Every vector so far was written in the usual axes: (4, 3) means 4 along x and 3 along y. Any two perpendicular directions, each 1 long, can serve as axes instead, and the same arrow then gets new numbers, its coordinates along them. Reading each one takes a single dot product. This matters because SVD and PCA, later in this module, hand back perpendicular unit directions. And researchers who study what a model has learned often find an idea stored as a direction in its vectors: how much of the idea a word has is its coordinate along that direction. This lesson tests directions, reads coordinates, rebuilds vectors from them, and takes a direction out.
What you'll be able to do
Test a set of vectors for being orthonormal: each 1 long, every pair perpendicular
Find a vector's coordinates in an orthonormal basis with one dot product each, and rebuild the vector from them
Say why orthonormal directions keep lengths and angles, and why reading coordinates and rebuilding undo each other
Take one direction out of a vector, and read how much of a made-up idea each word vector holds, in NumPy and PyTorch
Quick check
Have a guess before you start. Turn your map so that its "east" points along the unit arrow (0.6, 0.8). How far east does the arrow (3, 4) reach now?
4
5
3
7
Show the answer
5
(3, 4) is 5 × (0.6, 0.8): it points exactly along the new east, so all of its length, 5, is eastward. As a dot product: 3 × 0.6 + 4 × 0.8 = 1.8 + 3.2 = 5.
The math: new axes from perpendicular directions
Purpose. Describe a vector by how far it reaches along directions you choose.
The usual axes point along e₁ = (1, 0) and e₂ = (0, 1). Writing v = (4, 3) means 4 steps along e₁ plus 3 along e₂:
v=4e1+3e2
In plain words: a vector's numbers are amounts of the axis directions, added up.
A set of directions that builds every vector this way, each in exactly one way, is a basis (plural bases), and the amounts are the vector's coordinates in it. Any two directions that don't lie on one line make a basis of the plane. The usual axes have two extra properties: each is 1 long, and they meet at a right angle. Directions like that are orthonormal: "ortho" for right-angled, "normal" for length 1. Turn the usual axes and you get other orthonormal pairs, like q₁ = (0.6, 0.8) and q₂ = (−0.8, 0.6).
Quick check
From the last lesson: the shadow of v on b is v · b ÷ ‖b‖ long. How long is it when b is 1 unit long?
1
‖v‖
v · b
v · b ÷ 2
Show the answer
v · b
Dividing by ‖b‖ = 1 changes nothing, so the shadow is just v · b. That's why unit directions make every coordinate a single dot product.
Coordinates are shadows. The test for orthonormal, then the two formulas this lesson runs on:
qi⋅qi=1qi⋅qj=0(i=j)
ci=v⋅qiv=i∑ciqi
In plain words: each direction dotted with itself gives 1 (it's 1 long) and dotted with any other gives 0 (they're perpendicular). Then v's coordinate along each direction is one dot product: the length of its shadow on that direction's line. Each direction times its coordinate, all added, builds v back.
For q₁ = (0.6, 0.8) and q₂ = (−0.8, 0.6): q₁ · q₁ = 0.36 + 0.64 = 1, q₂ · q₂ = 0.64 + 0.36 = 1, and q₁ · q₂ = −0.48 + 0.48 = 0. They pass.
The picture. Step through v = (4, 3) in these turned axes. The drawing shows q₁ and q₂ as short arrows, with their lines dashed across the plane like a second pair of axes. The flow checks the two directions, drops v's shadow onto each line, then lays the two shadows tip to tail to rebuild v, and last compares the lengths.
Try it now
Then set directions q to 1, 1 and -1, 1 (one direction per line) and press Use these numbers. The two are perpendicular, but each is about 1.41 long, and the rebuild lands at (8, 6), twice v.
Why it works, and what it keeps. The two shadows meet at a right angle, so with v they make a right triangle, the flow's last picture. By Pythagoras, ‖v‖² = c₁² + c₂²: here 4.8² + 1.4² = 23.04 + 1.96 = 25 = 5². So the coordinates c = (4.8, −1.4) are exactly as long as v. Dot products are kept too, and so are angles, because a cosine is built from dot products and lengths. (4, 3) and (0, 5) have a cosine of 0.6, and so do their coordinates, (4.8, −1.4) and (4, 3), since (0, 5) · q₁ = 4 and (0, 5) · q₂ = 3. Turning the axes moves the numbers, never the arrows.
Quick check
Which pair of directions is orthonormal?
(1, 1) and (−1, 1)
(0, 1) and (−1, 0)
(0.6, 0.8) and (0.8, 0.6)
(1, 0) and (2, 0)
Show the answer
(0, 1) and (−1, 0)
(0, 1) · (0, 1) = 1, (−1, 0) · (−1, 0) = 1, and (0, 1) · (−1, 0) = 0: both 1 long, and perpendicular. They're the usual axes turned a quarter turn.
Reading undoes rebuilding. Rebuild v from its coordinates, then read its first coordinate back. Dot each piece of the rebuild with q₁ on its own: 4.8q₁ = (2.88, 3.84) gives 2.88 × 0.6 + 3.84 × 0.8 = 4.8, and −1.4q₂ = (1.12, −0.84) gives 1.12 × 0.6 + (−0.84) × 0.8 = 0. So the first coordinate comes back as exactly 4.8.
In plain words: the 1 and the 0 are the orthonormal test. Because every direction passes it, reading the coordinates of a rebuilt vector gives back the coordinates you built it from, and rebuilding from the coordinates you read gives back v. Each job undoes the other.
You'll see this written as… From lesson 53 on, the directions stand side by side as the columns of one table, Q. The rebuild is written Qc, reading is written Qᵀv (Qᵀ, "Q transpose", is Q turned so the directions are its rows), and the orthonormal test is QᵀQ = I, the identity: 1s down the diagonal, 0s elsewhere. So for orthonormal columns, the transpose is the inverse. This lesson needs none of that: just dot products.
Quick check
In the basis q₁ = (0, 1), q₂ = (−1, 0), what are the coordinates of v = (2, 5)?
The named mistake: directions that aren't 1 long. (1, 1) and (−1, 1) are perpendicular, but each is √2 long, so each dot product comes out √2 times too big, and the rebuild multiplies by √2 again. It lands at twice v. Divide each direction by its length first (lesson 49).
Taking a direction out. Subtract v's part along a unit direction q, and what's left has no shadow on q at all:
v⊥=v−(v⋅q)q
In plain words: find how far v reaches along q, step that far back along q, and you're left with the part of v that's perpendicular to q. The ⊥ sign means perpendicular.
For v = (4, 3) and q₁: 4.8 × (0.6, 0.8) = (2.88, 3.84), and (4, 3) − (2.88, 3.84) = (1.12, −0.84). That's exactly −1.4 × q₂, the other coordinate's part: in an orthonormal basis, taking one direction out leaves the others.
In your own words
In your own words: why does a unit direction make a coordinate just one dot product, and what goes wrong if a direction is 2 long?
Write a sentence or two, then compare it with a good answer in the lesson.
In Python: coordinates, rebuilds and removals
Purpose. Check a basis, read coordinates, rebuild, and take a direction out, one dot product at a time.
The checks print 1.0, 1.0 and 0.0, the coordinates 4.8 and −1.4, and the rebuild [4. 3.]: v again. The rounding hides float noise in the last digit, as in lesson 51.
Code flow. The same work as two loops: one reads the coordinates, one rebuilds. np.zeros(2) starts as two 0s. Step through and watch the boxes of c fill with 4.8 and −1.4, then rebuilt grow one direction at a time.
You'll see this written as one table. Code that works with many directions keeps them as the columns of one table, Q, and does each job in one line. You'll learn these lines properly in lesson 54 (.T, the transpose) and lesson 55 (@ between a table and a vector). For now, read them as shorthand for the loops you just stepped through: nothing later in this lesson needs them.
import numpy as np
Q = np.array([[0.6, 0.8], [-0.8, 0.6]]).T
v = np.array([4.0, 3.0])
print(np.round(Q.T @ Q, 6))
c = Q.T @ v
print(np.round(c, 6), np.round(Q @ c, 6))
Q.T @ Q is the table of every qᵢ · qⱼ: 1s and 0s. Q.T @ v is both coordinates, 4.8 and −1.4, and Q @ c is the rebuild, [4. 3.]: the stepper's numbers again.
Quick check
What does this print?
import numpy as np
v = np.array([2, 5])
q1 = np.array([0, 1])
q2 = np.array([-1, 0])
print(v @ q1, v @ q2)
5 -2
2 5
-2 5
5 2
Show the answer
5 -2
v @ q1 = 2 × 0 + 5 × 1 = 5 and v @ q2 = 2 × (−1) + 5 × 0 = −2: the coordinates from the math part's check.
Lengths and angles, kept. Here are the math part's claims in code: v = (4, 3) and w = (0, 5), read in the turned axes, and their lengths and the cosine of the angle between them, before and after.
import numpy as np
defcosine(a, b):
return (a @ b) / (np.linalg.norm(a) * np.linalg.norm(b))
q1 = np.array([0.6, 0.8])
q2 = np.array([-0.8, 0.6])
v = np.array([4.0, 3.0])
w = np.array([0.0, 5.0])
cv = np.array([v @ q1, v @ q2])
cw = np.array([w @ q1, w @ q2])
print(np.round(cv, 6), np.round(cw, 6))
print(round(np.linalg.norm(v), 6), round(np.linalg.norm(cv), 6))
print(round(cosine(v, w), 3), round(cosine(cv, cw), 3))
w's coordinates are (4, 3), both lengths are 5.0, and the cosine is 0.6 both ways: the arrows didn't move, only the axes.
The named mistake, in code. (1, 1) and (−1, 1) are perpendicular, but each is √2 long:
The coordinates come out as 7.0 and −1.0, each √2 times too big, and the rebuild lands at [8. 6.], twice v. Dividing each direction by its length makes the pair orthonormal, and the rebuild is v again.
Worked example
Worked example: a basis in three dimensions
Three perpendicular unit directions in 3D, each a third of a whole-number arrow. Check them, read the coordinates of (6, 3, 0), rebuild it, and compare lengths.
import numpy as np
qs = [np.array([1.0, 2.0, 2.0]) / 3,
np.array([2.0, 1.0, -2.0]) / 3,
np.array([2.0, -2.0, 1.0]) / 3]
for i inrange(3):
print(round(qs[i] @ qs[0], 6), round(qs[i] @ qs[1], 6), round(qs[i] @ qs[2], 6))
v = np.array([6.0, 3.0, 0.0])
c = np.zeros(3)
rebuilt = np.zeros(3)
for i inrange(3):
c[i] = v @ qs[i]
rebuilt = rebuilt + c[i] * qs[i]
print(np.round(c, 6), np.round(rebuilt, 6))
print(round(np.linalg.norm(v), 4), round(np.linalg.norm(c), 4))
In the lesson you can change this and run it.
The first three lines are the table of every qᵢ · qⱼ: 1s down the diagonal, 0s elsewhere. Each direction is 1 long, since (1 + 4 + 4) ÷ 9 = 1, and each pair is perpendicular. The coordinates are 4, 5 and 2, and the loop rebuilds (6, 3, 0). Both lengths are 6.7082, because 4² + 5² + 2² = 45 = 6² + 3² + 0²: Pythagoras with one more side. The code is the 2D code with a third direction, and it would be the same for three hundred.
Taking a direction out, in code. The math part's formula, one line, and a check that nothing of q is left: np.isclose(x, 0) says whether x is 0 apart from float noise.
import numpy as np
v = np.array([4.0, 3.0])
q = np.array([0.6, 0.8])
rest = v - (v @ q) * q
print(np.round(rest, 6), np.isclose(rest @ q, 0))
What's left is (1.12, −0.84), and its dot product with q is 0. Note the brackets: (v @ q) * q is a vector, how far v reaches along q times q. v - (v @ q) alone would subtract one number from every entry.
Quick check
What does this print?
import numpy as np
v = np.array([3.0, 4.0])
q = np.array([1.0, 0.0])
print(v - (v @ q) * q)
[0. 4.]
[0. 1.]
[-0. 4.]
[3. 0.]
Show the answer
[0. 4.]
v @ q = 3, so the line subtracts 3 × (1, 0) = (3, 0) and leaves (0, 4): taking out the across direction keeps only the up part.
Drill
Check a basis, read and rebuild
Fix the three functions. is_orthonormal(qs) should check that every direction is 1 long and every pair is perpendicular (allow 1e-9 of float noise). coords(v, qs) should give one dot product per direction, and rebuild(c, qs) the sum of each direction times its coordinate. There's one slip in each.
import numpy as np
defis_orthonormal(qs):
for i inrange(len(qs)):
ifabs(qs[i] @ qs[i] - 1) > 1e-9:
returnFalsereturnTruedefcoords(v, qs):
c = np.zeros(len(qs))
for i inrange(len(qs)):
c[i] = v @ qs[0]
return c
defrebuild(c, qs):
total = np.zeros(len(qs[0]))
for i inrange(len(qs)):
total = total + c[i]
return total
qs = [np.array([0.6, 0.8]), np.array([-0.8, 0.6])]
v = np.array([4.0, 3.0])
print(is_orthonormal(qs), coords(v, qs), rebuild(coords(v, qs), qs))
In the lesson, your code is checked as you write it.
In your own words
Your rebuild of (4, 3) came back exactly. What about the directions made sure it would, and what would you check first if a rebuild missed?
Write a sentence or two, then compare it with a good answer in the lesson.
In AI and ML: an idea as a direction
Purpose. Read how much of an idea each word vector holds, then take the idea out.
Researchers who study what a model has learned, a field called interpretability, often find an idea stored as a direction in its vectors: a word holds the idea to the extent its vector reaches along that direction. That reach is this lesson's coordinate, one dot product with a unit direction.
Here are the camp log's six words from lesson 51, three numbers each, and a made-up "tent" direction: the shelter idea that tent and tarp share. A simple way to make such a direction is to add the vectors of words that have the idea and make the sum 1 long.
import numpy as np
names = ["tent", "tarp", "stove", "kettle", "map", "compass"]
vecs = [np.array([0.9, 0.1, 0.0]), np.array([0.8, 0.0, 0.2]), np.array([0.1, 0.9, 0.0]),
np.array([0.0, 0.8, 0.1]), np.array([0.1, 0.0, 0.9]), np.array([0.0, 0.1, 0.8])]
t = vecs[0] + vecs[1]
t = t / np.linalg.norm(t)
print(np.round(t, 3))
for i inrange(len(names)):
print(names[i], round(vecs[i] @ t, 3))
t is (0.991, 0.058, 0.117): mostly the first number, shelter. tent holds the most of it, 0.898, then tarp, 0.816. The cooking words hold little: stove 0.152 and kettle 0.058.
Quick check
map scores 0.204 along the tent direction, more than stove's 0.152. Why?
map is a kind of shelter, like a tent
Dot products favor longer vectors
0.204 was rounded up from 0.152
t picks up tarp's bit of finding the way
Show the answer
t picks up tarp's bit of finding the way
t is tent plus tarp, made 1 long, and tarp has 0.2 in its third number, finding the way. So t carries 0.117 there, and map's big third number, 0.9, picks that up. A made-up direction carries whatever its example words carry.
Taking the idea out. Remove t from every word with the math part's formula, then compare pairs of words by cosine before and after:
Nothing of t is left in tarp (True). Taking the idea out shrinks tent from 0.906 long to 0.115: the idea was nearly all of it. tent and tarp were 0.964 alike, and what's left of them scores −1.0, exactly opposite. That's no accident: t was built from these two words alone, so what's left of tent and what's left of tarp add up to nothing, and only how they differ remains. stove and kettle hardly change, 0.986 before and 0.991 after: taking one direction out leaves everything perpendicular to it alone.
Researchers use this move in two ways. To test whether a model relies on an idea, they take the idea's direction out of the model's vectors and see what the model now gets wrong. To strip an unwanted idea, such as gender, from words that shouldn't carry it, they take its direction out of those words' vectors: an early way to reduce bias in word vectors. One direction rarely holds all of an idea, though, so traces of it can stay in what's left.
Quick check
After the tent direction is taken out, tent and tarp score −1.0. Why exactly opposite?
Any two short vectors have a cosine of −1
Taking a direction out always turns words around
t came from just these two, so their leftovers add to nothing
tent and tarp pointed opposite ways from the start
Show the answer
t came from just these two, so their leftovers add to nothing
t points along tent + tarp. Take it out of both and the two leftovers add up to (tent + tarp) with its own direction taken out, which is nothing. Two arrows that add to nothing point exactly opposite ways.
In PyTorch. The same lines on tensors, with the six words as the rows of one table, as in lesson 51's search. F.normalize(..., dim=0) makes the tent direction 1 long, words @ t reads every word's amount at once (one dot product per row), and the removal line is the NumPy one. The last four lines meet a tool from lesson 63: torch.linalg.svd takes the table of words apart into three pieces, named here U, S and Vh. The rows of Vh are three directions in the words' own 3-number space, the kind PCA reads data along, and the last two lines give them this lesson's test. abs keeps float noise from printing as −0.0.
The amounts are the NumPy cell's numbers, 0.898 for tent down to 0.058 for kettle, and nothing of t is left in tarp. SVD's three directions pass the test: each is 1 long, and every pair is perpendicular. They're an orthonormal basis that SVD found from the words alone.
Drill
Take an idea out
Fix amount(v, t), how far v reaches along the direction t, and remove(v, t), v with that direction taken out. t may not be 1 long, so both must make it 1 long first. Then set cleaned to tarp with the tent direction removed.
import numpy as np
defamount(v, t):
return v @ t
defremove(v, t):
t = t / np.linalg.norm(t)
return v - (v @ t)
tent = np.array([0.9, 0.1, 0.0])
tarp = np.array([0.8, 0.0, 0.2])
t = tent + tarp
cleaned = tarp
print(round(amount(tarp, t), 3), np.round(cleaned, 3))
In the lesson, your code is checked as you write it.
In your own words
Taking a direction out of every word vector removes one idea. What does it keep, and what can it lose by accident?
Write a sentence or two, then compare it with a good answer in the lesson.
Goes deeper in
Eigenvectors (intuition): the directions a matrix only stretches, which come out orthonormal for symmetric matrices.
SVD and PCA: orthonormal directions ranked by how much of the data lies along each.
Interpretability: looking inside models, often by reading their vectors along directions.
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.
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