A GENERAL CONNECTED SUM FORMULA FOR THE

FAMILIES BAUER-FURUTA INVARIANT

JOSHUA TOMLIN
arXiv:2510.14201v1 [math.DG] 16 Oct 2025

Abstract. The Bauer-Furuta invariant of a family of smooth 4-manifolds is

a stable cohomotopy refinement of the families Seiberg-Witten invariant and
is constructed from a finite dimensional approximation of the Seiberg-Witten
monopole map. We prove a general formula for the families Bauer-Furuta
invariant of a fibrewise connected sum, extending Bauer’s non-parameterised
formula [8]. In a subsequent paper [28], we will use this formula to derive a
general connected sum formula for the families Seiberg-Witten invariant which
incorporates both the families blow-up formula of Liu [22] and the gluing
formula of Baraglia-Konno [5].

1. Introduction

The Bauer-Furuta invariant [9] of a 4-manifold is a stable cohomotopy refinement

of its integer valued Seiberg-Witten invariant. Specifically, it is the equivariant
stable cohomotopy class of a finite dimensional approximation of the Seiberg-Witten
monopole map. This approach takes a new perspective of studying the monopole
map, rather than its moduli space of solutions. It is possible to recover the Seiberg-
Witten invariant from the Bauer-Furuta invariant, hence techniques from algebraic
topology can be used to circumvent laborious analytical arguments.

In subsequent work, Bauer derived a formula [8] for the Bauer-Furuta invariant of
a connected sum of 4-manifolds. His idea was to analyse behaviour
` of monopoles
on a 4-manifold with an n-component separating neck N (L) = n S 3 × [−L, L]
of varying length 2L. He showed that given a 4-manifold with a separating neck,
ends of the necks can be permuted without changing the Bauer-Furuta class of the
monopole map. The key insight was that monopoles decay exponentially towards
the middle of the neck, hence stretching the neck could be used to control the
dynamics in the middle.

Since Donaldson’s suggestion in 1996 [14], there has been much interest in studying
the Seiberg-Witten equations of 4-manifold families. Several authors including
Li-Liu, Nakamura and Ruberman have generalised Seiberg-Witten theory to the
families setting [21, 23, 25]. This body of work involves wall crossing formulas, non-
existence of positive scalar curvature metrics, and a particularly noteworthy families
1
2 JOSHUA TOMLIN

blow-up formula [22] due to Liu. One striking application of families Seiberg-
Witten theory applied to mapping tori is the construction of 4-manifolds with
diffeomorphisms that are continuously homotopic to the identity, but not smoothly
homotopic [24].

Since 2019, Baraglia has contributed to the theory of families Seiberg-Witten in-
variants in several papers [3–5]. In [5], Baraglia-Konno proved a connected sum
formula for the families Seiberg-Witten invariant under some restrictive assump-
tions. These assumptions simplified the moduli space of one of the summands and
avoided cases involving chambers. The overarching goal of this paper and upcom-
ing work [28] is to derive a completely general connected sum formula for families
Seiberg-Witten invariants extending both Baraglia-Konno’s formula and Liu’s fam-
ilies blow-up formula.

This is accomplished by first proving a similar result for the families Bauer-Furuta
invariant. Szymik illustrated in [27] that the Bauer-Furuta invariant naturally ex-
tends to the families setting. In this paper, we prove the following families connected
sum formula, generalising Bauer’s formula for the unparameterised case.

Theorem 1.1. For j ∈ {1, 2}, let Ej → B be a smooth family of closed, oriented 4-
manifolds equipped with a spinc structure sj on the vertical tangent bundle. Assume
a section ij : B → Ej exists with normal bundle Vj and suppose that φ : V1 → V2
is an orientation reversing isomorphism satisfying
∼ i∗ (sE ).
φ(i∗ (sE )) =
1 1 2 2

Then the families Bauer-Furuta class of the fiberwise connected sum E = E1 #B E2

is
(1.1) [µE ] = [µE1 ] ∧J [µE2 ].

In 2021, Baraglia-Konno demonstrated how to recover the families Seiberg-Witten

invariant from the families Bauer-Furuta invariant via a formulation of the families
Seiberg-Witten invariant in equivariant cohomology [6]. In upcoming work [28], we
will use this formulation and the above formula to prove a connected sum formula
for the families Seiberg-Witten invariant.

2. Finite dimensional approximation

The Bauer-Furuta invariant is obtained from the stable homotopy class of an ap-
proximation of the Seiberg-Witten monopole map by finite dimensional subspaces.
In [9], two methods of finite dimensional approximation are described, one method
due to Schwarz [29] and one due to Bauer-Furuta. The Bauer-Furuta method is
useful for formally defining the invariant, while the Schwarz method is more use-
ful for practical calculations. Bauer further clarifies their construction in [7] using
Spanier-Whitehead spectra. We begin by reviewing these two constructions and
showing that they are equivalent.
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
3

Let X and Y denote pointed topological spaces. We will assume that all maps
f : X → Y are continuous and basepoint preserving. Denote by [X, Y ] the set of
based homotopy classes of maps between X and Y . Let S n denote the unit sphere
in R ⊕ Rn with ∞ = (1, 0) ∈ S n as the basepoint. The n-th homotopy group of X
is
πn (X) = [S n , X].

The suspension functor ΣX = S 1 ∧ X defines a map of homotopy groups

Σ : πn (Σn X) → πn+1 (Σn+1 X).

The Freudenthal suspension theorem [16] states that this map is an isomorphism
for large enough n and the n-th stable homotopy group is defined by
πns (X) = Colim πn+k (Σk X).
−→k

In the stable range, the homotopy group πn+k (Σk X) = [S n+k , Σk X] does not
depend on the dimension of the domain and codomain, but only on the difference
in dimensions. In the opposite fashion, the n-th cohomotopy set of X is given by
π n (X) = [X, S n ]. The functor π n is now contravariant, but suspension still defines
a map Σ : π n (X) → π n+1 (ΣX). The n-th stable cohomotopy group of X is defined
as
πsn (X) = Colim π n+k (Σk X).
−→k

The stable cohomotopy groups define a generalised cohomology theory, and Brown’s
representability theorem [12] guarantees that this cohomology theory is repre-
sentable. The natural objects for representing stable cohomotopy groups are spec-
tra, in particular, the sphere spectrum Sn represents the above groups. In order to
define the Bauer-Furuta invariant, it will be more convenient to work with spaces
that are indexed by finite dimensional subspaces of an infinite dimensional Hilbert
space. This is more general than indexing by the natural numbers and allows us to
keep track of coordinates when taking suspensions.

Let G be a compact lie group. For our purposes, G will always be a product of
circles. A G-space is a pointed topological space X with a continuous left action
G × X → X that fixes the basepoint. For two G-spaces X and Y , let [X, Y ]G
denote the set of homotopy classes through equivariant pointed maps. The diagonal
subgroup of G × G naturally defines a G-action on the smash product X ∧ Y .

Definition 2.1. A G-universe U is an infinite dimensional separable Hilbert space

which G acts on by isometeries. It is required that U contains the trivial repre-
sentation and that for any irreducible G-module M , HomG (M, U) is either zero or
infinite dimensional.

The above condition on HomG (M, U) guarantees that if we ever suspend by an

irreducible representation M , then we can suspend by M an arbitrary number of
times. A G-universe is called complete if it contains a copy of every irreducible
representation [19].
4 JOSHUA TOMLIN

For any subspace U ⊂ U let SU denote the unit sphere in R⊕U , which has a natural
basepoint ∞ = (1, 0) ∈ R ⊕ U . If U is finite dimensional, then SU is the one-point
compactification of U . For any direct sum V ⊕ U , we have SV ⊕U = SV ∧ SU . We
say that U is a subrepresentation if it is G-invariant. In this case the G-action can
be extended to SU ⊂ R ⊕ U by acting trivially on the R component. Since G acts
orthogonally, this fixes the basepoint of SU .

Definition 2.2. A G-spectrum A = {AU } (indexed by U) is a collection of G-

spaces indexed by subrepresentations U ⊂ U. Additionally, for any subrepresenta-
tion W ⊃ U with orthogonal decomposition W = V ⊕ U , there is an equivariant
structure homeomorphism

σU,W : SV ∧ AU → AW .

The structure maps have the property that for any other subrepresentation W ′ ⊃ W
with W ′ = V ′ ⊕ W orthogonally, the following diagram commutes up to homotopy.
σU,W ′
SV ′ ⊕V ∧ AU AW ′
(2.1) = σW,W ′
id∧σU,W
SV ′ ∧ SV ∧ AU SV ′ ∧ AW

Definition 2.3. The set of morphisms HomU (A, B) between two G-spectra A and
B, both indexed by U , is

HomG,U (A, B) = Colim[AU , BU ]G .

U ⊂U

This colimit is taken over morphisms of the form

idS ∧−
[AU , BU ]G −→
V
[SV ∧ AU , SV ∧ BU ]G = [AW , BW ]G

for W = V ⊕U orthogonally. The identification of [SV ∧AU , SV ∧BU ] with [AW , BW ]

A B
is given by the structure maps σU,W and σU,W .

From the above definition, we see that morphisms between spectra are only defined
stably and up to homotopy. This means to define a G-spectrum A up to isomor-
phism, it is enough to specify AU only for subrepresentations U in an indexing set
that is cofinal in the directed system of subrepresentations of U .

Example 2.4 (Suspension Spectrum): For any G-space A, define the suspension
spectrum ΣA by

(ΣA)U = SU ∧ A.

For W = V ⊕ U orthogonally, the structure map σU,W : SV ∧ (SU ∧ A) → SW ∧ A

is just the identity. Further, a map f : A → B induces a map Σf : ΣA → ΣB
of spectra by taking smash products with the identity. Thus Σ embeds pointed
topological spaces as a full subcategory inside the category of spectra. We write
SnG to denote the suspension spectrum of S n .
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
5

More generally, for any finite dimensional subrepresentation V ⊂ U define the

suspension ΣV A of a G-spectrum A by
(ΣV A)U = SV ∧ AU .
The associated structure maps are the obvious ones induced by smash products
with the identity.

Example 2.5 (Desuspension): Fix a finite dimensional subrepresentation V ⊂ U .

For any subrepresentation W containing V , write W = V ⊕ U orthogonally and
define the desuspension Σ−V A by
(Σ−V A)W = AU .
This defines Σ−V A up to isomorphism since the set of subrepresentations containing
V is cofinal in the directed system of subrepresentations of U . The set of morphisms
between Σ−V A and another G-spectrum B is given by
HomU (Σ−V A, B) = HomU (A, ΣV B).
That is, Σ−V is the left adjoint of ΣV .

Example 2.6 (Smash product of spectra): Let A be a G1 -spectrum indexed by U

and B be a G2 -spectrum indexed by V. The smash product A ∧ B is a G1 × G2 -
spectrum indexed by the universe U ⊕ V and, for subrepresentations U ⊂ U and
V ⊂ V,
(A ∧ B)U ⊕V = AU ∧ BV .
Let WU = U ′ ⊕U and WV = V ′ ⊕V orthogonally. The structure map σU ⊕V,WU ⊕WV
is defined by the following diagram.
σU ⊕V,WU ⊕WV
SU ′ ⊕V ′ ∧ (A ∧ B)U ⊕V (A ∧ B)WU ⊕WV
(2.2) = =
σU,WU ∧σV,WV
(SU ′ ∧ AU ) ∧ (SV ′ ∧ BV ) AWU ∧ BWV

The motivating principle behind defining these objects is that spectra represent
equivariant stable cohomology theories. In this case, let B be a compact topological
space and fix a universe U. Let λ be an equivariant K-theory element λ ∈ RO(B).
Write λ = E − F where E and F are honest finite dimensional vector bundles
over B. Assume without loss generality that F = B × V is trivial with V ⊂ U a
subrepresentation. Let T E be the Thom space of E and define the Thom spectrum
of λ by
T λ = Σ−V T E.

Definition 2.7. The n-th equivariant stable cohomotopy group of B with coeffi-
cients in λ is
n
πG,U (B; λ) = HomG,U (T λ, Sn )
(2.3) = Colim[SU ∧ T E, SU ∧ SV ∧ S n ]G .
U ⊥V
6 JOSHUA TOMLIN

2.1. Bauer-Furuta Approximation. Fix a G-universe U. For simplicity, we will

assume that HomG (M, U ) is only non-zero for finitely many isomorphism classes of
irreducible G-modules M . Now the isotypical decomposition of U guarantees that
any finite dimensional subspace V ⊂ U is contained in a G-invariant subspace. Let
B be a finite CW complex, which implies that B is compact and Hausdorff. We let
G act on B trivially.

Let H ′ , H → B be G-Hilbert bundles, by which we mean locally trivially fibre

bundles over B with standard fibre U and fibre preserving, fibrewise orthogonal
G-action. Fix an equivariant bundle map l : H ′ → H that is fibrewise linear
Fredholm.

Definition 2.8. An equivariant bundle map f : H ′ → H is Fredholm (relative to l)

if c = f − l is continuous and compact. That is, c maps disk bundles to precompact
sets.

A disk bundle D ⊂ H is a subbundle where each fibre is a closed disk of constant

finite radius. We say that a Fredholm map f is bounded if the preimage of any disk
bundle is contained in a disk bundle. For any subbundle V ⊂ H we write S(V ) to
denote the unit sphere of V and set SV = S(R ⊕ V ). The fibre (SV )b over b ∈ B
is a sphere with natural choice of basepoint ∞b = (1, 0) ∈ (SV )b . Let B∞ ⊂ SH
denote the image of the section at infinity. We identify H = SH \ B∞ through
fibrewise stereographic projection. The boundedness condition for f is equivalent
to f admitting a continuous, basepoint preserving extension f : SH ′ → SH . Note
that this extension is equivariant since G acts orthogonally.

Kupier’s theorem [20] applied to the isotypical decomposition of H implies that

there is an equivariant trivialisation H → U × B and all such trivialisations are
homotopic. Fix a trivialisation and let p : H → U be projection onto the first
factor. For f : H ′ → H bounded Fredholm, we will often abuse notation by writing
f : H ′ → U to also denote f composed with this projection. With this notation in
mind, the extension f : SH ′ → SU factors through the Thom space T H ′ = SH ′ /B∞ .

Let V ⊂ U be a closed subrepresentation with pV : U → V the orthogonal projec-

tion. The orthogonal decomposition U = V ⊕V ⊥ identifies SV = S(R⊕V ⊕0) ⊂ SU
and SV ⊥ = S(R ⊕ 0 ⊕ V ⊥ ) ⊂ SU . The spheres S(V ⊥ ) and SV are disjoint subsets
of SU and there is a deformation retraction ρV : SU \ S(V ⊥ ) → SV defined by
1
(2.4) ρV (t, v, v ′ ) = p (t, v, 0).
t2 + |v|2
This deformation retraction has the property that if h ∈ U \ V ⊥ , then ρV (h) =
λ(h)p(h) for some positive and continuous function λ : U \ V ⊥ → R.

For any finite dimensional subrepresentation V ⊂ U , set V ′ = l−1 (V ) and V =

V × B. Let pV , pV ′ be orthogonal projections onto V and V ′ respectively. We say
that V surjects onto coker l if for each b ∈ B, the projection π : U → U/ im lb is still
surjective when restricted to Vb . In this case, V + (im lb )⊥ spans U for all b ∈ B and
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
7

V ′ → B is a vector bundle of rank dim V ′ = dim V + ind l. In particular, V ′ − V

represents the virtual index bundle ind l.

Assume for the moment that the image of f |SV ′ is disjoint from the sphere S(V ⊥ ) ⊂
SU . Composing with the above deformation retraction, we obtain a map ρV f |SV ′ :
SV ′ → SV which factors through the Thom space.

Definition 2.9. The map φf = ρV f |SV ′ : T V ′ → SV is called the (Bauer-Furuta)

finite dimensional approximation of f .

Notice that this definition of finite dimensional approximation depends on the choice
of subspace V such that f |SV ′ is valued in SU \ S(V ⊥ ) and the choice of decom-
position f = l + c. We will show that such subspaces exist and that the stable
homotopy class of φf is independent of V, l and c.

For any finite dimensional subspace W ⊃ V , write W as an orthogonal sum W =

U ⊕V with U the orthogonal complement of V inside W . Assuming that V surjects
onto coker l, let W ′ = l−1 (W ) with W ′ = Ue ⊕V ′ where U
e is the fibrewise orthogonal
′ ′ e → U is an isomorphism of vector
complement of V in W . Notice that l|Ue : U
bundles, hence U is trivial and T (U ⊕ V ) = SU ∧ T V ′ .
e e ′

Definition 2.10. A finite dimensional subrepresentation V ⊂ U is admissible (with

respect to f ) if it satisfies the following three conditions:

(1) V surjects onto coker l.

(2) For any finite dimensional subspace W ⊃ V , the image of f |SW ′ : SW ′ →
SU is disjoint from the unit sphere S(W ⊥ ) in W ⊥ . Consequently the de-
formation retract ρW : SU \ S(W ⊥ ) → SW defines a map
ρW f |SW ′ : T W ′ → SW .
(3) The maps ρW f |SW ′ and id∧ρV f |SV ′ are homotopic under the identifications
T W ′ = SU ∧ T V ′ and SW = SU ∧ SV .
ρW f |S
W′
TW′ SW
(2.5) = =
id∧ρV f |S
V′
SU ∧ T V ′ SU ∧ SV

Proposition 2.11 ([9] Lemma 2.3). For any bounded Fredholm map f = l + c :
H ′ → H, there exists an admissible subrepresentation V ⊂ U.

Proof sketch. To construct one such V , let D ⊂ U be the closed unit disk in U. By
the boundedness condition, f −1 (D) is contained in a closed disk bundle DR ′
⊂ H′
′
of radius R. Consequently, if |h | > R, then |f (h)| > 1. Set C to be the closure of
′
c(DR ), which is compact. Let 0 < ε ≤ 14 and choose a finite covering of C by balls
of radius ε with centers vi for i = 1, ..., N . By [2, Proposition A5] there is a finite
8 JOSHUA TOMLIN

dimensional subspace V0 ⊂ U with (im lb )⊥ ⊂ V0 for all b ∈ B. Let V be a finite

dimensional G-invariant subspace containing both V0 and span{v1 , ..., vn }, which
can be obtained using isotypical decomposition.

By construction V satisfies (1). Further, V has the property that for any subspace
′
W ⊃ V and h ∈ DR ,
|(1 − pW )c(h)| < ε.
′
Property (2) follows from this bound and the fact that |f (h)| = 1 implies h ∈ DR .
′ ′
Let S be the bounding sphere bundle of DR . Property (3) follows by defining a
′
homotopy ht : DR ∩ W ′ → SU \ S(W ⊥ ) between f |SW ′ and id ∧ ρV f |SV ′ on the
′
restricted domain DR ∩ W ′ . This homotopy is constructed so that the image of
ht |S ′ does not intersect W ⊥ for any t. Thus ht |S ′ is valued in SU \ (D ∩ W ⊥ ), which
is a contractible subset of SU \ S(W ⊥ ). Hence ht extends over the complementary
′
disk SW ′ \ (DR ∩ W ′ ) and composing with ρW gives a homotopy between ρW f |SW ′
and id ∧ ρV f |SV ′ on SW ′ . □

Definition 2.12 ([9] Theorem 2.6). Let f = l + c : H ′ → H be an equivariant,

bounded Fredholm map and fix an equivariant trivialisation H ∼
= U ×B. The Bauer-
Furuta class of f is the stable homotopy class
0
[φf ] ∈ πG,U (B; ind l)
where φf = ρV f |SV ′ : T V ′ → SV for any choice of admissible subrepresentation
V ⊂ U. This cohomotopy class is independent of V and the presentation f = l + c.

Proof. Fix an admissible subrepresentation V ⊂ U and recall that ind l = V ′ − V ,

hence the Thom spectrum T (ind l) is given by T (ind l) = Σ−V T V ′ . It follows that
0
πG,U (B; ind l) = Hom(T (ind l), S0 )
= Colim[SU ∧ T V ′ , SU ∧ SV ].
U ⊂V ⊥

Here U ⊂ U is orthogonal to V and the connecting morphisms are given by smash

products with the identity. For any other admissible subrepresentation W , there
is an admissible subrepresentation containing both V and W . Hence property
(3) implies that the Bauer-Furuta classes corresponding to V and W are stably
0
homotopic, therefore [φf ] ∈ πG,U (B; ind l) is well defined.

To see that [φf ] does not depend on the choice of decomposition f = l + c, let
f = li + ci be two Fredholm decompositions for i = 0, 1. Let Ft = lt + ct for
lt = (1 − t)l0 + tl1 and ct = (1 − t)c0 + tc1 , noting that Ft = f for all t. The maps lt
are linear Fredholm and the maps ct are compact. Now F is a Fredholm map over
B × [0, 1] which is certainly bounded. Applying finite dimensional approximation
to F gives a homotopy between finite dimensional approximations of f using the
two different presentations f = l0 + c0 and f = l1 + c1 . □

2.2. Schwarz approximation. In [29], Schwarz details an alternative approach

to finite dimensional approximation. Let D′ ⊂ H ′ be a closed disk bundle with
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
9

boundary sphere bundle S ′ . Fix a trivialisation H ∼

= U × B and let Cl (D′ , H)
′
denote the set of continuous maps f : D → U such that c = f − l|D′ is compact
and f |S ′ is non-vanishing.

Definition 2.13. Two Fredholm maps f0 , f1 : D′ → H are compactly homotopic

(relative to l) if there is a homotopy ft = l + ct with ct compact and (ft )|S ′ non-
vanishing for all t ∈ [0, 1]. More generally, we say that two bounded Fredholm maps
f0 , f1 : H ′ → H are compactly homotopic if there exists a disk D′ ⊂ H ′ containing
f0−1 (0) ∪ f1−1 (0) on which the restrictions f0 |D′ and f1 |D′ are compactly homotopic.

Give Cl (D′ , H) the uniform convergence topology so that π0 (Cl (D′ , H)) is the set
of compact homotopy classes relative to l. The homotopy class of a Fredholm map
f : H ′ → H is dull since it is classified by ind l [13], but restricting to homotopies
through Cl (D′ , H) uncovers more interesting behaviour.

Let f = l + c ∈ Cl (D′ , H). Suppose for now that c(D′ ) is contained in a finite
dimensional subrepresentation V ⊂ U. Without loss of generality, we can assume
that (im lb )⊥ ⊂ V for all b ∈ B. Let V ′ = l−1 (V ), which is a vector bundle of rank
dim V ′ = dim V + ind l. Denote the restriction f |D′ ∩V ′ by
ψf,V = f |D′ ∩V ′ : (D′ ∩ V ′ , S ′ ∩ V ′ ) → (V, V \ {0})
Let W ⊃ V be a finite dimensional subrepresentation containing V with W = U ⊕V
orthogonally. Let W ′ = l−1 (W ) so that W ′ = Ue ⊕V orthogonally with l| e : U
U
e →U
′ ′ ′ ′
an isomorphism. For any map g : (D ∩ V , S ∩ V ) → (V, V \ {0}), define a
suspension map

ΣU g : (D′ ∩ W ′ , S ′ ∩ W ′ ) → (W, W \ {0})

e

ΣU g(u + v) = l(u) + g(v).

e

Note that for w = u + v, if ΣU g(w) = 0 then g(v) = 0 and u = 0, which implies

e

/ S ′ ∩ W ′ . Let [(A, B); (C, D)] denote the set of homotopy classes of maps
that w ∈
from (A, B) to (C, D) where the homotopies are through maps of pairs. Then ΣU
e

descends to a map of homotopy classes

(2.6) ΣU : [(D′ ∩ V ′ , S ′ ∩ V ′ ); (V, V \ {0})] → [(D′ ∩ W ′ , S ′ ∩ W ′ ); (W, W \ {0})].

e

Define
(2.7) Πl (D′ , H) = Colim[(D′ ∩ V ′ , S ′ ∩ V ′ ); (V, V \ {0})]
V ⊂U

where the colimit is taken over the maps given by (2.6). Any map g : (D′ ∩
V ′ , S ′ ∩ V ′ ) → (V, V \ {0}) defines a class [g] ∈ Πl (D′ , H) by suspension. The map
ψf,V depends on the choice of subrepresentation V , however the class of [ψf,V ] ∈
Πl (D′ , H) does not.

Lemma 2.14. For f = l + c ∈ Cl (D′ , H), suppose that V and W are finite dimen-
sional subrepresentations which both contain c(D′ ) and surject onto coker l. Then
[ψf,V ] and [ψf,W ] are equal classes of Πl (D′ , H).
10 JOSHUA TOMLIN

Proof. Assume without loss of generality that V ⊂ W . As before write W = U ⊕ V

and W ′ = Ue ⊕ V ′ orthogonally with l| e : U
U
e → U an isomorphism. For any element
′ ′
u + v ∈ W with u ∈ U and v ∈ V , we have
e
f |S ′ ∩W ′ (u + v) = l(u) + l(v) + c(u + v).
Define a homotopy
Ft (u + v) = l(u) + l(v) + (1 − t)c(v) + tc(v + u).
This is a homotopy from F0 = ΣU ψf,V to F1 = ψf,W . Additionally, Ft is non-zero
e

on S ′ ∩ W ′ for all t ∈ [0, 1]. To see this, recall that c(D′ ) ⊂ V , hence Ft (u + v) = 0
implies that l(u) = 0. It follows that u = 0 and |v| = 1. But f |S ′ ∩W ′ (v) =
f |S ′ ∩V ′ (v), which does not vanish. Thus the classes [ψf,V ] and [ψf,W ] are equal. □

Lemma 2.15. Suppose ft = l + ct : [0, 1] → Cl (D′ , H) is a compact homotopy

with c0 (D′ ) ∪ c1 (D′ ) ⊂ V for some finite dimensional subrepresentation V ⊂ U that
surjects onto coker l. Then ψf0 ,V and ψf1 ,V are homotopic as maps of pairs.

Proof. This follows immediately from Definition 2.13 since ft |S ′ is non-vanishing,

hence the restriction
(ft )|D′ ∩V ′ : (D′ ∩ V ′ , S ′ ∩ V ′ ) → (V, V \ {0})
is a map of pairs for all t with f0 = ψf0 ,V to f1 = ψf1 ,V . □

Not all elements f = l + c ∈ Cl (D′ , H) are nice enough to have c(D′ ) contained
in a finite dimensional subrepresentation, however it is true that every compact
homotopy class has such a representative.

Lemma 2.16. For any f ∈ Cl (D′ , H), there exists δ > 0 such that |f (h)| > δ for
all h ∈ S ′ .

Proof. Fix b ∈ B and suppose that there is a sequence hn ∈ Sb′ with |f (hn )| → 0.
By the weak compactness of Sb′ , after passing to a subsequence it can be assumed
that hn → h weakly for some h ∈ Hb′ . By the compactness of c, after passing to
a further subsequence it can be assumed that c(hn ) → a strongly for some a ∈ U.
Now l(hn ) = f (hn ) − c(hn ) → −a strongly. Since lb is Fredholm, its image is
closed and a = l(v) for some v ∈ (ker lb )⊥ . Write hn = xn + yn for xn ∈ ker lb
and yn ∈ (ker lb )⊥ . Now l(hn ) = l(yn ) → −l(v). Since lb is an isomorphism from
(ker lb )⊥ onto its image, it follows that yn → −v. Further xn = hn − yn → h + v
weakly, but ker lb is finite dimensional so xn → h + v strongly as well. Thus hn → h
strongly and h ∈ Sb′ since Sb′ is closed. However f (hn ) → 0 implies that f (h) = 0,
contradicting the assumption that f |Sb′ ̸= 0. Since B is compact, such a delta can
be chosen simultaneously over all fibres. □

Remark 2.17: In fact, suppose that f : H ′ → H is a bounded Fredholm map

with f −1 (0) ∩ S ′ = ∅. The above argument can be extended to show that there
is a δ > 0 with |f (h)| > δ for every h ∈ H ′ − D′ . First choose a closed disk E ′
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
11

/ E ′ , which we can assume contains D′ . Now the

such that |f (h)| ≥ 1 for all h ∈
argument in the lemma easily extends to the closed, bounded set E ′ − D′ .

Corollary 2.18. Every element f = l + c0 ∈ Cl (D′ , H) is compactly homotopic

to a map g = l + c1 ∈ Cl (D′ , H) with c1 (D′ ) contained in a finite dimensional
subrepresentation.

Proof. From Lemma 2.16, choose δ > 0 such that |f (h)| > δ for all h ∈ S ′ . Let
ε = 2δ . Since D′ is bounded and c is compact, the closure of c(D′ ) can be covered by
finitely many balls of radius ε with centers v1 , ..., vn . Let V be a finite dimensional
subrepresentation that contains span{vi } and surjects onto coker l. Set V ′ = l−1 (V )
and let g = l + pV c. By construction, |(1 − pV )c(h)| < ε for all h ∈ D′ . Define a
homotopy for t ∈ [0, 1] by
Ft = l + (1 − t)c + tpV c.

Notice that for h ∈ S ′ ,

|Ft (h)| = |l(h) + c(h) − t(1 − pV )c(h)|
≥ |f (h)| − t|(1 − pV )c(h)|
δ
> .
2
Thus Ft is a compact homotopy from F0 = f to F1 = g. □

For any f ∈ Cl (D′ , H), define ψf = ψg,V for some choice of g compactly homotopic
to f with V a finite dimensional subrepresentation that contains c(D′ ) and surjects
onto coker l. The map f 7→ [ψf ] identifies π0 (Cl (D′ , H)) with a subset of Πl (D′ , H),
which is a result originally due to Schwarz [29].

Theorem 2.19 ([10] Theorem 5.3.20). Let l : H ′ → H be a linear Fredholm op-

erator and fix a closed disk bundle D′ ⊂ H ′ with bounding sphere bundle S ′ . The
map
ΨD′ : π0 (Cl (D′ , H)) → Πl (D′ , H)
(2.8) [f ] 7→ [ψf ]

is well-defined and injective.

Proof. Lemma 2.14 and 2.15 show that ΨD′ : π0 (Cl (D′ , H)) → Πl (D′ , H) is well
defined. To prove injectivity, suppose f = l + c0 and g = l + c1 are elements of
Cl (D′ , H) with [ψf ] = [ψg ]. After applying Σ if necessary, we can assume that there
is a compact homotopy F : (D′ ∩V ′ )×[0, 1] → V with F0 = f and F1 = g for V ⊂ U
a finite dimensional subrepresentation that contains c0 (D′ ) ∪ c1 (D′ ) and surjects
onto coker l. To show that f and g are compactly homotopy, we must extend F to
D′ × [0, 1].
12 JOSHUA TOMLIN

Pn
Let v1 , ..., vn be an orthonormal basis for V and write Ft (v) = l(v) + i=1 cit (v)vi
with cit (v) = ⟨ct (v), vi ⟩. Since (D′ ∩ V ′ ) × [0, 1] is a closed subset of D′ × [0, 1],
the Tietze extension theorem guarantees the existence of a continuous extension
cit : D′ × [0, 1] → R for all t ∈ [0, 1]. Define
Ht : D′ × [0, 1] → U
n
X
Ht (v) = l(v) + cit (v)vi
i=1
It remains to show that Ht is non-vanishing on S ′ for all t. If Ht (h) = 0 for
h ∈ S ′ , then l(h) ∈ V . Thus h ∈ l−1 (V ) = V ′ and h ∈ S ′ ∩ V ′ . Therefore
Ht (h) = Ft (h) ̸= 0. Thus Ht is a compact homotopy from f to g. □

2.3. Equivalence. Let f = l+c : H ′ → H be a bounded Fredholm map and V ⊂ U

an admissible subrepresentation with V ′ = l−1 (V ). Recall that the Bauer-Furuta
finite dimensional approximation φf is given by
φf = ρV f |SV ′ : (SV ′ , B∞ ) → (SV , ∞)
This maps factors through the Thom space T V ′ . Let Pl (H ′ , H) denote the set of
equivariant bounded Fredholm maps f : H ′ → H relative to l. Equip Pl (H ′ , H)
with the topology induced by the uniform metric on SH . Bauer-Furuta approxi-
mation defines a map
Φ : Pl (H ′ , H) → πG,U
0
(B; ind l)
f 7→ [φf ].
Alternatively, let D ⊂ H be a closed disk bundle with bounding sphere bundle S ′
′ ′

such that f −1 (0) ⊂ D′ and f −1 (0) ∩ S ′ = ∅, which is guaranteed to exist since f

is bounded. Recall that pV : H → V is the orthogonal projection and assume for
now that pV f does not vanish on S ′ ∩ V ′ . Then the Schwarz approximation of f is
given by
ψf = pV f |D′ ∩V ′ : (D′ ∩ V ′ , S ′ ∩ V ′ ) → (SV , SV \ {0}).
Schwarz approximation defines another map
ΨD′ : π0 (Cl (D′ , H)) → Πl (D′ , H)
(2.9) [f ] 7→ [ψf ]
We will leverage the properties of Schwarz approximation to prove that Φ descends
to a well defined map from π0 (Pl (H ′ , H)) to πG,U
0
(B; ind l) and that this map is a
bijection. At a surface level, it looks as if Schwarz approximation depends on the
appropriately chosen disk bundle D′ ⊂ H ′ . However, enlarging D′ does not change
the Schwarz approximation of f by the following lemma.

Lemma 2.20. Two elements f0 , f1 ∈ Pl (H ′ , H) are homotopic through bounded

Fredholm maps if and only if they are compactly homotopic on some disk bundle
D′ ⊂ H ′ that contains f0−1 (0) ∪ f1−1 (0).

Proof. Suppose ft : [0, 1] → Pl (H ′ , H) is a homotopy so that ft is a bounded

Fredholm map for each t ∈ [0, 1]. Compactness of the unit interval and continuity
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
13

of the homotopy guarantees the existence of a disk D′ ⊂ H ′ such that ft−1 (0) ⊂ D′
and ft−1 (0)∩S ′ = ∅ for all t ∈ [0, 1]. Thus f0 |D′ and f1 |D′ are compactly homotopic.

Suppose instead that there is a disk bundle D′ ⊂ H ′ of radius R′ on which f0 |D′

and f1 |D′ are compactly homotopic. Let Ft : D′ → H be such a homotopy with
R′
Ft−1 (0) ∩ S ′ = ∅ for all t ∈ [0, 1]. For any x ∈ H ′ − D′ , let s = |x| x ∈ S ′ and extend
Ft on H ′ − D′ by
|x|
Ft (x) = Ft (s).
R′
Now for each t ∈ [0, 1], Ft : H ′ → H is Fredholm and since 0 ∈
/ Ft (S ′ ), Lemma 2.16
guarantees that Ft is bounded. Hence [F0 ] = [F1 ] as elements of π0 (Pl (H ′ , H)).
We claim that f0 is homotopic to F0 through bounded Fredholm maps. Such a
homotopy ht : H ′ → H is given by ht |D′ = f0 |D′ and, for x ∈ H ′ − D′ ,
 t  −t !
|x| |x|
ht (x) = f0 x .
R′ R′
Similarly, [F1 ] = [f1 ] in π0 (Pl (H ′ , H)) and the result follows. □

To simplify notation, set

′
D− = D′ ∩ V ′
′ ′
D+ = SV ′ − D−
S0′ = S ′ ∩ V ′ .
′
That is, D± are the two hemispheres of SV ′ with S0′ the equator. Define an inter-
mediary map
′
ϕf = ρV f |SV ′ : (SV ′ , D+ ) → (SV , SV \ {0}).
′
This definition of ϕf assumes that ρV f does not vanish on D+ . The following
lemma shows that the finite dimensional subrepresentation V can be chosen to
simultaneously make φf , ψf and ϕf maps of pairs.

Lemma 2.21. Let f = l + c : H ′ → H be a bounded Fredholm map and fix a disk

bundle D′ ⊂ H ′ such that f −1 (0) ⊂ D′ and f −1 (0) ∩ S ′ = ∅. There exists a finite
dimensional subrepresentation V ⊂ U such that:

(1) V is an admissible subrepresentation as in Definition 2.10,

(2) pV f is non-vanishing on S0′ ,
′
(3) ρV f |SV ′ is non-vanishing on D+ .

These properties translate to any finite dimensional subrepresentation W ⊃ V .

Proof. Since f is bounded, we can assume that f −1 (D) ⊂ D′ where D ⊂ U is the

closed unit disk. As explained in Remark 2.17, choose a δ > 0 such that |f (h)| > δ
for h ∈ H ′ − D′ . Let ε = min{ 14 , 2δ }. Cover the closure of c(D′ ) by finitely many
14 JOSHUA TOMLIN

ε-balls with centres v1 , ..., vn and set V = span{vi }. As seen before, we can enlarge
V to be a subrepresentation that surjects onto coker l. Now V has the property
that |(1 − pV )f (h)| < ε for all h ∈ D′ ∩ V ′ and is an admissible subrepresentation
by Proposition 2.11. Since |f (h)| > δ for h ∈ S ′ , it follows that |pV f (h)| > 2δ for
h ∈ S′ ∩ V ′.

Suppose that ρV f (h) = 0 for some h ∈ SV ′ . Notice from the definition of ρV in

(2.4) that this implies that f (h) is finite with pV f (h) = 0 and |(1 − pV )f (h)| < 1.
This means that |f (h)| < 1 and h ∈ D′ ∩ V ′ . Therefore |(1 − pV )f (h)| < ε < δ and
′ ′
h∈/ D+ since |f (h)| < δ. That is, ρV f (h) is non-vanishing on D+ . For any finite
dimensional W ⊃ V , it is still the case that |(1 − pW )f (h)| < ε for h ∈ D′ ∩ W ′
and the argument can be repeated. □

Consider the following diagram where a, b and c are the obvious inclusions:
a ′ b ′
(SV ′ , B∞ ) (SV ′ , D+ ) (D− , S0′ )
ϕf
φf ψf
c
(SV , ∞) (SV , SV \ {0})

The dashed arrow ψf does not make the diagram commute, but we will show
that it does commute up to homotopy. These inclusions induce functions between
homotopy classes of maps of pairs:
c∗
[(SV ′ , B∞ ) ; (SV , ∞)] [(SV ′ , B∞ ) ; (SV , SV \ {0})]
a∗
∗
 ′ b
(D− , S0′ ) ; (SV , SV \ {0}) ′
  
(SV ′ , D+ ) ; (SV , SV \ {0})

Proposition 2.22. The maps a∗ , b∗ and c∗ induced by inclusions are bijections.

The composition ξ = b∗ (a∗ )−1 c∗ defines a bijection
 ′
, S0′ ) ; (SV , SV \ {0})

ξ : [(SV ′ , B∞ ) ; (SV , ∞)] → (D−
which identifies [φf ] with [ψf ].

′
Proof. Contracting D+ radially to ∞ fibrewise defines a homotopy Ft : SV ′ → SV ′
′ ′ ′
with F0 = id and F1 (D+ ) = B∞ . The compositions aF1 : (SV ′ , D+ ) → (SV ′ , D+ )
and F1 a : (SV ′ , B∞ ) → (SV ′ , B∞ ) are both homotopy equivalent to the identity
through maps of pairs, hence a is a homotopy equivalence of pairs and a∗ is bijection.

Since SV \ {0} is contractible, any f : SV ′ → SV with f (B∞ ) ⊂ SV \ {0} can be

composed with a homotopy that contracts f (B∞ ) to ∞. Hence c∗ is surjective. For
injectivity let g0 , g1 : SV ′ → SV be maps with gi (B∞ ) = ∞ and suppose that there
is a homotopy gt from g0 to g1 with gt (B∞ ) ⊂ SV \ {0}. Since SV ′ × I is compact,
there is an open neighbourhood U ⊂ SV of 0 such that gt (B∞ ) ⊂ SV \ U for all t.
Thus [g0 ] = [g1 ] as elements of [(SV ′ , B∞ ); (SV , SV \ U )]. By the same reasoning as
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
15

above, the inclusion (SV , ∞) → (SV , SV \ U ) is a homotopy equivalence of pairs.

Hence [g0 ] = [g1 ] as elements of [(SV ′ , B∞ ), (SV , ∞)].

To see that b∗ is surjective, suppose f : D− ′

→ SV is a map with f |S0′ valued in
′ ′
SV \{0}. Locally, SV ′ is obtained from D− by attaching D+ over S0′ . Since SV \{0}
′
is contractible, f |S0′ can be extended to D+ by a null homotopy while remaining
valued in SV \ {0}. This construction can be globalised using a partition of unity,
′
thus f extends to SV ′ with f (D+ ) ⊂ SV \ {0}.

For injectivity, let b′ : D−

′
→ SV ′ and b′′ : S0′ → D+
′
be inclusions with mapping
′
cones Cb and Cb . Recall that the cofibersequence (D−
′ ′′ , S0′ ) → (SV ′ , D+
′
) →
(Cb , Cb ) induces an exact sequence [1, III Prop 3.9]
′ ′′

′ b∗ ′
[(Cb′ , Cb′′ ); (SV , SV \ {0})] → [(SV ′ , D+ ); (SV , SV \ {0})] → [(D− , S0′ ); (SV , SV \ {0})].
The cone Cb′ deformation retracts onto Cb′′ , hence
[(Cb′ , Cb′′ ); (SV , SV \ {0})] ∼
= [(Cb′′ , Cb′′ ); (SV , SV \ {0})]
= [Cb′′ , SV \ {0}].

However [Cb′′ , SV \ {0}] is trivial since SV \ {0} is contractible. Thus b∗ is injective

by the exactness of the cofibersequence.

It remains to show that [ψf ] = [b∗ (a∗ )−1 c∗ (φf )]. We have that c∗ φf = a∗ ϕf , thus it
is enough to show that [ψf ] = [b∗ ϕf ]. Note that both ψf |S0′ and b∗ ϕf |S0′ are valued
in V \ {0} ⊂ H \ V ⊥ . Recall that ρV f |S0′ = λpV f |S0′ for some positive continuous
function λ : H \ V ⊥ → R, hence the straight line homotopy from b∗ ϕf to ψf never
vanishes. □

Corollary 2.23 ([7] Theorem 2.1). Given a choice of trivialisation H ∼

= U × B,
the map Φ descends to a bijection
(2.10) Φ : π0 (Pl (H ′ , H)) → πG,U
0
(B; ind l).

Proof. First suppose that f0 , f1 ∈ Pl (H ′ , H) are homotopic through bounded Fred-

holm maps. Then by Lemma 2.20, f0 and f1 are compactly homotopic on an
appropriately chosen D′ and [f0 ] = [f1 ] in π0 (Cl (D′ , H)). Thus Ψ|D′ f0 = ΨD′ f1
and applying ξ −1 gives Φf0 = Φf1 , hence Φ is well defined. This also proves injec-
tivity since if Φf0 = Φf1 , then Ψ|D′ f0 = ΨD′ f1 by applying ξ. Hence f0 and f1
are compactly homotopic on D′ by Theorem 2.19, and [f0 ] = [f1 ] in Pl (H ′ , H) by
Lemma 2.20.

0
For surjectivity, a class [f ] ∈ πG,U (B; ind l) is represented by a pointed map f :
T V → SV for V ⊂ U an admissible subspace with V ′ = l−1 (V ). After possible
′

suspension we can assume that f −1 (∞) = [B∞ ] ∈ T V ′ . For π : SV ′ → T V ′ the

projection, this means that (f ◦ π)−1 (∞) = B∞ and the restriction f ◦ π : V ′ → V
is proper. Since V is admissible, l defines an isomorphism from (V ′ )⊥ to V ⊥ .
16 JOSHUA TOMLIN

Hence f ◦ π can be extended to a bounded Fredholm map f˜ : H ′ → H such that

Φf˜ = [f ]. □

3. The families Bauer-Furuta invariant

Let B be a compact, connected smooth manifold. A 4-manifold family is a smooth,

locally trivial, oriented fibre bundle π : E → B with each fibre diffeomorphic to
a closed, oriented 4-manifold X. In particular, E → B has transition functions
valued in Diff+ (X). For b ∈ B, denote the fibres of E as Xb = π −1 (b).

Let T (E/B) → E be the vertical tangent bundle T (E/B) = ker π∗ , which is a

4-dimensional real vector bundle over E. Let g be a metric on T (E/B) with ∇ the
associated Levi-Civita connection. One can think of g and ∇ as smoothly varying
families of metrics {gb }b∈B and connections {∇b }b∈B on the fibres Xb . Let sE
be a spinc structure on T (E/B) with associated spinor bundles W ± → E. This
induces a smoothly varying family of spinc structures {sb }b∈B on the fibres of E.
Let L = det(W + ) be the determinant line bundle of W + , which is a family of U (1)-
bundles over B. A U (1)-connection 2A on L defines a family of spinc connections
∇A on W + .

Let Λi T ∗ (E/B) → E denote the i-th exterior power of T ∗ (E/B). A section of

Λi T ∗ (E/B) is a family of i-forms on the fibres Xb . Write ΩiB (E) = C ∞ (E, Λi T ∗ (E/B))
to denote the set of families of smooth i-forms, which has the structure of a vector
bundle ΩiB (E) → B. Similarly, C ∞ (E, W + ) → B denotes the bundle of families
of smooth spinors over B. We write Λ2+ T ∗ (E/B) to denote the bundle of self-dual
2-forms determined by the Hodge star.

3.1. Families with separating necks. Let V0 → B be a rank 4 oriented Rie-

mannian vector bundle equipped with a spinc structure sV0 . Denote by S(V0 ) ⊂ V0
the unit sphere sub-bundle of V0 . When performing a families connected sum, S(V0 )
will be obtained as the normal bundle of a section of the vertical tangent bundle of
one of the summands. For any L > 0, Let NB (L) denote the family of cylinders
NB (L) = S(V0 ) × [−L, L].
We write Nb (L) to denote the fibre of NB (L) → B over b ∈ B. Denote the families
of positive and negative fiberwise boundary components by
∂NB (L)+ = S(V0 ) × {L}
∂NB (L)− = S(V0 ) × {−L}.
Since the transition maps of V0 are valued in SO(4), the vertical tangent bundle
T (S(V0 )/B) can be equipped with a metric gS(V0 ) that restricts to the standard
round metric on each fibre. Equip the vertical tangent bundle of NB (L) with the
metric gNB (L) = gS(V0 ) +dt2 which on each fibre is the product of the standard round
metric on S 3 and the standard interval metric on [−L, L]. The spinc structure sV0
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
17

determines a 3-dimensional spinc structure on the vertical tangent space of S(V0 ).

Pulling this back to NB (L) defines a spinc structure sNB (L) on T (NB (L)/B).

Definition 3.1. Let E → B be a family of 4-manifolds with connected fibre X

and fix L > 1. A separating neck of length 2L on a E → B is an embedding
ι : NB (L) → E covering the identity. It is required that the neck complement
M = E − ι(NB (L − 1)) has fibres Mb which decompose as
a
Mb = Mb− Mb+
where ∂Mb− = ι(∂Nb (L − 1)− ) and ∂Mb+ = ι(∂Nb (L − 1)+ ), both with reversed
orientation. It is assumed that E is given a metric and spinc structure that extends
gNB (L) and sNB (L) .

Given a 4-manifold family E → B with a separating neck of length 2L, we identify

NB (L) with its image ι(NB (L)). If X has n connected components, then a sepa-
rating neck on E is just a separating neck on each component. In this case, the
neck is a disjoint union
an
NB (L) = NB (L)i .
i=1
Assume for convenience that L > 2. For each 1 ≤ i ≤ n, define collar subbundles
Ci± ⊂ NB (L)i by
Ci− = S(V0 ) × [−L, −L + 1]
Ci+ = S 3 (V0 ) × [L − 1, L].
Let C = i (Ci− ∪ Ci+ ). Each fibre Cb is a collar neighbourhood of the boundary
`
of Nb (L). Removing NB (L − 1) from E gives a family of manifolds Mb with fibres
Xb − Nb (L − 1) and a natural inclusion ι : C → M . For any other neck length
L′ > 2, there is a natural isometric inclusion C → N (L′ ) identifying C has a collar
neighbourhood of ∂N (L′ ). Let E(L′ ) = M ∪C NB (L′ ). That is, E(L′ ) is defined
by the following pushout
C NB (L′ )
ι

M E(L′ ).
Let τ ∈ Sn be an even permutation on n objects. Define a permuted inclusion map
ιτ : C → M such that ιτ |C − = ι|C − and ιτ |C + = ι|C + . That is, Ci− is mapped
i i i τ (i)

to ι(Ci− ) but Ci+ is mapped to ι(Cτ+(i) ). Define the permuted family E τ by the
following pushout
C NB (L)
ιτ

M Eτ .
Fiberwise, each boundary component of the form ι(Ci− )b ⊂ Mb has been connected
by a cylinder S 3 × [−L, L] to ι(Cτ+(i) )b . We write X τ to denote the standard fibre
of E τ .
18 JOSHUA TOMLIN

3.2. The families Seiberg-Witten monopole map. Fix a reference spinc connection
A0 on E. Any other connection A can be written as A = A0 + ia for some family
of one-forms a ∈ C ∞ (E, T ∗ (E/B)). Let n be the number of connected components
of X and fix an integer k ≥ 4. The metric and orientation of E determines an
L2 -inner product of spinors and forms through integration. We write L2k (E, −) to
denote the L2k -Sobolev space of k-times weakly differentiable sections, with weak
derivatives in L2 .

To define the families monopole map, we follow the construction in [6, Example 2.1
and 2.4]. For now assume that b1 (X) = 0. Define Hilbert space bundles A and C
over B by
A = L2k (E, W + ⊕ T ∗ (E/B)) ⊕ Rn
(3.1) C = L2k−1 (E, W − ⊕ Λ2+ T ∗ (E/B) ⊕ R).

The Rn term in A is identified with the space of locally constant functions H 0 (X; R)
on X. Denote by Tn = (S 1 )×n the group of locally constant gauge transformations.
Let Tn act on A and C in the usual manner, on spinors by multiplication and on
forms trivially. This action is fibre-preserving and orthogonal. The monopole map
µ : A → C is the Tn -equivariant bundle map given by the formula
µ(ψ, a, f ) = (DA0 +ia ψ, −iFA+0 +ia + iσ(ψ), d∗ a + f ).

The map σ is defined by the equation σ(ψ) = ψ ⊗ ψ ∗ − 12 Id where the traceless,

Hermitian endomorphism σ(ψ) is identified as an imaginary valued self-dual 2-form.
A solution (ψ, a, f ) ∈ µ−1 (0) must have f = 0, hence we will suppress the third
component. This solution corresponds to the Seiberg-Witten monopole (ψ, A0 + a).
The gauge fixing condition d∗ a = 0 determines the with gauge class of (ψ, A0 + a)
up to a harmonic gauge transformation. Since b1 (X) = 0, the only harmonic gauge
transformations are the locally constant ones.

There is a decomposition µ = l + c with

l(ψ, a, f ) = (DA0 ψ, d+ a, d∗ a + f )
(3.2) c(ψ, a, f ) = (ia · ψ, −iFA+0 + iσ(ψ), 0).
The map l is linear Fredholm and c is compact, hence µ is a Fredholm map. There
is a somewhat standard argument (e.g [9, Proposition 3.1]) in ordinary Seiberg-
Witten theory that shows that µ is a bounded Fredholm map when B is a point.
Assuming that B is compact means that this argument can be extended fibrewise.

In the case that b1 (X) > 0, it will be necessary to assume that a smooth section
x : B → E exists. In general the families Bauer-Furuta invariant will depend on
the homotopy class of x. Let H1 (R) ⊂ C ∞ (E, T ∗ (E/B)) denote the subbundle of
real harmonic forms. That is, H1 (R) → B is a vector bundle with fibre H1 (Xb ; R)
over b ∈ B. Now pull back the bundles defined in (3.1) to bundles over H1 (R):
à = L2k (E, W + ⊕ T ∗ (E/B)) ⊕ Rn → H1 (R)
C˜ = L2k−1 (E, W − ⊕ Λ2+ T ∗ (E/B) ⊕ R) ⊕ H1 (R) → H1 (R).
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
19

The tilde notation is used because we are yet to quotient out by harmonic gauge
transformations. Let Aθ = A0 + iθ denote the connection associated to θ ∈ H(R).
Note that since θ is harmonic, FAθ = FA0 . Define µ̃ : Ã → C˜ by
(3.3) µ̃θ (ψ, a, f ) = (DAθ +ia ψ, −iFA0 +ia + iσ(ψ), d∗ a + f, pr(a)).
This is the monopole map with gauge fixing, before dividing out by the harmonic
gauge transformations. The bundle map pr : L2k (E, T ∗ (E/B)) → H1 (R) is defined
as follows. Let {Uβ } ⊂ B be a trivialising open cover of B with E|Uβ ∼ = Uβ × X.
1 b1 (X)
Choose cycles α , ..., α that restrict to a homology basis on each fibre of E|Uβ .
Define a map prβ : Ω1B (E)|Uβ → H1 (R)|Uβ on each fibre above b ∈ Uβ by
Z
(3.4) (pr(a)b )(αbi ) = ab .
αib

Extend pr(a)b linearly so that pr(a)b ∈ Hom(H1 (Xb ), R) = H 1 (Xb ; R). Now let
{ρβ } be aPpartition of unity subordinate to {Uβ } and define pr : ΩB (E) → H1 (R)
by pr = β ρβ prβ . This map has the property that if a ∈ Ω1B (E) is a family of
closed one forms, then pr(a) ∈ H1 (R) is the cohomology class of a in each fibre.
This extends continuously to a map pr : L2k (E, T ∗ (E/B)) → H1 (R).

To account for the harmonic gauge transformations, let H(2πZ) → B be the bundle
of groups over B with fibre H 1 (Xb ; 2πZ). For each ω ∈ H(2πZ) and b ∈ B, define
a map gω,b : Xb → S 1 by
Z y !
gω,b (y) = exp i ω .
x(b)

This map is well defined since the periods of ω are multiples of 2π. Further, gω,b
−1
is the unique harmonic gauge transformation with the property that gω,b dgω,b =
iω and gω,b (x(b)) = 1. The gauge transformation gω acts on a connection A by
gω · A = A + iω.

Let the bundle of groups H(2πZ) act on H(R) fiberwise by ω·θ = ω+θ. The quotient
bundle J = H(R)/H(2πZ) is the b1 (X)-dimensional Jacobian torus bundle over B.
That is, each fibre Jb is the Jacobian torus J (Xb ) = H(Xb ; R)/H(Xb ; 2πZ) of Xb .
Define an action of H(2πZ) on elements (ψ, a, f ) ∈ Ãθ and (ϕ, η, g, α) ∈ C˜θ by
ω · (θ, (ψ, a, f )) = (θ + ω, (gω−1 ψ, a, f ))
ω · (θ, (ϕ, η, g, α)) = (θ + ω, (gω−1 ϕ, η, g, α)).
This is the free action of the based harmonic gauge transformations gω . Under
this action, µ̃ is equivariant. The fiberwise quotients A = Ã/H(2πZ) and C =
˜
C/H(2πZ) are Hilbert bundles over J with a residual Tn -action of the constant
gauge transformations. The map µ̃ descends to a Tn -equivariant Fredholm map
µ : A → C over J . This is the families monopole map in the setting b1 (X) > 0. In
a similar fashion to (3.2), µ = l + c is a bounded Fredholm map with
lθ (ψ, a, f ) = (DAθ ψ, d+ a, d∗ a + f, pr(a))
(3.5) cθ (ψ, a, f ) = (ia · ψ, −iFA+0 + iσ(ψ), 0, 0).
20 JOSHUA TOMLIN

Define a Tn universe U by
(3.6) U = L2k−1 (X, W |− 2 ∗ 1
X ⊕ Λ+ (T X) ⊕ R) ⊕ H (X; R).

This universe can be identified with each fibre of C. The map l defines a family of
linear Fredholm maps over J , so let indJ l denote the corresponding virtual index
bundle. Let H + → J denote the rank b+ (X) trivial bundle with fibre H+ 2
(X; R)
+
so that the relation indJ l = indJ D − H holds.

Definition 3.2. The families Bauer-Furuta invariant of a 4-manifold family E →

B is the cohomotopy class
[µ] ∈ πT0 n ,U (J , indJ l)
+
(3.7) = πTb n ,U (J , indJ D).

Now suppose that E is a family of 4-manifolds X(L) with necks of length 2L.
To construct an appropriate reference connection, let {ρβ } be a partition of unity
subordinate to a trivialising open cover {Uβ } of B. Let Aβ0 be a flat connection on
NUβ (L) that is identical on each neck component NUβ (L)i = Uβ × (S 3 × [−L, L]).
Such a connection exists since H 2 (S 3 × [−L, L]; R) = 0. Extend Aβ0 to E|Uβ and
set A0 = β ρβ Aβ0 . Then A0 defines a connection on both X and X τ which is flat
P
on the neck.

Moreover, let GN (1) → B be the bundle of Gauge groups with fibre maps (GN (1) )b ⊂
C ∞ (Xb , S 1 ) that fix the short neck N (1)b . Let ker dN (1) ⊂ Ω1B (E) be the sub-
set of families of forms a ∈ ker d that vanish on N (1). The inclusion (A0 +
i ker dN (1) )/GN (1) → JE is a smooth bundle map over B that restricts to a dif-
feomorphism on each fibre, hence we can identify (A0 + i ker dN (1) )/GN (1) = JE .
Now for any even permutation τ , JE = JE τ which means that µE and µE τ can be
treated as bundle maps over the same space J = JE = JE τ .

Denote by W c + → S(V0 ) × [−L, L] the restriction of W + → NB (L) to one of the

connected components of NB (L). Define F = ⊕ni=1 W c + to be the direct sum of
+
c over S(V0 ) × [−L, L]. Since NB (L) has n connected components, a
n-copies of W
section ψ : NB (L) → W + can be identified with a vector of sections

(3.8) ⃗ : S(V0 ) × [−L, L] → F.

ψ
That is, the restriction ψi to the ith component of NB (L) is identified with the ith
component of ψ.⃗ Let T : S(V0 )×[−L, L] → SO(n) denote a matrix valued function.
For a section ψ : NB (L) → W along NB (L), define an action by T ·ψ = T ψ ⃗ where T
acts pointwise on ψ⃗ and T ψ⃗ is identified with a section of W → NB (L). The same
+

process defines an action on forms along the neck a : NB (L) → Λi (T ∗ (NB (L)/B)).

Let γ : [0, 1] → SO(n) be a smooth path from the identity to τ , which exists under
the assumption that τ is even. Let φ : [−L, L] → [0, 1] be a smooth map that
vanishes on [−L, 1] and is identically equal to 1 on [1, L]. Define a matrix valued
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
21

function V : S(V0 ) × [−L, L] → SO(n) by

(3.9) V (x, t) = γ(φ(t)).

Note that V is constant along the S(V0 ) factor. Let (ψ, a) : NB (L) → W + ⊕
T ∗ (NB (L)/B) be a spinor-form pair along NB (L) and define (ψ, a)τ = (V · ψ, V · a)
by the action described above. The pair (ψ, a)τ has the property that (ψ, a)τi =
(ψ, a)i on C − and (ψ, a)τi = (ψ, a)τ (i) on C + . Now given a section (ψ, a) : E →
W + ⊕ T ∗ (E/B) defined on all of E, this permutation process defines a section
(ψ, a)τ on E τ with the property that (ψ, a) and (ψ, a)τ agree outside of NB (1).

This construction defines an isomorphism VA : AE → AE τ of Hilbert bundles over

J . Similarly for C, the action of V defines a map VC : CE → CE τ that on the
n
H1 (X; R) factor is just the identity. Thus VA and VC identify π0 (Pl (A, C)T ) and
n
π0 (Pl (Aτ , C τ )T ) by the map [f ] 7→ [VC f VA−1 ]. Moving forward we will suppress
the subscripts. Since all the permutation occurs in NB (1), there is a constant CV
independent of L such that

(3.10) ∥V (ψ, a)∥L2k ≤ CV ∥(ψ, a)∥L2k .

Theorem 3.3 (Families Permutation Theorem). Let E → B be a family of closed

4-manifolds that admits an n-component separating neck. Let τ ∈ Sn be an even
permutation with E τ the corresponding permuted family. Then

(3.11) [µE ] = [µE τ ]

+
as elements of πTb n ,U (J , ind D).

Remark 3.4: In the construction of E τ it is assumed that τ is an even permutation,

however Remark 6.7 explains how this assumption is unnecessary.

In [8], Bauer gave a proof of Theorem 3.3 in the unparameterised case where B
is a single point. While the ideas used in the proof of his formula were sound, we
were not able to reproduce some of his arguments and have deemed the proof to be
incomplete. Instead, we revisit his ideas to formulate a new proof that extends to
the families setting.

4. Monopoles on the neck

To prove the permutation theorem it is enough to show that µE is homotopic

to V −1 µE τ V through compact perturbations of l (see Corollary 2.23). Such a
homotopy is constructed in three stages, and at each stage it is important to check
that the boundedness conditions outlined in Definition 2.13 are satisfied. This is
accomplished using techniques from the theory of Sobolev spaces, elliptic operators
and monopoles on a cylinder with a varying neck length.
22 JOSHUA TOMLIN

4.1. Sobolev estimates. Two fundamental theorems in the theory of Sobolev

spaces are the Sobolev embedding theorem [11] and the Sobolev multiplication
theorem [30]. These theorems give estimates that relate different Sobolev norms on
a spinor-form pair (ψ, a) on X. For two neck lengths L1 and L2 , we require estimates
that apply to spinor-form pairs on both X(L1 ) and X(L2 ). The following results
achieve this goal in the situations necessary for Theorem 3.3.

Lemma 4.1 ([8] Proposition 3.1). Let k and p be non-negative integers such that
k − p4 > 0. There is a constant CS such that, for any L ≥ 2,
|(ψ, a)|C 0 ≤ CS ∥(ψ, a)∥Lpk
for any Lpk -pair (ψ, a) on X(L).

Proof. Fix a neck length L ≥ 2. For each x ∈ X let δx : X → [0, 1] be a smooth

bump function in a small neighbourhood of x. Let X0 = X(2) and use the Sobolev
embedding Lpk (X0 , W + ⊕T ∗ X0 ) ⊂ C 0 (X0 , W + ⊕T ∗ X0 ) [30, Theorem B.2] to choose
a constant C1 with
|(ψ ′ , a′ )|C 0 (X0 ) ≤ C1 ∥(ψ ′ , a′ )∥Lpk (X0 )

for any Lpk -pair (ψ ′ , a′ ) on X0 . Note that such a constant exists since k − p4 > 0.
For any spinor ψ on X, δx ψ can be identified as a spinor on X0 . The same is true
for δx a for a one-form a on X. Now for each x ∈ X,
|(δx ψ, δx a)|C 0 (X) ≤ C1 ∥(δx ψ, δx a)∥Lpk (X) .
Since δx is smooth and defined locally, it has bounded Ck norm which is independent
of L. Thus there exists a constant C2 such that for all x ∈ X,
∥(δx ψ, δx a)∥Lpk (X) ≤ C2 ∥(ψ, a)∥Lpk (X) .
It follows that
|(ψ, a)|C 0 (X) = sup |(δx ψ, δx a)|C 0 (X)
x∈X
≤ C1 sup ∥(δx ψ, δx a)∥Lpk (X)
x∈X
≤ C1 C2 ∥(ψ, a)∥Lpk (X) .
Setting CS = C1 C2 gives the result. □

The next lemma demonstrates that Sobolev multiplication bounds only depend
linearly on the length of the neck.

Lemma 4.2. Let k ≥ 0 and p ≥ 1 be integers. There is a constant CSM such that,
for any neck length L ≥ 2,
∥a · ψ∥Lpk ≤ CSM L∥a∥L2p ∥ψ∥L2p
k k

for any L2p

k -pair (ψ, a) on X(L).
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
23

Proof. For notational simplicity, assume that X is connected. Recall that M ±

denotes the two halves of M = X − N (L − 1) with tubular ends of the form
N (L)− ∩ M = S 3 × [−L, −L + 1]
N (L)+ ∩ M = S 3 × [L − 1, L].
We will cut N (L) into pieces that can be identified on X0 = X(2), then use Sobolev
multiplication on X0 . Let ϕ : X → [0, 1] be a smooth function such that ϕ ≡ 1 on
X − N (L − 54 ) and ϕ ≡ 0 on N (L − 2). Define a function χ : R → [0, 1] such that
χ ≡ 1 on [0, 1] and χ ≡ 0 outside [− 41 , 54 ]. Let χi be χ shifted by i so that χi ≡ 1
on [i, i + 1] and χi ≡ 0 outside [i − 14 , i + 54 ]. Let m = ⌊L − 54 ⌋. For i an integer
with −(m + 1) ≤ i ≤ m, extend χi to N (L) = S 3 × [−L, L] by projection onto the
interval factor. Let
v
u X m
χ2i .
u
φ = t ϕ2 +
i=−(m+1)

χi ϕ
Notice that φ is positive on X. Let φi = φ for −(m + 1) ≤ i ≤ m with φm+1 = φ.
By construction,
m+1
X
φ2i = 1.
i=−(m+1)

For each −(m + 1) ≤ i ≤ m + 1, set ψi = φi ψ and ai = φi a. Both ψi and ai can

be identified as sections on X0 . For −(m + 1) ≤ i ≤ m, this is accomplished by
shifting the interval [i − 41 , i + 54 ] to [− 14 , 54 ]. We can assume that the C k norm of φi
is bounded, which implies that there exists a constant C1 , independent of L, such
that
∥ψi ∥L2p (X0 ) ≤ C1 ∥ψ∥L2p (X)
k k

(4.1) ∥ai ∥L2p (X0 ) ≤ C1 ∥a∥L2p (X) .

k k

For the purposes of elliptic bootstrapping, the Lpk -Sobolev norm on X0 is defined
as
k
X
∥(ψi , ai )∥Lpk (X0 ) = ∥(Dj ψi , (d∗ + d+ )j ai )∥Lp (X0 ) .
j=0

Equivalently, the Lpk -norm

on X0 can instead be defined by differentiating spinors
with the spinc connection ∇A0 and forms with the Levi-Civita connection ∇. Thus
there are constants 0 < c ≤ C such that
k
X
c∥(ψi , ai )∥Lpk (X0 ) ≤ ∥(∇jA0 ψi , ∇j ai )∥Lp (X0 ) ≤ C∥(ψi , ai )∥Lpk (X0 ) .
j=0

Calculating with repeated applications of the Leibniz rule gives

k
1X j
∥ai · ψi ∥Lpk (X0 ) ≤ ∥∇A (ai · ψi )∥Lp (X0 )
c j=0
k j
1 XX
≤ Kj,l ∥Γ(∇l ai ) · (∇j−l
A ψi )∥Lp (X0 )
c j=0
l=0
24 JOSHUA TOMLIN

for some non-negative constants Kj,l . Here Γ(∇l ai ) ∈ End(W ) is the matrix corre-
sponding to spinor multiplication by the (l + 1)-form ∇l ai . The operator norm of
Γ(∇l ai ) is equal to |∇l ai |, hence applying Sobolev multiplication [30, Lemma B.3]
it follows that
(4.2) ∥ai · ψi ∥Lpk (X0 ) ≤ C2 ∥ai ∥L2p (X0 ) ∥ψi ∥L2p (X0 )
k k

for some constant C2 . This constant depends on c, Kj,l and Sobolev multiplication
on X0 , hence is independent of L. Combining (4.1) and (4.2) produces the result.
m+1
X
∥a · ψ∥Lpk (X) ≤ ∥ai · ψi ∥Lpk (X0 )
i=−m−1
m+1
X
≤ C2 ∥ai ∥L2p (X0 ) ∥ψi ∥L2p (X0 )
k k
i=−m−1
m+1
X
≤ C2 C12 ∥a∥L2p (X) ∥ψ∥L2p (X)
k k
i=−m−1
≤ CSM L∥a∥L2p (X) ∥ψ∥L2p (X) .
k k

The same argument applied to σ(ψ) instead gives the following result.

Lemma 4.3. Let k ≥ 0 and p ≥ 1 be integers. There is a constant Cσ such that,

for any neck length L ≥ 2,
∥σ(ψ)∥Lpk ≤ Cσ L∥ψ∥2L2p
k

for and ψ ∈ L2p +

k (X(L), W ).

4.2. Elliptic inequality. To analyse the properties of monopoles on a neck of

varying length, it is useful to apply Yang Mills theory on cylinders as in Chapter 2
of [15]. Fix a neck length L with X = X(L). For notational simplicity, assume that
X only has one connected component. Recall that M + and M − are the two halves
of M = X − N (L − 1). Attach infinite tubes to M + and M − to get manifolds with
tubular ends Y ± of the form
Y − = M − ∪ S 3 × [−L + 1, ∞)
Y + = S 3 × (−∞, L − 1] ∪ M + .

One-forms on the tubular component of Y ± can be analysed by studying forms on

the product S 3 × R. Let π : S 3 × R → S 3 be projection onto the S 3 factor. All
elements of Ω1 (S 3 × R) are of the form ωt + f dt for ωt ∈ Ω1 (S 3 ) a smooth family of
one-forms on S 3 and f : S 3 × R → R a smooth function. That is, we can identify
Ω1 (S 3 × R) = C ∞ (S 3 × R, R ⊕ π ∗ T ∗ S 3 ).
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
25

Similarly, self-dual 2-forms Ω2+ (S 3 × R) can be identified with time-dependent 1-

forms ξ ∈ C ∞ (S 3 × R, π ∗ T ∗ S 3 ) by the isomorphism
ξ 7→ ξ ∧ dt + ∗3 ξ.
Here ∗3 is the hodge star operator on S 3 . Thus we can interpret the elliptic operator
d∗ + d+ : Ω1 (S 3 × R) → Ω0 (S 3 × R) ⊕ Ω2+ (S 3 × R) as
(4.3) d∗ + d+ : C ∞ (S 3 × R, R ⊕ π ∗ T ∗ S 3 ) → C ∞ (S 3 × R, R ⊕ π ∗ T ∗ S 3 ).
Consider the operator L : Ω0 (S 3 ) ⊕ Ω1 (S 3 ) → Ω0 (S 3 ) ⊕ Ω1 (S 3 ) defined by
0 d∗
 
(4.4) L= .
d ∗d
This is a self-adjoint elliptic operator that squares to the Laplacian L2 = dd∗ +
d∗ d on Ω0 (S 3 ) ⊕ Ω1 (S 3 ). It can be shown by direct calculation that under the
identification (4.3),
∂
d∗ + d+ = +L
∂t
∂
where ∂t is the derivative in the R direction.

∂
Since the tubular ends of Y are not compact, solutions to the operator ∂t +L will be
studied in weighted Sobolev spaces. Weighted Sobolev spaces consist of functions
that have a controlled exponential increase towards the tubular ends. To define
them, fix a parameter α < 0 and let fα− be a smooth function on S 3 × [−L, ∞)
that is zero on S 3 × [−L, −L + 2] and decreases with slope α on S 3 × [−L + 3, ∞).
Similarly, define fα+ on S 3 × (−∞, L] to be zero on S 3 × [L − 2, L] and decrease
with slope α on S 3 × (−∞, L − 3]. Since α < 0, both functions fα± are non-positive.
Define the weighted Sobolev space Lp,α ± p ±
k (Y ) to be the completion of L (Y ) with
respect to the norm
∥g∥Lp,α
k
= ∥ exp(fα± )g∥Lpk .

Note that exp(fα± ) is decreasing exponentially towards the infinite end of Y ± . More-
over, the spaces Lp,α ±
k (Y ) are independent of the original neck length L.

It is shown in [15] that d∗ + d+ = ∂t ∂

+ L is a linear Fredholm operator on Lp,α 1 -
forms if α is not in the spectrum of L. Since L is self-adjoint and elliptic it has
discrete spectrum away from infinity, so choose α < 0 to be greater than the
maximal negative eigenvalue of L. As in (3.4), define a harmonic projection map
pr± : Ω1 (Y ± ) → Ω1 (Y ± ) by integrating a homology basis of curves away from the
neck. The image of pr± is H 1 (X; R), identified as the space of harmonic forms
H1 (Y ± ) ⊂ Ω1 (Y ± ). Fix p > 4 so that Lp1 (Y ± , T ∗ Y ) ⊂ C 0 (Y ± , T ∗ Y ) by Sobolev
embedding and extend pr± continuously to a map on Lp,α 1 forms. The operator
d∗ + d+ : Lp,α ± ∗ ±
1 (Y , T Y ) → L
p,α
(Y ± , R ⊕ Λ2+ T ∗ Y ± )
is Fredholm with kernel H1 (Y ± ) and cokernel H 0 (Y ± ; R) ⊕ H+
2
(Y ± ; R). Let H ± =
ker pr , which is a complement of ker(d + d ). Thus the restriction of d∗ + d+
± ∗ +

to H ± is a linear bijection onto the closed Lp,α -image of d∗ + d+ . The bounded

26 JOSHUA TOMLIN

inverse theorem guarantees that there are constants C ± > 0 such that for b ∈
Lp,α ± ∗ ±
1 (Y , T Y ),

≤ C ± ∥(d∗ + d+ )b∥Lp,α + ∥ pr± (b)∥ .

(4.5) ∥b∥Lp,α
1

Importantly, the constants C ± are independent of the neck length L. That is, for
another choice of neck length L′ and manifolds with tubular ends (Y ′ )± , there is
an isometry from Lp,α ± ∗ ± p,α ′ ± ∗ ′ ±
k (Y , T Y ) to Lk ((Y ) , T (Y ) ) defined by shifting the
′
interval component by L − L.

To analyse the behaviour of forms away from the middle of the neck, define smooth
cut-off functions β ± : X → [0, 1] which vanish on X ∓ ∪ N (2) and are equal to 1 on
M ± . To ensure such β exist, we will assume that L ≥ 3.

Lemma 4.4 ([8] Proposition 3.1). Let β ± be cutaway functions as described above
and fix p > 4. There exists a constant C such that, for any neck-length L > 3,
|a|C 0 (M ) ≤ C ∥(d∗ + d+ )β + a∥Lp (X) + ∥(d∗ + d+ )β − a∥Lp (X) + ∥ pr(a)∥

for any Lp1 -form a on X(L).

Proof. The Sobolev embedding Lp1 (X, T ∗ X) ⊂ C 0 (X, T ∗ X) guarantees the exis-
tence of a constant CS such that
(4.6) |a|C 0 (X) ≤ CS ∥a∥Lp1 (X) .
for all a ∈ Lp1 (X, T ∗ X).
Lemma 4.1 ensures that CS can be chosen independently
±
of L. To apply the elliptic bound, let b± = β ± a and notice that efα b± = a on M ± .
±
|a|C 0 (M ± ) = |efα b± |C 0 (M ± )
±
≤ |efα b± |C 0 (Y ± )
±
≤ CS ∥efα b± ∥Lp1 (Y ± )
(4.7) = CS ∥b± ∥Lp,α
1 (Y ± )

Note that b± is compactly supported in M ± ∪ N (L − 1) ⊂ Y ± , so the Sobolev

±
bound (4.6) applies to |efα b± |C 0 (Y ± ) . Now (4.5) gives
|a|C 0 (M ± ) ≤ CS C ± ∥(d∗ + d+ )b± ∥Lp,α (Y ± ) + ∥ pr(b± )∥

≤ CS C ± ∥(d∗ + d+ )b± ∥Lp (X) + ∥ pr(b± )∥ .

This inequality follows since fα± ≤ 0 and b± is compactly supported on M ± ∪N (L−

1) ⊂ Y ± . Putting this together with C = max{C s C + , C s C − } yields
|a|C 0 (M ) ≤ |a|C 0 (M + ) + |a|C 0 (M − )
≤ C ∥(d∗ + d+ )b+ ∥Lp (X) + ∥(d∗ + d+ )b− ∥Lp (X) + (∥ pr(b+ )∥ + ∥ pr(b− )∥) .

Recall that pr(b± ) is defined by integration over an orthonormal basis of curves

contained in M . Since b± vanishes on M ∓ we have
∥ pr(b+ )∥ + ∥ pr(b− )∥ = ∥ pr(b+ + b− )∥
= ∥ pr(a)∥.
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
27

It follows that
|a|C 0 (M ) ≤ C ∥(d∗ + d+ )b+ ∥Lp (X) + ∥(d∗ + d+ )b− ∥Lp (X) + ∥ pr(a)∥ .

Proposition 4.5 (Adapted from [8] Lemma 3.3). Fix p > 4. There exists a neck
length L0 and a constant CE such that the following holds: For any L ≥ L0 , let
a ∈ Lp1 (X, T ∗ X) be an Lp1 -form on X(L) such that pr(a) = 0. If (d∗ +d+ )a vanishes
on N (L − 1), then
|a|C 0 (M ) ≤ CE |(d∗ + d+ )a|C 0 (M ) .

Proof. Let β ± be cut-off functions as described in Lemma 4.4. Assume without

loss of generality that |dβ ± |C 0 (X) < L2 , which is possible when L > 6. Lemma 4.4
gives a constant C1 , independent of L, such that
(4.8) |a|C 0 (M ) ≤ C1 (∥(d∗ + d+ )β + a∥Lp (X) + ∥(d∗ + d+ )β − a∥Lp (X) ).
Calculating with the Leibniz rule yields
∥(d∗ + d+ )β ± a∥Lp (X) ≤ ∥β ± (d∗ + d+ )a∥Lp (X) + ∥dβ ± ∧ a∥Lp (X) .
The product β ± (d∗ + d+ )a is supported inside M ± , thus
∥β ± (d∗ + d+ )a∥Lp (X) = ∥(d∗ + d+ )a∥Lp (M ± ) .
Since N (L − 1) has non-negative Ricci curvature, the Weitzenböck formula [26, Ex
2.31] implies that |a| is a harmonic function when restricted to N (L − 1). Thus
the maximum principle holds and supN (L−1) |a| = sup∂N (L−1) |a|. Let N (2, L − 1)
denote N (L − 1) − N (2). Then dβ ± is supported inside X ± ∩ N (2, L − 1) and
∥dβ + ∧ a∥Lp (X) + ∥dβ − ∧ a∥Lp (X) ≤ ∥dβ + + dβ − ∥Lp (N (L−1)) sup |a|
N (2,L−1)
1 1
(4.9) ≤ 4L p −1 vol(S 3 ) p sup |a|.
∂N (L−1)

Combining this with (4.8) gives

1 1
|a|C 0 (M ) ≤ C1 ∥(d∗ + d+ )a∥Lp (M ) + 4C1 L p −1 vol(S 3 ) p |a|C 0 (∂N (L))
1 1 1
≤ C1 vol(M ) p |(d∗ + d+ )a|C 0 (M ) + 4C1 L p −1 vol(S 3 ) p |a|C 0 (M ) .
1 1
Set C2 = C1 vol(M ) p and C3 = 4C1 vol(S 3 ) p to obtain
1
|a|C 0 (M ) (1 − C3 L p −1 ) ≤ C2 |(d∗ + d+ )a|C 0 (M ) .
1
1 −1 p
− 1−p
Since p > 4, p1 − 1 < 0 and L ≥ L0 implies L p −1 ≤ L0p . Set L0 = (2C3 ) ,
which we can assume is larger than 6, so that L ≥ L0 implies
1
1−p
1
(1 − C3 L p −1 ) ≥ (1 − C3 L0 p ) = .
2
When L ≥ L0 it follows that
1
|a|C 0 (M ) ≤ C2 |(d∗ + d+ )a|C 0 (M ) (1 − C3 L p −1 )−1
≤ 2C2 |(d∗ + d+ )a|C 0 (M ) .
Let CE = 2C2 , which is independent of L. □
28 JOSHUA TOMLIN

Remark 4.6: Suppose instead that (d∗ + d+ )a only vanishes on N (2, L − 1). Then
the maximum of (d∗ + d+ )a could be obtained on ∂N (2) instead of ∂N (L − 1). To
overcome this, assume that there is a constant C, independent of L, such that
sup |a| ≤ C sup |a|.
N (2,L−1) ∂N (L−1)

Since (d∗ +d+ )a = 0 on N (2, L−1) and β ± is supported in X ± −N (2), the product
β ± (d∗ + d+ )a is supported in M ± . We can still execute the above argument with
(4.9) becoming
∥dβ + ∧ a∥Lp (X) + ∥dβ − ∧ a∥Lp (X) ≤ ∥dβ + + dβ − ∥Lp (N (L−1)) sup |a|
N (2,L−1)
1 1
≤ 4CL p −1 vol(S 3 ) p sup |a|.
∂N (L−1)
1
Setting C3 = 4C1 Cvol(S 3 ) p , there still exists constants CE and L0 such that, when
L ≥ L0 ,
|a|C 0 (X) ≤ CE |(d∗ + d+ )a|C 0 (M ) .

4.3. Elliptic bootstrapping. For a fixed connection A ∈ J (X) on X(L), an

elliptic bootstrapping argument can be used to produce a polynomial L2k -bound on
a monopole (ψ, a) of the form
∥(ψ, a)∥L2k ≤ CB (1 + |(ψ, a)|)dC 0 .
The constant CB depends on the curvature of A and the length of the neck L. To
cooperate with neck stretching, we show that CB only increases polynomially in L.

Lemma 4.7. Let A ∈ JX be a connection on X(L) and fix an integer k ≥ 2. There

are positive constants CB and d such that, for any L ≥ 2, if (ψ, a) is an L2k -pair
with
DA ψ = −ia · ψ
(4.10) d+ a = iFA+ − iσ(ψ),
then
∥(ψ, a)∥L2k ≤ CB Ld (1 + |(ψ, a)|C 0 )d .

Proof. Use the first order differential operators DA and d+ to define the Lpk -norm
so that
∥(ψ, a)∥pLp − ∥(ψ, a)∥pLp = ∥(DA ψ, d+ a)∥pLp .
i i−1

k+1
For any 0 ≤ i ≤ k and 2 ≤ p ≤ 2 , (4.10) ensures that
∥(DA ψ, d +
a)∥pLp ≤ ∥a · ψ∥pLp + (∥σ(ψ)∥Lpi−1 + ∥FA+ ∥Lpi−1 )p
i−1 i−1

By Lemma 4.2 and 4.3, there are constants CSM and Cσ independent of L such
that
∥(ψ, a)∥Lpi ≤ CSM L∥a∥L2p ∥ψ∥L2p + Cσ L∥ψ∥2L2p + ∥FA+ ∥Lpi−1 + ∥(ψ, a)∥Lp .
i−1 i−1 i−1
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
29

Since A is flat on the neck, ∥FA+ ∥Lpi−1 is a constant independent of L. Thus there
is a constant C1 such that
∥(ψ, a)∥Lpi ≤ C1 L(∥(ψ, a)∥2L2p + ∥(ψ, a)∥Lp )
i−1

for all 0 ≤ i ≤ k and 2 ≤ p ≤ 2k+1 . Starting with i = k and p = 2, inductively

applying this inequality gives a bound
∥(ψ, a)∥L2k ≤ Ld1 f (∥(ψ, a)∥L2 , ..., ∥(ψ, a)∥L2k+1 )
for some natural number d1 and polynomial f , both independent of L. Letting d2
be the degree of f , there is a constant C2 such that
|f (x1 , ..., xk )| ≤ C2 (1 + |x1 | + ... + |xk |)d2 .
Since vol(X(L)) increases linearly with L, there is a bound
1
∥(ψ, a)∥Lp ≤ vol(X(L)) p |(ψ, a)|C 0
≤ C3 L|(ψ, a)|C 0 .
Here C3 is a constant independent of L and p. Letting d = d1 + d2 , it follows that
∥(ψ, a)∥L2k ≤ C2 Ld1 (1 + ∥(ψ, a)∥L2 + ... + ∥(ψ, a)∥L2k+1 )d2
≤ C2 Ld1 Ld2 (1 + C3 |(ψ, a)|C 0 + ... + C3 |(ψ, a)|C 0 )d2
≤ CB Ld (1 + |(ψ, a)|C 0 )d
for some constant CB independent of L. □

Remark 4.8: Assume that there is a smooth function ρ : X → R and constant C

such that the pair (ψ, a) instead satisfies
DA ψ = −iρa · ψ
∥d+ a∥Lpi ≤ C(∥σ(ψ)∥Lpi + ∥FA+ ∥Lpi )
∥ρa∥Lpi ≤ C∥a∥Lpi

for all 0 ≤ i ≤ k, 2 ≤ p ≤ 2k+1 . The same argument can be repeated, the only
difference being that the constant C1 now depends on C. Thus there still exists
positive constants CB and d such that
∥(ψ, a)∥L2k ≤ CB Ld (1 + |(ψ, a)|C 0 )d .
These constants depend on C, but are independent of L so long as C is.

4.4. Exponential decay. Since X(L) is compact, there are Lp -bounds on spinors
and one-forms of the form
(4.11) ∥(ψ, a)∥Lp ≤ Cp |(ψ, a)|C 0
1
with Cp = vol(X(L)) p . This constant Cp grows linearly with the length of the
neck. However, we will demonstrate that monopoles decay exponentially towards
the middle of the neck, which will counteract this and other polynomial growth.
The following work is adapted from Chapter 3 of [15].
30 JOSHUA TOMLIN

Let E → S 3 be a vector bundle over S 3 , equipped with a metric gE and compatible

connection ∇E . For notational simplicity, we will assume that N (L) = S 3 × [−L, L]
has one connected component. Let π : N (L) → S 3 be projection onto the S 3
component. Fix k > 2 and let A : C ∞ (S 3 , E) → C ∞ (S 3 , E) be a first order, self-
adjoint, elliptic pseudo-differential operator on E. By spectral theory of elliptic
operators, there is an orthonormal basis of eigenvectors {ϕn }∞ 2 3
n=−N ⊂ L (S , E) for
A with discrete real eigenvalues {λn }. Label the eigenvalues so that the non-zero
eigenvalues have a positive index and the zero eigenvalues (of multiplicity N + 1)
have a non-positive index. Thus there is a δ > 0 such that |λn | > δ for all n ≥ 1.
Also ensure that the labeling is chosen so that |λn | ≥ |λm | when n ≥ m.

Let f0 ∈ C ∞ (S 3 , E) be a smooth section with eigen decomposition f0 = n f0n ϕn

P
2
convergentPin L for f0n ∈ R. Then Af0 is also smooth and its eigen decomposition
n
is Af0 = n λn f0 ϕn since A isP self-adjoint. A smooth section f of π ∗ E → N (L)
also has a decomposition ft = n f n (t)ϕn for some functions f n : [−L, L] → R.
The smoothness of ft implies the smoothness of the component functions f n by the
Leibniz integral rule.

Define a pseudo-differential operator by

D : C ∞ (N (L), π ∗ E) → C ∞ (N (L), π ∗ E)
∂
(4.12) D= + A.
∂t
Assume that D is elliptic and extend D to an operator on L2 sections. Recall that
C + = S 3 × [L − 1, L] and C − = S 3 × [−L, −L + 1] denote collar neighbourhoods of
the boundary of N (L).

Proposition 4.9 (Adapted from [15] Lemma 3.2). Fix constants r ≥ 1 and L ≥ 2r.
Suppose f ∈ L2 (N (L), π ∗ E) such that ft is orthogonal to ker A for all t ∈ [−L, L].
If Df = 0 then
Z  −2δ(L−2r)  Z Z 
2 e 2 2
(4.13) |f | ≤ |f | + |f |
N (2r) 1 − e−2δ C− C+

and
(4.14) sup |f | ≤ Cδ e−δ(L−2r) sup |f |.
N (r) N (L)

where δ and Cδ are positive constants independent of L and r.

Proof. Note that since D is assumedPto be elliptic, Df = 0 implies that f is smooth

by elliptic regularity. Write Aft = n λn f n (t)ϕn so that
X
∂t f + λn f n ϕn = 0.
n
2
Taking the L −inner product with ϕn yields
∂t f n (t) + λn f n (t) = 0.
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
31

Since ft is orthogonal to ker A it can be assumed that n ≥ 1 and λn ̸= 0 so that

f n (t) = e−λn t f n (0).

Notice that if λn > 0 then f n decays exponentially as t increases and if λn < 0

then f n decays exponentially as t decreases. To capture this behaviour, split f n =
n n
f− + f+ defined by
( (
n 0 if λn > 0 n f n if λn > 0
f− = n
f+ =
f if λn < 0. 0 if λn < 0
P∞ n
Also let f± = n=1 f± ϕn so that f = f− + f+ . Each half of |f |2 is integrated
separately.
Z Z 2r
|f+ |2 = ∥f+ (t)∥2L2 dt
N (2r) −2r
Z 2r ∞
X
= e−2λn t |f+
n
(0)|2 dt
−2r n=1
∞
X sinh(4rλn ) n
(4.15) = |f+ (0)|2
n=1
λ n

Here the monotone convergence theorem has been used to swap the sum and the
integral. Integrating instead over the band C − gives
∞  2λn L
− e2λn (L−1)
Z 
X e
(4.16) |f+ |2 = n
|f+ (0)|2
C −
n=1
2λ n

Choose a δ > 0 such that |λn | > δ for n ≥ 1. When λn > 0, notice that
sinh(4rλn ) e4rλn
≤
λn 2λn
e2λn L e−2λn (L−2r)
 
= (1 − e−2λn )
2λn 1 − e−2λn
 −2λn (L−2r)   2λn L
− e2λn (L−1)

e e
=
1 − e−2λn 2λn
 −2δ(L−2r)   2λn L
− e2λn (L−1)

e e
(4.17) ≤
1 − e−2δ 2λn
The last line follows since λn > δ > 0 and L − 2r ≥ 0. Combining (4.15), (4.16)
and (4.17) gives
∞  −2δ(L−2r)   2λn L
− e2λn (L−1)
Z 
X e e
|f+ |2 ≤ −2δ
n
|f+ (0)|2
N (2r) n=1
1 − e 2λ n
 −2δ(L−2r)  Z
e
= |f+ |2
1 − e−2δ C −
 −2δ(L−2r)  Z
e
≤ |f |2 .
1 − e−2δ C−
32 JOSHUA TOMLIN

Similarly when λn < 0,

∞  −2λn (L−1)
− e−2λn L
Z 
2
X e n 2
|f− | = |f− |
C+ n=1
2λ n
∞  2|λn |L
− e2|λn |(L−1)

X e n 2
= |f− | .
n=1
2|λn |

Now (4.17) can be applied to get

∞
sinh(4r|λn |) n
Z X
|f− |2 = |f− (0)|2
N (2r) n=1
|λ n |
 −2δ(L−2r)  Z
e
≤ |f− |2
1 − e−2δ C +
 −2δ(L−2r)  Z
e
≤ |f |2 .
1 − e−2δ C+

It follows that
e−2δ(L−2r)
Z   Z Z 
(4.18) |f |2 ≤ |f |2 + |f |2 .
N (2r) 1 − e−2δ C− C+

This proves the first inequality (4.13).

The supremum and essential supremum of |f | agree because f is continuous. Since

PN
the sequence ( i=1 f n (t)ϕn )∞ 2
N =1 converges to ft in L as N → ∞, there is a
subsequence that converges to ft pointwise almost everywhere. Let (x0 , t0 ) ∈ S 3 ×
[−r, r] be any point such that
∞
X
ft0 (x0 ) = e−λn t0 f n (0)ϕn (x0 ).
n=1

Since t0 ∈ [−r, r] it follows that

∞
X
|ft0 (x0 )| ≤ er|λn | |f n (0)||ϕn (x0 )|.
n=1

The Sobolev embedding L22 (S 3 , E) → C 0 (S 3 , E) gives a constant CS such that

|ϕn |C 0 ≤ CS ∥ϕn ∥L22 for all n. Further the second order elliptic operator A2 :
L22 (S 3 , E) → C 0 (S 3 , E) provides an elliptic inequality

∥ϕn ∥L22 ≤ CE (∥A2 ϕn ∥L2 + ∥ϕn ∥L2 )

= CE (λ2n + 1)∥ϕn ∥L2 .

Note that CS and CE are independent of L. Since ∥ϕn ∥L2 = 1, we have |ϕn |C0 ≤
CE CS (λ2n + 1) and
∞
X
|ft0 (x0 )| ≤ CE CS (λ2n + 1)er|λn | |f n (0)|.
n=1
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
33

Lemma 4.12 provides a bound

∞
!2 ∞
X
r|λn |
X sinh(4r|λn |) n
(λ2n + 1)e n
|f (0)| ≤ C′ |f (0)|2
n=1 n=1
|λ n |

for some constant C ′ which depends only on {λn }. Combining this with (4.15)
produces
∞
X sinh(4r|λn |) n
|ft0 (x0 )|2 ≤ C |f (0)|2
n=1
|λ n |
Z
=C |f |2
S 3 ×[−2r,2r]

where C = C ′ CS2 CE
2
. Applying (4.18) and taking the essential supremum over N (r)
yields
 −2δ(L−2r)  Z Z 
e
sup |f |2 ≤ C |f |2
+ |f |2
N (r) 1 − e−4δ C− C+
3
 
2Cvol(S ) −2δ(L−2r)
≤ e sup |f |2 .
1 − e−4δ N (L)
q
3)
Let Cδ = 2Cvol(S
1−e−4δ
so that

sup |f | ≤ Cδ e−δ(L−2r) sup |f |.

N (r) N (L)

Corollary 4.10. Suppose that a ∈ L2 (N (L − 1), T ∗ N (L − 1)) is a 1-form such that

(d∗ + d+ )a = 0. Then for any r ≥ 1 and L ≥ 2r + 1,
(4.19) sup |a ∧ dt| ≤ Cδ e−δ(L−2r) sup |a|
N (r) N (L−1)

for some positive constants δ and Cδ independent of L and r.

Proof. It is shown in (4.3) that d∗ +d+ can be identified as an operator on C ∞ (N (L−

1), R⊕π ∗ T ∗ S 3 ) and that d∗ +d+ = ∂t
∂
+L. Here L is a self-adjoint, elliptic operator
on Ω (S ) ⊕ Ω (S ) with L = dd + d∗ d. Note that d∗ + d+ is also self-adjoint
0 3 1 3 2 ∗

and elliptic. Since b1 (S 3 ) = 0, the kernel of L is one dimensional consisting of

only constant functions. Thus there is an eigenbasis {ϕn }∞ n=0 of L with eigenvalues
{λn }∞ 3
n=0 such that ϕ0 is a non-zero constant function on S , λ0 = 0 and λn ̸= 0 for
n ≥ 1. Write
∞
X
at = a0 (t)ϕ0 dt + an (t)ϕn
n=1

for some smooth functions an : [−L + 1, L − 1] → R. As in Proposition 4.9,

∂t a0 + λ0 a0 = 0 and therefore a0 is a constant function. Now a′ = a − a0 ϕ0 dt is L2 -
orthogonal to ker L for all t. Since (d∗ + d+ )(a0 ϕ0 dt) = 0 we have (d∗ + d+ )a′ = 0.
34 JOSHUA TOMLIN

Proposition 4.9 gives constants C1 and δ, independent of L and r, such that

sup |a′ | ≤ C1 e−δ(L−2r−1) sup |a′ |
N (r) N (L−1)
!
≤ C1′ e−δ(L−2r) sup |a| + sup |a0 ϕ0 dt| .
N (L−1) N (L−1)

Since a0 and ϕ0 are constants, we can calculate

Z
2
∥a0 ϕ0 dt∥L2 = |a0 ϕ0 dt|2
N (L−1)

= 2vol(S )(L − 1)|a0 |2 |ϕ0 |2 .

3

The decomposition a = a′ + a0 ϕ0 dt is L2 -orthogonal, hence ∥a0 ϕ0 dt∥2L2 = ∥a∥2L2 −

∥a′ ∥2L2 . It follows that
2vol(S 3 )(L − 1)|a0 |2 |ϕ0 |2 = ∥a0 ϕ0 dt∥2L2
≤ ∥a∥2L2
≤ 2(L − 1)vol(S 3 ) sup |a|2 .
N (L−1)

1
Thus |a0 | ≤ |ϕ0 | supN (L−1) |a| and there is a constant Cδ with

sup |a′ | ≤ Cδ e−δ(L−2r) sup |a|.

N (r) N (L−1)

′ ′
Finally, |a ∧ dt| = |a ∧ dt| ≤ |a | and (4.19) follows. □

Corollary 4.11. Let A0 be a flat reference connection on N (L). Suppose ψ ∈

L2 (N (L), W + ) is a spinor such that DA0 ψ = 0. Then for any r ≥ 1 and L ≥ 2r,
′
(4.20) sup |ψ| ≤ Cδ′ e−δ (L−2r) sup |ψ|
S 3 ×[−r,r] S 3 ×[−L,L]

for some positive constants δ ′ and Cδ′ independent of L and r.

Proof. The spinc structure on X is defined so that, on the neck, Clifford multipli-
cation Γ : T N (L) → End(W ) is induced by the Clifford multiplication γ : T S 3 →
End(WS 3 ) on S 3 .
   
0 γ(∂xi ) 0 id
(4.21) Γ(∂xi ) = , Γ(∂ t ) = .
−γ(∂xi )∗ 0 −id 0
Here {∂t , ∂x1 , ∂x2 , ∂x3 } is a basis for T N (L) corresponding to local coordinates
(x1 , x2 , x3 , t) of N (L). The spinc connection ∇A0 for the reference connection A0
is given by the formula
∂ 3
(4.22) ∇A0 = dt ⊗ + ∇S .
∂t
3
Here ∇S is a spinc connection on WS 3 → S 3 . Since b2 (S 3 ) = 0, it can be as-
3
sumed that ∇S is flat. This equation is understood by treating a spinor ψ ∈
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
35

C ∞ (N (L), W + ) as a time-dependent family of spinors {ψt } on S 3 . Over the neck

N (L), the Dirac operator DA0 : C ∞ (X, W + ) → C ∞ (X, W − ) takes the form
3
∂ X 3
DA0 = Γ(∂t ) · + Γ(∂xi ) · ∇Sxi
∂t i=1
3
∂ X 3
= Γ(∂t ) · − Γ(∂t ) · γ(∂xi )∇Sxi
∂t i=1
 
∂ S3
(4.23) = Γ(∂t ) −D .
∂t
3 3
Here DS is the self-adjoint Dirac operator associated to ∇S . Note that both
3
DA0 and DS are elliptic. Since A0 is flat and S 3 has positive scalar curvature,
3
the Weitzenböck formula implies that ker DS = 0. Therefore ψ is automatically
3
orthogonal to ker DS and the result follows from Proposition 4.9. □

To complete the proof of Proposition 4.9, it remains to prove the following lemma.

Lemma 4.12. Let A : C ∞ (S 3 , E) → C ∞ (S 3 , E) be an elliptic, self-adjoint, pseudo-

differential operator of positive order. Let 0 < |λ1 | ≤ |λ2 | ≤ ... denote the non-zero
eigenvalues of A, ordered by magnitude. There exists a constant C such that, for
any r ≥ 1,
∞
!2 ∞
X
2 r|λn |
X sinh(4r|λn |)
(4.24) (λn + 1)e |an | ≤C |an |2
n=1 n=1
|λ n |
for any real number sequence {an }.

Proof. First, apply the Cauchy-Schwarz inequality to obtain

∞
!2 ∞ p ! p !!2
X
2 r|λn |
X (λ2n + 1) |λn |er|λn | sinh(4r|λn |)
(λn + 1)e |an | = p p |an |
n=1 n=1 sinh(4r|λn |) |λn |
∞
! ∞ !
X (λ2n + 1)2 |λn |e2r|λn | X sinh(4r|λn |)
2
≤ |an |
n=1
sinh(4r|λn |) n=1
|λn |
P∞ (λ2 +1)2 |λn |e2r|λn | e4x
It suffices to bound n=1 n sinh(4r|λ n |)
. Fix 0 < δ < |λ1 |. The function sinh(4x)
is bounded on [δ, ∞), therefore there is a constant C1 such that, for all x ≥ δ,
e2x
≤ C1 e−2x
sinh(4x)
Apply this to r|λn | to produce
∞ ∞
X e2r|λn | (λ2n + 1)2 |λn | X
≤ C1 (λ2n + 1)2 |λn |e−2r|λn |
n=1
sinh(4r|λn |) n=1
∞
X
≤ C1 (λ2n + 1)2 |λn |e−2|λn | .
n=1
36 JOSHUA TOMLIN

Similarly, there exists a constant C2 such that x(x2 + 1)2 e−x ≤ C2 for all x ≥ 0. It
follows that
∞
X ∞
X
(4.25) C1 (λ2n + 1)2 |λn |e−2|λn | ≤ C1 C2 e−|λn | .
n=1 n=1

Since A is elliptic and self-adjoint, Weyl’s law [18, Lemma 1.6.3] implies that there
exists a constant C3 and an exponent α > 0 such that |λn | ≥ C3 nα for large enough
n. Thus to show that (4.25) is finite, it is enough to show that
∞
X a
e−n < ∞.
n=1

This follows from the integral test. Let u = xα so that

Z ∞
1 ∞ 1−α −u
Z
α
e−x dx = u α e du
1 α 1
 
1 1
≤ Γ
α α
< ∞.
P∞
Therefore C = n=1 C1 C2 e−|λn | is a suitable constant. □

5. Proof of the Families Permutation Theorem

Now we construct a homotopy from µX to V −1 µX τ V that, after restricting to

a suitably chosen disk bundle, is a homotopy through compact perturbations of l.
Such a homotopy proves Theorem 3.3 because of Corollary 2.23. The final homotopy
is a concatenation of three compact homotopies, each dealing with problematic
quadratic terms of µX separately. The idea to use these particular homotopies
comes from Bauer’s proof in [8], however great care is taken to ensure that these
homotopies satisfy the necessary boundedness conditions and that these conditions
are compatible with stretching the neck length.

Fix L > 2 and let E = E(L) → B be a family of closed 4-manifolds X with a

separating neck of length 2L. Fix a reference connection A0 , which can be assumed
to be flat on the neck NB (L). Recall that for θ ∈ J , Aθ denotes the associated
connection A0 + iθ. Note that Aθ is also flat on the neck. For a given R ≤ L, let
ρR : E → [0, 1] be a smooth function that vanishes on NB (R − 1) and is identically
1 on E − NB (R). Along NB (R) − NB (R − 1), we require that ρR only depends on
the interval coordinate. For s ∈ [0, 1], let ρsR be a linear homotopy ending at ρR of
the form

ρsR = (1 − s) + sρR .

Since ρsR is constant outside of NB (R) − NB (R − 1), the C k -norm of ρsR is indepen-
dent of L for all R and s.
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
37

5.1. The first homotopy. To define the first homotopy F : A → C fiberwise, let
θ ∈ H 1 (Xb ; R) for some b ∈ B and set
Fsθ (ψ, a) = (DAθ ψ + ia · ψ, d+ a − iFA+θ + iρsL σ(ψ), d∗ a, pr(a)).
Notice that F0 = µX and that the quadratic term in the second factor of F1 vanishes
on N (L − 1). The proof that (Fs )−1 (0) is L2k -bounded uses variations on techniques
that show compactness of the moduli space in ordinary Seiberg-Witten theory.

Proposition 5.1. Fix a connection Aθ for Xb (L) with θ ∈ Jb for some b ∈ B. For
s ∈ [0, 1], the preimage (Fsθ )−1 (0) is uniformly L2k -bounded.

Proof. Let (ψ, a) ∈ (Fsθ )−1 (0) so that DAθ +ia ψ = 0 and FA+θ +ia = ρsL σ(ψ). The
Weitzenböck formula [26, Theorem 6.19] applied to the connection Aθ + ia gives a
pointwise bound
sX 2 ∗
∆g |ψ|2 + |ψ| + FA+θ +ia ψ, ψ ≤ 2 DA θ +ia
DAθ +ia ψ, ψ
2
sX 2 1 s 4
∆g |ψ|2 + |ψ| + ρL |ψ| ≤ 0.
2 2
Here sX is the scalar curvature of X = Xb (L) and ∆g is the positive definite
Laplace-beltrami operator, which is non-negative at a maximum. Let S = supX {0, −sX }
and note that sX is positive along the neck. Thus ∆g |ψ|2 ≤ 0 on N (L) and |ψ|2
achieves a maximum on M = X − N (L − 1). At such a maximum x ∈ M , we have
|ψ(x)|2 (sX (x) + |ψ(x)|2 ) ≤ 0.
It follows that |ψ|2C 0 ≤ S. To bound |a|C0 , notice that d+ a = −iρsL σ(ψ) + iFA+θ and
|d+ a| ≤ |σ(ψ)| + |FA+θ |.
Fix some p ≥ 4 so that the Sobolev embedding Lp1 (X, T ∗ X) ⊂ C 0 (X, T ∗ X) gives a
constant CS with |a|C 0 ≤ CS ∥a∥Lp1 . Since d∗ + d+ is a self-adjoint elliptic operator,
[17, Theorem 4.12] guarantees the existence of a constant Ce such that
|a|C 0 ≤ CS CE (∥σ(ψ)∥Lp + ∥FA+θ ∥Lp ).
This shows that |a|C 0 is bounded by a constant since |ψ|C 0 is. For bootstrapping,
DAθ ψ = −ia·ψ and ∥d+ a∥Lpi = ∥−ρsL σ(ψ)+FA+θ ∥Lpi . The C k -norm of ρsL determines
a constant C such that, for any 0 ≤ i ≤ k and 2 ≤ p ≤ 2k+1
∥d+ a∥Lpi ≤ C∥σ(ψ)∥Lpi + ∥FA+θ ∥Lpi .
From Proposition 4.7 and Remark 4.8 there a constant CB and integer d ≥ 1 such
that
∥(ψ, a)∥L2k ≤ CB Ld (1 + |(ψ, a)|C 0 )d .
The norm |(ψ, a)|C 0 is bounded by a constant, hence so is ∥(ψ, a)∥L2k . This bound
is independent of s, but depends on the connection Aθ and neck length L. □

Proposition 5.2. The map Fs : A → C is a homotopy through compact perturba-

tions of l.
38 JOSHUA TOMLIN

Proof. For each s ∈ [0, 1], it is clear that Fs = l + cs with cs compact. Proposition
5.1 gives for each [θ] ∈ J a radius Rθ > 0 such that, for any (ψ, a) ∈ (Fsθ )−1 (0),
∥(ψ, a)∥L2k ≤ Rθ .
This bound does not depend on s ∈ [0, 1]. Let R be the supremum of Rθ over J ,
which exists since J is compact. Let D ⊂ A be a disk bundle over J with L2k -radius
2R. This shows in fact that each preimage Fs−1 (0) is contained in a bounded disk
bundle, a stronger result than required. □

5.2. The second homotopy. The second homotopy Gs for s ∈ [0, 3] is constructed
in three stages. For s ∈ [0, 1] define
Gθs (ψ, a) = (DAθ ψ + iρsr a · ψ, d+ a − iFA+θ + iρL σ(ψ), d∗ a, pr(a))
This homotopy eliminates the other quadratic term ia · ψ from NB (r − 1). The
constant r ≥ 3 will be defined later. It is assumed without loss of generality that
L ≥ 2r + 1.

To define the second stage of G, let P = G1 . This stage will transform P to

P τ = V −1 P V where the action of V was defined in equation 3.9. Restricting to
NB (r − 1), P is a first order linear differential operator given by the formula
P θ (ψ, a) = (DAθ ψ, d+ a, d∗ a, pr(a)).
Note that FA+θ = 0 since Aθ is flat on the neck. For s ∈ [0, 1], let
Vs (x, t) = γ((s − 1) · φ(t)) : S(V0 ) × [−L, L] → SO(n).
Define Qs : A → C by
Qθs (ψ, a) = Vs−1 ∂t Vs (dt · ψ, (dt ∧ a)+ , ∗(∗⃗a ∧ dt), 0).
⃗ (dt ∧ ⃗a)+ and
Here V −1 ∂t V is a matrix functions which acts on each vector dt · ψ,
∗(∗⃗a ∧ dt). Notice that Q vanishes outside of N (1) since ∂t V = 0 away from the
short neck. Applying the Leibniz rule, it follows that
Vs−1 P Vs (ψ, a) = P (ψ, a) + Qs (ψ, a).
For s ∈ [1, 2], define Gs by
(5.1) Gs = P + Q s .
Each Qs has the property that Qs = 0 outside of N (1), hence this formula is well
defined globally. Restricted to the neck N (L − 1), equation (5.1) is equivalent to
Gs = Vs−1 P Vs . For the final stage s ∈ [2, 3], let Gs = V −1 G3−s V . Now G is a
homotopy from G0 = F1 to G3 = V −1 F1 V .

Since G alters the DA+ia ψ = 0 equation, the previous argument fails to bound
G−1
s (0). However to show that G is a compact homotopy, it is only necessary
to find an L2k -disk bundle containing G−1 −1
0 (0) and G3 (0) for which its bounding
−1
sphere bundle does not intersect Gs (0) for any s ∈ [0, 3]. The following results
help accomplish this by proving similar results for the C 0 -norm of zeroes of Gs .
For any [θ] ∈ J with θ ∈ H 1 (Xb ; R), we set X = Xb (L).
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
39

Lemma 5.3. Let (ψ, a) ∈ (Gθs )−1 (0) for some s ∈ [0, 3] and [θ] ∈ J . If supX |ψ|
is achieved at some x ∈ M , then |ψ|2C 0 (X) ≤ S for S = supX {0, −sX }.

Proof. Restricted to M = X − N (L − 1), the pair (ψ, a) satisfies DA+ia ψ = 0 and

+
FA+ia = σ(ψ). As in Proposition 5.1, the Weitzenböck formula on M gives
sX 2 1 4
∆g |ψ|2 + |ψ| + |ψ| ≤ 0
2 2
Since X is a closed 4-manifold, ∆g |ψ|2 ≥ 0 at x. Since x ∈ M , it follows that
|ψ(x)|2 (sX (x) + |ψ(x)|2 ) ≤ 0.
Therefore |ψ|2 ≤ S since |ψ(x)| = |ψ|C 0 (X) . □

Lemma 5.4. Let (ψ, a) be a spinor-from pair along the n-component neck N (L).
For any 0 ≤ R ≤ L, we have
sup |ψ| ≤ n sup |Vs ψ| ≤ n2 sup |ψ|
N (R) N (R) N (R)
2
sup |a| ≤ n sup |Vs a| ≤ n sup |a|.
N (R) N (R) N (R)

Proof. We prove only the spinor case. Let ψ ⃗ be the vectorised version of ψ as in
⃗ 3 n +
(3.8). That is, ψ : S ×[−L, L] → ⊕i=1 W with the i-th component ψ ⃗i correspond-
ing to the restriction of ψ to the ith connected component of N (L). The restriction
of Vs ψ to the ith connected component of N (L) is given by the ith component of
⃗ Inside N (R), we have
Vs ψ.

X
⃗ i| =
|(Vs ψ) ⃗j
(Vs )ij ψ
j
X
≤ ⃗j |
|(Vs )ij ||ψ
j
 
X
≤ |(Vs )ij | sup |ψ|
j N (R)

= n sup |ψ|.
N (R)

The last line follows since Vs is valued in SO(n), hence the absolute value of each
of its entries is less than 1. Therefore supN (R) |Vs ψ| ≤ n supN (R) |ψ|. The same
calculation shows that supN (R) |ψ| = supN (R) |Vs−1 Vs ψ| ≤ supN (R) n|Vs ψ|. □

Remark 5.5: For any R ≤ L, the same calculation can be used to show that
sup |ψ| ≤ n sup |Vs ψ| ≤ n2 sup |ψ|
∂N (R) ∂N (R) ∂N (R)
2
sup |a| ≤ n sup |Vs a| ≤ n sup |a|.
∂N (R) ∂N (R) ∂N (R)
40 JOSHUA TOMLIN

Lemma 5.6. There exists positive constants L0 , CE , δ and Cδ such that the fol-
lowing holds. For any s ∈ [0, 3], let (ψ, a) ∈ (Gθs )−1 (0) be a spinor-form pair on
Xb (L). If L ≥ L0 , then
|a|C 0 (X) ≤ CE |(d∗ + d+ )a|C 0 (M )
(5.2) sup |a ∧ dt| ≤ Cδ e−δ(L−2r) sup |a|.
N (r) N (L−1)

Proof. For s ∈ [0, 1], we have

d+ a = iFA+θ − iρL σ(ψ)
d∗ a = 0
pr(a) = 0.

Along N (L − 1), d+ a = iFA+θ and therefore d+ a = 0 since Aθ is flat on the neck.

Thus (d∗ + d+ )a vanishes on N (L − 1). Hence Proposition 4.5 gives constants CE′

and L1 such that, if L ≥ L1 then

′
|a|C 0 (X) ≤ CE |(d∗ + d+ )a|C 0 (M ) .

Further, Corollary 4.10 applies to a∧dt yielding, for some δ > 0 and Cδ′ independent
of L,

sup |a ∧ dt| ≤ Cδ′ e−δ(L−2r) |a|C 0 (N (L)) .

N (r)

If s ∈ [1, 2], the condition pr(a) = 0 still holds. Restricting to N (L − 1) we have

Vs−1 P Vs (ψ, a) = 0 and therefore Vs (ψ, a) is a solution to P . Note that Vs (ψ, a) is
only defined on the neck when s ∈ (0, 1) and that (d∗ + d+ )Vs a = 0 on N (L − 1).
This means that supN (L−1) |Vs a| = sup∂N (L−1) |Vs a| by the maximum principle.
Lemma 5.4 implies that
sup |a| ≤ n sup |Vs a|
N (L−1) N (L−1)

=n sup |Vs a|
∂N (L−1)

(5.3) ≤ n2 sup |a|.

∂N (L−1)

Thus |a|C 0 (X) ≤ n2 |a|C 0 (M ) . Restricting to X − N (1) instead, we have P (ψ, a) = 0.

This means that (d∗ + d)a = 0 along N (2, L). Now (5.3) with Remark 4.6 implies
′′
the existence of constants L2 and CE such that, if L ≥ L2 ,

|a|C 0 (X) ≤ n2 |a|C 0 (M )

′′
(5.4) ≤ n2 CE |(d∗ + d+ )a|C 0 (M ) .

To obtain the exponential bound on a ∧ dt, note that Vs (a ∧ dt) = (Vs a) ∧ dt. We
have (d∗ + d+ )Vs a = 0 on N (L − 1) and Corollary 4.10 applies to Vs a ∧ dt, yielding

sup |Vs a ∧ dt| ≤ Cδ′ e−δ(L−2r) sup |Vs a|.

N (r) N (L−1)
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
41

By Lemma 5.4, it follows that

sup |a ∧ dt| ≤ n sup |Vs a ∧ dt|
N (r) N (r)

≤ nCδ′ e−δ(L−2r) sup |Vs a|

N (L−1)

(5.5) ≤n 2
Cδ′ e−δ(L−2r) sup |a|.
N (L−1)

For the third stage s ∈ [2, 3], we have V −1 G3−s V (ψ, a) = 0. Thus V (ψ, a), which
is defined globally, is a solution of G3−s . The argument for the second stage can
′ ′′
be repeated to establish (5.4) and (5.5). Setting CE = max{CE , n2 CE }, L0 =
2 ′
max{L1 , L2 } and Cδ = n Cδ ensures that (5.2) is satisfied for any s ∈ [0, 3]. □

Proposition 5.7. Let [θ] ∈ Jb for some b ∈ B. There exists positive constants
U0 , L0 , C, δ and r such that the following holds. If L ≥ L0 , then for any s ∈ [0, 3],
there are no solutions (ψ, a) ∈ (Gθs )−1 (0) with C 0 -norm in the interval [U0 , U (L)],
where
(5.6) U (L) = Ceδ(L−2r) .

Proof. Let (ψ, a) ∈ (Gθs )−1 (0) for some s ∈ [0, 3]. Notice that for any stage of Gs ,
on X − N (r) the pair (ψ, a) satisfies
DAθ +ia ψ = 0
d+ a = iFA+ − iρL σ(ψ)
d∗ a = 0
pr(a) = 0.
Lemma 5.6 gives constants CE and L0 such that, for L ≥ L0 ,
|a|C 0 (X) ≤ CE |(d∗ + d+ )a|C 0 (M ) .
Applying the Seiberg-Witten style equations above gives
|a|C 0 (X) ≤ CE (|FA+θ |C 0 + |σ(ψ)|C 0 (M ) )
1
= CE (|FA+θ |C 0 + |ψ|2C 0 (M ) ).
2
Recall that S = supX {−sX , 0} where sX is the scalar curvature of X. Let
√ 1
U0′ = 1 + S + CE (|FA+θ |C 0 + S).
2
+
Note that |FAθ |C 0 and S do not depend on L. To show that |(ψ, a)|C 0 (X) < U0′
it is enough to show that |ψ|2C 0 (X) ≤ S. By Lemma 5.3, it suffices to show that
supX |ψ| = supM |ψ|.

For now assume that s ∈ [0, 1] so that ψ satisfies DAθ +iρsr a ψ = 0 and d+ a =
iFA+θ −iρL σ(ψ). Inside N (L−1), the Weitzenböck formula applied to the connection
A′ = Aθ + iρsr a gives
D
∗ sN E
∆g |ψ|2 ≤ DA ′ DA′ ψ − ψ − FA+′ ψ, ψ .
2
42 JOSHUA TOMLIN

Here sN is the scalar curvature of the neck, which is a positive constant. Since Aθ
is flat on the neck, FA+′ = d+ (iρsr a). But d+ a = 0 on N (L − 1), so it follows that
sN
∆g |ψ|2 ≤ − |ψ|2 + ∥(dρs2 ∧ a)+ ∥|ψ|2
2
2
√ sN 
(5.7) = |ψ| 2|(dρsr ∧ a)+ | − .
2
Here ∥(dρsr ∧ a)+ ∥ is the operator norm of d+ (ρsr a) = (dρsr ∧ a)+ identified as an
element of End0 (W + ) and
√ |(dρsr ∧ a)+ | is the norm of (dρsr ∧ a)+ as a 2-form. The
relation ∥(dρsr ∧ a)+ ∥ = 2|(dρsr ∧ a)+ | is shown in [26, Lemma 7.4].

Since dρsr is supported in N (r), (5.7) guarantees that ∆g |ψ|2 < 0 on N (L−1)−N (r).
It remains to show that ∆g |ψ|2 < 0 on N (r). Since ρsr is constant on spheres,
dρsr = ∂t ρsr dt. Define
√
R = 2 sup |∂t ρsr |N (r) .
s∈[0,1]

If follows that
 sN 
(5.8) ∆g |ψ|2 ≤ |ψ|2 R|a ∧ dt| − .
2
Lemma 5.6 provides constants δ, Cδ such that if L ≥ L0 , then
sup |a ∧ dt| ≤ Cδ e−δ(L−2r) sup |a|.
N (r) N (L−1)

Define the constant C > 0 by

sN
(5.9) C= .
4RCδ
This is positive since sN , R and Cδ are. Define U ′ (L) by
U ′ (L) = Ceδ(L−2r) .
Note that the definition of C is independent of L and Aθ . Further, it can be assumed
that L is large enough to ensure that U ′ (L) > U0′ . When |(ψ, a)|C 0 ≤ U ′ (L) and
L ≥ L0 , inside N (r) we have
R|a ∧ dt| ≤ RCδ e−δ(L−2r) sup |a|
N (L−1)
−δ(L−2r)
≤ RCδ e U ′ (L)
sN
(5.10) ≤ .
4
From (5.8) it follows that ∆g |ψ|2 < 0 on all of N (L − 1). Therefore supN (L−1) |ψ| =
sup∂N (L−1) |ψ| because ∆g |ψ|2 is non-negative at an interior local maximum. Con-
sequently supX |ψ| = supM |ψ|, thus |ψ|C 0 ≤ S and |(ψ, a)|C 0 < U0 . It remains to
shows that |ψ|C 0 ≤ S for s ∈ [1, 3].

Now suppose (ψ, a) ∈ G−1 ′

s (0) for some s ∈ [1, 2] with |(ψ, a)|C 0 ≤ U (L). Recall
that Gs = P + Qs and Qs = 0 outside of N (1), hence P (ψ, a) = 0 on X − N (1).
Alternatively, Gs = Vs−1 P Vs on the neck so Vs (ψ, a) is a solution to P on N (L − 1).
Again we prove that |ψ|2C 0 ≤ S by showing that supX |ψ| = supM |ψ|.
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
43

Restricting to N (1, L − 1) = N (L − 1) − N (1), the Weitzenböck formula as before

for the connection A′ = Aθ + iρr a gives
 sN 
∆g |ψ|2 ≤ |ψ|2 R|a ∧ dt| − .
2
For L ≥ L0 , Lemma 5.6 still applies to (ψ, a) yielding
(5.11) sup |a ∧ t| ≤ Cδ e−δ(L−2r) sup |a|.
N (r) N (L−1)

Thus the calculation in (5.10) guarantees ∆g |ψ|2 < 0 on N (1, L − 1). This implies
that
(5.12) sup |ψ| = max{ sup |ψ|, sup |ψ|}.
X N (1) M

Notice that DAθ Vs ψ = 0 on N (r − 1). Thus Corollary 4.11 implies the existence of
constants δ ′ , Cδ′ > 0 such that
′
(5.13) sup |Vs ψ| ≤ Cδ′ e−δ (r−2) sup |Vs ψ|.
N (1) N (r−1)

Fix a large enough r to ensure that

′ 1
(5.14) Cδ′ e−δ (r−2) ≤.
n2
Note that this definition of r is independent of L, and we can assume that L0 ≥ 2r.
Since Vs ψ is a solution to P along N (L − 1), we have that
sup |Vs ψ| = sup |Vs ψ|.
N (L−1) ∂N (L−1)

This follows from the the argument presented in the s ∈ [0, 1] case. It follows from
Lemma 5.4, (5.13) and (5.14) that
sup |ψ| ≤ n sup |Vs ψ|
N (1) N (1)
′
≤ nCδ′ e−δ (r−2) sup |Vs ψ|
N (r−1)
1
≤ sup |Vs ψ|
n N (r−1)
1
≤ sup |Vs ψ|
n ∂N (L−1)
≤ sup |ψ|.
∂N (L−1)

That is, supN (1) |ψ| ≤ supM |ψ| and therefore supX |ψ| = supM |ψ| by (5.12). Thus
Lemma 5.3 guarantees |ψ|2 ≤ S and |(ψ, a)| < U0′ .

For the third stage s ∈ [2, 3], we have Gs (ψ, a) = V −1 G3−s V (ψ, a) = 0. Note that
V (ψ, a) is defined globally and thus V (ψ, a) is a solution of G3−s . Further, by the
same calculation as Lemma 5.4, |V (ψ, a)|C 0 ≤ n|(ψ, a)|C 0 ≤ n2 |V (ψ, a)|C 0 . This
implies that if |(ψ, a)|C 0 ≤ n1 U ′ (L), then |V (ψ, a)|C 0 ≤ U ′ (L) and |(ψ, a)| ≤ nU0′ .
The result follows by taking U (L) = n1 U ′ (L) and U0 = nU0′ , ensuring that L0 is
large enough so that U (L) > U0 for L ≥ L0 . □
44 JOSHUA TOMLIN

The above lemma shows that given a neck length L and a connection Aθ , there
are no elements of (Gθs )−1 (0) with C 0 -norm in the interval [U0 , U (L)]. This will
be used to find an L2k -disk in Aθ with boundary that does not intersect (Gθs )−1 (0)
for any s ∈ [0, 3]. The L2k -norm of a pair (ψ, a) ∈ (Gθs )−1 (0) can be bounded by a
polynomial in |(ψ, a)|C 0 and L. The exponential increase of U (L) counteracts this
polynomial growth. First we show that the endpoints (Gθ0 )−1 (0) and (Gθ3 )−1 (0) are
contained in an L2k -disk with radius that increases polynomially with L.

Lemma 5.8. Let [θ] ∈ Jb for some b ∈ B. There exists positive constants C, d and
L0 such that, for any L ≥ L0 ,
∥(ψ, a)∥L2k ≤ CLd
for any solution (ψ, a) ∈ (Gθ0 )−1 (0) ∪ (Gθ3 )−1 (0) on Xb (L).

Proof. For (ψ, a) ∈ (Gθ0 )−1 (0) we have

DAθ +ia ψ = 0
d+ a = iFA+θ − iρL σ(ψ)
d∗ a = 0.
As in Proposition 5.1, the Weitzenböck formula gives
|ψ|2C 0 ≤ S.
Since (d + d∗ )a = 0 on N (L − 1), Proposition 4.5 provides constants L0 and C ′
such that L ≥ L0 implies
|a|C 0 ≤ C ′ |(d∗ + d+ )a|C 0
1
≤ C ′ (|FA+ |C 0 + S).
2
√
Let U = 1 + S + C ′ (|FA+ |C 0 + 21 S) so that |(ψ, a)|C 0 < U . Notice that |ρL σ(ψ)| ≤
|σ(ψ)| and that dρL is supported on N (L) − N (L − 1). Therefore the C k -norm of
ρ can be used to obtain a constant Cρ such that ∥ρL σ(ψ)∥Lpi ≤ Cρ ∥σ(ψ)∥Lpi with
Cρ independent of L. Now applying elliptic bootstrapping as in Remark 4.8, there
are constants CB and d such that
∥(ψ, a)∥L2k ≤ CB Ld (1 + U )d
≤ C1 Ld .
The constant C1 is independent of L since CB , d and U are.

The argument for (ψ, a) ∈ G−13 (0) is similar. Recall G3 = V

−1
G0 V so that V (ψ, a)
is a solution to G0 and therefore
∥V (ψ, a)∥L2k ≤ C1 Ld .
Applying V −1 gives
∥(ψ, a)∥L2k = ∥V −1 V (ψ, a)∥L2k
≤ CV −1 ∥V (ψ, a)∥L2k
≤ C1 CV −1 (1 + L)d .
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
45

Here CV −1 is a constant from (3.10) that is independent of L. The result follows

with C = max{C1 , CV −1 C1 }. □

It remains to find an L2k -disk bundle D with bounding sphere bundle S that does
not intersect G−1
s (0) for any s ∈ [0, 3]. This is done by combining Proposition 5.7
with the following elliptic bootstrapping result.

Lemma 5.9. Let θ ∈ Jb for some b ∈ B. There are constants CB and d such that,
for any L ≥ 2, if (ψ, a) ∈ (Gθs )−1 (0) for some s ∈ [0, 3] then
∥(ψ, a)∥L2k ≤ CB Ld (1 + |(ψ, a)|C 0 )d .

−1
Proof. First assume that s ∈ [0, 1] so that (ψ, a) ∈ (GA
s) (0) implies
DA ψ = −iρsr a
d+ a = iFA+ − iρL σ(ψ).

For any 0 ≤ i ≤ k and 2 ≤ p ≤ 2k+1 , there is a constant C1 such that

(5.15) ∥ρsr a∥Lpi ≤ C1 ∥a∥Lpi .

This constant comes from the C k -norm of ρsr . Since a and ρsr a only differ on
N (r) − N (r − 1), C1 is independent of L. Taking the supremum over s ∈ [0, 1], we
can assume that (5.15) holds for any s. Similarly,
∥d+ a∥Lpi ≤ ∥FA+ ∥Lpi + ∥ρL σ(ψ)∥Lpi
≤ C2 (∥FA+ ∥Lpi + ∥σ(ψ)∥Lpi ).
Once again C2 can be chosen independent of L. Now apply bootstrapping as in
Remark 4.8 to obtain
′
∥(ψ, a)∥L2k ≤ CB Ld (1 + |(ψ, a)|C 0 )d
′
for some constants CB > 0 and d ≥ 1, both independent of L. This proves the
result for s ∈ [0, 1].

If s ∈ [1, 2], we have P (ψ, a) = 0 on X − N (1) and P Vs (ψ, a) = 0 on N (1). On

N (1), the fact that DAθ Vs ψ = 0 and (d+ + d∗ )Vs a = 0 implies that
∥Vs (ψ, a)∥2L2 (N (1)) = ∥Vs (ψ, a)∥2L2 (N (1))
k

≤ 2vol(S 3 )|Vs (ψ, a)|2C 0 (N (1)) .

From Lemma 5.4 and (3.10) it follows that

∥(ψ, a)∥2L2 (N (1)) ≤ CVs−1 ∥Vs (ψ, a)∥2L2 (N (1))
k k

≤ 2CVs−1 vol(N (1)) · sup |Vs (ψ, a)|2

N (1)

(5.16) ≤ C3 |(ψ, a)|2C 0 .

46 JOSHUA TOMLIN

The elliptic bootstrapping argument of Lemma 4.7 can be applied to (ψ, a) over
X − N (1) to obtain
∥(ψ, a)∥2L2 = ∥(ψ, a)∥2L2 (X−N (1)) + ∥(ψ, a)∥2L2 (N (1))
k k k

≤ C4 Ld (1 + |(ψ, a)|C 0 )d + C3 |(ψ, a)|2C 0

′′ d
≤ CB L (1 + |(ψ, a)|C 0 )d .
Here we have assumed without loss of generality that d ≥ 2. For s ∈ [2, 3], we have
Gs (ψ, a) = V −1 G3−s V (ψ, a) = 0. Thus G3−s V (ψ, a) = 0 globally and Lemma 4.7
applies to V (ψ, a). Lemma 5.4 and (3.10) imply
∥(ψ, a)∥L2k ≤ CV −1 ∥V (ψ, a)∥L2k
≤ CV −1 C5 Ld (1 + |V (ψ, a)|C 0 )d
′′′ d
≤ CB L (1 + |(ψ, a)|C 0 )d .
′ ′′ ′′′
Hence the result follows with CB = max{CB , CB , CB }. □

Proposition 5.10. There are constants r and L0 such that, if L ≥ L0 , then Gs :

A → C is a homotopy through compact perturbations of l.

Proof. For any [θ] ∈ J , Lemma 5.8 provides constants C1θ and d such that, for large
enough L,
∥(ψ, a)∥L2k ≤ C1θ Ld

for any (ψ, a) ∈ (Gθ0 )−1 (0) ∪ (Gθ3 )−1 (0). The constant d from the bootstrapping
argument only depends on k, hence the same d can be used for each θ. Let C1 =
supθ∈J C1θ so that
(5.17) ∥(ψ, a)∥L2k ≤ C1 Ld

for (ψ, a) in any fibre of G−1 −1

0 (0) ∪ G3 (0).

Again for each [θ] ∈ J , Proposition 5.7 provides constants U0θ , C, δ and r such that,
for large enough L,
|(ψ, a)|C 0 ≤ U (L) ⇒ |(ψ, a)|C 0 < U0θ

so long as (ψ, a) ∈ (Gθs )−1 (0) for some s ∈ [0, 3]. Recall that U (L) = Ce−δ(L−2r) .
The constant δ is chosen based on the eigenvalues of the first order elliptic operator
L on S 3 defined in (4.4). Thus the same δ can be used for any θ on any fibre Xb (L)
of E. Further, from (5.9) we can see that C only depends on δ, the scalar curvature
of S 3 × [−L, L], and the derivative of ρ. Hence C is also independent of θ and b.
By similar reasoning, r can also be chosen independently from θ and b by (5.14).

Letting U0 = supθ∈J U0θ , it follows that

(5.18) |(ψ, a)|C 0 ≤ U (L) ⇒ |(ψ, a)|C 0 < U0
so long as (ψ, a) is an element of some fibre of G−1
s (0) for some s ∈ [0, 3].
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
47

By taking a supremum over fibrewise Sobolev embeddings, there is a constant

CS = supb∈B CSb such that, for any L2k -pair (ψ, a) on any fibre Xb (L),
(5.19) |(ψ, a)|C 0 ≤ CS ∥(ψ, a)∥L2k .
Lemma 4.1 ensures that CS can be chosen independently from L. Finally, to facil-
θ
itate bootstrapping, for each [θ] ∈ J Lemma 5.9 gives a constant CB such that
θ d
∥(ψ, a)∥L2k ≤ CB L (1 + |(ψ, a)|C 0 )d
This holds so long as (ψ, a) ∈ (Gθs )−1 (0) for some s ∈ [0, 3]. Once again let CB =
θ
supθ∈J CB so that
(5.20) ∥(ψ, a)∥L2k ≤ CB Ld (1 + |(ψ, a)|C 0 )d
so long as (ψ, a) is an element of some fibre of G−1
s (0) for some s ∈ [0, 3].

Set R(L) = UC(L)

S
and let D ⊂ A be a disk bundle with L2k -radius R(L). Let S
denote the bounding sphere bundle of D. Choose L0 large enough so that L ≥ L0
implies
R(L) ≥ max{C1 Ld , 2CB Ld (1 + U0 )d }.
This is achievable since R(L) increases exponentially. By (5.17), R(L) contains
G−1 −1
0 (0) ∪ G3 (0). Further, suppose (ψ, a) ∈ (Gs )
θ −1
(0) ∩ D for some s ∈ [0, 3]
and [θ] ∈ J . Then ∥(ψ, a)∥L2k ≤ R(L) and by (5.19), |(ψ, a)|C 0 ≤ U (L). Thus
|(ψ, a)|C 0 < U0 by (5.18) and (5.20) implies that
∥(ψ, a)∥L2k ≤ CB Ld (1 + U0 )d
1
≤ R(L).
2
That is, (Gθs )−1 (0) does not intersect S for any θ ∈ J and s ∈ [0, 3]. □

5.3. The third homotopy. The third homotopy Hs for s ∈ [0, 1] is given by
Hs = V −1 F1−s V.
This homotopy starts at H0 = G3 = V −1 F1 V and ends at H1 = V −1 µE τ V .

Proposition 5.11. The homotopy Hs is a homotopy through compact perturbations

of l.

Proof. A solution (ψ, a) ∈ (Hs )−1 (0) satisfies F1−s

θ
V (ψ, a) = 0 for some b ∈ B and
[θ] ∈ Jb . Proposition 5.2 provides a constant R > 0, independent of s and θ, such
that
∥V (ψ, a)∥L2k ≤ R.
It follows from (3.10) that
∥(ψ, a)∥L2k = ∥V −1 V (ψ, a)∥L2k
≤ CV −1 R.
48 JOSHUA TOMLIN

The constant CV −1 can be chosen independently of θ ∈ J . The disk bundle D ⊂ Ak

with fibres of 2L2k -radius CV −1 R contains Hs−1 (0) for all s ∈ [0, 1]. □

Proof of Theorem 3.3. The concatenation F · G · H is a homotopy from µE to

V −1 µE τ V through compact perturbations of l. By Corollary 2.23, the Bauer-
+
Furuta classes [µE ] and [µE τ ] are equal in πTb n ,U (JE , ind D), where the class [µE τ ]
is represented by the bounded Fredholm map V −1 µE τ V . □

Remark 5.12: The definition of the separating neck NB (L) required that the
fibres of the neck components are of the form S 3 × [−L, L], with the application to
connected sums in mind. However in Section 4, no particularly special properties
of S 3 were used. We only used that fact that S 3 has a positive scalar curvature
metric and that b1 (S 3 ) = 0. Thus Theorem 3.3 will extend to the case that the
fibres of the neck are a product M 3 × [−L, L] with M 3 any spherical 3-manifold.

6. The Families Bauer-Furuta Connected Sum Formula

For j ∈ {1, 2}, let Ej → B be a family of closed, oriented 4-manifolds Xj . To define

the families connected sum, it is necessary to have sections ij : B → Ej with normal
bundles Vj → B and an orientation reversing isomorphism φ : V1 → V2 . Since the
fibre of Ej is 4-dimensional, Vj is a real 4-dimensional vector bundle. Fix a metric
on Vj and identify the open unit disk bundle D(Vj ) as a tubular neighbourhood of
ij with S(Vj ) the bounding unit sphere bundle. Let Uj = Ej − D(Vj ) so that
E1 = U1 ∪S(−V1 ) D(V1 )
(6.1) E2 = D(V2 ) ∪S(V2 ) U2 .
Here we are interpreting S(V2 ) as the outgoing boundary of D(V2 ) and S(−V1 ) as
the ingoing boundary of D(V1 ), hence the negative sign. Thus φ identifies S(−V1 )
with S(V2 ). Topologically the families connected sum E = E1 #B E2 is defined as
(6.2) E = U1 ∪S(−V1 ) U2 .
We write S(V ) ⊂ E to denote S(−V1 ) ⊂ U1 , which has been identified with
φ(S(−V1 )) = S(V2 ) ⊂ U2 . To define a metric on E, attach cylinders to E1 and E2
to get
Ê1 = U1 ∪S(−V1 ) (S(V1 ) × [0, ∞))
Ê2 = (S(V2 ) × (∞, 0]) ∪S(V2 ) E2 .
Let g1 be the metric on S(V1 ) × [0, ∞) which restricts to a product of the standard
round metric and interval metric on the fibres. The metric g1 can be smoothly
extended to Ê1 using a collar neighbourhood. Repeat the same process to get a
metric g2 on Ê2 . For L > 0, let
Ê1 (L) = Ê1 − (S(V1 ) × (L + 1, ∞))
Ê2 (L) = Ê2 − (S(V2 ) × (−∞, −L − 1)) .
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
49

For gluing along the cylindrical ends, define a smooth map

f : S 3 × [L − 1, L + 1] → S 3 × [−L − 1, −L + 1]
f (x, t) = (x, t − 2L).
Now let E(L) = E1 (L)∪f E2 (L) with metric gE(L) = g1 ∪f g2 . By construction E(L)
is a 4-manifold family with standard fibre X(L) = X1 #X2 that has a separating
neck of length 2L. Up to diffeomorphism, the families connected sum E(L) depends
only on the given sections i1 and i2 and the orientation reversing diffeomorphism
of the normal bundles φ.

To get a spinc structure on E = E(L), let sj be a spinc structure on the vertical tan-
gent bundle T (Ej /B) for j ∈ {1, 2}. Write S(E) to denote the set of isomorphism
classes of spinc structures on E. There is a restriction map defined by
r : S(E) → S(E1 ) × S(E2 )
r(s) = (s|E1 , s|E2 )

Lemma 6.1. The restriction map r : S(E) → S(E1 ) × S(E2 ) is a bijection onto
the subset T ⊂ S(E1 ) × S(E2 ) defined by
T = {(s1 , s2 ) ∈ S(E1 ) × S(E2 ) | s1 |S(V ) ∼
= s2 |S(V ) }.

Proof. From (6.2) is it clear that the image of r is contained in T . Given (s1 , s2 ) ∈ T ,
a spinc structure s on E can be obtained from gluing, hence r is surjective. It
remains to prove injectivity. Suppose s, s′ are spinc structures on E with r(s) =
r(s′ ). That is, there are isomorphisms φj : s|Ej → s′ |Ej for j ∈ {1, 2}. If φ1 |S(V ) =
φ2 |S(V ) , then φ1 and φ2 would glue to give an isomorphism s → s′ .

Let ψ = φ−1
1 |S(−V ) ◦ φ2 |S(V ) so that φ2 |S(V ) = φ1 |S(V ) ◦ ψ. The map ψ is an auto-
morphism of spinc structures over S(V ) and therefore is determined by a smooth
map f : S(V ) → S 1 . We claim that f extends to a smooth map f˜ : E1 → S 1 .
Assuming this claim implies that ψ extends to an automorphism ψ̃ of s|E1 . Set-
ting φ′1 = φ1 ◦ ψ̃ : s|E1 → s′ |E1 gives an isomorphism of spinc structures with the
property that φ′1 |S(V ) = φ2 |S(V ) and the result follows by gluing.

To prove the claim, recall that the set of homotopy class of maps [S(V ), S 1 ] are in
bijection with H 1 (S(V ); Z). The Serre spectral sequence implies that H 1 (S(V ); Z)
is isomorphic to H 1 (B; Z) by pullback. That is, the homotopy class of f corre-
sponds to the pullback of an element α ∈ H 1 (B; Z). Pulling back α to H 1 (E1 ; Z)
corresponds to a homotopy class of [E1 , S 1 ] and we can choose a representative f˜
that restricts to f on S(V ). □

Corollary 6.2. For j ∈ {1, 2}, let Ej → B be a 4-manifold family equipped with a
spinc structure sj on the vertical tangent bundle. Let ij : B → Ej be a section with
normal bundle Vj and assume that an orientation reversing isomorphism φ : V1 →
V2 is given. An extension of s1 and s2 to the families connected sum E = E1 #B E2
50 JOSHUA TOMLIN

exists if and only if

φ(i∗1 (sE1 )) ∼
= i∗2 (sE2 ).

6.1. Families Bauer-Furuta formula. The families Bauer-Furuta connected sum

formula follows from the
`nTheorem 3.3 by the following observations. For a disjoint
union of families E = i=1 Ei the monopole map µE : A → C is the direct sum
n
M n
M n
M
µE = µEi : AEi → CEi .
i=1 i=1 i=1

Assume that each Ei is connected and let Ui be an S 1 -universe for Ei as in (3.6).

Then U = ⊕i Ui is a Tn -universe with Tn acting component-wise and the Bauer-
Furuta class of µE is an element of πTn ,U (J ; ind l).
`n
Proposition 6.3. If E = i=1 Ei is a disjoint union of families of 4-manifolds
over B, then the Bauer-Furuta class [µE ] ∈ πTn ,U (J ; ind l) is given by the fibrewise
smash product

[µE ] = [µE1 ] ∧J · · · ∧J [µEn ].

The above proposition follows directly from the definition of [µE ] outlined in Defini-
tion 2.12. The next observation demonstrates a method for calculating the Bauer-
Furuta invariant in the simplest cases. Recall that H + → J is the rank b+ (X)
2
trivial bundle with fibre H+ (X; R) and that SH + → J denotes the unit sphere
+
bundle in H ⊕ R. In the case that b1 (X) = 0, the Jacobian torus J(X) is just a
point and H + is a bundle over B.

Proposition 6.4. Let E → B be a 4-manifold family with fibre X such that

b1 (X) = 0 and assume a spinc structure on T (E/B) is given. Suppose there exists
a family of metrics {gb }b∈B on E with positive scalar curvature and that E admits
a family of flat spinc connections {Ab }b∈B . Then the class [µE ] is stably homotopic
to the inclusion

ι : B × S 0 → SH + .

Proof. Let n be the number of connected components of E. For t ∈ [0, 1] define a

homotopy

µt : L2k (E, W + ⊕ T ∗ (E/B)) ⊕ Rn → L2k−1 (E, W − ⊕ Λ2+ T ∗ (E/B) ⊕ R)

by the formula

µt (ψ, a, f ) = (DA+ta ψ, d+ a − tσ(ψ), d∗ a + f ).

Since b1 (X) = 0 and FA = 0, we have µ1 = µE . Further, µ0 is the linearised

monopole map l = DA ⊕ d+ ⊕ d∗ . We show that µt is a homotopy through compact
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
51

perturbations of l. Suppose that µt (ψ, a, f ) = 0 for some t ∈ [0, 1]. This implies
that
DA+ta ψ = 0
d+ a = tσ(ψ)
d∗ a = 0
f = 0.
It follows from the Weitzenböck formula that
s t2
∆g |ψ|2 + |ψ|2 + |ψ|4 ≤ 0.
2 2
At a maximum of |ψ| we obtain
s 2 t2
|ψ|C 0 + |ψ|4C 0 ≤ 0.
2 2
Since s > 0 we have ψ = 0. This in turn implies that d+ a = 0. Since d∗ a = 0 and
b1 (X) = 0, a is harmonic and therefore a = 0. Thus µ−1 (0) contains only one point
and certainly is bounded. That is, µ is a compact homotopy.

Recall that ind l = ind DA − b+ (X). The positive scalar curvature and the fact
that FA = 0 implies that both ker DA = 0 and coker DA = 0. Thus DA is an
isomorphism and therefore the Bauer-Furuta finite dimensional approximation of l
is stably homotopic to the inclusion ι. □

Let V → B be an SO(4)-vector bundle with a spinc structure s on the vertical

tangent space T (V /B). This induces a spinc structure on SV = S(R ⊕ V ) in the
following way. Let Fr(V ) denote the vertical oriented frame bundle of V . The
spinc structure on V determines a principle Spinc (4)-bundle PV → Fr(V ) which
pulls back to a principle Spinc (5)-bundle PR⊕V → Fr(R ⊕ V ). Let i : Fr(S(V )) →
Fr(R ⊕ V ) be the inclusion map of frames defined by the outward normal first
convention. Then i∗ (PR⊕V ) → Fr(S(V )) is the spinc structure on SV induced by s.

Corollary 6.5. Let V → B be an SO(4)-bundle with a spinc structure and give

π : SV → B the induced spinc structure on the vertical tangent bundle T (SV /B).
Then the class [µSV ] is stably homotopic to the identity id : B × S 0 → B × S 0 .

Proof. Since b1 (S 4 ) = b2 (S 4 ) = 0, the pullback map π ∗ : H 2 (B; Z) → π ∗ (SV ; Z)

is an isomorphism by the Serre spectral sequence. Let L → SV be the canonical
line bundle of the induced spinc structure on T (SV /B). Then the first chern class
c1 (L) ∈ H 2 (SV ; Z) is in the image of π ∗ . Thus there exists a connection A on L
with curvature FA = π ∗ (ω) for some 2-form ω ∈ Ω2 (B). Let ib : π −1 (b) → SV be
the inclusion of the fibre over b ∈ B. Then the restriction Ab = i∗b A is flat since
FAb = i∗b π ∗ ω = 0.

Since the structure group of V is SO(4), the fibres of SV can be equipped with
the standard round metric which has positive scalar curvature. By Proposition 6.4,
[µSV ] = [id]. □
52 JOSHUA TOMLIN

Finally, we have all the necessary tools to derive Bauer-Furuta connected sum
formula. We begin with the unparameterised case, which was first formulated by
Bauer in [8]. Afterwards, we prove the families formula which is a new result.

Theorem 6.6 ([8] Theorem 1.1). Let X = #i Xi be a connected sum of n closed,

oriented, 4-manifolds. The Bauer-Furuta invariant [µX ] is given by the formula
n
^
(6.3) [µX ] = [µXi ].
i=1

Proof. It is enough to prove the result for a connected sum of two 4-manifolds.
Define
Y1 = X1 #S 4
Y2 = S 4 #X2
(6.4) Y3 = S 4 #S 4 .
`
Set Y = i Yi . By the connected sum construction outlined in 6, we can choose a
metric that gives Y the structure of a separating neck. The negative components of
Y are given by the left summands of (6.4) and the positive components by the right
summands. Further, any choice of spinc structure on X1 and X2 extends uniquely
to a spinc structure on Y . Now [µY1 ] = [µX1 ], [µY2 ] = [µX2 ] and Proposition 6.4
implies that [µY3 ] = [id]. By Proposition 6.3 we have
[µY ] = [µX1 ] ∧ [µX2 ].
Let τ be the even permutation τ = (123) so that
Y τ = (X1 #X2 ) ⨿ (S 4 #S 4 ) ⨿ (S 4 #S 4 ).
Applying Propositions 6.3 and 6.4 again yields
[µY τ ] = [µX1 #X2 ].
Thus Theorem 3.3 implies that [µX ] = [µX1 ] ∧ [µX2 ]. □

Remark 6.7: In the construction of X τ it is assumed that τ is an even permuta-

tion, however this assumption is unnecessary for Theorem 3.3. If τ happens to be
odd, then replace X with the disjoint union
X ′ = X ⨿ (S 4 #S 4 ) ⨿ (S 4 #S 4 ).
Now include an extra transposition in τ that swaps the last two S 4 components.
As shown in the argument above, [µX ] = [µX ′ ].

Theorem 6.8 (Families Bauer-Furuta Connected Sum Formula). For j ∈ {1, 2},
let Ej → B be a 4-manifold family equipped with a spinc structure sj on the vertical
tangent bundle. Let ij : B → Ej be a section with normal bundle Vj and assume
that φ : V1 → V2 is an orientation reversing isomorphism satisfying
φ(i∗1 (sE1 )) ∼
= i∗2 (sE2 ).
A GENERAL CONNECTED SUM FORMULA FOR THE FAMILIES BAUER-FURUTA INVARIANT
53

Then the families Bauer-Furuta class of the fiberwise connected sum E = E1 #B E2

is
(6.5) [µE ] = [µE1 ] ∧J [µE2 ].

Proof. By Corollary 6.2, there is a unique spinc structure on the vertical tangent
space of E that extends s1 and s2 . Let Uj = Ej − D(Vj ) as in (6.1) so that
E1 = U1 ∪S(−V1 ) D(V1 )
E2 = D(V2 ) ∪S(V2 ) U2 .
Recall that S(V ) ⊂ E denotes S(−V1 ) ⊂ E1 and S(V2 ) = φ(S(−V1 )) ⊂ E2 . For any
L > 0, we can choose a metric on E1 and E2 that gives both of them a separating
neck of length 2L. Let F = E1 ⨿ E2 so that [µF ] = [µE1 ] ∧J [µE2 ] by Proposition
6.3. Let τ be the transposition (12) so that
F τ = U1 ∪S(V ) U2 ⨿ D(V2 ) ∪S(V ) D(V1 ) .

That is, F τ = E ⨿ SV2 . The spinc structure on SV2 is induced by s2 and therefore
[µSV2 ] = [id] by Corollary 6.5. Thus [µF τ ] = [µE ] by Proposition 6.3. Theorem 3.3
implies that [µF ] = [µF τ ] and therefore
[µE ] = [µE1 ] ∧J [µE2 ].
Note that the fact that τ is an odd permutation is not an issue by Remark 6.7. □

Of course, this formula extends to a connected sum of arbitrarily many families.

Further, the diffeomorphism type of the connected sum E = E1 #B E2 depends on
the sections i1 , i2 and the isomorphism φ, however the class [µE1 ] ∧J [µE2 ] does
not. That is, if E ′ is obtained as a connected sum of E1 and E2 for different i1 , i2
and φ, then [µE ] = [µE ′ ].

References

[1] J. F. Adams. Stable homotopy and generalised homology. Chicago Lectures in Mathematics.
University of Chicago Press, Chicago, IL, 1995. Reprint of the 1974 original.
[2] M.F. Atiyah and D.W. Anderson. K-Theory; Lectures by M.F. Atiyah. Notes by D.W. An-
derson, Fall, 1964. 1967.
[3] D. Baraglia. Obstructions to smooth group actions on 4-manifolds from families Seiberg-
Witten theory. Adv. Math., 354:106730, 32, 2019.
[4] D. Baraglia. Constraints on families of smooth 4-manifolds from Bauer-Furuta invariants.
Algebr. Geom. Topol., 21(1):317–349, 2021.
[5] D. Baraglia and H. Konno. A gluing formula for families Seiberg-Witten invariants. Geom.
Topol., 24(3):1381–1456, 2020.
[6] D. Baraglia and H. Konno. On the Bauer-Furuta and Seiberg-Witten invariants of families
of 4-manifolds. J. Topol., 15(2):505–586, 2022.
[7] S. Bauer. Refined Seiberg-Witten invariants. 3:1–46, 2004.
[8] S. Bauer. A stable cohomotopy refinement of Seiberg-Witten invariants. II. Invent. Math.,
155(1):21–40, 2004.
[9] S. Bauer and M. Furuta. A stable cohomotopy refinement of Seiberg-Witten invariants. I.
Invent. Math., 155(1):1–19, 2004.
[10] M.S. Berger. Nonlinearity and Functional Analysis: Lectures on Nonlinear Problems in
Mathematical Analysis. Pure and Applied Mathematics. Elsevier Science, 1977.
54 JOSHUA TOMLIN

[11] H. Brezis. Functional analysis, Sobolev spaces and partial differential equations. Universitext.
Springer, New York, 2011.
[12] E.H Brown-Jr. Cohomology theories. Ann. of Math. (2), 75:467–484, 1962.
[13] N. Doll, H. Schulz-Baldes, and N. Waterstraat. Spectral flow—a functional analytic and
index-theoretic approach, volume 94 of De Gruyter Studies in Mathematics. De Gruyter,
Berlin, [2023] ©2023.
[14] S.K. Donaldson. The Seiberg-Witten equations and 4-manifold topology. Bull. Amer. Math.
Soc. (N.S.), 33(1):45–70, 1996.
[15] S.K. Donaldson. Floer homology groups in Yang-Mills theory, volume 147 of Cambridge
Tracts in Mathematics. Cambridge University Press, Cambridge, 2002. With the assistance
of M. Furuta and D. Kotschick.
[16] H. Freudenthal. Über die Klassen der Sphärenabbildungen I. Große Dimensionen. Compositio
Math., 5:299–314, 1938.
[17] O. Garcia-Prada and R.O. Wells. Differential Analysis on Complex Manifolds. Graduate
Texts in Mathematics. Springer New York, 2007.
[18] P.B. Gilkey. Invariance theory: the heat equation and the Atiyah-Singer index theorem, vol-
ume 16. CRC press, 2018.
[19] J.P.C. Greenlees and J.P. May. Equivariant stable homotopy theory. In Handbook of algebraic
topology, pages 277–323. North-Holland, Amsterdam, 1995.
[20] N.H. Kuiper. The homotopy type of the unitary group of Hilbert space. Topology, 3:19–30,
1965.
[21] T-J Li and A-K Liu. Family Seiberg-Witten invariants and wall crossing formulas. Comm.
Anal. Geom., 9(4):777–823, 2001.
[22] A-K Liu. Family blowup formula, admissible graphs and the enumeration of singular curves.
I. J. Differential Geom., 56(3):381–579, 2000.
[23] N. Nakamura. The Seiberg-Witten equations for families and diffeomorphisms of 4-manifolds.
Asian J. Math., 7(1):133–138, 2003.
[24] D. Ruberman. An obstruction to smooth isotopy in dimension 4. Math. Res. Lett., 5(6):743–
758, 1998.
[25] D. Ruberman. Positive scalar curvature, diffeomorphisms and the Seiberg-Witten invariants.
Geom. Topol., 5:895–924, 2001.
[26] D. Salamon. Spin Geometry and Sieberg-Witten Invariants. ETH, 2000.
[27] M. Szymik. Characteristic cohomotopy classes for families of 4-manifolds. Forum Math.,
22(3):509–523, 2010.
[28] J. Tomlin. A general connected sum formula for the families Seiberg-Witten invariant. 2025
(in preparation).
[29] A.S. Švarc. On the homotopic topology of Banach spaces. Dokl. Akad. Nauk SSSR, 154:61–63,
1964.
[30] K. Wehrheim. Uhlenbeck compactness. EMS Series of Lectures in Mathematics. European
Mathematical Society (EMS), Zürich, 2004.