Multiplicative Selection Model

• Multiplicative Selection Model is a framework where selection or attrition is implemented by multiplicative reweighting of a current state, linking evolutionary genetics, survival analysis, and missing data theories.
• It utilizes discrete replicator dynamics and coordination game formulations to model allele frequency updates and statistical selection, highlighting key methodologies like multiplicative weights updates.
• The model bridges diverse applications—from genetic fitness landscapes to Bayesian sparsity selection—by providing analytical tractability and unifying principles across disciplines.

Searching arXiv for the supplied papers to ground the article and verify the relevant usages of “multiplicative selection model.” arxiv_search(query="(Chastain et al., 2012)", max_results=5, sort_by="relevance") arxiv_search(query="(Chastain et al., 2012)", max_results=5) Searching for (Chastain et al., 2012). {"query":"(Chastain et al., 2012)","max_results":5} A multiplicative selection model is a family of formal constructions in which selection, attrition, or response acts through multiplicative reweighting of a current state. In population genetics, the canonical instance is the weak-selection haploid model, where allele frequencies evolve by the discrete replicator equation and coincide exactly with multiplicative-weights updates in a coordination game between genes (Chastain et al., 2012, Mehta et al., 2014). Closely related multiplicative formulations also appear in rank-ordered survival modeling (Fenner et al., 2017), in missing-not-at-random identification with instrumental variables (Zhang et al., 26 Sep 2025), and in statistical selection procedures based on multiplicative weights or multiplicative priors (Chaturvedi et al., 2020, Cao et al., 2019).

1. Terminological scope

The term is not reserved for a single universally standardized model. In the evolutionary literature, it refers to fitness-proportional updating of allele frequencies under haploid selection. In sociophysics and electoral modeling, it refers to a multiplicative-decrease process generating a rank-ordered survival function. In missing-data theory, it refers to a multiplicative restriction on the nonresponse mechanism. A plausible unifying description is that the state variable is transformed by proportional factors rather than by additive increments.

Domain	Core multiplicative form	Representative source
Weak-selection haploid evolution	Allele-frequency update by multiplicative fitness weighting	(Chastain et al., 2012, Mehta et al., 2014)
Rank-ordered survival	$F(i+1,s+1)=(1-\mu(i,s))F(i,s)$	(Fenner et al., 2017)
MNAR missing data with IVs	$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$	(Zhang et al., 26 Sep 2025)
Graphical-model learning and sparsity selection	Multiplicative-weights or multiplicative-prior updates	(Chaturvedi et al., 2020, Cao et al., 2019)

This multiplicative viewpoint matters because it links disparate areas to common analytical machinery: discrete replicator dynamics, regret bounds, transport equations, influence-function theory, and hierarchical Bayesian sparsity selection.

2. Weak selection in haploid population genetics

In "Multiplicative Updates in Coordination Games and the Theory of Evolution" (Chastain et al., 2012), the model is specified for two genes with allele-frequency vectors

$$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$

and a fitness-landscape matrix $W=(w_{ij})$ with $w_{ij}\in[1-s,1+s]$ for small $s>0$. Writing

$\Delta_{ij}:=\frac{w_{ij}-1}{s},$

the weak-selection dynamics on the Wright manifold are

$$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$

where

$$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$

and

$$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$

Since $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$0, one can also write

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$1

This is exactly the weak-selection form of the discrete replicator equation. If $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$2 and $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$3, then

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$4

The biological interpretation is that alleles are boosted in proportion to their expected fitness against the current allelic distribution at the other locus.

The 2014 analysis of haploid selection uses the same multiplicative structure in the standard single-locus form

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$5

and then lifts it to a two-locus coordination-game representation. There, natural selection is described as applying the discrete multiplicative-weights update separately to each gene (Mehta et al., 2014).

3. Coordination games, replicator dynamics, and mixability

The game-theoretic reformulation in (Chastain et al., 2012) treats gene $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$6 and gene $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$7 as the two players of a coordination game. Gene $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$8 chooses an allele $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$9, gene $$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$0 chooses an allele $$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$1, and both players receive payoff $$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$2. The mixed strategies are precisely the allele-frequency vectors $$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$3 and $$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$4. The expected payoff to pure strategy $$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$5 is

$$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$6

and symmetrically

$$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$7

Under multiplicative weights with step size $$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$8,

$$x(t)=(x_1(t),\dots,x_m(t))^T,\qquad y(t)=(y_1(t),\dots,y_n(t))^T,$$9

with an analogous update for $W=(w_{ij})$0. Thus natural selection under weak selection is tantamount to multiplicative updates in a coordination game.

A central quantity in this formulation is mixability. The mixability of allele $W=(w_{ij})$1 of gene $W=(w_{ij})$2 at time $W=(w_{ij})$3 is

$W=(w_{ij})$4

and that of allele $W=(w_{ij})$5 of gene $W=(w_{ij})$6 is

$W=(w_{ij})$7

These utilities are exactly the terms that drive the multiplicative updates. In this sense, the model gives a rigorous realization of the idea that natural selection under sex favors alleles with high average fitness against varied genetic backgrounds (Chastain et al., 2012).

The 2014 coordination-game treatment makes the same identification in the discrete replicator form

$W=(w_{ij})$8

where $W=(w_{ij})$9 is the common-payoff matrix of a symmetric coordination game. When $w_{ij}\in[1-s,1+s]$0 is identified with the genotype-fitness matrix $w_{ij}\in[1-s,1+s]$1, the term $w_{ij}\in[1-s,1+s]$2 is the average fitness, or mixability, of allele $w_{ij}\in[1-s,1+s]$3 at gene 1 (Mehta et al., 2014).

4. Fixed points, support size, and the fate of diversity

Two major results define the long-run theory. First, (Chastain et al., 2012) shows that under weak selection the discrete dynamics remain within $w_{ij}\in[1-s,1+s]$4 of the Wright manifold and converge. Equivalently, the multiplicative-weights regret bound implies that in the limit $w_{ij}\in[1-s,1+s]$5, the empirical payoffs of the evolving $w_{ij}\in[1-s,1+s]$6 do no worse than the best fixed pure pair $w_{ij}\in[1-s,1+s]$7 in hindsight. In particular, the population focuses on alleles of maximal cumulative mixability.

Second, (Mehta et al., 2014) proves a pointwise convergence theorem for coordination games with all row and column entries distinct: the replicator map is a diffeomorphism on the interior of $w_{ij}\in[1-s,1+s]$8, every orbit converges pointwise to a fixed point, and, except for an initial set of Lebesgue measure zero, every trajectory converges to a pure Nash equilibrium. The proof uses the Losert–Akin convergence theorem, a linearization argument showing that stable fixed points are weakly stable Nash equilibria, and the Center–Stable-Manifold Theorem to show that unstable fixed points have measure-zero basins of attraction.

A common misconception is that the existence of many mixed equilibria by itself implies long-run maintenance of polymorphism. The two analyses separate existence from attraction. At a fixed point, all alleles in the support of $w_{ij}\in[1-s,1+s]$9 must have the same expected payoff, and likewise for $s>0$0. If $s>0$1 and $s>0$2 have size $s>0$3, then necessarily both supports have size $s>0$4, and the corresponding $s>0$5 submatrix $s>0$6 of $s>0$7 satisfies

$s>0$8

for some scalar $s>0$9 and probability vectors $\Delta_{ij}:=\frac{w_{ij}-1}{s},$0. Writing $\Delta_{ij}:=\frac{w_{ij}-1}{s},$1, one obtains

$\Delta_{ij}:=\frac{w_{ij}-1}{s},$2

This yields explicit conditions for nontrivial mixed equilibria and shows that equilibria in two-person coordination games can have large supports (Chastain et al., 2012).

The same paper derives a probabilistic lower bound: if the $\Delta_{ij}:=\frac{w_{ij}-1}{s},$3 are iid symmetric continuous on $\Delta_{ij}:=\frac{w_{ij}-1}{s},$4, then a random $\Delta_{ij}:=\frac{w_{ij}-1}{s},$5 matrix $\Delta_{ij}:=\frac{w_{ij}-1}{s},$6 has $\Delta_{ij}:=\frac{w_{ij}-1}{s},$7 with all row- and column-sums positive with probability at least $\Delta_{ij}:=\frac{w_{ij}-1}{s},$8. Hence any $\Delta_{ij}:=\frac{w_{ij}-1}{s},$9 submatrix $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$0 has probability at least $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$1 of supporting a nontrivial equilibrium, and the expected number of such equilibria is at least

$$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$2

This suggests a two-level picture: mixed equilibria can be combinatorially abundant, but under the genericity assumptions of (Mehta et al., 2014) they do not attract a positive-measure set of initial conditions. Biologically, natural selection alone therefore leads almost surely to fixation, so long-term preservation of genetic diversity requires additional mechanisms such as mutation, recombination, or speciation (Mehta et al., 2014).

5. Rank-ordered survival and beta-like selection

In "A multiplicative process for generating a beta-like survival function with application to the UK 2016 EU referendum results" (Fenner et al., 2017), the multiplicative selection idea is recast as a generative model for rank-ordered vote shares. Let $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$3 be the expected survival in district $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$4 after processing up to stage $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$5, and let $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$6 be an attrition or mortality function. The discrete dynamics are

$$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$7

with boundary condition $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$8 for $$x_i(t+1)=x_i(t)\cdot \frac{1+s\cdot q_A(i;t)}{Z_x(t)},\qquad y_j(t+1)=y_j(t)\cdot \frac{1+s\cdot q_B(j;t)}{Z_y(t)},$$9. In a continuous approximation,

$$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$0

and the solution is

$$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$1

For the mixture attrition

$$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$2

the solution becomes

$$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$3

Fixing $$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$4 and introducing a scale constant $$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$5 yields the beta-like survival function

$$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$6

For large $$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$7, the asymptotic form is

$$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$8

The parameters have distinct shape roles: $$q_A(i;t):=\sum_{j=1}^n y_j(t)\Delta_{ij},\qquad q_B(j;t):=\sum_{i=1}^m x_i(t)\Delta_{ij},$$9 controls the intermediate-rank power-law decay, $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$0 controls the cutoff near $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$1, and $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$2 scales the curve. Estimation is performed by nonlinear least squares or direct maximum likelihood using

$$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$3

Applied to the UK result with $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$4 Local Authorities, the nonlinear fit for rank-ordered Leave shares gave $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$5, $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$6, $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$7, and $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$8; for Remain shares it gave $$Z_x(t)=\sum_{i=1}^m x_i(t)[1+s\cdot q_A(i;t)],\qquad Z_y(t)=\sum_{j=1}^n y_j(t)[1+s\cdot q_B(j;t)].$$9, $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$00, $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$01, and $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$02. The paper also rank-orders census covariates and reports that this method outperformed simple linear regression on raw census percentages, especially for covariates with strong geographic clustering (Fenner et al., 2017).

6. Missing-data identification and related selection frameworks

In missing-data theory, the multiplicative selection model appears as a restriction on the nonresponse mechanism rather than on a dynamical update. In "A Multiplicative Instrumental Variable Model for Data Missing Not-at-Random" (Zhang et al., 26 Sep 2025), each subject contributes outcome $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$03, missingness indicator $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$04, covariates $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$05, instrument $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$06, and unobserved factor $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$07. The core multiplicative selection assumption is

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$08

combined with $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$09 and $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$10. This excludes interaction between $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$11 and $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$12 on the log scale and leaves the degree of selection bias on the outcome scale unrestricted. Identification proceeds through a “single-arm Wald ratio”: if

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$13

and

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$14

then

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$15

The paper derives the efficient influence function, develops semiparametric multiply robust IV estimators, extends the framework to polytomous and continuous instruments, and reports that in a Botswana HIV survey application the IF-based MIV estimate of HIV prevalence among nonresponders was $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$16 $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$17, versus $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$18 among responders, leading to an adjusted overall estimate of $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$19 $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$20 (Zhang et al., 26 Sep 2025).

Related multiplicative formulations also appear in statistical selection and structure learning. For Gaussian graphical models, "Learning Gaussian Graphical Models via Multiplicative Weights" uses a Sparsitron update in which a positive weight vector is updated coordinatewise by

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$21

with per-node runtime $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$22, total runtime $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$23, and sample complexity

$P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$24

under the stated edge-strength, $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$25-norm, variance, and diagonal bounds (Chaturvedi et al., 2020). For high-dimensional Gaussian DAG models, "Consistent Bayesian Sparsity Selection for High-dimensional Gaussian DAG Models with Multiplicative and Beta-mixture Priors" introduces a multiplicative prior in which node-specific latent weights $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$26 determine edge sparsity in the Cholesky factor of the precision matrix, and proves posterior-mode consistency under assumptions A1–A6; in the reported simulation table at $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$27 sparsity, SSC–Multiplicative attained PPV $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$28, TPR $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$29, and MCC $P\{R=0\mid Z,U,X\}=\exp\{\alpha_z+\alpha_u\}$30 (Cao et al., 2019).

These later uses are not identical to the haploid evolutionary model, but they preserve the same structural idea: selection or recovery is implemented through multiplicative reweighting, multiplicative factorization, or multiplicative priors. Across domains, that structure makes the models analytically tractable and connects them to established methods in online learning, dynamical systems, semiparametric inference, and Bayesian high-dimensional statistics.