Motivation
Products let us compute the torus group from two circle coordinates. Gluing asks a different question: how do loops in two overlapping pieces combine, and which identifications does the overlap impose? A figure eight has two circle loops, but they do not commute. Attaching a disk along a loop makes that loop contractible. A useful theorem must distinguish these two effects: creating generators and imposing relations.
Van Kampen's theorem does this using open sets. Openness permits sufficiently small path segments and homotopy squares to stay in one piece. Path-connectedness of the intersection lets us bring transition points back to one common basepoint. The proof below constructs the resulting homomorphism and checks every choice; it does not assume that small drawings automatically preserve a loop class.
Definitions
Definition 7.1.1
(Free product)
For groups , their free product consists of alternating words of nonidentity elements from the two factors. Multiplication concatenates and reduces by multiplying adjacent letters from one factor and deleting identities. The elementary word construction gives its universal property: homomorphisms and extend uniquely to , by evaluating the letters in order. We use this group-theoretic construction; the order of evaluation is never discarded.
Definition 7.1.2
(Normal closure)
The normal closure of in a group is the smallest normal subgroup containing it. Concretely its elements are finite products of conjugates of elements of and their inverses. Quotienting by it imposes the equations , including all consequences under multiplication and conjugation.
Suppose , where are open and path-connected, and is path-connected with chosen . Write for the inclusion-induced homomorphisms from to the two groups, and for those from into . None of these maps is assumed injective.
We concatenate paths from left to right, as in the preceding chapters. A based homotopy of loops fixes both endpoints throughout. Finite interval subdivisions and square grids use the Lebesgue-number lemma on their compact parameter spaces, not a metric on .
Theorem/proposition
Theorem 7.1.3
(Van Kampen: the universal property)
Under the hypotheses above, if homomorphisms and satisfy , there is exactly one homomorphism with and .
Equivalently, the homomorphism induced by the inclusions is an isomorphism
Factor labels distinguish the two copies in the numerator. The overlap imposes equality of their images, rather than requiring either image map to be injective.
Theorem 7.1.4
(Attaching a disk imposes one normal relation)
Let be a connected finite graph and attach a closed disk by a continuous boundary map . For the resulting adjunction space , the inclusion induces
with basepoints connected along a radial path if necessary. The subgroup is a normal closure, not merely the cyclic subgroup generated by . The finite graph and disk have their usual Hausdorff gluing topology.
Proof sketch or proof idea
Proof
(A well-defined value for each loop)
We first construct the homomorphism in the universal statement. Subdivide a loop so each closed segment lies in or , and label the segment by a piece containing it. Such a subdivision exists because the inverse images of form an open cover of . At a cut where the labels differ, the cut point lies in ; choose a path from to it inside . Where both labels are the same piece, choose inside that piece. At the initial and final point take constant paths. Then segment , labeled , determines a based loop
Assign the product in the displayed order.
This value is independent of the connectors. If a connector changes where adjacent labels agree, both affected factors lie in one group and the inserted path followed by its reverse cancels there. Where the labels differ, the change gives a loop in ; the agreement hypothesis makes its two values identical in , so the neighboring inserted factors again cancel. Initial and final connectors remain constant.
Inserting a subdivision point into one labeled segment changes nothing: choose its connector in that piece and use its homomorphism to cancel the extra reverse path. To change the label of a segment lying in both pieces, first choose its two endpoint connectors in , which is allowed by the connector-independence just proved. The resulting based segment loop lies in , so the two assigned values agree. Any two labeled subdivisions have a common refinement. On each refined interval, different labels mean its image lies in both pieces. Refinement and this label-change argument prove complete independence of the subdivision.
Proof
(Homotopy invariance, the homomorphism and uniqueness)
We must still prove invariance under a based homotopy . Take a sufficiently fine grid so that the image of each closed cell lies in one piece and label each cell accordingly. At a vertex incident to both labels, its image lies in ; choose a connector there in . At a vertex incident only to one label, choose its connector in that piece. On the left and right boundary, whose image is , choose constant connectors.
Assign an oriented grid edge the value of its connector-based loop, evaluated in either adjacent cell's group. This is unambiguous: if the labels differ, the entire edge and both connectors lie in . Reversing an edge inverts its value. Around each cell, the four edge values multiply to one in boundary order. Indeed the intermediate connectors cancel inside its one piece, and the remaining loop extends over that cell by , so is null-homotopic there.
These cell equations say that traveling along the bottom and then the right side of a cell has the same value as traveling along its left and then its top side. Adjacent swaps of these two-step routes compare all monotone lattice paths from the lower-left to the upper-right corner. Thus the value along the whole bottom and right boundary equals that along the whole left and top boundary. The two side boundaries have value one, giving equality of the bottom and top loop values. This proves homotopy invariance without commuting any factors. An informal instruction to multiply all cell boundaries and cancel interior edges would miss their noncommutative order.
The assignment therefore defines on loop classes. Concatenating labeled subdivisions proves it is a homomorphism. A loop wholly in one piece can use a single segment, so extends the prescribed maps. For uniqueness, every loop class in equals the product of the connector-based segment classes: the successive reverse connectors cancel up to based homotopy. Hence any extension has the product value already defined.
Proof X-ray
Where each hypothesis of van Kampen is used
Openness of and is used only through the Lebesgue-number lemma: it supplies subdivisions of and grids on whose pieces each lie in one set. Path-connectedness of supplies connectors at transition points and at vertices where the two labels meet, and that is what makes the two images of an overlap loop comparable. The agreement is used exactly when a connector or a label changes at such a point. Path-connectedness of and is needed only for connectors at points where a single label occurs.
The key move is the cell equation combined with swaps of monotone lattice paths. It carries the value along the bottom and right sides to the value along the left and top sides without commuting any two factors. The argument closes because the vertical sides of a based homotopy have value one. If is disconnected, connectors may not exist: in the cover of a circle by two arcs, a transition point in the other component of the overlap cannot be joined to inside , and this is the exact step that breaks.
Proof
(The quotient presentation)
To obtain the quotient formula, form the displayed quotient of the free product. Its two factor maps agree on by construction, so the universal statement gives . Conversely the inclusions into satisfy the same relations, giving . Their composites fix each factor and, by the uniqueness just proved, all loop classes. They are inverse isomorphisms.
Proof
(Attaching a disk through van Kampen)
For the disk-attachment theorem, use and let be a small open concentric disk in the attached face. The punctured face in retracts radially onto its boundary, fixing , so deformation retracts onto . This homotopy is continuous in the adjunction topology: the radial formula fixes the attached boundary and descends from the graph and disk. The finite compact Hausdorff gluing also ensures that the quotient map times is a quotient map, so joint continuity is retained. The intersection is a punctured disk, with fundamental group , and is contractible. Under the radial retraction, its positive generator becomes the attaching loop , up to the chosen basepoint path. Van Kampen therefore kills exactly its normal closure. A different basepoint path conjugates the loop and leaves that normal closure unchanged.
Worked examples
Worked example 7.1.5
(Higher-dimensional spheres are simply connected)
Cover , , by the complements of its north and south poles. For a point with the north pole removed, stereographic projection is , with inverse . Direct substitution verifies both compositions; the south-pole formula is analogous. Thus each piece is homeomorphic to and contractible. The intersection is homeomorphic to and is path-connected: every sphere of positive dimension is path-connected, by circular arcs, using an intermediate point for antipodal endpoints. Both factor groups are trivial, so the theorem gives . We did not need to know the intersection's fundamental group. For this same intersection has two components, so the hypothesis fails and no contradiction with arises.
Worked example 7.1.6
(A figure eight has a free, noncommutative group)
Let be two circles joined at one point. Take to contain the first whole circle and a small open arc of the second around the joining point; take with the roles reversed. These are open in the graph. Each retracts onto its whole circle, while the intersection consists of small arms meeting at the joining point and is contractible. Thus , the free group on . The words and are different reduced alternating words; the commutator is nontrivial. Two visible circular routes do not automatically produce two commuting coordinates.
Worked example 7.1.7
(The torus relation fills a commutator)
A square with opposite sides identified gives the torus. Its edge graph is a figure eight, and its interior is an attached disk. Traversing the oriented boundary follows . The attachment theorem gives
For the second isomorphism, the relation lets all letters move past all letters, so every class has form . Sending it to is a homomorphism, and the map is its inverse. This agrees with the earlier product calculation but explains precisely which formerly nontrivial loop the two-dimensional face removes.
Worked example 7.1.8
(The projective-plane disk model)
In the disk model of the real projective plane, antipodal boundary points are identified. The boundary quotient is a circle, and the map from the original boundary to that circle has degree two. One can identify it explicitly with on . Attaching the interior disk therefore gives
The single generator survives but two traversals become trivial. This calculation uses the disk model; a separate identification with the space of lines through the origin is not required for the argument.
Common mistakes
Common mistake
(A piece need not inject into the union)
The theorem does not say that each inclusion of a piece induces an injection. A disk attached to a circle can kill that circle's generator. Also, killing a loop means quotienting by its normal closure; conjugating a null-homotopic loop by another path must still give a null-homotopic loop.
Common mistake
(Closed pieces and disconnected overlaps)
A decomposition into closed pieces is not automatically an eligible open cover. Use collars or suitable open neighborhoods and justify their retractions. The graph-and-face constructions above provide them explicitly. Finally, the usual based theorem needs a path-connected overlap; disconnected-overlap versions require a fundamental groupoid or additional connecting data and are not proved here.
Counterexample mode
Two simply connected open pieces need not give a simply connected union
Claim: if with and open and simply connected, then is simply connected. On the circle take and . Each is homeomorphic to an open interval, through a choice of argument in an open interval of length , so each is contractible and in particular simply connected. Yet by the earlier lifting computation. The overlap consists of the open upper and lower semicircles, two components, so the hypothesis that is path-connected fails. With that hypothesis restored, the statement is true, as the first exercise below shows.
Summary
Subdivision expresses loops through the two pieces, agreement on overlap loops removes connector choices, and a labeled grid verifies homotopy invariance in the correct multiplication order. The resulting quotient presentation turns gluing into generators and relations. Attaching a disk imposes its boundary word, giving computations that distinguish a figure eight, a torus and a projective plane.
Exercises
Quick check
(If are open in , simply connected, and their intersection is nonempty and path-connected, prove is simply connected. Must the intersection itself be simply connected?)
Solution · Answer
The union is path-connected because paths in either piece can be joined through a point of the overlap. Its group is a quotient of , so it is trivial. The overlap need not be simply connected: the two-puncture cover of has intersection homeomorphic to , whose group is by the product theorem.
Quick check
(Attach a disk to a circle by , with . Compute the fundamental group, including .)
Solution · Answer
The circle generator is sent to , so the quotient is when , with order . For the attaching loop is null-homotopic, the normal closure is trivial and the group stays . Negative degree reverses the relation's orientation without changing its generated subgroup.
Quick check
(Explain why the circle cover by complements of two different points cannot be used to prove its fundamental group trivial.)
Solution · Answer
Each complement is an open interval, but their intersection has two components. A transition point in the other component cannot be connected to the chosen basepoint inside the overlap. The connector construction fails there, so the theorem has not supplied a group computation.
Quick check
(Before the torus disk is attached, is trivial in the edge graph? After attachment, why are all its conjugates trivial too?)
Solution · Answer
It is a reduced nonempty word in , hence nontrivial. The disk makes it null-homotopic. Conjugating a based null-homotopy by a fixed loop gives another null-homotopy, or algebraically the kernel of the inclusion-induced homomorphism is normal. Both explanations show why merely removing one isolated word is insufficient.
Quick check
(Use the sphere and projective-plane computations to show that and the projective-plane disk model are not homotopy equivalent.)
Solution · Answer
Their fundamental groups are respectively trivial and of order two. A homotopy equivalence would induce an isomorphism by the preceding chapter, which is impossible. This distinguishes them even at the weaker level of homotopy type, not only up to homeomorphism.