: Extensions to unbounded domains.
\textit{BIO-Complexity} \textbf{2020}(3):1-6. \url{doi:10.5048/BIO-C.2020.3}.
\bibitem{Durrett2008}
Durrett R. (2008)
\textit{Probability Models for DNA Sequence Evolution}.
Springer.
\bibitem{Etheridge2011}
Etheridge A. (2011)
\textit{Some Mathematical Models from Population Genetics}.
Springer.
\bibitem{Edwards2011}
Edwards, A. W. F. (2011)
Mathematizing Darwin.
\textit{Behavioral Ecology and Sociobiology} \textbf{65}(3), 421-430.
\bibitem{EwertDembskiMarks2009}
Ewert W., Dembski W. A, Marks II R. J. (2009)
Evolutionary Synthesis of Nand Logic: Dissecting a Digital Organism
\textit{Proceedings of the 2009 IEEE International Conference on Systems, Man, and Cybernetics. San Antonio, TX, USA, October}, 3047-3053.
\bibitem{EwertDembskiMarks2012}
Ewert W., Dembski W. A., Marks II R. J. (2012)
Climbing the Steiner Tree---Sources of Active Information in a Genetic Algorithm for Solving the Euclidean Steiner Tree Problem.
\textit{Bio-Complexity} \textbf{2012}(1), 1-14.
\bibitem{EwertDembskiMarks2013}
Ewert W., Dembski W. A., Marks II R. J. (2013)
Active Information in Metabiology.
\textit{BIO-Complexity}, \textbf{2013}(4), 1-10. \url{doi:10.5048/BIO-C.2013.4}.
\bibitem{Fisher1922}
Fisher R. A. (1922).
On the Dominance Ratio.
\textit{Proceedings of the Royal Society of Edinburgh} \textbf{42}, 321-341.
\bibitem{Kingman1982}
Kingman R. F. C. (1982)
The Coalescent.
\textit{Stochastic Processes and their Applications} \textbf{13}, 235-248.
\bibitem{Klebaner2012}
Klebaner F. (2012)
\textit{Introduction to Stochastic Processes with Applications.} Third edition.
Imperial College Press.
\bibitem{LenskiEtAl2003}
Lenski R. E., Ofria C., Pennock R. T., Adami C.
The evolutionary origin of complex features. (2003)
\textit{Nature} \textbf{423}, 139-144.
Supplementary material is available at \url{http://myxo.css.msu.edu/papers/nature2003}. An ``Index of Avida Documentation," is available at \url{http://dllab.caltech.edu/avida/v2.0/docs}.
\bibitem{MarksDembskiEwert2017}
Marks II R. J., Dembski W. A., Ewert W. (2017)
\textit{Introduction to Evolutionary Informatics}
World Scientific.
\bibitem{MontanezEtAl2010}
Monta\~{n}ez G., Ewert W., Dembski W. A., Marks II R. J. (2010)
Vivisection of the ev Computer Organism: Identifying Sources of Active Information.
\textit{BIO-Complexity}, \textbf{2010}(3), 1-6. \url{doi:10.5048/BIO-C.2010.3}.
\bibitem{Schneider2000}
Schneider T.D. (2000)
Evolution of biological information.
\textit{Nucleic Acids Res} \textbf{28}, 2794-2799. \url{doi:10.1093/nar/28.14.2794}.
\bibitem{Wolpert1997}
Wolpert D. H., Macready W. G. (1997)
No free lunch theorems for optimization.
\textit{IEEE transactions on evolutionary computation}\textbf 1(1), 67-82.
\end{thebibliography}
\end{document}
%%
%% End of file `elsarticle-template-num.tex'.