21.02.06 · number-theory / quadratic-forms-local-fields

Witt's theorem: cancellation and the Witt decomposition

shipped3 tiersLean: none

Anchor (Master): Witt 1937 *J. reine angew. Math.* 176, 31-44 (originator: cancellation, the extension theorem, and the Witt ring); Serre 1973 *A Course in Arithmetic* Ch. IV §§1-2; Lam 2005 *Introduction to Quadratic Forms over Fields* (AMS GSM 67) Ch. I-II (modern textbook treatment of the Witt ring and Witt decomposition); Scharlau 1985 *Quadratic and Hermitian Forms* (Springer Grundlehren 270) Ch. 1-2; Milnor-Husemoller 1973 *Symmetric Bilinear Forms* (Springer Ergebnisse 73)

Intuition Beginner

A quadratic form is a homogeneous degree-two polynomial such as $3 x^{2} - 2 y^{2} + 5 z^{2}$ . Two forms are the same form in disguise — isometric — when a linear change of variables turns one into the other. The basic question of the whole subject is: how do we tell two forms apart, and what is the shortest list of numbers that pins a form down up to isometry?

Witt's theorem answers this with a single clean picture. Every form is built out of two kinds of pieces. The first kind is a form that never takes the value zero except at the origin — an anisotropic form, a genuinely rigid core. The second kind is a hyperbolic plane, the form $x y$ (equivalently $u^{2} - v^{2}$ ), which is as far from rigid as possible: it has whole lines of inputs on which it vanishes.

Witt's discovery is that the rigid core and the count of hyperbolic planes are not a matter of how you happen to write the form. They are intrinsic. Two forms with the same core and the same number of hyperbolic pieces are isometric, and you may cancel a shared summand from both sides of an isometry. So "the anisotropic core" and "how many hyperbolic planes split off" are two well-defined measurements of any quadratic form.

Visual Beginner

Picture a quadratic form as a stack of blocks. The bottom block is the anisotropic kernel: a solid, indivisible core that records the form's true shape. On top sit identical thin slabs, one for each hyperbolic plane, each slab being the split form $x y$ that vanishes along two crossing lines.

The number of slabs is the Witt index. Witt's theorem says the height of the solid core and the number of slabs do not depend on how you cut the stack apart: every way of separating core from slabs gives the same answer. Over the real numbers the core is recorded by a single integer, the signature, recovering the older Sylvester picture as the simple case.

Worked example Beginner

Take the real form $f (x, y, z) = x^{2} + y^{2} - z^{2}$ . Does a hyperbolic plane split off, and what is left over?

Look for a line on which $f$ vanishes. Setting $x = 0$ gives $y^{2} - z^{2} = (y - z) (y + z)$ , which is zero when $y = z$ . So the vector $(0, 1, 1)$ satisfies $f = 0$ without being the origin; the form is isotropic.

Whenever a non-degenerate form has such a zero vector, the subspace it spans pairs up with a partner vector to form a copy of the hyperbolic plane. Here the pair $(0, 1, 1)$ and $(0, 1, - 1)$ span a plane on which $f$ restricts to a multiple of $u^{2} - v^{2}$ . What remains orthogonal to that plane is the line spanned by $(1, 0, 0)$ , where $f = x^{2}$ .

So $f ≅ ⟨ 1 ⟩ ⊥ H$ : an anisotropic core $⟨ 1 ⟩$ of rank one, orthogonal to one hyperbolic plane. The Witt index is $1$ . The core $x^{2}$ never vanishes off the origin, so it cannot be cut down further. This matches the signature count: $f$ has signature $(2, 1)$ , and pairing one plus with one minus into a hyperbolic plane leaves a lone plus.

Check your understanding Beginner

Formal definition Intermediate+

Throughout, $k$ is a field of characteristic different from two, and $(V, q)$ is a finite-dimensional quadratic space: $V$ is a $k$ -vector space and $q : V \to k$ is a quadratic form, with associated symmetric bilinear form $b (x, y) = \frac{1}{2} (q (x + y) - q (x) - q (y))$ , so that $q (x) = b (x, x)$ . The space is non-degenerate when $b (x, y) = 0$ for all $y$ forces $x = 0$ .

An isometry between quadratic spaces $(V, q)$ and $(V^{'}, q^{'})$ is a linear isomorphism $σ : V \to V^{'}$ with $q^{'} (σ x) = q (x)$ for all $x \in V$ ; we write $q ≅ q^{'}$ when one exists. The isometries of $(V, q)$ to itself form the orthogonal group $O (q) = O (V, q)$ . Two subspaces are orthogonal, written $⊥$ , when $b$ pairs them to zero; an orthogonal direct sum of quadratic spaces is denoted $q_{1} ⊥ q_{2}$ , and corresponds to a block-diagonal Gram matrix. Diagonalisation over $k$ gives $q ≅ ⟨ a_{1} ⟩ ⊥ \dots ⊥ ⟨ a_{n} ⟩$ with $a_{i} \in k^{\times}$ , abbreviated $⟨ a_{1}, \dots, a_{n} ⟩$ .

A non-zero vector $v$ is isotropic when $q (v) = 0$ , and the space $(V, q)$ is isotropic when it contains an isotropic vector; otherwise $(V, q)$ is anisotropic. The hyperbolic plane $H$ is the two-dimensional non-degenerate space with form $q (x, y) = x y$ , isometric to $⟨ 1, - 1 ⟩$ over a field of characteristic not two. A space isometric to an orthogonal sum $H ⊥ \dots ⊥ H$ of $m$ copies, written $H^{⊥ m}$ , is hyperbolic (also called split) of dimension $2 m$ .

The Witt index of $(V, q)$ , written $ind (q)$ , is the dimension of a maximal totally isotropic subspace — a subspace $W \subseteq V$ on which $q$ vanishes identically. The discriminant $d (q) = a_{1} \dots a_{n} \in k^{\times} / (k^{\times})^{2}$ and, over $R$ , the signature are isometry invariants carried over from 01.01.15.

Counterexamples to common slips

Characteristic two is excluded. There the formula $b (x, y) = \frac{1}{2} (\dots)$ has no sense, the quadratic form and its polarisation decouple, and the entire diagonalisation-and-cancellation package needs a separate treatment (Arf invariant in place of discriminant).
Non-degeneracy is required for cancellation. The degenerate form $⟨ 0 ⟩$ satisfies $⟨ 0 ⟩ ⊥ ⟨ 1 ⟩ ≅ ⟨ 0 ⟩ ⊥ ⟨ 1 ⟩$ in many spurious ways once a radical is present, and cancellation can fail. One always splits off the radical first and works with the non-degenerate quotient.
"Isotropic" is not "degenerate". The hyperbolic plane $H$ is isotropic (it has zero vectors) yet non-degenerate (its Gram determinant is $- 1 \neq = 0$ ). Isotropy is about zero values of $q$ ; degeneracy is about the radical of $b$ .

Key theorem with proof Intermediate+

Theorem (Witt's extension and cancellation theorems; Witt 1937). Let $(V, q)$ be a non-degenerate quadratic space over a field $k$ of characteristic not two. (Extension) Every isometry $τ : U \to U^{'}$ between subspaces $U, U^{'} \subseteq V$ extends to an isometry $τ \in O (q)$ of the whole space. (Cancellation) If $q ⊥ q_{1} ≅ q ⊥ q_{2}$ for non-degenerate forms, then $q_{1} ≅ q_{2}$ .

Proof. We first prove extension, then derive cancellation. The engine is the reflection. For an anisotropic vector $v$ (so $q (v) \neq = 0$ ), the reflection $τ_{v} (x) = x - \frac{2 b ( x , v )}{q ( v )} v$ lies in $O (q)$ , fixes the hyperplane $v^{⊥}$ , and sends $v \mapsto - v$ .

The extension theorem is proved by induction on $dim U$ . For $dim U = 1$ , write $U = k u$ and $U^{'} = k u^{'}$ with $q (u) = q (u^{'}) = a \neq = 0$ (the degenerate case $a = 0$ is folded in below). Consider $q (u - u^{'}) + q (u + u^{'}) = 2 q (u) + 2 q (u^{'}) = 4 a \neq = 0$ , so at least one of $u - u^{'}$ and $u + u^{'}$ is anisotropic. If $u - u^{'}$ is anisotropic, the reflection $τ_{u - u^{'}}$ swaps $u$ and $u^{'}$ , sending $u \mapsto u^{'}$ ; if instead $u + u^{'}$ is anisotropic, then $τ_{u + u^{'}}$ sends $u \mapsto - u^{'}$ , and composing with $τ_{u^{'}}$ sends $u \mapsto u^{'}$ . Either way a product of at most two reflections carries $u$ to $u^{'}$ , extending $τ$ .

For the inductive step with $dim U \geq 2$ , choose a decomposition $U = U_{0} ⊥ k u$ with $q (u) \neq = 0$ (a diagonalising basis of $U$ supplies one once we handle the isotropic case by first splitting off hyperbolic planes from $U$ and matching them across, which the rank-one case and non-degeneracy permit). Apply the rank-one case to extend $τ ∣_{k u}$ to an isometry $σ \in O (q)$ with $σ (u) = τ (u)$ . Then $σ^{- 1} τ$ maps $U_{0}$ into $u^{⊥}$ and fixes $u$ ; by induction inside the non-degenerate space $(τ u)^{⊥}$ it extends, and composing the extensions gives $τ$ . This proves extension.

Cancellation follows. Suppose $φ : q ⊥ q_{1} \to q ⊥ q_{2}$ is an isometry, where the first copy of $q$ sits on a subspace $U$ and the target copy on a subspace $U^{'}$ , both isometric to $(V_{q}, q)$ via the inclusion. Let $τ = (incl_{U^{'}})^{- 1} \circ φ ∣_{U} : U \to U^{'}$ be the induced isometry of the two copies of $q$ . By extension, $τ$ lifts to $τ \in O (q ⊥ q_{2})$ of the ambient target. Then $τ^{- 1} \circ φ$ is an isometry carrying $U$ identically onto the standard copy of $q$ , hence carrying its orthogonal complement $q_{1}$ isometrically onto the orthogonal complement $q_{2}$ . Therefore $q_{1} ≅ q_{2}$ . $□$

Bridge. Witt cancellation builds toward the entire local-global theory and is the foundational reason the Hasse-Minkowski classification can speak of "the anisotropic part" of a rational form at all. The central insight is that orthogonal summands may be cancelled, which is exactly the statement that the monoid of isometry classes under $⊥$ has cancellation and so embeds in a group; this group is the Witt-Grothendieck group, and quotienting by hyperbolic planes is the bridge to the Witt ring. The extension theorem generalises the Sylvester rigidity of real forms to an arbitrary field, and is dual to the Cartan-Dieudonne fact that reflections generate $O (q)$ . Putting these together, the well-definedness of the Witt index and the anisotropic kernel — which 21.02.08 uses freely when it writes $V = V_{an} ⊥ H^{⊥ m}$ — appears again in the proof that the Hasse-Witt invariant is an isometry invariant, and this is exactly the cancellation step that lets the local classification reduce every form to its anisotropic core. The bridge is the recognition that "two measurements of a form are well-defined" is a cancellation statement in disguise.

Exercises Intermediate+

Exercise 4 (medium, short-answer).

State the Cartan-Dieudonne theorem and explain its relation to Witt's extension theorem.

Hint

Cartan-Dieudonne is about generating the orthogonal group; the extension proof builds the extending isometry out of the same generators.

Answer

Cartan-Dieudonne theorem. Over a field of characteristic not two, every isometry of an $n$ -dimensional non-degenerate quadratic space is a product of at most $n$ reflections $τ_{v}$ in anisotropic vectors. Relation. Witt's extension theorem builds the ambient isometry extending a subspace isometry as a product of reflections, one rank at a time, so it is a constructive companion to Cartan-Dieudonne: the extension lives inside the group that Cartan-Dieudonne says is reflection-generated. The rank-one step of the extension proof is exactly the observation that a single anisotropic vector can be moved to another of the same length by one or two reflections. Rubric: full credit for the at-most- $n$ -reflections statement and the reflection-generation link.

Exercise 5 (hard, short-answer).

Prove that the Witt index of a non-degenerate form is well-defined: any two maximal totally isotropic subspaces have the same dimension.

Hint

Use Witt's extension theorem applied to an isometry between the two isotropic subspaces, or use cancellation on the hyperbolic parts they generate.

Answer

Let $W_{1}, W_{2}$ be maximal totally isotropic subspaces, with $dim W_{1} = m_{1} \leq m_{2} = dim W_{2}$ . A totally isotropic subspace of dimension $m$ inside a non-degenerate space generates a hyperbolic subspace $H^{⊥ m}$ (each basis vector pairs with a dual partner), so $q ≅ H^{⊥ m_{i}} ⊥ q_{i}^{'}$ with $q_{i}^{'}$ non-degenerate for $i = 1, 2$ . Choose a totally isotropic subspace $W_{1}^{'} \subseteq W_{2}$ of dimension $m_{1}$ ; the inclusion-induced isometry $W_{1} \to W_{1}^{'}$ extends to $O (q)$ by Witt's extension theorem, carrying the hyperbolic envelope of $W_{1}$ onto that of $W_{1}^{'}$ . By Witt cancellation on the $H^{⊥ m_{1}}$ summand, the complements are isometric, so $q_{1}^{'} ≅ H^{⊥ (m_{2} - m_{1})} ⊥ q_{2}^{'}$ . If $m_{2} > m_{1}$ then $q_{1}^{'}$ contains a hyperbolic plane, contradicting the maximality of $W_{1}$ (its isotropic dimension could be raised). Hence $m_{1} = m_{2}$ . Rubric: full credit for the extension-plus-cancellation argument forcing equal dimension.

Exercise 6 (hard, symbolic).

Compute the Witt ring $W (F_{5})$ , identifying it as an explicit ring of order four.

Hint

For $F_{q}$ with $q \equiv 1 (mod 4)$ , every binary form is isotropic, the anisotropic forms have rank at most $1$ , and there are two square classes. Track rank modulo $2$ and discriminant.

Answer

$W (F_{5}) ≅ Z /2 [F_{5}^{\times} / (F_{5}^{\times})^{2}] ≅ Z /2 [ε] / (ε^{2} - 1)$ , order four. In $F_{5}$ the squares are ${1, 4}$ , so $F_{5}^{\times} / (F_{5}^{\times})^{2} = {1, u}$ with $u = 2$ a non-square. Every binary form over a finite field is isotropic (Chevalley-Warning, or the pigeonhole on values), hence universal, so anisotropic forms have rank $0$ or $1$ : the classes are $0$ , $⟨ 1 ⟩$ , $⟨ u ⟩$ , and $⟨ 1, u ⟩$ . Since $q \equiv 1 (mod 4)$ we have $- 1$ a square, so $⟨ 1, 1 ⟩$ is isotropic ( $= H = 0$ in $W$ ); the additive group is $(Z /2)^{2}$ and the ring is $Z /2 [ε] / (ε^{2} - 1)$ with $ε = ⟨ u ⟩$ . Rubric: full credit for order four and the ring identification with the two square classes.

Exercise 7 (hard, short-answer).

Explain how Witt cancellation recovers Sylvester's law of inertia over $R$ .

Hint

Sylvester says the signature $(p, q)$ is an invariant. Phrase it as cancellation of shared $⟨ 1 ⟩$ and $⟨ - 1 ⟩$ summands.

Answer

Over $R$ every non-degenerate form diagonalises to $⟨ 1 ⟩^{⊥ p} ⊥ ⟨ - 1 ⟩^{⊥ q}$ . Suppose two such presentations of the same form gave $(p, q)$ and $(p^{'}, q^{'})$ with $p + q = p^{'} + q^{'} = n$ . Witt cancellation lets us cancel the common $⟨ 1 ⟩$ 's and $⟨ - 1 ⟩$ 's: if $p < p^{'}$ then after cancelling $p$ positive and the matched negative summands we would obtain an isometry between a positive-definite form and an indefinite one, impossible since one is anisotropic-positive and the other has an isotropic vector. So $p = p^{'}$ and $q = q^{'}$ : the signature is well-defined. This is precisely Sylvester's law of inertia, now a corollary of cancellation rather than a separate eigenvalue argument. Rubric: full credit for casting Sylvester as a cancellation of signed summands.

Advanced results Master

Theorem (Witt decomposition; Witt 1937). Every non-degenerate quadratic space $(V, q)$ over a field $k$ of characteristic not two decomposes as $$ V ;\cong; V_{\mathrm{an}} ;\perp; \mathbb{H}^{\perp m}, $$ where $V_{an}$ is anisotropic and $m = ind (q)$ is the Witt index. The anisotropic kernel $V_{an}$ is determined up to isometry, and $m$ is determined, by $(V, q)$ .

The existence half is a descending induction: while the form is isotropic it contains a hyperbolic plane (an isotropic vector plus a non-degenerate partner), which splits off as an orthogonal summand, lowering the dimension by two; the process halts at an anisotropic remainder. The uniqueness half is exactly Witt cancellation: two decompositions $V_{an} ⊥ H^{⊥ m} ≅ V_{an}^{'} ⊥ H^{⊥ m^{'}}$ with $m \leq m^{'}$ cancel the $H^{⊥ m}$ to give $V_{an} ≅ V_{an}^{'} ⊥ H^{⊥ (m^{'} - m)}$ , and anisotropy of the left side forces $m = m^{'}$ and $V_{an} ≅ V_{an}^{'}$ .

Theorem (the Witt ring $W (k)$ ; Witt 1937). The set of isometry classes of anisotropic forms over $k$ , equivalently the set of equivalence classes of non-degenerate forms modulo hyperbolic planes, carries a commutative ring structure with addition induced by orthogonal sum $⊥$ and multiplication induced by tensor product $\otimes$ , in which the class of every hyperbolic form is the additive zero. This is the Witt ring $W (k)$ .

Two non-degenerate forms are Witt-equivalent when their anisotropic kernels are isometric, equivalently when they become isometric after adding hyperbolic planes to each. The orthogonal sum is well-defined on classes because cancellation makes $⊥$ a cancellative, hence group-completable, operation; the additive inverse of $⟨ a_{1}, \dots, a_{n} ⟩$ is $⟨ - a_{1}, \dots, - a_{n} ⟩$ , since their sum $⟨ a_{1}, - a_{1}, \dots ⟩$ is a sum of hyperbolic planes. The tensor product of forms, with $⟨ a ⟩ \otimes ⟨ b ⟩ = ⟨ ab ⟩$ , descends to a ring multiplication because $H \otimes q$ is hyperbolic for every $q$ . The augmentation $dim mod 2 : W (k) \to Z /2$ has kernel the fundamental ideal $I (k)$ , whose powers $I^{n}$ filter $W (k)$ and connect, by the Milnor conjectures (Voevodsky), to Galois cohomology and Milnor K-theory.

Theorem (computation of the basic Witt rings). The Witt rings of the real, complex, and finite fields are: $$ W(\mathbb{R}) \cong \mathbb{Z}, \qquad W(\mathbb{C}) \cong \mathbb{Z}/2, \qquad W(\mathbb{F}_q) \cong \begin{cases} \mathbb{Z}/2[\mathbb{F}_q^\times/(\mathbb{F}_q^\times)^2] & q \equiv 1 !!\pmod 4, \ \mathbb{Z}/4 & q \equiv 3 !!\pmod 4. \end{cases} $$

For $R$ the only anisotropic forms are the definite ones $⟨ 1 ⟩^{⊥ p}$ and $⟨ - 1 ⟩^{⊥ q}$ , and the Witt class is recorded by the signature $p - q \in Z$ ; the signature is the ring isomorphism $W (R) \sim Z$ , so the Witt ring over $R$ is exactly Sylvester's invariant promoted to a ring. For $C$ every form is determined by rank alone (every element is a square), so $⟨ 1, 1 ⟩$ is already a hyperbolic plane and $W (C) = {0, ⟨ 1 ⟩} ≅ Z /2$ . For finite fields the two square classes and the universality of binary forms give a ring of order four, of one of the two listed shapes according to whether $- 1$ is a square.

Synthesis. The central insight of Witt's theory is that two measurements of a quadratic form — its anisotropic kernel and its Witt index — are well-defined, and the foundational reason is cancellation, which is exactly the statement that orthogonal sum is a cancellative operation on isometry classes. Putting these together, the Witt decomposition $V ≅ V_{an} ⊥ H^{⊥ m}$ generalises Sylvester's law of inertia from $R$ to an arbitrary field, and is dual to the Cartan-Dieudonne generation of the orthogonal group by reflections: the existence of the decomposition splits hyperbolic planes off geometrically, while uniqueness is the algebraic cancellation that the reflection calculus supplies. This is exactly the structure that 21.02.08 presupposes when it writes a rational form as an anisotropic core plus hyperbolic planes and classifies it by rank, discriminant, signature, and the Hasse-Witt invariant; the bridge is that each of those invariants is constant on Witt classes, so they are functions on $W (k)$ , and the local-global theory becomes a statement about the family of localisation maps $W (Q) \to \prod_{v} W (Q_{v})$ . This pattern recurs in algebraic K-theory through the Milnor conjectures relating the filtration $I^{n} / I^{n + 1}$ to Galois cohomology $H^{n} (k, μ_{2})$ , and the central insight that a form is determined by an anisotropic core appears again in the theory of quadratic forms over schemes and the Grothendieck-Witt spectrum.

Full proof set Master

Proposition 1 (isotropic vector yields a hyperbolic plane). Let $(V, q)$ be non-degenerate over $k$ of characteristic not two, and let $v \in V$ be a non-zero isotropic vector. Then $v$ lies in a subspace isometric to $H$ .

Proof. Since $(V, q)$ is non-degenerate and $v \neq = 0$ , the linear functional $b (v, -)$ is non-zero, so there is $w_{0}$ with $b (v, w_{0}) = 1$ after scaling. Set $w = w_{0} - \frac{1}{2} q (w_{0}) v$ . Then $q (w) = q (w_{0}) - q (w_{0}) b (v, w_{0}) + \frac{1}{4} q (w_{0})^{2} q (v) = q (w_{0}) - q (w_{0}) + 0 = 0$ , using $q (v) = 0$ and $b (v, w_{0}) = 1$ , so $w$ is isotropic as well, and $b (v, w) = b (v, w_{0}) = 1$ . The plane $P = k v \oplus k w$ has Gram matrix $\begin{psmallmatrix} 0 & 1 \\ 1 & 0 \end{psmallmatrix}$ with determinant $- 1 \neq = 0$ , so $P$ is non-degenerate with two isotropic basis vectors paired to $1$ ; this is precisely the hyperbolic plane $H$ . $□$

Proposition 2 (reflections are isometries fixing a hyperplane). For anisotropic $v$ , the map $τ_{v} (x) = x - \frac{2 b ( x , v )}{q ( v )} v$ lies in $O (q)$ , fixes $v^{⊥}$ pointwise, sends $v \mapsto - v$ , and satisfies $τ_{v}^{2} = id$ .

Proof. Write $c = \frac{2 b ( x , v )}{q ( v )}$ . Then $$ q(\tau_v x) = q(x) - 2 c, b(x, v) + c^2 q(v) = q(x) - 2 c, b(x, v) + \frac{4 b(x, v)^2}{q(v)} = q(x), $$ since $c^{2} q (v) = \frac{4 b ( x , v ) ^{2}}{q ( v ) ^{2}} q (v) = \frac{4 b ( x , v ) ^{2}}{q ( v )}$ cancels the middle term $2 c b (x, v) = \frac{4 b ( x , v ) ^{2}}{q ( v )}$ . So $τ_{v}$ preserves $q$ . If $b (x, v) = 0$ then $τ_{v} (x) = x$ , so $v^{⊥}$ is fixed; and $τ_{v} (v) = v - \frac{2 q ( v )}{q ( v )} v = - v$ . A direct substitution gives $τ_{v}^{2} = id$ . $□$

Proposition 3 (Witt cancellation in the rank-one case). If $⟨ a ⟩ ⊥ q_{1} ≅ ⟨ a ⟩ ⊥ q_{2}$ with $a \in k^{\times}$ and $q_{1}, q_{2}$ non-degenerate, then $q_{1} ≅ q_{2}$ .

Proof. Let $φ$ be the isometry, and let $u_{1}, u_{2}$ be the vectors spanning the two copies of $⟨ a ⟩$ in source and target, so $q (u_{1}) = q (u_{2}) = a$ and $φ (u_{1}) =: u_{1}^{'}$ has $q (u_{1}^{'}) = a$ . By the rank-one extension argument, one of $u_{1}^{'} - u_{2}$ and $u_{1}^{'} + u_{2}$ is anisotropic, and a product $ρ$ of at most two reflections in the target sends $u_{1}^{'} \mapsto u_{2}$ . Then $ρ \circ φ$ is an isometry of the ambient spaces sending the source copy of $⟨ a ⟩$ exactly onto the target copy spanned by $u_{2}$ . An isometry matching the $⟨ a ⟩$ summands carries orthogonal complements to orthogonal complements, so it restricts to an isometry $q_{1} ≅ q_{2}$ . $□$

Proposition 4 (Witt decomposition is unique). If $V_{an} ⊥ H^{⊥ m} ≅ V_{an}^{'} ⊥ H^{⊥ m^{'}}$ with both kernels anisotropic, then $m = m^{'}$ and $V_{an} ≅ V_{an}^{'}$ .

Proof. Without loss of generality $m \leq m^{'}$ . Iterating Witt cancellation on the $m$ shared hyperbolic planes (each a $⟨ 1, - 1 ⟩$ , cancelled one anisotropic summand at a time via Proposition 3) yields $V_{an} ≅ V_{an}^{'} ⊥ H^{⊥ (m^{'} - m)}$ . If $m^{'} > m$ , the right-hand side contains the isotropic vectors of a hyperbolic plane, so it is isotropic; but the left-hand side $V_{an}$ is anisotropic, a contradiction. Hence $m^{'} = m$ and the leftover hyperbolic part is empty, giving $V_{an} ≅ V_{an}^{'}$ . $□$

Proposition 5 (additive inverses in $W (k)$ ). In the Witt ring, $- [⟨ a_{1}, \dots, a_{n} ⟩] = [⟨ - a_{1}, \dots, - a_{n} ⟩]$ .

Proof. It suffices to show $⟨ a ⟩ ⊥ ⟨ - a ⟩$ is Witt-zero for each $a \in k^{\times}$ , since orthogonal sum is additive on classes. The form $⟨ a, - a ⟩$ is $a x^{2} - a y^{2}$ , which vanishes at $(1, 1) \neq = 0$ and is non-degenerate, so by Proposition 1 it is a hyperbolic plane, hence zero in $W (k)$ . Summing over the coordinates, $⟨ a_{1}, \dots, a_{n} ⟩ ⊥ ⟨ - a_{1}, \dots, - a_{n} ⟩ ≅ H^{⊥ n}$ is Witt-zero, so the second class is the additive inverse of the first. $□$

Proposition 6 ( $W (R) ≅ Z$ via signature). The signature map $⟨ 1 ⟩^{⊥ p} ⊥ ⟨ - 1 ⟩^{⊥ q} \mapsto p - q$ is a well-defined ring isomorphism $W (R) \to Z$ .

Proof. Over $R$ every element of $R^{\times}$ is $\pm$ a square, so every form diagonalises to $⟨ 1 ⟩^{⊥ p} ⊥ ⟨ - 1 ⟩^{⊥ q}$ , and $⟨ 1, - 1 ⟩ = H$ is Witt-zero, so the class depends only on $p - q$ ; the anisotropic kernel is $⟨ 1 ⟩^{⊥ (p - q)}$ when $p \geq q$ and $⟨ - 1 ⟩^{⊥ (q - p)}$ otherwise. The map is additive since signatures add under $⊥$ , and multiplicative since $⟨ \pm 1 ⟩ \otimes ⟨ \pm 1 ⟩ = ⟨(\pm 1) (\pm 1)⟩$ multiplies signs, matching multiplication in $Z$ . It is surjective ( $p - q$ ranges over all of $Z$ ) and injective (equal signatures give isometric anisotropic kernels by Witt's uniqueness), hence a ring isomorphism. $□$

Connections Master

Bilinear and quadratic forms 01.01.15. This unit is the arithmetic continuation of the foundational linear-algebra treatment, which stops at diagonalisation, discriminant, signature, and Sylvester's law of inertia over $R$ . Witt's theory promotes those invariants to an arbitrary field of characteristic not two: Sylvester's law becomes the special case $W (R) ≅ Z$ of the Witt-ring computation, and the signature reappears as the ring isomorphism. The cancellation theorem is the general-field reason the signature is well-defined, replacing the eigenvalue continuity argument used over $R$ .
Hilbert symbol and the product formula 21.02.05. The Hilbert symbol $(a, b)_{v}$ records whether the ternary form $⟨ 1, - a, - b ⟩$ is isotropic over $Q_{v}$ , which by the Witt decomposition is the question of whether that form has a hyperbolic plane splitting off. The Hasse-Witt invariant $ε_{v} (f) = \prod_{i < j} (a_{i}, a_{j})_{v}$ is constant on Witt classes precisely because cancellation makes it well-defined, so the symbol calculus of 21.02.05 is a calculation inside the local Witt rings $W (Q_{v})$ .
Hasse-Minkowski theorem 21.02.08. Hasse-Minkowski writes every rational form as an anisotropic kernel orthogonal to hyperbolic planes and classifies it by rank, discriminant, signature, and Hasse-Witt invariant — every step of which presupposes the Witt decomposition and cancellation proved here. The local-global classification is the injectivity of $W (Q) \to \prod_{v} W (Q_{v})$ , a statement about the Witt rings of $Q$ and its completions; this unit supplies the rings, and 21.02.08 supplies the reciprocity constraint cutting out the image.
Theta series of quadratic forms 21.04.04. The representation numbers $r_{Q} (n)$ counted by a theta series depend only on the isometry class of the positive-definite integral form $Q$ , and the Witt-theoretic invariants (rank, discriminant, and the anisotropic kernel of $Q \otimes Q_{v}$ at each place) are the genus data that organise forms with the same theta behaviour. The co-produced lateral unit on theta series builds on the structure theory established here, using the Witt decomposition over $Q_{p}$ to read off local representability.

Historical & philosophical context Master

Ernst Witt's 1937 paper Theorie der quadratischen Formen in beliebigen Körpern (J. reine angew. Math. 176, 31-44) ^{[Witt 1937]} reshaped a subject that had been tied to the real and rational fields into a structural theory over any field. Before Witt, the classification of quadratic forms proceeded field by field: Sylvester's 1852 law of inertia handled $R$ , and the Minkowski-Hasse local-global apparatus handled $Q$ , but each rested on field-specific computations. Witt's insight was to isolate the purely formal content — that orthogonal sum admits cancellation, that an isometry of subspaces extends to the whole space, and that every form decomposes into an anisotropic core and a controlled number of hyperbolic planes — and to show that this content holds over an arbitrary field of characteristic not two. The resulting Witt ring $W (k)$ turned the classification problem into the study of a single commutative ring functorial in the field, an early and influential instance of the structural turn that would define twentieth-century algebra.

The geometric engine, the generation of the orthogonal group by reflections, traces to Cartan's 1938 Leçons sur la théorie des spineurs ^{[Cartan 1938]} and was put in general-field form by Dieudonne's 1948 Sur les groupes classiques ^{[Dieudonné 1948]}, giving the Cartan-Dieudonne theorem that every isometry is a product of at most $dim V$ reflections. Serre's 1973 A Course in Arithmetic (Ch. IV) made Witt's package the standard prelude to the Hasse-Minkowski theorem, and Milnor and Husemoller's 1973 Symmetric Bilinear Forms connected the fundamental-ideal filtration $I^{n} (k)$ of the Witt ring to algebraic K-theory and Galois cohomology, a thread that culminated in Voevodsky's proof of the Milnor conjecture. Philosophically, Witt's theorem is a paradigm of how isolating the right invariants — here the anisotropic kernel and the Witt index — converts a tangle of case-by-case computations into a clean statement of well-definedness, and how a cancellation law silently underwrites the very possibility of speaking about "the" core of a structure.

Bibliography Master

@article{Witt1937,
  author  = {Witt, Ernst},
  title   = {Theorie der quadratischen Formen in beliebigen K{\"o}rpern},
  journal = {Journal f{\"u}r die reine und angewandte Mathematik},
  volume  = {176},
  year    = {1937},
  pages   = {31--44}
}

@book{Serre1973,
  author    = {Serre, Jean-Pierre},
  title     = {A Course in Arithmetic},
  publisher = {Springer-Verlag},
  series    = {Graduate Texts in Mathematics},
  volume    = {7},
  year      = {1973},
  note      = {English translation of \emph{Cours d'arithm{\'e}tique}, Presses Universitaires de France, 1970}
}

@book{Lam2005,
  author    = {Lam, T. Y.},
  title     = {Introduction to Quadratic Forms over Fields},
  publisher = {American Mathematical Society},
  series    = {Graduate Studies in Mathematics},
  volume    = {67},
  year      = {2005}
}

@book{Scharlau1985,
  author    = {Scharlau, Winfried},
  title     = {Quadratic and Hermitian Forms},
  publisher = {Springer-Verlag},
  series    = {Grundlehren der mathematischen Wissenschaften},
  volume    = {270},
  year      = {1985}
}

@book{MilnorHusemoller1973,
  author    = {Milnor, John and Husemoller, Dale},
  title     = {Symmetric Bilinear Forms},
  publisher = {Springer-Verlag},
  series    = {Ergebnisse der Mathematik und ihrer Grenzgebiete},
  volume    = {73},
  year      = {1973}
}

@book{Cartan1938,
  author    = {Cartan, {\'E}lie},
  title     = {Le{\c c}ons sur la th{\'e}orie des spineurs},
  publisher = {Hermann},
  address   = {Paris},
  year      = {1938}
}

@book{Dieudonne1948,
  author    = {Dieudonn{\'e}, Jean},
  title     = {Sur les groupes classiques},
  publisher = {Hermann},
  address   = {Paris},
  year      = {1948}
}

Prerequisites

01.01.15
01.01.01
21.02.01

Tier anchors

beginner: Serre 1973 *A Course in Arithmetic* (Springer GTM 7) Ch. IV §1 informal opening; Lam 2005 *Introduction to Quadratic Forms over Fields* (AMS GSM 67) Ch. I §1 informal opening
intermediate: Serre 1973 *A Course in Arithmetic* Ch. IV §§1-2; Lam 2005 *Introduction to Quadratic Forms over Fields* Ch. I-II; Scharlau 1985 *Quadratic and Hermitian Forms* (Springer Grundlehren 270) Ch. 1
master: Witt 1937 *J. reine angew. Math.* 176, 31-44 (originator: cancellation, the extension theorem, and the Witt ring); Serre 1973 *A Course in Arithmetic* Ch. IV §§1-2; Lam 2005 *Introduction to Quadratic Forms over Fields* (AMS GSM 67) Ch. I-II (modern textbook treatment of the Witt ring and Witt decomposition); Scharlau 1985 *Quadratic and Hermitian Forms* (Springer Grundlehren 270) Ch. 1-2; Milnor-Husemoller 1973 *Symmetric Bilinear Forms* (Springer Ergebnisse 73)

References

Witt, E. — Theorie der quadratischen Formen in beliebigen Körpern · *Journal für die reine und angewandte Mathematik* 176 (1937), 31-44. The originator paper. Introduces the extension theorem (an isometry of subspaces extends to the whole space), the cancellation theorem, the decomposition into an anisotropic kernel plus hyperbolic planes, and the ring of equivalence classes now called the Witt ring.
Serre, J.-P. — A Course in Arithmetic · Springer Graduate Texts in Mathematics 7 (1973), translation of *Cours d'arithmétique* (PUF 1970). Ch. IV §1 (quadratic forms over a field, orthogonality, isotropy, the hyperbolic plane); §2 (Witt's theorem, Witt decomposition, the Witt group).
Lam, T. Y. — Introduction to Quadratic Forms over Fields · American Mathematical Society Graduate Studies in Mathematics 67 (2005). Ch. I (foundations, diagonalisation, hyperbolic spaces), Ch. II (Witt's chain-equivalence theorem, Witt cancellation, the Witt decomposition, the Witt ring $W(k)$ and Witt-Grothendieck ring $\widehat{W}(k)$), with the computations of $W(\mathbb{R})$, $W(\mathbb{C})$, and $W(\mathbb{F}_q)$.
Scharlau, W. — Quadratic and Hermitian Forms · Springer Grundlehren der mathematischen Wissenschaften 270 (1985). Ch. 1-2 give a careful modern account of the orthogonal group, the Cartan-Dieudonne theorem, Witt's theorems, and the fundamental ideal filtration of $W(k)$.
Milnor, J. and Husemoller, D. — Symmetric Bilinear Forms · Springer Ergebnisse der Mathematik 73 (1973). A concise account of symmetric bilinear forms over fields and rings, the Witt ring, and its connection to algebraic K-theory and the Milnor conjectures.
Cartan, É. — Leçons sur la théorie des spineurs · Hermann, Paris (1938). Develops the structure of the orthogonal group via reflections; the Cartan-Dieudonne generation of $O(q)$ by reflections is the geometric engine behind Witt's extension theorem.
Dieudonné, J. — Sur les groupes classiques · Hermann, Paris (1948). The general-field version of the theorem that every isometry of a non-degenerate quadratic space is a product of at most $\dim V$ reflections, including the characteristic-two caveats.

Estimated time

beginner: 16m
intermediate: 42m
master: 85m