Max-Albert Knus Quadratic and Hermitian Forms over Rings "| Springer-Verlag , Berlin Heidelberg New York /| London Paris Tokyo HongKong BarcelonaTable of Contents Chapter I. Hermitian Forms over Rings... . . . Leb ce ete ee eee 1 §1. Rings with Involution..............- -. 1 §2. Sesquilinear and Hermitian Forms. . . . 5 §3. Hermitian Modules . ul §4. Symplectic Spaces . . 2 §5. Unitary Rings and Modules . : 22 §6. Hermitian Spaces over Division Rings . 28 §7. Change of Rings............-0. . 36 §8. Products of Hermitian Forms... ... . 4 §9. Morita Theory for Hermitian Modules . 51 §10. Witt Groups . : 59 $11. Cartesian Diagrams and Patching of Hermitian Forms. . . 64 Chapter II. Forms in Categories. n §1. Additive Categories n §2. Categories with Duality . 5 §3. Transfer. 81 §4. Reduction 85 §5. The Theorem of Krull-Schmidt for Additive Categories 93 §6. The Krull-Schmidt Theorem for Hermitian Spaces . 96 §7. Some Applications. . . . . . : 101 Chapter III. Descent Theory and Cohomology... ....-0.0+0++ 106 §1. Descent of Elements. §2. Descent of Modules and — §3. Discriminant Modules . . . §4. Quadratic Algebras . §5. Azumaya Algebras. . . §6. Graded Algebras and Modules . §7. Universal Norms ...... 2... §8. Involutions on Azumaya Algebras §9. The Pfaffianx Contents Chapter IV. The Clifford Algebra... 2... e ee ee eee 193 §1. Construction of the Clifford Algebra... 2... . : . 193 §2. Structure of the Clifford Algebra, the Even Rank Case 199 §3. Structure of the Clifford Algebra, the Odd Rank Case . - 207 §4. The Discriminant and the Arf Invariant . . 21 §5. The Special Orthogonal Group... . . 2B §6. The Spinors.......-...- . .227 §7. Canonical Isomorphisms . . 236 §8. Invariants of Quadratic Spaces we . 242 §9. Quadratic Spaces with Trivial Arf Invariant _. 247 Chapter V. Forms of Low Rank....... 2220-0000 eee eens 251 §1. Quadratic Modules of Rank 1 251 §2. Quadratic Modules of Rank 2. . «252 §3. Quadratic Modules of Rank 3. . . 260 §4. Quadratic Modules of Rank 4. . 264 §5. Quadratic Spaces of Rank 5 and 6... +. 295 §6. Hermitian Modules of Low Rank. . . . - - 300 §7. Composition of Quadratic Spaces. . . . . 305 Chapter VL. Splitting and Cancellation Theorems .........-....- 312 §1. Semilocal Rings, the Stable Range . . . . 312 §2. ThefRank................04 cece eee eee 322 §3. Serre’s Splitting Theorem and Cancellation . 327 6 UnmnyGop .. ..... . 335 §5. Cancellation for Unitary Spaces over Semilocal Rings . . - - 360 §6. Cancellation and Stability for Unitary Spaces 3/1 §7. A Splitting Theorem .. 1.2.0... ee eee ee eee 379 Chapter VII. Polynomial Rings §1. Principal Ideal Domains §2. Polynomial Rings . §3. Bundles over PS. . §4. The Theorem of Karoubi §5. Quillen’s Theorem. §6. A Rigidity Theorem and the Horrocks Theorem . §7. Isotropic Hermitian Spaces 425 §8 Projective Modules over Polynomial Rings . . 429 §9. Hermitian Spaces of Low Rank... . . - 440 §10. Indecomposable Anisotropic Spaces . . . 448 §11. Hermitian Modules over Projective Spaces... 453Contents xi Chapter VIII. Witt Groups of Affine Rings §1. Witt Group of Schemes. . §2. Domains of Dimension <3. §3. Regular Local Rings Essentially of Finite Type §4. Real Smooth Surfaces... . . §5. Real Curves. . §6. Examples . . . §7. Symplectic Bundles over Affine Surfaces.Foreword From its birth (in Babylon?) till 1936 the theory of quadratic forms dealt almost exclusively with forms over the real field, the complex field or the ring of integers. Only as late as 1937 were the foundations of a theory over an arbitrary field laid. This was in a famous paper by Ernst Witt. Still too early, apparently, because it took another 25 years for the ideas of Witt to be pursued, notably by Albrecht Pfister, and expanded into a full branch of algebra. Around 1960 the development of algebraic topology and algebraic K-theory led to the study of quadratic forms over commutative rings and hermitian forms over rings with involutions. Not surprisingly, in this more general setting, algebraic K-theory plays the role that linear algebra plays in the case of fields. This book exposes the theory of quadratic and hermitian forms over rings in a very general setting. It avoids, as far as possible, any restriction on the characteristic and takes full advantage of the functorial aspects of the theory. The advantage of doing so is not only aesthetical: on the one hand, some classical proofs gain in simplicity and transparency, the most notable examples being the results on low-dimensional spinor groups; on the other hand new results are obtained, which went unnoticed even for fields, as in the case of involutions on 16-dimensional central simple algebras. The first chapter gives an introduction to the basic definitions and properties of hermitian forms which are used throughout the book. Chapter IT distills the categorical aspects of hermitian forms. The beginning is rather formal but towards the end the reader is rewarded by very explicit and concrete results about hermitian forms on projective varieties. This, by the way, is one of the two single places where the author departs from affine schemes. Chapter III recalls the theory of descent and introduces the cohomological tools that are needed in the subsequent chapters. A good portion of it is dedicated to Azumaya algebras, thus paving the way for the study of Clifford algebras in Chapter IV. In the latter, following, I believe, an idea of Martin Kneser, semiregular forms are introduced, to replace regular forms of odd rank when 2 is not invertible. Clifford algebras are used in Chapter V to study forms of rank up to 6. An interesting tool here is the so called reduced pfaffian, which is to forms of rank 5 or 6, what the reduced norm of a quaternion algebra is to forms of rank 3 or 4.vi Foreword The next two chapters, VI and VII, have a distinctive K-theoretic flavour. Chapter VI proves stability- and cancellation theorems in the linear case and the analogous results in the unitary case. It contains, in particular, a complete, detailed and (most probably) correct proof, due to Maria Saliani, of injective stability for the unitary group. The results of Chapter VI are applied to polynomial rings in Chapter VII, which studies the quadratic analogues of the theorem of Quillen and Suslin on projective modules. The last chapter contains results on the Witt group of regular rings and some amusing computations of Witt groups of curves and surfaces. Authors are frequently accused by nasty reviewers of inadmissible omissions, so let me indulge for a while in this pleasure. Where did the author leave L-theory? Doesn't he think that arithmetic also deserves a little place here and there? Is he, by any chance, totally unaware of all the recent results about quadratic bundles over projective varieties? And what about lattices, packings and codes? Well, this is like bemoaning the absence of canard 4 orange in a vegetarian re- staurant. Human life and endurance being limited, the author has chosen to stress the algebraic aspects of the theory and to avoid—within reason—the overlapping with other remarkable books on quadratic forms, like those of Lam, Milnor-Husemoller and Scharlau. Within these limits his book gives a clear, detailed and complete picture of the theory of quadratic and hermitian forms. Finally, two words about prerequisites. Although the level of the exposition is variable and tends to increase towards the last chapters, unproved theorems are always quoted, unless they are basic results in commutative algebra or homo- logical algebra. Occasionally, some familiarity with algebraic geometry may be helpful, as in the proof of Horrocks’ theorem, without being indispensable. As I said before, algebraic K-theory is both a tool and a source of inspiration, hence a previous knowledge of its more elementary results will allow the reader to proceed faster through Chapter VI and VII. Summer 1990 I. BertuccioniForeword vii Acknowledgements This volume would not have been written without the help of my friend Manuel Ojanguren. Even if | am responsible for the final draft, his influence is ubiquitous and I hardly can give him sufficient credit for his ideas and advice. Many results presented in this book were developed over years by (or in collaboration with) R. Parimala, R. Sridharan and M. Ojanguren. I am deeply grateful to them for many wonderful hours doing mathematics in Bombay, Lausanne or Zurich. Special thanks go to my former teacher Beno Eckmann for his continued encouragement, to I. Bertuccioni, who offered to write the foreword, and to Springer-Verlag for publishing the volume. I enjoyed the help of many other friends and colleagues, to mention only some: E. Bayer-Fluckiger, M. Kneser, M. Kolster, A. Paques, H. G. Quebbemann, C. Richm, W. Scharlau and C. B. Thomas. Further, I received assistance from former students: C. Arlettaz-Brack, W. Bichsel, B. Keller, T. Mittelholzer, Y. Pitteloup, M. Saliani, N. Spriano, M. Schneider. Mrs. D. Spérri typed the first version, Mrs. P. Keller read the final version and corrected some of my worst “french” english. It is a pleasure to thank all of them. Conventions A ring will always be associative with an identity element 1. Modules will be right modules and maps are written on the left like functions. The composition of two maps f, gis denoted by fg (first g, then f). The commutative base ring of an algebra will be usually denoted by R and unadorned tensor products are taken over R. The ring of (n x n-matrices over an algebra A is denoted by M,(A) and the group of units of A by A’. Fall 1990 M.-A. KnusChapter I. Hermitian Forms over Rings ‘This first chapter gives the basic theory of hermitian and quadratic forms. Once the general formalism is introduced, we restrict to forms over finitely genetated projective modules, In characteristic 2 it is convenient to view a quadratic form as the class of a bilinear form modulo alternating forms. This idea, which already occurs in Klingeuberg-Witt [1954], is used by Bak [1969] to define a very general notion of forms, Bak considers equivalence classes of sesquilinear forms depending on two parameters ¢ and A, We introduce these classes in section 5, following closely the presentation of Bass [1973]. Bass calls the classes (¢, A}-quadratic spaces. We propose to change the terminology to (¢, A}unitary spaces. ‘The first five soctions are of general level. Thea, as a first application, we study forms over division rings. In the rest of the chapter, we develop different tools which are used later: change of rings, products of forms, Morita theory, patching diagrams and the Witt group. § 1. Rings with Involution (1.1) An involution on a ring A is an antiautomorphism o:at+a of A of order < 2. Thus: a+b=a+5b, ab=ba, I=1 and G=a forabeaA. Let R be a commutative ring with involution ¢,. A R-algebra with involution is an R-algebra A with an izivolution o, that extends the involution of R: o,(ra) = o4(")3,(a) for ac A,reR. If A ig a ring with involution ¢ and Z = Z(A) = {© Alxy — yx Vye A} is the centre of A, then ¢ gives by restriction an involution on Z and Ais a Z-algebra with involution. In most cases gg will be the identity (which is an myvolution since R is commutative), We denote by Fix(e,) the set of elements fixed by o,. Let (A, 9,) and (B, og) be rings with involution. A morphism of rings with involution @: (A, 6,)~»(B, 6) is a homomorphism of rings g: A» B such that 65(¢(@)) = g(a,(a)), a€ A. For example, if J < A is two-sided ideal such that o,(1) Xq — %i— x2 J ~ xpi, ER, 1-0, 3, is an involution of IH such that tr(x) =x + ¥ = 2xpeR and a(x) = xt = }) x2 eR. The real number tr (x) is reo called the reduced trace of x and n(x) is called the reduced norm of x. As we shail see in (1.3.4) the conjugation map x++¥ on HH is the unique involution o of Hi such that Rc Fix(o), x + o(x)eR and xe(x)eR for ail xe lH. Notice that has many involutions ¢ with Fix (o) > R. Let xt+ ¥, cH be as above. Then any map xr+ute™', ua unit of HH with @= tu, is such an involution. Conversely, any involution « of H with Fix(e)=R can he expressed as a(x) = uxu" for # = +4 {sce IL, (8.1), (1.24) Let G bea finite group and RX a commutative ring with involution x+-+ X, xe R. The map G -+ G given by g++ g' extends to an mvolution o of the group ting RG: o{La,g) = Lag" (1.25) Let M,(R) be the ring of nx mmatrices over a commutative ting R. The map ++ x’ which associates to any matrix x = (a,j) its transpose x! = (@) = 5) is an involution of M,(R). (1.26) Let R be a commutative ring with involution og. The teusor product A® ,H of the R-algebras with involution (4, 4,) and (B, 6) is an algebra with involution ¢, © ¢,. For example any matrix ring M,(A), where 4 is a ring with inyolution a, , hag the inyolution o: (a,)+ (o,{@,)}, Which corresponds to the tensor product of the inyolution (1.2.5) on M,(Z) with the involution ¢, on A. (1.2.1) For any ring A, let A°” be the opposite ring, i. the same additive group with reversed multiplication: ab? = (bay, where a stands for a as an element of 4°*. Involutions on A arc isomorphismsF, Rings with Involuion 3 g: A» A® of rings such that 9? = 14. For any ring A, the product A x 4? has an involution {a B?)i+ (5, a?) . This product is called the hyperbolic ring of A and is noted H(A). We observe that H is a functor from the category of rings to the category of rings- with-involution which is right adjoint to the forgetful functor. (1.28) Let (4,¢) be a semisimpie ring with involution. We can decompose Aasa product A, x ,.. x A,x B, x... x B,, where the A, are simple and o(4;)= 4, and the B, are products of two simple algebras 4) x Aj with o(4i) = AJ. The rings B, are isomorphic to hyperbolic rings, B, x H(A‘) sece (1.2.7). We call a ring with involution which is either simple or of the form H(A) with A simple, a simple ring-with-invotution. (1.3) Standard Involutions, Let R be a commutative ring. For any R-algebra A with involution «, we define the wace tr and the norm 1 with respect to the involution as trZ)=xto%) nx)=x0(x), xed. We say that ¢ is a standard involution if 4) R is fixed uader ¢, ie. Fix(o) > R 2) tr@eR and nQje R for all xe A. Since . nd + x)=nlx) +1 +t, we see that (xjeR for all xeA follows from n(xjeR for all xe A, More generally, we have 3.) nx + ¥)— A(x) + a() + Hxo(y)) x, YEA- The notin is multiplicative for a standard involution: 3.3 AEKY) = XVI (XY) = XVO( YOR) — xe (X)yo(V) = AK)AL) - Let ¢ be an inyoiution of A. To check that g is a staadard involution, it suffices in view of (1.3.1) and (1.3.2) to verify 1) and 2) for a set of generators of the algebra A. Any element xe A satisfies the quadratic equation (13.3) x? —tr)xta@y=0. It follows from (1.3.3) that n(x) = xe(x) = a(x)x for all xe A. The conjugation maps on the R-algebras € and H are standard involutions, Before giving more examples, we prove the following uniqueness property. (1.3.4) Proposition. Let A be an R-algebra which is a finitely generated projec- twe and faithful R-module. Then there exists at most one standard invalution on A.4 1. Hermitian Formas over Rings Proof, Localizing, we may assume that 4 is a finitely generated free R-module. By the next lemma, we have a hasis {e,,..., ¢,} of A with e, — I. Now let 6, and a, be two standard involutions of A. It follows from (1.3.3) that tr, (¢,)e; = trale,)e; (and m, (e;) = nz{e,)) for all i = 2,. ..,», where tr, is the trace of g, and n, the nonin of g,. Since tr, (e, ) = 7, (¢,), We have tr, (¢,) = tr(¢,) for ail #=1,2,..., m thus 6,(¢,) =42€,), ey a (13.5) Lemma. Let A be an R-algebra. If A is finitely generated projective and faithful as an R-module, then R-1, is an R-direct summand of A. Proof. The unit map &: R +A, e(x) = x-1, xR, has a section if and only if Hom (¢, 1): Hom,(4, R) + Hom,(R, R) is surjective. But Hom (g, 1) is surjective modulo any maximal ideal m of R because A, has positive rank. Hence Hom ¢, {) is surjective. a (136) Quadratic Algebras. We say that an R-algebra S is quadratic if S is a projective R-module of rank 2. If 2 is a local ring, then by (1.3.5), S has a basis {1,2} and 2? =az +6, abeR. It follows that a quadratic algebra is always commitative. We define an involution on S by o(2) -a—z or a(x) =a + ua — wz ifx - A+ yz, 4, weR, We have tr(x) =x + 0%) = 24+ uoeR (x) = xo(x) = 4? - Apa — pbER for the corresponding trace and norm. This shows that any free quadratic Realgebra has a standard involution. In particular all the localizations S,, peSpec(R) of a quadratic R-algebra S have standard involutions o(p). For any peSpec(R), there is feR\p such that the localization S, is free over Ry and has a standard iyolution o(f). Let now {U,=SpecRy,, i-1,...,r} be a covering of Spec (R), where the f, are as above. By the uniqueness of standard involutions, the o(j;) can be glued together to give a standard involution for S. Thus any quadratic algebra has a standard involution. Another way to con- struct the standard mvolution of a quadratic algebra is to usc the existence of a reduced trace 1 on End a(S) (see IF, (1.2)) and to define tr (x) = t(p(x)), xe 5, where p: S -+End,(S) is the regular representation. An important example of a quadratic algebra is the product R x 2 (with the diagonal action of R). The standard involution is given by (x, y)r+(y, x). (1.3.7) Quaternion Algebras, We say that an R-algebra A is a quaternion algebra it 1) A is a projective R-module of rank 4, 2) A has a standard involution. ‘There are two important classes of quaternion algebras: crossed products and twisted forms of matrix algebras. As we shail see later these classes are strongly2 Sesquilinear and Hermitian Forms 5 related. First we describe the crossed products. Let S be a quadratic R-algebra with standard involution a, and let 4¢R. We define 4 multiplication on the free S-module of rank 2 A=S@us by u? = 4, su — uos(s}, seS and the rule that the multiplication extends the multiplication of S. Then 4 is an associative R-algebra which is projective of rank 4. We have an involution ax + uyrrosx)—uy, x,yeS with te g(t + uy) = X + o9{x} = ts (xX) als + uy) — Xas(x) — Ayas(y) = Mts(x) — dns) for the corresponding trace and norm. Thus a, is a standard involution. We denote the algebra A by (4,5/R]. For example (—1, €/R] is the usual real quaternion algebra. Next we consider twisted forms of matrix algebras. The algebra M.(R) has a standefd involution given by ab an d -b cd ~e a}” ‘The corresponding trace is the usual trace of a matrix and the nonin is the determinant. We observe that My{R)~ (LR x R/R) a _{R 0 O14 taking S = Rx x = (4 nya -(4 a): We say that an R-algebra A is a twisted form of M,(R} if there exist a faithfully fat R-algebra T and an isomorphism a: A@ 15 M,(T) of T- algebras. It follows by descent theory (see III, §1) that the standard involution of M,,(R) induces a standard javolution on 4, so that 4 is 4 quaternion algebra. We have tr4(x) = tr(a(x®1)) and n(x) = det (a(x @1)) {Compare with TM, (1.2)) In view of V, (7.1.6), quadratic algebras and quaternion algebras are essentially all the associative R-algebras which are firite generated projective R-modules and have a standard involution. $2. Sesquilinear and Hermitian Forms Q.1) Let A be a ring with an involution denoted by ar+d. Any left A-modnle M can be conyerted to a right A-module Mf by the rule ma= am, meM6 1. Hermitian Forms over Rings More precisely, we define M = M as additive group and we let A act on M on the right as above. For example, the involution of 4 is the isomorphism of right A-modules A -> A given by L++1. We call M the opposite of M. For any right A-module M, the dual MY = Hom,(M, A) is a left A-module through the operation (af}(m) = afm), fe M” . Similarly, M” is a right A-module if M is a Jeft A-module. Let Dtob-A be the category of right 4-nodules and let 4-Mod be the category of left 4-nodules, We have two functors ~ and * from Dtob-A to A-Mod (or 4-Mod to Wod- A). These functors commute: (24.1) Lemma. Let M be a right A-module. The map STS): ries fom, for m =m as element of M and fe M”, defines an isomorphism of A-modules Ty: MY 3, which gives @ natural isomorphism of functors “= Proof. We first verify that Ty(f)eM”. In fact, (Ty, f (am) — f (md) = fem). {aTf ea). The map Ty is Adinear: Ty {fa)(em) = Tyas (re) = af gn) S(n)a = Tyg fa (i). The inverse imap Ty" is given by Tix'(g)¢n) = g(m) 01 The module M* = M” is aright 4-nodule for any right A-module M. We have a contravariant functor *: Dtod-A — Mob-A (or A-Mobd -+ A-Wod) if we put o* = @", where @’: NY -+ MY is the transpose of @: MN, ie. 0” (f) = fo, SEN’. Let M’” be the douhic dual of M. In view of 2.1.1), there is a canonical isomorphisin Ty MYY = (MM) = Mi given by m’*r+(frem’*(f), feM’ = M*. Composing Ty with the canonical map M—>M**, mism’”, mv” (f) = fem), we obtain a natural transforination (2.1.2) can: M+ M** We haye cany(m)—m**, m*(f)—f(m). For any A-homomorphism g: M +N, we check that g**can,, = cany@, Thus the diagram M—~3N a) ~~. | I~ wee 2 Now counnutes. Further, the transformation “can” satisfies the following identity: {Cat yg) (CAM ig) = Lae -2. Sesquiinear and Herminan Forms 7 That is, the diagram Mt = M* (2.14) cays J Mone is commutative. Let feM*, xeM** and yeM***, We have cane(f)Q)= £2") = x(f)and (cany)*(¥) = yoany. Thus (cany,)*canye( f}(m) — m**(f)— ‘flen) as claimed. (2.2) Let M be a right A-module, A sesquilimear form on M js a biadditive map b:Mx MA such that (xa, ya’) = ab(x, y)a’ for x, ye M and a, a’e A. The form 6 induces a map fyi M—oM*, iy@)Q)- by), SEM - which is A-linear: h,(xa)(y) = b(xa, y) = ab(x, y} = Ghy(x)(y) = (h,(x)2)(y). Conversely an A-linear homomorphisin 4: M + M* defines a sesquilinear forun AX Y= AE)O, x, ye. We call h, the adjoint of b. Since b,, => and h,, = 4, 2 sesquilinear form deterinined by its adjoint and con versely. Later we shaill frequently use the same notation for a sesquilinear forin or for its adjoint. Let Z be the centre of A. The set Sesq,,(M) of sesquilinear forms on M is a Zanodule for the addition b, +b, of fortis and the scalar multiplication (£0) 9) = €b(x, 9), €Z. The map br-+A, is an isoworphisin of Z-modules (22:1) Sesq,(M) % Hon, (M, M*), if we define the action of Z on Hom, (M, M*) by (£9)(x) = (22). Ke M, £EZ. For any homomorphism g: M, ~»M; of A-modules, we have a Z-linear map Sesq(g): Sesq,(M)—+ Sesq (Mi) given by Sesq(o)(0)(x, y) = 2(@(), 9(y), x,y © M,. The corresponding inap Hom ,(M,, MF) + Hom,(M,, MP) for the adjoints is given by h is pth. Let b, ¢ Sesq,(M,) and bz € Sesq,(M;). We say that geHom,(M,,Mz) is a marphism (M,,b,)+(M2,bz) if Sesq(p){bz) = bs, ke. 529 (x), PO) = bi, y) for all x, ye M, Tf, further, g is an isomorphism, we call @ an isometry and say that the two forins are isometric. The isometnes of (M, 6) form a group with respect to composition8 —_L Hermutian Forms over Rings called the orthogonal or unitary group of (M,»). This group is denoted by U0, b) 2.3) For any sesquilinear form b on M, we define a sesquilinear form b* on M by bx yy=bo,x, xem. The map b ++b* is an automorphism of order < 2 of the additive group Sesq_,(4) and satisfies (£b)* = &b* for Fe Z. We compute the adjoint of b*: hye(x)(y) = DC, yp = BEY, *) = hy Vex) = AOD) = EVP) = A can ye C9(y). Thus ye — (hy) ean - Further, we have Sesq(p)(b*) = (Sesa(p)(b)]* for any pe Hom,(M, N) and be Sesq,(N). The set G(Z) of elements ¢ of the centre Z of A such that ef = | is a multipli- cative subgroup of the group of units of A. Let ¢¢ G(Z). We say that a form be Sesq,(M) is e-hermitian if b = eb*, ie. > y- b,x, x YEM- We call an element hc Hom,(M, M*) e-hermitian if h = ch* cay. Since the adjoint of 6* is h*can,,, a form b is e-hermitian if and only if its adjoint A, is e-hermitian, In the following, we agree to say that a sesquilinear form has a certain property if and oaly if its adjoint has the property. We are mostly interested in the cases ¢ = +1. 1-hermitian forms are classically called hermitian forms and (~ |)-hermitian forms skew-hermitian forms. If the involution is trivial, we obtain the usual symmetric and skew-symmetric bilinear forms over com- mutative rings. Let S,: Sesq,(M) — Sesq,(M), resp. S,: Hom,(M, M*) + Hom, (M, M*) be the map given by S,(b)—5 + cb*, resp. 5,(4)—h4 ch*can,; clearly Asay = S(h,)- Let Sesq,(M) = ImiS,) & Sesas(M) Sesq'(M) = Ker(S_,} < Sesq,(M). The latter is the group of e-hermitian forms on M. Since S_(b + 2b*) —0, Sesq,(M) is contained m Sesq'(M). If 2 is invertible in A, we have Sesq,(M) ~ Sesq’(M), since b = 45,(b) for b € Sesg'(M ), In general, Sesq,(M) is strictly contained in Sesq’(M), We call the elements of Sesq,(M) even s-hermitian forms. Thus b € Sesq,, (M)is even s-hermitian if and only ifb — k 4 ek* for some sesqquilinear form k € Sesq,,(M). We have Sesq,(M) ~ Sesq,(M)/KerS., so that 6 is determined by the class [K] of k modulo sesquilinear forms ky such that ko + ekg = 0. It is convenient to denote an even e-hermitian form on M by its class [Kk]. Let (M,,b,), (M2,6,) be even hermitian forms and let 9: (M,, b,) + (Mp, 63) be a morphism. Ifb, = k, + ekf and by = ky + ekf, it2, Sesquinear and Hermitian Forms = 9 follows from Sesq(9)(b,) = b, that Sesq(y) (to) — &, — —e(Sesa(o)(ks) — ky)*. Thus we get [Sesq(g)(k,)] — [k,] and conversely, for any g: M,—»M, such that [Sesq(9)(ka)] = [kr], we have Sesq()(b,) = by. Thus a morphism 9: (My, bi)» (Mz, b2) is a homomorphism of A-modules p: M,~> Mz such that [Scsq(p)(é2)] = [ky]. We now describe cven e-hermitians forms through their adjoints. Let Hom,(M, M*) = Im(S,) Hom,(M, M*) Hom'(M, M*) = Ker(S_,)< Hom,(M, M*), where §,: Hom,(M, M*)> Hom,(M,M*) is as above. The group Hom‘(M, M*) consists of all e-hermitian elements of Hom ,(M,M*), and Hom,(M, M*) is the set of adjomts of even ehermitian forms. Since Hom,(M, M*) ~ Hom,(M, M*)/Hom~(M, M*), we have a 1-1-correspond- ence between even e-hermitian forms and classes [g] of maps g « Hom,(M, M*) modulo homomorphisms gy: M —» M* such that go + eg cany = 0. Using this interpretation of even forms as classes of maps, we sce that a morphism 9: (Hy, [ox ]) + Ma, [a2]) is a homomorphism @: M,—+M, such that the diagram M, 2 .M, 23.0 o Mt. Mt commutes modulo a map go: M, + M¥ which satisfies gy + eg§ canur, = 0. (24) Sesquitinear Forms Over Free Modules. Lot M,(A) be the set of ( > k}- matrices with entries in A, For any «eM, ,(A) we denote by a* ¢ M,, »(A) the {k x nj} matrix 2, where a’ is the transpose of x and & = (%,) if = (x,)} Let M be a finitely generated free 4-module with basis {e,,..., e,}. We have a map Sesq,(M) > M,(A) given by br+f = (by) with by =b{e,, ¢;) The map is an isomorphism of Z-modules, since we can associate with any matrix f =(by)€M,(A) the sesquilinear form dix, y= 2B, x yeM; where £, 7 arc the column vectors with entries the components of x, y with respect to the given basis {e,,...,¢,} of M. Let {f1,..-, A}, fe M*, be the dual basis, ic. fi(¢) = 6). The adjoint 4: M -+ M* of b is given by he) = Lb ted 8) 710 1, Hermitian Forms over Rings where, as above, by = b(e;.¢,). If {ei,..., &4} is another basis with a elements (two bases necd not have the same number of elements!) and if p = (u,)€ GL, (A) is the transformation matrix: = Lem. then f= p*f'p, where f’ is the mactix of b with respect to the basis fei, ---e4} Let M, and M, be free A-modules of the same rank m and let f: (M,, b,)—>(M,, bz) be an isometry given by fe@)= Leite, 2 = (ua) e GL,(A) , where fe,,...,€) is a hasis of M, and {ej,..., e4} a basis of M,. We have b(n 4) = bx Fe, Fe) = Lane, én oF f, = 0*Bxo, B, being the matrix of b, with respect to the given basis of Mj, i= 1,2. We say that two matrices f,, 8, €.M,,{A) are congruent if 8, — p* Bp for some p € GL,(A). Thus the matrices of a form with respect to different bases (of the same cardinality) or the matrices of two isometric forins are congruent. If § = (bj), by; = b(e,, eis is the matrix of the forin b with respect to the basis {e,,... , 2}, then f* = (bj; obviously the matrix of b* with respect co the same basis. ‘Thus we have p= «ft for the matrix of an s-hermitian form. We call a matrix Be M,(A) e-hermitian if fi = 2f*. Let now b be an even s-hermitian forin on M. We have beet, €) = (er, 7) + k(C,, for some sesquilinear form k. Thus necessary conditions for a matrix = (by) to be the matrix of an cven s-hennitian for arc that n+, qeA, im l...n eby fori #j. These conditions are also sufficient, since we can write byaky + eke, BJ=.o om, with ki = a;, ky = by fori < jand k, = Ofori > j. Thus the forin is given by the class [(k.,)] of the matrix ky kaa 0 haw Okay 0 0 0... Oke We may say tbat the matrix (k,) is the “half” of the matrix of the form,3, Hormitian Modules = 11 §3. Hermitian Modules (31) Let A bea ring with involution and let § — A be the category of finitely gencfated projective tight A-modules. We call a pair (M,b) with Meg — A and be Sesq'{M) an ¢-hermitian module, We also use the notation (M, h) for an c-hermitian module, with he Hom,(M, M*) the adjoint of 6. If the form b is even, ie. b= k+sk* for some keSesqs(M)] or h=g + eg? cany for g¢Hom,(M, M*), we call (M,h) an even e-hermitian module and use the notation (M, []}. 12 4*, any hermitian module is even (the prefix s will often be omitted), If the involution is trivial, 1-hermitian modules are called symmetric bilinear farm modufes and (—1)-hermitian modules skew-symmetric bilinear form modules. For any aed, we define S,:A+A by S{a)=a-+ea and put S*(A)= Kor S_, = {xe A|x = ex). We clearly have S,(A) = ImS, < S*(A). G.1.1) Proposition. Let (M, b) be an e-hermitian module. Then (M, b) is even if and only if b(x, xyeS,(A) for all xe M. Proof. If b is even, then b(x, x) = k(x, x) + ek(x, x) = S,(k(x,x)) for some keSesq(Mf), The converse was proved in (2.4) for a free module M. For M projective, we choose M’ such that N = M © M’ is free and define an even form 4, on N by b, = bon Mand b, = 0 otherwise. We deduce from the free case that b, = S,{k,) and b = S,(k), where & is the restriction of k, to M. a A module M such that cany is an isomorphism is calfed reflexive, Finitely generated projective modules are reflexive: B.1.2) Proposition. For any Me sp — A, the canonical map cany: M+ M** Is an isomorphism. Proof. The claim is cvident for the A-module A. The general case follows from the additivity of the functor **. a Even if it is usual to identify Mf and 4f** through the isomorphism can,,, we prefer not to. However we shall often omit the index M in cany. (3.2) Let (.M,»b) be an e-bermitian module. We say that the form b is non- Singular or regular b if the adjoint h: M+ M* is an isomorphism, We call b nondegenerate if h is injective. Modules (M;) with b nonsmgular are called spaces. Thus, a nonsingular e-hermitian module is called an s-hermitian space and a nonsingular even e-hermitian module an even z-hermitian space.124, Hermitian Formas over Rings 2.1} Example, Let (M, 5) be a symmetric bilinear form module such that M 1s free with a basis {e,, . . . ,¢,}. Since the adjoint of b is described hy the matrix 5(e,, ¢;), (M, 6) is nonsingular if and only if det(b(e,, ¢,)) is a unit of R. We call dey, ..., &,) = det(h(e;, €,}) the discriminant of b with respect to the basis {e,,...,&,)- If {e,,..., &%} is another basis of M with base change e,=) uye, we have de, ..., 0) =d(er,-...,¢&)- dettu,)?. Thus the class 4 of d(e,...,¢,) modulo R* is independent of the choice of basis. We call 4 the discriminant of (M, b). For an arbitrary symmetric bilinear form moduie (M, b) of constant rank we define the discriminant ideal d(M, b) of (M, b) as the ideal generated hy all elements d(x,,..., %4), EM, i , n Localizing (and applying (7.1.3)), we see that d(’M, 5) = if and only if (a, dy is nonsingular. We denote by Serm'(A) the category of «-hermitian A-modules and by Serr’, (A) the catcgory of even é-hermitian A-modulea. The corresponding categories of spaces are denoted by $'(A) and §¥,(4). Morphisms in these categories are usually taken to be isometries. G3) We now recall some formal properties of finite direct sums of modules. These properties will be useful in the definition of orthogonal sums and hyperholic spaces. For any finite direct sum @ M,, tet q; be the canonical injections and p, the canonical projections; * M, 4 é M;,— M,. We recall that pig, = 5,1, and S440, — toa For any ge Hom, (é M,, é neh we have n? maps ay = PggsMj~N, FH A.yn where {4),p,} are the canonical injections and projections for 6 Mand (4), #4) 2 those for @ N,. Conversely, we can recover » from the @;,: “1 e= z Tay; - If @,,) is the matrix of maps associated with ge Hom( ® Mj, @N,) and (b,,) the one associated with ye Hom(® Nj, @ L,), then the matrix (c,} associated with the composition ¥@ is given hy the usual product formula of the two matrices (Gj) and (b,;) 655 = z b44,,. dentifying the dual (@ M,)" with @ M)’, we have PY = q;. Thos if (au) i is the matrix of g, then (aj) is the matrix of p’3. Heemitian Modules 13 Some useful results follow easily from this formalism: (33.1) Lemma, Let g¢Hom,(N,@N;z, L, ® £2) and let a, = Peay Nj L,- Then 1) If and ay, are isomorphisms, then py@~ ‘gL + N is an isomorphism. 2) Assume that az, = 0 or ay, = 0. If two of the three maps 9, 11; G22 are isomorphisms, the third alsa is an isomorphism. 3) [fay, =O and a, = 0, then ¢ is an isomorphism if and only if a,, and ay, are isomorphisms. 4) Assume that 4,1 ~ 0 oF 4, =0. Then {8 an isomorphism if'and only if, and @,, are isomorphisms. Proof. 1) If (by) is the matrix of g~', then a matrix computation shows that DogiLy + Ny is an isomorphism with inverse a2, — @21¢;;'412. 2) Assume that g and a,, are isomorphisms, then it follows from the proof of 1) that a3 is an isomorphism, The other cases are similar, (at oO 3) The inverse of y is given hy the matrix ( “ } az 4) If ay, — 0, the inverse of @ is given by the matrix ~1 - 4,102,077 ay! az} 0 ‘The other case is similar. a LM = ® M,, then M* is canonically isomorphic to é Mf# and a homomor- phism hy M — M* is given by a matrix (h,,} with hy: My —> MP. ‘The map h* can, is described by the matrix (h* cany), — kkcany,,: MM; —» MP For example, an hermitian module (M, © M;, hy has a matrix Ge aoa) fy Ia with f,,:M;— M¥ such that chifcamy, = Ay, i— 1, 2 and hy, €Hom ,(M, M2} An element & of Sesq~*(M, © M,) is given hy ( ky, — 6h, cama, ka Kap with k, H*(M) and is an isomorphism if and only (M, [g]) is nonsingular. Proof. We have to check that the diagram M®M —. M @M* (4) (4) M*@ M* .— Mt @ Me commutes modulo Sesq~*(M © M). We have O 1) _f 9 sgtean\_ f(g 0 0 eg*can (0 o)e(5 ran)“ ( aan) (9 Wo")16 1, Herautan Forms over Rings Since g — eg*caneSesq_.(M) < Sesq*{M) < Sesq”*(M@® M), we see that @ = eg*can (modulo Sesq-*(M @ M)). Further a olea(o a)*(. “e")=(0 1) * sine ( 0 eg yn eseng © M). Finally, the factorization o 0 _ft oven -1 =( (6 ¢+uren) shows, in view of (3.3.1), that a is an isomorphism if and only if g + eg*can is an isomorphism. a (3.5.4) Corollary. Any even s-hermitian space is isometric to an orthogonal summand of a hyperbolic space H*(A"), Proof. By (3.5.3) (M, [g]) is an orthogonal summand of H(M), Ifwe choose M’ such that M@M~ A®, then (M,[g]) is an orthogonat summand of (ML) 1. He(M!) = FEA"), a 3.6) In the rest of this section, « is fixed and the prefix ¢ is omitied. Let (M, 5) be an hennitian module. For any suhmodule U — M we denote the restriction of bio U hy bly and the adjoin of bjy by hi. [fb is even, then bly is even. 3.6.1) Lemma. Let (M, b) be an hermitian space. Let U and V be submodules of M such that M= U4 V and b(U,¥)=0. Then (U, bly) and (V, bly) are hermitian spaces and (M, by= (U, bluy LK bly) - Proof. Wf xe UV, then bix, U) = b(x, V) =0, henes b(x, M) = 0. Since b is nonsingular, we have x = 0. This shows that Un V = {0} and M=U@YV. Now let =hyU@V— Ur@Vt, i=1,2 be the adjoint of b. The map ha: U — V* is given hy w+ b(u, —)lv,30 hg, = 9. Since » is hermitian we have h,, = h},can and h,, — 0. Thus 4 -(° Ok ) 22 and we have the claimed orthogonal decomposition. Wo say that two subsets X < M and ¥ c M of an hermitian module (M, 4) are orthoganal if b(x, y) = 0 for all ¥¢.X and ye ¥, For any suhmodale U < M we define + = (xe M|b(x,t)- 0 VueU} =(yeM|b(wy)=0 VucU}. 3. Herminan Modules 17 0.6.2) Lemma, 1) Let (M, b) be an hermitian module and let U be a submodule of M which is finitely generated and projective. If (U, bly) is nonsingular, then (M,b) = (U, blu) LU, bly). 2) If (M, b) = (M,, by) 1(Mz, bz), for some hermitian modules (M,, by), i = 1, 2, then the b, are nonsingular if and only if b is nonsingular. Proof. |) Since bly is nonsingular, UU! = {0} and by G61) it suf- fices to show that M=U+ U+. For any meM, we have a linear form h(n) = b(n, —)eM*, hence by restriction a linear form A(mjlyeU*. Since hly 38 an isomorphism, there is we U such that b(m, —jly — ou, —)|y. This shows that b(m — 2, —) is zero on U or that m — ue U+, Therefore we have mautm—uel + Ut. 2) is am immediate consequence of (3.3.1), a G.63) Remork. @.6.1) and (3.6.2) are chviously also valid for even forms. ‘We say that an hermitian space (M’, b’) is a subspace of (M,) if M’ is a submodule of M and b’ = Bly. @.7) An clement x #0 of an hermitian module (M, b) is called iso:ropic if (x, x) — Oand an benmitian module is called isotropic if it contaius an isotropic vector #0. We say that (M,b) is anisotropic if b{x, x) — 0 implies x ~ 0, A submodule U of an hermitian module (M, b) is called totally isotropic if b|,, == 0 or equivalently if U < U+, An element x of an even hermnitian module (M, [g]) is isotropic if g(x, x)eS~*(A) = {a€ Ala = —ea} and (M, [g}) is anisotropic if0 is the only isotropic vector of M_A submodule U of (M, [g}) is totally isotropic if S,)ly = 0. For example, the suhmodnke P of a hyperbolic space H"(P) is totally isotropic. Conversely, we have: 3.7.1) Proposition. Let (M, [g]) be an even hermitian space and let U be a totally isotropic direct summand of M. There is a submodule V of M such that M=U‘@¥V and (U®Y, [alverv])~ H(U). In particular, there exists a morphism H'(U)— (M, [@]) which restricts to the identity on U. Thus (M, [g]) contains an orthogonal summand isometric to H*(U), Proof. Since U is a direct summand of M, the embedding U < M induces a surjective map M* -» U+. Composing with the adjoint h: M + M* of b = S,(g), we obtain a surjective map M —» U* with kemei U+, ie, the sequence. 0-+U'SM—U*+0, where mrrbim —Jle is exact. The module U* is finitely generated projective because U is finitely gcnorated projective. Hence the sequence splits and U+ also is a direct summand of M: let M=U!® W. Then ¢:W—» U* giyen hy wr+b(w, ~J[py is an iso- morphism and @*cany: U —» W’* is an isomorphism. We haye g*can(u)(w) = u**(p(w)) = p(w)(i#} = b(w, u) = Zb(u,w), Thus the map U -» W* given hy18 T Hermitian Forms over Rings. urrb(u, Iq is an isomorphism. The map hip gw: @ W-» U* @ W* i 0 chf,can fing Igy urs bw, —)|w- By 3:3. 1, dle w is an isomorphism and bly » y is nonsmgular. Since b is even, the restriction b|y g w is also even. We now show that W can be replaced by another direct summand V for which there exists an isomorphism ({U® ¥; [alv@yv]) = H'(U) which induces the identity on U. Let U @ W =N and let ':N—-N* be the adjoint of b’ = bly. Since b’ is even, we have given by a matrix where 4,,:U + W* is the isomorphism ht = S,(g'), 9’ € Hom(N, N*). Let 1) be the matrix of g’ with respect to the decomposition N = U @ W. Since hily = 0, we have a + sa*cany = 0 and there- fore wi-[¢~ a0 0 0) Thus, we may assume that a = 0. Similarly, we: may also assume that b = 0, since [( orcan 3) [-2 We then have 0 ectean op oo, . i -( 44 wean Since his nonsingular, ¢ is an isomorphism. For any automorphism f- (1 *) ot w= U@ W, we have ae o1 an Ce Pit-(o aa): Let x — ~e7!dand let 7 = f(W). Westill have N = U ® / and fis an isometry 672) wenn >(ve “2 ol) which induces the identity on U. We have [C LG ®)! with d= sctcaneHom,(¥,U*} The isomorphism ( °)\-u gus ver oa finally gives an isometry H*(U) + (U @ v,[9']) = (U @ V, [alu v1), which is the identity on U. As im the first part of the proof, we check that 7" is a direct summand of M. Since N is a subspace of M, we have by (36.2) M=N LW4, herice the orthogonal complement of V in M is the orthogonal sum of N+ with the orthogonal complement of ¥ in N, In view of (3.7.2) the orthogoual complement of V in N is V itself, hence V+— Y@N+. This shows that M =U @ V* and (3.7.8) is proved. a (3.7.3) Corollary. Let (M, [g]) be an even hermitian space. Then M has a direct summand U such that U-= U* in M ifand only if (M, [g]) is isometric to H'(U).3, Hermitian Modules 19 Proof. We verify that M = U ® V, where Vis as in (3.7.1). By (3.6.2) (for even. forms) M = U @ VLU’ for some U' c U". The condition U ~ U+ then implies U’ = {0}, The converse follows from the definition of hyperbolic spaces. 0 (3.74) Corollary. Let (M,[g}) be an even hermitian space. Let xeM be a unimodular element, i.e. a basis of a free direct summand of M, which is isotropic. ‘Then there is an isotropic element yEM such that S,(g)(x, y) = 1. In particulor, there is a hyperbolic plane H°( A) ~ (M,[g]) such that x¢H*(A). Proof. If we take U = xd in 3.7.1), then ¥ = yA and y can be normalized so that S,(g)(x ¥) = L. a A pair (x, y) of elements of M as im (3.7.4) is called a hyperbolic pair of M. A direct summand U in an hermitian space (M, 4) such that U+ = U is called a lagrangian. By (3.7.3) any even hermitian space with a Jagrangian is hyperbolic, This is not truc im gencral for hermitian spaces. For example the symmetric bilinear space 2/9 1 (er over a field K of characteristic 2 has a lagrangian but is not hyperbolic. An hermitian space (M, A) with a lagrangian U is called a metabolic space. The notion of 2 metabolic space and the following results are due to Knebusch [1969/70}. To any hermitian module (U, k), we may associate the metabolic space. M(U,kKy={ UG U* 0 1 (G4) *\ecan ko) }* If kis even hermitian (for example k = 01), then M(U, k)is an even hermitian space and M(U,k)~ H*(U). This follows from (3.7.3) or directly from the equation 673) (4 Deo wo a a6 5) putting k = d + ed*can. We now check that any metabolic space is isometric to € space of the type M(U, k). @.7.6) Proposition, Let (M, h) be @ metabolic space with lagrangian U. Then 1) (4, ~( veUs, ‘for some hermitian form kU» U*, can kt 2) (M,h) LM, ly ~ HU) LM, ~h),20 IL. Herminan Forms over Rings Proof..Let M = U@ V and ket 0 d he (caren 4) be the matrix of h:M = U® ¥ > M* = U* ® V*. Since h is an isomorphism, d:¥ + U* is an isomorphism, Identifying U* with V by means of d, we see that 0 1 ainr=(vour(,2. )) for some hermitian form k:U -> U*. This shows 1), To prove 2), we may assume that (Mf, A) is as in 1), The claim then foliows from the matrix equation 1000 0 ol 9 9 LO fF gcan-fek o1o08 ecan k O 0 910 0 _ 1010 0 0 Oo -t oo14 0 - Koo. 0 0 -ean -k/\o1 0 1 o 1 0 0 ecan 0 0 a Q o 68 0 -i 0 O -ecan —-k G17) Corollary, The orthogonal sum of metabolic spaces is metabolic. Proof. M(U,k)LM(U’,K) =~ M(U@U,KLK). a The next results are generalizations to hermitian spaces of previous results on even hermitian spaces. @.78) Lemma. For any hermitian space (M,h), the space (M,h).L(M, —h) is metabolic. 1 1\fa oO 10 oO rot (0 a) SGC). 2 ‘Let (M, k) be an hermitian space and tei U be a totally isotropic subspace of M. ‘The hermitian form h induces in a canonical way an hermitian form on U*/U. We denote it by h. (3.79) Propesition. Let M, b be an hermitian space. If U is a totally isotropie direct summand of M, there exists a submodule V of M such that M =U ®V4. Symplecuc Spaces 21 and(U ® V. ily @ v) is ametabolic space, The orthogonal complement of U @ V is isometric to (U+)U,h). Proof. As in the proof of (3.7.1), we get that U+ is a direct summand of M, M- U'@V and that hy 9 is given by a matrix -( 0 nom) fay dag] where 4, ,:U + ¥* is given by ut h(u)ly and is an isomorphism. By (3.3.1) the form h’ is nonsingular and is, obviously, imetabolic. By (3.6.1) we can pick an orthogonal complement U’ to U @ V. We have U’ c U4 and U4 = U' @U. The last claisn follows fron this decomposition. ao §4 Sympiectic Spaces (4.1) Let R be a commutative ring witb the trivial involution and tet M be an Ranodule, We call an even (—1)}-hermitian form b:M x M— R an alternating (form. Thus b is skew-symmetric and (x, x) = 0 for all xe M.A pair (M, 5) with Me ® — Ris calied a symplectic module. We say that (M, b) is a symplectic space if(M, 6)is noasingular, Lf M is free, the anatrix of b with respect to any basis of M is skew-syimmetric with zeroes on the diagonal. We call such a inatrix altern- ating, Tt follows froin (4.1.2) that a symplectic space snust have even rank. The following result shows that the structure of symplectic spaces is, for inany rings, very simple. (41.1) Theorem. if for every symplectic space (M, b) of rank > 2, M has a direct summand isomorphic to R, then every symplectic space of constant rank 2m over R is isometric to HR), (412) Corollary. [f every MeB — R is free, every symplectic space over R is isometric to HW }\(R™) for some m. Proof of (4.1.1): Let M = M' @ Rey, Re, free of rank 1. Let fe M* be such that f(e.)=1 and {(M’)=0. Since (M, 6) is regular, there is e)¢.M such that {= b(-, e,), The elements ¢, and e, are linearly independent and the restric- 0 1 o} BY G62) and 363), (M, 6) = HO(R) 1 (M’, 6’) and the claim follows by induction (and te formula HO 9(R)L HORT) = HON Ry, tion of b to Re; @Re, has matrix (_ (41.3) Remark, Let w be an invertible (2m x 2m)altornating matrix with entries in a ting R satisfying the assumptions of (4.1.2), By (4.1.2) there exists22 1. Hermuan Forms aver Rings ve GLaa(Rysuch tate = ing (( 9 oe (on 9) Ta dete a square in R. Let J,,, be the alternating (2m x 2m)-matrix : Oo 1, ta-(f ; ). where 1, is the (a x n}identity matrix. Let {e1,..., én} be the standard basis of R™ and let {ef,..., ef )e(R™)* be the dual basis. Then J,,, is the matrix of the symplectic space H("")(R™) with respect to the basis {e,,..., em, ef,...,¢E}. (4.1.4) Proposition, Let ueGL;,(R) be an alternating matrix. There is a faithfully flat R-algebra T and veGLz,(T) such that v'uv = Jz, it GLon(T} Proof. If R is local, we may take T= in view of (4.1.2) Thus for any pe Spec(R), there is v(pJeGL(R,) such that v(pur(p) — La». For any fe R, Jet R, be the localization $+ R at 5 = {1,f,f7,. ..}. Choosing a common de- nominator for all entries of v(p), we get ge R\p and v(g)eM;,(R,) such that v(g¥wu(g) = Jaq, over Ry. To have equality over Ry for some fe R\p and to got an invertible matrix v(/), we may have to invert formally two more elements 7.9’ R\ p and to replace g by f = gq’g”. The ideal generated by the set E of the J's such that there exists »( fje GL(R,) with v( f¥uv(f) — Jo, is not contained in any maximal ideal of R, hence is R. Thus we getf,,...,f,¢ E such that Rf, +... Rf = R. We put T= Ry, x... x Ry, and v= ((f,)---, H(f-)} ‘The type of argument used in the proof of (4.1.4) is applied at different places of this book. Observe that T is in fuct etale (ste IIT, (2.5)) §5. Unitary Rings and Modules (3.1) Let A be a ring with inyolution and let Me B — A, In section 2, we have identified the additive group Sesq,(M) of even e-hermitian forms on M with a quotient of Sesq(M) ~ Hom ,(M, M*) Sesq.(M) — Sesq( MM /Sesq~*(M) = Hom,(M, M*)/Hom7"(M, M*), where Sesq‘(M) = {bc Sesq(M)Ib + eb* = 0} Hom -(M, M*) = {heHom,(M, M*)|h 4 ch*eatg = 0) .5. Unitary Rings and Modules 23 jt is useful to haye a theory of “forms” corresponding to various quotients of Sesq(M) or Hom ,(M, M*). This idea, due to Bak (see Bak [1981] for more jnformation), bas a nice unification role for the theory of even bermitian and quadratic spaces (sve (5.3.5) and is very useful is characteristic 2. Another reason for us to introduce this formalism is to obtain a good reduction theorem (sce Remark (4.6.3) in Chapter TI) Let S-4{A) = (a — 24,424} S-*(A)= {acAla + od = 0}. We have S_(A) < S-"(A), In general, if 2 is not a unit of A, the twa sets are different. For example, $_,(A) = 2A and S~*(A) = 4 ife = —1 and the involu- tion is trivial, He ~ | and the involution is trivial, then S_,(A) = 0 and S~*(A) = Ker(A —2+ A) We observe that $_,(4) = S~*(4) if there is an element a in the centre of A such that a + d= I:if x + 2% — 0, then x = ax ~ sax, Let 4 be an additive subgroup of A such that 1 S_(A)e Ae SA) 2) ax@e A for all x¢ A and ae A A triple (A, 5, 4) with A as above is called a unitary ring and the pair (2, A) is called a form parameter for A. We shall sometimes denote $~*(A) by Ang, and S_{A) by Agia. A morphism of unitary rings f:(A,a A)—(4’.¢,, A’) is a homomorphism f:A— A’ of rings with mvolution such that f(s)=«’ and f(A} A’. We say that fis a unitary surjection iff (A) = A‘ and f(A) = A’. If Ais an R-algebra with an K-linear inyolution, we define a unitary R-algebra as a triple (A, e, A) with A an R-submodule of A such tharS_(A)< Ac S7*{A)and axac A for all xe A and acd. For any commutative R-algehra R’, the triple (A. =A@gR, & =6@ 1, A’ = A @_R’) is a unitary R’-algebra. Let (A, , A) be a unitary ring and fet M be a right A-module, We define Sesqi"(M) = {beSesq-*(M)|b(m, mje A for all meM}, and A(M) = {heHom,(M, M*)|h + ch*cany, = 0, h(m\m)e A, Ve M} : A(M)is the image of Sesq7‘(M) through the isomorphism which associates to any form 6 its adjoint 4,. Since S_.(A) M be a homomorphism of A-modules. Since Sesq(f b*) — (Sesq(/)(b))® and Sesq(/}(S.(b)) = S.(Sesq(f Xo), Sesa(f) maps Sesq,*(M) to Sesqs*(M’).24. Herminan Forms aver Rings (5.1.1) Example. Let M = A* with the standard basis {e1....,¢.}. The group Sesq(A") can be identified with the additive group of M,(A} of (« x n}matrices over A (sec (24}) through the map 6++f — (by) with b,; — b(e,,¢)). ‘Then Sesq;‘(A") is the set of (n x n}-matrices # = (by) such that b= — eb, and bye A for al i,j =1,-.. 5%. (5.2) We define an (¢, A)-unitary module as a pair (M, [k]) where Me 3 — A and [K] is the class of ke Sesq(M } (modulo Sesq;*(M)})- Equivalently, we shall use the notation (M,[g]). where [g] is the class of geHom,(M, M*) (mod A(M)). The prefix (¢, A} will usually be suppressed if the form parameter (cA) is fixed. Since Sesq7'{M) c Sesq”*(M), the even hermitian module {M, S.(k)) is uniquely determined hy tbe unitary module (M, []). We say that the unitary module (M, [K]) is nonsingujar ot that it is a unitary space if $,(k} is nonsingular. If A = S~*(A), then (M, [k]) — (M, S.(k)) but in general the same even hermitian space may correspond to different unitary spaces. A morphism of unitary modules f:(M,,£4,})7 (M2 Ek2]) is an A- linear map f:M,>M, such that Sesq(f)fk.]} =[k,], ic. such that k, —S*kfe Sesa,*(M,). It mduces a morphism of the associated even mo- dules. A morphism fis called an isometry if fis an isomorphism. Observe that nonisometric unitary modules may have isometric associated even modules, (52.1) Example. If M is free with basis {e,,. .. , ¢}, we may identify as above a sesquilinear form k on M with the matrix (ky}@M,(A), ky =S(e, €) Since the matrix (s,) with ky for i>j sy= 4 0 inj — ely i Ha(Q) o1 or since Sat (9 7G op): The map (55.1) EEA(P,) 1. Ha(P2) > Ha(Ps @ Pa) given by ((7,,f,)(™2,f2)}> (71, M2), (Ai, /2)} is an isometry of unitary spaces. As in (3.5.3), there is a morphism (5.5.2) (M, [e}) 1M, 0 — 1) > Fa (M) which is an isometry if and only if 5.(k} is nonsingular. It follows that: (553) Proposition. Any (c, A}unitary space is an orthogonal summand of a hyperbolic space H(A". (5.5.4) Remark. We say that an (c,A}unitary space (M, [4}) is metabolic if there exists a direct sum decomposition M=N@P with N = N‘, that is k= ( 5) wth ae A(N), In fact these spaces are hyperbolic: adding to k the —2 —yecan S 0 claim follows from the equations (0 o)(0 S)0 #)-C a) ti nara ( 4 Jeatan, we may assume that & -(0 5) and the5 Unitary Rings and Modules 27 1 O\/0 1 1 oF o1 (2 M0 alo 1")-C6 a): (5.6) An clement xe(M, [k]} is called isotropic if [q}(x} = 0 aad (M, £}) is called anisotropic if 0 is the only isotropic vector of M. A submodule U of (M, [}) is called totally isotropic if [g]{ U} = 0 and $,(k}|y = 0 or equivaiently if the adjoint A, of k restricted to U belongs to 4(U}. We observe that the even hermitian space (M, S,(k)} is not necessarily anisotropic if (M, [k]) is aniso- tropic. For example the polar 6, of an anisotropic quadratic form q is isotropic in characteristic 2. ‘As for even hermitian spaces (see (3.7.1)) and with the same proof, we have (5.6.1) Proposition. Let (M, [k}) be a (e, A)}unitary space and let U be a totally isotropte direct summand of M. There is a morphism H'(U)— (M, [kl]) which induces the identity on U. (56.2) Corollary. Le: (M, [k}) be a unitary space and let xeM be unimodular, iz. @ basis of a free direct summand of M, and isotropic. There is a unimodular yeM such that [g}(o)=0 and S,(4)(x ») in particular, {x, y} Is the basis af a hyperbolic plane H',(A) in (M, []). Proof. Let U = xA. Since k(xa, xa) = ak(x, xjae A, we have [4](u) = 0 for all ue U. Further, 3,(k)lc = 0, so U is a totally isotropic direct summaod of M and the claim follows from (5.6.1) o A pair (x, y) of clements of M as in (5.6.2) is called a hyperbolic pair for the unitary space (M, []). (5.7) We denote by £%,( 4} the category of (¢, 4}-unitary spaces oyer A with isometries as morphisms. This is a category with products and we have obvious product preserving functors Hie) —> O4(4) and S,:24(4) —> §4(A)} (58) Scaling. Let (4, «, 4) be a unitary ring and let 1 be a unit in the centre Z of A. Then (4,6, A’) with = ji-te and A’ = 2A is again a unitary ring. For any sesquitinear form b:M x M— A, we have p5,(b) = S,(ub). Thus yb is &- hermitian if and only if b is e-hermitian. Moreover, ub €Sesaz (M) if and only if beSesqi*(M). If b,:M, x My A is another sosquilincar form and if SM, > M is A-tinear, then Sesq(f} is Z-lincar. It follows that fis a morphism28, Hermitian Forms over Rings (M,, ub.) > (M, ub) if and only iff is a morphism (M ,,,)— (M, b) Thus, the functor . (M, by (M, ub), which is the identity on morphisms, induces an isomorphism from the category of e-hermitian modules to the category of ¢-hermitian modules. In particular, we have isomorphisms of categories: Serm*(A) 5 Germ" (A) Germ’, (4) Germ! (A) Simnilarly (M, [£})> (M, Luk) defines an isomorphism (A) O4-(A). (5.8.1) Example, Suppose that there is weZ such that c= yf !. Then the categories Serm‘(A) and Serm'{A), resp. Sernt',(A) and Germ! (A) are iso- morphic via scaling. Let R= {xeZo,(x)= x}. Such a p exists if Z is a separable quadratic R-algebra with H' (Gal(Z/R) Z*) — © (see ITM, (281)) The condition H'(Gal(Z/R), Z*)=0 is satisfied if Pic(R) = 0. (5.8.2) Remark. Let u be a unit of A such that area? = uau-t is also an jnvolution of A. Then » — ui! is a central element of A and pi ~ 1. 1f (A, 8 A}is a unitary ring with respect to the involution a+-» 4, then (A, pe, uA} is a unitary ring with respect to the involution at+a°. Further, if (M, b) is an e- hermitian space over A with respect to the involution ad of A, then (M, ub), with (uby(x, ¥) = ub (x, ¥) is an ep-hermitian space over A with respect to the involution a-+a® — udu}. Similarly, (M, [nk]) is an (ey, uA}unitary space with respect to the involution asa" if (M, [k}) is an (, A)-unitary space with respect to the involution a a. As above, we have isomorphisms of the corresponding categories which respect to orthogonal sums and hyperbolic spaces. §6. Hermitian Spaces over Division Rings (6.1) We first show that an hermitian space {M, b) over a division ring can be decomposed into the orthogonal sum of an anisotropic space and a metabolic: space,6, Hermitian Spaces over Division Rings 29 (6.1.1) Proposition, Let (M,b) be an s-hermitian space over D. There exists an orthogonal decomposition (atop cya (7, (? a) tao :): with (M’, b") anisotropic. Proof. Let e, # Oe M be an isotropic clement. Since } is nonsingular we have Si€M such that b(e,,f,) = 1. The matrix of the restriction of b to the plane N =e,D+/,D with respect to the basis {e, ,f,} is the metabolic space ool 4) with d, =5(/,,f,). Thus dy is nonsingular and we have a decomposition (M, b) = (M, bly). (N, dlw+) in view of (3.6.2), Repeating the procedure, we we the wanted decomposition. ‘We claim that (M’, b’)is uniquely determined up to isometry by (M, b). We need some notations, Let V be a totally isotropic subspace of M, ic. Ve V+, The hermitian form induces in a canovical way an hermitian form 5 on ¥+/V, which, ohviously, is nonsingular (see (3.7.9)} (6.1.2) Lemma. Let (N,,C;) and (Nz, Cz) be metabolic spaces. If we have (Mi, b)L (Ny, C1) = (Na, Ca) then (M, 6) is metabolic. Proof. We assume that M and N, ate subspaces of a metabolic space N2. Let V;, bea jagrangian m N, and V, a lagrangian in N. Let V+ be the orthogonal of VY, im Ny. The induced form ¢, on V+ /¥; is isometric to (M, 6) so it is enough to check that ¢, is metabolic. The subspace L = {re Vt jV,\xe His alagramgian in VisV, sce = Vd in Ny. (6.1.3) Remark. This elegant proof of (6.1.2) comes from Barge-Lannes-Lat- our-Vogel [1974] and was communicated to us hy E, Bayer. (6.1.4) Proposition. Let (M,, bi), (M2, 42), (N1,¢1) and (N2, ¢2) be hermitian spaces over a division ring. Suppose that M,, My are anisotropic and Ny, Nz metabolic. If M,1N, =~ M,1.Ni, then M, ~M,. Proof. By (6.1.2) and (3.7.8) the space (M, b) = (M,,>,).1.(M,, —ba) is meta- bolic. Let 7 c M be a lagrangian, ie. V = V+. The orthogonal projections of (Mf, by onto [Mab yand ts, —b,)induce maps p,:¥ + M, and po:V—+ M2. Since (M,,,) and (M2, —b,) are anisotropic, p, and p, are injective maps: Lom, +m, meM,, we get 0 = b, (et, m,)— b (rz, mz), 80 that m, = 0 implics b2(mz,7t2)=0 and m,=0. Therefore DimpM, > DimpV,30. Hermitian Forms over Rings DimpM, 2 Dim,V. On the other hand Dim,M, + DimpM, = Dim)M = 2Dimy V, so that Dimy V = DimpM, = DimpM, and p,, p2 are isomorphisms. The map p2pz? is an isometry (M,, 61) = (Mz, 62). a {6.1.5} Remark. The metabolic part in (6.1.1) is not uniquely determined up to isometry. For example 11 0\s1 0 O\s110 $oo 1114/7001 111 )]=;0 01 0106 o1tj/\O 106 010 ot over F;, but the metabolic spaces t i) and C 9.) ate not isometric. However this will be the case if 2 # 0 by (63.2) (6.2) We say that an c-hermitian module (M, b) over A is diagonalizable if M is free with a basis {e,,...,¢,} such that the matrix of b with respect to {é1,-»+ + &} is diagonal. Such a basis is called an orthogonal basis. We denote the e-hermitian moduie (A",b), where the matrix of b with respect to the standard basis of A" js a diagonal matrix diag(a,,..., a,), by (ay, ..-, a,)- Thus Me, by is diagonalizable if and only if (4,8) = Ca, ..., 0,) for some ay, «+, Gy Such that a; —28;,8=1,..., (62.1) Lemma, Any anisotropic e-hermitian space (M, b) over @ division ring D has an orthogonal basis. Proof. For any xe M, x #0, the form |,» is nonsingular. Thus, by (3.6.2) we have a decomposition (M, 6) ~ (xD, Blan) 1 (M,, 61). The claim then follows ty induction on the dimension of M. Combining (6.2.1) with (6.1.1), we get (6.22) Corollary. Any s-hermitian space (M,b) over a division ring D is iso- metric to an orthogonal decomposition 1 (M,, b,) with Dimp M; < 2. i An e-hermitian space such that all its elements are isotropic cannot have an orthogonal basis. Such spaces only occur in very special cases: (6.23) Lemma, Let (M,b) be an c-hermitian space over a division ring D such that all its elements are isotropic. Then e = +1, the involution on D is trivial (hence D is commutative) and b is antisymmetric. In particular, the rank of M is even and, ifs = 1, then D must have characteristic 2.6. Hermitian Spaces over Drvinon Rings 31 Proof. Assume that the involution on Dis not trivial and let d, D with d, # d,. We have d= —d for d = d, — d,. Choosing x, ye M such that h(x, y} = 4, wo obtain 0 = b(x + yx + y) — B(x, x} — bly, y) = d + ed, thus d(1 — = 0 and e= 1. Thus, ifs # 1, the involution is trivial. If the involution js trivial, we have e? ~ 1 since cé — 1. Therefore fe 4 i, then ¢ — — 1, the involution is trivial and antisymmetric. Now let ¢ = 1. Choosing x, ye M such that B(x, y) = 1, we obtain as above 142-0. Thus ¢=1 implies 2=0. Let de PD. Choosing x, yEM such that d(x, y) = d, we see that O=d +ad (N’, 0’) of hermitian spaces and for any subspace V < N, we have u(V4) — u(V}/, the claim follows from the following iemma: (6.3.5) Lemma. Let (M, 6) be an hermittan space over a division ring D of characteristic not 2. Let ¢,e'€M he such that ble, e) = bie’, e¢) # 0. There exists an isometry u: M > M such that u(e) = e’. Proof. Let N=(e—¢)D. If e¢N+, then M=eD@N* and u(ed + x)= eld +x, xeN+, isan isometry. ee N+ then also e’€ N+ and b(e — ¢,¢ — e)—0. But then bie +e, + ¢) = 4b(e, 0) # 0 and e¢(N")!, where N’=(e + DD. Replacing NV hy N’, we are reduced to the first case, a (64) In characteristic 2, the reduction to the case dim, M = 1 in the proof of the Witt cancellation is not possible. We shall deduce Witt cancellation for even e-hermitian spaces from the following analogue of (63.5. (64.1) Lemma, Let (M, 6) be an even e-hermitian space over a ditision ring D and let {x,y}, {x,y} be hyperbolic pairs in (Mb). There exists an isometry ui MM such that u(x) = x, u(y) = y’- (642) Corollary. The Witt cancejlation (6.3.2) holds for even e-hermitian spaces over a division ring D. Proof of (6.4.2). By 3.5.4) and induction, we may assume that (M, 6), (M’, ') are hyperbolic pianes in a space (M, 6) and that we have two orthogonal decompositions (M,, b,)L H = (M, 5) = (Mz, by) LH’. It follows trom (64.1) that there is an isometry « of M such that u(H) =H’. Thus # induces an isometry of even hermitian spaces (M,,b,) = H+ 73 (M,, b:) =(H"}-. a We introduce some definitions before proving (6.4.1), We say that two hyper bolic pianes H, H’ in (M, b) are adjacent if there is a hyperbolic pair {x, y} in H and a hyperbolic pair {2, y"} in H’ with a common clement. We say that H and LH are related if there is a finite chain of adjacent hyperbolic planes connecting H and H’.6, Hermitian Spaces aver Division Rings 33 (64.3) Lemma. Two hyperbolic planes H, H' in an e-hermition module (M, b) over a division ring D are always related, Proof. Let {x, y}, {x’, y’'} be hyperbolic pairs m H, H’. If one of the 4 elements d(x, x'), B(x, Y), B(y, x’), O(y, ¥') of D is not zero, say b(x, x’), we replace the pair {x,y} fot H’ by the pair {x" = x(x, x’)! y" — y’bQ%, x)}. Then H and H' are adjacent to the hyperbolic plane generated hy the hyperbolic pair {x, x""}. If all of the above 4 elements are zero, then {x,y +x} and {y,y + x’} are hyperbolic pairs generating hyperbolic planes H, and H,. The chain H, H,, 2, H' shows thal H and H’ are related. a Proof of (6.4.1). In view of (64.3), we may assume that x = x’ in (6.4.1), Let z=y'— yand et N =2D. If y¢N4, thon M = yD@ N+. The map u: MM given by u(yd + w) = (y + 2)d + wis an isometry such that u(y) = y’, u(x) = x. If ye N+, then b(z, 2) ~ O and ze H+, Let vc H~ such that {z, v} is a hyperbolic pair im H+ and let Hy, be the hyperbolic plane generated by {z, 0}. The linear map 4! LH, + HLH, given by xtox, yoy 4 2, 2-42 and o++0 — ex has the matrix 1000 _f{ 9110 “=| oo10 -- 001 with respect to the basis {», z, y, x}. Since 010 O11 100 w{2 900), { 000 0|_{o 000 000 1f**|-« 00 1/=[o 002]: 0000, 0000; \0 000 we see that u, is an isometry H 1 H, > H 1 H, of even e-hermitian spaces such that u,(x) =x and #,(y)= y". Smee (M,b)~ HLH, 1(H LH,), 4 can be extended to an isometry u: (M, 6) (M, 5) with the desired property. a The lemma (6.4.1), used for the proof of (6.4.2), is of dependent interest. Lemma (6.4.3), used for the proof of (6.4.1), may, on the other hand, be applied directly to prove a more general version of Witt cancellation: (644) Proposition, Let A be a ring with involution and let (M, b) be an hermitian module over A. If H and H' are two adjacent hyperbolic planes in M, then the s- hermitian modujes H~ and H'* are isometric. Proof. In view of (3.6.2) we have orthogonal sum decompositions (M, b) ~ HLN=H'LN' with N=H! and N’ =H’, Let {qu p;} be the injection and projection maps corresponding to the first decomposition and let {qi pi} be those corresponding to the second decomposition (see (3.3)). We compute the