48b6792
THE HILBERT-SMITH CONJECTURE IN EVERY FINITE DIMENSION
a complete proof, rewritten to be read
========================================================================
HOW TO READ THIS DOCUMENT
========================================================================
This is the full proof of the Hilbert-Smith conjecture in every finite
dimension. It is long, but the shape of the argument is simple, and we
say the shape first so that no reader gets lost in the machinery.
There are two very different halves.
HALF ONE is algebra and geometry: we build a machine that turns a
space into an abelian group (a "signature group") which is rigid ---
it does not have room for fractions with huge denominators. This is
where the heavy sheaf theory lives.
HALF TWO is the group action: a faithful action of the p-adic integers
forces a family of classes in this group whose rational values are
4/p^k of a fixed generator, with p^k as large as we like. That is a
fraction with unbounded denominator, which the machine forbids.
So the contradiction at the end is arithmetic (4/p^k versus a fixed
grid of size 1/L). Everything else is there only to justify that the
arithmetic is legal. If you remember one sentence, remember this:
A faithful p-adic action would put 4/p^k on a lattice of mesh 1/L,
for every k at once. For p^k > 4L that is impossible.
Notation is standard. We write Z_p for the p-adic integers, R and C for
the reals and complexes. A "manifold" always means a connected,
second-countable, Hausdorff, finite-dimensional topological manifold
without boundary. Sheaf coefficients are real unless stated.
========================================================================
THE TWO THEOREMS
========================================================================
MAIN THEOREM (Hilbert-Smith). Let n >= 1. A locally compact,
second-countable, Hausdorff group acting faithfully and jointly
continuously on a connected n-manifold is a Lie group.
Here "faithful" means the action map G -> Homeo(M) has trivial kernel:
no nonidentity element fixes every point. It does NOT mean point
stabilizers are trivial; stabilizers may be arbitrary.
Because we work in every dimension with no regularity and no
compactness, the manifold may be noncompact, nonorientable, or even
nontriangulable, and fixed points are allowed.
THEOREM (p-adic exclusion). Let n >= 1 and let p be prime. Every
jointly continuous action of Z_p on a connected n-manifold has nonzero
kernel. Equivalently, such an action factors through a finite quotient
of Z_p.
The second theorem is the engine. The first follows from it and from
classical structure theory, as we explain next. So almost the entire
document is devoted to proving the p-adic exclusion.
========================================================================
WHY Z_p IS THE ONLY ENEMY
========================================================================
The p-adic integers are the profinite group
Z_p = lim_k Z / p^k Z,
additive, compact, totally disconnected. They are the standard example
of a compact group that is not a Lie group: around 0, the subgroups
p^k Z_p are open and shrink to nothing, so there is no neighborhood of
the identity free of nontrivial subgroups.
Classical structure theory, due independently to Gleason and to
Montgomery-Zippin (1952) and extended by Yamabe (1953), gives the exact
converse:
REDUCTION LEMMA. A locally compact group is a Lie group if and only
if it has no small subgroups. Consequently, a locally compact group
which is NOT a Lie group contains a closed subgroup isomorphic to Z_p
for some prime p.
This is the passage from "no small subgroups" to "a copy of Z_p": if
the group is not Lie and acts faithfully on a connected manifold, the
failure of the no-small-subgroup property, together with connectedness,
produces a nontrivial compact subgroup, and inside the appropriate
totally disconnected corner sits a copy of Z_p. We will not reprove
this classical theorem; it is the standard doorway into the Hilbert-
Smith problem (see Montgomery-Zippin 1955, Lee 1997, or the introduction
of Pardon 2013).
Granting it, the argument is:
Suppose G acts faithfully on a manifold M and is not a Lie group.
Then G contains Z_p, acting faithfully and continuously on the same
M. That contradicts the p-adic exclusion. Hence G is a Lie group.
So from now on the task is: no faithful Z_p action on a connected
manifold, in any dimension. Everything that follows serves that task.
========================================================================
THE PLAN IN ONE PAGE (the lattice contradiction)
========================================================================
Fix an odd integer d >= 3; for an action on an n-manifold we will pick
d > n. Half One constructs, for the sphere S^d, an abelian group
E_d(S^d) and a distinguished element u_d in it, together with a fixed
positive integer L_d, such that TWO things hold.
(L1) E_d(S^d) (x) Q = Q * u_d. Rationally the group is one line.
(L2) For the natural map into the rational line, the image of the
INTEGRAL group lies in (1/L_d) Z * u_d. The integral elements
sit on a lattice of mesh 1/L_d.
Here E_d(S^d) is a Witt group of symmetric forms on a certain category
of sheaves on S^d (Part I), and u_d is the class of the orientation
form. L_d is a power of 2, fixed by d alone, independent of any group
action. Properties (L1), (L2) are the content of Part IV.
Half Two takes a faithful Z_p action and, for every k >= 1, manufactures
p^k integral classes z_0(a), ..., z_{p^k - 1}(a) in E_d(S^d) with two
properties.
(A1) Their rational images are all EQUAL.
(A2) Their rational images ADD UP to 4 u_d.
From (A1) and (A2) each rational class equals (4 / p^k) u_d. By (L2)
each of these lies in (1/L_d) Z * u_d, so 4/p^k is an integer multiple
of 1/L_d, i.e. 4 L_d / p^k is an integer. But 4 L_d is a fixed finite
number and p^k is as large as we want, so for p^k > 4 L_d the number
4 L_d / p^k lies strictly between 0 and 1 and cannot be an integer.
Contradiction. Therefore no faithful Z_p action exists.
Everything in Half One exists to make (L2) legal: it is the statement
that the integral signature group on the sphere does not have slots at
(1/p^k) for large k. Everything in Half Two exists to fill those slots.
========================================================================
ROADMAP
========================================================================
PART I. Witt groups and the duality we use. Self-contained: what a
symmetric form on a category with duality is, how orthogonal
sum and neutral forms give a group, why exact sequences of
categories give exact sequences of Witt groups. Verdier
duality on sheaves is stated, not reproved.
PART II. The compact-source sheaf categories T(X) and P(X). These are
the "test categories". We prove Verdier duality restricts to
them (so forms make sense) and prove the cylinder relation
(so their Witt groups are homotopy invariant).
PART III. Localization: cutting a category along an open set, proving
quotient morphisms are germs, and proving that an open cover
gives a dense inclusion with kernel and cokernel killed by 2.
This is the source of every factor of 2 and of the integer
L_d.
PART IV. The lattice theorem on the sphere: (L1) and (L2). Compare the
continuous category T with the semialgebraic category P, where
signature is an honest integer. The comparison is onto after
inverting 2, and a fixed power of 2 suffices.
PART V. Local tests. From a faithful Z_p action, Newman's theorem
produces a faithful orientation-preserving action on an open
subset of R^n. We build the compact sphere source, the orbit
space, and a degree-one test a : Y -> S^d.
PART VI. Character classes. Decompose the orbit sheaf by characters of
Z_p, build the classes z_i, prove they are natural and
homotopy invariant, move a positive-degree multiple of the
test onto p^k sheets, and run the contradiction.
PART VII. Assembly: the p-adic exclusion and the Lie group theorem,
proved from Parts I-VI.
PART VIII. Appendix on signs. The one place where sign conventions
could silently go wrong, fixed once and checked.
Items you can safely skim on a first reading (everything is used later,
but the logic is clear without the verification): the sign appendix;
the cokernel half of the diagram chases in Part IV; the explicit
Tietze/bump construction in Part V. Items you should NOT skim: the
definition of E_d(S^d) and the two properties (L1), (L2); the identity
sum z_i = 2[A,alpha]; the p^k-sheet equalization.
We now begin.
========================================================================
PART I. WITT GROUPS AND DUALITY
========================================================================
This part is the algebra we will feed the geometry into. Nothing here
needs sheaves; we state it for an abstract linear category so that in
Part II we can just check the axioms. A reader happy with Witt groups
can skim to "1.5 The localization sequence", which is the one imported
theorem we cannot do without.
1.1 Categories with duality.
Let C be an R-linear triangulated category (an additive category with a
built-in notion of "cone" of a map, obeying the usual axioms; the only
example we use is a bounded derived category of sheaves). A DUALITY on
C is the data of
(a) an exact functor # : C -> C ("take the dual");
(b) a natural isomorphism eta : id -> ## ("biduality"), that is, for
every object A a chosen isomorphism eta_A : A -> ##A;
(c) an exactness sign delta in {+1, -1} recording how # acts on the
connecting map of a triangle (this is a convention; see Part VIII).
The slogan: # is a well-behaved "linear duality" with a chosen way to
identify A with its double dual. In Part II, # is Verdier duality and
eta is the Verdier biduality map, proved invertible on the categories we
use.
We need forms of a fixed cohomological degree, so we also allow a SHIFT.
For an integer r, put
#_r = [-r] o # (take the dual, then shift down by r).
The number r is the DEGREE of the forms we are about to build.
1.2 Symmetric forms.
A (nonsingular) symmetric form of degree r in (C, #, eta) is a pair
(A, alpha), alpha : A -> #_r A an isomorphism,
that is its own transpose. "Transpose" is the only subtle word:
the TRANSPOSE of a map f : A -> B is the map
#_r f : #_r B -> #_r A,
and we then use the biduality isomorphisms eta to view #_r A and
#_r B as dual to A and B themselves. Composing these, a map
A -> #_r A can be transposed back to a map A -> #_r A, and a form is
SYMMETRIC when alpha equals this transpose of alpha.
Equivalent, more familiar phrasing (used from Part II on): a symmetric
form of degree r is a pairing
beta : A (x) A -> Omega[r] (Omega the dualizing object)
with beta equal to beta after swapping the two factors (with the Koszul
sign) and with adjoint A -> #_r A an isomorphism. The map alpha is that
adjoint. The two descriptions agree by tensor/Hom adjunction. We will
use whichever is convenient, and Part VIII checks that the sign
conventions are coherent.
An ISOMETRY of forms (A, alpha) -> (B, beta) is an isomorphism f : A -> B
in C with beta f = #_r(f) alpha. Isomorphic forms are the same for
everything below.
The ORTHOGONAL SUM of forms (A, alpha) and (B, beta) is
(A (+) B, alpha (+) beta) of the same degree r,
using the natural dualities on a direct sum.
1.3 Neutral forms and the Witt group.
A form (A, alpha) of degree r is NEUTRAL if it has a LAGRANGIAN: an
object L and a map i : L -> A that fits into a triangle
L --i--> A --#_r i . alpha--> #_r L --> L[1],
compatible with the duality (in the sense of triangular Witt theory).
The clean intuition: a lagrangian is a "totally isotropic direct
half" whose dual is the other half, so the form it defines is
"hyperbolic" and should count as zero. The standard example is the
HYPERBOLIC form
H(L) = ( L (+) #_r L , [[0,1],[delta_r * id, 0]] ),
with lagrangian L; it is neutral.
The WITT GROUP W_r(C) of degree-r forms on C is the abelian group with
generators: isometry classes [A, alpha] of symmetric forms;
relations : [ (A,alpha) (+) (B,beta) ] = [A,alpha] + [B,beta],
[neutral form] = 0.
The negation is realized by multiplying the pairing by -1 (equivalently,
using the additive inverse form). W_r is a Z-graded family as r varies.
1.4 Three facts we will actually use.
(F1) Hyperbolic forms are neutral, so H(L) = 0 in W_r for every L.
(This is what makes "adding a hyperbolic summand" harmless.)
(F2) Scaling a real form by a positive real p > 0 does not change its
class: multiply the underlying object by the inverse square root of
p to get an isometry. Hence in R-coefficient settings we may
normalize constant factors freely as long as they are nonzero.
(F3) Multiplication by 2 is invertible on every morphism group (R is a
field of characteristic zero; in fact we work over R). Balmer's
symmetric-cone construction needs only this and the fact that C is
triangulated with a compatible duality. This is the reason every
later "dense inclusion" loses exactly a factor of 2 and no more.
1.5 The localization sequence (imported).
The one truly external algebra theorem we use is the localization
sequence for triangular Witt groups, proved by Balmer (2000) and in the
general form we need by Balmer-Walter (2002, Theorem 2.1). We state the
consequence.
WITT LOCALIZATION. Let C be a triangulated category with duality in
which multiplication by 2 is invertible on morphism groups and the
enriched octahedral axiom (TR4+) holds. Let A be a full thick
triangulated subcategory stable under the duality, so the Verdier
quotient C/A is again a category with duality. Then for every r there
is a natural long exact sequence
... -> W_r(A) -> W_r(C) -> W_r(C/A) -> W_{r-1}(A) -> W_{r-1}(C) -> ...
and it is natural for exact duality-preserving functors of pairs.
This is the workhorse of Part III and Part IV: it lets us compute the
Witt group of a big category from the group on an open piece and the
group of the quotient "near the closed complement", with a boundary map
shifting degree by one.
What is C/A? It is the Verdier quotient: we formally invert every map
whose cone lies in A. Geometrically, when A is "objects supported in a
compact piece of an open set U", the quotient C/A forgets exactly the
part of an object living inside U and remembers only its behaviour near
the complement of U. Part III makes this precise as "germs".
1.6 What remains.
Parts II-IV supply the examples: C = T(X) (sheaves generated by proper
images of compact polyhedra), # = Verdier duality, and the subcategories
cut out by open sets. The Witt group E_r(X) = W_r(T(X)) is the signature
group of the space X. The whole proof is about E_d(S^d) for one odd d.
We need two further facts about these concrete categories, proved where
they are used:
- the CYLINDER relation (homotopy invariance of W_r), in Part II;
- DENSE DOUBLING (an open cover inclusion has kernel and cokernel
killed by 2), in Part III.
With those, Part IV pins the lattice (L1), (L2), and Part VI fills it.
========================================================================
PART II. THE COMPACT-SOURCE SHEAF CATEGORIES
========================================================================
We build the concrete categories whose Witt groups are the signature
groups. The guiding idea is finiteness of SOURCE, not of stalk: each
object is generated, by finitely many exact operations, from proper
direct images of a constant sheaf on a compact polyhedron. Such objects
can have wild fibres and infinite-dimensional stalks, but Verdier
duality still sees them as finite, and their Witt groups are homotopy
invariant.
2.1 Test spaces.
A TEST SPACE is a locally compact, finite-dimensional, separable
metrizable space that is one of: a finite polyhedron; a homeomorph of
one; an open subset of either; or a finite product of these. The point
and the empty set are test spaces. Verdier duality is only ever applied
on test spaces.
Why the finite-c-soft hypothesis holds automatically. Sheaf duality for
Rf_! and f^! asks that the map f have finite c-soft dimension (the
compactly supported direct image has bounded cohomological dimension on
fibres). For a test space this is free: embed it as a locally closed
subset of R^N (realize each polyhedral factor as a closed Euclidean set,
take the prescribed relatively open pieces, then the product). A
continuous map g : X -> Y into a Hausdorff Y has fibres closed in X,
hence locally closed in R^N; for such an embedded fibre Z the extension
i_! is exact and R Gamma_c(Z; F) = R Gamma_c(R^N; i_! F), which vanishes
above degree N. So every allowed map has finite c-soft dimension,
uniformly, with a bound N independent of the fibre.
Write D(X) = D^b(Sh_R(X)) for the bounded derived category of sheaves of
real vector spaces.
2.2 The four categories.
Fix a test space X.
T^0(X) = the smallest full replete triangulated subcategory of D(X)
containing every Rf_* const_R(B), where B is a compact
finite polyhedron and f : B -> X is any continuous map.
T(X) = the closure of T^0(X) under direct summands in D(X).
For a test space with a chosen semialgebraic model, define P^0(X) and
P(X) the same way but requiring f to be continuous SEMIALGEBRAIC (in the
model). For an open U in a semialgebraic model, "semialgebraic
f : B -> U" means the composite B -> model is semialgebraic; U itself
need not be semialgebraic.
Each generator Rf_* const_R(B) has compact support (its support lies in
the compact image f(B)), is bounded, and is proper-pushed, since B is
compact and X is Hausdorff. Sums, shifts, cones and summands keep this
compact support. Membership in T(X), P(X) is thus a finiteness
condition on the source that survives all our operations.
An ALLOWED MAP for T is any continuous map of test spaces; for P it is a
continuous semialgebraic map (or a restriction/inclusion of relatively
open pieces). The only property we need is that composing an allowed
map with a generator gives another allowed generator, which holds in
both categories.
Essential smallness is harmless: there is a set of compact polyhedra and
of maps from each into X up to isomorphism, the derived category is
locally small, and finite constructions preserve the set of isomorphism
classes. Take representatives and replete subcategories.
2.3 Simplex generators and summands.
LEMMA (simplex generators). Closed simplices suffice in place of all
compact polyhedra; and every object of T(X) (respectively P(X)) is a
direct summand of an object of T^0(X) (respectively P^0(X)). In
particular these categories are thick and triangulated.
Sketch. For a simplex sigma with interior sigma^0 and boundary d(sigma),
the open-complement triangle
(j_!) const_R(sigma^0) --> const_R(sigma) --> (i_*) const_R(d sigma)
--> (j_!) const_R(sigma^0)[1]
builds the constant sheaf on the closed simplex from constants on its
closed faces, by induction on dimension (the interiors of a simplicial
filtration form a disjoint union of open simplices). Pushing along f
expresses const_R(B) using constants on closed simplices of a
triangulation, all semialgebraic in the P case.
For the summand statement: the summands of a full triangulated
subcategory are again triangulated. If A (+) A' = C and B (+) B' = D are
in the subcategory and u : A -> B, then u (+) 0 : C -> D is in the
subcategory (fullness), and a cone on it is Cone(u) (+) B' (+) A'[1]; so
Cone(u) is a summand of an object of the subcategory. Shifts and finite
sums behave the same, and an iterated summand is a summand. Shifts,
sums and summands preserve thickness.
2.4 Verdier duality restricts.
Let a_X : X -> pt. Set the dualizing complex and dual functor
omega_X = a_X^! R, D_X A = R Hom(A, omega_X).
Precisely what we use from the sheaf formalism (locally compact spaces,
maps of finite c-soft dimension; all from Schapira's lecture notes):
- trace: Rf_! has right adjoint f^!, composition gives
f^! omega_Y = omega_X, and the counit is Tr_f : Rf_! omega_X -> omega_Y.
- internal duality: Rf_* D_X A = D_Y Rf_! A, adjoint to evaluation then
trace, and compatible with eta_A : A -> D_X^2 A whenever the terms are
bounded. For an open inclusion j one has j^! = j^{-1}, so duality
commutes with open restriction.
- proper base change, the projection formula, and compact-support
Kuenneth, in their natural forms.
None of this asserts that an ARBITRARY bounded complex is reflexive; we
prove reflexivity only on the generated categories.
THEOREM (duality on the test categories). The Verdier operation
restricts to a triangulated duality on each of T^0(X), T(X), P^0(X),
P(X). For every object A in one of these, D_X A lies in the same
category and eta_A : A -> D_X^2 A is an isomorphism. Moreover there is a
compact K in X off which both A and D_X A vanish.
Sketch. Compute on a closed n-simplex Delta with an orientation of its
interior. The local calculation of compactly supported cohomology of
balls (interior: R in degree n; boundary: half-balls give zero) yields
D_Delta const_R(Delta) = omega_Delta = (j_!) const_R(Delta^0)[n],
and, dually, D of the right side is const_R(Delta). The second
isomorphism uses Rj_* const_R(Delta^0)[n]; the unit
const_R(Delta) -> Rj_* const_R(Delta^0) is an isomorphism because small
relative neighbourhoods meet the interior in something contractible.
Under these identifications the biduality map is that unit, hence
invertible, boundary included.
For a general simplex generator A = Rf_* const_R(Delta) with f proper,
the internal duality formula and the simplex computation give
D_X A = Rf_* omega_Delta. Since omega_Delta is built from constants on
compact subpolyhedra (boundary triangle), D_X A stays in the generated
category. A commutative diagram expressing compatibility of Rf_* with
biduality reduces, under tensor/Hom adjunction, to the fact that both
routes push the same evaluation pairing through the same trace; so
eta_{Rf_*} is invertible whenever eta on the source is.
Finally, "dual bounded and in the category" and "eta invertible" are both
closed under shifts and triangles (naturality of eta on a morphism of
triangles, two-out-of-three), and pass to summands by dualizing the
splitting maps. The simplex-generator lemma gives all four categories.
The common compact support K is the finite union of the compact images
f(Delta) over a generating presentation; open restriction of duality
shows D_X A vanishes off the same K.
2.5 Direct images and open extension.
PROPOSITION (pushforward). For an allowed map g : X -> Y, direct image
induces exact functors g_* : T(X) -> T(Y) (and P(X) -> P(Y)). On these
categories Rg_! = Rg_*, this functor commutes coherently with Verdier
duality and its shifts, and it is compatible with composition. For an
open inclusion j : U -> X the functor is j_! and is fully faithful.
Sketch. If A vanishes outside a compact K in X and i : K -> X, then
A = i_* i^{-1} A and g i is proper, so the comparison Rg_! A -> Rg_* A is
the invertible comparison for the proper map g i; compactness of K
suffices, no polyhedral structure needed. On a generator
Rg_* Rf_* const_R(B) = R(gf)_* const_R(B), again a generator, and
exactness plus preservation of splittings give category membership.
Because both A and D_X A have compact support, internal duality gives
Rg_! D_X A = D_Y Rg_! A naturally, defined by evaluation then trace, and
so compatible with biduality. A pairing beta : A (x) A -> omega_X[-r]
is sent to the pairing on Rg_! A obtained by pushing beta and applying
Tr_g; under tensor/Hom adjunction this equals the pushed duality
comparison, so compatibility with biduality is automatic. Proper
pushforwards compose, and for an open inclusion j, j^! = j^{-1} with
(j_!, j^{-1}) adjunction makes j_! fully faithful. The comparison with
Rj_* applies because test objects on U have compact support in U.
2.6 Shifted forms and the signature groups.
We use cochain shifts and homological indexing:
D_{X,r} = [-r] o D_X, W_r(C) = W^{-r}(C), E_r(X) = W_r(T(X)).
The exactness sign of D_{X,r} is delta_r = (-1)^r; these are the
translated-duality signs of triangular Witt theory, and Part VIII checks
that they are coherent. With these signs a symmetric form of degree r
is equivalently (Part I, 1.2) a pairing
beta : A (x) A -> omega_X[-r], beta o tau = beta,
whose adjoint alpha : A -> D_{X,r} A is invertible. Here tau is tensor
symmetry with the Koszul sign. In particular, on an oriented d-manifold
the orientation identification omega_X[-d] = const_R(X) makes the
multiplication pairing of the constant sheaf a symmetric degree-d form;
when its underlying object lies in a test category it defines a class
there. That class, for X = S^d, will be u_d.
2.7 The symmetric cone.
The next algebraic fact is the engine of homotopy invariance.
LEMMA (symmetric cone). Let (#, eta) be a triangulated duality with
exactness sign delta in {+1,-1} on a category in which multiplication by
2 is invertible on morphism groups. Let u : L -> #L be symmetric, with
no invertibility assumed. In a triangle
L --u--> #L --v--> C --w--> L[1],
the cone carries a nonsingular symmetric form psi : C -> [1]#C for the
translated duality. It is neutral, and it can be chosen so that
psi v = -[1](#w), [1](#v) psi = delta [1](eta_L) w,
with the dual shift comparisons understood. Taking # = D_{X,r+1} this is
a neutral form of degree r.
Sketch. Dualize and rotate the triangle, so its first arrow is
delta #u. Symmetry of u makes the first two terms compatible with
delta eta_L and the identity of #L; complete to a morphism of triangles
with third map psi_0. The signed transpose of psi_0 is another such
completion. Replacing psi_0 by the average
(1/2) ( psi_0 - delta [1](#psi_0) eta_C )
preserves the first two squares (2 is invertible) and imposes symmetry
for the translated biduality -delta eta. The result psi satisfies the
two displayed identities. It is an isomorphism by two-out-of-three for
the morphism of triangles, and it is neutral by construction: rotating
the triangle to start with #L -> C makes its third object L[1], which is
[1]#(#L), and the displayed identities supply the triangular lagrangian.
For # = D_{X,r+1}, [1]D_{X,r+1} = D_{X,r} with exactly the translated
signs, giving the degree.
2.8 The interval and homotopy invariance.
Let I = [0,1], and e_t : {t} -> I for t = 0, 1. Set
H = const_R(I), J = (j_!) const_R((0,1)), E_t = (e_t)_* R,
E = E_0 (+) E_1,
where j : (0,1) -> I. Let C_I be the full thick subcategory generated by
H, E_0, E_1; it is contained in P(I) and T(I), and contains J through the
triangle
J --u--> H --v--> E --d--> J[1]. (interval triangle)
The dualities we use are D_{I,1} J = H, D_{I,1} H = J, D_I E = E; the
first arrow u, read as J -> D_{I,1} J, is symmetric, its pairing being
multiplication J (x) J -> J = omega_I[-1].
LEMMA (products with the interval). For A in T(X) (or P(X)) and B in
C_I, the external product B (x) A lies in T(I x X) (or P(I x X)). For all
s, r the natural comparison
mu_{s,r} : D_{I,s} B (x) D_{X,r} A -> D_{I x X, s+r}( B (x) A )
is an isomorphism, adjoint to the product of evaluations with the tensor
shift signs; it commutes with proper pushforward, duality, and trace.
Hence the product of symmetric pairings is symmetric of the summed degree.
(Proof: verify on generators B = H, E_t and A = Rf_* const_R(Delta),
where external duality is the product-orientation isomorphism on
I x Delta; then close under triangles and summands. The sign of the
comparison is fixed in Part VIII.)
LEMMA (the two endpoint forms). Let (A, alpha) be a symmetric degree-r
form in T(X) or P(X), and let i_t : X -> I x X, x |-> (t,x). There is a
neutral degree-r form on
E (x) A = (i_0)_* A (+) (i_1)_* A
whose two diagonal forms are c (i_0)_* alpha and -c (i_1)_* alpha for a
common nonzero real scalar c, with no off-diagonal terms.
Sketch. First compute the interval boundary. Compactly supported
cohomology of the interval triangle is
0 -> R --x|->(x,x)--> R^2 --(x_0,x_1)|-> x_1 - x_0--> R -> 0,
the last identification using the orientation of I (it is the coboundary
of endpoint values for the pair (I, {0,1})). Proper duality therefore
makes the shifted dual of d a map H -> E with opposite nonzero
coefficients at the two ends; closed-image adjunction Hom(H, E_t) = R
shows these coefficients determine the map. Now form the symmetric
morphism
J (x) A --u (x) alpha--> H (x) D_{X,r} A --mu_{1,r}-->
D_{I x X, r+1}( J (x) A ),
symmetric by the product lemma, with cone triangle
J(x)A -> H(x)A --v (x)1--> E(x)A --d (x)1--> (J(x)A)[1].
The symmetric cone lemma gives a neutral form psi of degree r on E (x) A.
The two slices are disjoint closed subsets, so by closed-image adjunction
every morphism between their supported complexes vanishes: psi is
diagonal. Precomposing with v (x) 1 determines each diagonal, and the
computation of the dual of d (x) 1 above gives coefficients that are a
common nonzero scalar times alpha and its negative. The absolute
normalization is fixed in Part VIII.
THEOREM (homotopy invariance). If h : I x X -> Y is a continuous homotopy
of maps of test spaces, then its endpoints induce equal homomorphisms
h_{0*} = h_{1*} : E_r(X) -> E_r(Y), r in Z;
and likewise for W_r(P(-)) under allowed semialgebraic homotopies. In
particular homotopy equivalences induce isomorphisms, and a nonempty
contractible test space has the Witt groups of a point.
Proof. For a form (A, alpha), the endpoint-forms lemma gives, in the
degree-r Witt group of the cylinder,
[ (i_0)_* A, c (i_0)_* alpha ] + [ (i_1)_* A, -c (i_1)_* alpha ] = 0.
Multiplying a pairing by a positive real scalar changes neither its
isometry class nor its class in W_r (Fact F2). If c < 0 the displayed
relation is the negative of the one with |c|, so in all cases
(i_0)_*[A, alpha] = (i_1)_*[A, alpha].
Push this equality along h via the pushforward proposition; the composite
maps are h_0 and h_1, giving the desired equality on classes, hence on
E_r. Properness of the whole homotopy is irrelevant: every object in the
relation is supported in I x K for a compact K in X. The P case is
identical, with the compact-source maps and homotopy images semialgebraic.
Apply the endpoint equality to the two composites of a homotopy
equivalence and its inverse for the final assertions.
REMARK. At a point, both test categories equal the bounded derived
category of finite-dimensional real vector spaces: a compact polyhedron
has finite-dimensional bounded cohomology, and the point generator with
its finite sums and shifts already generates every such complex. We use
this identification and its elementary Witt computation in Part IV; the
point group is Z in degree multiple of 4 and 0 otherwise (Lemma 4.1).
========================================================================
PART III. LOCALIZATION, GERMS, AND DENSE SUBCATEGORIES
========================================================================
We cut our categories along an open set, identify quotient morphisms with
germs near the complement, and compare the quotients of an open cover.
The open-cover comparison has kernel and cokernel killed by 2, and that
integral factor-of-2 bound is exactly what produces the denominator L_d
in Part IV.
Throughout, B denotes either T or P, and B^0 its triangulated
subcategory before adjoining summands. For U open in X we identify
B(U) with its image in B(X) under extension by zero, which is fully
faithful (Part II, 2.5). An open subset of a semialgebraic model need
not be semialgebraic; every construction below cuts polyhedral sources
inside its inverse image, so this never becomes an extra hypothesis.
3.1 Cutting a finite construction off an open set.
LEMMA (cutoff). Let K contained in O contained in X with K closed in X
and O open. For every A in B^0(X) there is a morphism
s : A' -> A
such that A' is extended from B^0(O), its cohomology vanishes outside a
compact subset of O, and s is an isomorphism on some open neighbourhood
of K.
Sketch. Take first A = Rf_* const_R(Delta), Delta a compact simplex.
The compact set F = f^{-1} K lies in the relatively open f^{-1} O.
Finite subdivision gives a relatively open polyhedral neighbourhood E
of F with
C = closure(E) contained in f^{-1} O, C and C \ E compact polyhedra.
(Choose the subdivision so the closed stars of all vertices meeting F
lie in f^{-1} O; positive distance from F to the complement plus small
mesh does it. Assign vertex value 0 on those stars and 1 elsewhere,
extend linearly to h, and set E = {h < 1/2}, C = {h <= 1/2}; all vertex
values in {0,1} give C = closure(E). If F is empty take E = C = empty.)
For e : E -> Delta the natural map e_! const_R(E) -> const_R(Delta)
induces s : A' = Rf_* e_! const_R(E) -> A. Pushing the open-closed
triangle e_! const_R(E) -> const_R(C) -> i_* const_R(C \ E) along the
restrictions of f shows: both compact sources map into O, so A' is
extended from B^0(O) with support bounded by the compact f(C) in O
(semialgebraic throughout in the P case). Moreover
N = X \ f(Delta \ E) is open and contains K, and s|N is an isomorphism.
The statement is preserved by shifts and finite sums. For cones,
suppose a : A -> B is a morphism between objects where it holds. Choose
a cutoff b : B' -> B over O, invertible on an open N containing K, then
cut off A over the smaller open O cap N to get c : A' -> A. Since A' is
supported in a compact subset of N, for j : N -> X we have A' = j_! j^* A',
and the adjunction Hom(A', B') = Hom(A'|N, B'|N) lifts
(b|N)^{-1}(a|N)(c|N) to a morphism a' : A' -> B'. The same adjunction
with target B gives b a' = a c globally. Completing this square to a
morphism of triangles yields Cone(a') -> Cone(a) invertible on the
intersection of the two invertibility neighbourhoods. Open-extension
fullness identifies a' with a morphism in B^0(O), so its cone is there,
and the union of the two compact support bounds is a compact support
bound inside O. Finite triangulated generation of B^0(X) finishes.
3.2 Recognizing objects supported in an open set.
PROPOSITION (objects supported in an open set). For an open inclusion
j : U -> X, the essential image of j_! : B(U) -> B(X) is exactly those
objects whose cohomology vanishes outside some compact subset of U.
This image is a thick, duality-stable subcategory.
Sketch. An extended object has such a support bound (Part II, 2.5).
Conversely, let S in B(X) vanish off a compact C contained in U, and
write it as an actual summand S --i--> A --r--> S with r i = id_S and
A in B^0(X). By the cutoff lemma, choose s : A' -> A over U invertible
on a neighbourhood N of C. Since S is extended by zero from N,
open-extension adjunction lifts i to i' : S -> A' with s i' = i globally,
so (r s) i' = id_S. Restricting this splitting to U makes S|U a summand
of A'|U in B^0(U); thus S|U is in B(U) and extends back to S.
Finite unions of compact subsets of U are compact in U, and the long
cohomology sequence of a triangle together with stalkwise splitting of a
summand show the support condition is thick. If S vanishes off C, open
restriction of Verdier duality shows D_X S vanishes off C too, giving
duality stability.
3.3 The quotient as a category of germs.
Fix an open inclusion U in X and put K = X \ U. The neighbourhoods of K
(in X) are ordered by shrinking, a filtered system. For A, B in B(X)
define the group of germ morphisms by
colim_{N sup K, N open in X} Hom_{D(N)}( A|N , B|N ),
with D(N) the ambient bounded derived category; neither restricted
object is required to lie in B(N), and composition happens after
restricting two representatives to a common neighbourhood.
THEOREM (germ description). Restriction induces natural isomorphisms
Hom_{B(X)/B(U)}(A, B) = colim_{N sup X\U} Hom_{D(N)}(A|N, B|N)
for all A, B in B(X), respecting composition, open restriction of
duality, and restriction of distinguished triangles. Here B(X)/B(U) is
the Verdier quotient.
Sketch. An object of B(U) vanishes outside a compact C contained in U,
hence vanishes on the neighbourhood X \ C of K; so a morphism with cone
in B(U) becomes invertible in germs, and restriction factors through the
quotient.
Fullness, first for A in B^0(X). Represent a germ by v : A|N -> B|N on
N sup K, and apply cutoff with O = N to get s : A' -> A with A' compactly
supported inside N and s invertible near K. Adjunction extends
v (s|N) to t : A' -> B on X. The cone of s has a compact support bound C
and vanishes on some open M sup K, so it is supported in the compact
C \ M contained in U; by the support proposition it lies in B(U). The
roof A <--s-- A' --t--> B is a quotient morphism with germ v. For
general A in B(X), write it as a summand A --i--> A_0 --p--> A, lift the
germ v(p|N) by the case just done, and compose with i; this gives
fullness for all sources.
Faithfulness. First let t : A -> B be an actual morphism with zero germ.
If A in B^0(X), take N sup K where t is zero and a cutoff s : A' -> A
supported in N; open-extension adjunction makes t s = 0 globally, and s
is invertible in the quotient, so t is zero there. For a summand source
apply this to t p : A_0 -> B and use t = (t p) i. Finally every quotient
morphism is a roof A <--s-- C --t--> B with Cone(s) in B(U), so s is
invertible near K; a zero-germ roof has zero-germ numerator t, hence is
zero in the quotient by the previous case. This proves faithfulness for
all roofs. The functor is restriction; the cutoff choices only proved it
is full and faithful.
3.4 Witt localization.
The support characterization supplies the kernel category.
THEOREM (Witt localization). For every open U in X and every r there is
a natural long exact sequence
... -> W_r(B(U)) -> W_r(B(X)) -> W_r(B(X)/B(U))
--d--> W_{r-1}(B(U)) -> W_{r-1}(B(X)) -> ...
natural for P -> T and for pushforwards: a map g : X -> X' with
g(U) in U' induces a morphism of sequences.
Sketch. Apply Balmer-Walter (2002, Theorem 2.1) with J = B(U),
K = B(X). By the support proposition J is a full thick subcategory
stable under the exact duality, hence saturated. The categories are
essentially small, morphism groups are real vector spaces so 2 is
invertible, and the enriched octahedral axiom (TR4+) holds in a bounded
derived category and passes to full triangulated subcategories and
localizations. With S the class of morphisms with cone in J, the Verdier
quotient S^{-1}K = K/J has full kernel J and the duality preserves S
(dualizing a cone triangle shifts the dual cone). Applying the theorem
to every shifted duality and rewriting W_r = W^{-r} gives the displayed
sequence. Naturality: P -> T preserves support subcategories and
duality, and a compact support bound C contained in U maps to the compact
bound g(C) contained in U', so the kernel subcategories are preserved.
The boundary is the Witt class of the symmetric cone of a representing
symmetric morphism, transported diagram-by-diagram, so independence of
representative gives equality.
3.5 An open cover comparison, killed by 2.
PROPOSITION (open cover is a dense inclusion). If X = U union V with
U, V open, then extension by zero induces a fully faithful exact
duality-preserving functor
B(V)/B(U cap V) -> B(X)/B(U),
whose replete image is dense: every object of the target is an actual
summand of an object in the image. It commutes with P -> T.
Sketch. Extension from V sends B(U cap V) into B(U), so it induces the
functor, exact with duality. Put K = X \ U; then K is contained in V and
K = V \ (U cap V). Open neighbourhoods of K contained in V are cofinal
among all its neighbourhoods in X (intersect any with V), so the germ
theorem on V and on X identifies the morphism groups on the two sides:
full faithfulness. For A_0 in B^0(X), cutoff with O = V and this K gives
A'_0 -> A_0 with cone compactly supported inside U, hence an isomorphism
in B(X)/B(U), while A'_0 comes from B^0(V). Every A in B(X) is a
summand of such an A_0, proving density; fullness and exactness make the
image triangulated and it is duality stable. All functors are the same
open extensions for P and T.
THEOREM (dense doubling). Let A contained in C be a full replete
triangulated subcategory stable under duality, with A dense: every object
of C is a summand of an object of A. Assume triangular Witt groups
exist. Then in every degree the inclusion
iota : W_r(A) -> W_r(C)
has kernel and cokernel annihilated by 2. More precisely there is a
homomorphism d : W_r(C) -> W_r(A) with
iota d = 2 id, d iota = 2 id.
No closure of A under arbitrary summands is required.
Sketch (the doubling construction). Fix the degree and write # for its
duality. Density plus fullness first gives
y (+) y[1] in A, for all y in C.
Indeed, write c = y (+) z with c in A; the endomorphism 0_y (+) id_z of
c is in A by fullness, and its cone is y (+) y[1]. Applying this to
y[-1] also gives y (+) y[-1] in A.
For a nonsingular symmetric form q = (x, alpha) in C, define
D(q) = q (+) q (+) H(x[1]),
H the hyperbolic form. Its underlying object is
x (+) x (+) x[1] (+) #(x[1])
= (x (+) x[1]) (+) (x (+) x[-1]),
using alpha : x = #x and #(x[1]) = (#x)[-1]; by the previous paragraph it
lies in A, and repleteness plus fullness make D(q) a form in A. The
assignment respects isometries and orthogonal sums (apply them to the two
copies of q; hyperbolic respects direct sums).
Neutral descent. Suppose q has a triangular lagrangian a : l -> x, with
lagrangian triangle l --a--> x -> #l -> l[1]. The object z = l (+) l[1]
is in A, so H(z) is neutral already in A. In C the form D(q) (+) H(z)
has lagrangian
L = l (+) l (+) x[1] (+) #l (+) l[1]
(first two factors the two lagrangians of the copies of q; x[1] the
first lagrangian of H(x[1]); in H(z) = H(l) (+) H(l[1]) take the second
lagrangian #l of H(l) and the first l[1] of H(l[1])). Check L in A using
only fullness and triangles: both objects of a (+) 0 : l (+) l[1] ->
x (+) x[1] lie in A, so its cone, and hence
C = #l (+) x[1] (+) l[2],
lies in A; adding l (+) l[1] in A and rearranging gives
C (+) l (+) l[1] = (#l (+) x[1] (+) l) (+) (l[1] (+) l[2]); the second
summand l[1] (+) l[2] is in A, and a split triangle (cone of its split
inclusion, all objects in A) puts #l (+) x[1] (+) l in A; finally
L = (l (+) l[1]) (+) (#l (+) x[1] (+) l) is in A. Now the underlying
object of D(q) (+) H(z), its lagrangian L, and #L all lie in A, and
fullness supplies all arrows of the lagrangian diagram; the inherited
triangulation and duality make it a lagrangian diagram in A. So
D(q) (+) H(z) is neutral in A, and since H(z) is neutral there,
[D(q)] = 0 in W_r(A). Thus D descends to a homomorphism d : W_r(C) ->
W_r(A).
Finally, H(x[1]) is neutral in C, so iota d[q] = 2[q]; if q starts in A,
then x[1] is in A and the same hyperbolic form is neutral there, so
d iota[q] = 2[q]. The first identity kills the cokernel of iota by 2,
the second its kernel.
COROLLARY (open cover bound). The map on Witt groups induced by the open
cover inclusion B(V)/B(U cap V) -> B(X)/B(U) has kernel and cokernel
annihilated by 2, in every degree and for either test category.
Proof. Apply dense doubling to the replete image of the open cover
inclusion.
That a fixed power of 2 (not just inverting 2) suffices, assembly into
L_d, and the whole lattice theorem are the content of Part IV.
========================================================================
PART IV. A BOUNDED LATTICE ON THE TEST SPHERE
========================================================================
Set F_r(X) = W_r(P(X)) and E_r(X) = W_r(T(X)); the inclusion P -> T gives
a natural comparison F_r(X) -> E_r(X). Spheres and their pole
complements carry their usual semialgebraic models. Rationalization
means tensoring with Q; sheaf coefficients stay real. We fix an odd
integer d >= 3 and prove that the rational image of E_d(S^d) sits in a
lattice with ONE denominator L_d, fixed by d alone and independent of any
action, test map, prime or exponent.
4.1 Point groups and two diagram chases.
LEMMA (the point). For every r, the comparison at a point is an
isomorphism, and
F_r(pt) = E_r(pt) = Z if r = 0 mod 4, 0 otherwise.
In degree 0 the isomorphism is signature. On S^0 (two points) both
theories are a direct sum of two copies of the point theory, and the map
to the point is addition.
Sketch. A generator over a point is the finite-dimensional bounded
cohomology complex of a compact polyhedron, and conversely the point
constant with its shifts generates every bounded finite-dimensional
complex (split a complex into cohomology plus a contractible complex):
both point categories are just the bounded derived category of
finite-dimensional real vector spaces.
For a degree-r form, cohomology pairs H^i with H^{r-i}. The complex
splitting identifies the form with the orthogonal sum of these pairings.
Terms with i != r-i occur in hyperbolic pairs (one member is a
lagrangian, the other its shifted dual), so they die in the Witt group.
If r is odd nothing remains; if r = 2k the remaining form is on H^k with
symmetry sign (-1)^{k^2} = (-1)^k from tensor symmetry. For odd k it is
alternating, hence hyperbolic by a symplectic basis; for even k it is a
real symmetric form, and diagonalization with cancellation of one
positive and one negative line gives its class as signature times the
positive line. Signature respects neutral relations: in a lagrangian
triangle at the middle degree, exactness makes the image of H^k(L) its
own perpendicular in H^k(A), which has signature zero. The positive line
has signature one, so the group is infinite cyclic. On the two-point
space, sheaves and dualities split over the two clopen components.
LEMMA (finite diagram chases). In a commutative diagram of exact rows
A1 -> A2 -> A3 -> A4 -> A5
| | | | |
B1 -> B2 -> B3 -> B4 -> B5
suppose the kernels and cokernels of f1, f2, f4, f5 are annihilated by q.
Then the kernel and cokernel of f3 (the middle vertical) are annihilated
by q^3. Second: in a commutative square with horizontal maps i, j, if
both horizontal kernels and cokernels are annihilated by 2 and the
kernel and cokernel of the left vertical f are annihilated by q, then
those of the right vertical g are annihilated by 4q.
Proof idea. Both are finite element chases using exactness: chase
q x through the rows and lift with q, tracking the accumulating factor
of q on the kernel side and the double-lift on the cokernel side. The
square chase multiplies the existing q by 4 because it lifts twice and
uses the annihilating 2 twice. We record the conclusions only; the
constructions are mechanical.
4.2 The bounded comparison.
THEOREM (bounded comparison). For every j >= 0 there is a nonnegative
integer e_j, independent of r, such that the kernel and cokernel of
F_r(S^j) -> E_r(S^j)
are annihilated by 2^{e_j} for every integer r.
Proof. Induct on j. For j = 0 use the point lemma, e_0 = 0. Assume the
bound 2^e for S^{j-1}. Let X = S^j, let U, V be the complements of two
distinct poles, and put I = U cap V. The spaces U, V are
semialgebraically contractible, and I is semialgebraically homotopy
equivalent to S^{j-1} (for j = 1, to two points). By homotopy invariance
the comparisons on U, V are isomorphisms and the comparison on I has
the induction bound in every degree; these homotopies need not be proper
because the pushforward functors use compact supports.
Write Q_r^B(Z, A) = W_r( B(Z)/B(A) ) for A open in Z. Naturality of Witt
localization gives two exact rows
W_r(B(I)) -> W_r(B(V)) -> Q_r^B(V,I) -> W_{r-1}(B(I)) -> W_{r-1}(B(V)).
The four outer comparisons have the uniform induction bound, so the
first diagram chase bounds the comparison on Q_r(V, I) in all degrees,
with exponent 3e.
Next, by the open-cover proposition and dense doubling, the two
horizontal maps in the natural square
Q_r^P(V,I) -> Q_r^P(X,U)
| |
Q_r^T(V,I) -> Q_r^T(X,U)
have kernel and cokernel killed by 2. The square chase then controls the
right-hand comparison up to a bounded additive loss (the horizontals are
2-killed, costing a fixed amount on top of the running exponent).
Finally use the two localization rows
Q_{r+1}^B(X,U) -> W_r(B(U)) -> W_r(B(X)) -> Q_r^B(X,U) -> W_{r-1}(B(U)).
A third five-term chase (cubing the running exponent, as in the first chase)
gives a bound for the middle comparison W_r(B(U)) -> W_r(B(X)); combined
with the known bound on W_r(B(U)) this yields a bound on F_r(S^j) -> E_r(S^j). Only finitely many
exact diagrams are used, and the same exponents serve all r.
REMARK (bookkeeping). The crude exponents produced are, step by step,
3e, then 3e + 2, then 3(3e + 2) = 9e + 6 (the first and third chases
multiply the running exponent by 3; the square step adds the fixed loss
2). Without optimizing, one may take
e_0 = 0, e_j = 9 e_{j-1} + 6, equivalently e_j = 3(9^j - 1)/4.
This depends on the sphere dimension j, works for every degree and every
class, and is what we need. A comparison only after inverting 2 would
NOT suffice: the inclusion Z subset Z[1/2] has cokernel with unbounded
2-power torsion, exactly the bounded denominator we must avoid.
4.3 The rational sphere group.
PROPOSITION (rational sphere). For G_r = F_r or E_r and j >= 0,
G_r(S^j) (x) Q = ( G_r(pt) (x) Q ) (+) ( G_{r-j}(pt) (x) Q ).
In particular, for the fixed odd integer d, both F_d(S^d) (x) Q and
E_d(S^d) (x) Q are one-dimensional rational vector spaces.
Sketch. Let G~_r(Z) = ker( G_r(Z) -> G_r(pt) ) be the reduced group; a
point inclusion splits the projection whenever Z is nonempty. Use the
pole cover X = U union V with I = U cap V. Since G_r(U) -> G_r(pt) is an
isomorphism, G_r(U) -> G_r(X) is split injective, and localization gives
Q_r(X,U) = G~_r(X) with zero next boundary (because G_{r-1}(U) ->
G_{r-1}(X) is injective). Likewise G_r(I) -> G_r(V) is split surjective
(project to the point, use contractibility of V, section by a point of I),
so the other localization sequence identifies its boundary as
Q_r(V,I) = G~_{r-1}(I). The dense map between these relative groups is an
isomorphism after tensoring with Q, by dense doubling. Hence
G~_r(S^j) (x) Q = G~_{r-1}(S^{j-1}) (x) Q.
For S^0 the reduced group is the kernel of addition of two point groups,
i.e. one point group. Induction plus the split point summand gives the
formula. At (r,j) = (d,d) the two summands are 0 and Q, because d is odd
and the point lemma gives G_d(pt) = 0, G_0(pt) = Z.
4.4 Integral signature and the semialgebraic generator.
Orient S^d. The constant real sheaf with its orientation pairing into
omega_{S^d}[-d] determines the orientation classes
u_{F,d} in F_d(S^d), u_d in E_d(S^d),
and the comparison sends u_{F,d} to u_d. Write a bar for the rational
image.
PROPOSITION (signature on a finite semialgebraic relation). Every finite
collection of objects of P(S^d) has finite-dimensional locally constant
cohomology on a common dense open subset of S^d. On a sufficiently small
oriented ball there, a degree-d form has an integral signature on its
degree-zero cohomology. For any finite Witt relation one can choose one
such ball simultaneously for all objects and lagrangian diagrams in the
relation, and the signatures respect it. Consequently u_{F,d} is
nontorsion and
F_d(S^d) (x) Q = Q * u_{F,d},
im( F_d(S^d) -> F_d(S^d)(x)Q ) = Z * u_{F,d}.
Sketch. For a continuous semialgebraic map from a compact polyhedron,
Hardt's semialgebraic local triviality theorem gives a finite
semialgebraic partition of the target over whose pieces the map is
trivial; refining into cells, the top-dimensional cells form an open
dense subset. Each fibre is a compact semialgebraic set, hence has the
finite cohomology of a finite triangulation; proper base change and the
trivializations show a generator has finite-dimensional locally constant
cohomology there.
This passes to the finite constructions defining P. Shifts and summands
are immediate. For a cone the cohomology long exact sequence writes each
cohomology sheaf as an extension of a kernel and cokernel of morphisms of
finite local systems; those are local systems, and on a small ball the
positive Ext groups between finite constant sheaves vanish
(Ext^k_B(const V, const W) = H^k(B; const Hom(V,W)) = 0 for k > 0), so
the extension is locally constant. Take the finite intersection of the
dense opens needed for the finitely many generators.
Choose a ball B in the common open set. Each bounded complex with finite
locally constant cohomology restricts on B to the constant complex of its
cohomology (splitting successive truncation triangles, since the relevant
Ext vanish), so the constant-complex functor identifies that full
subcategory with the bounded derived category of finite-dimensional real
vector spaces. The orientation identification omega_B = const_R(B)[d]
identifies D_{B,d} with ordinary derived linear duality there, compatibly
with evaluation and biduality; the orientation form becomes the positive
form xy. This degree-zero local description holds in both odd congruence
classes of d mod 4.
So a nonsingular degree-d form on A gives an ordinary symmetric form b on
H^0(A)_x for x in B, with integer signature, additive under orthogonal
sums and invariant under isometry. For neutrality, include a lagrangian
triangle before choosing B; writing l : H^0(L)_x -> H^0(A)_x, the middle
exact sequence H^0(L)_x --l--> H^0(A)_x --l^* b--> (H^0(L)_x)^* gives
im(l) = ker(l^* b) = im(l)^{perp}, an ordinary lagrangian, so the middle
signature is zero. Since each Witt equality is a finite combination of
isometries, orthogonal sums and neutral relations, signatures respect any
fixed finite relation.
The class u_{F,d} has signature one on every oriented ball, so if a
nonzero multiple of it vanished a witnessing finite relation would
contradict its signature; hence u_{F,d} is nontorsion and spans the
rational line. For x in F_d(S^d) write x (x) 1 = (a/b) u_{F,d} with
a in Z, b > 0; the difference b x - a u_{F,d} is torsion, so v b x =
v a u_{F,d} actually in the Witt group for some integer v > 0. Pick a
common ball for this finite equality; the signature of x there, call it
sigma, gives v b sigma = v a, hence a/b = sigma in Z. Thus every
integral multiple of u_{F,d} occurs, giving the asserted image equality.
4.5 The fixed sphere lattice.
THEOREM (the fixed sphere lattice). Let d >= 3 be odd, with orientation
class u_d as above. There is a positive power of 2, denoted L_d, such
that
E_d(S^d) (x) Q = Q * u_d,
im( E_d(S^d) -> E_d(S^d)(x)Q ) is contained in L_d^{-1} Z * u_d.
One may take L_d = 2^{e_d}. This denominator is fixed by the sphere
dimension and the comparison; it is independent of every later action,
test map, character partition, prime p and exponent k.
Proof. By the bounded comparison the comparison F -> E is a rational
isomorphism, so the signature proposition identifies its nonzero rational
orientation class with the image of the semialgebraic generator, namely
u_d. Take L_d = 2^{e_d}, which annihilates the comparison cokernel in
degree d. For every x in E_d(S^d) there is y in F_d(S^d) with L_d x equal
to the image of y; the rational image of y is an integral multiple of
u_{F,d} by the signature proposition, so x (x) 1 lies in
L_d^{-1} Z * u_d.
This is exactly (L1) and (L2). What remains, to connect to a group
action, are two further facts (4.6) and then the geometric half of the
proof (Parts V-VII).
4.6 Degree action and integral localization.
Two last properties connect the lattice to a group action: degree compares
sphere tests as multiplication by an integer, and a contractible open set
can be localized away integrally.
PROPOSITION (action of degree). Let d >= 3 be odd. If f : S^d -> S^d is
continuous of degree s in Z, then
f_* ( u_d bar ) = s * u_d bar in E_d(S^d) (x) Q,
and f_* acts on the whole rational group by multiplication by s. More
generally, if N is a connected closed oriented triangulable d-manifold
and f : N -> S^d has degree s, the pushed orientation form has rational
class s * u_d bar.
Sketch. Choose a finite polyhedral model for the source and a finite
semialgebraic triangulation of the target, semialgebraically identifying
its model with S^d. Take a simplicial approximation g to f (after
subdivision, as in Hatcher AT 2C.1); it is semialgebraic and homotopic to
f, and homotopy invariance identifies f_* and g_* on the continuous
theory. The constant sheaf on the source model is an identity-map
generator, so its orientation form is semialgebraic; orient both models by
their identifying homeomorphisms (these preserve fundamental classes and
orientation forms).
Fix a top-dimensional target simplex and, in its interior, a small ball B
avoiding the images of lower-dimensional and collapsed top-dimensional
source simplices. Each source simplex mapping onto the target simplex
contributes one open sheet over B, mapped homeomorphically onto B; these
are all of g^{-1}(B), finitely many, and the restricted map is proper (base
change of a compact-source map). On B the direct image is the constant
vector space with one basis vector per sheet. The pushed pairing is
diagonal, and the trace on one sheet is +1 or -1 according as the sheet
preserves or reverses orientations (the trace identifies source and target
dualizing complexes; relative to the two orientation trivializations the
induced map on local top homology is the orientation sign). Its signature
on B is the sum of these local signs. For a sphere source this sum is the
degree s (local-degree formula, Hatcher 2.30). For a general N, fix y in B
and g^{-1}(y) = {x_1,...,x_t}; excision identifies
H_d(N, N \ g^{-1}(y); Z) = direct sum_i H_d(N, N \ {x_i}; Z),
and the fundamental class [N] maps to the tuple of positive local
generators; the relative map g sends the i-th to epsilon_i times the
orientation generator of H_d(S^d, S^d \ {y}; Z), epsilon_i the orientation
sign of that sheet. Naturality of absolute-to-relative maps plus
g_*[N] = s[S^d] gives s = sum_i epsilon_i. Only interiors of
top-dimensional simplices occur (open balls), so no PL structure of the
triangulation is needed.
This local computation fixes the rational coefficient without a globally
defined generic signature: if the rational image of the pushed class is
rho * u bar, clear denominators and torsion to get a finite Witt relation,
choose its signature ball inside B (possible, the common open set is
dense), and the signature says rho = s. Apply the comparison for E. Since
u_d bar spans the rational group, the whole group is multiplication by s.
PROPOSITION (a contractible open set can be localized away). Let d >= 3
be odd. Let X be a test space and U contained in X a nonempty
contractible open subset. Then the quotient functor gives an INTEGRAL
isomorphism
E_d(X) -> Witt_d( T(X)/T(U) ).
It is natural for continuous g : X -> X' with g(U) contained in U', when U'
is also nonempty contractible open.
Proof. Homotopy invariance and the point lemma give E_d(U) = E_d(pt) = 0
(d is odd, not 0 mod 4). Let i : U -> X and let p_Z : Z -> pt. The
composite E_{d-1}(U) --i_*--> E_{d-1}(X) --(p_X)_*--> E_{d-1}(pt) equals
(p_U)_*, an isomorphism (in degree d-1: if d = 1 mod 4 this is Z, if
d = 3 mod 4 it is 0 -- split injectivity holds in either case). So i_* is
split injective, with retraction ((p_U)_*)^{-1}(p_X)_*. All these maps
exist by compact-support pushforward, proper or not. The relevant segment
of Witt localization is
0 -> E_d(X) -> Witt_d(T(X)/T(U)) -> E_{d-1}(U) --i_*--> E_{d-1}(X).
Exactness plus injectivity of the last map gives the displayed
isomorphism over Z. Naturality: g_* preserves the support subcategory
because g(U) contained in U', so the natural localization sequence gives a
commutative square. No rationalization and no splitting of quotient
idempotents is involved; this is why the character classes below lift
INTEGRALLY.
Part IV is complete: (L1), (L2), degree action, and integral
localization are all available. Parts V and VI use them to extract the
contradiction from a faithful Z_p action.
========================================================================
PART V. PREPARING LOCAL TESTS
========================================================================
Fix n >= 1 and a prime p, and suppose Z_p acts faithfully and jointly
continuously on a connected n-manifold M (no boundary), as in the p-adic
exclusion. The plan: reduce to a faithful orientation-preserving action of
an open subgroup G = Z_p on a connected open O in R^n, thicken it in an odd
dimension d > n, form the orbit space, and build a degree-one map from that
orbit space to S^d. Then a faithful action forces unbounded denominators.
5.1 An invariant chart subset.
PROPOSITION (invariant chart subset). There are an open subgroup G = Z_p
and a nonempty connected G-invariant open subset O of a coordinate chart of
M such that G acts faithfully and preserves orientation on O. In
particular O may be taken as the actual open subset of R^n inside that
chart.
Proof. Use Newman's theorem: a finite-order homeomorphism of a connected
manifold that is the identity on a nonempty open set is the identity
everywhere (Newman 1931; modern proof in Pardon 2019, and the local
reduction is Pardon 2013, 4.1).
First claim: there is x in M such that no open subgroup of Z_p fixes a
neighborhood of x pointwise. Suppose otherwise. With H_j = p^j Z_p and
F_j = interior of Fix(H_j) in M, some F_j is nonempty. Then F_j is
closed: if x is in closure(F_j), H_j fixes x (fixed sets are closed), and
by the supposition some neighborhood N of x is fixed pointwise by H_l.
Compactness of H_j and continuity at H_j x {x} give a coordinate ball B
around x with H_j B contained in N. The saturation N' = H_j B is open and
connected (every translate of B contains x), H_j-invariant, and the
restricted action factors through the finite group H_j/(H_j cap H_l), so
each of its homeomorphisms is finite-order and fixes the nonempty open
N' cap F_j; Newman makes each the identity on N', so x in F_j and F_j is
closed. Connectedness makes F_j = M, contradicting faithfulness. This
proves the claim.
Choose such x, a chart C around x, and a coordinate ball B around x with
closure(B) compact in C. Joint continuity and compactness of closure(B)
give an open subgroup H with H closure(B) contained in C; shrinking H, also
H x contained in B. Then O = H B is open and H-invariant, and connected
since every g B meets B at g x. Its restriction kernel in H is closed. A
nonzero closed subgroup of Z_p is p^j Z_p for some j (take an element of
smallest finite p-adic valuation; its integer multiples are dense in the
subgroup). So a nonzero restriction kernel would be an open subgroup
fixing the neighborhood O of x pointwise, impossible. The chart orients
O; every homeomorphism of connected O either preserves or reverses it, so
every square preserves it. Replace H by G = 2H, an open subgroup = Z_p
with faithful restricted action; for odd p this changes nothing, for p = 2
it has index two.
5.2 Thickening and the one-point compactification.
Fix an odd integer d > n with d >= 3 for the rest of the proof. Give
W = O x R^{d-n} contained in R^d
the product orientation and let G act on the O factor and trivially on the
R^{d-n} factor; this stays faithful and orientation-preserving. Let
W+ = W union {infinity} be the one-point compactification, and let
S^d = R^d union {infinity_amb} with the matching orientation be the compact
test source; its orientation form represents u_d of the lattice theorem.
The choice d > n lets us use the odd-sphere calculation of 5.4.
LEMMA (collapse and compactification). The map
c : S^d -> W+, c(z) = z for z in W, c(z) = infinity for z not in W,
is continuous. The action on W extends to a jointly continuous action on
W+ fixing infinity. Both W+ and its orbit space by any closed subgroup of
G are compact metrizable.
Proof. Preimage of an open subset of W is open in S^d; a basic
neighborhood of infinity is W+ \ K for compact K in W, with preimage the
open S^d \ K. This gives continuity of c for any open W. For compact K in
W the saturation G K is compact (image of G x K); every g in G maps
W+ \ G K into W+ \ K, giving joint continuity at all (g, infinity), while
continuity on G x W is already there. W is locally compact, noncompact,
second countable, so W+ is compact metrizable. If d_0 is a compatible
metric on W+, then d_1(x,y) = max_{g in G} d_0(g x, g y) is a compatible
invariant metric (uniform continuity on the compact G x W+). For a closed
subgroup H of G, d_H(Hx, Hy) = min_{h in H} d_1(x, h y) induces the
quotient topology, positive on distinct orbits by compactness, triangle
inequality by invariance.
5.3 An averaged degree-one test.
LEMMA (averaged degree-one test). After replacing G by an open subgroup
and identifying it with Z_p, put q : W+ -> Y = W+/G and y_inf = q(infinity).
There is a continuous a : Y -> S^d with
deg( a q c ) = 1.
Moreover a q is homotopic on W+ to a degree-one pinch supported in a round
ball compactly contained in W.
Proof. Fix xi in S^d and a map P : R^d -> S^d constant with value xi
outside the unit ball and inducing an orientation-preserving homeomorphism
closure(B^d)/boundary(B^d) -> S^d. Choose z in W and r > 0 with
closure(B(z,r)) in W. Then P_{z,r}(w) = P((w-z)/r), P_{z,r}(infinity) = xi,
is continuous on W+ and P_{z,r} c has degree one. View S^d as the unit
sphere in R^{d+1}. By continuity and compactness a small open subgroup H
of G satisfies sup_{h in H, w in W+} |P_{z,r}(h w) - P_{z,r}(w)| < 1/2.
With Haar probability mu_H set A(w) = integral_H P_{z,r}(h w) d mu_H(h), a
continuous H-invariant R^{d+1}-valued map with |A(w) - P_{z,r}(w)| < 1/2.
Every vector (1-t)P_{z,r}(w) + t A(w) has norm at least 1/2, so normalizing
gives a homotopy from P_{z,r} to the invariant A/|A|, fixing infinity and
constant equal to xi outside the compact H closure(B(z,r)) in W. Replace G
by H. The normalized map descends to a : Y -> S^d, and the homotopy gives
deg(a q c) = 1; faithfulness and orientation preservation persist.
Henceforth G, q, Y, a are fixed as in this lemma.
5.4 Rational orbit cohomology and maps to an odd sphere.
To move the test later, detect rational cohomology on the orbit space by
pulling it back to the compactification. Rational cohomology of compact
metrizable spaces means Cech cohomology (from finite open-cover nerves). It
agrees with constant-sheaf cohomology (Bergfalk-Lupini-Panagiotopoulos,
sections 1.6 and 2); the sheaf-Cech comparison is an isomorphism on
compact Hausdorff spaces and commutes with pullback (Stacks, Lemmas 20.16.2
and 20.15.1). To relate it to nerve cochains, shrink a finite cover so
each closure lies in its original member; on each compact intersection of
closures any locally constant cochain value has finite image and a positive
constancy scale, so for finitely many cochains a fine common refinement
makes every pulled-back value constant on each new intersection. Applying
this to cocycles and primitives identifies the two direct limits, naturally
and with no covering-dimension bound.
LEMMA (rational orbit pullback). Let E be a compact metrizable space with
a continuous G-action and r : E -> E/G its orbit map. For any field k, in
the bounded-below derived category of sheaves,
R r_* const_k(E) = r_* const_k(E).
For k = Q the unit const_Q(E/G) -> r_* const_Q(E) has a natural sheaf
retraction. Consequently r^* : H^j_cek(E/G; Q) -> H^j_cek(E; Q) is
injective for every j.
Sketch. Each orbit is G/G_x for a closed subgroup G_x, a compact
zero-dimensional space (for Z_p: a point, a finite quotient, or Z_p).
Global sections of a sheaf of vector spaces on a compact zero-dimensional
space form an exact functor (lift through an epimorphism on a finite cover,
glue over a finite disjoint clopen refinement), so the higher cohomology of
each orbit vanishes. Then r is closed, separated, with compact fibres, and
bounded-below proper base change (Stacks 20.18.1, 20.18.2) gives
(R^j r_* const_k(E))_y = H^j(r^{-1}(y); const_k) = 0 for j > 0; no covering-
dimension bound on E or E/G enters.
For an open U in E/G and a locally constant phi : r^{-1}(U) -> Q, define
A_U(phi)(r(x)) = integral_G phi(g x) d mu_G(g).
This is independent of the representative x; for fixed x the locally
constant g |-> phi(g x) factors through some finite quotient G/p^l G, so
the average is rational. It is locally constant in x: for each g,
continuity gives a product neighborhood of (g, x) on which phi(h x') is
constant, and finitely many group neighborhoods cover G; intersecting the
corresponding neighborhoods of x shows phi(h x') = phi(h x) for all h and
all nearby x', hence equal averages. An invariant locally constant
function descends because r is an open quotient map. These operators
commute with restriction and fix pulled-back functions, giving the sheaf
retraction. Derived global sections, using R r_* const_Q(E) = r_* const_Q(E),
make the cohomology pullback split injective; the Cech-sheaf comparison
gives the stated conclusion.
Let gamma_d in H^d_cek(S^d; Q) be the rational orientation class. On an odd
sphere, rational triviality upgrades to nullhomotopy after a positive-degree
postcomposition.
LEMMA (killing a rationally trivial sphere map). Let d >= 3 be odd. If K
is compact metrizable and v : K -> S^d satisfies v^* gamma_d = 0, then there
is a positive integer s and a degree-s map D_s : S^d -> S^d for which D_s v
is nullhomotopic.
Sketch. The case K = empty is trivial. Choose a finite open cover
{U_i} of K on each of whose members v has image of Euclidean diameter < 1/4,
and x_i in U_i. For a simplex sigma of the nerve N U, choose a point in the
corresponding intersection; all v(x_i) for vertices of sigma are within 1/4
of its image, so their convex combinations are nonzero and
b(sum_i t_i i) = (sum_i t_i v(x_i)) / |sum_i t_i v(x_i)| defines a continuous
b : |N U| -> S^d. A subordinate partition of unity gives a nerve map
p_U : K -> |N U| with v homotopic to b p_U.
For finite covers of a compact space, Cech cohomology is the direct limit of
the cohomologies of their nerves, the map to the limit being pullback by
p_U. Since the image of b^* gamma_d is zero, some finite refinement V with
simplicial refinement map rho : |N V| -> |N U| satisfies
(b rho)^* gamma_d = 0 in H^d(|N V|; Q), and p_U is homotopic to rho p_V.
We have reduced to b rho : P -> S^d from a finite complex P = |N V|.
Now use d odd, d >= 3. Using based cellular degree maps, form the mapping
telescope S = Tel( S^d --D_2--> S^d --D_3--> S^d --D_4--> ... ). Serre
finiteness gives finite pi_j(S^d) for j > d (Serre 1951, V.3 Prop 3), and
his degree-map assertion (Serre 1953, V.1 Cor 2 to Prop 1): if q x = 0 in a
homotopy group, postcomposition with a degree-2q map kills x. The telescope
has CW type, and its homotopy groups are direct limits of those of its
spheres (maps and homotopies from spheres have bounded height, and each
finite subtelescope retracts onto its last sphere). Any element of
pi_j(S^d), j > d, has finite order q; a long enough subsequent product of
the degrees 2,3,4,... is divisible by 2q, and its map is homotopic to a
composite through a degree-2q map, hence kills that element. So all higher
direct limits vanish; in degree d the limit is
colim(Z --2--> Z --3--> Z -> ...) = Q, and below d it is zero. Thus S has
type K(Q, d). (The same odd-sphere telescope appears in
Dranishnikov-Ferry-Weinberger 2003, proof of Prop 4.5.)
Let j : S^d -> S include the first sphere, and choose gamma_S in
H^d(S; Q) with j^* gamma_S = gamma_d (j_* in degree d is Z into Q; universal
coefficients identify this restriction with Hom(Q,Q) -> Hom(Z,Q), an
isomorphism). Then (j b rho)^* gamma_S = (b rho)^* gamma_d = 0, so
Eilenberg-MacLane representability on the finite complex P makes j b rho
nullhomotopic. A nullhomotopy has compact image, so the continuous height
function on the telescope is bounded there and the image lies in a finite
subtelescope; retracting that subtelescope onto its last sphere gives a
nullhomotopy of D_s b rho with s > 0 the product of the finitely many
displayed degrees. Composing with p_V and v homotopic to b rho p_V proves
the lemma.
5.5 Concentrating on finite quotient sheets.
Faithfulness supplies a region with finite quotient sheets for every
quotient order, even with fixed points and nontrivial stabilizers; the
degree-one test a stays fixed while the region varies.
LEMMA (finite quotient sheets). For every integer k >= 1, put m = p^k and
zeta = exp(2 pi i / m). There is a nonempty open V contained in
Y \ {y_inf} such that the finite quotient map
W/(p^k G) -> W/G = Y \ {y_inf}
over V is a disjoint union of m homeomorphisms onto V. There is a locally
constant function
beta : q^{-1}(V) -> {1, zeta, ..., zeta^{m-1}} in S^1
satisfying
beta(g w) = zeta^{g bar} beta(w), g bar in G/p^k G = Z/m.
Proof. Some x in W has stabilizer G_x contained in p^k G. Otherwise the
classification of closed subgroups of Z_p would force p^{k-1}G contained in
G_x for every x in W, contradicting faithfulness. Put H = p^k G, Z = W/H,
and let z be the image of x; Z is a Hausdorff orbit space, and the
stabilizer of z for the finite action of G/H is (G_x + H)/H = 0. Let tau be
the element induced by 1 in G.
Choose pairwise disjoint open neighborhoods U_j of tau^j z, 0 <= j < m, and
set N = intersection_{j=0}^{m-1} tau^{-j} U_j; then the translates tau^j N
are pairwise disjoint. If r : Z -> W/G is the finite quotient and V = r(N),
openness of r makes V nonempty open, and r^{-1}(V) = disjoint union_{j} tau^j
N, each restriction r : tau^j N -> V an open continuous bijection hence a
homeomorphism. Pulling back under W -> Z partitions q^{-1}(V) into
relatively clopen subsets; give the pullback of tau^j N the constant value
zeta^j. The resulting beta is locally constant, and the cyclic permutation
of sheets proves the multiplier identity.
5.6 Moving the test into a prescribed open set.
LEMMA (extending a homotopy from a closed subset). Let Y_0 be compact
metrizable, F_0 contained in Y_0 closed, and f : Y_0 -> S^d continuous.
Every homotopy H : F_0 x [0,1] -> S^d starting at f|F_0 is the restriction
of a homotopy on Y_0 starting at f.
Proof. Put I = [0,1] and paste H with f on the closed set
C = (F_0 x I) union (Y_0 x {0}) in Y_0 x I. View the result as R^{d+1}-
valued and extend its d+1 coordinates by the Tietze theorem to
E : Y_0 x I -> R^{d+1}. The open set N = { |E| > 1/2 } contains C, and
E/|E| extends the pasted sphere map to N. Compactness of I gives an open
U containing F_0 with U x I contained in N (complement of the projection of
the compact (Y_0 x I) \ N). Choose continuous lambda : Y_0 -> [0,1] equal
to one on F_0 and zero off U. Then H~(y,t) = E(y, t lambda(y)) / |E(y, t
lambda(y))| is defined everywhere (the argument lies in U x I when y in U,
and is (y,0) in C otherwise), is continuous, starts at f, and agrees with H
on F_0 x I.
PROPOSITION (moving a test into a prescribed open set). For every nonempty
open V contained in Y \ {y_inf}, set F = Y \ V. There are a positive
integer s, a degree-s map D_s : S^d -> S^d, and h : Y -> S^d with
h homotopic to D_s a, h|F identically b, some b in S^d.
In particular h(y_inf) = b and q^{-1} h^{-1}(S^d \ {b}) is contained in
q^{-1}(V) contained in W.
Sketch. First show (a|F)^* gamma_d = 0. Choose a round ball with closure in
the nonempty open q^{-1}(V) in W. The pinch of the averaged-test lemma can
be moved to this ball by a homotopy on W+: connect the two centers by a path
in connected open W (path connected), whose compact image has positive
distance from R^d \ W; transport a small fixed radius along it, shrinking
and enlarging radially at the ends. The family P_{z(t), r(t)} is
continuous, all radii positive, all closed balls in one compact subset of W,
hence continuous at infinity (constant value xi). So a q is homotopic to a
map constant on q^{-1}(F); thus the pullback of (a|F)^* gamma_d to q^{-1}(F)
is zero. This preimage is a closed invariant compact subset of W+ with
orbit space F, so the orbit pullback lemma makes (a|F)^* gamma_d = 0. By the
killing lemma, some D_s a|F is nullhomotopic. Apply homotopy extension to
this nullhomotopy and f = D_s a on Y; its endpoint is the required h.
Finally y_inf in F and every point of F maps to b, giving the preimage
inclusion.
REMARK (the source of a restricted test). Let t : Y -> X be continuous into
a compact Hausdorff test space X, and write f = t q c : S^d -> X, b = t(y_inf).
The source over X \ {b} is exactly q^{-1} t^{-1}(X \ {b}) contained in W,
with its given orientation and G-action. The restriction of f to this
source is proper, being the base change of the proper f to an open subset of
X, so its oriented direct-image pairing is the restriction of the pairing
pushed from S^d. These manifold sources are what carry duality in Part VI;
Y enters only through ordinary sheaf direct image. In the move-test
proposition this source lies over the finite quotient sheets of 5.5 when V
is chosen there; the positive degree s may depend on V and k.
Part V is complete. We have the fixed data c, q, Y, a and, for every
nonempty open V in Y \ {y_inf}, a positive-degree postcomposition of a
concentrated into V.
========================================================================
PART VI. CHARACTER CLASSES AND THE CONTRADICTION
========================================================================
Fix n >= 1 and an odd d > n with d >= 3, as in Part V. We use its data
W = O x R^{d-n} contained in R^d, c : S^d -> W+,
q : W+ -> Y = W+/G, a : Y -> S^d,
where O in R^n is connected open, G = Z_p acts faithfully and preserves
orientation on W, and deg(a q c) = 1. Put y_inf = q(infinity), and take
the orientation of S^d used for u_d. All sheaves and pairings are over R;
the symbol C denotes the real coefficient plane with its complex scalar
operations. We attach integral classes to finite partitions of the
character set, prove they are natural under maps and homotopies of tests,
use the finite quotient sheets to equate their rational images, and then
compare with the fixed lattice.
6.1 Characters and proper direct images.
Let Lambda = union_{j>=0} mu_{p^j} in S^1 be the group of complex roots of
unity of p-power order. Define the orbit sheaf H = q_* const_C(W+). Let T
be the action of 1 in G on H, with the convention (T s)(x) = s((-1)x), and
let M_lambda denote multiplication by lambda in C. Put
H_lambda = ker(T - M_lambda), H_D = (+)_{lambda in D} H_lambda.
LEMMA (character coproducts). There are canonical isomorphisms
R q_* const_C(W+) = H = (+)_{lambda in Lambda} H_lambda.
For every continuous t : Y -> X into a compact Hausdorff space and every
D contained in Lambda, the natural map
(+)_{lambda in D} R t_* H_lambda -> R t_* H_D
is an isomorphism in the bounded-below derived category of real sheaves on
X; the sum on the left is a coproduct in the ambient derived category.
Sketch. Apply the orbit pullback lemma to W+ with k = R (so const_C is the
sum of two real constant sheaves); additivity gives the first isomorphism,
with no dimension bound on Y. The stalk of H at an orbit is the complex
vector space of locally constant functions on G/H; the pullback of any such
function to G factors through some G/p^j G (a finite clopen partition on
which it is constant refines by cosets of a common open subgroup). So the
representation is the union of finite cyclic representations, and each
decomposes into T-eigenspaces because z^{p^j} - 1 has distinct complex
roots. Kernels and sheaf direct sums are checked on stalks, giving the
second isomorphism (an algebraic sheaf sum; a section over a noncompact
open set need not have finite character support).
For the derived sum, use compactness. If K is compact Hausdorff and
(F_lambda) a family of sheaves, every section of (+)_lambda F_lambda on K
involves only finitely many indices (locally so by sheafification; a finite
subcover makes the index set finite on K), so global sections commute with
direct sums. The same on every closed subset makes direct sums of SOFT
sheaves soft (extend the finitely many component sections separately), so
soft resolutions and their acyclicity on compact Hausdorff spaces give
H^r( K, (+)_lambda F_lambda ) = (+)_lambda H^r( K, F_lambda ),
r >= 0,
(the compact case of Schapira, Props 11.2.4, 11.2.5, 11.2.8). Apply this to
K = t^{-1}(x) and the restricted eigensheaves: proper base change identifies
the stalk cohomology of the derived sum with these isomorphisms. All the
derived images have common lower bound zero, hence stay bounded below, and
countable sheaf sums are exact; termwise sums compute coproducts in the
derived category (Stacks 13.33.5). This proves both assertions.
6.2 The compact-test data and real orthogonality.
A COMPACT TEST is a compact test space X as in Part II. For a continuous
t : Y -> X whose value b_t = t(y_inf) has a basis of contractible open
neighborhoods, write
f_t = t q c : S^d -> X, A_t = R f_{t*} const_C(S^d),
alpha_t : A_t = D_{X,d} A_t,
where alpha_t is the pushforward of the sphere orientation pairing with
coefficient form B(z,w) = Re(z wbar). Then A_t lies in T(X) (two copies of a
compact-polyhedron generator), and the form is supplied by Part II's duality
and pushforward.
Set X_t^o = X \ {b_t}. Since c sends S^d \ W to infinity, the preimage
f_t^{-1}(X_t^o) lies in W and equals (t q)^{-1}(X_t^o) under the identity on
W; the restricted map to X_t^o is proper (base change over an open set of
the proper f_t). Composition and open base change give
A_t | X_t^o = ( R t_* H ) | X_t^o,
and the pairing on this open set is the direct image of the orientation
pairing on the actual oriented manifold f_t^{-1}(X_t^o) contained in W, so
duality is used only on the manifold and on the test space, never on Y.
For D contained in Lambda put B_D = ( R t_* H_D ) | X_t^o. Each B_D is a
retract of A_t|X_t^o, hence bounded with the same amplitude bounds; its
Verdier dual is bounded too (a retract of D_{X_t^o}(A_t|X_t^o)). These
statements live in the ambient derived category on X_t^o; B_D need not lie
in a test subcategory there. Bounded and bounded-below categories sit fully
faithfully in the ambient category (Stacks 13.11.3, 13.11.6), so the
coproduct property there detects the morphisms between bounded objects used
below.
LEMMA (real orthogonality). If D, E in Lambda are disjoint, the cross
pairing between B_D and B_E induced by alpha_t|X_t^o is zero. In
particular the projectors of any finite partition of Lambda are
self-adjoint.
Sketch. Write A = A_t|X_t^o, alpha = alpha_t|X_t^o, and for an endomorphism
R of A let R^dagger = alpha^{-1} D_{X_t^o,d}(R) alpha be its adjoint. The
orientation-preserving G-action on the proper manifold map acts by isometries
of alpha (naturality of orientation and proper duality), and complex scalar
multiplication by a unit preserves B. With J = M_i,
J^dagger = -J, M_lambda = Re(lambda) id + Im(lambda) J,
T^dagger = T^{-1}, M_lambda^dagger = M_{lambda^{-1}} (lambda in Lambda).
Fix lambda not in E. On B_{lambda}, T - M_lambda = 0. Let R_E be the
endomorphism of B_E induced by T^{-1} - M_{lambda^{-1}} on H_E; it is
invertible (on the mu-summand its inverse is multiplication by
(mu^{-1} - lambda^{-1})^{-1}), and these componentwise inverses assemble to
one endomorphism of the algebraic sheaf sum, hence to R_E^{-1} on B_E after
Rt_* and restriction. For inclusions i_lambda : B_{lambda} -> A and
i_E : B_E -> A, adjointness gives
0 = D_{X_t^o,d}(i_E) alpha (T - M_lambda) i_lambda
= D_{X_t^o,d}(R_E) D_{X_t^o,d}(i_E) alpha i_lambda.
The first factor is invertible, so the cross pairing from B_{lambda} to
D_{X_t^o,d} B_E vanishes; this also covers mu = conj(lambda) != lambda,
since the complex scalar operators separate these two real subrepresentations.
By the character-coproduct lemma and exact open restriction, B_D is the
derived coproduct of the B_{lambda}, lambda in D; a morphism out of a
coproduct is determined by its composites with the coprojections, so the
cross pairing from B_D vanishes. Symmetry gives the opposite order, which
is exactly the statement that every projector is self-adjoint.
6.3 Integral classes without splitting quotient idempotents.
The projectors above live on the punctured target. Their germs turn them
into integral classes on the whole target, using the integral localization
of 4.6 so that NO idempotent need actually split.
Fix a finite partition Lambda = D_0 disjoint-union ... disjoint-union
D_{m-1}. Choose a contractible open neighborhood U of b_t and set
Q_U(X) = T(X)/T(U).
The open projectors of 6.2 define endomorphisms e_i of the image of A_t in
Q_U(X) by the germ theorem, since X_t^o contains X \ U. In this quotient
they satisfy
e_i e_j = delta_ij e_i, sum_i e_i = id, e_i^dagger = e_i,
by real orthogonality (only endomorphisms of the existing object A_t are
asserted; the projector need not extend to a summand over b_t).
DEFINITION (character classes). Let
l_U : E_d(X) -> Witt_d( Q_U(X) )
be the integral isomorphism of 4.6. For 0 <= i < m define z_i(t) in E_d(X)
uniquely by
l_U( z_i(t) ) = [ A_t, alpha_t ] + [ A_t, alpha_t (2 e_i - id) ].
This is a class in the ordinary quotient Witt group: the second form is
nonsingular and symmetric because 2 e_i - id is a self-adjoint involution
(it is +id on a selected summand and -id on its complement; if such a
summand existed, the sum would cancel the complement and double the selected
part). An image object for e_i is NOT part of the definition.
The definition is independent of U. For U_1, U_2 choose a contractible
U_0 contained in U_1 cap U_2. The quotient functors Q_{U_0}(X) ->
Q_{U_j}(X) carry the projectors to each other by their germ descriptions and
commute with localization from E_d(X); since both l_{U_j} are injective, the
lifts via U_0, U_1, U_2 coincide.
PROPOSITION (sum of the character classes). Let t : Y -> X be a compact test
map with a contractible open neighborhood basis at b_t. For every finite
partition Lambda = D_0 disjoint-union ... disjoint-union D_{m-1} the
character classes satisfy the INTEGRAL identity
sum_{i=0}^{m-1} z_i(t) = 2 [ A_t, alpha_t ] in E_d(X).
Proof. Work in Q_U(X) and abbreviate (A_t, alpha_t) to (A, alpha). On
A^{direct-sum m} put
Q = diag( 2 e_0 - id, ..., 2 e_{m-1} - id ),
Q_0 = diag( id, -id, ..., -id ).
Let P_j be the permutation matrix interchanging the 0th and jth copies
(P_0 = id) and E_j the diagonal matrix with e_j in every diagonal entry.
Define S = sum_{j=0}^{m-1} E_j P_j. All E_j commute with the permutation
matrices, and using only the projector identities we get S^2 = id and
S^dagger = S (adjoint for alpha^{direct-sum m}). Moreover
S^dagger Q_0 S = sum_j E_j P_j Q_0 P_j = Q,
because the i-th diagonal entry is e_i - sum_{j != i} e_j = 2 e_i - id and
all off-diagonal entries vanish. Thus S is an isometry from the sum of the
twisted forms to (A, alpha) perp (A, -alpha)^{perp (m-1)}. The sum of their
Witt classes is (2 - m)[A, alpha]. Adding the m untwisted terms in the
definition gives 2[A, alpha]. Injectivity of l_U proves the identity.
6.4 Covariance and homotopy.
PROPOSITION (covariance on compact tests). Let t : Y -> X be a compact test
map and let g : X -> X' be a continuous map of compact tests. Suppose X has
a contractible neighborhood basis at b_t and X' has one at
b_{gt} = g(b_t). For every FIXED finite partition of Lambda,
g_* z_i(t) = z_i(gt) in E_d(X').
Sketch. Choose a contractible open U' around g(b_t) and then a contractible
open U around b_t with U contained in g^{-1}(U'). Pushforward takes an
object supported in a compact subset of U to one supported in a compact
subset of U', so Part II pushforward and Witt localization give a
duality-preserving quotient functor
gbar_* : Q_U(X) -> Q_{U'}(X')
and a commuting square on Witt groups with the maps l_U, l_{U'}. Composition
of proper direct images identifies the pushed total form with
(A_{gt}, alpha_{gt}).
Check the endomorphisms. Represent a projector e_i by a quotient roof,
agreeing with the character projector on some open neighborhood N of X \ U.
Shrink N so N is contained in X \ {b_t} and the roof's denominator is
invertible there. Define
N' = X' \ g(X \ N).
Since X is compact, g(X \ N) is closed; since X \ N is contained in U and g(U)
contained in U', N' contains X' \ U'; also g^{-1}(N') contained in N and
g(b_t) notin N'. Over N' the pushed roof is computed entirely where the
original projector agrees with its character description. Open base change
and composition identify that description with the projector on
R(gt)_* H for D_i; every upstairs preimage avoids infinity, so the open
identification of 6.2 identifies it with the projector for gt on A_{gt}.
The germ theorem gives equality of these endomorphisms in Q_{U'}(X'). The
quotient functor thus carries the two forms in the definition to their
counterparts for gt, and applying the integral localization isomorphisms
proves the covariance identity.
PROPOSITION (homotopy of character classes). For a fixed partition of
Lambda, homotopic maps t_0, t_1 : Y -> S^d satisfy z_i(t_0) = z_i(t_1) in
E_d(S^d).
Proof. The value at y_inf may move during the homotopy; combine two
sufficiently close stages into a single compact test and recover them by
homotopic projections. Consider the compact test
C_d = { (x_0, x_1) in S^d x S^d : <x_0, x_1> >= 0 }.
It is compact semialgebraic, hence a finite polyhedron, with a contractible
open neighborhood basis at each point. At the boundary, the differential of
F(x_0, x_1) = <x_0, x_1> on S^d x S^d is nonzero where F = 0 (its derivative
in the tangent direction x_0 at x_1 equals 1), so C_d is a smooth 2d-manifold
with boundary, with local Euclidean or half-space charts and small balls as
the required basis. Its test forms have degree d, inherited by pushforward
from S^d (target dimension 2d).
The projections pi_0, pi_1 : C_d -> S^d are homotopic by
(x_0, x_1, r) |-> ((1-r) x_0 + r x_1) / |(1-r) x_0 + r x_1|, 0 <= r <= 1,
whose denominator squared is (1-r)^2 + r^2 + 2r(1-r)<x_0, x_1> >= 1/2.
Let H : Y x [0,1] -> S^d be the given homotopy. Uniform continuity on its
compact domain gives a finite subdivision 0 = r_0 < ... < r_N = 1 such that
each pair t^{(j)} = (H(-,r_j), H(-,r_{j+1})) : Y -> C_d lands in C_d, with
distinguished value having the neighborhood basis just verified. By
covariance and homotopy invariance (Part II),
z_i( H(-,r_j) ) = pi_{0*} z_i(t^{(j)}) = pi_{1*} z_i(t^{(j)})
= z_i( H(-,r_{j+1}) ).
Iteration over the finitely many stages proves the assertion.
6.5 Equal parts on the finite quotient sheets.
Fix k >= 1, put m = p^k, and let zeta = exp(2 pi i / m). Choose a
transversal D_0 for the cosets of mu_m in Lambda, and set
D_i = zeta^i D_0, 0 <= i < m.
These are disjoint and partition Lambda. By the finite-sheets lemma there
are a nonempty open V contained in Y \ {y_inf} and a locally constant
function beta : q^{-1}(V) -> S^1 with
beta(g x) = chi(g) beta(x), chi(1) = zeta,
for the character chi : G -> mu_m.
PROPOSITION (equality for a concentrated test). Let k >= 1, and let V and
its character multiplier beta be as above. If h : Y -> S^d is constant on
Y \ V, then the classes for the cyclic partition satisfy
z_0(h) = z_1(h) = ... = z_{m-1}(h) in E_d(S^d).
Proof. Let b be the constant value. Since y_inf notin V it equals b_h. The
source of the restricted test over S^d \ {b} lies in q^{-1}(V) contained in W.
Multiplication by beta on its coefficient sheaf is defined (beta locally
constant) and is a real isometry (|beta| = 1). With our convention for T,
T M_beta = M_{zeta^{-1}} M_beta T.
So multiplication by beta takes the lambda-eigensheaf to the
zeta^{-1} lambda-eigensheaf and cyclically permutes the subsums in the
cyclic partition. After the proper restricted manifold pushforward it
remains an isometry of (A_h, alpha_h)|S^d \ {b} and conjugates e_i to
e_{i-1}, indices mod m.
The isometry and its inverse define quotient morphisms (germ theorem). Its
conjugation identity makes the twisted forms for e_i and e_{i-1} isometric
in the quotient; their untwisted terms agree too. The integral isomorphism
of 4.6 and the definition of the character classes now give z_i(h) =
z_{i-1}(h), hence all are equal.
6.6 The contradiction on the invariant chart.
THEOREM (contradiction on the invariant chart). For every n >= 1 and prime
p, an orientation-preserving continuous Z_p-action on a connected open
subset of R^n has nonzero kernel.
Proof. Suppose the action on such a connected open O in R^n is faithful.
Choose an odd integer d > n with d >= 3 and form the data of Part V:
W = O x R^{d-n}, c, q : W+ -> Y, and a : Y -> S^d with deg(a q c) = 1 (the
averaged-test lemma may replace G by an open subgroup; its action stays
faithful and orientation-preserving).
First fix the lattice. The lattice theorem gives a positive power of 2, L_d,
with
E_d(S^d) (x) Q = Q u_d, im( E_d(S^d) -> E_d(S^d)(x)Q ) in L_d^{-1} Z u_d,
depending only on the sphere comparison. Choose k so large that
m = p^k > 4 L_d, then choose V, beta and the cyclic partition of 6.5. These
give integral classes z_i(a) in E_d(S^d).
Write a bar for rational image. By the move-test proposition there are an
integer s > 0 and a map h, constant on Y \ V, with h homotopic to D_s a,
where D_s : S^d -> S^d has degree s. Thus equal parts, homotopy, and
covariance give
D_{s*} z_i(a) = z_i(D_s a) = z_i(h) = z_j(h) = D_{s*} z_j(a).
On the rational line Q u_d the map D_{s*} is multiplication by s (degree
action). Since s > 0, all the z_i(a) bar are equal. This cancels s only as
a nonzero rational scalar: the original classes z_i(a) remain integral and
their rational images lie in the same fixed lattice; the bound L_d is
unchanged, whatever the size or prime factors of s.
Finally compute the sum. The coefficient form B is the orthogonal sum of
two positive real lines, so [A_a, alpha_a] = 2 (a q c)_* u_d, and deg(a q c)
= 1 gives [A_a, alpha_a] bar = 2 u_d bar. By the sum-of-classes identity,
sum_{i=0}^{m-1} z_i(a) bar = 4 u_d bar, hence z_i(a) bar = (4/m) u_d bar
in E_d(S^d) (x) Q.
Each z_i(a) is integral, so z_i(a) bar lies in L_d^{-1} Z u_d. The lattice
inclusion would force 4 L_d / m in Z. But 0 < 4 L_d / m < 1 by choice of k.
Contradiction. This applies to every prime, including p = 2.
========================================================================
PART VII. ASSEMBLY: THE TWO HEADLINE THEOREMS
========================================================================
Every ingredient is now in place. We assemble the p-adic exclusion and
then the Hilbert-Smith conjecture itself.
7.1 The p-adic exclusion.
THEOREM (p-adic exclusion). Let n >= 1 and let p be prime. Every jointly
continuous action of Z_p on a connected Hausdorff second-countable
topological n-manifold without boundary has nonzero kernel. Equivalently,
the action factors through a finite quotient of Z_p.
Proof. Suppose the action on a connected n-manifold M (no boundary) is
faithful. The invariant-chart proposition of 5.1 produces a faithful,
orientation-preserving action of an open subgroup G = Z_p on a connected
G-invariant open subset O of one coordinate chart, i.e. on a connected open
subset of R^n. This contradicts the contradiction theorem 6.6. Hence the
original kernel is nonzero.
The kernel is closed, being the intersection of the closed stabilizers of the
points of M. Every nonzero closed subgroup of Z_p is open: if it contains an
element x of p-adic valuation j, it contains the closure of the integer
multiples of x, which is p^j Z_p, an open subgroup. So the kernel is open,
and the action factors through the finite quotient Z_p / kernel.
(The chart reduction used none of compactness, global orientability or
triangulability of M, and permitted arbitrary stabilizers and fixed points;
all of that was arranged in Part V. The odd stabilization d > n was the only
dimensional choice, and it treated both odd congruence classes of d mod 4.)
7.2 The Hilbert-Smith conjecture.
THEOREM (Hilbert-Smith). Let n >= 1. A locally compact second-countable
Hausdorff group acting faithfully and jointly continuously on a connected
Hausdorff second-countable topological n-manifold without boundary is a Lie
group.
Proof. The classical reduction of Hilbert-Smith to p-adic exclusion is the
no-small-subgroups structure theory of Gleason and Montgomery-Zippin (1952),
extended by Yamabe (1953): a locally compact group which is NOT a Lie group
contains a topologically embedded copy of Z_p for some prime p (see the
introduction of Pardon 2013, or Lee 1997, or Montgomery-Zippin 1955).
Apply it to the given locally compact Hausdorff second-countable group G. If
G were not a Lie group, it would contain a closed subgroup Z_p, which would
then act faithfully and jointly continuously on the same n-manifold by
restriction --- joint continuity and faithfulness persist. That contradicts
the p-adic exclusion. Therefore G is a Lie group. Dimension zero is
immediate: a connected 0-manifold is a point, so a faithful acting group is
trivial and hence Lie.
REMARK (almost-periodic homeomorphisms). As in Pardon's Conjecture 1.4, the
theorem has a dynamical consequence. For a manifold M as above, a
homeomorphism f is almost periodic if the cyclic subgroup { f^j : j in Z } has
compact closure K in Homeo(M). Then some positive iterate of f is the
time-one map of a continuous real flow: there are m >= 1 and a continuous
homomorphism Phi : R -> Homeo(M) with Phi(1) = f^m. Indeed K is a compact
abelian group acting faithfully and jointly continuously on M; evaluation on
a countable dense subset embeds K in a countable product of copies of M, so K
is second countable; by the theorem K is a compact abelian Lie group; its
identity component K^0 is a torus of finite index, so f^m in K^0 for some m,
and the surjective exponential map of the torus places f^m on a one-parameter
subgroup, giving Phi.
This completes the proof of the Hilbert-Smith conjecture in every finite
dimension. What remains is the appendix fixing sign conventions.
========================================================================
PART VIII. APPENDIX ON SIGNS
========================================================================
Parts II-IV leaned several times on a specific sign ("the exactness sign",
"the product-shift sign", "the (-1)^r in the cone matrix"). This appendix
fixes ONE coherent convention and checks those signs, so no reader has to
worry that a hidden parity was mismanaged. The translated-duality
convention is that of Woolf 2008, section 2.1; the cone identities are
Woolf's Lemma 2.1 (their delta-version is the symmetric-cone lemma of Part
II). Only complex signs are involved, so everything commutes with
restriction and with the derived-functor comparisons of Parts II-IV.
8.1 Hom, shifts, biduality.
For a cochain complex A put A[j]^i = A^{i+j} and d_{A[j]} = (-1)^j d_A. If
a has degree i in A, write a[j] for the element of degree i-j in A[j]
corresponding to it. For a fixed complex K, the Hom complex DA = Hom(A,K)
has differential d(F) = d_K F - (-1)^n F d_A for |F| = n. The shift
comparison is
kappa_j : D(A[j]) -> (DA)[-j],
kappa_j(F)(a) = (-1)^{ j n - j(j-1)/2 } F(a[j]), F in D(A[j])^n.
For j = 1 this is the factor (-1)^n. It is a chain map: raising n by one
changes the exponent by j, matching the target shift's differential sign,
while the other Hom differential term picks up the source shift's (-1)^j;
these are the two terms of the Hom differential. Iterating the one-step
comparison gives exponent n + (n-1) + ... + (n-j+1) for positive j, and
negative j by inversion. We use these iterated comparisons whenever dual
functors and shifts are identified.
The unshifted biduality is the tensor-symmetry evaluation
eta_A(a)(F) = (-1)^{ |a| |F| } F(a),
a chain map (substitute the two Hom differentials); its two evaluation signs
make D(eta_A) eta_{DA} = id_{DA} whenever the duality is defined. These
comparisons make the unshifted dual a +1-exact duality: dualizing a cone
triangle and identifying its shifts gives the dual triangle with positive
final arrow.
Identify D_j = [j]D with Hom(-, K[j]) via the output-shift comparison
f[j](a) = f(a)[j] (no extra sign); both differential expressions equal
(-1)^j d_K f - (-1)^n f d_A for shifted degree n. Tensor symmetry into K[j]
gives the shifted biduality eta^{(j)}, whose expression relative to the
unshifted biduality is
eta^{(j)} = (-1)^{ j(j+1)/2 } eta.
Indeed, on a of degree i, moving the inner j-shift through the outer Hom
contributes (-1)^{ ji + j(j+1)/2 }; in the evaluation, a dual element now has
its degree changed by j, contributing (-1)^{ji}; the two i-dependent factors
cancel. The exactness sign of D_j is (-1)^j.
From tensor/Hom adjunction, a morphism A -> Hom(A, K[j]) is symmetric for
this biduality exactly when its pairing A (x) A -> K[j] is invariant under
Koszul interchange. With K = omega_X and j = -r this proves the pairing form
of 2.6 and its agreement with W_r = W^{-r}. If X is an oriented d-manifold,
the target omega_X[-d] is the constant sheaf in degree zero, and
1 (x) 1 |-> 1 is symmetric and positive in an oriented local trivialization;
no parity sign attaches to this orientation form.
8.2 Evaluation of a product.
Fix the convention for combining output shifts by
K[p] (x) L[q] -> (K (x) L)[p+q],
k[p] (x) l[q] |-> (-1)^{ q |k| } (k (x) l)[p+q],
|k| the degree in the unshifted K. It is a chain map (first differential term
picks up q; the second exponents differ by 2p), and associative (three
factors give q|k| + t|k| + t|l| either way).
Let f, g have degrees l, n in D_{I,s}B, D_{X,r}A; their degrees before the
output shifts are l-s, n-r. For homogeneous b, a the product map of the
interval-products lemma evaluates as
mu_{s,r}(f boxtimes g)(b boxtimes a)
= (-1)^{ n|b| + r(|b| + l - s) } f(b) boxtimes g(a),
followed by the product trace. The exponent n|b| moves g past b; the first
unshifted output has degree |b| + l - s, and the output shifts -s, -r give the
second exponent. This derives the formula without any finiteness or local
constancy of the stalks of A.
The product-shift-sign lemma (used to dualize the interval boundary): with
gamma_s(B) = [-s] kappa_1 : D_{I,s}(B[1]) -> (D_{I,s}B)[-1] and the first-
factor comparisons sigma, epsilon introducing no sign, the two routes to the
suspended total dual satisfy
gamma_{s+r}(B (x) A) mu~_{s,r} = (-1)^r mu_{s,r}^{B,A}[-1] lambda.
The scalar (-1)^r is independent of A, of its cohomological degrees, and of
the point of I. (Discrepancy: (l+n-s-r) - (l-s) - n = -r mod 2, all degree
dependence cancelling.) This is exactly the suspension discrepancy used
when dualizing d (x) 1.
8.3 Normalizing the endpoint calculation.
Recall H, J, E of 2.8 and the interval triangle
J --u--> H --v--> E --d--> J[1]. Choose the orientation identification so
the connecting map is the difference map x |-> (x, x) then (x_0, x_1) |->
x_1 - x_0 (2.8). Proper duality to the point identifies the shifted dual of
d with
[1](D_{I,1} d) : H -> E, h |-> (-h|_0, h|_1),
and the symmetric-cone identity psi_E v = -[1](D_{I,1} d) gives
psi_E = diag(1, -1) : E -> D_I E,
with no off-diagonal maps (disjoint endpoint supports) and diagonal
coefficients forced by precomposition with v (the restriction of v to each
endpoint is the identity). Thus cohomology determines the full
derived-category morphism.
For a degree-r form (A, alpha) on X, the product cone of the cylinder
morphism, the cone identity, naturality of the external-duality comparison,
and the product-shift-sign lemma give
psi_r = (-1)^r diag( (i_0)_* alpha, -(i_1)_* alpha )
: E (x) A -> D_{I x X, r}(E (x) A),
a neutral degree-r form on E (x) A.
Check it by precomposing with v (x) 1; combining the shifted interval and
total dualities uses the product-shift-sign lemma, and desuspending the first
dual output leaves the output-shift convention unchanged, so exactly the
factor (-1)^r appears.
Reversing the orientation, or changing all cone normalizations coherently,
may negate BOTH entries of the cone matrix; their RELATIVE sign stays
opposite. Since the form is neutral, both normalizations give the same
equality of the two endpoint Witt classes, which is the only sign
consequence used in the homotopy-invariance theorem. This closes the sign
bookkeeping: the exactness sign delta_r = (-1)^r (2.6), the translated
biduality eta^{(j)} = (-1)^{j(j+1)/2} eta, the product-shift discrepancy
(-1)^r, and the cone factor (-1)^r are all the same convention viewed four
ways.
========================================================================
SOURCES
========================================================================
The proof is self-contained given the background cited here; each item is
used for one identifiable job.
Triangular Witt groups, localization, dense doubling:
P. Balmer, Triangular Witt Groups I (K-Theory 2000);
P. Balmer, C. Walter, A Gersten-Witt spectral sequence for regular
schemes (Ann. Sci. ENS 2002, Thm 2.1).
Used: Witt axioms (Part I), Witt localization (3.4), dense doubling (3.5).
Sheaf duality:
M. Kashiwara, P. Schapira, Sheaves on Manifolds (1990);
P. Schapira, An Introduction to Categories and Sheaves (2023).
Used: Verdier formalism, finite c-soft dimension, internal duality,
soft resolutions (Part II; 6.1).
Cylinder relation and translated-duality signs:
J. Woolf, Witt groups of sheaves on topological spaces
(Comment. Math. Helv. 2008).
Used: cone/homotopy invariance, sign conventions (Parts II, VIII).
Semialgebraic triviality and integral signature:
R. Hardt (Amer. J. Math. 1980); M. Coste, Semialgebraic geometry (2000);
Hardt-Lambrechts-Turchin-Volic (Algebr. Geom. Topol. 2011).
Used: finite partition, locally constant cohomology, signature (4.4).
Degree, nullhomotopy, and the odd-sphere telescope:
J.-P. Serre (Ann. of Math. 1951, 1953); Eilenberg-Steenrod (1952);
Dranishnikov-Ferry-Weinberger (Ann. of Math. 2003);
A. Hatcher, Algebraic Topology (local degree, simplicial approximation).
Used: 4.6 and the killing lemma of 5.4.
Cech/sheaf cohomology, proper base change, bounded derived categories:
The Stacks Project (Tags 01FD, 09V2, 09V4, 05RR, 0A5L);
Bergfalk-Lupini-Panagiotopoulos (Forum Math. Pi 2024).
Used: 5.4 and 6.1.
Structure theory, reduction, and rigidity:
A. Gleason (1952); D. Montgomery, L. Zippin (1952, 1955); H. Yamabe
(1953); J. S. Lee (1997); J. Pardon (JAMS 2013; PSPM 2019);
M. H. A. Newman (1931).
Used: the reduction of Part 0, the chart reduction and rigidity of
Part V (5.1), and the assembly of Part VII.
This completes the document. It proves the Hilbert-Smith conjecture in
every finite dimension, with the complete argument: the p-adic exclusion in
all dimensions and every prime, and the classical passage to the Lie group
theorem.
END.