07.06.29 · representation-theory / lie-algebraic

Composition algebras and the octonions

shipped3 tiersLean: none

Anchor (Master): Springer-Veldkamp *Octonions, Jordan Algebras and Exceptional Groups* (2000) Ch. 1-2 (Hurwitz's theorem, the doubling, automorphisms); Conway-Smith *On Quaternions and Octonions* (2003) Ch. 6-9; Jacobson 1958 *Composition algebras and their automorphisms* (Rend. Circ. Mat. Palermo 7); Hurwitz 1898 *Über die Composition der quadratischen Formen* (Nachr. Ges. Wiss. Göttingen)

Intuition Beginner

We are used to four operations on numbers: add, subtract, multiply, divide. We are also used to measuring size — the absolute value of a real number, the modulus of a complex number. There is a remarkable compatibility between size and multiplication: the size of a product is the product of the sizes. For ordinary numbers this is the familiar rule that the modulus of a product equals the product of the moduli.

A number system in which you can add, subtract, multiply, and divide is called a division algebra. If on top of that it carries a notion of size obeying the size-of-product rule, it is a composition algebra. The question this unit answers is: how many such systems are there, if we insist the size be the ordinary positive length built from coordinates?

The answer is exactly four, and they have dimensions 1, 2, 4, and 8. They are the real numbers, the complex numbers, the quaternions, and the octonions. Each is built from the previous one by a doubling recipe that stacks two copies side by side. You cannot go further: there is no compatible sixteen-dimensional system.

As you climb the ladder you lose a familiar comfort at each rung. The complex numbers give up the ordering of the reals. The quaternions give up commutativity, so the order of multiplication matters. The octonions give up associativity, so even the grouping of a triple product matters. What survives all the way up is the size-of-product rule and the ability to divide — and that is exactly what pins the list to four.

Visual Beginner

Picture a four-rung ladder. The bottom rung is a single line: the real numbers, dimension 1. The next rung is a plane: the complex numbers, dimension 2, built by stacking two copies of the line. The third rung is dimension 4, the quaternions, built by stacking two copies of the plane. The top rung is dimension 8, the octonions, built by stacking two copies of the quaternions. Each step doubles the dimension and adds a label warning what was lost: "no ordering", then "no commuting", then "no grouping".

Above the top rung, a fifth rung is drawn and then crossed out: dimension 16, the sedenions. The doubling recipe still produces them, but the size-of-product rule breaks and you can multiply two nonzero things and get zero. The crossed-out rung is the picture of why the list stops at four.

Worked example Beginner

Take the complex numbers and check the size-of-product rule by hand. Write a complex number as a pair of real coordinates: $3 + 4 i$ has coordinates $(3, 4)$ and length $5$ , since $3 \times 3 + 4 \times 4 = 25$ and the square root of $25$ is $5$ . Take a second number $1 + 2 i$ , coordinates $(1, 2)$ , length the square root of $1 + 4 = 5$ .

Multiply them: $(3 + 4 i) (1 + 2 i) = 3 + 6 i + 4 i + 8 i^{2} = 3 + 10 i - 8 = - 5 + 10 i$ , coordinates $(- 5, 10)$ . Its length squared is $5 \times 5 + 10 \times 10 = 25 + 100 = 125$ , so its length is the square root of $125$ .

Now compare. The product of the two lengths is $5 \times$ the square root of $5$ , and squaring that gives $25 \times 5 = 125$ . The two answers agree: the length of the product equals the product of the lengths. This is the size-of-product rule in action, and it is the single property that, demanded in every dimension, forces the ladder to have exactly four rungs.

What this tells us: the compatibility we take for granted in the complex numbers is a strong constraint. Most ways of multiplying coordinate lists do not respect length. The few that do are special, and there are only four of them.

Check your understanding Beginner

Formal definition Intermediate+

Let $k$ be a field, here taken to be $R$ unless stated otherwise (the notion of a field and an algebra over it is that of 01.01.01). A composition algebra over $k$ is a unital (not necessarily associative) $k$ -algebra $A$ equipped with a non-degenerate quadratic form $N : A \to k$ — its norm — satisfying the multiplicative law $$ N(xy) = N(x),N(y) \qquad \text{for all } x, y \in A. $$ The norm is a quadratic form in the sense of 01.01.15; its polarisation is the symmetric bilinear form $⟨ x, y ⟩ = \frac{1}{2} (N (x + y) - N (x) - N (y))$ , and non-degeneracy of $N$ means this bilinear form is non-degenerate. The form $N$ being positive-definite (anisotropic over $R$ , $N (x) > 0$ for $x \neq = 0$ ) is the case of the four classical algebras; dropping definiteness admits the split companions discussed below.

Every element satisfies $⟨ 1, 1 ⟩ = N (1) = 1$ , and conjugation is the linear map $\overset{x}{ˉ} = 2 ⟨ x, 1 ⟩ 1 - x$ , fixing $1$ and negating the orthogonal complement $A_{0} = (k \cdot 1)^{⊥}$ of imaginary elements. Each $x$ satisfies the minimal relation $$ x^2 - 2\langle x, 1\rangle,x + N(x),1 = 0, \qquad x\bar x = \bar x x = N(x),1, $$ so over $R$ with $N$ positive-definite every nonzero $x$ has the two-sided inverse $x^{- 1} = \overset{x}{ˉ} / N (x)$ : a positive-definite composition algebra is a division algebra.

The Cayley-Dickson construction. Given a composition algebra $A$ with conjugation and norm, the doubling $A^{'} = A \oplus A$ carries the multiplication, conjugation, and norm $$ (a, b)(c, d) = (ac - \bar d, b,; d a + b \bar c), \qquad \overline{(a, b)} = (\bar a, -b), \qquad N(a, b) = N(a) + N(b). $$ Applied in sequence from $R$ this produces $R \to C \to H \to O$ , the reals, complexes, quaternions, octonions, of dimensions $1, 2, 4, 8$ . The quaternions $H$ are the even Clifford algebra $Cl_{3}^{0} ≅ Cl_{2}$ of 03.09.02; the octonions $O$ are the first doubling that escapes any Clifford algebra, because the next property to fall is associativity itself.

Alternativity. Octonion multiplication is not associative, but it is alternative: the associator $$ [x, y, z] := (xy)z - x(yz) $$ is an alternating function of its three arguments. Equivalently, $A$ satisfies $x (x y) = (xx) y$ and $(y x) x = y (xx)$ — any subalgebra generated by two elements is associative (Artin's theorem). Alternativity is sharpened by the three Moufang identities $$ (xy)(zx) = x\big((yz)x\big), \qquad (x(yz))x = (xy)(zx), \qquad x(y(xz)) = (x(yx))z, $$ which package every way a product of three or four octonions can be reassociated. The reals, complexes, and quaternions are associative; the octonions are alternative but not associative; the sedenions, the fifth doubling, are not even alternative and carry zero divisors.

Counterexamples to common slips

The doubling formula does not stop at $O$ : applying it once more yields the $16$ -dimensional sedenions $S$ . But $S$ has $N (x y) \neq = N (x) N (y)$ in general and contains nonzero $x, y$ with $x y = 0$ . It is not a composition algebra and not a division algebra. The chain of composition algebras genuinely terminates at $O$ .
"Alternative" is strictly weaker than "associative" and strictly stronger than "power-associative". The octonions are power-associative ( $x^{n}$ is unambiguous) and alternative, yet a generic triple has $(x y) z \neq = x (y z)$ . Do not conflate the partial grouping that Artin's theorem provides — valid only inside a two-generated subalgebra — with full associativity.
Dropping positive-definiteness does not break composition: the split algebras $C_{s}, H_{s}, O_{s}$ have isotropic norms (signature $(1, 1), (2, 2), (4, 4)$ ), still satisfy $N (x y) = N (x) N (y)$ , but possess zero divisors and so are not division algebras. They are composition algebras without the division property; Hurwitz's classification with definiteness dropped lists exactly the split companions alongside the four definite ones.

Key theorem with proof Intermediate+

Theorem (Hurwitz, 1898). The only finite-dimensional real composition algebras with positive-definite norm are $R, C, H, O$ , of dimensions $1, 2, 4, 8$ . Equivalently, the only positive-definite real normed division algebras are these four.

Proof. Let $A$ be a real composition algebra with positive-definite norm $N$ and associated inner product $⟨, ⟩$ . Polarising the multiplicative law $N (x y) = N (x) N (y)$ in each variable yields the two scaling identities $$ \langle xy, xz\rangle = N(x)\langle y, z\rangle, \qquad \langle xz, yz\rangle = N(z)\langle x, y\rangle, $$ and a further polarisation in mixed variables gives $⟨ x y, u z ⟩ + ⟨ u y, x z ⟩ = 2 ⟨ x, u ⟩ ⟨ y, z ⟩$ . Specialising these produces the conjugation relations $\overline{x y} = \overset{y}{ˉ} \overset{x}{ˉ}$ and $x \overset{x}{ˉ} = N (x) 1$ already recorded.

The engine of the classification is the doubling lemma: if $B ⊊ A$ is a composition subalgebra (closed under multiplication and conjugation, containing $1$ , with $N ∣_{B}$ non-degenerate), pick any $j \in B^{⊥}$ with $N (j) = 1$ . Then $B \oplus B j$ is a composition subalgebra, $B j ⊥ B$ , and the multiplication on $B \oplus B j$ is exactly the Cayley-Dickson formula with parameter $N (j) = 1$ . The proof is a direct computation using the scaling identities: one checks $b_{1} (b_{2} j) = (b_{2} b_{1}) j$ , $(b_{1} j) b_{2} = (b_{1} \overset{ˉ}{b}_{2}) j$ , and $(b_{1} j) (b_{2} j) = - \overset{ˉ}{b}_{2} b_{1}$ , and that $N$ restricted to $B \oplus B j$ remains the sum $N (b) + N (b^{'})$ .

Now build $A$ from the bottom. The subalgebra $R \cdot 1$ is always present. If $A \neq = R \cdot 1$ , choose $j_{1} \in (R 1)^{⊥}$ with $N (j_{1}) = 1$ ; the doubling lemma produces $R 1 \oplus R j_{1} ≅ C$ . If $A \neq = C$ , double again to get $H$ ; if $A \neq = H$ , double once more to get $O$ .

The chain must stop at $O$ . The Cayley-Dickson formula preserves the composition property only as long as the algebra being doubled is associative: the computation $(b_{1} j) (b_{2} j) = - \overset{ˉ}{b}_{2} b_{1}$ used to verify $N (x y) = N (x) N (y)$ on the double requires associativity of $B$ . The reals, complexes, and quaternions are associative, so the first three doublings give composition algebras. But $O$ is not associative, so doubling $O$ produces an algebra ( $S$ , dimension $16$ ) on which $N (x y) = N (x) N (y)$ fails. Concretely, the failure is measured by the associator: for the double, $N (x y) - N (x) N (y)$ is a nonzero multiple of $⟨[a, b, c], \cdot ⟩$ , which vanishes identically iff the doubled algebra is associative. Hence no composition algebra of dimension $16$ exists over $R$ , and the construction terminates after producing $R, C, H, O$ . Positive-definiteness guarantees at each stage that a unit-norm $j$ in the orthogonal complement exists, so the tower is forced — there is no room for an intermediate algebra of a non-doubling dimension. $□$

Bridge. This classification builds toward the entire theory of exceptional symmetry and appears again in 07.06.26, where the automorphism group $Aut (O) = G_{2}$ is the first exceptional Lie group, manufactured as the symmetries of the terminal algebra this theorem isolates. The foundational reason the list has length four is the failure of associativity past $O$ : the doubling lemma's composition check is exactly an associativity check, so the central insight is that "normed division algebra" and "associative-enough-to-double" coincide precisely up to dimension $8$ . This is exactly the structural fact that makes $O$ special rather than accidental, and it generalises the loss-of-property pattern — ordering, then commutativity, then associativity — into a theorem rather than a slogan. Putting these together, the bridge is the identification of the norm form $N$ of 01.01.15 with the quadratic form whose multiplicativity is the whole constraint: the quaternions arise here as the Clifford algebra $Cl_{3}^{0}$ of 03.09.02, and the octonions are precisely the first composition algebra that no Clifford construction can reach.

Exercises Intermediate+

Exercise 3 (medium, numeric).

In $O$ with imaginary basis $e_{1}, \dots, e_{7}$ and Fano rules including $e_{1} e_{2} = e_{3}$ , $e_{1} e_{4} = e_{5}$ , $e_{2} e_{4} = e_{6}$ , compute the associator $[e_{1}, e_{2}, e_{4}]$ and report its squared norm.

Hint

Compute $(e_{1} e_{2}) e_{4}$ and $e_{1} (e_{2} e_{4})$ separately. Distinct imaginary units anticommute, and the Fano line through two of them gives the third.

Answer

$(e_{1} e_{2}) e_{4} = e_{3} e_{4}$ . The Fano line through $e_{3}, e_{4}$ gives $e_{3} e_{4} = e_{7}$ (with the standard orientation). Meanwhile $e_{1} (e_{2} e_{4}) = e_{1} e_{6}$ , and the line through $e_{1}, e_{6}$ gives $e_{1} e_{6} = - e_{7}$ . So $[e_{1}, e_{2}, e_{4}] = e_{7} - (- e_{7}) = 2 e_{7}$ , with squared norm $N (2 e_{7}) = 4$ . The nonzero associator is direct evidence of non-associativity, and that $e_{1}, e_{2}, e_{4}$ lie on no common Fano line is exactly why grouping fails for this triple. Rubric: full credit for both groupings and the nonzero associator.

Exercise 4 (medium, short-answer).

Prove that the associator $[x, y, z]$ vanishes whenever two of its arguments are equal, and deduce that any two octonions generate an associative subalgebra.

Hint

Alternativity says $[x, y, z]$ is alternating. Use the defining identities $x (x y) = x^{2} y$ and $(y x) x = y x^{2}$ .

Answer

Alternativity is the statement that the associator is an alternating trilinear form, so it changes sign under any transposition of arguments; a form that is antisymmetric under swapping two equal arguments must vanish when they coincide, giving $[x, x, z] = [x, y, x] = [x, y, y] = 0$ . Equivalently these are the identities $x (x y) = x^{2} y$ , $(x y) y = x y^{2}$ , $(x y) x = x (y x)$ . Now for two elements $x, y$ : every product of $x$ 's and $y$ 's can be reassociated freely because each associator that arises has a repeated argument (only $x$ and $y$ are available), hence vanishes. Therefore the subalgebra generated by $x, y$ is associative — Artin's theorem. Rubric: full credit for the alternating-form vanishing and the repeated-argument reduction.

Exercise 5 (medium, short-answer).

Show that a finite-dimensional real division algebra cannot have a zero divisor, and use this to explain why the sedenions are not a division algebra.

Hint

In a division algebra every nonzero element has a two-sided inverse. A zero divisor $x$ has nonzero $y$ with $x y = 0$ .

Answer

If $A$ is a division algebra and $x \neq = 0$ , then $x^{- 1}$ exists; from $x y = 0$ we get $y = x^{- 1} (x y) = x^{- 1} \cdot 0 = 0$ (associativity or alternativity suffices to reassociate $x^{- 1} (x y)$ since only $x$ and $y$ appear). So nonzero $x$ is never a zero divisor. The sedenions $S$ contain nonzero $u, v$ with $uv = 0$ — concretely $(e_{1} + e_{10}) (e_{5} + e_{14}) = 0$ in the standard basis — so $S$ has zero divisors and is not a division algebra. This is the structural breakdown at the fifth doubling: norm-multiplicativity fails, and with it both the division and the composition property. Rubric: full credit for the no-zero-divisor argument and the sedenion zero divisor.

Exercise 6 (hard, short-answer).

Prove the doubling lemma's composition check fails exactly when the doubled algebra $B$ is non-associative: the Cayley-Dickson double of $B$ satisfies $N (x y) = N (x) N (y)$ for all $x, y$ if and only if $B$ is associative.

Hint

Write elements of the double as $x = (a, b)$ , expand $N (x y)$ using the doubling formula, and compare with $N (x) N (y) = (N (a) + N (b)) (N (c) + N (d))$ . The discrepancy is governed by an associator in $B$ .

Answer

Let $x = (a, b)$ , $y = (c, d)$ in $B \oplus B$ . The doubling formula gives $x y = (a c - \overset{ˉ}{d} b, d a + b \overset{c}{ˉ})$ , so $N (x y) = N (a c - \overset{ˉ}{d} b) + N (d a + b \overset{c}{ˉ})$ . Expanding each norm and using $N (uv) = N (u) N (v)$ on the associative-pair terms, the cross terms collect to $N (x) N (y) + 2 ⟨ a c, \overset{ˉ}{d} b ⟩$ -type expressions that cancel precisely when conjugation reverses products and reassociation is free. The residual is $2 (⟨(a c) ?, \cdot ⟩)$ proportional to the associator $[a, b, \overset{ˉ}{d}]$ in $B$ : it vanishes for all $a, b, d$ iff every such associator vanishes, i.e. iff $B$ is associative. Hence $R, C, H$ (associative) double to composition algebras, while doubling $O$ (non-associative) breaks $N (x y) = N (x) N (y)$ . Rubric: full credit for expanding the norm on the double, isolating the associator obstruction, and concluding the equivalence.

Advanced results Master

The full classification with split companions. Over a field $k$ of characteristic not $2$ , a composition algebra has dimension $1, 2, 4,$ or $8$ , and is determined up to isomorphism by its norm form. Over $R$ this gives, alongside the four positive-definite algebras, their split forms: $C_{s} = R \oplus R$ (norm of signature $(1, 1)$ ), the split quaternions $H_{s} ≅ M_{2} (R)$ (signature $(2, 2)$ ), and the split octonions $O_{s}$ (signature $(4, 4)$ ). The split algebras are composition algebras but not division algebras: their isotropic norms produce zero divisors, and $H_{s}$ is the genuinely associative $2 \times 2$ matrix algebra. Hurwitz's theorem in its general form lists exactly these — four definite and three split (with $R$ shared) — as the complete inventory of real composition algebras.

The Frobenius theorem as a corollary. Restricting to algebras that are associative and division, the doubling tower of the Hurwitz proof stops one rung earlier. Doubling $H$ produces $O$ , which is non-associative; so an associative real division algebra cannot reach dimension $8$ . This is the Frobenius theorem ^{[Frobenius 1878]}: the only finite-dimensional associative real division algebras are $R, C, H$ . The composition-algebra proof gives it cleanly: associativity is exactly the hypothesis the doubling lemma needs, and demanding it permanently caps the tower at $H$ . Dropping commutativity but keeping associativity buys one extra rung past $C$ ; dropping associativity for alternativity buys one more, to $O$ ; nothing buys a fourth.

The automorphism group and $G_{2}$ . Every automorphism of a composition algebra fixes $1$ and preserves the norm, hence acts orthogonally on the space of imaginary elements. For $O$ the imaginary part $Im O$ is $7$ -dimensional, and $Aut (O)$ is the compact exceptional Lie group $G_{2}$ of dimension $14$ , with Lie algebra the derivation algebra $Der (O) = g_{2}$ . The automorphism groups descend the ladder: $Aut (R) = 1$ , $Aut (C) = Z /2$ , $Aut (H) = SO (3)$ acting on the three imaginary quaternions, and $Aut (O) = G_{2}$ — the only one large enough to be exceptional. The split octonions $O_{s}$ have automorphism group the split real form $G_{2}^{(2)}$ . The detailed construction of $G_{2} = Der (O)$ as the stabiliser of a generic alternating $3$ -form on $R^{7}$ is carried out in 07.06.26.

Triality and $Spin (8)$ . The norm form on $O ≅ R^{8}$ has orthogonal group $O (8)$ and spin group $Spin (8)$ , whose Dynkin diagram $D_{4}$ has the symmetric triple node permuted by the outer automorphism group $S_{3}$ . This triality acts on the three $8$ -dimensional representations — the vector and the two half-spinors — and the octonion multiplication $O \times O \to O$ is precisely a triality-equivariant trilinear form: for $g \in Spin (8)$ there exist $g_{1}, g_{2}$ with $g_{1} (x) g_{2} (y) = \overline{g (\overline{x y})}$ . The composition law $N (x y) = N (x) N (y)$ is the bilinear shadow of this triality, and $G_{2}$ is recovered as the subgroup of $Spin (8)$ fixing the triality-defining vector. The triality picture is developed in 03.09.13.

Synthesis. The composition law $N (x y) = N (x) N (y)$ is the foundational reason the four-rung ladder exists and the foundational reason it stops: the doubling lemma turns "is a composition algebra" into "is built by doubling an associative algebra", so the central insight is that the failure of associativity at $O$ is identically the failure of the composition property at the sixteenth dimension. This is exactly the mechanism behind both Hurwitz and Frobenius — the latter is the former with the associativity demand made permanent, capping the tower at $H$ rather than $O$ . Putting these together, the terminal algebra $O$ is dual to the exceptional group $G_{2} = Aut (O)$ in the precise sense that the symmetries of the last division algebra are the first exceptional Lie group, and this generalises through the Freudenthal magic square into $f_{4}, e_{6}, e_{7}, e_{8}$ . The norm form of 01.01.15, the quaternions as the Clifford algebra of 03.09.02, and the triality of 03.09.13 are the three windows onto a single object: the unique non-associative normed division algebra, whose existence and uniqueness this unit proves and whose symmetry the rest of exceptional Lie theory inherits.

Full proof set Master

The Hurwitz theorem and the doubling lemma are proved in full in the Key theorem section. The remaining Master claims are recorded here.

Proposition (the doubling lemma, precise statement). Let $B$ be an associative composition algebra over $k$ with non-degenerate norm, and let $j \in B^{⊥}$ with $N (j) = μ \neq = 0$ inside a larger composition algebra $A$ . Then $B \oplus B j$ is a composition subalgebra with multiplication $(b_{1} + b_{1}^{'} j) (b_{2} + b_{2}^{'} j) = (b_{1} b_{2} + μ \overset{ˉ}{b}_{2}^{'} b_{1}^{'}) + (b_{2}^{'} b_{1} + b_{1}^{'} \overset{ˉ}{b}_{2}) j$ and norm $N (b + b^{'} j) = N (b) + μ N (b^{'})$ .

Proof. Orthogonality $B j ⊥ B$ follows from $⟨ b, b^{'} j ⟩ = ⟨ b \overline{b^{'}}, j ⟩ \cdot$ (const) $= 0$ since $b \overline{b^{'}} \in B ⊥ j$ . The product formulas come from the scaling and adjoint identities $⟨ x y, z ⟩ = ⟨ y, \overset{x}{ˉ} z ⟩ = ⟨ x, z \overset{y}{ˉ} ⟩$ , applied to the four cases $b_{1} b_{2}$ , $b_{1} (b_{2}^{'} j)$ , $(b_{1}^{'} j) b_{2}$ , $(b_{1}^{'} j) (b_{2}^{'} j)$ ; associativity of $B$ is used to move conjugations across products, e.g. $(b_{1}^{'} j) (b_{2}^{'} j) = μ \overline{b_{2}^{'}} b_{1}^{'}$ . The norm is additive because $B j ⊥ B$ and $N (b^{'} j) = N (b^{'}) N (j) = μ N (b^{'})$ by composition on the ambient $A$ . Verifying $N (x y) = N (x) N (y)$ on $B \oplus B j$ reduces, after expansion, to the vanishing of associators $[b_{1}, b_{2}, b^{'}]$ in $B$ , which holds because $B$ is associative. $□$

Proposition (Frobenius theorem). The only finite-dimensional associative division algebras over $R$ are $R, C, H$ .

Proof. Let $A$ be such an algebra. Each $x \in A$ generates a commutative subfield $R [x]$ , a finite real field extension, hence $R$ or $C$ ; so every element satisfies a real quadratic. Define $N (x)$ and $⟨, ⟩$ from this quadratic (the reduced norm and trace); non-degeneracy and positive-definiteness follow from $A$ being a division algebra over $R$ . One checks $N$ is multiplicative, so $A$ is a composition algebra, and associativity is given. By the doubling structure of the Hurwitz proof, an associative composition algebra arises by doublings of $R$ that never reach the non-associative stage: $R$ , then $C$ , then $H$ . The next doubling produces $O$ , which is not associative, contradicting the hypothesis. Hence $dim A \leq 4$ and $A \in {R, C, H}$ . $□$

Proposition (the Moufang identities hold in any alternative algebra). In an alternative algebra, the three Moufang identities $(x y) (z x) = x ((y z) x)$ , $(x (y z)) x = (x y) (z x)$ , and $x (y (x z)) = (x (y x)) z$ are satisfied.

Proof. In an alternative algebra the associator $[,,]$ is alternating, and the linearised alternative laws give $[x, y, z] + [y, x, z] = 0$ and cyclic variants. The Moufang identities follow by expanding each side into associators and using the alternating property to cancel. For the central identity, write $(x (y z)) x - (x y) (z x)$ as a sum of associators $[x, y z, x]$ and $[x, y, z x]$ -type terms; alternation forces $[x, w, x] = 0$ for all $w$ , and the remaining terms cancel pairwise under the sign-change of the alternating associator. The same bookkeeping, applied to the left and right Moufang expressions, gives the other two. The identities are exactly the obstruction-free reassociations available once the associator is alternating but not zero. $□$

Proposition (uniqueness of the norm). On a composition algebra the norm $N$ is the unique multiplicative non-degenerate quadratic form with $N (1) = 1$ , recovered from the algebra structure by $N (x) 1 = x \overset{x}{ˉ}$ where $\overset{x}{ˉ} = 2 ⟨ x, 1 ⟩ 1 - x$ and $⟨ x, 1 ⟩$ is the coefficient of the minimal quadratic of $x$ .

Proof. Each $x$ satisfies $x^{2} - t (x) x + n (x) 1 = 0$ for scalars $t (x), n (x)$ determined by the two-dimensional associative subalgebra $R [x]$ . Setting $⟨ x, 1 ⟩ = \frac{1}{2} t (x)$ and $N (x) = n (x)$ reconstructs the norm purely from multiplication. Multiplicativity then pins $N$ uniquely: any other multiplicative non-degenerate $N^{'}$ with $N^{'} (1) = 1$ agrees with $N$ on each $R [x]$ , hence everywhere by polarisation. $□$

Connections Master

The Clifford algebra 03.09.02 supplies the associative half of this ladder: the quaternions are $H ≅ Cl_{3}^{0} ≅ Cl_{2}$ , the even part of the Clifford algebra of $R^{3}$ , and the complex numbers are $Cl_{1}$ . The octonions are the first composition algebra that no Clifford algebra realises, because Clifford algebras are associative by construction while $O$ is not. This unit picks up exactly where the Clifford construction must stop, and the norm form here is the same quadratic form whose Clifford algebra produces $H$ there.
The bilinear / quadratic form 01.01.15 is the substrate of the whole theory: the norm $N$ is a quadratic form and its polarisation is the bilinear pairing $⟨, ⟩$ , and the entire content of "composition algebra" is the multiplicativity $N (x y) = N (x) N (y)$ of that form. Hurwitz's theorem is, read through this lens, a classification of multiplicative quadratic forms, and the split companions correspond to the indefinite-signature forms catalogued there.
The field 01.01.01 is the ground over which an algebra is defined and over which division is asked: a composition algebra is an algebra over $R$ , and the division property is the statement that it is a (non-associative) division ring extending the field $R$ . Frobenius's theorem, proved here as a corollary, is precisely the classification of the finite-dimensional associative division algebras over this field.
The construction $G_{2}$ via the octonions 07.06.26 is the immediate downstream payoff: it takes the terminal algebra $O$ isolated here and builds its automorphism group $Aut (O) = G_{2}$ as the first exceptional Lie group, reading the $G_{2}$ root system off the multiplication table. This unit supplies the existence and uniqueness of $O$ that that construction presupposes; without Hurwitz, "the octonions" would not be a well-defined object to take automorphisms of.
The triality unit 03.09.13 constructs $O$ by the complementary spinor route — squaring half-spinors of $Spin (8)$ — and exhibits octonion multiplication as a triality-equivariant trilinear form on the three $8$ -dimensional $D_{4}$ representations. The composition law $N (x y) = N (x) N (y)$ proved here is the bilinear shadow of that triality, and the two units present the same algebra through the division-algebra window and the spin-representation window.

Historical & philosophical context Master

The octonions were discovered by Arthur Cayley in $1845$ , shortly after Hamilton's $1843$ quaternions, and independently a little earlier by John Graves, to whom Hamilton had written about the quaternions; Graves called them "octaves". For half a century they were a curiosity without a structural home. The decisive theorem is Adolf Hurwitz's $1898$ Über die Composition der quadratischen Formen von beliebig vielen Variablen (Nachr. Ges. Wiss. Göttingen, 309–316) ^{[Hurwitz 1898]}, which proved that a sum-of-squares identity $(\sum x_{i}^{2}) (\sum y_{j}^{2}) = \sum z_{k}^{2}$ with $z_{k}$ bilinear in $x, y$ exists only in dimensions $1, 2, 4, 8$ — the algebraic core of the composition law. Earlier, Ferdinand Georg Frobenius's $1878$ Über lineare Substitutionen und bilineare Formen (J. reine angew. Math. 84) ^{[Frobenius 1878]} had classified the associative real division algebras as $R, C, H$ , the associative shadow of Hurwitz's later result.

The unification of the octonions with Lie theory is due to Élie Cartan and, structurally, to Nathan Jacobson, whose $1958$ Composition algebras and their automorphisms (Rend. Circ. Mat. Palermo 7) ^{[Jacobson 1958]} gave the modern account of composition algebras over arbitrary fields, the doubling construction now named for Cayley and Leonard Dickson, and the identification of the automorphism groups with the two real forms of $G_{2}$ . The Cayley-Dickson doubling itself was systematised by Dickson in the $1910$ s as a uniform mechanism producing the whole ladder and its sedenion sequel. The structural reason the ladder terminates — that each doubling sheds one law (ordering, commutativity, associativity, and finally the composition law itself) — was made precise in the composition-algebra framework, where the doubling lemma's composition check is identically an associativity check on the algebra being doubled.

What the classification settles is whether the exceptional structures of mathematics are accidents or are forced. The chain $R, C, H, O$ is forced: positive-definiteness supplies a unit-norm vector to double at each stage, and non-associativity halts the tower at dimension $8$ with no room for an intermediate algebra. That the symmetry group of the terminal algebra is the smallest exceptional Lie group, and that the same algebra reappears in triality, the magic square, and the holonomy classification, is the first evidence for the recurring thesis — argued at length by Baez and earlier by Freudenthal — that the exceptional objects across mathematics are the octonionic objects.

Bibliography Master

@article{hurwitz1898composition,
  author  = {Hurwitz, Adolf},
  title   = {{\"U}ber die Composition der quadratischen Formen von beliebig vielen Variablen},
  journal = {Nachrichten von der Gesellschaft der Wissenschaften zu G{\"o}ttingen, Mathematisch-Physikalische Klasse},
  pages   = {309--316},
  year    = {1898}
}

@article{frobenius1878lineare,
  author  = {Frobenius, Ferdinand Georg},
  title   = {{\"U}ber lineare Substitutionen und bilineare Formen},
  journal = {Journal f{\"u}r die reine und angewandte Mathematik},
  volume  = {84},
  pages   = {1--63},
  year    = {1878}
}

@article{jacobson1958composition,
  author  = {Jacobson, Nathan},
  title   = {Composition algebras and their automorphisms},
  journal = {Rendiconti del Circolo Matematico di Palermo},
  volume  = {7},
  pages   = {55--80},
  year    = {1958}
}

@article{baez2002octonions,
  author  = {Baez, John C.},
  title   = {The octonions},
  journal = {Bulletin of the American Mathematical Society},
  volume  = {39},
  number  = {2},
  pages   = {145--205},
  year    = {2002}
}

@book{springer2000octonions,
  author    = {Springer, Tonny A. and Veldkamp, Ferdinand D.},
  title     = {Octonions, Jordan Algebras and Exceptional Groups},
  series    = {Springer Monographs in Mathematics},
  publisher = {Springer-Verlag, Berlin},
  year      = {2000}
}

@book{conway2003quaternions,
  author    = {Conway, John H. and Smith, Derek A.},
  title     = {On Quaternions and Octonions: Their Geometry, Arithmetic, and Symmetry},
  publisher = {A K Peters, Natick, MA},
  year      = {2003}
}

Prerequisites

03.09.02
01.01.15
01.01.01

Tier anchors

beginner: Baez 2002 *The octonions* (Bull. AMS 39) §1-2 (the four normed division algebras, Cayley-Dickson doubling, the Fano-plane multiplication table); Stillwell *Naive Lie Theory* Ch. 1-2 (informal account of the number systems $\mathbb{R}, \mathbb{C}, \mathbb{H}$)
intermediate: Conway-Smith *On Quaternions and Octonions* (2003) Ch. 6-7 (composition algebras, the Cayley-Dickson construction, alternativity); Springer-Veldkamp *Octonions, Jordan Algebras and Exceptional Groups* (2000) Ch. 1 (the norm form $N(xy)=N(x)N(y)$, Hurwitz's theorem)
master: Springer-Veldkamp *Octonions, Jordan Algebras and Exceptional Groups* (2000) Ch. 1-2 (Hurwitz's theorem, the doubling, automorphisms); Conway-Smith *On Quaternions and Octonions* (2003) Ch. 6-9; Jacobson 1958 *Composition algebras and their automorphisms* (Rend. Circ. Mat. Palermo 7); Hurwitz 1898 *Über die Composition der quadratischen Formen* (Nachr. Ges. Wiss. Göttingen)

References

Springer-Veldkamp — Octonions, Jordan Algebras and Exceptional Groups (Springer Monographs in Mathematics, 2000) · Ch. 1 (composition algebras, the Cayley-Dickson doubling, Hurwitz's theorem 1.6.2), Ch. 2 (the structure group and triality), §2.2 ($\mathrm{Aut}(\mathbb{O}) = G_2$)
Baez — The octonions (Bull. AMS 39, 2002) · §1-2 (Cayley-Dickson construction, the four normed division algebras, Hurwitz and Frobenius); §4 ($G_2 = \mathrm{Aut}(\mathbb{O})$); §2.3 (triality and $\mathrm{Spin}(8)$)
Conway-Smith — On Quaternions and Octonions (A K Peters, 2003) · Ch. 6 (the composition law and the doubling), Ch. 7 (the Moufang identities and alternativity), Ch. 8 (the automorphism group)
Hurwitz — Über die Composition der quadratischen Formen von beliebig vielen Variablen (Nachr. Ges. Wiss. Göttingen, 1898) · pp. 309--316; the multiplicativity $N(xy)=N(x)N(y)$ forces dimension $1, 2, 4, 8$
Frobenius — Über lineare Substitutionen und bilineare Formen (J. reine angew. Math. 84, 1878) · the only associative real division algebras are $\mathbb{R}, \mathbb{C}, \mathbb{H}$

Estimated time

beginner: 20m
intermediate: 50m
master: 85m