Abstract
We introduce a general solution concept for the Fokker–Planck (Kolmogorov) equation representing the diffusion limit of the Wright–Fisher model of random genetic drift for an arbitrary number of alleles at a single locus. This solution will continue beyond the transitions from the loss of alleles, that is, it will naturally extend to the boundary strata of the probability simplex on which the diffusion is defined. This also takes care of the degeneracy of the diffusion operator at the boundary. We shall then show the existence and uniqueness of a solution. From this solution, we can readily deduce information about the evolution of a Wright–Fisher population.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
Introduction
The random genetic drift model developed implicitly by Fisher in [11] and explicitly by Wright in [25], and henceforth called the Wright–Fisher model, is one of the most popular stochastic models in population genetics [2, 9]. In its simplest form, it is concerned with the evolution of the probabilities between non-overlapping generations in a population of fixed size of two alleles at a single diploid locus that are obtained from random sampling in the parental generation, without additional biological mechanisms like mutation, selection, or a spatial population structure. Generalizations to multiple alleles, several loci, inclusion of mutations and selection etc. then constituted an important part of mathematical population genetics. It is our aim to develop a general mathematical perspective on the Wright–Fisher model and its generalizations. In the present paper, we treat the case of multiple alleles at a single site. In a companion paper [23], we have discussed the simplest case of two alleles in more detail. Generalizations will be addressed in subsequent papers.
Let us first describe the basic mathematical contributions of Wright and Kimura. In 1945, Wright approximated the discrete process by a diffusion process that is continuous in space and time (continuous process, for short) and that can be described by a Fokker–Planck equation. In 1955, by solving this Fokker–Planck equation derived from the Wright–Fisher model, Kimura obtained an exact solution for the Wright–Fisher model in the case of two alleles (see [15]). Kimura [16] also developed an approximation for the solution of the Wright–Fisher model in the multi-allele case, and in 1956, he obtained ([17]) an exact solution of this model for three alleles and concluded that this can be generalized to arbitrarily many alleles. This yields more information about the Wright–Fisher model as well as the corresponding continuous process. Kimura’s solution, however, is not entirely satisfactory. For one thing, it depends on very clever algebraic manipulations so that the general mathematical structure is not very transparent, and this makes generalizations very difficult. Also, Kimura’s approach is local in the sense that it does not naturally incorporate the transitions resulting from the (irreversible) loss of one or more alleles in the population. Therefore, for instance the integral of his probability density function on its defined domain is not equal to 1.
As mentioned, while the original model of Wright and Fisher works with a finite population in discrete time, many mathematical insights into its behavior are derived from its diffusion approximation. After the original work of Wright and Kimura just described, a more systematic approach was developed within the theory of strongly-continuous semigroups and Markov processes. In this framework, the diffusion approximation for the multi-allele Wright–Fisher model was derived by Ethier and Nagylaki [6, 7], and a proof of convergence of the Markov chain to the diffusion process can be found in [5]. (In this paper, we are not concerned with deriving the diffusion approximation, but actually, this can be derived in a rather direct manner without having to appeal to the general theory, as we shall show elsewhere.) One may then derive existence and uniqueness results for solutions of the Fokker–Planck equation from the theory of strongly continuous semigroups [5, 6, 14]. As the diffusion operator of the diffusion approximation becomes degenerate at the boundary, the analysis at the boundary becomes difficult, and this issue is not addressed by the aforementioned results. Recent work of Epstein and Mazzeo [3, 4], however, treats the boundary regularity with general PDE methods.
The full structure of the Wright–Fisher model and its diffusion approximation, however, is only revealed when one can connect the dynamics before and after the loss of an allele, or in analytic terms, if one can extend the process from the interior of the probability simplex to all its boundary strata. In particular, this is needed to preserve the normalization of the probability distribution. Therefore, in this paper, we develop the definition of a general solution that naturally includes the transitions resulting from the disappearance of alleles and derive the formalism for its solution. Since this formalism is rather explicit, it will allow us to derive and generalize the known explicit formulas for the quantities associated with the Wright–Fisher diffusion model like expected waiting times for the loss of one or several alleles in a systematic manner. The key for our approach are evolution equations for the moments of the probability density and the duality between the forward and backward Kolmogorov equations. We show that there exists a unique global solution of the Fokker–Planck equation. Since, as explained, our concept of a solution is different from (and, as we believe, better adapted to the structure of the Wright–Fisher model than) those treated in the literature, insofar as it extends to the boundary, these results do not follow from the general results of the literature mentioned above.
In the present paper, we only treat genetic drift in the absence of mutation, selection, and recombination. Extensions that can be obtained on the basis of the formalism presented here, in particular to general recombination schemes, will be presented in subsequent publications.
The Global Solution of the Wright-Fisher Model
In this section, we shall first establish some notation, and then prove some propositions as well as the main theorem of this paper.
Notations
\(\Delta _n:=\{(x^1,x^2,\dots ,x^{n+1}):\, x^i\ge 0, \sum _{i=1}^{n+1} x^i=1\}\) is the standard n-simplex in \(\mathbb {R}^{n+1}\) representing the probabilities or relative frequencies of alleles \(A_1,\dots , A_{n+1}\) in our population. Often, however, it is advantageous to work in \(\mathbb {R}^n\) instead of \(\mathbb {R}^{n+1}\), and with \(e_0:= (0,\ldots ,0) \in \mathbb {R}^n,\, e_k:=(0,\ldots ,\underbrace{1}_{k{\mathrm{th}}},\ldots ,0)\in \mathbb {R}^n\), we therefore define
Moreover, we shall need the subsimplices corresponding to subsets of alleles, using the following notations
We shall also need some function spaces:
We can now define the differential operators for our Fokker–Planck equation:
where the coefficients are defined by
Finally, we shall need
Proposition 2.1
For each \(1\le k\le n\), \(m\ge 0, |\alpha |=\alpha ^1+\cdots +\alpha ^k=m\), the polynomial of degree m in k variables \(x=(x^{i_1},\ldots ,x^{i_k})\) in \(\overline{V_k^{(i_0,\ldots ,i_k)}}\)
where the \(a^{(k)}_{m,\beta }\) are inductively defined by
is the eigenvector of \(L_k^{(i_0,\ldots ,i_k)}\) corresponding to the eigenvalue \(\lambda ^{(k)}_m=\frac{(m+k)(m+k+1)}{2}\).
Proof
We have
By equalizing coefficients we obtain
and
This completes the proof. \(\square \)
Remark 2.2
-
When \(k=1\), \(X^{(1)}_{m,m}(x^1)\) is the \(m\mathrm{th}-\)Gegenbauer polynomial (up to a constant). Thus, the polynomials \(X_{m,\alpha }^{(k)}(x)\) can be understood as a generalization of the Gegenbauer polynomials to higher dimensions.
-
Because of this representation of eigenvectors, we can easily see that \(X_{m,\alpha }^{(n)}(x)\) is a basis of \(C^2(\overline{V_n})\).
Proposition 2.3
If \(X\in C^2(\overline{V_k^{(i_0,\ldots ,i_k)}})\) is an eigenvector of \(L_k^{(i_0,\ldots ,i_k)}\) corresponding to \(\lambda \), then \(w_k^{(i_0,\ldots ,i_k)}X\) is an eigenvector of \((L_k^{(i_0,\ldots ,i_k)})^*\) corresponding to \(\lambda \).
Proof
If \(X\in C^2(\overline{V_k^{(i_0,\ldots ,i_k)}})\) is an eigenvector of \(L_k^{(i_0,\ldots ,i_k)}\) corresponding to \(\lambda \), it follows that
This completes the proof. \(\square \)
Proposition 2.4
Let \(\nu \) be the exterior unit normal vector of the domain \(V_k^{(i_0,\ldots ,i_k)}\). Then we have
Proof
In fact, on the surface \((x^{s}=0)\), for some \(s\in \{i_1,\ldots , i_k \}\) we have \(\nu =-e_s\), and hence \(\sum \nolimits _{j\in \{i_1,\ldots , i_k \}} a_{ij}\nu ^j=a_{is}=x^s(\delta _{si}-x^i)=0\). On the surface \((x^{i_0}=0)\) we have \(\nu =\frac{1}{\sqrt{k}}(e_{i_1}+\cdots +e_{i_k})\), hence \(\sum \nolimits _{j\in \{i_1,\ldots , i_k \}} a_{ij}\nu ^j=\frac{1}{\sqrt{k}}\sum \nolimits _{j\in \{i_1,\ldots , i_k \}} a_{ij}=\frac{1}{\sqrt{k}}x^i x^{i_0} =0\). This completes the proof. \(\square \)
Proposition 2.5
\(L_k^{(i_0,\ldots ,i_k)}\) and \((L_k^{(i_0,\ldots ,i_k)})^*\) are weighted adjoints in \(H_k^{(i_0,\ldots , i_k)}\), i.e.
Proof
We put \(F^{(k)}_i(x):=\sum \nolimits _{j\in \{i_1,\ldots , i_k \}} \frac{\partial (a_{ij}(x)X(x))}{\partial x^j}\). Because of \(w_k^{(i_0,\ldots ,i_k)} Y \in C^\infty _0(\overline{V}_k^{(i_0,\ldots ,i_k)})\), the second Green formula, and Proposition 2.4, we have
\(\square \)
Proposition 2.6
In \(\overline{V}_k^{(i_0,\ldots ,i_k)}\), \(\{X^{(k)}_{m,\alpha } \}_{m\ge 0,|\alpha |=m}\) is a basis of \(H_k^{(i_0,\ldots , i_k)}\) which is orthogonal with respect to the weights \(w_k^{(i_0,\ldots ,i_k)}\), i.e.,
Proof
\(\{X^{(k)}_{m,\alpha }\}_{m\ge 0,|\alpha |=m}\) is a basis of \(H_k^{(i_0,\ldots , i_k)}\) because \(\{x^\alpha \}_{\alpha }\) is a basis of this space. To prove the orthogonality we apply the Propositions 2.1, 2.3, 2.7 as follows
Because \(\lambda ^{(k)}_{m}\ne \lambda ^{(k)}_{j}\), this finishes the proof. \(\square \)
Proposition 2.7
-
(i)
The spectrum of the operator \(L_k^{(i_0,\ldots ,i_k)}\) is
$$\begin{aligned} Spec(L_k^{(i_0,\ldots ,i_k)})=\bigcup _{m\ge 0} \left\rbrace \lambda _m^{(k)}=\frac{(m+k)(m+k+1)}{2}\right\lbrace =:\Lambda _k \end{aligned}$$and the eigenvectors of \(L_k^{(i_0,\ldots ,i_k)}\) corresponding to \(\lambda ^{(k)}_m\) are of the form
$$\begin{aligned} X=\sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } X^{(k)}_{m,\alpha }, \end{aligned}$$i.e., the eigenspace corresponding to \(\lambda ^{(k)}_m\) is of dimension \(k+m-1 \atopwithdelims (){k-1}\);
-
(ii)
The spectrum of the operator \(L_k\) is the same.
Proof
-
(i)
Proposition 2.1 implies that \(\Lambda _k\subseteq Spec(L_k^{(i_0,\ldots ,i_k)})\). Conversely, for \(\lambda \notin \Lambda _k\), we will prove that \(\lambda \) is not an eigenvalue of \(L_k^{(i_0,\ldots ,i_k)}\). In fact, assume that \(X \in H_k^{(i_0,\ldots , i_k)}\) such that \(L_k^{(i_0,\ldots ,i_k)} X= -\lambda X\) in \(H_k^{(i_0,\ldots , i_k)}\). Because \(\{X^{(k)}_{m,\alpha }\}_{m,\alpha }\) is an orthogonal basis of \(H_k^{(i_0,\ldots , i_k)}\) with respect to the weights \(w_k^{(i_0,\ldots ,i_k)}\) (Proposition 2.4), we can represent X by \(X=\sum \limits _{m=0}^\infty \sum \nolimits _{|\alpha |=m} d^{(k)}_{m,\alpha } X^{(k)}_{m,\alpha }\). It follows that
$$\begin{aligned} \begin{aligned} \sum \limits _{m=0}^\infty \sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } (-\lambda ^{(k)}_m)X^{(n)}_{m,\alpha }&=\sum \limits _{m=0}^\infty \sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } L_k^{(i_0,\ldots ,i_k)} X^{(k)}_{m,\alpha }\\&=L_k^{(i_0,\ldots ,i_k)} X\\&=-\lambda \sum \limits _{m=0}^\infty \sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } X^{(k)}_{m,\alpha }. \end{aligned} \end{aligned}$$For any \(j\ge 0\), \(|\beta |=j\), multiplying by \(w_k X^{(k)}_{j,\beta }\) and then integrating on \(\overline{V}_n\) we have
$$\begin{aligned} \begin{aligned}&\sum \limits _{|\alpha |=j} d^{(k)}_{j,\alpha } \lambda ^{(k)}_j\left( X^{(k)}_{j,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{j,\beta }\right) \\&\quad =\sum \limits _{|\alpha |=j} d^{(k)}_{j,\alpha } \lambda \left( X^{(k)}_{j,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{j,\beta }\right) , \quad \forall j\ge 0, |\beta |=j,\\&\Rightarrow \left( X^{(k)}_{j,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{j,\beta }\right) _{\beta ,\alpha } \left( d^{(k)}_{j,\alpha } \lambda ^{(k)}_j\right) _{\alpha }\\&\quad =\left( X^{(k)}_{j,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{j,\beta }\right) _{\beta ,\alpha } \left( d^{(k)}_{j,\alpha } \lambda \right) _{\alpha },\quad \quad \forall j\ge 0, |\beta |=j,\\&\Rightarrow d^{(k)}_{j,\alpha } \lambda ^{(k)}_j=d^{(k)}_{j,\alpha } \lambda ,\quad \forall j\ge 0, |\beta |=j, \text { because } det \left( X^{(k)}_{j,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{j,\beta }\right) _{\beta ,\alpha } \ne 0\\&\Rightarrow d^{(k)}_{j,\alpha } = 0,\quad \forall j\ge 0, |\alpha |=j, \text { because } \lambda \ne \lambda ^{(k)}_j. \end{aligned} \end{aligned}$$It follows that \(X=0\) in \(H_k^{(i_0,\ldots , i_k)}\). Therefore
$$\begin{aligned} Spec(L_k^{(i_0,\ldots ,i_k)})=\bigcup _{m\ge 0} \left\rbrace \lambda _m^{(k)}=\frac{(m+k)(m+k+1)}{2}\right\lbrace =\Lambda _k. \end{aligned}$$Moreover, assume that \(X\in H_k^{(i_0,\ldots , i_k)}\) is an eigenvector of \(L_k^{(i_0,\ldots ,i_k)}\) corresponding to \(\lambda ^{(k)}_j\), i.e., \(L_k^{(i_0,\ldots ,i_k)} X=-\lambda _j X\). We represent X by
$$\begin{aligned} X=\sum \limits _{m=0}^\infty \sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } X^{(k)}_{m,\alpha }. \end{aligned}$$It follows that
$$\begin{aligned} \begin{aligned} \sum \limits _{m=0}^\infty \sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } (-\lambda ^{(k)}_m)X^{(k)}_{m,\alpha }&=\sum \limits _{m=0}^\infty \sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } L_k^{(i_0,\ldots ,i_k)}X^{(k)}_{m,\alpha }\\&=L_k^{(i_0,\ldots ,i_k)} X\\&=-\lambda ^{(k)}_j \sum \limits _{m=0}^\infty \sum \limits _{|\alpha |=m} d^{(k)}_{m,\alpha } X^{(k)}_{m,\alpha }. \end{aligned} \end{aligned}$$For any \(i\ne j\), \(|\beta |=i\), multiplying by \(w_k X^{(k)}_{i,\beta }\) and then integrating on \(\overline{V}_n\) we have
$$\begin{aligned} \begin{aligned}&\sum \limits _{|\alpha |=i} d^{(k)}_{i,\alpha } \lambda ^{(k)}_i\left( X^{(k)}_{i,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{i,\beta }\right) \\&\quad = \sum \limits _{|\alpha |=i} d^{(k)}_{i,\alpha } \lambda ^{(k)}_j \left( X^{(k)}_{i,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{i,\beta }\right) , \quad \forall i\ne j, |\beta |=i,\\&\Rightarrow \left( X^{(k)}_{i,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{i,\beta }\right) _{\beta ,\alpha } (d^{(k)}_{i,\alpha } \lambda ^{(k)}_i)_{\alpha }\\&\quad =\left( X^{(k)}_{i,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{i,\beta }\right) _{\beta ,\alpha } (d^{(k)}_{i,\alpha } \lambda ^{(k)}_j)_{\alpha },\quad \forall i\ne j, |\beta |=i,\\&\Rightarrow d^{(k)}_{i,\alpha } \lambda ^{(k)}_i=d^{(k)}_{i,\alpha } \lambda ^{(k)}_j,\quad \forall i\ne j, |\beta |=i, \text { because } det \left( X^{(k)}_{i,\alpha },w_k^{(i_0,\ldots ,i_k)} X^{(k)}_{i,\beta }\right) _{\beta ,\alpha } \ne 0,\\&\Rightarrow d^{(k)}_{i,\alpha } = 0,\quad \forall i\ne j, |\alpha |=i, \text { because } \lambda ^{(k)}_i \ne \lambda ^{(k)}_j. \end{aligned} \end{aligned}$$It follows that
$$\begin{aligned} X=\sum \limits _{|\alpha |=j}d^{(k)}_{j,\alpha } X^{(k)}_{j,\alpha }. \end{aligned}$$This completes the proof.
-
(ii)
is obvious.\(\square \)
Definition of the Solution
We shall now formally derive the Fokker–Planck equation as the diffusion limit of the Wright–Fisher model and introduce our solution concept for this equation. We consider a diploid population of fixed size N with \(n+1\) possible alleles \(A_1,\ldots ,A_{n+1},\) at a given locus. Suppose that the individuals in the population are monoecious, that there are no selective differences between these alleles and no mutations. There are 2N alleles in the population in any generation, so it is sufficient to focus on the number \(Y_m=(Y_m^1,\ldots ,Y_m^n)\) of alleles \(A_1,\ldots ,A_n\) at generation time m. Assume that \(Y_0=i_0=(i_0^1,\ldots ,i_0^n)\) and according to the Wright–Fisher model, the alleles in generation \(m+1\) are derived by sampling with replacement from the alleles of generation m. Thus, the transition probability is
where
and
After rescaling
we have a discrete Markov chain \(X_t\) valued in \(\{0, \frac{1}{2N},\ldots ,1\}^n\) with \(t=1\) now corresponding to 2N generations. It is easy to see that
We now denote by \(m_\alpha (t)\) the \(\alpha ^{th}-\)moment of the distribution about zero at the \(t^{th}\) generation, i.e.,
Then
Expanding the right hand side and noting (3) we obtain the following recursion formula, under the assumption that the population number N is sufficiently large to neglect terms of order \(\frac{1}{N^2}\) and higher,
Under this assumption, the moments change very slowly per generation and we can replace this system of difference equations by a system of differential equations:
With the aim of finding a continuous process which approximates the above discrete process, we should look for a continuous Markov process \(\{X_t\}_{t\ge 0}\) valued in \([0,1]^n\) with the same conditions as (3) and (5). Denoting by u(x, t) the probability density function of this continuous process, the condition (3) implies (see for example [9, p. 137], or for a more rigorous analysis [6–8]) that u is a solution of the Fokker–Planck (Kolmogorov forward) equation
and the condition (5) implies
and hence, since the polynomials are dense in \(H_n\) w.r.t. the product \([.,.]_n\),
This leads us to the following definition of a solution.
Definition 2.8
We call \(u\in H\) a solution of the Fokker–Planck equation associated with the Wright–Fisher model if
We point out that the last of these equations implicitly contains the boundary behavior that we wish to impose upon our solution. This will become clear from our construction in the next section.
The Global Solution
In this subsection, we shall construct the solution and prove the existence as well as the uniqueness of the solution. The process of finding the solution is as follows: We first find the general solution of the Fokker–Planck equation (21) by the separation of variables method. Then we construct a solution depending on certain parameters. We then use the conditions of (9, 10) to determine the parameters. Finally, we verify that we have indeed found the solution.
Step 1 Working on \(V_n\), assume that \(u_n(\mathbf {x},t)=X(\mathbf {x})T(t)\) is a solution of the Fokker–Planck equation (21). Then we have
Clearly \(\lambda \) is a constant which is independent on T, X. From the Proposition 2.7 we obtain the local solution of the Eq. (21) of the form
where
is the eigenvalue of \(L_n\) and
are the corresponding eigenvectors of \(L_n\).
For \(m\ge 0, |\beta |=m\), we conclude from Proposition 2.3 that
It follows that
Therefore
Thus,
It follows that
Step 2 The solution \(u\in H\) satisfying (21) will be found in the following form
We use the condition (10) to obtain iteratively values of \(u_k, \, k=n-1,\ldots ,0\). In fact, assume that we want to calculate \(u^{(0,\ldots ,n-1)}_{n-1}(x^1,\ldots ,x^{n-1},0,t)\).
We note that, if we choose
then \(\phi (\mathbf {x})\) vanishes on faces of dimension at most \(n-1\) except the face \(V_{n-1}^{0,\ldots ,n-1}\). Therefore, the expectation of \(\phi \) will be
The left hand side can be calculated easily by the condition (10)
It follows that
The first part of the right hand side is
Therefore we can expand \(u^{(0,\ldots ,n-1)}_{n-1}(x^1,\ldots ,x^{n-1},0,t)\) as follows
Putting this formula into Eq. (12), we shall obtain all the coefficients \(c^{(n-1)}_{m,l,\alpha }\). Thus, we shall obtain \(u^{(0,\ldots ,n-1)}_{n-1}(x^1,\ldots ,x^{n-1},0,t)\). Similarly we shall obtain \(u_{n-1}\). And finally we shall obtain all \(u_k, k=n-1,\ldots ,0\). Thus, we shall obtain the global solution in the form
It is not difficult to show that u is a solution of the Fokker–Planck equation associated with the WF model.
Step 3 We can easily see that this solution is unique. In fact, assume that \(u_1,u_2\) are two solutions of the Fokker–Planck equation associated with the WF model. Then \(u=u_1-u_2\) will satisfy
It follows that
Therefore
Since \(\{1,\{x^i\}_i,\{w_k^{(i_0,\ldots , i_k)} X^{(k)}_{j,\alpha }\chi _{{V_k}^{(i_0,\ldots , i_k)}}\}_{1\le k\le n,(i_0,\ldots ,i_k)\in I_k,j\ge 0,|\alpha |=j} \}\) is also a basis of \(H_n\) it follows that \(u=0 \in H\).
In conclusion, we have established:
Theorem 2.9
The Fokker Planck equation associated with the Wright–Fisher model with \(n+1\) alleles possesses the unique solution
Example 2.10
To illustrate this process, we consider the case of three alleles.
We shall construct the global solution for the problem
where the global solution of the form
and the product is
Step 1: We find the local solution \(u_2\)
To define the coefficients \(c^{(2)}_{m,\alpha ^1,\alpha ^2}\) we use the initial condition and the orthogonality of the eigenvectors \(X^{(2)}_{m,\alpha ^1,\alpha ^2}\) to get
Since the matrix
is positive definite, we therefore have unique values of the \(c^{(2)}_{m,\alpha ^1,\alpha ^2}\). It follows that we have a unique local solution \(u_2\).
Step 2: We will use the moment condition to define all other coefficients of the global solution.
Firstly, we define the coefficients of \(u_1^{1,2}\) as follows
We note that
Therefore
It follows that
Thus we have
By equating the coefficients of \(e^{\alpha t}\) we obtain \(u_1^{1,2}\). Similarly we obtain \(u_1\). Then, we define the coefficients of \(u_0^1\) from the first moment.
Note that when \(\phi = x^i\), \(L_2^*(\phi )=0\), therefore \([u_t,\phi ]_2=0\) or
It follows that
Thus we obtain \(u_0^1(1,0,t)\). Similarly we get all \(u_0\). Therefore we obtain the global solution u.
It is easy to check that u is a global solution. To prove the uniqueness we proceed as follows: Assume that u is the difference of any two global solutions, i.e., u satisfies
We shall prove that
In fact,
It implies that
Therefore
We need only to prove that Eq. (17) holds for all
-
(1)
If \(n=0, m\ge 0\), we see that \(\phi \) can be generated from \(\{1,x^1,w_1(x^1)X^{(1)}_m(x^1)\}\), therefore \([u,\phi ]_2=0\);
-
(2)
If \(m=0, n\ge 0\), we see that \(\phi \) can be generated from \(\{1,x^2,w_1(x^2)X^{(1)}_m(x^2)\}\), therefore \([u,\phi ]_2=0\);
-
(3)
If \(n=1, m\ge 1\), we expand \((x^1)^{m-1}\) by
$$\begin{aligned} (x^1)^{m-1}=\sum _{k\ge 0} c_k X_k^{(1)}(x^1). \end{aligned}$$Note that
$$\begin{aligned} L_2^*\Big (x^1 x^2 X_k^{(1)}(x^1)\Big )=-\lambda _k^{(1)}x^1 x^2 X_k^{(1)}(x^1). \end{aligned}$$Therefore
$$\begin{aligned} \Big [u_t, x^1 x^2 X_k^{(1)}(x^1)]_2=[u, L_2^*\Big (x^1 x^2 X_k^{(1)}(x^1)\Big )\Big ]_2=-\lambda _k^{(1)} [u,x^1 x^2 X_k^{(1)}(x^1)]_2. \end{aligned}$$It follows that
$$\begin{aligned}{}[u,x^1 x^2 X_k^{(1)}(x^1)]_2=[u(0),x^1 x^2 X_k^{(1)}(x^1)]_2 e^{-\lambda ^{(1)}_k}=0. \end{aligned}$$Therefore
$$\begin{aligned}{}[u,\phi ]_2=\sum _{k\ge 0} c_k [u,x^1 x^2 X_k^{(1)}(x^1)]_2=0; \end{aligned}$$ -
(4)
If \(n\ge 2, m\ge 1\) we proceed by induction w.r.t. n. We have
$$\begin{aligned} (x^1)^m(x^2)^n&= x^1x^2(x^1+x^2-1)(x^1)^{m-1}(x^2)^{n-2}+(x^1)^m(1-x^1)(x^2)^{n-1}\\&=-w_2(x^1,x^2)(x^1)^{m-1}(x^2)^{n-2}+(x^1)^m(1-x^1)(x^2)^{n-1}. \end{aligned}$$By the inductive assumption, we have
$$\begin{aligned} \left[ u, (x^1)^m(1-x^1)(x^2)^{n-1}\right] _2=0. \end{aligned}$$Then, we expand \((x^1)^{m-1}(x^2)^{n-2}\) by
$$\begin{aligned} (x^1)^{m-1}(x^2)^{n-2}=\sum _{m,\alpha }c^{(2)}_{m,\alpha }X^{(2)}_{m,\alpha }(x^1,x^2). \end{aligned}$$Therefore
$$\begin{aligned} \left[ u,w_2(x^1,x^2)(x^1)^{m-1}(x^2)^{n-2}\right] _2=\sum _{m,\alpha }c^{(2)}_{m,\alpha }\left[ u,w_2(x^1,x^2)X^{(2)}_{m,\alpha }(x^1,x^2)\right] _2=0. \end{aligned}$$It follows that \([u,(x^1)^m(x^2)^n]_2=0\).
Thus, \(u=0\).
Applications
In this section, we present some applications of our global solution to the evolution of the process \((X_t)_{t\ge 0}\) such as the expectation and the second moment of the absorption time, the probability distribution of the absorption time for having \(k+1\) alleles, the probability of having exactly \(k+1\) alleles, the \(\alpha \mathrm{th}\) moments, the probability of heterogeneity, and the rate of loss of one allele in a population having \(k+1\) alleles. Several of our formulas are known from other methods, see [9, 15–17, 19, 20], but we emphasize here the general and unifying approach.
The Absorption Time for Having \(k+1\) Alleles
The moments of the sojourn and absorption times were derived by Nagylaki [22] for two alleles, and by Lessard and Lahaie [18] in the multi-allele case. We denote by \(T^{k+1}_{n+1}(p)=\inf \{{t>0: X_t\in \overline{V}_k}|X_0=p \}\) the first time when the population has (at most) \(k+1\) alleles. \(T^{k+1}_{n+1}(p)\) is a continuous random variable valued in \([0,\infty )\) and we denote by \(\phi (t,p)\) its probability density function. It is easy to see that \(\overline{V}_k\) is invariant under the process \((X_t)_{t \ge 0}\), i.e. if \(X_s \in \overline{V}_k\) then \(X_t\in \overline{V}_k\) for all \(t\ge s\) (once an allele is lost from the population, it can never again be recovered). We have the equality
It follows that
Therefore the expectation for the absorption time of having \(k+1\) alleles is (see [9, p. 194]):
and the second moment of this absorption time is (see [1, 20]):
In order to see what this means, we consider the case of three alleles:
and the product is
By expansion of eigenvectors, we have
where \(c^{(2)}_{m,\alpha }(\mathbf {p})\) is uniquely defined. We represent \(u_1(\mathbf {x};t)\) by
where the coefficients \(a^{\cdot ,\cdot }_{m}(x^1)\) are defined as follows:
Putting
we note that \(\psi _n(0)=\psi _n(1)=0\) and
It follows that
Therefore
We obtain by equating the coefficients in terms of \(e^{-\lambda t}\)
Remark 3.1
The coefficients of \(u_2\) occur in the representation of the coefficients of \(u_1\) because of the probability flux.
Similarly because of
We have
Thus,
The expectation for the absorption time of having only 1 allele is
We first calculate the first term; the other terms will be obtained similarly. To do this, we expand \(x^1\) in terms of the \(\psi _n(x^1)\)
We construct a sequence of entropy functions on [0, 1] as follows.
-
\(E_0(x)=-x\)
-
\(E_r(x)\) is the unique solution of the boundary value problem
$$\begin{aligned} \left\{ \begin{array}{l} L_1^*(E_r(x))=-r E_{r-1}(x)\\ E_r(0)=E_r(1)=0. \end{array}\right. \end{aligned}$$
By some simple calculations, we obtain some first entropy functions
-
(1)
\(E_0(x)=-x\)
-
(2)
\(E_1(x)=-2(1-x) \log (1-x) \)
-
(3)
\(E_2(x)=-8xz(x)+8(1-x)\log (1-x)\)
-
(4)
\(E_3(x)=48(1-x)u(x)+96[xz(x)-(1-x)\log (1-x)]\)
where
Lemma 3.2
The entropy functions satisfy
and more generally,
Proof
We have
Similarly we have
The proof for all r is similar. \(\square \)
From the Lemma, we can expand \(E_1(x^1)\) as
Therefore we have
Thus, we have
Remark 3.3
We can obtain the rth moments of this absorption time by the same method, i.e.
The Probability Distribution of the Absorption Time for Having \(k+1\) Alleles
We note that \(X_{T^{k+1}_{n+1}(p)}\) is a random variable valued in \(\overline{V_k}\). We consider the probability that this random variable takes its value in \(V_k^{(i_0,\ldots ,i_k)}\), i.e., the probability of the population at the first time having at most \(k+1\) alleles to consist precisely of the \(k+1\) alleles \(\{A_{i_0},\ldots ,A_{i_k}\}\). Let \(g_k\) be a function of k variables defined inductively by
Then we have
Theorem 3.4
Proof
Method 1: By proving that
and elementary combinatorial arguments, we immediately obtain the result (see [19]).
Method 2: By proving that it is the unique solution of the classical Dirichlet problem
\(\square \)
The Probability of Having Exactly \(k+1\) Alleles
The probability of having only the particular allele \(A_i\) is (see [12]):
The probability of having exactly the \((k+1)\) alleles \(\{A_0,\ldots ,A_k\}\) (the coexistence probability of alleles \(\{A_0,\ldots ,A_k\}\)) is (see [16, 20]):
The \(\alpha \mathrm{th}\) Moments
The \(\alpha \mathrm{th}\)-moments are (see [15–17]):
The Probability of Heterogeneity
The probability of heterogeneity is (see [16]):
The Rate of Loss of One Allele in a Population Having \(k+1\) Alleles
We have the solution of the form
The rate of loss of one allele in a population with \(k+1\) alleles equals the rate of decrease of
which is \(\lambda ^{(k)}_0=\frac{k(k+1)}{2}\). This means that the rate of loss of alleles in the population decreases as k gets smaller in the course of the process (see [10, 13, 16]).
Population Genetics
The Wright–Fisher model as the basic theoretical model in population genetics then can also be applied to population genetics data, although in most cases, the basic model needs to be suitably extended. For instance, single nucleotide polymorphisms (SNPs) can be modeled in this way. As the name indicates, in a SNP, there is a genetic variant in the population that differs from the rest of the population at a single nucleotide position in the genetic sequence. Thus, we may consider the background sequence and its variant as two different alleles, \(A_1, A_2\), and ask for the chances of the variant allele \(A_2\) to sweep the population, for the expected when this happens or when the variant will go extinct, etc., and apply the formulae given in Sect. 3. A detailed bioinformatical analysis of SNPs in the human genome was carried out in [21]. In that paper, the effects of recombination and of varying population size were analyzed, again on the basis of the multinomial distribution as in Sect. 2.2. In particular, a population bottleneck in the human history around 40,000 years ago could be identified on the basis of the current SNP distributions in the human population. For our methods to apply, we thus need to extend the basic Wright–Fisher model to include recombination—which we shall present in another paper—as well as varying population size. Similarly, data about mitochondrial DNA [24] indicate a population bottleneck in human history. Since mitochondrial DNA is solely inherited from the mother, we can treat this as a haploid Wright–Fisher model without recombination. Again, however, to be applicable, our model needs to be extended to handle variable population sizes.
Conclusion
We have developed a new global solution concept for the Fokker–Planck equation associated with the Wright–Fisher model, and we have proved the existence and uniqueness of this solution (Theorem 2.9). From this solution, we can easily read off the properties of the considered process, like the absorption time of having \(k+1\) alleles, the probability of having exactly \(k+1\) alleles, the \(\alpha \mathrm{th}\) moments, the probability of heterogeneity, and the rate of loss of one allele in a population having \(k+1\) alleles.
References
Buss, S.R., Clote, P.: Solving the Fisher–Wright and coalescence problems with a discrete Markov chain analysis. Adv. Appl. Probab. 36(4), 1175–1197 (2004)
Bürger, R.: The Mathematical Theory of Selection, Recombination, and Mutation. John Wiley, New York (2000)
Epstein, C.L., Mazzeo, R.: Wright–Fisher diffusion in one dimension. SIAM J. Math. Anal. 42, 568–608 (2010)
Epstein, C.L., Mazzeo, R.: Degenerate Diffusion Operators Arising in Population Biology. Princeton University Press, Princeton (2013)
Ethier, S.N.: A class of degenerate diffusion processes occurring in population genetics. Commun. Pure Appl. Math. 29, 483–493 (1976)
Ethier, S.N., Kurtz, T.: Markov Processes: Characterization and Convergence. Wiley (1986), New York (2005)
Ethier, S.N., Nagylaki, T.: Diffusion approximations of Markov chains with two time scales and applications to population genetics. Adv. Appl. Probab. 12, 14–49 (1980)
Ethier, S.N., Nagylaki, T.: Diffusion approximations of the two-locus Wright–Fisher model. J. Math. Biol. 27, 17–28 (1989)
Ewens, W.J.: Mathematical population genetics I. Theoretical introduction. In: Interdisciplinary Applied Mathematics, 2nd edn. Springer, New York (2004)
Felsenstein, J.: The rate of loss of multiple alleles in finite haploid populations. Theor. Popul. Biol. 2, 391–403 (1971)
Fisher, R.A.: On the dominance ratio. Proc. R. Soc. Edinb. 42, 321–341 (1922)
Gill, W.: Modified fixation probability in multiple alleles models in the asymmetric sharply-peaked landscape. J. Korean Phys. Soc. 55(2), 709–717 (2009)
Gladstien, K.: Subdivided populations the characteristic values and rate of loss of alleles. J. Appl. Prob. 241, 241–248 (1977)
Karlin, S., Taylor, H.M.: A Second Course in Stochastic Processes. Academic Press, New York (1981)
Kimura, M.: Solution of a process of random genetic drift with a continuous model. PNAS-USA 41(3), 144–150 (1955)
Kimura, M.: Random genetic drift in multi-allele locus. Evolution 9, 419–435 (1955)
Kimura, M.: Random genetic drift in a tri-allelic locus; exact solution with a continuous model. Biometrics 12, 57–66 (1956)
Lessard, S., Lahaie, P.: Fixation probability with multiple alleles and projected average allelic effect on selection. Theor. Popul. Biol. 75, 266–277 (2009)
Littler, R.A., Good, A.J.: Ages, extinction times, and first passage probabilities for a multiallele diffusion model with irreversible mutation. Theor. Popul. Biol. 1(3), 214–225 (1978)
Littler, R.A.: Loss of variability at one locus in a finite population. Math. Biosci. 25, 151–163 (1975)
Marth, G., et al.: Sequence variations in the public human genome data reflect a bottlenecked population history. Proc. Natl. Acad. Sci. 100.1, 376–381 (2003)
Nagylaki, T.: The decay of genetic variability in geographically structured populations. PNAS 71, 2932–2936 (1974)
Tran, T.D., Hofrichter, J., Jost, J.: An introduction to the mathematical structure of the Wright–Fisher model of population genetics. Theory Biosci. 1–10 (2012). doi:10.1007/s12064-012-0170-3
Walker, A.C., Smith, S.D., Smith, S.D.: Mitochondrial DNA and human evolution. Nature 325, 1–5 (1987)
Wright, S.: Evolution in Mendelian populations. Genetics 16, 97–159 (1931)
Acknowledgments
We would like to thank the anonymous referee for helpful comments, in particular suggestions concerning SNPs and mitochondrial DNA, as mentioned in Sect. 4. The research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013)/ERC Grant agreement no. 267087.
Author information
Authors and Affiliations
Corresponding author
Additional information
The research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013)/ERC Grant agreement no. 267087. T. D. Tran and J. Hofrichter have also been supported by the IMPRS “Mathematics in the Sciences”, and the material presented in this paper is largely based on the first author’s thesis.
Rights and permissions
About this article
Cite this article
Tran, T.D., Hofrichter, J. & Jost, J. A General Solution of the Wright–Fisher Model of Random Genetic Drift. Differ Equ Dyn Syst 27, 467–492 (2019). https://doi.org/10.1007/s12591-016-0289-7
Published:
Issue Date:
DOI: https://doi.org/10.1007/s12591-016-0289-7