Contents
27 — The four forces and effective population size Part 5 — Population genetics

27 — The four forces and effective population size

Before this: Ch 26 · Ch 16 · Time: ~55 min

Hardy–Weinberg says nothing changes. This chapter is about the four things that make it change, how fast each one works, and which one wins when they disagree.

What you'll be able to do

The core idea

Chapter 26 established a null model. Random mating, no forces: frequencies snap to p², 2pq, q² in one generation and stay there forever. That is a fixed point, and it is boring by construction — its whole purpose is to be violated.

Exactly four processes violate it. creates . Selection changes their frequency according to their consequences. Migration imports frequencies from elsewhere. Genetic changes them for no reason at all, because a finite population is a finite sample.

The first three are directional: given the parameters, you can write down Δp and integrate it. Drift has expectation zero and only a variance. So a population's trajectory is a stochastic process with a deterministic term and a noise term, and almost every question in population genetics reduces to their ratio.

A population is a noisy optimiser. Selection is the gradient, genetic drift is the sampling noise, and Ne·s is the signal-to-noise ratio. Below |Nes| ≈ 1 the gradient is invisible and what the allele actually does stops mattering — which is why the same mutation can be under strong selection in a fly and effectively neutral in a human.

If you have written a stochastic optimiser, you already own the intuition. What you do not own is the biology-specific part: the noise amplitude is set by Ne, which is not the number of organisms, and the gradient is filtered through dominance, which decides whether selection can see the allele at all.

Notation throughout: two alleles A₁ and A₂ at one autosomal locus, frequencies p and q = 1 − p, discrete non-overlapping generations, random mating within a population.


1. Mutation: the only source of novelty, and hopeless on its own

Let μ be the forward rate A₁ → A₂ per copy per generation, and ν the back rate A₂ → A₁. Every A₁ copy has probability μ of leaving, every A₂ copy probability ν of returning:

p' = p(1 − μ) + qν
Δp = p' − p = −μp + νq

Substituting q = 1 − p gives Δp = ν − (μ + ν)p, which is linear in p, so the recursion is a contraction with a single fixed point. Setting Δp = 0:

p̂ = ν/(μ + ν), and q̂ = μ/(μ + ν) — the equilibrium is the ratio of rates, not of anything biological.

The approach is geometric. Write dt = pt − p̂; then dt+1 = pt+1 − p̂ = ν − (μ+ν)pt + pt − p̂ = (1 − μ − ν)dt, so

p_t − p̂ = (p_0 − p̂)(1 − μ − ν)^t          half-life  t½ = ln2/(μ + ν)

Now put numbers in it. Per-locus are of order 10⁻⁶ to 10⁻⁵ per generation (a coding sequence of ~1.5 kb, at the pinned genome-wide rate of ~1.3 × 10⁻⁸ per bp per generation (verified facts), takes ~2 × 10⁻⁵ point changes per generation, a fraction of which inactivate the gene). Take μ + ν = 10⁻⁵:

t½ ≈ 69,000 generations ≈ 1.9 million years at the pinned human generation time of 27 years (verified facts).

That is the entire lesson. alone is a negligible force on any timescale you care about. It matters in exactly two ways: it is the sole origin of every allele that any other force acts on, and it becomes important when paired with something else — with selection (§3) or with (§4, where the compound parameter 4Neμ governs how much variation a population carries).

Assumptions: infinite population, constant rates, no selection. Drop the first and the deterministic equilibrium becomes a stationary distribution rather than a point.

2. Selection: fitness, s, h, and the general one-locus recursion

Absolute W is the expected number of offspring an individual of a given contributes to the next generation — viability × fertility, one number absorbing survival, mating success and fecundity. Only ratios matter, so we use relative fitness w = W/Wmax. Parameterise:

GenotypeA₁A₁A₁A₂A₂A₂
Frequency (after HWE)p²2pqq²
Relative fitness11 − hs1 − s

s is the : the fitness cost of being homozygous A₂A₂. h is the dominance coefficient: the fraction of that cost paid by the heterozygote. h = 0 means A₂ is fully (heterozygote unaffected), h = 1 fully , h = ½ additive, h < 0 overdominant, h > 1 underdominant. Dominance is not a property of the — it is a number describing where the heterozygote sits, and it is the single hardest parameter in this chapter to measure.

Mean fitness is the normaliser:

w̄ = p²·1 + 2pq(1 − hs) + q²(1 − s) = 1 − 2pq·hs − q²s

Selection reweights genotypes by fitness, then extracts alleles. The frequency of A₂ after selection is its share of the fitness-weighted gene pool — A₂A₂ contributes all its alleles, heterozygotes half:

q' = [ q²(1 − s) + pq(1 − hs) ] / w̄  =  q[ q(1 − s) + p(1 − hs) ] / w̄

Subtract q = q·w̄/w̄ and expand the numerator, using p + q = 1:

q·[ q(1−s) + p(1−hs) − w̄ ]
 = q·[ (p+q) − qs − phs − 1 + 2pq·hs + q²s ]
 = q·[ −qs(1 − q) − phs(1 − 2q) ]
 = −pq·s·[ q + h(1 − 2q) ]

Δq = −pq·s·[ q + h(1 − 2q) ] / w̄

That is the whole of one-locus selection theory. Everything below is a special case.

Two structural facts before the cases. First, the factor pq — selection needs variation to act on, and stops entirely at q = 0 or q = 1. Second, differentiate w̄ and compare:

dw̄/dq = −2hs(1 − 2q) − 2qs        ⇒        Δq = (pq/2) · d(ln w̄)/dq

Selection is gradient ascent on log mean fitness, with a step size of pq/2 — that is, proportional to the genetic variance available. This is Wright's adaptive topography, and it is the seed of the in Ch 31. It also immediately tells you the limits: with a single locus w̄ increases monotonically, so selection cannot cross a fitness valley — a fact that returns as underdominance below.

The four regimes

RegimehΔq when A₂ is rare (q → 0)Speed
Deleterious recessive0−s q²O(q²) — collapses as q falls
Deleterious additive½−sq/2O(q) — geometric
Deleterious dominant1−sqO(q) — geometric, fastest
Overdominant< 0Δq > 0pushes toward a stable interior point
Underdominant> 1Δq < 0 near 0, > 0 near 1interior point is unstable

Selection against a recessive is the case that matters clinically, and it is brutal. Set h = 0: Δq = −pq²s/w̄. The q² is the whole story — selection can only see A₂ in homozygotes, and homozygotes are q² of the population. Equivalently: of the 2q copies of A₂ per individual, the fraction sitting in homozygotes is 2q²/2q = q. At q = 0.01, 99% of all copies of the allele are hiding in heterozygotes, invisible to selection.

Take the most extreme possible case — a fully recessive lethal, s = 1, h = 0. Then w̄ = 1 − q², and

q' = pq/(1 − q²) = q(1−q)/[(1−q)(1+q)] = q/(1 + q)
⇒  1/q' = 1/q + 1        ⇒        1/q_t = 1/q_0 + t

The reciprocal is linear in time. To go from q₀ = 0.01 to q₁ = 0.005 takes t = 1/0.005 − 1/0.01 = 100 generations — roughly 2,500 years — of every affected individual leaving zero offspring. To halve it again: 200 more generations. This is the quantitative refutation of every proposal to eliminate recessive disease alleles by discouraging reproduction: the arithmetic says it does not work even under assumptions far more favourable than reality.

Selection against a dominant is fast. With h = 1 and q small, Δq ≈ −sq, so qt ≈ q₀(1 − s)t — geometric decay with a per-generation factor, independent of how rare the allele is. Every copy is exposed, because every carrier shows the . A dominant lethal (s = 1) is gone in one generation.

(heterozygote advantage) is the case that maintains variation. Reparameterise so the heterozygote is best: w₁₁ = 1 − s₁, w₁₂ = 1, w₂₂ = 1 − s₂, both s > 0. Then w̄ = 1 − s₁p² − s₂q², and repeating the algebra above:

q' = q(1 − s₂q)/w̄
Δq = q[ (1 − s₂q) − w̄ ]/w̄ = q[ s₁p² − s₂q(1−q) ]/w̄ = pq(s₁p − s₂q)/w̄

Setting Δq = 0 with 0 < q < 1 requires s₁p = s₂q, and with p = 1 − q:

q̂ = s₁/(s₁ + s₂)

It is stable: for q < q̂ the bracket s₁p − s₂q is positive so q rises, and for q > q̂ it is negative so q falls. The equilibrium is a ratio of the two homozygote costs — the worse homozygote's allele sits at the lower frequency. Worked numerically for sickle cell below.

Underdominance uses the same algebra with the heterozygote worst (w₁₂ < both homozygotes). The interior equilibrium exists at the same kind of ratio but the sign of Δq flips on each side, so it repels: whichever allele starts above the threshold goes to , the other is lost. This is the structure of most chromosomal rearrangements — a heterozygous or carrier mis-segregates at meiosis (Ch 20) — and it explains why a new rearrangement almost never spreads: it must cross a fitness valley, which selection cannot do. The only route across is , which is why rearrangements fix in small or founder populations and essentially nowhere else.

3. Mutation–selection balance

A deleterious is removed by selection and recreated by . Equate the two rates.

Per generation, mutation adds Δqmut = μp ≈ μ (since p ≈ 1 for a rare deleterious allele; back mutation is negligible against selection). Selection removes Δqsel ≈ −sq(q + h) after setting p ≈ 1, w̄ ≈ 1 and (1 − 2q) ≈ 1. Balance:

μ = s·q̂·(q̂ + h)

case (h = 0): μ = s q̂² ⇒ q̂ = √(μ/s)

or partially dominant case (h ≫ q̂): μ = s h q̂ ⇒ q̂ = μ/(hs)

The square root in the recessive case is the point. With μ = 10⁻⁶ and s = 1, a recessive sits at q̂ = 10⁻³ — a 2q̂ = 1/500, a disease incidence q̂² = 1 in a million. A fully dominant allele with the same μ and s sits at q̂ = 10⁻⁶, a thousandfold lower. Recessiveness is a hiding place, and the size of the hiding place is 1/√μ.

Note the crossover condition carefully. The recessive formula requires q̂ ≫ h, i.e. h ≪ √(μ/s) ≈ 10⁻³. Most alleles called "recessive" clinically are not that recessive — measured h for human loss-of-function and disease alleles is typically ~0.05–0.2, not ~0 (genes behind autosomal recessive disease average about 0.2; genome-wide inference for nonsynonymous mutations puts the strongly deleterious class no lower than ~0.05) — roughly two orders of magnitude above the 10⁻³ threshold. In that regime the dominant formula μ/(hs) governs and the allele is far rarer than √(μ/s) predicts. When a mutation–selection calculation is off by an order of magnitude, mis-specified h is the usual culprit.

Where the model works, and where it doesn't:

Achondroplasia is the textbook success. It is dominant (h = 1), reproductive is roughly 0.2 so s ≈ 0.8, and the is unusually high — a single hotspot nucleotide in FGFR3 with an estimated ~1.4 × 10⁻⁵ per gamete per generation (estimates range to ~5 × 10⁻⁵), inflated further by selfish selection among spermatogonia. Then

q̂ = μ/(hs) = 1.4×10⁻⁵ / 0.8 = 1.75×10⁻⁵
incidence ≈ 2q̂ = 3.5×10⁻⁵ ≈ 1 in 29,000        (observed: ~1 in 15,000–30,000)

And a second, sharper prediction falls out for free. At equilibrium, new mutations produce 2μ affected zygotes per zygote, while total incidence is 2μ/s; so the fraction of cases that are de novo equals s = 0.8. About 80% of achondroplasia cases do arise from new mutation. A one-line model predicting two independent observables is doing real work.

Cystic fibrosis is the instructive failure. Incidence in European-ancestry populations is ~1/2,500, so q ≈ 0.02 and the carrier frequency 2pq ≈ 1/25. Before modern treatment s ≈ 1. then demands μ = sq² = 4 × 10⁻⁴ per generation — 400-fold higher than any plausible per-locus rate. The model is not merely imprecise here; it is refuted. Something else is holding q up: heterozygote advantage (cholera, typhoid and tuberculosis have all been proposed, none convincingly demonstrated), or and in a historically bottlenecked population — which §7 will argue is a plausible alternative that does not require any heterozygote advantage. Both remain live.

4. Drift: a variance process, not a force

Model a finite population as Wright–Fisher: N individuals produce an effectively infinite gamete pool at frequency p, and the next generation is 2N gametes sampled from it with replacement. Then the count of A₂ is Binomial(2N, p), and

E[p'] = p                        Var(p') = p(1 − p)/(2N)

The expectation is unchanged. does not push toward loss, toward , or toward 0.5. It has no direction. It is — over an ensemble of replicate populations the mean frequency stays put while the distribution spreads, until probability mass piles up against the absorbing boundaries at 0 and 1.

 20 replicate populations, N = 20, p₀ = 0.5, one column per generation
 gen:      0     5    10    20    40    80
 mean p  0.50  0.50  0.49  0.51  0.50  0.50    ← unchanged
 sd(p)   0.00  0.17  0.23  0.31  0.39  0.47    ← grows toward its maximum, 0.5
 fixed    0/20  0/20  1/20  4/20 10/20 17/20   ← mass accumulates at the boundaries

Heterozygosity decays deterministically even though frequency does not. Take two allele copies at random from generation t+1. With probability 1/(2N) they are copies of the same parental allele and are therefore identical; otherwise they are two independent draws from generation t. So the probability they differ is

H_{t+1} = (1 − 1/2N)·H_t        ⇒        H_t = H_0(1 − 1/2N)^t ≈ H_0·e^{−t/2N}

Variation is lost at a rate 1/(2N) per generation, and the timescale of drift is 2N generations. This is the definition that §7 will invert to define Ne.

Fixation probability of a neutral allele = its current frequency. Two derivations, both one line. (i) Exchangeability: exactly one of the 2N allele copies present today is the ancestor of every copy in the distant future, and by neutrality each is equally likely; a fraction p of them are A₂. (ii) Martingale: pt has constant expectation and is bounded, so it converges, and the only absorbing states are 0 and 1; hence P(fix)·1 + P(loss)·0 = p₀.

Consequences worth holding:

5. Migration: how little it takes

Continent–island model. Each generation a fraction m of the island's gene pool is replaced by immigrants from a large mainland with frequency pm:

p' = (1 − m)p + m·p_m          Δp = −m(p − p_m)
p_t − p_m = (p_0 − p_m)(1 − m)^t

Same geometric structure as , same fixed point logic — but m is typically 10⁻² to 10⁻³, two to four orders of magnitude larger than the per-locus μ of §1. At m = 0.01 the difference between island and mainland halves every ln2/m ≈ 69 generations. Migration is a fast homogeniser.

Now the fact that surprises people. pushes two populations apart at rate ~1/(2N); migration pulls them together at rate m. Balance them properly. Wright's island model: a large number of demes, each of size N, each drawing a fraction m of its genes each generation from the pool at large.

Let f be the probability that two copies drawn from the same deme are identical by descent — that they trace back to one ancestral copy with no migration event in between. Two things have to happen. First, neither lineage arrived as an immigrant in the previous generation: probability (1 − m)². Second, given both are resident, they either came from the same parental copy — probability 1/(2N), the drift rate of §4, in which case they are identical by descent for certain — or from different copies, probability 1 − 1/(2N), in which case they are identical with the previous generation's probability f. So

f' = (1 − m)²·[ 1/(2N) + (1 − 1/(2N))·f ]

Set f′ = f for the equilibrium and write λ = (1 − m)². The recursion is linear in f, so

f = λ / [ 2N(1 − λ) + λ ]

Now use m ≪ 1: λ ≈ 1 − 2m, hence 1 − λ ≈ 2m, and the denominator becomes 4Nm + 1. That is it:

F_ST ≈ 1/(1 + 4Nm)

This f is FST: it is the homozygosity a deme carries in excess of the metapopulation, which is exactly what Ch 28 §7 (coming next) defines F-statistics to measure. The approximation error tracks m, not Nm — about 1.5% at m = 0.01 — so it is safe wherever migration is a small fraction of the deme per generation, which is the only regime the model is about.

The parameter is Nm — the absolute number of migrants per generation — not the proportion m. A population of 100 and a population of 10⁷ both need the same number of immigrants to stay genetically homogeneous with their neighbours.

Nm (migrants/generation)FST
0.10.71 — strong divergence
10.20 — modest divergence
100.024 — effectively one population

One migrant per generation — one individual, whatever the population size — holds FST near 0.2 and prevents drift-driven divergence from running away. This is why genetic differentiation between human populations is small, and why genetics worries about corridors rather than absolute isolation — Ch 35A §7 makes that a management decision, and the Isle Royale wolves are what happens when the migrant arrives once instead of every generation.

6. The inequality that decides everything: |Nes|

Selection is deterministic and is stochastic, so their relative importance depends on the timescale you compare them over. Do it properly.

Scaling argument. Over T generations, selection's cumulative deterministic change is about T·s·pq. Drift's cumulative standard deviation is about √(T·pq/2N) — a random walk with per-step variance pq/2N. The relevant T is the timescale on which drift resolves anything, namely T ≈ 2N (§4). Substituting:

selection:  2N·s·pq            drift:  √(pq)
ratio    ≈  2Ns·√(pq)   →   order Ns

Note this is not the same as the per-generation ratio, which scales as s√N. The compound parameter is Ns because drift needs 2N generations to act, and selection gets all of them.

Diffusion argument (exact). Kimura's probability for an at frequency p under genic selection, with α ≡ 4Nes the scaled :

u(p) = (1 − e^{−αp}) / (1 − e^{−α})

Everything depends on p and α alone — Ne and s never appear separately. Take the two limits.

Strong selection, α ≫ 1. The denominator → 1 and for a new p = 1/(2N), u ≈ 1 − e−2s ≈ 2s. Haldane's result: a new beneficial mutation with a 1% advantage fixes with probability ~2%, and is lost 98% of the time. Compared with the neutral 1/(2N), selection has multiplied its chances by ~4Nes.

Weak selection, |α| ≪ 1. Expand both exponentials to second order:

u(p) ≈ [αp − (αp)²/2] / [α − α²/2] = p(1 − αp/2)/(1 − α/2) ≈ p·[1 + (α/2)(1 − p)]

The fractional deviation from the neutral answer u = p is of order α/2 = 2Nes.

When |Nes| ≪ 1 an allele behaves exactly as if it were neutral, no matter what it does to the organism. When |Nes| ≫ 1 drift is a rounding error and the deterministic equations of §2 are accurate. The band around |Nes| ≈ 1 is the nearly neutral zone, and it is where most functionally relevant variation actually sits.

graph TD M["new mutation<br/>p = 1/2N"] --> C{"|N<sub>e</sub>s| ?"} C -->|"≪ 1"| N["<b>effectively neutral</b><br/>P(fix) ≈ 1/2N<br/>fate set by drift alone"] C -->|"≈ 1"| NN["<b>nearly neutral</b><br/>drift and selection comparable<br/>outcome depends on N<sub>e</sub> history"] C -->|"≫ 1, s &gt; 0"| B["<b>beneficial</b><br/>P(fix) ≈ 2s<br/>still lost ~1−2s of the time"] C -->|"≫ 1, s &lt; 0"| D["<b>deleterious</b><br/>P(fix) ≈ 0<br/>held at mutation–selection balance"]

The consequences are large and not intuitive:

7. Effective population size

Everything above used N as if it were the number of organisms. It is not. Define:

Ne is the size of an ideal Wright–Fisher population that would experience the same rate of as the population in question.

"The same rate of what" needs specifying, and the answers differ:

FlavourDefined byFormula
Variance Nematching Var(Δp)Ne = pq/(2·Var(Δp))
Inbreeding Nematching the rate of heterozygosity loss1/(2Ne) = −ΔH/H
Nematching the expected time to common ancestryE[T₂] = 2Ne

They coincide for an ideal constant population and diverge otherwise — a real source of confusion when comparing estimates across papers.

Why Ne < N

Unequal sex ratio. Every autosomal in an offspring came half from a male and half from a female, so the sexes contribute equally to the gene pool regardless of their numbers. The rarer sex is a bottleneck:

N_e = 4·N_m·N_f/(N_m + N_f)
500 males + 500 females  →  N_e = 1,000    (= N)
 10 males + 990 females  →  N_e = 4·10·990/1000 = 39.6

A thousand breeding adults with ten breeding males drifts like a population of forty.

Variance in reproductive success. Wright–Fisher assumes offspring number is Poisson, so Vk = k̄ = 2 in a constant population. For general Vk:

N_e ≈ (4N − 2)/(V_k + 2)
V_k = 2  (Poisson, ideal)  →  N_e ≈ N
V_k = 10 (harems, sweepstakes reproduction) →  N_e ≈ N/3
V_k = 0  (every adult contributes exactly 2) →  N_e ≈ 2N

Note the last row: Ne can exceed N, if reproduction is more even than random. Managed breeding programmes exploit exactly this.

Fluctuating size — and why the harmonic mean. This is the one to internalise, and it falls straight out of §4. Heterozygosity multiplies across generations:

H_t/H_0 = Π_i (1 − 1/2N_i) ≈ exp(−Σ_i 1/(2N_i))

Equate to the ideal population's exp(−t/2Ne):

t/N_e = Σ_i 1/N_i        ⇒        1/N_e = (1/t)·Σ_i 1/N_i     ← harmonic mean

Drift accumulates as a sum of reciprocals, so a single small generation dominates everything around it. Sizes 1,000 / 1,000 / 10 / 1,000 / 1,000 have arithmetic mean 802 and

(1/5)(0.001 + 0.001 + 0.1 + 0.001 + 0.001) = 0.0208   →   N_e = 48

One crash generation to 10 individuals makes a population of ~800 drift like a population of 48. Harmonic means ignore large values; a population's genetic memory is a record of its worst moments.

Non-random mating and structure. Inbreeding reduces the effective number of independently sampled gametes, roughly Ne ≈ N/(1 + F) (Ch 28). Subdivision cuts both ways: it lowers Ne locally within demes while raising the species-wide coalescent Ne, because lineages in different demes take a long time to meet.

Linked selection. A or ongoing removes neutral variants hitchhiking nearby, which locally looks exactly like a reduced Ne. So Ne is not even constant along a : it dips in regions of low recombination and high gene density (Ch 33).

Human Ne ≈ 10⁴ versus 8 billion

Estimate it from diversity. At neutral drift– equilibrium, is θ = 4Neμ. Two random human haploid genomes differ at roughly 0.8–1.0 per 1,000 sites, and the pinned rate is ~1.3 × 10⁻⁸ per bp per generation (verified facts):

N_e = θ/(4μ) = 0.0009 / (4 × 1.3×10⁻⁸) ≈ 1.7 × 10⁴

Order 10⁴, consistent with the 10,000–20,000 range from independent sequence-based estimates. The census population is ~8 × 10⁹, so Ne/N ≈ 10⁻⁶.

What that does not mean. It is not a claim that only 10,000 humans were ever alive at once. Ne here is a harmonic-mean-like summary over hundreds of thousands of years of history dominated by small, structured, fluctuating populations — and harmonic means are insensitive to large values, so the last few thousand years of explosive growth contribute almost nothing to it. It is also inflated by ancient population structure, which lengthens coalescence times without any single population ever being large.

Two consequences that matter downstream. First, humans are genetically depauperate for a large mammal — two humans from opposite sides of the world are more similar than two chimpanzees from neighbouring forests. Second, the recent explosion has generated an enormous excess of very rare variants that the equilibrium Ne does not describe at all: most human variation is young, rare, and population-specific, which is precisely why rare-variant association is hard (Ch 54) and why transfer poorly across populations (Ch 53).

The three corrections above are not alternatives — they multiply. Applied together to real wildlife populations they land on a ratio Ne/N of about 0.1, so a census of 500 breeding adults drifts and inbreeds like 50. Ch 35A §7 does that arithmetic on a managed population, chains it to ΔF = 1/(2Ne), and turns it into the decision the numbers exist for.

Bottlenecks and founder effects

A bottleneck is a sharp temporary reduction in N. A is the special case where a new population is established by a few individuals. Both are Ne events, and both leave a characteristic asymmetric signature.

Consider a crash to N = 10 (20 gametes) for one generation:

Bottlenecks destroy allelic richness far faster than they destroy heterozygosity, because rare alleles carry almost no heterozygosity but are the overwhelming majority of alleles. Detecting recent bottlenecks exploits exactly this discrepancy: an excess of heterozygosity relative to the number of alleles observed.

Founder effects also shift frequencies permanently, not just reduce them. A variant that was rare in the source population can be common in the founders by chance, and then rides the new population's growth. This is the documented origin of the Finnish disease heritage, of the BRCA1 and Tay–Sachs founder alleles in Ashkenazi Jewish populations, and of Ellis–van Creveld syndrome among the Old Order Amish — and it is one of the two live explanations, alongside heterozygote advantage, for the cystic fibrosis frequency that §3 showed cannot produce.


Common misconceptions

What people thinkWhat's actually true
Evolution means selectionSelection is one of four forces. Most molecular evolution is acting on variants that do nothing (Ch 33)
Drift pushes toward loss, or toward 0.5E[p′] = p. Drift has zero mean and only variance. Alleles are lost often because 0 is absorbing and most alleles start near it, not because drift aims there
Selection eliminates harmful allelesAt q = 0.01, 99% of copies hide in heterozygotes. Halving the frequency of a fully recessive lethal takes 100 generations. Selection sees q², not q
A lethal disappears, so dominant disease alleles are vanishingly rareThey persist at μ/(hs) and are continually recreated. Achondroplasia is dominant, largely -reducing, and still ~1 in 25,000 — because ~80% of cases are new
Ne ≈ the number of breeding adultsIt is a harmonic mean over time, discounted by sex ratio and reproductive variance. It is routinely 10–100× below N, occasionally above it
Human Ne = 10,000 means only 10,000 humans existedIt is a summary statistic of drift across deep time, dominated by the smallest and most structured epochs. Census size today is irrelevant to it
A small proportion of migrants can't stop populations divergingFST depends on Nm, not m. One migrant per generation holds FST ≈ 0.2 at any population size
Selection maintains genetic variationUsually it removes it — both directional selection and reduce diversity. maintains it but is genuinely rare; unambiguous examples are few
A beneficial mutation will spreadA new mutation with s = 0.01 is lost ~98% of the time. Selection changes odds, it does not guarantee outcomes
explain allele frequenciesMutation alone moves frequencies with a half-life of ~10⁵ generations. It sets frequencies only in balance with selection or drift

Worked example: sickle cell, with and without malaria

The HBB HbS causes sickle cell disease in homozygotes and confers resistance to falciparum malaria in heterozygotes. Classic estimates from malaria-endemic West Africa (approximate, and historical — they vary by region and era):

w(AA) = 0.89     w(AS) = 1.00     w(SS) = 0.20
s₁ = 1 − 0.89 = 0.11   (cost of AA: malaria)
s₂ = 1 − 0.20 = 0.80   (cost of SS: sickle cell disease)

Step 1 — equilibrium. q̂ = s₁/(s₁ + s₂) = 0.11/0.91 = 0.1209.

Observed HbS frequencies in high-transmission regions run ~0.10–0.15. The model lands inside the observed range from two measurements.

Step 2 — verify it is stable. Use Δq = pq(s₁p − s₂q)/w̄ with w̄ = 1 − s₁p² − s₂q².

At q = 0.05, p = 0.95:

s₁p = 0.11 × 0.95   = 0.1045
s₂q = 0.80 × 0.05   = 0.0400        bracket = +0.0645
w̄   = 1 − 0.11(0.9025) − 0.80(0.0025) = 1 − 0.099275 − 0.002 = 0.898725
Δq  = 0.95 × 0.05 × 0.0645 / 0.898725 = 0.00306375/0.898725 = +0.00341

At q = 0.20, p = 0.80:

s₁p = 0.11 × 0.80   = 0.0880
s₂q = 0.80 × 0.20   = 0.1600        bracket = −0.0720
w̄   = 1 − 0.11(0.64) − 0.80(0.04)  = 1 − 0.0704 − 0.0320 = 0.8976
Δq  = 0.80 × 0.20 × (−0.0720)/0.8976 = −0.01152/0.8976 = −0.01284

Below q̂ it rises, above q̂ it falls. Stable, as derived.

Step 3 — the cost. At equilibrium, p̂ = 0.8791:

w̄ = 1 − 0.11(0.8791²) − 0.80(0.1209²)
  = 1 − 0.11(0.772854) − 0.80(0.014617)
  = 1 − 0.085014 − 0.011694 = 0.9033

Mean fitness is 0.903: a genetic load of 9.7%, paid every generation. Concretely, q̂² = 0.0146 of newborns are SS — 1 in 68 births with sickle cell disease. This is what "heterozygote advantage" costs. does not optimise anything; it parks the population at the top of a fitness hill whose summit is well below either homozygote's best case.

Step 4 — remove malaria. Set s₁ = 0. The equilibrium collapses to q̂ = 0, and Δq becomes −s₂pq²/w̄ — pure selection against a . Integrating dq/dt ≈ −s₂q²(1 − q), with antiderivative −1/q + ln[q/(1−q)]:

from q = 0.1209 → 0.0605:
  [−1/0.1209 + ln(0.1209/0.8791)] = −8.2713 − 1.9839 = −10.2552
  [−1/0.0605 + ln(0.0605/0.9395)] = −16.5289 − 2.7420 = −19.2709
  Δ = −9.016 = −s₂·t     ⇒     t = 9.016/0.80 ≈ 11 generations

from q = 0.0100 → 0.0050 (same s₂, same process):
  [−1/0.01 + ln(0.01/0.99)]   = −100.000 − 4.595 = −104.595
  [−1/0.005 + ln(0.005/0.995)]= −200.000 − 5.293 = −205.293
  Δ = −100.70 = −s₂·t    ⇒     t = 100.70/0.80 ≈ 126 generations

The first halving takes 11 generations; a halving at q = 0.01 takes 126 — with identical selection. That factor of eleven is the q² term making itself felt, and it is why populations that left malarial regions generations ago still carry HbS at appreciable frequency. Selection against recessives is fast where the allele is common and grinds to a halt exactly where you would want it to work.

Connections