Every probability formula can be asked backward. A test report gives P(positive | sick), but you want the base rate. A game's average is known, but one payout is missing. A system must be 99% reliable, and you need the number of checks. Each is a transposition with the moves you know, plus one new condition: the answer must be a chance, between 0 and 1.
What you'll be able to do
Turn the complement, "and", "or" and Bayes rules around for any chance, the prior and the likelihood included
Free a chance that appears twice in the total-probability rule by expanding, collecting and dividing
Rearrange expected value, variance, binomial and z-score formulas, and choose the root that fits
Solve pⁿ ≥ target for n with logs, flipping the sign because ln p is negative
Check that every answer lies between 0 and 1, with exact fractions and a seeded simulation
Quick check
Have a guess before you start. 30% of the trail club's hikers carry a map, and 27% carry a map and finish. Of the map carriers, what share finish?
8.1%
27%
90%
57%
Show the answer
90%
P(map and finish) = P(map) × P(finish | map). Undo the multiplication by dividing: P(finish | map) = 0.27 ÷ 0.3 = 0.9.
The math: run each rule backward
Purpose. Make any chance in a probability formula the subject, and check the answer can be a chance.
Each rule from the probability module turns around with the four moves of lesson 30:
Rule
Turned around
P(not A) = 1 − P(A)
P(A) = 1 − P(not A)
P(A and B) = P(A) × P(B given A)
P(B given A) = P(A and B) ÷ P(A)
P(A or B) = P(A) + P(B) − P(A and B)
P(A and B) = P(A) + P(B) − P(A or B)
P(A given B) = P(B given A) × P(A) ÷ P(B)
P(A) = P(A given B) × P(B) ÷ P(B given A)
P(B | A) is read "the chance of B given A", as in lesson 90. In Bayes' rule, P(A) is the prior: the chance before any evidence. P(B | A) is the likelihood: the chance of the evidence if A is true. Either can be the subject, so P(B | A) = P(A | B) × P(B) ÷ P(A) too.
Conditions. Divide only by a chance above 0. And every answer must land from 0 to 1. Say P(A) = 0.5, P(B) = 0.4 and P(A or B) = 0.95. The overlap comes out as −0.05, so these three numbers can't all be true. That's a check on the data, not a rounding slip.
Quick check
From an earlier lesson: P(A) = 0.5, P(B) = 0.4 and P(A and B) = 0.2. What is P(A or B)?
0.9
0.2
1.1
0.7
Show the answer
0.7
Add the two, then take away the overlap so it isn't counted twice: 0.5 + 0.4 − 0.2 = 0.7.
A chance that appears twice. In the trail club, 90% of hikers with a map finish, and 50% without one do. Overall, 62% finish. What share carry a map? Call it a. Finishers come by two paths, with a map or without:
0.62=0.9a+0.5(1−a)
The letter a appears twice, so use the move from lesson 32: expand, collect, divide. That gives 0.62 = 0.5 + 0.9a − 0.5a = 0.5 + 0.4a. So 0.4a = 0.12, and a = 0.3. Collected, the rule is a two-step machine: multiply by 0.4, which is 0.9 − 0.5, then add 0.5. Watch it built and peeled.
Try it now
Try a = 0 and a = 1: the finish rate runs from 0.5 to 0.9. A rate outside that range has no map share that fits.
Symbols and formula. The total-probability rule, solved for the prior:
In plain words: how far the overall rate sits above the "not A" rate, as a share of the gap between the two rates.
Conditions. The two rates must differ. If they're equal, B happens at the same rate either way, so it says nothing about A, and the formula divides by 0. And P(B) must lie between the two rates, or the answer leaves 0 to 1.
Named mistakes.
Dividing by the wrong chance: P(A and B) ÷ P(B) is P(A | B), not P(B | A).
Writing 0.5a for the second path. Hikers without a map are 1 − a of the club, not a.
Keeping an answer of 1.3 or −0.2 instead of rechecking the data.
Quick check
A spam filter flags posts with a certain word. P(spam | word) = 0.8, P(word) = 0.05 and P(spam) = 0.2. Solve Bayes' rule for the likelihood P(word | spam).
A missing value. E[X] = Σxp is a sum of products. Tent rentals of 10, 20 or x, with chances 0.2, 0.5 and 0.3, have E[X] = 21. So 2 + 10 + 0.3x = 21, giving 0.3x = 9 and x = 30.
A missing chance. All the chances add to 1, so write the last one as 1 minus the others. Rentals of 10, 20 or 30 with chances 0.2, q and 0.8 − q, and E[X] = 25: 2 + 20q + 30(0.8 − q) = 25. Collect q: 26 − 10q = 25, so q = 0.1. Check that no chance goes below 0: q must lie from 0 to 0.8.
The second moment. E[X²], the mean of the squares, comes from Var(X) = E[X²] − μ² turned around: E[X²] = Var(X) + μ². For the first rentals, 49 + 21² = 490.
A binomial counts the successes in n independent tries, each with chance p. The mean μ = np gives n = μ/p and p = μ/n (p above 0, and n must come out whole). The variance np(1 − p) = v is a quadratic in p, np² − np + v = 0. By lesson 34, p = (1 ± √(1 − 4v/n))/2. Both roots lie from 0 to 1, and they add to 1: p and 1 − p give the same spread. The variance can't choose between them, but the mean can. A real root needs v ≤ n/4.
Quick check
A model answers 100 questions, each right with chance p. Its number right has variance 21, so np(1 − p) = 21 gives p = 0.3 or p = 0.7. The model gets most questions right. Which fits?
Neither: 21 > 100/4
Both, with mean 50
p = 0.3, mean 30
p = 0.7, mean 70
Show the answer
p = 0.7, mean 70
Both roots give variance 21, since 100 × 0.3 × 0.7 = 21 either way. "Most right" means p above 0.5, so p = 0.7 and the mean is np = 70.
Distances and repeats, turned around. A z-score z = (x − μ)/σ has three other letters to free. Multiply by σ and add μ: x = μ + zσ, the value z standard deviations above the mean. Take zσ away: μ = x − zσ. Or divide: σ = (x − μ)/z, which needs z ≠ 0.
Independent repeats multiply. If each step goes right with chance p, all n steps do with chance pⁿ. For the most steps that keep this at or above a target, take logs (lesson 35): n·ln p ≥ ln(target). But ln p is negative for p between 0 and 1, so dividing by it flips the sign (lesson 36): n ≤ ln(target)/ln p. With p = 0.9 and a target of 0.5, n ≤ 6.58. Round down to 6, since 0.9⁷ ≈ 0.478 falls short.
In your own words
A rearranged chance comes out as 1.2. What should you conclude, and why?
Write a sentence or two, then compare it with a good answer in the lesson.
In Python: exact fractions and a seeded simulation
Purpose. Check a rearrangement two ways: exactly, with fractions, and by building the world forward with random numbers.
It runs right here in your browser, with nothing to install.
Fraction keeps every number exact. The map share is 3/10, and putting it back gives 31/50, which is 0.62, with no rounding error at all.
The seeded generator makes 100,000 hikers, about 30% with a map. Each one's chance to finish is 0.5 + 0.4 × has_map: True counts as 1, so that's 0.9 with a map and 0.5 without.
The simulated rate gives back about 0.3, not exactly. The rate wobbles a little, and the solve divides that wobble by 0.4, which makes it 2.5 times bigger.
Worked example
Worked example: a binomial's p from its spread
A model answers n = 100 questions. Its number right has variance 21 and mean 70. Solve np(1 − p) = 21 for both roots, keep the one the mean agrees with, and check E[X²] = Var + μ² by simulation.
import math
import numpy as np
defbinomial_p(n, var):
disc = 1 - 4 * var / n
if disc < 0:
returnNone
root = math.sqrt(disc)
return (round((1 - root) / 2, 6), round((1 + root) / 2, 6))
print(binomial_p(100, 21), binomial_p(100, 25), binomial_p(100, 30), round(70 / 0.7))
rng = np.random.default_rng(1)
x = rng.binomial(100, 0.7, size=100000)
mu, var = float(x.mean()), float(x.var())
print(round(mu, 1), round(var, 1))
print(round(float((x ** 2).mean()), 2), round(var + mu ** 2, 2), 21 + 70 ** 2)
In the lesson you can change this and run it.
A variance of 21 gives the pair 0.3 and 0.7, and the mean of 70 picks 0.7. Then n = μ/p = 70 ÷ 0.7 gives back the 100 questions. A variance of 25 is the largest any 100 trials can have, so both roots meet at 0.5. A variance of 30 has no real p, and the function says so with None. In the simulation, the last line's first two numbers match: for any list of numbers, the mean of the squares is exactly its variance plus its mean squared. They sit near 4,921, not on it, because the simulated mean and variance only come close to 70 and 21.
The "or" rule turned around gives P(A and B) = 1/2 + 2/5 − 7/10 = 1/5. Dividing by P(A) = 1/2 gives P(B | A) = 2/5. Fractions print as fractions.
Quick check
An agent gets each step right with chance 0.95, independently. Solving 0.95ⁿ ≥ 0.6 for n, what is the longest task it finishes at least 60% of the time?
At most 10 steps
At most 9 steps
At most 12 steps
At least 10 steps
Show the answer
At most 9 steps
n·ln 0.95 ≥ ln 0.6. ln 0.95 is negative, so dividing flips the sign: n ≤ 9.96. Round down, since 0.95¹⁰ ≈ 0.599 falls just short of 0.6.
Drill
Fix the backward rules
prior_from_total divides by the wrong thing and never checks its answer. most_steps rounds the wrong way. Fix both. Return None when the rates are equal or the answer leaves 0 to 1.
In the lesson, your code is checked as you write it.
In your own words
The simulated hikers gave a map share of about 0.3, not exactly 0.3. Why does solving make the wobble bigger, and what does each check tell you?
Write a sentence or two, then compare it with a good answer in the lesson.
In AI and ML: base rates, checks, thresholds and costs
Purpose. Four everyday questions in testing and running a model are probability formulas run backward.
The base rate behind a classifier. The base rate is the share of cases that really are positive. A spam filter on the club's message board has recall r = 0.8 (it catches 80% of spam), a false-positive rate f = 0.05 (it flags 5% of good posts) and precision 0.8 (80% of its flags are spam). Precision is Bayes' rule, with the base rate π inside it twice:
prec=rπ+f(1−π)rπ⇒π=r(1−prec)+prec⋅fprec⋅f
In plain words: clear the fraction, gather the π terms on one side, and divide by their bracket.
One post in five is spam, and putting 1/5 back into the precision formula gives 4/5 exactly.
Quick check
The same filter moves to another board. Its recall and false-positive rate stay the same, but its precision drops to 0.5. What happened to the base rate?
It rose, to about 50%
It stayed at 20%
It fell, to about 6%
It fell to 0
Show the answer
It fell, to about 6%
π = 0.5 × 0.05 ÷ (0.8 × 0.5 + 0.5 × 0.05) = 0.025 ÷ 0.425 ≈ 0.059. Spam is rarer here, so the 5% of good posts it flags make up more of its flags.
How many checks? A checking step catches a wrong answer with chance 0.6, and the checks are independent. All n miss with chance 0.4ⁿ, so by the complement at least one catches it with chance 1 − 0.4ⁿ. For 99%, 0.4ⁿ ≤ 0.01. Take logs, and divide by ln 0.4 < 0 to flip the sign: n ≥ 5.03. Round up to 6 checks, since 5 still miss 1.02%.
An anomaly threshold. A model server's response times have mean μ = 200 ms and σ = 12 ms. Flag any time more than 3σ above the mean: x = μ + zσ = 236 ms. Turned back, z = (x − μ)/σ says how unusual a time was.
Where two actions cost the same. The filter gives a post a spam chance p, then hides it or lets it through. Hiding a good post costs 1 (an annoyed member). Letting spam through costs 4. Letting through has expected cost 4p, and hiding has expected cost 1 × (1 − p). They're equal when 4p = 1 − p. Collect p: 5p = 1, so p = 0.2. Hide every post above 0.2. In general, with c_fp for a false alarm and c_fn for a miss, the threshold is c_fp/(c_fp + c_fn).
31Formulas with fractions: clear them firstFreeClearing denominators, cross-multiplying, a letter in the bottom, and reciprocal formulas like 1/R = 1/R₁ + 1/R₂.35 min
37Inverse functions, lines and scalingFreeUndoing a function by making x the subject, domains that allow it, line forms, and unscaling predictions.40 min
38Sequences and sums: find n, a, d and rFreeEvery letter of the arithmetic and geometric formulas as the subject, and an average's window from its decay.40 min
40Review: rearrange anything (algebra)FreeBroken rearrangements to fix, mixed practice from the whole algebra block, a strategy chart, and its formula sheet.35 min
41Coordinates and distanceFreePoints as (x, y), Pythagoras, distance in any number of coordinates, and nearest neighbors.42 min
42Angles, degrees and radiansFreeMeasuring a turn, converting units, arcs, and the angle between two arrows.36 min
43Sine and cosineFreeAcross and up: right triangles, the unit circle, and directions and rotations.42 min
44Waves: height, speed and shiftFreeAmplitude, period and phase, sampling a wave, and position codes in transformers.37 min
68Vector equations: rearranging with arrowsFreeSolving for a vector or a scale, recovering a gradient from two weight snapshots, and an angle from a dot product.40 min
87Counting: choices, orders and groupsFreeThe counting principle, factorials, ordered picks and groups, and why attention's cost grows with the square of the length.43 min
88Probability: from counts to chancesFreeEqually likely outcomes, the complement, distributions, seeded simulation, and a next-word model.43 min
89Combining events: and, or, independentFreeThe or rule, independence, probability trees, drawing with and without replacement, and chains of steps.42 min
100Maximum likelihoodFreeLikelihoods as products, log-likelihoods as sums, grid search, underflow, and the negative log-likelihood loss.45 min
101Linear regression: the best lineFreeResiduals, least squares by formula, three roads to one line, the matrix form, and train versus test error.45 min
109One neuron, start to finishFreeWeighted sum, activation, loss, slopes and update, by hand and in code, and what one straight line can and can't split.45 min
111Training a small networkFreeThe backward pass in matrix form, gradient checks, a full training loop, and reading loss curves.45 min
112The same network in PyTorchFreenn.Sequential with the NumPy weights loaded, the same loss and slopes, then SGD and mini-batches.45 min
113Attention as matrix mathFreeScaled dot-product scores, the causal mask, a row-by-row softmax, and blending value vectors.45 min
114Residual connections and layer normFreeAdding the input back and why its slope keeps learning alive, layer norm by hand, and add and norm.40 min
115A transformer block, start to finishFreeA sentence through embeddings, attention, add and norm, a small network and next-word scores, every shape named.45 min
117Final test: transpose any formulaFreeSlips to fix from every topic, mixed practice of every kind of rearranging, the strategy chart and the full sheet.45 min
Learn the math behind AI from arithmetic up: numbers, algebra, vectors and matrices, calculus, probability and statistics, each shown as numbers moving and checked in Python, ending with a network you build and train from the formulas.
Learn to make any letter of any formula the subject: layers undone in reverse order, fractions, a letter that appears twice, powers, roots, quadratics, logs, inequalities, trigonometry, vectors, matrices, calculus, probability, statistics and the formulas inside AI models, each checked in Python.
See the course →
Never trained an AI? Teach a small one to write, right in your browser. It takes about 20 seconds and needs no account. Teach it to write