Hilbert Space and the Hypercomplex Tower: ℝ, ℂ, ℍ, 𝕆, 𝕊 and Beyond

Table of Contents

Interactive explainer: Inner products, completeness, and the Cayley–Dickson ladder

Two great constructions run through modern mathematics. One goes up in dimension without end: Hilbert space, where vectors can be whole functions and sums can have infinitely many terms. The other doubles dimension in discrete jumps (ℝ → ℂ → ℍ → 𝕆 → 𝕊 → the 32-dimensional trigintaduonions) and pays a toll at every jump, surrendering one cherished law of arithmetic each time. This page builds both, step by step, and shows exactly where they meet.

1. Hilbert space: geometry with infinitely many directions

A Hilbert space is a vector space equipped with an inner product — a machine that takes two vectors and returns a number measuring how much they point the same way — and one extra demand: the space must be complete, meaning every sequence of vectors that ought to converge to something actually converges to a vector inside the space. Inner product gives geometry (lengths, angles, projections); completeness guarantees the geometry has no holes.

Everything starts with the inner product. In the plane, for u=(u1,u2) and v=(v1,v2), it is the familiar dot product. The definition, and the three quantities it buys you, are collected in one table:

QuantityFormulaWhat it means
Inner product⟨u,v⟩=u1v1+u2v2A single number: positive when u and v lean together, zero when perpendicular, negative when opposed.
Length (norm)‖u‖=⟨u,u⟩=u12+u22Pythagoras, restated: a vector’s length is its inner product with itself, square-rooted.
Anglecos⁡θ=⟨u,v⟩‖u‖‖v‖The inner product, normalised by the two lengths, is exactly the cosine of the angle between the vectors.
Projectionprojuv=⟨u,v⟩⟨u,u⟩·uThe shadow of v along u, the part of v that u can “see”. What remains, v−projuv, is perpendicular to u.

Figure 1: The inner product as geometry in the plane

θu = (4, 0)v = (3.2, 1.9)proj of v onto uO
Reading the picture. Here ⟨u,v⟩=4·3.2+0·1.9=12.8. The dashed drop from the tip of v lands at distance 3.2 along u, exactly the first component of v, and the leftover vertical segment is perpendicular to u. That splitting of a vector into “parallel part + perpendicular part” is the single most-used move in Hilbert space theory, and it works in every dimension, including infinitely many.

Try it: the inner-product explorer

One slider turns v around the origin; a second stretches it. The diagram, its moving labels, and the readouts below all follow the same two controls, so you can watch the inner product change sign as v swings past perpendicular — and check the Cauchy–Schwarz inequality |⟨u,v⟩|≤‖u‖·‖v‖, which says the inner product can never exceed the product of the lengths, with equality only when the vectors are parallel.

Figure 2: Inner-product explorer: v at 45°, |v| = 3.0

θ = 45°u, |u| = 3.8v, |v| = 3.0proj at 2.12O
What to try. At θ = 90° the inner product is exactly 0: perpendicular vectors are “orthogonal”, the Hilbert-space word for it. Past 90° the projection points backwards along u and the inner product goes negative. At θ = 0° Cauchy–Schwarz becomes an equality.

Completeness: no holes in the space

Geometry alone does not make a Hilbert space. Consider the rational numbers ℚ on a line: you can add them and measure distances, yet the sequence 1, 1.4, 1.41, 1.414, … (rationals closing in on 2) is a Cauchy sequence (its terms get arbitrarily close to each other) whose limit is missing from ℚ. A space with such a hole is an inner product space but not a Hilbert space. Completeness is the demand that every Cauchy sequence converges to a point of the space itself.

Figure 3: A Cauchy sequence closing in on its limit

limit Lx₁x₂x₃x₄x₅x₇gaps shrink
Why it matters. Completeness is what licenses infinite processes: Fourier series, limits of approximations, solutions of differential equations built as limits. ℝ is complete (that is essentially its construction from ℚ); ℚ is not. Every finite-dimensional inner product space is automatically complete — which is why the finite hypercomplex algebras later in this page are all, quietly, Hilbert spaces already. The interesting completeness questions begin at infinite dimension.

The two flagship infinite-dimensional examples

SpaceIts “vectors” and inner productWhy it is complete / useful
ℓ2: square-summable sequencesx=(x1,x2,x3,…) with ∑|xi|2<∞, and ⟨x,y⟩=∑xiy¯iThe direct infinite generalisation of ℂn. The condition ∑|xi|2<∞ is exactly “finite length”, and the inner-product series always converges by Cauchy–Schwarz.
L2: square-integrable functionsf a function on [a,b] with ∫|f|2<∞, and ⟨f,g⟩=∫abf(x)g(x)dxFunctions become points in a space; the integral is their dot product. Quantum mechanics lives here: a wavefunction ψ is a unit vector, and ⟨φ,ψ⟩ is a probability amplitude.

In L2, the trigonometric functions 1,cos⁡x,sin⁡x,cos⁡2x,sin⁡2x,… form an orthonormal basis — infinitely many mutually perpendicular unit directions. Any square-integrable function is a (possibly infinite) combination of them: its Fourier series. Completeness is precisely the theorem that the partial sums converge to the function in the L² sense. The explorer below builds a square wave out of sine waves, one basis direction at a time.

Figure 4: Fourier synthesis: a square wave built from 3 sine terms

target square wave (dashed)partial sum S_N(x)one period
How to read it. Each added term is one more basis direction of L2 being switched on. The approximation overshoots near the jumps by about 9% no matter how many terms you add (the Gibbs phenomenon) — pointwise it misbehaves there, yet the total squared error still shrinks to zero, which is convergence in the Hilbert-space sense. Infinite dimensions are not a luxury; a sharp edge genuinely needs all of them.

2. The hypercomplex tower at a glance

A hypercomplex number system is what you get when you try to keep doing arithmetic — add, subtract, multiply, take lengths — in more dimensions than the familiar number line and plane. There is a classical construction that doubles the dimension at every step (Section 3), producing an infinite ladder. But arithmetic gets worse as you climb. The whole story of the tower is in this ledger of what survives:

AlgebraDimBasis unitsOrderedCommutative
a·b=b·a
Associative
(a·b)·c=a·(b·c)
AlternativeDivision
(no zero divisors)
Normed
‖a·b‖=‖a‖‖b‖
ℝ reals11yesyesyesyesyesyes
ℂ complex21,ilostyesyesyesyesyes
ℍ quaternions41,i,j,klostlostyesyesyesyes
𝕆 octonions8e0…e7lostlostlostyesyesyes
𝕊 sedenions16e0…e15lostlostlostlostlostlost
𝕋 trigintaduonions32e0…e31lostlostlostlostlostlost

“Alternative” is the weakened associativity law a·(a·b)=(a·a)·b and (a·b)·b=a·(b·b): associativity when two of the three factors coincide. It is strong enough to keep octonion division working. Below the sedenions nothing in the table is left except addition, a conjugation, and a notion of length that no longer respects multiplication.

Figure 5: The tower: each step doubles the dimension and drops one law

ℝℂℍ𝕆𝕊𝕋dim 1dim 2dim 4dim 8dim 16dim 32ordering lostcommutativity lostassociativity lostalternativity, division, norm lostnothing left to lose but structure itself
The pattern. Dimension doubles; one law falls per step, in a fixed order. This is not an accident of one construction: Hurwitz’s theorem (1898) proves that ℝ, ℂ, ℍ and 𝕆 are the only normed division algebras over the reals, in any dimension, by any method. The ladder of well-behaved arithmetics provably ends at 8. Everything above is built anyway — and Section 4 shows what still works up there.

3. The Cayley–Dickson construction: how each level is built from the last

The doubling machine is due to Cayley and Dickson. Its input is an algebra A with a conjugation (an operation a↦a* that flips the sign of every imaginary part, like complex conjugation). Its output is a new algebra A2 whose elements are ordered pairs (a, b) of elements of A — so the dimension exactly doubles. Three definitions do all the work. They are stated once, in full generality, and then every algebra in the tower is just this recipe applied repeatedly:

IngredientDefinition in A2Notes
Elements(a,b),a,b∈AThink “a is the old part, b is the coefficient of one brand-new imaginary unit ℓ”. Formally (a,b) means a+bℓ.
Conjugation(a,b)=(a,−b)Conjugate the old part; negate the new part. For ℝ, a=a, so ℂ’s rule (a,b)=(a,−b) is ordinary complex conjugation.
Multiplication(a,b)(c,d)=(ac−db,da+bc)Every symbol on the right is an operation inside A. This one line, applied recursively, generates the multiplication tables of ℂ, ℍ, 𝕆, 𝕊, 𝕋, …
The new unitℓ=(0,1)Check it from the multiplication rule: (0,1)(0,1)=(0·0−1·1,1·0+1·0)=(−1,0). Each doubling manufactures a fresh square root of −1, perpendicular to all previous ones.
ℓ2=(−1,0)=−1

Worked check: the recipe really gives ℂ from ℝ

Take A = ℝ, where conjugation does nothing (a* = a). Apply the multiplication rule with all terms kept visible, in three steps — no step silently drops or rearranges anything:

StepComputationWhat happened
1. Substitute(a,b)(c,d)=(ac−db,da+bc)The general Cayley–Dickson rule, with the pair components named.
2. Use ℝ’s conjugation=(ac−db,da+bc) since d=dand c=c in ℝOnly the two conjugates are replaced; every product term is still present, in the same order.
3. Reorder real products=(ac−bd,ad+bc)Real multiplication commutes, so db=bd and da=ad. This is exactly the school rule for (a+bi)(c+di). The familiar formula was the doubling rule all along.

Run the same recipe on ℂ and the output is ℍ: writing a quaternion as a pair of complex numbers, q=(a,b) with a,b∈ℂ, is the standard “complex pair” form q=a+bj. Run it on ℍ to get 𝕆, on 𝕆 to get 𝕊, on 𝕊 to get the trigintaduonions 𝕋. One caveat explains the falling laws of Figure 5: the recipe’s output inherits a law only if the input had a stronger one. Doubling a commutative algebra yields a merely associative one waiting to happen, and so on — each application spends one law to buy twice the dimensions.

The doubling explorer

Figures 6 and 7 are driven by one shared slider: the doubling step k. Figure 6 shows the pair construction at that step; Figure 7 shows the basis it produces and the property ledger that results. Both titles track the slider.

Figure 6: Doubling step k = 3: building 𝕆 (8 dimensions) from pairs of ℍ

a ∈ ℍfirst componentb ∈ ℍsecond component→(a, b) = a + bℓa single new element→𝕆dim 8

ℓ = new unit e4: old units enter as (ei,0), new units as (0,ei−4).

How doubling feels from inside. Half of the new algebra is an exact copy of the old one — elements (a, 0) multiply exactly as a did. The other half, elements (0, b), is the old algebra “times ℓ”. Every multiplication in the new algebra is four multiplications in the old one, glued by the rule in the table above.

Figure 7: Basis and property ledger at k = 3: 𝕆, the octonions

real unit e₀ = 1inherited units (from the previous algebra)new units (added by this doubling)
e0e1e2e3e4e5e6e7
The ledger is the lesson. Slide back to k = 0 and everything is green; each step forward turns exactly one more property red, until at k = 4 (sedenions) division itself fails. The basis count, meanwhile, just keeps doubling — dimension is cheap, laws are not.

4. Walking the tower, one algebra at a time

ℝ: the real numbers (dimension 1)

The ground floor is a Hilbert space already: ℝ is ℝ1 with inner product ⟨a,b⟩=ab. It is the only level with a total ordering compatible with arithmetic (a<b implies a+c<b+c, and products of positives are positive). Everything above ℝ¹ will have to give that up — there is no way to say whether i is “positive” that survives multiplication, since i2=−1 would then have to be positive too.

ℂ: the complex numbers (dimension 2): pay ordering, gain roots

ℂ=ℝ2 as a Hilbert space, with the extra gift of multiplication. The imaginary unit i=(0,1) is the first Cayley–Dickson ℓ. Multiplying by i rotates the plane by 90° counter-clockwise, which is why i2=−1 is forced geometrically: two quarter-turns make a half-turn, and a half-turn sends every number to its negative.

Figure 8: Multiplication by i is a quarter-turn

1i−1−i×i×i×i×ireal axis →
The cycle closes. Four multiplications by i rotate through 360° and return every number to itself: i4=1. In exchange for ordering, ℂ gains the fundamental theorem of algebra — every polynomial has all its roots — and becomes the scalar field of quantum mechanics.

ℍ: the quaternions (dimension 4): pay commutativity, gain 3D rotation

Hamilton’s quaternions (1843) write q=a+bi+cj+dk with three imaginary units obeying the relations he carved into Brougham Bridge: i2=j2=k2=ijk=−1. From those, the full multiplication of units follows — and it is ordered: ij=k but ji=−k. Rotations in 3D do not commute either (tip a book forward-then-yaw versus yaw-then-forward), and that is precisely why quaternions, not despite but because of non-commutativity, are the working language of 3D rotation in graphics, robotics and spacecraft attitude control.

Figure 9: The quaternion unit cycle, and a multiplication bench

ijki·j = kj·k = ik·i = jfollow the arrows: product is the next unit;go against them: the negative
First factor a
Second factor b
Associativity survives. ℍ loses commutativity but keeps (a·b)·c=a·(b·c), keeps division (every non-zero q has inverse q−1=q*‖q‖2), and keeps the norm law ‖pq‖=‖p‖·‖q‖. By Hurwitz, it is the last level where multiplication is fully associative.

𝕆: the octonions (dimension 8): pay associativity, keep division

An octonion has seven imaginary units e1…e7 (with e0=1), each squaring to −1. Their multiplication is encoded by the Fano plane: seven points, seven lines of three points each. The rule: the three units on any line form a quaternion-style triple — the product of any two of them is the third, with the sign fixed by the line’s cyclic direction (click a line in Figure 10). But with three different lines involved, brackets start to matter. Here is a complete, checkable witness, with every intermediate product evaluated:

StepComputationWhat happened
1. Left bracketing, inside first(e1·e2)·e4=e3·e4e1·e2=e3, because {1, 2, 3} is a Fano line with cyclic order 1 → 2 → 3.
2. Left bracketing, finish(e1·e2)·e4=e7e3·e4=e7, from the line {3, 4, 7} with cyclic order 3 → 4 → 7.
3. Right bracketing, inside firste1·(e2·e4)=e1·e6e2·e4=e6, from the line {2, 4, 6} with cyclic order 2 → 4 → 6. Same three factors, different grouping.
4. Right bracketing, finishe1·(e2·e4)=−e7e1·e6=−e7: on the line {1, 7, 6} the pair (1, 6) runs against the cyclic order 1 → 7 → 6, so the sign flips.
5. Compare(e1e2)e4=+e7The two bracketings differ — not by a little, but by a sign on the whole answer. Associativity is dead.
e1(e2e4)=−e7

What survives is alternativity: bracketings still agree whenever two of the three factors are equal, and that is enough to keep inverses and the norm law working. 𝕆 is the last division algebra, and the last normed one — Hurwitz’s theorem says there is no ninth dimension where this could continue.

Figure 10: The Fano plane: octonion multiplication at a click

e₁e₂e₃e₄e₅e₆e₇
Seven lines, seven quaternionic triples. Each line of the Fano plane, together with e₀ = 1, spans a copy of ℍ inside 𝕆. The diagram is the entire multiplication table of the imaginary units, compressed: pick any two units, and the line through them names their product. Click each of the seven lines (the circle counts as a line) to read off its triple.

𝕊: the sedenions (dimension 16): pay division itself

Double once more and the failure is no longer about the order of operations — it is about existence. The sedenions contain zero divisors: non-zero elements whose product is exactly zero. That kills division (if xy=0 with x≠0, then y has no inverse), kills the norm law (‖x‖·‖y‖>0 but ‖xy‖=0), and kills alternativity too. Here is a fully explicit witness, expanded with every term kept, and every basis product evaluated using the Cayley–Dickson rule of Section 3:

StepComputationWhat happened
1. Set upx=e3+e10Both factors are plainly non-zero: ‖x‖=‖y‖=2.
y=e6−e15
2. Expandx·y=e3e6−e3e15+e10e6−e10e15Distributivity gives four terms; all four are kept, signs attached. Nothing has cancelled yet.
3. Evaluate each terme3e6=+e5Each product of basis units from the recursive Cayley–Dickson table for 𝕊. Note the repeats: e5 and e12 each appear twice.
e3e15=+e12
e10e6=+e12
e10e15=+e5
4. Substitute back, keeping all four termsx·y=(+e5)+(−e12)+(+e12)+(−e5)The signs from step 2 attach to the values from step 3. The four terms now visibly pair up.
5. Cancel in pairsx·y=(e5−e5)+(e12−e12)=0Total cancellation. Two non-zero sedenions multiply to zero — impossible in ℝ, ℂ, ℍ and 𝕆, routine in 𝕊.

Figure 11: Zero-divisor cancellation: four terms, two annihilating pairs

+e₅−e₁₂+e₁₂−e₅e₅ pair cancelse₁₂ pair cancels
Why this can only happen from 16 dimensions up. The cancellation needs two independent “copies” of octonion-like structure whose products can mirror and annihilate each other. Inside 𝕆 there is no room for the mirror; the doubling to 𝕊 creates it. Notice the ingredients: e3 and e6 live in the original 𝕆 half of 𝕊, while e10 and e15 live in the new ℓ half — a zero divisor is always a conspiracy between the old half and the new half.

𝕋: the trigintaduonions (dimension 32): the construction outlives the arithmetic

Nothing stops the machine. One more doubling gives the trigintaduonions: 32-dimensional numbers x=∑xiei with e0=1 and thirty-one imaginary units, built as pairs of sedenions, (a,b) with a,b∈𝕊, multiplied by the same one-line rule. (One proposed naming scheme also calls this level the pathions, the 64-dimensional level the chingons, the 128-dimensional the routons and the 256-dimensional the voudons — playful names signalling that these are now objects of study, not of calculation.)

What is true up here:

FeatureSituation in the trigintaduonions (dim 32)
Still standingAddition, subtraction, real scalar multiplication; conjugation x↦x; the quadratic norm N(x)=x·x=∑xi2, a non-negative real; power-associativity: powers xn are unambiguous because the subalgebra generated by any single element is commutative and associative (a general Cayley–Dickson theorem, true at every level, forever).
Long goneCommutativity, associativity, alternativity, division, and the multiplicative norm law — all lost at or before the sedenions and never regained. The sedenion zero divisors of Figure 11 embed unchanged: the same (e3+e10)(e6−e15)=0 holds verbatim in 𝕋, since 𝕊 sits inside 𝕋 as the pairs (a,0).
What proliferatesZero divisors become abundant and structured — in 𝕊 and beyond they organise into interlocking families (in 𝕊, the zero divisors built from basis units form patterns related to the “box” of 84 distinguished pairs). At 32 dimensions and above, classifying them is a research topic, not a table.
Why anyone caresThe tower’s well-behaved floors explain deep facts elsewhere: 𝕆 underlies the exceptional Lie groups and appears in string-theory compactifications; the pattern of losses (Hurwitz, and Bott–Milnor’s result that only dimensions 1, 2, 4, 8 admit the relevant parallelisable spheres S0,S1,S3,S7) ties the tower to topology. The higher floors are studied for the zero-divisor geometry itself and for proposed applications in signal processing and physics — with the honest caveat that without division, “solving equations” up here is a different and harder game.

Figure 12: The ladder never ends: dimension 2k against step k

1248163264128256512ℝℂℍ𝕆𝕊𝕋64128256k=0k=1k=2k=3k=4k=5k=6k=7k=8dimension (log₂ scale)
Exponential growth, linear chart. On a log2 axis the doubling ladder is a straight line — each step costs one more bit of “address” for a basis unit (an index from 0 to 2k−1) while the multiplication table grows as the square of the dimension: a trigintaduonion product is 32 × 32 = 1,024 basis products glued together by five nested layers of the Cayley–Dickson rule.

5. Where the two constructions meet

The tower and Hilbert space were never separate stories. Equip any level of the tower with its standard inner product: for x=∑xiei and y=∑yiei, take ⟨x,y⟩=∑xiyi, equivalently ⟨x,y⟩=Re(x·y*), and:

FactStatementConsequence
Every floor is a Hilbert space, as an inner product spaceℝ≅ℝ1Finite-dimensional inner product spaces are automatically complete, so each floor is a Hilbert space with extra structure (a multiplication) laid on top of the geometry.
ℂ≅ℝ2
ℍ≅ℝ4
𝕆≅ℝ8
𝕊≅ℝ16
𝕋≅ℝ32
The norm is the Hilbert norm‖x‖2=⟨x,x⟩=x·x*=∑xi2The “length” that Hurwitz’s theorem is about is exactly the Hilbert-space length. Normed division algebra = Hilbert space whose multiplication never stretches lengths unexpectedly.
The tower stops cooperating, Hilbert space doesn’tdim can be any n, or ∞, for a Hilbert space, but normed division only for n∈{{1,2,4,8}}Hilbert space asks only for linear structure + inner product + completeness, all of which scale to any dimension. The tower additionally asks for a well-behaved multiplication, and that is the scarce resource.
ℂ graduatesQuantum theory = Hilbert space over ℂThe second floor of the tower supplies the scalars for the infinite-dimensional theory: states are unit vectors in a complex Hilbert space, observables are its self-adjoint operators, and ⟨φ,ψ⟩ is the transition amplitude. The two constructions are, in the end, one subject.
The one-paragraph summary

Start with ℝ. Ask for more directions while keeping arithmetic perfect, and Cayley–Dickson doubling gives you ℂ, ℍ, 𝕆, charging ordering, then commutativity, then associativity, after which Hurwitz’s theorem proves the shop is closed: 𝕊 and the trigintaduonions beyond it keep the construction but lose division to zero divisors. Ask instead for more directions while keeping only geometry — inner product, length, angle, projection — plus the completeness that makes limits land, and you get Hilbert space, which happily extends to infinitely many dimensions, houses Fourier series and quantum states, and contains every floor of the tower as a finite-dimensional special case.

All diagrams are drawn directly in this page as SVG. Worked examples were chosen so every step can be checked by hand: the octonion associator (e1e2)e4=+e7 versus e1(e2e4)=−e7, and the sedenion zero divisor (e3+e10)(e6−e15)=0, both following from the Cayley–Dickson multiplication rule stated in Section 3, the latter holds verbatim inside the trigintaduonions as well.