Contents
S2 — The distributions you'll actually meet Statistics track

S2 — The distributions you'll actually meet

Read before: Ch 09 · Time: ~45 min

A couple has four children and three of them have cystic fibrosis. Is that a red about the diagnosis, or an ordinary run of luck? You sequence a bacterial genome to 6× average depth and 0.85% of it has no reads at all — is that expected, or did the library fail? You measure a gene in three biological replicates of the same yeast strain and get 12,528, 8,756 and 9,724 counts — is that gene noisy, or is your pipeline broken?

None of these is answerable without a model of how the numbers were generated. That is what a probability distribution is: not a curve to memorise, but a claim about a generating process. Choose the process correctly and the arithmetic is trivial. Choose it wrongly and you will get an answer that is confident, precise, and off by ninety-five orders of magnitude — as §5 demonstrates on real data.

There are about six generating processes that produce nearly everything you will meet in genomics. This chapter is a field guide to them.

All code in this chapter runs from the repository root — the directory holding README.md and .venv — with source .venv/bin/activate, and addresses data as labs/data/…. The three real datasets it uses are built by lab-02, lab-06 and lab-07, and everything it imports is in lab-00's base install.

What you'll be able to do

The core idea

Every distribution in this chapter is the answer to a counting story.

The storyThe distribution
Flip a biased coin onceBernoulli
Flip it n times, count headsBinomial
Very many flips, each very unlikely, expected count fixedPoisson
Add up many small independent contributionsNormal
Add up squared standardised normalsChi-square
Poisson counting, but the rate itself varies between samples
Wait for the first successGeometric / Exponential
A probability that is itself uncertainBeta

Learn the stories, not the formulas. When you meet a new measurement, ask which story generated it. If none of them did, the distribution you are about to assume is wrong, and that is worth knowing before you compute a p-value from it.

Two numbers summarise any of them: the mean (where it sits) and the variance (how far it spreads). The single most useful diagnostic in genomics is the ratio of the two — the variance-to-mean ratio, or VMR. Poisson has VMR exactly 1. Nearly all real count data has VMR above 1, and §5 is about what to do when it does.

All code below assumes:

import numpy as np, pandas as pd
from scipy import stats

1. Bernoulli and binomial: n independent tries, count the successes

A Bernoulli trial has two outcomes with probability p and 1 − p. Mean p, variance p(1 − p) — maximised at p = 0.5, and zero at p = 0 or 1, which is the formal statement that a certain event carries no information.

Do n of them independently, with the same p, and count successes: that is Binomial(n, p).

P(X = k) = C(n,k) · p^k · (1−p)^(n−k)        mean = np      variance = np(1−p)

The binomial coefficient C(n,k) counts the orderings, and forgetting it is the single most common error in arithmetic. Both assumptions are load-bearing: independent and identical p. Nearly every failure of the binomial in genomics is one of those two breaking.

1.1 Family composition

Two carrier parents, Aa × Aa. Each child independently has probability ¼ of being aa (Ch 10). Four children:

for k in range(5):
    print(f"exactly {k} affected: {stats.binom.pmf(k, 4, 0.25):.4f}")
print("at least one:", round(stats.binom.sf(0, 4, 0.25), 4))
exactly 0 affected: 0.3164
exactly 1 affected: 0.4219
exactly 2 affected: 0.2109
exactly 3 affected: 0.0469
exactly 4 affected: 0.0039
at least one: 0.6836

Those are 81/256, 108/256, 54/256, 12/256, 1/256 — the fourth row of the expansion of (¾ + ¼)⁴. Three affected out of four has probability 0.047: uncommon, not remarkable. You would see it in about one in twenty-one such families.

Note the difference between 0.10547 — first three unaffected, then one affected, in that order — and 0.42188, exactly one affected in any order. The ratio is C(4,1) = 4. Pedigree questions almost always want the second, and the phrasing rarely says so.

Mean 4 × 0.25 = 1 affected child, variance 4 × 0.25 × 0.75 = 0.75, sd 0.87. A family of four carrying a ¼ risk expects one affected child and routinely has none — an intuition worth having before counselling anyone about pedigrees.

1.2 A genotype is a binomial draw

Here is the reframing that makes Hardy–Weinberg obvious. Under random mating, an individual's is two independent draws from the gamete pool. Count ALT : the genotype dosage 0/1/2 is Binomial(2, p), and p², 2pq, q² is nothing but that binomial's pmf.

Real data — the 503 European-ancestry samples in the 1000 Genomes chr22 subset:

raw  = pd.read_csv("labs/data/chr22_qc.raw", sep="\t")   # plink2 --pfile labs/data/chr22_qc --export A --out labs/data/chr22_qc
raw  = raw[raw.IID.isin(open("labs/data/EUR.ids").read().split())]   # 2504 -> 503 EUR
G    = raw.iloc[:, 6:].to_numpy(float)         # 503 samples x 3564 SNPs, dosage 0/1/2
p    = G.mean(0) / 2
i    = np.argmin(np.abs(p - 0.5))              # the SNP closest to p = 0.5
g, n = G[:, i], G.shape[0]

obs = np.array([(g == k).sum() for k in (0, 1, 2)])
exp = n * stats.binom.pmf([0, 1, 2], 2, p[i])
print("observed:", obs, " Binomial(2,p):", exp.round(1))
SNP chr22:20682316 G>A  p_ALT = 0.5000
observed: [123 257 123]  Binomial(2,p): [125.8 251.5 125.8]

Two independent Bernoulli draws per person, 503 people, and the genotype counts land within three individuals of the prediction. Everything Hardy–Weinberg says is contained in that one line of stats.binom.pmf.

1.3 ALT reads at a heterozygous site

At a het site, each read independently samples one of the two , so the ALT read count is Binomial(depth, 0.5). This is the model every variant caller uses (Ch 46). It has sharp consequences at low depth:

for d in (5, 10, 20, 30):
    print(f"{d:2d}x  P(<=2 ALT reads) = {stats.binom.cdf(2, d, 0.5):.4f}"
          f"   P(0 ALT) = {stats.binom.pmf(0, d, 0.5):.5f}")
 5x  P(<=2 ALT reads) = 0.5000   P(0 ALT) = 0.03125
10x  P(<=2 ALT reads) = 0.0547   P(0 ALT) = 0.00098
20x  P(<=2 ALT reads) = 0.0002   P(0 ALT) = 0.00000
30x  P(<=2 ALT reads) = 0.0000   P(0 ALT) = 0.00000

At 5× depth, 3.1% of true heterozygotes have zero ALT reads and are called homozygous reference with no ambiguity whatsoever in the data. That number, not the sequencer's error rate, is why low- genotyping needs . And to test a candidate site for allelic imbalance you want the exact binomial test, not a normal approximation:

r = stats.binomtest(9, 30, 0.5)          # 9 ALT reads out of 30
print(round(r.pvalue, 4), np.round(r.proportion_ci(), 3))
0.0428 [0.147 0.494]

Where the binomial breaks in practice. Reads carrying the ALT allele mismatch the reference and align slightly worse, so p is a little under 0.5 — . Set p = 0.45 and P(≤2 ALT reads at 10×) rises from 0.055 to 0.0996. A 5% shift in one parameter nearly doubles the dropout rate. The failure is not in the binomial; it is in the claim that p = 0.5.

2. Poisson: rare events in a large window

The derivation is the specification

Take Binomial(n, p). Let n → ∞ and p → 0 with the product λ = np held fixed. The binomial coefficient and the powers conspire, and the limit is

P(X = k) = e^(−λ) · λ^k / k!          mean = λ      variance = λ

Watch it happen, holding λ = 3:

for k in range(4):
    row = [stats.binom.pmf(k, n, 3.0/n) for n in (10, 100, 1000, 100000)]
    print(k, " ".join(f"{v:.6f}" for v in row), f"| {stats.poisson.pmf(k, 3.0):.6f}")
k    n=10      n=100     n=1000    n=100000  | Poisson(3)
0  0.028248  0.047553  0.049563  0.049785   | 0.049787
1  0.121061  0.147070  0.149137  0.149359   | 0.149361
2  0.233474  0.225153  0.224154  0.224043   | 0.224042
3  0.266828  0.227474  0.224379  0.224045   | 0.224042

By n = 1,000 the two agree to three decimals. Poisson is the binomial when there are enormous numbers of opportunities and each is individually almost impossible — which is precisely the structure of a genome. That is why it appears everywhere:

Genomic quantityn opportunitiesp per opportunityλ
Reads starting at a given baseevery read in the library1/genome sizedepth / read length
Reads covering a given baseevery readread length / genomedepth
per genome6.2 × 10⁹ bases~1.2 × 10⁻⁸~70
Crossovers on a every base pair~10⁻⁸ per bpmap length in Morgans
Reads assigned to a geneevery readgene's expression shareexpected count

Note the one parameter. Poisson has no separate variance knob: variance = mean, VMR = 1. That rigidity is what makes it so easy to falsify, and §5 is about what happens when it is falsified.

De novo mutations, straight from the pinned rate (Ch 16): λ = 2 × 1.1 × 10⁻⁸ × 3.1 × 10⁹ = 68 per genome per generation, with sd √68 = 8.3. Much of the variability between siblings is a sampling property of the Poisson process rather than any difference in their parents' mutational quality. But not all of it: paternal age adds a genuine second variance component of ~1.3–1.5 extra per year of the father's age, so siblings born five years apart differ by ~7 in expected count — comparable to the Poisson sd of 8.3, and a systematic shift rather than a draw.

Sequencing coverage: the real thing

Every base of a genome is covered by a Poisson number of reads with λ = depth. This is the Lander–Waterman model, and the piece everyone uses is P(0 reads) = e^(−λ):

genome_bp = 4_629_812                             # E. coli REL606 (not `G` — that is the
                                                  # genotype matrix from §1, still in scope)
for d in (1, 5, 10, 30):
    print(f"{d:3d}x  P(uncovered) = {np.exp(-d):.3e}   bases missed = {genome_bp*np.exp(-d):12,.0f}")
  1x  P(uncovered) = 3.679e-01   bases missed =    1,703,213
  5x  P(uncovered) = 6.738e-03   bases missed =       31,195
 10x  P(uncovered) = 4.540e-05   bases missed =          210
 30x  P(uncovered) = 9.358e-14   bases missed =            0

improves the gaps exponentially: each extra 1× multiplies the uncovered fraction by 1/e. This is the whole argument for why assemblies stopped being coverage-limited long ago (Ch 43) and why 30× became the human standard.

Now check it against a real alignment — the E. coli from lab 02, 199,861 mapped 150 bp reads:

import pysam
bam = pysam.AlignmentFile("labs/data/aln606.bam")
cov = np.sum(bam.count_coverage("NC_012967.1", 0, 4_629_812, quality_threshold=0), axis=0)
print(f"mean {cov.mean():.3f}  variance {cov.var():.3f}  VMR {cov.var()/cov.mean():.3f}")
print(f"observed P(depth=0) {(cov==0).mean():.5f}   Poisson {stats.poisson.pmf(0, cov.mean()):.5f}")
mean 6.047  variance 9.031  VMR 1.493
observed P(depth=0) 0.00852   Poisson 0.00236

Overlay the observed histogram on the Poisson pmf (# = observed, | = Poisson prediction):

depth  observed  Poisson
    0    0.0085   0.0024   #|
    1    0.0280   0.0143   ####|##
    2    0.0622   0.0432   ###########|####
    3    0.0986   0.0871   ######################|##
    4    0.1286   0.1317   ################################ |
    5    0.1406   0.1593   ###################################
    6    0.1350   0.1606   ##################################
    7    0.1166   0.1387   #############################      |
    8    0.0925   0.1049   #######################   |
    9    0.0678   0.0705   ################# |
   10    0.0468   0.0426   ###########|
   11    0.0301   0.0234   ######|#
   12    0.0187   0.0118   ###|#
   13    0.0111   0.0055   #|#
   14    0.0066   0.0024   #|

The shape is right and the spread is wrong. Both tails are too heavy, the peak too low; VMR is 1.49, not 1. Real coverage is overdispersed because λ is not constant across the genome — GC content, mappability, repeats and library chemistry all modulate it. The consequence is not cosmetic: Poisson says 0.24% of the genome is uncovered, and 0.85% actually is — a 3.6-fold underestimate of the gaps. §5 gives the fix and the Worked example turns it into a sequencing budget.

When a derivation's assumptions fail, the distribution fails in a predictable direction. The Poisson limit needs constant p across opportunities. Vary p and you get overdispersion (VMR > 1). Make events repel each other and you get underdispersion (VMR < 1) — which is exactly what crossovers do. and the obligate crossover (Ch 09) mean crossovers are more evenly spaced than Poisson: a Poisson with λ = 1 would have no crossover 37% of the time, and real bivalents essentially always have one. That is the whole difference between the Haldane map function (Poisson, no interference: r = 0.316 at 0.5 M) and Kosambi (interference: r = 0.381) in Ch 14.

3. Normal: sums of many small effects

The central limit theorem says: if X₁ … Xₙ are independent, identically distributed, and have finite variance, then their standardised sum converges in distribution to Normal(0, 1) as n → ∞. Mean μ, variance σ², and — the fact that gets used — the sum of independent normals is normal, so the sample mean of n draws has variance σ²/n.

Be precise about what that is and is not a promise:

The CLT promisesThe CLT does not promise
The sum or mean approaches normalThat your data are normal — heights are, individual read counts are not
Convergence in the body of the distributionAccuracy in the far tails, where it can be badly wrong at any n
Convergence for any finite-variance parentA useful n; heavily skewed parents need far more
Nothing at all if variance is infinite— (heavy-tailed parents never converge to normal)

The tail caveat is not academic. Standardise the mean of 1,000 draws from a Poisson(0.5) — badly skewed, but n = 1,000. No simulation is needed: the sum of 1,000 independent Poisson(0.5) draws is exactly Poisson(500), so the true tail probabilities are available in closed form.

lam, n = 0.5, 1000
mu, sd = n*lam, np.sqrt(n*lam)                 # sum ~ Poisson(500), sd = sqrt(500) = 22.361
for z in (1.96, 4.0, -4.0):
    thr   = np.floor(mu + z*sd)
    exact = stats.poisson.sf(thr, mu) if z > 0 else stats.poisson.cdf(thr, mu)
    nrm   = stats.norm.sf(abs(z))
    print(f"P(Z {'>' if z > 0 else '<'} {z:5.2f})  exact {exact:.6f}   "
          f"normal {nrm:.6f}   ratio {exact/nrm:.2f}")
P(Z >  1.96)  exact 0.027084   normal 0.024998   ratio 1.08
P(Z >  4.00)  exact 0.000048   normal 0.000032   ratio 1.53
P(Z < -4.00)  exact 0.000019   normal 0.000032   ratio 0.59

The 5% cut-off is nearly right; the 4σ right tail is 1.5× too heavy and the left tail 1.7× too light. Genome-wide testing lives at 5σ and beyond (Ch 51), which is exactly where the approximation you were taught to trust stops being trustworthy — and why permutation and exact tests survive in genomics (S4).

Why quantitative traits are normal

Ch 30 opens with "quantitative genetics is Mendelian genetics plus the central limit theorem". Here it is literally. Take the real chr22 , give each SNP a random additive effect, and sum:

X = G[:, np.minimum(p, 1-p) > 0.05]                 # 503 people x 1784 common SNPs
rng = np.random.default_rng(0)
for L in (1, 5, 50, X.shape[1]):
    idx  = rng.choice(X.shape[1], L, replace=False)
    y    = X[:, idx] @ rng.normal(0, 1, L)
    y    = (y - y.mean()) / y.std()
    print(f"L={L:5d}  skew={stats.skew(y):+.3f}  Shapiro p={stats.shapiro(y).pvalue:.3g}")
L=    1  skew=+0.406  Shapiro p=1.03e-25
L=    5  skew=-0.175  Shapiro p=3.99e-05
L=   50  skew=-0.163  Shapiro p=0.232
L= 1784  skew=-0.011  Shapiro p=0.365

One locus gives three discrete genotype classes — emphatically not normal. By 50 loci the normality test has nothing left to reject. Nobody had to assume a normal trait: it fell out of adding up small independent contributions, which is what a polygenic architecture is. The same argument, run on environmental contributions too, is why the residual is also treated as normal — and it is the entire justification for the variance decomposition that rests on.

Note what this does not justify: multiplicative traits. If effects combine by multiplication rather than addition, the logarithm is normal, not the trait. That is why expression levels, odds ratios and concentrations get log-transformed as a reflex.

4. Chi-square: adding up squared standardised normals

If Z₁ … Zₖ are independent standard normals, then ΣZᵢ² has a chi-square distribution with k degrees of freedom. Mean k, variance 2k. That is the entire definition, and every in genetics is an instance of it.

Why does Σ(O − E)²/E follow it? Not quite for the reason usually given. The cell counts are multinomial, not Poisson: with n fixed, Var(Oᵢ) = n pᵢ(1 − pᵢ) = Eᵢ(1 − Eᵢ/n). So each term (Oᵢ − Eᵢ)/√Eᵢ is asymptotically normal but has variance 1 − Eᵢ/n, a little under 1 — for a heterozygote cell at p = 0.5, where E/n = 0.5, its variance is only ½. Nor are the terms independent: the counts must sum to n, so they are negatively correlated, Cov(Oᵢ, Oⱼ) = −n pᵢpⱼ. The two departures cancel exactly. The shortfall in each term's variance is precisely offset by the negative covariance between cells, and the quadratic form comes out chi-square anyway — with cells − 1 df rather than one per cell. That lost degree of freedom is the fixed total, not an accounting convention.

Degrees of freedom = how many of those numbers were free to vary. Start with the number of cells, subtract one for every quantity you forced the expected counts to match:

TestCellsConstraintsdf
Mendelian 3:1, n fixed2total1
Mendelian 9:3:3:1, n fixed4total3
Hardy–Weinberg3total, and p estimated from these very counts1
HWE with p known independently3total only2

The Hardy–Weinberg row is the one people get wrong (Ch 26 §5). Three classes look like df = 2, but p̂ was computed from the same genotypes, so once you fix the total and the only one cell is still free. Using df = 2 makes the test conservative — you will miss real genotyping failures.

Settle it empirically. Compute the HWE statistic at all 1,784 common chr22 SNPs in the EUR samples and look at where the distribution sits:

def hwe_chi2(col, pp):
    o = np.array([(col == k).sum() for k in (0, 1, 2)])
    e = o.sum() * np.array([(1-pp)**2, 2*pp*(1-pp), pp**2])
    return ((o - e)**2 / e).sum()

s = np.array([hwe_chi2(G[:, j], p[j]) for j in np.where(np.minimum(p,1-p) > 0.05)[0]])
print("SNPs:", s.size, " mean statistic:", round(s.mean(), 3))
print("frac p<0.05 using df=1:", round((s > stats.chi2.ppf(.95, 1)).mean(), 4),
      "  using df=2:",           round((s > stats.chi2.ppf(.95, 2)).mean(), 4))
SNPs: 1784  mean statistic: 1.077
frac p<0.05 using df=1: 0.0807   using df=2: 0.0291

A chi-square with 1 df has mean 1; with 2 df, mean 2. The observed mean is 1.077. The question is not close. And the calibration confirms it: df = 1 gives 8.1% of SNPs below p = 0.05 — a little above nominal, as expected from residual structure and genotyping error inside a super-population — while df = 2 gives 2.9%, visibly conservative.

Related distributions you will meet, all built from the same normals: Student's t (a normal divided by an independently estimated sd — heavier tails than normal, converging to it as df grows), and the F (a ratio of two chi-squares over their df, the workhorse of ANOVA and variance-component estimation in S5).

5. Negative binomial: when the rate itself varies

This is the distribution that forced on genomics, and the reason is biological, not mathematical.

Sequencing a library is Poisson sampling: given a gene's true fraction of the transcript pool, the read count is Poisson. So counts across technical replicates — the same library, sequenced twice — really are Poisson. But biological replicates are different cultures, different mice, different patients. Each has its own true expression level. The rate λ is not a constant, it is a random variable.

Let λ vary between replicates with mean μ and variance α μ² (i.e. a constant coefficient of variation √α), and count Poisson-with-rate-λ. The law of total variance gives

Var(X) = E[Var(X|λ)] + Var(E[X|λ]) = μ + αμ²
          └ Poisson noise ┘   └ biological variation ┘

Make λ specifically gamma-distributed and the marginal count distribution is exactly the , with

mean = μ        variance = μ + α μ²        α = dispersion

That second term is the whole story. Poisson noise scales with μ; biological variation scales with μ². At low counts Poisson noise dominates and the negative binomial looks Poisson; at high counts biological variation dominates and it does not. So the fingerprint of overdispersion is not a constant inflation — it is a VMR that grows with expression.

The fingerprint, in real yeast RNA-seq

Three wild-type biological replicates from lab 06, size-factor normalised the way DESeq2 does it, binned by mean expression:

cnt  = pd.read_csv("labs/data/yeast_counts.tsv", sep="\t", index_col=0)
wt   = cnt[["wt_rep1", "wt_rep2", "wt_rep3"]]
lg   = np.log(wt.replace(0, np.nan))
sf   = np.exp(lg.sub(lg.mean(axis=1), axis=0).median(skipna=True))   # median-of-ratios
norm = wt / sf
N    = norm[norm.mean(axis=1) > 5]
m, v = N.mean(axis=1).to_numpy(), N.var(axis=1, ddof=1).to_numpy()
mean-count range   genes   med. mean   med. variance    VMR   implied alpha
      5 –    20     2357        11.8            10.0    0.85       -0.0127
     20 –    50     1589        29.5            33.2    1.12        0.0042
     50 –   150      825        75.9           151.7    2.00        0.0131
    150 –   500      277       232.9          1100.9    4.73        0.0160
    500 –  2000       74       726.0         11812.1   16.27        0.0210
   2000 –  ∞          11      4266.0        890284.4  208.69        0.0487

Read the VMR column downward: 0.85 → 209. At counts around 10 the data are indistinguishable from Poisson — the bottom bin's negative α is just estimation noise, since with three replicates each variance carries 2 degrees of freedom. At counts in the thousands the variance is two hundred times the mean. From 50 counts upward the α settles between 0.013 and 0.05 — a biological coefficient of variation of 11–22%, which for replicate yeast cultures is entirely reasonable. α is roughly constant while VMR is not, which is the sense in which the negative binomial is the right two-parameter family.

Now the consequence, on one gene. YGR192C (TDH3, glyceraldehyde-3-phosphate dehydrogenase) is the most highly expressed gene in the set. Its three wild-type replicates:

row   = norm.loc["YGR192C"]
mu, s2 = row.mean(), row.var(ddof=1)
alpha  = (s2 - mu) / mu**2
n_, p_ = 1/alpha, (1/alpha) / (1/alpha + mu)
print(np.round(row.values).astype(int), f"mean {mu:.0f} variance {s2:.0f}")
print("Poisson P(X >= 12528) =", f"{stats.poisson.sf(12527, mu):.3g}")
print("NegBin  P(X >= 12528) =", f"{stats.nbinom.sf(12527, n_, p_):.3g}")
[12528  8756  9724]   mean 10336   variance 3836708
Poisson P(X >= 12528) = 7.06e-97
NegBin  P(X >= 12528) = 0.133

Three replicates of the same yeast strain, and Poisson calls the spread between them a 7 × 10⁻⁹⁷ event. The negative binomial calls it p = 0.13 — ordinary. This is the single most important number in this chapter. Fit Poisson to biological replicates and every highly expressed gene becomes overwhelmingly significant, because you have modelled away the only source of variation that matters. The p-values are not slightly optimistic; they are meaningless.

This is why DESeq2 and edgeR exist, and why almost all of their sophistication goes into estimating α, not μ. With three replicates you cannot estimate a per-gene dispersion at all, so both packages shrink each gene's α toward a fitted mean–dispersion trend — borrowing strength across genes, an idea S7 generalises. In scipy, stats.nbinom(n, p) uses the "number of failures before n successes" parameterisation, so convert: n = 1/α, p = n/(n + μ). Verify with .mean() and .var() every time; this conversion is a reliable source of silent bugs.

The same logic applies wherever counts come from heterogeneous units: single-cell counts (Ch 48), and peak counts (Ch 49), and — as §2 showed — sequencing depth along a genome.

6. Three more, briefly

Geometric — how many trials until the first success. P(X = k) = (1−p)^(k−1)p, mean 1/p, variance (1−p)/p². Discrete waiting time; memoryless (having waited does not change what comes next). Number of clones you must before finding a positive.

Exponential — the continuous version. Density λe^(−λx), mean 1/λ, sd also 1/λ. It is the gap between consecutive events of a Poisson process, so wherever §2 gave you a count, this gives you the spacing: distance to the next crossover along a , distance to the next , time to the next event (Ch 27). The signature test is sd = mean.

Real check, on the 29,700 variant positions in a 1 Mb window of 1000 Genomes chr22:

mean gap 33.7 bp   sd 76.3 bp   (exponential predicts sd = mean)
P(gap > 100) observed 0.0508   exponential 0.0513
P(gap > 500) observed 0.0013   exponential 0.0000

Right in the body, badly wrong in the tail — the same lesson as everywhere else in this chapter. Variants are not thrown down at a uniform rate: conserved exons and mutational cold spots create long gaps that a single-rate exponential can never produce, which is why long gaps are interesting (Ch 33).

Beta — a distribution over a probability. Support [0, 1], parameters a, b, mean a/(a+b). Think of a − 1 prior successes and b − 1 prior failures. Beta(1,1) is uniform; Beta(20,20) is tight around 0.5; Beta(1,20) piles up near zero. It is the conjugate prior for the binomial: observe k ALT out of n and the posterior is simply Beta(a + k, b + n − k) (S6).

post = stats.beta(1 + 3, 1 + 1006 - 3)         # 3 ALT copies among 1006 EUR chromosomes
print(f"mean {post.mean():.5f}  95% interval {post.ppf(.025):.5f}–{post.ppf(.975):.5f}")
mean 0.00397  95% interval 0.00108–0.00868

The point estimate 3/1006 = 0.00298 is a single number pretending to certainty; the interval spans an eightfold range. That is the honest state of knowledge about a rare allele in 503 people, and mixing Beta with Binomial gives the beta-binomial — the overdispersed binomial, used for allele-specific expression and for pooled sequencing, on exactly the logic of §5.

Field guide

DistributionGenerating storyParametersMeanVariancescipy
Bernoullione trialppp(1−p)stats.bernoulli(p)
Binomialn independent trials, same pn, pnpnp(1−p)stats.binom(n, p)
Poissonn → ∞, p → 0, np fixedλλλstats.poisson(mu)
Normalsum of many small effectsμ, σμσ²stats.norm(loc, scale)
Chi-squaresum of k squared std normalskk2kstats.chi2(df)
Poisson with a gamma-varying rateμ, αμμ + αμ²stats.nbinom(1/a, ...)
Geometrictrials until first successp1/p(1−p)/p²stats.geom(p)
Exponentialwaiting time in a Poisson processλ1/λ1/λ²stats.expon(scale=1/lam)
Betauncertainty about a probabilitya, ba/(a+b)—stats.beta(a, b)

Every frozen distribution above exposes the same interface — .pmf/.pdf, .cdf, .sf (the upper tail, and always use it instead of 1 - cdf for small p-values), .ppf (quantiles), .rvs (simulate), .mean(), .var(). Learning one teaches you all of them. To check any claim in this chapter, simulate and overlay:

import matplotlib.pyplot as plt
d = stats.nbinom(20, 20/120)                       # mu = 100, alpha = 0.05
x = d.rvs(200_000, random_state=0)
grid = np.arange(0, 300)
plt.hist(x, bins=np.arange(-.5, 300), density=True, color="0.8")
plt.plot(grid, d.pmf(grid), color="C3")
print(f"theory mean {d.mean():.1f} var {d.var():.1f} | sample {x.mean():.1f} {x.var():.1f}")
theory mean 100.0 var 600.0 | sample 100.0 var 599.4

Common misconceptions

What people thinkWhat's actually true
Data are normal if you collect enough of itThe CLT is about the sum or mean, not the data. A million read counts are still not normal; their mean is
Poisson is a good model for read countsIt is exactly right for technical replicates and badly wrong for biological ones. The whole reason DESeq2 exists
Overdispersion inflates the variance by a constantIt scales as μ², so it is invisible at low counts and at high ones. That growing VMR is the diagnostic
Three classes means df = 2df = 1 when p was estimated from the same counts. Every estimated parameter costs a degree of freedom
The normal approximation is fine at n = 1,000Fine at p = 0.05; the 4σ right tail is 1.5× too heavy in the exact calculation above. Genome-wide thresholds live in the region where it fails
A binomial needs only "two outcomes"It needs independent trials with identical p. breaks the second; linked reads and break the first
The exponential and geometric are different ideasSame story, continuous vs discrete time. Both memoryless, both the waiting time in a Poisson process
Crossovers are Poisson along a They are underdispersed — and the obligate crossover space them out. That is the Haldane/Kosambi difference
A rare seen 3 times in 1,006 has frequency 0.003That is a point estimate with a 95% interval of 0.001–0.009. The Beta posterior says so out loud
Variance = mean is a property of countsIt is a property of Poisson. Real counts almost always have variance > mean

Worked example: how deep should I sequence?

The question. You are sequencing a 4.6 Mb bacterial genome and need at least 5× at every base to call variants confidently. What average depth do you buy?

Step 1 — name the process. Reads land at effectively random positions; each base has a huge number of chances to be covered, each individually tiny. That is the Poisson limit of §2, with λ = average depth.

Step 2 — the textbook answer.

for d in (10, 15, 20, 30):
    print(f"{d:2d}x  Poisson P(depth < 5) = {stats.poisson.cdf(4, d):.6f}")
10x  Poisson P(depth < 5) = 0.029253
15x  Poisson P(depth < 5) = 0.000857
20x  Poisson P(depth < 5) = 0.000017

15× leaves 0.09% of the genome under-covered; 20× leaves 0.002%. Buy 15×.

Step 3 — check the model against real data before spending money. The E. coli alignment from §2 has mean depth 6.05 and variance 9.03. Poisson requires them to be equal. Compare the observed tail with the prediction:

threshold   observed frac    Poisson(6.05)     ratio
depth < 1        0.00852         0.00236        3.60
depth < 2        0.03654         0.01666        2.19
depth < 5        0.32592         0.27878        1.17

The model is systematically optimistic, and it gets worse the further into the tail you go — which is where the decision is being made. Poisson is not merely imprecise here; it is biased in the direction that costs you.

Step 4 — refit with the observed overdispersion. Fit a to the real : α = (variance − mean)/mean² = (9.031 − 6.047)/6.047² = 0.0816. Redo the calculation:

alpha = 0.08159
for d in (15, 20, 30, 50):
    n_ = 1/alpha; p_ = n_/(n_ + d)
    print(f"{d:2d}x  Poisson {stats.poisson.cdf(4, d):.6f}   NegBin {stats.nbinom.cdf(4, n_, p_):.6f}")
15x  Poisson 0.000857   NegBin 0.012905
20x  Poisson 0.000017   NegBin 0.002470
30x  Poisson 0.000000   NegBin 0.000145
50x  Poisson 0.000000   NegBin 0.000002

Step 5 — answer, with the reason. At 15× the honest estimate of under-covered genome is 1.3%, not 0.09% — fifteen times worse than the Poisson answer, and at 4.6 Mb that is roughly 60,000 bases you cannot call. The negative binomial does not cross 0.1% until 23×, where Poisson claimed 15× (0.086%) already cleared that bar. Buy 30× and keep the margin.

Step 6 — what actually happened. The 30× convention in genomics is not superstition and it is not the Lander–Waterman calculation, which would have settled on half that. It is the Lander–Waterman calculation corrected for the fact that λ is not constant along a genome. Every step of the reasoning is a distribution choice: Poisson because reads are rare events in a large window, negative binomial because the rate varies, and the variance-to-mean ratio of the real data as the arbiter between them. That is the whole method of this chapter.

Where this is used

Next: S3 — Sampling, estimation and error turns these distributions into standard errors and confidence intervals; S4 turns them into tests.