Many questions ask "at least" or "at most", not "exactly": how many steps until the learning rate is small enough, or how many test examples keep an error bar under 2 points. Each is a formula set against a target. You rearrange it like an equation, with one extra job: keep track of which way the sign points. Most moves keep it. A few flip it, and this lesson names every one.
What you'll be able to do
Rearrange an inequality with the four moves, flipping the sign whenever you multiply or divide by a negative
Handle a letter of unknown sign by splitting into cases, or by using a known sign such as t > 0
Track the direction through reciprocals, squares, roots and logs, with the condition each one needs
Turn |x − a| < k into the range a − k < x < a + k, and round a boundary the right way to a whole number
Solve formula targets in training: steps until a decayed rate is small, a test-set size, clipping and a safe step size
Quick check
Have a guess before you start. 1/x < 1/4, and x is positive. Which x make it true?
x < 4
x > 4
x < 1/4
x > 1/4
Show the answer
x > 4
Try x = 8: 1/8 < 1/4 is true. Try x = 2: 1/2 < 1/4 is false. A bigger positive number has a smaller reciprocal, so taking 1 over both sides flips the sign: x > 4.
The math: rearrange, and track the direction
Purpose. Make a letter the subject of an inequality, and know at every step whether the sign stays or flips.
An inequality compares two sides with <, ≤, > or ≥. Transposing it, making a letter its subject, uses the same opposites as an equation. The four moves are adding, taking away, multiplying and dividing both sides by the same number:
Move on both sides
The sign
add or take away any number
stays
multiply or divide by a positive number
stays
multiply or divide by a negative number
flips
multiply by 0
not allowed: both sides become 0
Why the flip? Multiplying by −1 mirrors the number line around 0. So 3 < 5, but −3 > −5.
Take 7 − 2x ≥ 1. Take 7 from both sides: −2x ≥ −6. Divide by −2, a negative, so the sign flips: x ≤ 3.
Check it twice. Swap the sign for = to get the boundary, x = 3. Then test a number on each side. With x = 0: 7 ≥ 1 is true, and 0 ≤ 3. With x = 4: 7 − 8 = −1, and −1 ≥ 1 is false.
Quick check
From an earlier lesson: 2 < 5. Multiply both sides by −1. Which statement is true?
−2 < −5
−2 = −5
2 > 5
−2 > −5
Show the answer
−2 > −5
−2 sits to the right of −5 on the number line, so it's bigger. Multiplying by a negative reverses the order, so < becomes >.
Watch the boundary of 7 − 2x ≥ 1 built from x = 3, then peeled back. The way back runs "− 7" then "÷ −2". Ask at each step: does this move keep the sign, or flip it?
Try it now
Change the × −2 to × 2 and press Use these numbers: the way back now divides by a positive, and nothing flips.
Symbols and formula. The two flipping rules, for numbers a < b:
c<0⇒ca>cb,0<a<b⇒a1>b1
In plain words: multiplying by a negative flips the sign. Taking 1 over both sides flips it too, when both sides are positive.
A letter of unknown sign. In 12/t ≤ 3, you want to multiply by t. But is t positive or negative? Split into cases. If t > 0, the sign stays: 12 ≤ 3t, so t ≥ 4. If t < 0, it flips: 12 ≥ 3t, so t ≤ 4, and every negative t fits that. So 12/t ≤ 3 means t < 0 or t ≥ 4.
Often you know the sign already. A time, a count, a length or a speed is positive. Say so (t > 0), and you need only the first case: t ≥ 4. Never divide by a letter without knowing its sign.
Quick check
Solve 6/x > 2. Nothing is known about the sign of x.
x > 3
x < 3
x < 0 or x > 3
0 < x < 3
Show the answer
0 < x < 3
Case x > 0: multiply by x, sign stays: 6 > 2x, so x < 3. Together: 0 < x < 3. Case x < 0: the sign flips, 6 < 2x, so x > 3, which can't happen when x < 0. Check x = 1: 6 > 2 is true.
Squares and roots keep the sign only when both sides are 0 or more. 3 < 5 gives 9 < 25. But −5 < 3 gives 25 > 9: squaring broke it. So square both sides only when both are ≥ 0, and take square roots on the same terms. With no sign known, x² < 9 means −3 < x < 3, both roots kept, as in lesson 33.
Logs keep the sign when the log's base is above 1, as for ln (base e) and log₁₀. ln is increasing: a bigger number has a bigger log. Both sides must be positive, or the log doesn't exist. So 2ⁿ ≥ 1000 gives n·ln 2 ≥ ln 1000, and n ≥ 9.97.
The trap comes next. Take 0.9ⁿ ≤ 0.5. Logs keep the sign: n·ln 0.9 ≤ ln 0.5. Now divide by ln 0.9. It's about −0.105, a negative, because 0.9 is below 1. So the sign flips:
n≥ln0.9ln0.5≈6.58
Check it: 0.9⁶ ≈ 0.531, still above 0.5, while 0.9⁷ ≈ 0.478 is below. So n ≥ 7 whole steps.
Quick check
Solve 0.8ⁿ < 0.1 for n.
n < −10.32
n < 10.32
n > 10.32
n > −10.32
Show the answer
n > 10.32
Take logs: n·ln 0.8 < ln 0.1. ln 0.8 ≈ −0.223 is negative, so dividing by it flips: n > ln 0.1/ln 0.8 ≈ 10.32. Check: 0.8¹¹ ≈ 0.086, below 0.1.
Distance inequalities. |x − a| means the distance from x to a. So |x − a| < k says x is less than k away from a, on either side:
∣x−a∣<k⟺a−k<x<a+k
A bolt must be 20 mm long, give or take 0.5 mm: |x − 20| ≤ 0.5 means 19.5 ≤ x ≤ 20.5. Backward, the range 19.5 ≤ x ≤ 20.5 has center 20 and half-width 0.5. With a three-part range, do each move to all three parts:
Named mistakes.
Dividing by a negative and keeping the sign.
Dividing by a letter whose sign you don't know.
Squaring both sides when one side might be negative.
Dividing by ln γ with γ below 1, and forgetting the flip.
Writing |x − 20| ≤ 0.5 as x ≤ 20.5, and losing the lower end.
In your own words
List the moves that flip an inequality, and say what each one needs to be true.
Write a sentence or two, then compare it with a good answer in the lesson.
In Python: test both sides of the boundary
Purpose. Python can't rearrange for you, but it can catch a wrong direction in a second.
Write the original and your rearranged version as two functions that return True or False. Then test points on both sides of the boundary, and the boundary itself. They must agree everywhere.
deforiginal(x):
return7 - 2 * x >= 1defrearranged(x):
return x <= 3for x in [-10, 0, 2.9, 3, 3.1, 50]:
print(x, original(x), rearranged(x))
It runs right here in your browser, with nothing to install.
Every row agrees. If you had kept the sign as x ≥ 3, every row but one would disagree. The boundary row, 3, can't tell the two apart: x ≤ 3 and x ≥ 3 both include it. Only points on either side can.
Whole numbers round by the direction. "At most" rounds down, with math.floor. "At least" rounds up, with math.ceil.
Python also checks the numbers from the math part: the boundaries, ln 0.9, and the powers on each side.
The budget is "at most", so it rounds down. 10 nights cost $95, and an 11th would make $103. The ball is "at least", because of the flip: ln(0.05) and ln 0.6 are both negative, the boundary is 5.86, and n ≥ 5.86 rounds up to 6. The second line checks it: 0.0933 m after 6 bounces, but still 0.1555 m after 5.
Quick check
What does this print?
import math
b = math.log(0.5) / math.log(0.9)
print(round(b, 2), 0.9 ** 6 <= 0.5, 0.9 ** 7 <= 0.5)
6.58 False True
6.58 True False
6.58 True True
-6.58 False True
Show the answer
6.58 False True
Both logs are negative, so b ≈ 6.58 is positive. 0.9⁶ ≈ 0.531 is still above 0.5, and 0.9⁷ ≈ 0.478 is below. So 0.9ⁿ ≤ 0.5 needs n ≥ 6.58, the flipped direction.
Quick check
You rearranged an inequality and want to test it in Python. Which test can catch a missed flip?
A point on each side of the boundary
The boundary value on its own
Printing the rearranged rule
Rounding the boundary up
Show the answer
A point on each side of the boundary
x ≤ 3 and x ≥ 3 share the boundary 3, so only points on either side tell them apart.
Drill
Fix the direction slips
x_limit(c) should give the largest x with 7 − 2x ≥ c, and most_nights the most whole nights a budget buys. Each has one direction slip. Fix both.
In the lesson, your code is checked as you write it.
In your own words
Why can't checking only the boundary value catch a missed flip?
Write a sentence or two, then compare it with a good answer in the lesson.
In AI and ML: schedules, test sets, clipping and safe steps
Purpose. Training and evaluation are planned with "at least" and "at most". Each plan is an inequality turned around.
Steps until a decayed learning rate is small. Many training runs multiply the learning rate (the size of each training step) by a factor γ below 1 every step: lr = lr₀γᵗ. When is lr₀γᵗ ≤ target?
Divide by lr₀, which is positive: γᵗ ≤ target/lr₀. The sign stays.
Take ln, which is increasing: t·ln γ ≤ ln(target/lr₀). The sign stays.
Divide by ln γ, which is negative since γ < 1. The sign flips.
So t ≥ ln(target/lr₀)/ln γ. lesson 35 gave this formula; the flip is why it says "at least" and rounds up. With lr₀ = 0.1, γ = 0.95 and a target of 0.001, t ≥ ln 0.01/ln 0.95 ≈ 89.78, so 90 steps.
Quick check
A run starts at lr₀ = 0.2 and multiplies by γ = 0.9 each step. Which is right for lr₀γᵗ ≤ 0.002?
t ≥ ln 0.002/ln 0.9
t ≥ ln 0.01/ln 0.9
t ≥ 0.01/0.9
t ≤ ln 0.01/ln 0.9
Show the answer
t ≥ ln 0.01/ln 0.9
Divide by 0.2: 0.9ᵗ ≤ 0.01. Take logs: t·ln 0.9 ≤ ln 0.01. ln 0.9 is negative, so dividing flips: t ≥ ln 0.01/ln 0.9 ≈ 43.7, so 44 steps.
The test-set size for an error bar. An accuracy p, measured on n test examples, wobbles by about √(p(1 − p)/n). That's its standard error, which you'll meet properly in lesson 98. To keep it at or below a margin E:
Both sides are 0 or more, so square: p(1 − p)/n ≤ E². The sign stays.
Divide by p(1 − p), a positive number: 1/n ≤ E²/(p(1 − p)). The sign stays.
Take 1 over both sides, both positive. The sign flips: n ≥ p(1 − p)/E².
An accuracy near 0.8, with an error bar of 0.02 (2 points), needs n ≥ 0.16/0.0004 = 400 test examples.
Gradient clipping. Training takes steps along a gradient g. A rare huge g can wreck a run, so it's often clipped to |g| ≤ c, which means −c ≤ g ≤ c (clipping by value). With c = 5, a gradient of 7.5 becomes 5 and −12 becomes −5.
A safe step size. A standard result, taken as given here: for a bowl-shaped loss whose curvature (how fast its slope changes) is at most L, gradient descent with step η settles when 0 < η < 2/L. You can see it in the simplest bowl, (L/2)w², where each step multiplies w by (1 − ηL). It shrinks while |1 − ηL| < 1, so −1 < 1 − ηL < 1. Take 1 away: −2 < −ηL < 0. Divide by −L (L > 0), flipping both signs: 0 < η < 2/L. With L = 4, any step between 0 and 0.5 is safe.
A float trap. An accuracy near 0.95 with a 1-point error bar needs n ≥ 0.0475/0.0001 = 475. Python computes 475.0000000000004, a hair over, and math.ceil makes it 476. So round off that dust before you round up or down: math.ceil(round(x, 9)).
import math
defsteps_until(target, lr0, gamma):
return math.ceil(math.log(target / lr0) / math.log(gamma))
deftest_size(p, E):
return math.ceil(round(p * (1 - p) / E ** 2, 9))
defclip(g, c):
if g > c:
return c
if g < -c:
return -c
return g
defmax_step(L):
return2 / L
print(round(math.log(0.01) / math.log(0.95), 2), steps_until(0.001, 0.1, 0.95))
print(round(math.log(0.01) / math.log(0.9), 2), steps_until(0.002, 0.2, 0.9))
print(test_size(0.8, 0.02), test_size(0.5, 0.05), test_size(0.95, 0.01))
p, E = 0.95, 0.01print(p * (1 - p) / E ** 2, math.ceil(p * (1 - p) / E ** 2))
print(clip(7.5, 5), clip(-12, 5), clip(2, 5), max_step(4))
Quick check
What does this print?
import math
deftest_size(p, E):
return math.ceil(round(p * (1 - p) / E ** 2, 9))
print(test_size(0.5, 0.05), test_size(0.5, 0.025))
100 200
10 20
100 400
400 100
Show the answer
100 400
0.25/0.0025 = 100, and 0.25/0.000625 = 400. E is squared on the bottom, so halving the error bar needs four times the test examples.
Drill
Plan the run
steps_until rounds the wrong way, and test_size forgets a square. Fix both, so the first gives the first whole step with lr₀γᵗ ≤ target, and the second the fewest test examples for an error bar of at most E.
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