, ,

Higher analogues of Reidemeister torsion and Ray-Singer analytic torsion were developed by J. Wagoner, J.R. Klein, the author, M. Bismut, J. Lott, W. Dwyer, M. Weiss, E.B. Williams, S. Goette and many others ([20] , [17] , [16] , [3] , [9] , [2] , [10] , [11] , [13] ).

This paper develops higher torsion from an axiomatic viewpoint. There are three main objectives to this approach:

(1) Simplify the computation of these invariants.
(2) Isolate the key properties of higher torsion.
(3) Explain the theorems relating higher Franz-Reidemeister torsion, Miller-Morita-Mumford (tautological) classes and higher analytic torsion classes.

The following two theorems are examples of results which will make much more sense from the axiomatic viewpoint.

Theorem 0.1 (Hain,I,Penner). The higher Franz-Reidemeister torsion invariants for the Torelli group

τ_{2 k} (T_{g}) \in H^{4 k} (T_{g}; R)

are proportional to the Miller-Morita-Mumford classes.

This had been conjectured by J.R. Klein [18] . The precise proportionality constant was computed in [13] . We will see that this theorem is an example of the uniqueness theorem for even higher torsion invariants.

The next theorem is Theorem 0.2 in [10] . See [2] , [10] , [11] for more details.

Theorem 0.2 (S. Goette). Suppose that

p : M \to B

is a smooth bundle with closed oriented manifold fiber

X

ℱ

is a Hermitian coefficient system on

M

and

H^{*} (X; ℱ)

admits a

π_{1} B

invariant metric. Suppose further that there exists a fiberwise Morse function on

M

. Then the Chern normalizations (from [2] ) of the higher analytic torsion classes

T_{2 k} (E) \in H^{4 k} (B; R)

and the higher Franz-Reidemeister torsion class are defined and agree up to a correction term which is a multiple of the transfer to

B

of the Chern character of the vertical tangent bundle of

M

T_{2 k} (E) = τ_{2 k}^{c h} (E) + ζ^{'} (- 2 k) r k (ℱ) t r_{B}^{E} (c h_{4 k} (T^{v} M)) .

This theorem, together with the uniqueness theorem for odd higher torsion invariants proved below, suggests that nonequivariant higher analytic torsion classes are odd torsion invariants. Recently, Sebastian Goette has claimed that he can prove his theorem in general, i.e., without the existence of a fiberwise Morse function.

In this paper we define a higher torsion invariant to be a characteristic class

τ \in H^{4 k} (B; R)

of “unipotent” smooth bundles

E \to B

which satisfies two axioms.

We show that each such invariant is the sum of even and odd parts

τ = τ^{+} + τ^{-}

The main theorem (Theorem 3.3 below) is:

Theorem 0.3. Nontrivial even and odd torsion invariants

τ^{+}, τ^{-}

exist in degree

4 k

for all

k > 0

and they are uniquely determined up to scalar multiples.

The uniqueness statement is simply a reflection of the fact that, given the axioms, the higher torsion invariant is easy to compute in many cases. We will carry out this computation for any unipotent smooth bundle pair

(E, \partial_{0} E) \to B

which admits a fiberwise Morse function

(E, \partial_{0} E) \to (I, 0)

with the additional property that the critical points have distinct critical values.

To prove the existence we show that there are two linearly independent higher torsion theories given by the higher Miller-Morita-Mumford classes

M_{2 k} (E) = t r_{B}^{E} ((2 k)! c h_{4 k} (T^{v} E)) \in H^{4 k} (B; Z)

and the higher Franz-Reidemeister (FR) torsion invariants

τ_{2 k} (E) \in H^{4 k} (B; R)

Using basic properties of higher FR-torsion proved in [13] and [15] , in particular the framing principle, it is easy to show that

τ_{2 k}

satisfies the axioms. Basis properties of the transfer map imply that

M_{2 k}

also satisfies these axioms.

It is easy to see that

M_{2 k}

is an even higher torsion invariant, i.e., it is trivial when the fiber is a closed odd dimensional manifold. However,

τ_{2 k}

has both even and odd components. The uniqueness theorem implies that

τ_{2 k}

and

M_{2 k}

are proportional whenever the fiber is a closed oriented even dimensional manifold, for example an oriented surface. In order to determine the proportionality constant it suffices to compute one example.

Bismut and Lott [3] showed that nonequivariant analytic torsion classes are trivial for bundles with closed even dimensional fibers. Thus, we believe that they are odd higher torsion invariants. S. Goette's theorem above says that, assuming the existence of a fiberwise Morse function,

T_{2 k}

is proportional to the odd part of

τ_{2 k}

And the proportionality constant is 1 if they are normalized in the same way.

An earlier version of this paper was entitled “Axioms for higher torsion II.” The present paper also incorporates relevant elements from my lecture notes [14] . Thus this paper replaces both of these earlier works.

I am in great debt to John R. Klein and E. Bruce Williams for their help in completing the final crucial steps in the proof of the main theorem. I also benefitted greatly from conversations with Sebastian Goette, Xiaonan Ma and Wojciech Dorabiala.

1 Preliminaries

We consider smooth fiber bundles

F \to E - \to p B

where

E

and

B

are compact smooth manifolds,

p

is a smooth submersion and

F

is a compact orientable manifold with or without boundary. In the boundary case there is a subbundle

\partial F \to \partial^{v} E \to B

E

. We call

\partial^{v} E

the vertical boundary of

E

. (The boundary of

E

is the union of

\partial^{v} E

and

p^{- 1} (\partial B)

.) We assume that

B

is connected. We assume that the action of

π_{1} B

F

preserves some orientation of

F

We will assume that the bundle

E \to B

is unipotent in the sense that the rational homology of its fiber

F

is unipotent as a

π_{1} B

-module. In other words,

H_{*} (F; Q)

has a filtration by

π_{1} B

submodules so that the subquotients have trivial

π_{1} B

actions.

In particular,

π_{1} B

does not permute the components of

F

. Note that unipotent

π_{1} B

modules form a Serre category. In fact, it is the Serre category generated by the trivial modules.

Proposition 1.1. If

E \to B

is a unipotent bundle where

F

has boundary

\partial F

then

H_{*} (\partial F; Q)

and

H_{*} (F, \partial F; Q)

are also unipotent

π_{1} B

modules. In particular, the vertical boundary

\partial^{v} E \to B

is a unipotent bundle.

Proof. By Poincaré duality, $H_{*} (F; Q) \sim = H^{*} (F; \partial F; Q)$ is a unipotent $π_{1} B$ module.
Its dual $H_{*} (F, \partial F; Q)$ must also be unipotent. Since unipotent modules form a Serre category, the long exact homology sequence of $(F, \partial F)$ implies that $H_{*} (\partial F; Q)$ is also unipotent. □

Let

T^{v} E

denote the vertical tangent bundle of

E

. This is the bundle of all tangent vectors which lie in the kernel of

T p : T E \to T B

. The Euler class of the bundle

e (E) \in H^{n} (E, \partial^{v} E; Z)

is simply the Euler class of

T^{v} E

The transfer

\begin{matrix} t r_{B}^{E} : H^{*} (E; Z) \to H^{*} (B; Z) \end{matrix}

(1.1)

is given by

t r_{B}^{E} (x) = p_{*} (x \cup e (E))

where

p_{*} : H^{* + n} (E, \partial^{v} E; Z) \to H^{*} (B; Z)

is the push-down operator (given over

R

by integrating along fibers).

If the orientation of the fiber

F

is reversed, both

e (E)

and

p_{*}

change sign. Thus, the transfer is independent of the choice of orientation of

F

. For the basic properties of the transfer see [1] . The main property that we need is that, for closed fibers

F

t r_{B}^{E} = (- 1)^{n} t r_{B}^{E} .

So, rationally,

t r_{B}^{E} = 0

n = dim F

is odd.

2 Axioms

We define a higher torsion invariant (in degree

4 k > 0

) to be a real characteristic class

τ (E) \in H^{4 k} (B; R)

for unipotent smooth bundles

E \to B

with closed orientable fibers satisfying the additivity and transfer axioms described below.

When we say that

τ

is a “characteristic class” we mean it is a natural cohomology class. I.e.,

τ (f^{*} E) = f^{*} (τ (E)) \in H^{4 k} (B^{'}; R)

f^{*} E

is the pull-back of

E

along

f : B^{'} \to B

. Naturality implies that

τ

is zero for trivial bundles:

τ (B \times F) = 0

2.1 Additivity

E = E_{1} \cup E_{2}

where

E_{1}, E_{2}

are unipotent bundles over

B

with the same vertical boundary

E_{1} \cap E_{2} = \partial^{v} E_{1} = \partial^{v} E_{2}

then the Additivity Axiom says that

\begin{matrix} τ (E) = \frac{1}{2} τ (D E_{1}) + \frac{1}{2} τ (D E_{2}) \end{matrix}

(2.1)

where

D E_{i}

is the fiberwise double of

E_{i}

This wording of the Additivity Axiom comes from Ulrich Bunke.

2.2 Transfer

Suppose that

p : E \to B

is a unipotent bundle with closed fiber

F

and

q : D \to E

is an oriented

S^{n}

bundle (associated to an

S O (n + 1)

bundle over

E

). Then the Transfer Axiom says that the higher torsion invariants

τ_{B} (D) \in H^{4 k} (B; R)

and

τ_{E} (D) \in H^{4 k} (E; R)

are related by the formula:

\begin{matrix} τ_{B} (D) = χ (S^{n}) τ (E) + t r_{B}^{E} (τ_{E} (D)) \end{matrix}

(2.2)

Note that

χ (S^{n}) = 2

0

depending on whether

n

is even or odd respectively.

2.3 Examples

As stated in the introduction, two examples of higher torsion invariants are the higher Miller-Morita-Mumford classes

M_{2 k} (E)

and the higher Franz-Reidemeister torsion invariants

τ_{2 k} (E)

The Miller-Morita-Mumford classes, for closed fiber

F

, are given by

M_{2 k} (E) = t r_{B}^{E} ((2 k)! c h_{4 k} (T^{v} E))

where

c h_{4 k} (T^{v} E) = \frac{1}{2} c h_{4 k} (T^{v} E \otimes C)

. Although this is an integral cohomology class (for

k > 0

) we consider it as a real characteristic class. This invariant is defined for any smooth bundle with closed orientable fiber

F

. If

n = dim F

is odd then twice the transfer map

t r_{B}^{E}

is zero.

Proposition 2.1.

M_{2 k} (E) = 0

for closed odd dimensional fibers

F

. □

Theorem 2.2.

M_{2 k}

is a higher torsion invariant for every

k \geq 1

The higher FR torsion invariants ([13] , [15] )

τ_{2 k} (E, \partial_{0} E) \in H^{4 k} (B; R)

are defined for any relatively unipotent bundle pair

(E, \partial_{0} E) \to B

. By this we mean that the vertical boundary

\partial^{v} E

is a union of two subbundles

\partial_{0} E, \partial_{1} E

with the same vertical boundary

\partial_{0} E \cap \partial_{1} E = \partial^{v} \partial_{0} E = \partial^{v} \partial_{1} E

and that the rational homology of the fiber pair

(F, \partial_{0} F)

(E, \partial_{0} E)

is a unipotent

π_{1} B

module.

Theorem 2.3. The higher FR torsion invariants

τ_{2 k}

are higher torsion invariants for unipotent bundles with closed manifold fibers.

These two theorems will be proved later.

J-M. Bismut and J. Lott [3] constructed even differential forms on

B

called analytic torsion forms. In some cases these are closed and topological (i.e., independent, up to exact forms, of the metric and horizontal distribution used to defined the form).

For example, if

F

is a closed oriented manifold and

π_{1} B

acts trivially on

H_{*} (F; Q)

, then they obtain a (nonequivariant) analytic torsion class

T_{2 k}^{B L} (E) \in H^{4 k} (B; R) .

They showed that

Proposition 2.4.

T_{2 k}^{B L} (E) = 0

for closed even dimensional fibers

F

We also have the following theorems of X. Ma and U. Bunke.

Theorem 2.5 (Ma [19] ).

T_{2 k}

satisfies the transfer axiom.

Theorem 2.6 (Bunke [6] ). Let

E \to B

be the

S^{2 n - 1}

bundle associated to an

U (n)

bundle

ξ

over

B

. Then

T_{2 k}^{B u n k e} (E) = \frac{(4 k + 1)!}{2^{4 k} (2 k)!} ζ (2 k + 1) c h_{4 k} (ξ)

Remark 2.7. Bunke uses a different normalization of the analytic torsion. We should multiply by

(2 π i)^{- 2 k}

to get the Bismut-Lott normalization:

T_{2 k}^{B L} (E) = (- 1)^{k} (2 π)^{- 2 k} \frac{(4 k + 1)!}{2^{4 k} (2 k)!} ζ (2 k + 1) c h_{4 k} (ξ)

As we noted in the introduction, S. Goette has extended the definition of the higher analytic torsion class to the case when

π_{1} B

acts orthogonally on

H_{*} (F; Q)

, i.e., preserving some metric. However, we need to extend it to the unipotent case.

3 Statement

We give the statement of the main theorem. We begin with the following elementary observations.

Lemma 3.1. For each

k > 0

the set of all higher torsion invariants

τ

of degree

4 k

is a vector space over

R

Proof. The axioms are homogeneous linear equations in $τ$ . □

Lemma 3.2. If

τ

is a higher torsion invariant then so is

(- 1)^{n} τ

where

n = dim F

considered as a function of the bundle

E \to B

Proof. We need to show that $(- 1)^{n} τ$ satisfies the axioms: $(- 1)^{n} τ (E) = (- 1)^{n} \frac{1}{2} τ (D E_{1}) + (- 1)^{n} \frac{1}{2} τ (D E_{2})$ $(- 1)^{m + n} τ_{B} (S (ξ)) = (- 1)^{n} χ (S^{m}) τ (E) + (- 1)^{m} t r_{B}^{E} (τ_{E} (S (ξ))) .$ The additivity axiom (the first equation) is the same as before. The transfer axiom (the second equation) is the same as before if both $n = dim F$ and $m$ are even. If one or both are odd then the terms on the right with the wrong sign are zero since $χ (S^{m}) = 0$ for odd $m$ and $t r_{B}^{E} = 0$ for odd $n$ . □

These lemmas imply that higher torsion invariants can always be expressed as a sum of even and odd parts:

τ = τ^{+} + τ^{-}

where

τ^{+} = τ^{+} = \frac{τ + (- 1)^{n} τ}{2}, τ^{-} = τ^{-} = \frac{τ - (- 1)^{n} τ}{2} .

The even torsion invariant can only be nontrivial for even dimensional fibers and the odd torsion invariant can only be nontrivial for odd dimensional fibers.

The main theorem of this paper is the following.

Theorem 3.3 (Main Threorem). Nontrivial even and odd torsion invariants exist in degree

4 k

for all

k > 0

and they are unique up to a scalar factor.

Corollary 3.4. Every even torsion invariant is a scalar multiple of

M_{2 k}

and every odd torsion invariant is a scalar multiple of the odd part

τ_{2 k}^{-}

of the higher FR-torsion

τ_{2 k}

The scalar multiples can be determined as follows. Let

λ

be the universal

U (1) = S O (2)

bundle over

C P^{\infty}

. Let

S^{1} (λ) \to C P^{\infty}

be the circle bundle associated to

λ

and

S^{2} (λ) \to C P^{\infty}

the associated

S^{2}

bundle (the fiberwise suspension of

S^{1} (λ)

Given any odd torsion theory

τ^{-}

in degree

4 k

we have

τ^{-} (S^{1} (λ)) \in H^{4 k} (C P^{\infty}; R) \sim = R

Therefore,

τ^{-} (S^{1} (λ)) = 2 s_{1} c h_{4 k} (λ)

for some

s_{1} \in R

. The statement is that

τ^{-}

is uniquely determined by the scalar

s_{1}

Similarly,

τ^{+} (S^{2} (λ)) = 2 s_{2} c h_{4 k} (λ)

. And

τ^{+}

is uniquely determined by the scalar

s_{2}

. For example, we have the following calculations which will be explained later.

Proposition 3.5.

M_{2 k} (S^{2} (λ)) = 2 (2 k)! c h_{4 k} (λ)

Thus,

s_{2} = (2 k)!

for the higher Miller-Morita-Mumford classes

M_{2 k}

(and

s_{1} = 0

Proposition 3.6.

τ_{2 k} (S^{n} (λ)) = (- 1)^{k + n} ζ (2 k + 1) c h_{4 k} (λ)

where

ζ (s) = \sum_{m > 0} 1 / m^{s}

is the Riemann zeta function.

So,

s_{n} = \frac{1}{2} (- 1)^{k + n} ζ (2 k + 1)

for the higher FR torsion invariants

τ_{2 k}

The uniqueness of even torsion gives us the following.

Corollary 3.7 ([15] ). If

E \to B

is a unipotent bundle with closed even dimensional fibers then

τ_{2 k} (E) = \frac{(- 1)^{k} ζ (2 k + 1)}{2 (2 k)!} M_{2 k} (E) .

Theorem 0.1 in the introduction is a special case of this corollary.

The uniqueness of odd torsion can now be expressed as follows.

Corollary 3.8. Any odd torsion invariant is a scalar multiple of the odd part of higher Franz-Reidemeister torsion which is given by

τ_{2 k}^{-} = τ_{2 k} - \frac{(- 1)^{k} ζ (2 k + 1)}{2 (2 k)!} M_{2 k} .

Theorem 0.2 in the introduction says that analytic torsion classes are odd torsion invariants on certain bundles. We expect that the same formula should hold in general.

The rest of this paper is devoted to the proof of the main theorem. We show that, given the values of the scalars

s_{1}, s_{2}

above, any higher torsion invariant can be computed in sufficiently many cases to determine it completely.

4 Extension to relative case

In order to compute the higher torsion invariant

τ (E)

we need to cut

E

into simpler pieces and compute the relative torsion of each piece. To do this we need to extend

τ

first to the case when

F

has a boundary and then to the case of a unipotent bundle pair

(F, \partial_{0} F) \to (E, \partial_{0} E) \to B

4.1 Higher torsion in the boundary case

Suppose that

E \to B

is a unipotent smooth bundle with fiber

F

a compact orientable manifold with boundary. By Proposition 1.1 the vertical boundary

\partial^{v} E

is also unipotent. And it follows from the Mayer-Vietoris sequence that the fiberwise double

D E

is also unipotent. A higher torsion invariant

τ

can now be extended to the boundary case by the formula

τ (E) : = \frac{1}{2} τ (D E) + \frac{1}{2} τ (\partial^{v} E) .

We will show that this extension of

τ

satisfies boundary analogues of the additivity and transfer axioms. We need the following lemmas.

Lemma 4.1. Suppose that

E_{i}

are smooth unipotent bundles over

B

with the same vertical boundary. Then

τ (E_{1} \cup E_{2}) + τ (E_{3} \cup E_{4}) = τ (E_{1} \cup E_{3}) + τ (E_{2} \cup E_{4})

Proof. Both sides are equal to $\frac{1}{2} \sum τ (D E_{i})$ . □

Lemma 4.2.

τ (\partial^{v} E) = τ (\partial^{v} (E \times D^{2}))

assuming

E

is unipotent.

Proof. Since $\partial^{v} (E \times D^{2}) = \partial^{v} E \times D^{2} \cup E \times S^{1}$ we have by the additivity axiom that $τ (\partial^{v} (E \times D^{2})) = \frac{1}{2} τ (\partial^{v} E \times S^{2}) + \frac{1}{2} τ (D E \times S^{1}) .$ But $τ (\partial^{v} E \times S^{2}) = 2 τ (\partial^{v} E)$ and $τ (D E \times S^{1}) = 0$ by the transfer axiom. □

Lemma 4.3 (additivity of transfer). If

E = E_{1} \cup E_{2}

is a union of two smooth bundles along their common vertical boundary

\partial^{v} E_{1} = \partial^{v} E_{2} = E_{1} \cap E_{2}

then

t r_{B}^{E} (x) = t r_{B}^{E_{1}} (x | E_{1}) + t r_{B}^{E_{2}} (x | E_{2}) - t r_{B}^{\partial_{0} E_{1}} (x | \partial^{v} E_{1})

for all

x \in H^{*} (E; R)

Proposition 4.4 (additivity for boundary case). If

(E_{1}, \partial_{0}), (E_{2}, \partial_{0})

are unipotent bundle pairs over

B

with

E_{1} \cap E_{2} = \partial_{0} E_{1} = \partial_{0} E_{2}

then

τ (E_{1} \cup E_{2}) = τ (E_{1}) + τ (E_{2}) - τ (E_{1} \cap E_{2}) .

Proof. We expand each term using the defining equation:

\begin{matrix} τ (E_{i}) & : = \frac{1}{2} τ (D E_{i}) + \frac{1}{2} τ (\partial^{v} E_{i}) \end{matrix}

\begin{matrix} τ (E_{1} \cup E_{2}) & : = \frac{1}{2} τ (D (E_{1} \cup E_{2})) + \frac{1}{2} τ (\partial^{v} (E_{1} \cup E_{2})) \end{matrix}

\begin{matrix} τ (E_{1} \cap E_{2}) & : = \frac{1}{2} τ (\partial^{v} (E_{1} \cap E_{2})) + \frac{1}{2} τ (D (E_{1} \cap E_{2})) . \end{matrix}

\begin{matrix}  \end{matrix}

The order of the terms in the last equation is reversed so that it matches the following two examples of Lemma 4.1 :

τ (\partial^{v} E_{1}) + τ (\partial^{v} E_{2}) = τ (\partial^{v} (E_{1} \cup E_{2})) + τ (D (E_{1} \cap E_{2}))

τ (D E_{1}) + τ (D E_{2}) = τ (D (E_{1} \cup E_{2})) + τ (D (I \times (E_{1} \cap E_{2}))) .

However, the last term is

τ (D (I \times (E_{1} \cap E_{2}))) = τ (\partial^{v} (D^{2} \times (E_{1} \cap E_{2}))) = τ (\partial^{v} (E_{1} \cap E_{2}))

by Lemma 4.2 . The proposition follows. □

Proposition 4.5 (transfer for boundary case). If

X \to D \to E

is an oriented linear disk or sphere bundle then

τ_{B} (D) = χ (X) τ (E) + t r_{B}^{E} (τ_{E} (D)) .

Proof. We consider first the case when $D = D (ξ)$ is an oriented linear $D^{n}$ -bundle and $F$ is closed. I.e., we will show: $τ_{B} (D (ξ)) = τ (E) + t r_{B}^{E} (τ_{E} (D (ξ)))$ . This is just half of the sum of the following two examples of the original transfer axiom. $τ_{B} (S^{n} (ξ)) = χ (S^{n}) τ (E) + t r_{B}^{E} (τ_{E} (S^{n} (ξ)))$ $τ_{B} (S^{n - 1} (ξ)) = χ (S^{n - 1}) τ (E) + t r_{B}^{E} (τ_{E} (S^{n - 1} (ξ)))$ The transfer axiom takes care of the case when $D$ is a sphere bundle and $F$ is closed. The remaining case when $\partial F$ is nonempty is given by the following lemma. □

Lemma 4.6. With the fiber

X

D \to E

fixed, the transfer formula for

F

closed implies the transfer formula for

F

with boundary.

Proof. Write

D E = E \cup E^{'}

as the union of two copies of

E

along its vertical boundary. Let

D, D^{'}

be two copies of

D

with

D \cap D^{'} = q^{- 1} (\partial^{v} E)

. Then the transfer formula

τ_{B} (D) = χ (X) τ (E) + t r_{B}^{E} (τ_{E} (D))

is half the sum of the following two transfer formulas with closed fibers

D F, \partial F

respectively.

\begin{matrix} τ_{B} (D \cup D^{'}) & = χ (X) τ (E \cup E^{'}) + t r_{B}^{E \cup E^{'}} (τ_{E \cup E^{'}} (D \cup D^{'})) \end{matrix}

\begin{matrix} τ_{B} (D \cap D^{'}) & = χ (X) τ (E \cap E^{'}) + t r_{B}^{E \cap E^{'}} (τ_{E \cap E^{'}} (D \cap D^{'})) \end{matrix}

\begin{matrix}  \end{matrix}

The additivity of transfer (Lemma 4.3 ) is used here. □

4.2 Relative torsion

Suppose that

(F, \partial_{0} F) \to (E, \partial_{0} E) \to B

is a unipotent smooth bundle pair.

By this we mean that the vertical boundary

\partial^{v} E

is the union of two subbundles

\partial^{v} E = \partial_{0} E \cup \partial_{1} E

which meet along their common vertical boundary:

\partial_{0} E \cap \partial_{1} E = \partial^{v} \partial_{0} E = \partial^{v} \partial_{1} E

and that both

E

and

\partial_{0} E

are unipotent. This implies the weaker condition that the pair

(E, \partial_{0} E)

is relatively unipotent. We use the abbreviation

(E, \partial_{0})

for

(E, \partial_{0} E)

Suppose that

τ

is a higher torsion invariant which has been extended to the boundary case as above. Then for any unipotent smooth bundle pair

(E, \partial_{0}) \to B

we define the relative torsion by

τ (E, \partial_{0}) : = τ (E) - τ (\partial_{0} E) .

Proposition 4.7 (additivity in the relative case). Suppose that

E \to B

is a smooth bundle which can be written as a union of two subbundles

E = E_{1} \cup E_{2}

which meet along a subbundle of their respective vertical boundaries:

E_{1} \cap E_{2} = \partial_{0} E_{2} \subseteq \partial^{v} E_{1}

. Let

\partial^{v} E_{1} = \partial_{0} E \cup \partial_{1} E

be a decomposition

\partial^{v} E_{1}

so that

\partial_{0} E_{2} \subseteq \partial_{1} E_{1}

and

(E_{i}, \partial_{0}) \to B, i = 1, 2

are unipotent smooth bundle pairs. Then

(E, \partial_{0} E_{1}) \to B

is unipotent and

τ (E_{1} \cup E_{2}, \partial_{0} E_{1}) = τ (E_{1}, \partial_{0}) + τ (E_{2}, \partial_{0}) .

Proof. Both sides are equal to $τ (E_{1}) + τ (E_{2}) - τ (E_{1} \cap E_{2}) - τ (\partial_{0} E_{1})$ . □

Here is another variation of the additivity axiom which is also trivial to prove.

Proposition 4.8 (horizontal additivity). Suppose that

(E, \partial_{0}) \to B

is a union of two unipotent bundle pairs

(E_{i}, \partial_{0} E_{i})

in the sense that

E = E_{1} \cup E_{2}

and

\partial_{0} E = \partial_{0} E_{1} \cup \partial_{0} E_{2}

with

E_{1} \cap E_{2} \subseteq \partial_{1} E_{1} \cap \partial_{1} E_{2}

. Let

X = E_{1} \cap E_{2}

and

\partial_{0} X = X \cap \partial_{0} E

and suppose

(X, \partial_{0})

is a unipotent bundle pair. Then

(E, \partial_{0})

is unipotent and

τ (E, \partial_{0}) = τ (E_{1}, \partial_{0}) + τ (E_{2}, \partial_{0}) - τ (X, \partial_{0}) . □

To state the transfer axiom in the relative case we need the relative transfer :

t r_{B}^{(E, \partial_{0})} : H^{*} (E; Z) \to H^{*} (B; Z)

given by

\begin{matrix} t r_{B}^{(E, \partial_{0})} (x) = p_{*} (x \cup e (E, \partial_{0})) \end{matrix}

(4.1)

where

p_{*}

is the push-down operator (as before) and

e (E, \partial_{0}) \in H^{n} (E, \partial^{v} E; Z)

is the relative Euler class given by pulling back the Thom class of the vertical tangent bundle

T^{v} E

along any vertical tangent vector field which is nonzero along the vertical boundary

\partial^{v} E

and which points inward along

\partial_{0} E

, outward along

\partial_{1} E

and is tangent to

\partial^{v} E

pointing from

\partial_{0} E

\partial_{1} E

along

\partial_{0} E \cap \partial_{1} E

. As in the absolute case, the relative transfer is independent of the choice of orientation of the fiber. The relative transfer also satisfies the following two equations for any

x \in H^{*} (E; Z)

\begin{matrix} t r_{B}^{(E, \partial_{0})} (x) = t r_{B}^{E} (x) - t r_{B}^{\partial_{0} E} (x | \partial_{0} E) \end{matrix}

(4.2)

\begin{matrix} t r_{B}^{(E, \partial_{1})} (x) = (- 1)^{n} t r_{B}^{(E, \partial_{0})} (x) \end{matrix}

(4.3)

And, finally,

t r_{B}^{(E, \partial_{0})} \circ p^{*} : H^{*} (B) \to H^{*} (B)

is multiplication by the relative Euler characteristic of the fiber pair

(F, \partial_{0} F)

given by

χ (F, \partial_{0}) : = χ (F) - χ (\partial_{0} F) .

Proposition 4.9 (transfer in the relative case). Let

(F, \partial_{0}) \to (E, \partial_{0}) - \to p B

and

(X, \partial_{0}) \to (D, \partial_{0}) - \to q E

be unipotent smooth bundle pairs so that the second is an oriented linear

S^{n}

D^{n}

bundle with

\partial_{0} X = S^{n - 1}, D^{n - 1}

\emptyset

. Then

τ_{B} (D, \partial_{0} D \cup q^{- 1} \partial_{0} E) = χ (X, \partial_{0}) τ (E, \partial_{0}) + t r_{B}^{(E, \partial_{0})} (τ_{E} (D, \partial_{0})) .

Proof. We already did the case when both

\partial_{0} F

and

\partial_{0} X

are empty. The case when

\partial_{0} X

is empty follows easily from the following formula which holds by definition.

τ_{B} (D, q^{- 1} \partial_{0} E) = τ_{B} (D) - τ_{B} (q^{- 1} \partial_{0} E)

The general case follows from the following two examples of the

\partial_{0} X = \emptyset

case.

\begin{matrix} τ_{B} (\partial_{0} D, \partial_{0} D \cap q^{- 1} \partial_{0} E) & = χ (\partial_{0} X) τ (E, \partial_{0}) + t r_{B}^{(E, \partial_{0})} (τ_{E} (\partial_{0} D)) \end{matrix}

\begin{matrix} τ_{B} (D, q^{- 1} \partial_{0} E) & = χ (X) τ (E, \partial_{0}) + t r_{B}^{(E, \partial_{0})} (τ_{E} (D)) \end{matrix}

\begin{matrix}  \end{matrix}

Take the second formula minus the first to prove the proposition. □

A useful special case is the case when

D \to E

is the pull-back of a linear bundle over

B

. In this case

D

is the fiber product of two bundles.

Corollary 4.10 (product formula). Suppose that

(F, \partial_{0}) \to (E, \partial_{0}) \to B

and

(X, \partial_{0}) \to (E^{'}, \partial_{0}) \to B

are unipotent smooth bundle pairs so that the second is an oriented linear

S^{n}

D^{n}

bundle with

\partial_{0} X = S^{n - 1}, D^{n - 1}

\emptyset

. Let

D = E \times_{B} E^{'}

be the fiber product of these bundles and let

\partial_{0} D = \partial_{0} E \times_{B} E^{'} \cup E \times_{B} \partial_{0} E^{'}

. Then

τ (D, \partial_{0}) = χ (X, \partial_{0}) τ (E, \partial_{0}) + χ (F, \partial_{0}) τ (E^{'}, \partial_{0}) . □

Exercise 4.11. Show that for any higher torsion invariant

τ

and any unipotent bundle pair

(E, \partial_{0}) \to B

we have

τ (E, \partial_{0}) + (- 1)^{n} τ (E, \partial_{1}) = 2 τ^{+} (E, \partial_{0})

where

n = dim F

is the fiber dimension.

4.3 Further extension

Suppose that

(E, \partial_{0}) \to B

is a relatively unipotent smooth bundle pair. Then the torsion

τ (E, \partial_{0})

can be defined as follows.

(1) Let $D (ν) \to E$ be the normal disk bundle and let $\partial_{0} D = D (ν) | \partial_{0} E$ .
(2) Embed $\partial_{0} E$ fiberwise into the northern hemisphere of $B \times S^{N}$ for $N$ large.
(3) Thicken this embedding to a codimension zero embedding $\partial_{0} D \to B \times S^{N}$ . Then the union of $B \times D^{N + 1}$ with $D$ along $\partial_{0} D$ is unipotent.
(4) Define the higher torsion of $(D, \partial_{0})$ by $τ (D, \partial_{0}) : = τ (B \times D^{N + 1} \cup D) .$
(5) Define the higher torsion of $(E, \partial_{0})$ by $τ (E, \partial_{0}) : = τ (D, \partial_{0}) - (s_{1} + s_{2}) t r_{B}^{E} (c h_{4 k} (ν))$ where $s_{1}, s_{2}$ are given in the explanation of the Main Theorem 3.3 .

The main theorem (and Theorem 8.9 ) are needed to show that this is well defined.

The proof is that these formulas are well defined for

M_{2 k}

and for

τ_{2 k}

which span all possibilities by the main theorem. Therefore, we cannot use this extension to the relatively unipotent case to prove the main theorem. In the case when

(E, \partial_{0})

is unipotent these formulas hold by additivity and transfer.

5 Stability of higher torsion

Smooth bundle are stabilized by taking products with disks. The following special case of the product formula says that higher torsion is a stable invariant.

Corollary 5.1 (stability of torsion). If

(E, \partial_{0}) \to B

is a unipotent smooth bundle pair then so is

(E \times D^{n}, \partial_{0} E \times D^{n})

and the relative torsion is the same:

τ (E \times D^{n}, \partial_{0} E \times D^{n}) = τ (E, \partial_{0}) . □

M

is a compact smooth manifold then we recall that a concordance of

M

is a diffeomorphism of

M \times I

which is the identity on

M \times 0 \cup \partial M \times I

. Let

C (M)

be the space of concordances of

M

with the

C^{\infty}

topology.

C (M) = D i f f (M \times I r e l M \times 0 \cup \partial M \times I)

The classifying space

ℋ (M) = B C (M)

is the space of

h

-cobordisms of

M r e l \partial M

Recall that an

h

-cobordism of

M r e l \partial M

is a compact smooth manifold

W

with boundary

\partial W = M \times 0 \cup \partial M \times I \cup M^{'}

where

M^{'}

is another compact smooth manifold with

\partial M^{'} = \partial M \times I

. A mapping

B \to ℋ (M)

is a smooth bundle over

B

whose fibers are all

h

-cobordisms of

M r e l \partial M

so that the subbundle with fiber

M \times 0 \cup \partial M \times I

is trivial. We will call such a bundle an

h

-cobordism bundle over

B

There is a suspension map

σ : C (M) \to C (M \times I)

which is highly connected when

dim M

is large by the concordance stability theorem ([12] or the last chapter of [15] ). The limit is the stable concordance space

P (M) = {lim}_{\to} C (M \times I^{n})

which is well known to be an infinite loop space. Therefore, the set

[B, B P (M)]

of homotopy classes of maps from

B

to the classifying space

B P (M)

is an additive group. By the concordance stability theorem this group is isomorphic to

[B, ℋ (M \times I^{n})]

for sufficiently large

n

Proposition 5.2. For any

h

-cobordism bundle

E \to B

let

\partial_{0} E

be the trivial subbundle with fiber

M \times 0 \cup \partial M \times I

. Then the higher torsion invariant

E \mapsto τ (E, \partial_{0})

gives an additive map

τ : [B, ℋ (M \times I^{n})] \to R

Proof. This is an immediate consequence of the horizontal additivity of $τ$ (Proposition 4.8 ) since the H-space structure on the $h$ -cobordism space $ℋ (M \times I^{n})$ is given by lateral union. I.e., the sum of two mappings $B \to C (M \times I^{n})$ is given by lateral union of the corresponding $h$ -cobordism bundles followed by rescaling. (The lateral union is an $h$ -cobordism of $M \times I^{n - 1} \times [0, 2]$ . The last coordinate needs to be rescaled down to $[0, 1]$ .) □

Corollary 5.3. Suppose that

E_{α}

is a collection of

h

-cobordism bundles which spans the

Q

vector space

[B, ℋ (M \times I^{n})] \otimes Q

. Suppose also that

τ (E_{α}, \partial_{0}) = 0

for all

α

. Then

τ (E, \partial_{0}) = 0

for any

h

-cobordism bundle

E

classified by a map

B \to ℋ (M \times I^{n})

. □

Remark 5.4. This is equivalent to saying that, if

τ, τ^{'}

are two higher torsion invariants which agree on the rational generators

E_{α}

then they agree on all

h

-cobordism bundles

E

6 Computation of higher torsion

We will now show how the higher torsion invariants of unipotent bundles can be computed in many cases given the values of the parameters

s_{1}, s_{2}

. We recall that these parameters are given by

τ (S^{n} (λ)) = 2 s_{n} c h_{4 k} (λ)

where

S^{n} (λ)

is an

S^{1}

S^{2}

bundle associated to a complex line bundle

λ

over

B

6.1 Torsion of disk and sphere bundles

Theorem 6.1. The higher torsion of the

D^{n}

-bundle

D^{n} (ξ)

associated to an

S O (n)

-bundle

ξ

over

B

is given by

τ (D^{n} (ξ)) = (s_{1} + s_{2}) c h_{4 k} (ξ) .

Proof. For $n = 2$ this is by definition of $τ$ in the boundary case: $τ (D^{2} (ξ)) = \frac{1}{2} τ (S^{2} (ξ)) + \frac{1}{2} τ (S^{1} (ξ)) = (s_{2} + s_{1}) c h_{4 k} (ξ) .$ The general case follows from the product formula (Corollary 4.10 ) and the splitting principle. If $n = 2 m + 1$ then we may assume by the splitting principle that $ξ$ is a direct sum of $m$ complex line bundles $λ_{i}$ and a trivial real line bundle.
Then $τ (D^{n} (ξ)) = τ (D^{2} (λ_{1}) \times_{B} \dots \times_{B} D^{2} (λ_{m}) \times_{B} (B \times I))$ $= (s_{1} + s_{2}) \sum c h_{4 k} (λ_{i}) = (s_{1} + s_{2}) c h_{4 k} (ξ)$ by the product formula. The even case is similar. □

From the calculation

τ (D^{n} (ξ)) = \frac{1}{2} τ (S^{n}) + \frac{1}{2} τ (S^{n - 1} (ξ)) = (s_{1} + s_{2}) c h_{4 k} (ξ)

we get the following by induction on

n

Corollary 6.2. For

n > 0

the higher torsion of the

S^{n}

-bundle

S^{n} (ξ)

associated an

S O (n + 1)

-bundle

ξ

over

B

is given by

τ (S^{n} (ξ)) = 2 s_{n} c h_{4 k} (ξ)

where

s_{n}

depend only on the parity of

n

Comparing this with Proposition 8.4 and Theorem 8.11 below we get the following.

Lemma 6.3. Let

k \geq 1

and let

a, b \in R

be given by

a : = \frac{(- 1)^{k + 1} 2 s_{1}}{ζ (2 k + 1)}, b : = \frac{s_{1} + s_{2}}{(2 k)!}

Then

τ (E) = a τ_{2 k} (E) + b M_{2 k} (E)

for

E

any oriented linear sphere or disk bundle over

B

Remark 6.4. If the value of

τ

on a bundle

E

is determined by the known values of

τ

on disk and sphere bundles then this lemma implies that

τ (E) = a τ_{2 k} (E) + b M_{2 k} (E)

for that bundle.

6.2 Morse bundles

Given a Morse function

f : (M, \partial_{0} M) \to (I, 0)

, a compact

n

-manifold

M

will be decomposed as a union of handles

D^{i} \times D^{n - i}

attached along

S^{i - 1} \times D^{n - i}

to the union of lower handles and the base

\partial_{0} M \times [0, ε]

. Each such handle has a critical point at its center with index

i

. The core of the handle is

D^{i} \times 0

. This is also the union of trajectories of the gradient of

f

(with respect to some metric on

M

) which converge to the critical point. The tangent plane is the negative eigenspace of the second derivative

D^{2} f

at the critical point.

Suppose that

(E, \partial_{0}) \to B

is a smooth bundle pair and

f : E \to I = [0, 1]

is a fiberwise Morse function with

f^{- 1} (0) = \partial_{0} E

and with distinct critical values. In other words, for each

b \in B

, the restriction

f_{b} : (F_{b}, \partial_{0}) \to (I, 0)

is a Morse function with critical points

x_{1} (b), \dots, x_{m} (b)

having critical values

f_{b} (x_{1}) < f_{b} (x_{2}) < \dots < f_{b} (x_{m})

For

1 \leq j \leq m

let

ξ_{j}

be the negative eigenspace bundle of the fiberwise second derivative of

f

along

x_{j}

. Let

η_{j}

be the positive eigenspace bundle. Then

E

has a filtration

E = E_{m} \supset E_{m - 1} \supset \dots \supset E_{0}

where

E_{0} \sim = \partial_{0} E \times I

and

E_{j} = E_{j - 1} \cup D (ξ_{j}) \times_{B} D (η_{j})

where

E_{j - 1} \cap D (ξ_{j}) \times_{B} D (η_{j}) = S (ξ_{j}) \times_{B} D (η_{j})

. By additivity we get the following.

Lemma 6.5. If

\partial_{0} E \to B

is unipotent and the bundles

ξ_{j}, η_{j}

are oriented then

(E, \partial_{0}) \to B

is a unipotent bundle pair with torsion invariant

τ (E, \partial_{0}) = \sum τ ((D (ξ_{j}), S (ξ_{j})) \times_{B} D (η_{j}))

Each summand in the above lemma can be determined using the product formula (Corollary 4.10 ).

Lemma 6.6. The value of

τ

on the fiber product of an oriented linear disk bundle

D (η)

and the oriented relative

i

-disk bundle

(D^{i} (ξ), S^{i - 1} (ξ))

is given by

τ ((D^{i} (ξ), S^{i - 1} (ξ)) \times_{B} D (η)) = (- 1)^{i} τ (D (η)) + τ (D^{i} (ξ), S^{i - 1} (ξ))

= (- 1)^{i} (s_{1} + s_{2}) c h_{4 k} (η) + (s_{i} - s_{i - 1}) c h_{4 k} (ξ)

Remark 6.7. Since

s_{i} - s_{i - 1} = (- 1)^{i} s_{2} - (- 1)^{i} s_{1}

, the even and odd parts of the above formula are:

τ^{+} ((D^{i} (ξ), S^{i - 1} (ξ)) \times_{B} D (η)) = (- 1)^{i} s_{2} (c h_{4 k} (η) + c h_{4 k} (ξ))

τ^{-} ((D^{i} (ξ), S^{i - 1} (ξ)) \times_{B} D (η)) = (- 1)^{i} s_{1} (c h_{4 k} (η) - c h_{4 k} (ξ))

Putting these together we get the following theorem which is a mild improvement over the obvious. Namely, the negative eigenspace bundles need not be oriented.

However, the sum

ξ_{i} \oplus η_{i}

of negative and positive eigenspace bundles must be orientable since it is the vertical tangent bundle along the

i

th component of the Morse critical set.

Theorem 6.8. Suppose that

(E, \partial_{0}) \to B

is a unipotent smooth bundle pair and

f : (E, \partial_{0} E) \to (I, 0)

is a fiberwise Morse function with distinct critical values. Let

ξ_{i}, η_{i}

be the negative and positive eigenspace bundles associated to the

i

th critical point. then

τ (E, \partial_{0}) = \sum (- 1)^{i} (s_{1} + s_{2}) c h_{4 k} (η) + (- 1)^{i} (s_{2} - s_{1}) c h_{4 k} (ξ)

Proof. Suppose first that the bundles $ξ_{i}, η_{i}$ are oriented. Then the two lemmas above apply to prove the theorem. If these bundles are not oriented then there is a finite covering $\tilde{B}$ of $B$ so that, on the pull-back $\tilde{E}$ , the function $\tilde{E} \to E \to I$ is a Morse function with oriented eigenspace bundles ${\tilde{x i}}_{i}, {\tilde{η}}_{i}$ which are pull-backs of $ξ_{i}, η_{i}$ . Thus the theorem applies to $\tilde{E}$ . However, the induced map in real cohomology $H^{*} (B; R) \to H^{*} (\tilde{B}; R)$ is a monomorphism. The two expressions in our theorem are elements of $H^{4 k} (B; R)$ which go to the same element of $H^{4 k} (\tilde{B}; R)$ . So, they must be equal. □

6.3 Hatcher's example

One crucial example of a Morse bundle to which the above theorem holds is Hatcher's construction. This constructs an exotic disk bundle over

B = S^{n}

out of an element of the kernel of the

J

-homorphism

J : π_{n - 1} O \to π_{n - 1}^{s} (S^{0}) .

Bökstedt [4] interpretted this as a mapping from

G / O

to the stable concordance space of a point

ℋ (*) = B P (*)

. This means that a mapping

B \to G / O

gives an exotic disk

h

-cobordism bundle over

B

Since

G / O

is the homotopy fiber of the map

B O \to B G

, a map

B \to G / O

is given by an

n

-plane bundle

ξ : B \to S O (n)

together with a homotopy trivialization of the associated sphere bundle. I.e., we have a fiber homotopy equivalence

g : S^{n - 1} (ξ) ≃ B \times S^{n - 1} .

We write this as a family of homotopy equivalences

g_{t} : S_{t}^{n - 1} (ξ) \to S^{n - 1}, t \in B

and extend to the disk

D_{t}^{n} (ξ)

(the fiber over

t \in B

of the disk bundle

D^{n} (ξ)

)

{\bar{g}}_{t} : (D_{t}^{n} (ξ), S_{t}^{n - 1} (ξ)) \to (D^{n}, S^{n - 1}) .

Assuming that

n

and

m

are large enough, we can lift

{\bar{g}}_{t}

up to a family of embeddings

{\tilde{g}}_{t} : (D_{t}^{n} (ξ), S_{t}^{n - 1} (ξ)) \to (D^{n} \times D^{m}, S^{n - 1} \times D^{m})

If we let

η

be the complementary bundle to

ξ

, we can extend this to a family of codimension zero embeddings

G_{t} : (D_{t}^{n} (ξ) \times D_{t}^{m} (η), S_{t}^{n - 1} (ξ) \times D_{t}^{m} (η)) \to (D^{n} \times D^{m}, S^{n - 1} \times D^{m})

We will use this to construct a nontrivial

h

-cobordism bundle over

B

. We start this construction with the trivial

h

-cobordism bundle

E_{0} = B \times D^{n + m - 1} \times I

. Over each

t \in B

the fiber is

D^{n + m - 1} \times I

which the trivial

h

-cobordism of

D^{n + m - 1}

. We now add two canceling handles: an

n - 1

handle

H^{n - 1}

and an

n

handle

H^{n}

. We will then replace

H^{n}

with a variable

n

-handle

H_{t}^{n}

The

n - 1

handle

H^{n - 1}

is simply

S^{n - 1} \times D^{m + 1}

which is attached by boundary connected sum to the top boundary

D^{n + m - 1} \times 1

. The next cell

H^{n}

is a copy of

D^{n} \times D^{m}

which is attached along an embedding:

i d \times α : S^{n - 1} \times D^{m} \to S^{n - 1} \times \partial D^{m - 1}

where

α : D^{m} \to \partial D^{m - 1}

is some fixed embedding.

The new

h

-cobordism bundle

B \times (D^{n + m - 1} \times I \cup H^{n - 1} \cup H^{n})

is still a trivial bundle. However, it contains a nontrivial subbundle given by replacing

H^{n}

with the image of

G_{t}

for every

t \in B

. Let

H_{t}^{n}

denote this subset. Then

H_{t}^{n}

is a variable (“rotating”)

n

-handle which is attached to

\partial H^{n - 1}

by the mapping

(i d \times α) \circ G_{t} : S_{t}^{n - 1} (ξ) \times D_{t}^{m} (η) \to S^{n - 1} \times \partial D^{m - 1} \subset \partial H^{n - 1}

Let

(E_{1}, \partial_{0})

denote the bundle pair with fiber pair

(D^{n + m - 1} \times I \cup H^{n - 1} \cup H_{t}^{n}, D^{n + m - 1} \times 0 \cup \partial D^{n + m - 1} \times I) .

The main property of this bundle are given by the following theorem originally due to Böckstedt [4] (see also [5] ) and reproved using Morse theory in [13] .

Theorem 6.9. The bundles

(E_{1}, \partial_{0}) \to B

given by Hatcher's construction above are rational generators for the group

[B, ℋ (D^{n + m - 1})]

By construction, Hatcher's

h

-cobordism bundle admits a fiberwise Morse function with exactly two critical points on each fiber. The first critical point has index

n - 1

and has a trivial negative eigenspace bundle. The second has index

n

and its negative eigenspace bundle is

ξ

. By Theorem 6.8 , the torsion of Hatcher's bundle is:

τ (Δ) (ξ) = (- 1)^{n} (s_{1} + s_{2}) c h_{4 k} (η) + (- 1)^{n} (s_{2} - s_{1}) c h_{4 k} (ξ) .

Since

c h_{4 k} (η) + c h_{4 k} (ξ) = 0

(

ξ \oplus η

being trivial) this simplifies to:

\begin{matrix} τ (Δ) (ξ) = (- 1)^{n + 1} 2 s_{1} c h_{4 k} (ξ) . \end{matrix}

(6.1)

Since

s_{1} = 0

for

M_{2 k}

this implies the following.

Lemma 6.10. For any bundle

(D, \partial_{0}) \to B

h

-cobordisms of disks we have

τ (D, \partial_{0}) = τ (D) = a τ_{2 k} (D)

where

a

is given in Lemma 6.3 .

Proof. This follows from the calculation ( 6.1 ) above in the case of Hatcher's example.
By Böckstedt (Theorem 6.9 ) and Corollary 5.3 this implies that $τ (D) = a τ_{2 k} (D)$ for any disk $h$ -cobordism bundle of sufficiently large fiber dimension.
If $D$ is any disk $h$ -cobordism bundle then $τ (D) = τ (D \times D^{k})$ and $τ_{2 k} (D) = τ_{2 k} (D \times D^{k})$ by stability (Corollary 5.1 ). Furthermore, $D \times D^{k}$ will be a disk bundle with fiber dimension, say $N$ , which contains a trivial $N - 1$ disk boundary in its vertical boundary. Therefore it can be modified to be an $h$ -cobordism bundle where the formula holds. It follows from the stable result that $τ (D \times D^{k}) = a τ_{2 k} (D \times D^{k})$ .
The lemma follows. □

Proposition 6.11. For any smooth oriented disk bundle

D \to B

we have

τ (D) = a τ_{2 k} (D) + b M_{2 k} (D)

where

a, b

are given in Lemma 6.3 .

Proof. By Corollary 5.1 , we can stabilize $D$ without changing the value of $τ (D), τ_{2 k} (D)$ or $M_{2 k} (D)$ . Then the fiber of $D \to B$ will be a disk $D^{N}$ where $N > dim B$ . Next, choose a smooth section $B \to \partial^{v} D$ . Thicken this up to get a linear $N - 1$ disk bundle $D (ξ)$ which lies in $\partial^{v} D$ .
Suppose first that $ξ$ is a trivial bundle. Then the bundle $D$ is equivalent to an $h$ -cobordism bundle. We can see this by attaching $D$ along the image of $D (ξ)$ to the top of $B \times D^{N - 1} \times I$ . Then, by additivity, the resulting $h$ -cobordism bundle will have the same torsion ( $τ, τ_{2 k}$ and $M_{2 k}$ ) as the disk bundle. The equation relating these torsions holds for the $h$ -cobordism bundle by the lemma. Therefore it holds for the disk bundle.
If the linear bundle $ξ$ is nontrivial we simply take the fiber product of $D$ with the disk bundle of the complementary bundle $η$ . The bundles $ξ, η$ are necessarily oriented since $D$ is oriented. So, the new bundle $D \times_{B} D (η)$ has torsion $τ (D \times_{B} D (η)) = τ (D) + τ (D (η))$ by the product formula. Since the proposition holds for $D \times_{B} D (η)$ and $D (η)$ , it holds for $D$ . □

7 Uniqueness of higher torsion

We are now ready to prove the uniqueness part of the main theorem. Namely, we will show that

τ = a τ_{2 k} + b M_{2 k}

on all unipotent bundles. To do this we show that the difference is zero. We continue to assume that

τ_{2 k}, M_{2 k}

are higher torsion invariants, facts that we prove in the last section.

7.1 The difference torsion

We define the difference torsion by

τ^{δ} : = τ - a τ_{2 k} - b M_{2 k} .

This is a linear combination of higher torsion invariants and therefore a higher torsion invariant. Furthermore this new invariant has the property that

s_{1} = s_{2} = 0

So:

Lemma 7.1. The difference torsion is zero on all smooth oriented disk bundles and oriented linear sphere bundles.

Lemma 7.2 (thickening lemma). If

q : D \to E

is an oriented linear disk bundle then the difference torsion of any unipotent bundle pair

(E, \partial_{0})

is equal to that of

(D, \partial_{0})

as a bundle over

B

where

\partial_{0} D = q^{- 1} (\partial_{0} E))

. I.e.,

τ_{B}^{δ} (D, \partial_{0}) = τ^{δ} (E, \partial_{0}) .

Proof. By the relative transfer formula we have $τ_{B}^{δ} (D, \partial_{0}) = τ^{δ} (E, \partial_{0}) + t r_{B}^{E} (τ_{E}^{δ} (D)) .$ But $τ_{E}^{δ} (D) = 0$ since $D$ is a linear disk bundle over $E$ . □

Lemma 7.3. The difference torsion is zero on any unipotent bundle pair

(E, \partial_{0}) \to B

whose fibers

(F, \partial_{0})

are

1

-connected and homologically trivial (

H_{*} (F, \partial_{0} F; Z) = 0

Proof. First we note that the homological condition on the fiber implies that $F / \partial_{0} F$ is contractible (being $1$ -connected and acyclic).
Choose a smooth fibred embedding $(E, \partial_{0} E) \to (B \times S^{N - 1} \times I, B \times S^{N - 1} \times 0)$ for some large $N$ . Let $(D, \partial_{0})$ be a tubular neighborhood of the image of $(E, \partial_{0})$ .
Since $(D, \partial_{0})$ is the normal disk bundle it has the same difference torsion as $(E, \partial_{0})$ .
So it suffices to show that $τ^{δ} (D, \partial_{0}) = 0$ .
The union $D \cup B \times D^{N}$ is a bundle whose fiber is homotopy equivalent to $F / \partial_{0} F$ and thus contractible. Therefore it is a disk bundle when the corners are rounded.
So, its difference torsion is trivial. The additivity theorem implies that $τ^{δ} (D \cup B \times D^{N}) = τ^{δ} (D) + τ^{δ} (B \times D^{N}) - τ^{δ} (\partial_{0} D) = τ^{δ} (D, \partial_{0}) = 0$ as claimed. □

Lemma 7.4 (main lemma). The difference torsion

τ^{δ}

is a fiber homotopy invariant of unipotent smooth bundle pairs.

Proof. Suppose that $(E_{1}, \partial_{0})$ and $(E_{2}, \partial_{0})$ are unipotent smooth bundle pairs over $B$ that are fiber homotopy equivalent. Then we want to show that $τ^{δ} (E_{1}, \partial_{0}) = τ^{δ} (E_{2}, \partial_{0}) .$ By the thickening lemma we can make the fiber dimension of $(E_{2}, \partial_{0})$ arbitrarily large. Then we can approximate the fiber homotopy equivalence by a fiberwise smooth embedding $g : (E_{1}, \partial_{0}) \to (D_{2}, \partial_{0}) .$ Using the thickening lemma again we can assume that $g$ is a codimension $0$ embedding. The complement of the image of $g$ is a unipotent fiberwise $h$ -cobordism with trivial $τ^{δ}$ by the previous lemma. Thus $τ^{δ} (E_{1}, \partial_{0}) = τ^{δ} (D_{2}, \partial_{0})$ by additivity.
□

7.2 Vanishing of the fiber homotopy invariant

The final step in the proof of Theorem 3.3 is to show the following.

Lemma 7.5. Any higher torsion invariant

τ^{δ}

which is also a fiber homotopy invariant of unipotent smooth bundle pairs must be zero.

Since

τ^{δ}

is a fiber homotopy invariant, it is well defined on any fibration pair

(Z, C) \to B

with fiber

(X, A)

which is unipotent in the sense that

H_{*} (X; Q)

and

H_{*} (A; Q)

are unipotent as

π_{1} B

-modules and smoothable in the sense that it is fiber homotopy equivalent to a smooth bundle pair

(E, \partial_{0})

with compact manifold fiber

(F, \partial_{0})

. By definition, the relative torsion is related to the absolute torsion by

\begin{matrix} τ^{δ} (Z, C) = τ^{δ} (Z) - τ^{δ} (C) . \end{matrix}

(7.1)

We will examine what it means for a fibration

X \to Z \to B

to be unipotent and smoothable in terms of diffeomorphism spaces of compact manifolds. Suppose that

M \to E \to B

is a smoothing of

Z \to B

. Then

M

is a compact smooth manifold homotopy equivalent to

X

. By taking the product with a high dimensional disk we may assume that the inclusion map of the boundary

\partial M ↪ M

is highly (at least

4 k

) connected. The structure group of the bundle

E \to B

will be a subgroup of the diffeomorphism group of

M

consisting of orientation preserving diffeomorphisms which act unipotently with respect to a fixed flag

ℱ

in the rational homology of

M

, i.e., these are diffeomorphisms which preserve the flag and act as the identity on the successive quotients. Let

D i f f_{ℱ} (M)

denote the space of such diffeomorphisms with the

C^{\infty}

topology. Then the smooth bundle

E \to B

is classified by a mapping

B \to B D i f f_{ℱ} (M)

The difference torsion comes from a universal invariant

τ^{δ} \in H^{4 k} (B D i f f_{ℱ} (M); R)

associated to the universal bundle over

B D i f f_{ℱ} (M)

with fiber

M

. Let

M_{0} \subset M

be the closure of the complement of a tubular neighborhood of

\partial M

. Then

M_{0}

is a deformation retract of

M

. So, any diffeomorphism of

M

which is the identity on

M_{0}

will automatically induce the identity in homology. Thus we have an inclusion map

j : D i f f (M r e l M_{0}) ↪ D i f f_{ℱ} (M)

Proposition 7.6. The universal difference invariant

τ^{δ}

maps to zero under the induced map

j^{*} : H^{4 k} (B D i f f_{ℱ} (M); R) \to H^{4 k} (B D i f f (M r e l M_{0}); R) .

Proof. When the universal bundle with fiber $M$ is pulled back to $B D i f f (M r e l M_{0})$ it will contain a trivial $M_{0}$ bundle. The remainder is an $h$ -cobordism bundle for $\partial M$ . Therefore $τ^{δ}$ of this pull-back is zero by additivity and Lemma 7.3 . □

Proposition 7.7. There is a fibration sequence

D i f f (M r e l M_{0}) \to D i f f_{ℱ} (M) - \to r E m b_{ℱ} (M_{0}, i n t M)

where

E m b_{ℱ} (M_{0}, i n t M)

is the space of orientation preserving smooth embeddings

M_{0} \to i n t M

which is unipotent on

H_{*} (M_{0}; Q) \sim = H_{*} (M; Q)

with respect to the flag

ℱ

and

r

is the restriction map.

Proof. See the appendix of [7] . □

Lemma 7.8. Given unipotent smoothable fibrations

(Z, C)

and

Y

over

B

and any continuous mapping

f : C \to Y

over

B

the union

Y \cup_{C} Z

is also unipotent and smoothable with

τ^{δ} (Y \cup_{C} Z) = τ^{δ} (Y) + τ^{δ} (Z, C) .

Proof. Unipotence follows from the Mayer-Vietoris sequence for the fiber homology of $Y \cup_{C} Z$ . Additivity of torsion gives the formula. Smoothing is easy, very similar to the construction in subsection 4.3 . □

Lemma 7.9. For any unipotent smoothable fibration pair

(Z, C)

the fiberwise smash

Z /_{B} C

and fiberwise suspension

Σ_{B} (Z /_{B} C) \to B

(with fibers

X / A

and

Σ (X / A)

) are unipotent and smoothable with

τ^{δ} (Z /_{B} C) = τ^{δ} (Z, C) = - τ^{δ} (Σ_{B} (Z /_{B} C)) .

Proof. Taking $Y = B$ in the previous lemma we get $τ^{δ} (Z /_{B} C) = τ^{δ} (B \cup_{C} Z) = τ^{δ} (Z, C) = τ^{δ} (Z) - τ^{δ} (C) .$ Apply this to the cofibration sequence $X / A \lor X / A \to X / A \to Σ (X / A)$ to get $τ^{δ} (Σ_{B} (Z /_{B} C)) = τ^{δ} (Z /_{B} C) - 2 τ^{δ} (Z /_{B} C) = - τ^{δ} (Z /_{B} C) .$ Where we use the equation $τ^{δ} (X \lor Y) = τ^{δ} (X) + τ^{δ} (Y)$ which is another special case of Lemma 7.8 . □

The following lemma and its proof goes along the lines of a discussion I had with John R. Klein.

Lemma 7.10.

τ^{δ} (Z) = 0

for pointed unipotent smoothable fibrations with rationally acyclic fibers.

Proof. By pointed we mean that $Z$ has a section $B \to Z$ whose image is a copy of $B$ . By the previous lemma we may assume that the fiber $X$ has been suspended many more times than the dimension of $B$ . Since $X$ is a rationally trivial finite complex it has only finitely many homotopy classes of self maps. Thus there is a finite covering $\tilde{B}$ of $B$ so that the action of $π_{1} \tilde{B}$ on $X$ is trivial up to pointed homotopy. Since $H^{4 k} (B, R)$ maps monomorphically into $H^{4 k} (\tilde{B}; R)$ it suffices to show that the difference torsion of the pull back $\tilde{Z}$ of $Z$ to $\tilde{B}$ is trivial.
By assumption there is a smooth thickening $M \to E \to \tilde{B}$ of the bundle $\tilde{Z} \to \tilde{B}$ where $M$ is a compact $n$ -manifold homotopy equivalent to $X$ . By passing to the normal disk bundle we may assume that the vertical tangent bundle of $E$ is trivial.
Since the bundle has a section this means that $E$ contains a trivial $n$ -disk bundle $\tilde{B} \times D^{n}$ . Therefore the bundle is classified by a mapping $\tilde{B} \to B D i f f_{0} (M r e l D^{n})$ where $D i f f_{0}$ means diffeomorphisms tangentially homotopic to the identity.
Claim 1: We may assume that $D i f f_{0} (M r e l D^{n})$ is connected and therefore $B D i f f_{0} (M r e l D^{n})$ is simply connected.
Proof: Let $M_{0}$ be the closure of the complement of a collar neighborhood of $\partial M$ in $M$ . We may assume the disk $D^{n}$ lies in $M_{0}$ . By taking the product with a disk if necessary we may assume that $M_{0}$ contains as a deformation retract a spine $K ≃ X$ of dimension $< (n - 4 k) / 2$ . Let $f$ be any diffeomorphism of $M$ which is tangentially homotopic to the identity. By immersion theory, the restriction of $f$ to $M_{0}$ is isotopic by regular immersions to the inclusion map. Since $M_{0}$ has a small dimensional spine we can deform this immersion isotopy into an isotopy through embeddings. This extends to $M$ by isotopy extension. But $\partial M$ is simply connected.
So, its concordance space $C (\partial M)$ is connected by Cerf [8] . Thus $f$ is isotopic to the identity as claimed.
Claim 2: There is a fibration sequence: $E m b_{0} (M_{0}, i n t M r e l D^{n}) \to B D i f f (M r e l M_{0}) - - \to B j B D i f f_{0} (M r e l D^{n})$ where $E m b_{0}$ means embeddings isotopic to the inclusion map.
Proof: Choose a fixed embedding $φ_{0} : M \to R^{\infty}$ and let $E D i f f (M r e l D^{n})$ be the space of all embeddings $φ : M \to R^{\infty}$ which agree with $φ_{0}$ on $D^{n}$ . Then $E D i f f (M r e l D^{n})$ is contractible and admits a free action by $D i f f_{0} (M r e l D^{n})$ .
The quotient space $B D i f f_{0} (M r e l D^{n})$ is the space of all pairs $(W, [φ])$ where $W$ is a submanifold of $R^{\infty}$ containing $D^{n}$ and $[φ]$ is an isotopy class of diffeomorphisms $φ : M \to W$ which are equal to $φ_{0}$ on $D^{n}$ . The classifying space for $D i f f (M r e l M_{0})$ , $B D i f f (M r e l M_{0}) = E D i f f (M r e l D^{n}) / D i f f (M r e l M_{0}),$ is the space of all pairs $(W, ψ)$ where $W \subset R^{\infty}$ is as above and $ψ : M_{0} \to W$ is an embedding equal to $φ_{0}$ on $D^{n}$ and whose image has complement a collar neighborhood of $\partial W$ . Each $ψ$ determines an isotopy class $[φ]$ and, for $W = M$ , the space of all $ψ$ which give the isotopy class $[i d]$ of the identity is exactly $E m b_{0} (M_{0}, i n t M r e l D^{n})$ as claimed.
Claim 3: $E m b_{0} (M_{0}, i n t M r e l D^{n})$ is rationally trivial through degree $4 k$ .
Proof: By immersion theory and transversality this space has the $4 k$ homotopy type of the identity component of the space of all pointed maps $X \to X \times O$ which is rationally trivial since $X$ is rationally trivial.
One consequence of this is that the mapping $D j$ in Claim 2 is a rational homotopy equivalence through degree $4 k$ . Thus $D j$ induces an isomorphism in rational cohomology in degee $4 k$ . But $D j^{*} (τ^{δ}) = 0$ by Proposition 7.6 . Therefore, the universal class $τ^{δ}$ is trivial. □

The rest of the proof of the Lemma 7.5 follows suggestions of E. Bruce Williams.

Lemma 7.11.

τ^{δ} (Z) = 0

on all unipotent smoothable fibration.

Proof. Taking the fiberwise suspension we may assume that the fibers are pointed.
By Lemma 7.10 it suffices to reduce the rank of the rational homology of the fiber $X$ of a pointed smoothable fibration without changing the value of $τ^{δ} (Z)$ . We use the fact that stable homotopy groups are rationally the same as reduced homology.
Let $n$ be maximal so that the reduced rational homology ${\bar{H}}_{n} (X; Q)$ is nonzero.
Choose one generator which is fixed by the action of $π_{1} B$ . (This exists since the action is unipotent.) Assuming that $X$ has been suspended a large number of times, a multiple of this generator is represented by some $α \in π_{n} (X)$ . Multiplying by the order of the torsion subgroup of $π_{n} (X)$ we may assume that $α$ represents a rationally nontrivial element of $π_{n} (X)$ which is fixed by the action of $π_{1} B$ .
The homotopy groups $π_{i} (X)$ will be torsion and thus finite for all $n < i < n + dim B$ . Let $m$ be the product of the orders of all of these groups. Then, by obstruction theory, $m α$ is represented by a fiber preserving continuous mapping $f : (B \times S^{n}, B \times *) \to (Z, B) .$ The fiberwise mapping cone $C_{B} (f)$ of $f$ has the same torsion: $τ^{δ} (C_{B} (f)) = τ^{δ} (Z)$ by Lemma 7.8 and its fiber $X \cup e^{n + 1}$ has less rational homology than $X$ . The lemma follows by induction on the rank of the rational homology of $X$ . □

This completes the proof of Lemma 7.5 which implies the Main Theorem 3.3 .

8 Existence of higher torsion

In this section we show that higher Miller-Morita-Mumford classes

M_{2 k}

and higher Franz-Reidemeister torsion

τ_{2 k}

are linearly independent higher torsion invariants.

8.1 Miller-Morita-Mumford classes

p : (E, \partial_{0}) \to B

is any smooth bundle pair (not necessarily unipotent) with compact fiber

(F, \partial_{0})

, the (higher relative) Miller-Morita-Mumford class

M_{2 k}

is defined to be the integral cohomology classes given by

M_{2 k} (E, \partial_{0}) : = t r_{B}^{(E, \partial_{0})} ((2 k)! c h_{4 k} (T^{v} E)) \in H^{4 k} (B; Z)

for

k \geq 1

where

T^{v} E

is the vertical tangent bundle of

E

c h_{4 k} (ξ) : = \frac{1}{2} c h_{4 k} (ξ \otimes C)

is the Chern character and

t r_{B}^{(E, \partial_{0})}

is the relative transfer defined in ( 4.1 ). In degree 0 this formula gives a half integer:

M_{0} (E, \partial_{0}) = \frac{n}{2} χ (F, \partial_{0}) = \frac{n}{2} (χ (F) - χ (\partial_{0} F))

where

n = dim F

. The properties of the transfer [1] give the following properties of the higher Miller-Morita-Mumford classes.

Lemma 8.1 (vertical additivity of

M_{2 k}

). Suppose that

E

is the union of two subbundle

E = E_{1} \cup E_{2}

and

E_{1} \cap E_{2} = \partial_{0} E_{2} \subseteq \partial_{1} E_{1}

. Then

M_{2 k} (E, \partial_{0} E_{1}) = M_{2 k} (E_{1}, \partial_{0}) + M_{2 k} (E_{2}, \partial_{0}) .

Proof. Choose a vertical tangent vector field for $E$ which points inward along $\partial_{0} E_{1}$ , outward along the rest of $\partial^{v} E$ and points from $E_{1}$ to $E_{2}$ along $\partial_{0} E_{2}$ . Then the zero set of this vector field, say $Z$ , is a disjoint union $Z = Z_{1} \prod Z_{2}$ where $Z_{i} \subset E_{i} - \partial^{v} E_{i}$ .
$M_{2 k} (E, \partial_{0} E_{1})$ is given by restricting a multiple of the Chern character of $T^{v} E$ to $Z$ and pushing down to $B$ . (See [15] for a detailed discussion of this.) This is obviously the sum of the push-downs of the restrictions to $Z_{1}$ and $Z_{2}$ which give $M_{2 k} (E_{1}, \partial_{0})$ and $M_{2 k} (E_{2}, \partial_{1})$ . □

Proposition 8.2. For all

k > 0

we have

(1) (stability) $M_{2 k} (E \times I) = M_{2 k} (E)$ .
(2) (relative formula) $M_{2 k} (E, \partial_{0}) = M_{2 k} (E) - M_{2 k} (\partial_{0} E)$ .
(3) (additivity) If $E_{1}, E_{2}$ are smooth bundles over $B$ with the same vertical boundary $E_{1} \cap E_{2} = \partial^{v} E_{1} = \partial^{v} E_{2}$ then $M_{2 k} (E_{1} \cup E_{2}) = M_{2 k} (E_{1}) + M_{2 k} (E_{2}) - M_{2 k} (\partial^{v} E_{1}) .$
(4) (transfer) If $q : D \to E$ is a bundle with fiber $X$ then $M_{2 k} (D)_{B} \in H^{4 k} (B)$ and $M_{2 k} (D)_{E} \in H^{4 k} (E)$ are related by $M_{2 k} (D)_{B} = χ (X) M_{2 k} (E) + t r_{B}^{E} (M_{2 k} (D)_{E}) .$

Remark 8.3. The relative formula and additivity in the case

E_{1} \sim = E_{2}

are important since they imply that the relative invariant

M_{2 k} (E, \partial_{0})

is related to the closed fiber case in the way that higher torsion invariants are supposed to behave.

Proof. The transfer formula follows from the formula

T_{B}^{v} D = T_{E}^{v} D \oplus q^{*} T^{v} E

and the additivity of the Chern character:

\begin{matrix} M_{2 k} (D)_{B} & = t r_{B}^{D} ((2 k)! c h_{4 k} T_{B}^{v} D) \end{matrix}

\begin{matrix} = t r_{B}^{E} t r_{E}^{D} ((2 k)! c h_{4 k} (T_{E}^{v} D)) + t r_{B}^{E} t r_{E}^{D} (q^{*} (2 k)! c h_{4 k} (T^{v} E)) \end{matrix}

\begin{matrix} = t r_{B}^{E} (M_{2 k} (D)_{E}) + χ (X) M_{2 k} (E) \end{matrix}

\begin{matrix}  \end{matrix}

Stability is a special case of the transfer formula.

Vertical additivity (Lemma 8.1 ) applied to

E \sim = E \cup \partial_{0} E \times I

gives

M_{2 k} (E, \partial_{0}) = M_{2 k} (E) - M_{2 k} (\partial_{0} E \times I)

By stability this gives the relative formula.

Vertical additivity applied to

E = E_{1} \cup E_{2}

gives

M_{2 k} (E) = M_{2 k} (E_{1}, \partial^{v}) + M_{2 k} (E_{2})

This gives additivity by the relative formula. □

Lemma 8.4. If

S^{2 n} (ξ) \to B

is the

S^{2 n}

-bundle associated to an

S O (2 n + 1)

-bundle

ξ

over

B

then

M_{2 k} (S^{2 n} (ξ)) = 2 (2 k)! c h_{4 k} (ξ) .

Proof. This follows from additivity and the elementary calculation $M_{2 k} (D^{m} (ξ)) = (2 k)! c h_{4 k} (ξ)$ for any linear disk bundle $D^{m} (ξ)$ . □

Theorem 8.5. The higher Miller-Morita-Mumford class

M_{2 k}

is a nontrivial even higher torsion invariant for all

k > 0

with

s_{1} = 0

and

s_{2} = (2 k)!

Proof. The additivity axiom follows from additivity (Proposition 8.2 .3) applied to $E_{1} \cup E_{2}$ , $D E_{1}$ and $D E_{2}$ . The transfer axiom is a special case of the transfer formula (Proposition 8.2 .4). The calculation of $s_{2}$ follows from the above lemma. Finally, $s_{1} = 0$ since $e (E) = (- 1)^{n} e (E)$ for an oriented bundle with closed $n$ -dimensional fiber. □

8.2 Higher FR-torsion

Higher Franz-Reidemeister (FR)-torsion invariants are real characteristic class

τ_{2 k} (E, \partial_{0}) \in H^{4 k} (B; R)

for

k \geq 1

defined for smooth bundle pairs

(F, \partial_{0}) \to (E, \partial_{0}) \to B

which are relatively unipotent in the sense that

H_{*} (F, \partial_{0} F; Q)

is unipotent as a

π_{1} B

module.

(See [13] .) The main tool for computing the higher FR-torsion is the framing principle which is proved in a special case in [13] and in general in [15] .

We will state the framing principle leaving unanswered the difficult question:

What exactly is a “family of chain complexes” and how is the higher torsion of such a family defined? Suppose that

(E, \partial_{0}) \to B

is a relatively unipotent smooth bundle pair and

f : E \to R

is a fiberwise oriented generalized Morse function (i.e., it has Morse and birth-death critical point and an orientation of the negative eigenspace of

D^{2} f_{t}

along the Morse point sets so that they cancel with positive incidence along the birth-death points). Suppose that the vertical gradient

\nabla^{v} f

f

with respect to some metric points inward along

\partial_{0} E

and outward along

\partial_{1} E

and point from

\partial_{0} E

towards

\partial_{1} E

along

\partial_{0} E \cap \partial_{1} E

(i.e., the zero set

Σ

\nabla^{v} f

is dual to the relative Euler class

e (E, \partial_{0})

Theorem 8.6 (Framing Principle [15] ). Let

C (f_{t}), t \in B

be the family of cellular chain complexes obtained from the functions

f_{t}

. Then

τ_{2 k} (E, \partial_{0}) = τ_{2 k} (C (f_{t})) + (- 1)^{k} ζ (2 k + 1) p_{*}^{Σ} (c h_{2 k} (ξ))

Here

ξ

is the negative eigenspace bundle of

D^{2} f_{t}

along the singular set, the push-down is the alternating sum

p_{*}^{Σ} = \sum (- 1)^{i} p_{*}^{i}

where

p_{*}^{i}

is the restriction of the push-down on the set

Σ^{i} (f_{t})

of critical points of index

i

Remark 8.7. The higher FR-torsion of a family of based free chain complexes

C (f_{t})

is defined provided that its rational homology is unipotent. We call

τ_{2 k} (C (f_{t}))

the algebraic torsion. The other summand is called the correction term.

An easy example of this rule is given by a linear

n

-sphere bundle

S^{n} (ξ)

associated with an oriented

n

-plane bundle

ξ

over

B

. This has a fiberwise Morse function

f

having exactly two critical point in each fiber of index

0

and

n

. The family of chain complexes

C (f_{t})

is constant and therefore has trivial algebraic torsion. The negative eigenspace bundle is trivial over

Σ^{0} (f_{t})

and isomorphic to

ξ

over

Σ^{n} (f_{t})

Therefore the framing principle gives the following.

Corollary 8.8. The higher FR-torsion of an oriented linear sphere bundle is given by

τ_{2 k} (S^{n} (ξ)) = (- 1)^{n + k} ζ (2 k + 1) c h_{2 k} (ξ)

ξ

is an oriented

n

-plane bundle. □

Theorem 8.9. [13] The higher FR-torsion

τ_{2 k} (E, \partial_{0}) \in H^{4 k} (B; R)

is defined for relatively unipotent smooth bundle pairs

(E, \partial_{0}) \to B

and satisfies the following conditions.

(1) (relative formula) If $E$ and $\partial_{0} E$ are unipotent bundles then $τ_{2 k} (E, \partial_{0}) = τ_{2 k} (E) - τ_{2 k} (\partial_{0} E) .$
(2) (additivity) If $(E_{1}, \partial_{0}), (E_{2}, \partial_{0})$ are relatively unipotent smooth bundles over $B$ with $E_{1} \cap E_{2} = \partial_{0} E_{2} \subseteq \partial_{1} E_{1}$ then $τ_{2 k} (E_{1} \cup E_{2}, \partial_{0} E_{1}) = τ_{2 k} (E_{1}, \partial_{0}) + τ_{2 k} (E_{2}, \partial_{0}) .$
(3) (stability) If $q : D \to E$ is a linear disk bundle and $\partial_{0} D = q^{- 1} (\partial_{0} E)$ then the higher FR-torsion of $(D, \partial_{0})$ as a bundle pair over $B$ is equal to the higher torsion of $(E, \partial_{0})$ : $τ_{2 k} (D, \partial_{0})_{B} = τ_{2 k} (E, \partial_{0}) .$

Proof. The proofs of these theorems, given in great detail in [13] , can be summarized as follows. For additivity we may assume that there is a fiberwise oriented generalized Morse function (GMF) $f$ on $(E, \partial_{0})$ whose restriction to $(E_{1}, \partial_{0})$ and $(E_{2}, \partial_{0})$ is suitable for defining their torsion. Then the family of chain complexes $C (f_{t})$ fits into a short exact sequence $0 \to C (f | E_{1}) \to C (f_{t}) \to C (f | E_{2}) \to 0$ where $C (f | E_{i})$ is the family of chain complexes associated to the restriction of $f$ to $E_{i}$ . However, algebraic torsion is additive for short exact sequences, i.e., $τ_{2 k} (C (f_{t})) = τ_{2 k} (f | E_{1}) + τ_{2 k} (f | E_{2})$ . The correction term is also additive since the critical set of $f_{t}$ will be a disjoint union of the critical sets of $f | E_{1}$ and $f | E_{2}$ .
For stability we note that a fiberwise oriented GMF $f$ for $E$ gives a fiberwise oriented GMF $σ f$ on any linear disk bundle $p : D \to E$ by $σ f (x) = f (p (x)) + ∥ x ∥^{2}$ .
This has the same cellular chain complex and the same critical set with the same negative eigenspace bundles. So, the torsion is unchanged.
The relative formula follows formally from additivity in the case $E \sim = E \cup \partial_{0} E \times I$ and stability which gives $τ_{2 k} (\partial_{0} E \times I) = τ_{2 k} (\partial_{0} E)$ . □

Higher FR-torsion and the higher Miller-Morita-Mumford classes are related by:

Theorem 8.10. [15] If

E \to B

is a unipotent smooth bundle with closed even dimensional fibers and

k \geq 1

then

τ_{2 k} (E) = (- 1)^{k} \frac{ζ (2 k + 1)}{2 (2 k)!} M_{2 k} (E) .

Proof. We explain the proof only in the case when there exists a fiberwise oriented GMF $f_{t}$ on $E$ . Then $- f_{t}$ is also a fiberwise oriented GMF. The negative eigenspace bundle of $- D^{2} f_{t}$ is equal to the positive eigenspace bundle $η$ for $D^{2} f_{t}$ . Also, the index of the critical points change from $i$ to $n - i$ . So, the framing principle gives:
$τ_{2 k} (E) = τ_{2 k} (C (f_{t})) + (- 1)^{k} ζ (2 k + 1) p_{*}^{Σ} (c h_{2 k} (ξ))$ $τ_{2 k} (E) = τ_{2 k} (C (- f_{t})) + (- 1)^{n + k} ζ (2 k + 1) p_{*}^{Σ} (c h_{2 k} (η))$ However, the algebraic torsion has the property that $τ_{2 k} (C (- f_{t})) = (- 1)^{n} τ_{2 k} (C (f_{t}))$ .
This is the involution property. Therefore, for $n = dim F$ even, we get the following by adding the above two equations. $2 τ_{2 k} (E) = (- 1)^{k} ζ (2 k + 1) p_{*}^{Σ} (c h_{2 k} (T^{v} E)) = (- 1)^{k} ζ (2 k + 1) (\frac{1}{(2 k)!} M_{2 k} (E))$ The theorem follows by dividing by 2. □

This theorem says that

M_{2 k}

is proportional to the even part of

τ_{2 k}

. The following calculation shows that the odd part of

τ_{2 k}

has the same size (but opposite sign).

Theorem 8.11. [13] The

S^{n}

-bundle associated to the

S O (n + 1)

-bundle

ξ

over

B

has higher FR-torsion

τ_{2 k} (S^{n} (ξ)) = (- 1)^{n + k} ζ (2 k + 1) c h_{4 k} (ξ)

Proof. It follows from the stability of $τ_{2 k}$ that $τ_{2 k} = 0$ on all linear disk bundles over $B$ . Since the $S^{n + 1}$ bundle association to $ξ$ is the union of two disk bundles along their common boundary $S^{n} (ξ)$ we have by additivity (Theorem 8.9 .2) that $τ_{2 k} (S^{n + 1} (ξ)) = 2 τ_{2 k} (D^{n + 1} (ξ)) - τ_{2 k} (S^{n} (ξ)) = - τ_{2 k} (S^{n} (ξ))$ The left hand side is equal to $(- 1)^{n + 1 + k} ζ (2 k + 1) c h_{2 k} (ξ)$ by Corollary 8.8 ). The theorem follows. □

Corollary 8.12. Higher Franz-Reidemeister torsion

τ_{2 k}

is a higher torsion invariant with

s_{n} = \frac{1}{2} (- 1)^{n + k} ζ (2 k + 1)

Proof. The additivity axiom follows from Theorem 8.9 .2. For the transfer axiom we need to prove the following for any smooth unipotent bundle

E \to B

with closed fiber

F

and any oriented linear sphere bundle

S^{m} (ξ) \to E

\begin{matrix} τ_{2 k} (S^{m} (ξ))_{B} = χ (S^{m}) τ_{2 k} (E) + t r_{B}^{E} (τ_{2 k} (S^{m} (ξ))_{E}) \end{matrix}

(8.1)

There are four cases depending of the parity of

m

and

n = dim F

n, m

have the same parity then the transfer formula ( 8.1 ) is equivalent to the transfer formula for

M_{2 k}

which we already proved. (If

n, m

are both odd the RHS is zero for both

τ_{2 k}

and

M_{2 k}

.) If

n

is even and

m

is odd then, by the previous case we know that

\begin{matrix} τ_{2 k} (S^{m + 1} (ξ))_{B} = 2 τ_{2 k} (E) + t r_{B}^{E} (τ_{2 k} (S^{m + 1} (ξ))_{E}) . \end{matrix}

(8.2)

But, the LHS of this is

\begin{matrix} τ_{2 k} (S^{m + 1} (ξ))_{B} = 2 τ_{2 k} (D^{m + 1} (ξ))_{B} - τ_{2 k} (S^{m} (ξ))_{B} = 2 τ_{2 k} (E) - τ_{2 k} (S^{m} (ξ))_{B} \end{matrix}

(8.3)

by additivity. Combining ( 8.2 ) and ( 8.3 ) we get

τ_{2 k} (S^{m} (ξ))_{B} = - t r_{B}^{E} (τ_{2 k} (S^{m + 1} (ξ))_{E}) .

However,

τ_{2 k} (S^{m + 1} (ξ))_{E} = - τ_{2 k} (S^{m} (ξ))_{E}

by Theorem 8.11 . So, ( 8.1 ) holds in this case.

In the case

n

odd and

m

even, the LHS of ( 8.3 ) is zero since it is a multiple of

M_{2 k} (S^{m + 1} (ξ))_{B} = χ (S^{m + 1}) M_{2 k} (E) + t r_{B}^{E} (M_{2 k} (S^{m + 1} (ξ))_{E}) = 0 + 0 .

and the second term on the RHS of ( 8.1 ) is zero since the transfer map is zero.

So Equation ( 8.1 ) holds proving the transfer axiom in all four parity cases. □

Corollary 8.13. The odd part of higher FR-torsion is given by

τ_{2 k}^{-} (E) = τ_{2 k} (E) - (- 1)^{k} \frac{ζ (2 k + 1)}{2 (2 k)!} M_{2 k} (E) .

Remark 8.14. A corollary of the main theorem is that any odd torsion invariant must be proportional to the above expression. In particular it is expected that the nonequivariant analytic torsion classes of [3] and [2] are odd torsion invariants and therefore proportional to

τ_{2 k}^{-}

References

J.C. Becker and D.H. Gottlieb, The transfer and fiber bundles, Topology 14 (1975), 1–12.
Jean-Michel Bismut and Sebastian Goette, Families torsion and Morse functions, Astérisque (2001), no. 275, x+293.
Jean-Michel Bismut and John Lott, Flat vector bundles, direct images and higher real analytic torsion, J. Amer. Math. Soc. 8 (1995), no. 2, 291–363.
M. Bökstedt, The rational homotopy type of $Ω W h^{D i f f} (*)$ , LNM 1051, Springer-Verlag, 1984, pp. 25–37.
M. Bökstedt and F. Waldhausen, The map $B S G \to A (*) \to Q S^{0}$ , Algebraic Topology and Algebraic $K$ -theory (William Browder, ed.), Annals of Math. Studies, vol. 113, 1987, pp. 418–431.
U. Bunke, Higher analytic torsion of sphere bundles and continuous cohomology of $D i f f (S^{2 n - 1})$ , preprint, math.DG/9802100.
Jean Cerf, Sur les difféomorphismes de la sphère de dimension trois $(γ_{4} = 0)$ , Springer-Verlag, Berlin, 1968.
, La stratification naturelle des espaces de fonctions différentiables réelles et le théorème de la pseudo-isotopie, Inst. Hautes Études Sci. Publ. Math. (1970), no. 39, 5–173.
W. Dwyer, M. Weiss, and B. Williams, A parametrized index theorem for the algebraic $K$ -theory Euler class, Acta Math. 190 (2003), no. 1, 1–104.
Sebastian Goette, Morse theory and higher torsion invariants I, math.DG/0111222.
, Morse theory and higher torsion invariants II, math.DG/0305287.
Kiyoshi Igusa, The stability theorem for smooth pseudoisotopies, $K$ -Theory 2 (1988), no. 1–2, vi+355.
, Higher Franz-Reidemeister Torsion, AMS/IP Studies in Advance Mathematics, vol. 31, International Press, 2002.
, Axioms for higher torsion, preprint, 2003.
, Higher complex torsion and the framing principle, math.KT/0303047, to appear in Memoires of AMS, 2003.
Kiyoshi Igusa and John Klein, The Borel regulator map on pictures. II. An example from Morse theory, $K$ -Theory 7 (1993), no. 3, 225–267.
John Klein, The cell complex construction and higher R-torsion for bundles with framed Morse function, Ph.D. thesis, Brandeis University, 1989.
, Higher Franz-Reidemeister torsion and the Torelli group, Mapping Class Groups and Moduli Spaces, Contemp. Math., vol. 150, American Math. Soc., 1993, pp. 195–204.
Xiaonan Ma, Formes de torsion analytique et familles de submersions, C. R. Acad. Sci. Paris Sér. I Math. 324 (1997), no. 2, 205–210.
J. B. Wagoner, Diffeomorphisms, $K_{2}$ , and analytic torsion, Algebraic and Geometric Topology (Proc. Sympos. Pure Math., Stanford Univ., Stanford, Calif., 1976), Part 1, Amer. Math. Soc., Providence, R.I., 1978, pp. 23–33.

Kiyoshi Igusa