Manifolds with Boundary

Half-Space and the Need for Boundary

Every manifold we have constructed so far has been locally modeled on \(\mathbb{R}^n\), with no edges and no rim. That list runs from spheres, projective spaces and tori to the Grassmannians and the topological manifolds of the opening pages. Near every point, such a manifold looks like all of Euclidean space. Yet many of the most natural geometric objects do have edges. A closed interval \([a, b] \subseteq \mathbb{R}\) has its two endpoints. A closed disk in the plane has its bounding circle. A closed hemisphere of the sphere \(\mathbb{S}^n\) terminates at an equatorial \((n-1)\)-sphere. At an endpoint of \([a, b]\), no neighborhood inside the interval is homeomorphic to an open interval of \(\mathbb{R}\). Every such neighborhood is a half-open interval, with the endpoint sitting on its edge. These spaces are not manifolds in the sense developed so far, but they are manifolds in a slightly enlarged sense, one that admits a boundary.

We develop this enlarged notion now, as the closing topic of the topological and smooth foundations, because the boundary is precisely where the most important integral identities of geometry live. The classical theorems of vector calculus all relate an integral over a region to an integral over the edge of that region. Green's theorem in the plane, the divergence theorem of Gauss and the curl theorem of Stokes are the familiar instances. Their common generalization, the modern Stokes theorem, takes the form \[ \int_{M} d\omega = \int_{\partial M} \omega, \] in which the left-hand side integrates over a manifold and the right-hand side integrates over its boundary \(\partial M\). For this identity even to be stated, one needs a class of spaces that carry both a smooth structure in their interior and a well-defined boundary on which the second integral makes sense. The objects of this page are exactly those spaces. We do not develop integration here, since that requires the machinery of differential forms, taken up later in the series. The present page builds the stage on which it will be performed.

The need is not confined to classical analysis. Wherever a domain is constrained, a boundary appears. Boundary-value problems for partial differential equations are posed on regions whose edges carry the prescribed data, as in the Dirichlet and Neumann problems and in the free-boundary problems of the calculus of variations. The integration theory we developed on the measure-theoretic side has its natural geometric counterpart on manifolds with boundary. The same structure recurs in machine learning and robotics. A configuration space subject to constraints, such as joint angle limits on an articulated arm or obstacle-avoidance regions in a planning problem, is a manifold with boundary rather than a boundaryless one. The variational autoencoder page's robotic manipulation demonstration parametrizes a three-degree-of-freedom arm by a triple of joint angles, and it implicitly works over such a constrained space. The admissible joint angles fill a product of closed intervals \([\theta_{\min}, \theta_{\max}]^3\). Strictly, a product of closed intervals has corners as well as faces, which places it in the slightly larger class of manifolds with corners. The manifold series treats that class much later. For now we record only that boundaries are the rule, not the exception, once domains are constrained.

The Upper Half-Space

Just as a boundaryless \(n\)-manifold is modeled locally on \(\mathbb{R}^n\), a manifold with boundary is modeled locally on a half-space: the points of \(\mathbb{R}^n\) lying on one side of a hyperplane, together with the hyperplane itself. Fixing the last coordinate as the one cut off, we obtain the standard local model.

The half-space inherits two distinguished subsets from the position of a point relative to the bounding hyperplane. For \(n \ge 1\), the interior of the half-space consists of the points lying strictly off the hyperplane, \[ \operatorname{Int} \mathbb{H}^n = \bigl\{ (x^1, \ldots, x^n) : x^n \gt 0 \bigr\}, \] and the boundary of the half-space consists of the points lying on it, \[ \partial \mathbb{H}^n = \bigl\{ (x^1, \ldots, x^n) : x^n = 0 \bigr\}. \] The interior \(\operatorname{Int} \mathbb{H}^n\) is an open subset of \(\mathbb{R}^n\), and it is exactly the topological interior of \(\mathbb{H}^n\) as a subset of \(\mathbb{R}^n\). The boundary \(\partial \mathbb{H}^n\) is the hyperplane \(\{x^n = 0\}\). Discarding the final coordinate identifies it with \(\mathbb{R}^{n-1}\), so it is an \((n-1)\)-dimensional Euclidean space sitting inside \(\mathbb{R}^n\). The degenerate case \(n = 0\) follows the convention above. Here \(\mathbb{H}^0 = \{0\}\) is a single point with \(\operatorname{Int} \mathbb{H}^0 = \{0\}\) and \(\partial \mathbb{H}^0 = \varnothing\).

The symbols \(\operatorname{Int}\) and \(\partial\) call for a word of caution, and we return to it once manifolds enter the picture. As written here, \(\operatorname{Int} \mathbb{H}^n\) and \(\partial \mathbb{H}^n\) are defined by an inequality on the coordinate \(x^n\). They are subsets of \(\mathbb{R}^n\) singled out by their position relative to the hyperplane, and this is the description we use as a local model. A separate matter remains, and it is the central subtlety of the next section. The interior and boundary of an abstract manifold are defined intrinsically, without reference to any ambient space, and we must ask whether they can be detected chart by chart through this model. That they can is the content of an invariance theorem. Until it is established we keep the half-space picture and the manifold picture notationally distinct.

Topological Manifolds with Boundary

With the local model in hand, the definition of a manifold with boundary is obtained from the boundaryless definition by a single change. A neighborhood of a point is now permitted to look like an open subset of the half-space \(\mathbb{H}^n\), not only an open subset of \(\mathbb{R}^n\). The Hausdorff and second-countability requirements are unchanged.

The phrase "with boundary" is a permission, not a requirement. A point whose neighborhood happens to be homeomorphic to an open subset of \(\mathbb{R}^n\) is admitted exactly as before. Every boundaryless \(n\)-manifold is therefore a manifold with boundary in which the boundary turns out to be empty. We will be able to make that statement precise once the boundary is defined. The definition deliberately enlarges the class of admissible spaces instead of replacing one class with another.

Charts and Their Two Types

The chart vocabulary carries over verbatim, with the codomain now allowed to lie in either model space.

Charts split into two kinds according to where their image sits relative to the bounding hyperplane of the model. The distinction is the engine of the entire theory of the boundary.

A single subtlety deserves emphasis. By definition of the subspace topology, an open subset of \(\mathbb{H}^n\) that misses the hyperplane \(\partial \mathbb{H}^n\) is an open subset of \(\mathbb{R}^n\) as well. A chart with such an image is therefore an interior chart, even when the half-space was the model we had in mind when writing it down. The classification into interior and boundary charts is exclusive and exhaustive. A chart image that lies in \(\mathbb{H}^n\) and meets \(\partial \mathbb{H}^n\) is never open in \(\mathbb{R}^n\), and every chart image is of one kind or the other. It is precisely the points carried onto \(\partial \mathbb{H}^n\) by some chart that we are about to single out.

The codomain of a boundary chart often has a standard convenient shape, the half-space analogue of a coordinate ball.

Interior and Boundary Points

We can now define, intrinsically on \(M\), the two classes of points that the half-space model distinguishes locally.

The definitions are phrased existentially, in terms of lying in the domain of some chart of the stated kind, and this is exactly where a difficulty hides. A given point \(p\) may lie in the domains of many charts at once. Nothing in the definitions, as stated, forbids one chart from presenting \(p\) as an interior point while another presents the same \(p\) as a boundary point. Were that to happen, the labels "interior" and "boundary" would not be properties of the point at all, but artifacts of the chart chosen to view it, and the sets \(\operatorname{Int} M\) and \(\partial M\) would be ill-defined. A theorem rules this out and shows that the two classes are genuinely intrinsic.

We state this result without proof. The obstruction is genuine. At the purely topological level, the assertion that an open subset of \(\mathbb{R}^n\) cannot be homeomorphic to a neighborhood of a point on the edge of \(\mathbb{H}^n\) is a statement about the local topology of Euclidean space, and it lies beyond what point-set methods can reach. The standard proof uses the machinery of singular homology, and in particular the local homology groups that distinguish a point of \(\operatorname{Int} \mathbb{H}^n\) from a point of \(\partial \mathbb{H}^n\). That machinery is outside the scope of the manifold series at this stage. Remarkably, once a smooth structure is available, the same invariance can be proved by elementary calculus with no algebraic topology at all. That smooth counterpart is the technical climax of this page. One proof requires homology and the other requires only the chain rule, and the contrast between them is one of the instructive lessons of the subject.

Two Meanings of the Symbol \(\partial\)

The notation \(\partial M\) collides with another standard use of the same symbol, and the two must be kept apart. In point-set topology, when a space \(X\) sits inside an ambient space, the topological boundary of \(X\) is the set of points in the closure of \(X\) that are not in its topological interior. These are the points that touch both \(X\) and its complement. The manifold boundary just defined is an intrinsic notion. It depends only on \(M\) as an abstract manifold and not on any embedding into a larger space. The two need not coincide, and conflating them is a common source of error.

The closed half-space and the closed unit ball illustrate the agreement, and an interval placed in the plane illustrates the disagreement. Regarded as a manifold with boundary, \(\mathbb{H}^n\) has manifold boundary \(\partial \mathbb{H}^n = \{x^n = 0\}\). Regarded as a subset of \(\mathbb{R}^n\), its topological boundary is the same hyperplane, so here the two notions agree. The closed unit ball \(\overline{\mathbb{B}}^n \subseteq \mathbb{R}^n\) is a manifold with boundary whose manifold boundary is the unit sphere \(\mathbb{S}^{n-1}\), and its topological boundary as a subset of \(\mathbb{R}^n\) is also \(\mathbb{S}^{n-1}\). The two agree again. But the agreement is not automatic. It requires the manifold to sit in the ambient space as a closed subset of the same dimension, and either condition can fail. The half-open interval \([0,1)\) inside \(\mathbb{R}\) has manifold boundary \(\{0\}\) and topological boundary \(\{0, 1\}\), so a failure of closedness alone already separates the two. The open interval \((0,1)\), viewed as a \(1\)-manifold, has empty manifold boundary. Viewed as a subset of \(\mathbb{R}^2\) along the \(x\)-axis, it has empty topological interior and closure \([0,1] \times \{0\}\), so every one of its points is a topological boundary point. The manifold boundary is the notion this page concerns, and the symbol \(\partial\) always refers to it here.

Closed and Open Manifolds

Two further pieces of terminology, both concerning boundaryless manifolds, are conventional in the literature and worth recording before they appear unannounced. Both invite confusion with the point-set vocabulary of open and closed sets, and neither has anything to do with it.

The word "closed" here records compactness, not the topological property of being a closed set, and neither a closed manifold nor an open manifold has any boundary at all. The sphere \(\mathbb{S}^n\) is a closed manifold. The Euclidean space \(\mathbb{R}^n\) and the open ball are open manifolds. A closed interval \([0,1]\), by contrast, is neither. It is compact but does have a boundary, so it is a compact manifold with boundary and falls under none of this terminology. The terms are useful precisely because the property of having no boundary, combined with a compactness alternative, organizes much of the global theory of manifolds.

Topological Properties of Manifolds with Boundary

Granting the topological invariance of the boundary, the interior and the boundary of a manifold are well-defined sets, and we can ask what kind of spaces they are in their own right. The answer is the cleanest possible one. The interior is a boundaryless manifold of the same dimension, and the boundary is a boundaryless manifold of one dimension lower. The statement also records when the boundary is empty and what happens in the degenerate dimension zero.

The relation \(\partial M = M \setminus \operatorname{Int} M\) used in part (2) is exactly the disjoint-union decomposition supplied by the invariance theorem. Without that theorem, the complement of the interior would not be guaranteed to consist of boundary points, and the closedness of \(\partial M\) would not follow. This is the one place in the elementary theory where the deferred topological invariance is genuinely needed.

Carrying the Topological Properties Across

The properties established earlier for boundaryless manifolds all survive the introduction of a boundary. They include the existence of well-behaved bases, local compactness, paracompactness and connectivity. The reason is structural. Each of those proofs rested on the local model being Euclidean, and the half-space \(\mathbb{H}^n\) is just as well-behaved a local model as \(\mathbb{R}^n\). Wherever a coordinate ball was used, a coordinate half-ball serves the same purpose at boundary points, and the arguments are otherwise unchanged. We collect the results for the record.

Each statement is the boundary-aware version of a result proved for manifolds without boundary, and the proofs require only the substitution of half-balls for balls at boundary points. Statement (1) refines the earlier basis of precompact coordinate balls by allowing half-balls in the cover. At boundary charts the single-chart construction in \(\mathbb{R}^n\) is replaced by the identical construction inside \(\mathbb{H}^n\), and the countable assembly is unchanged. Statement (2) follows from (1) exactly as before, since local compactness needs only one precompact neighborhood at each point, and half-balls are precompact. Statement (3), paracompactness, is deduced from the basis of (1) and local compactness by the same compact-exhaustion argument, which never referred to the absence of boundary. Statements (4) and (5) are the connectivity properties. Coordinate half-balls, like coordinate balls, are path-connected, since a half-ball is convex and path-connectedness transfers under the chart homeomorphism. The manifold therefore again has a basis of path-connected open sets, and the component count is governed by second-countability just as before.

Statement (6), the countability of the fundamental group, is of a different character from the rest. It is not a point-set consequence of the basis but a fact about the algebraic topology of \(M\). Its proof for manifolds without boundary cuts a loop into finitely many arcs, each confined to a simply connected coordinate ball, and reads off the loop's class from the countable record of how consecutive arcs meet. Nothing in that argument uses the absence of a boundary. Coordinate half-balls are convex, hence simply connected, and the cover by half-balls is countable piece by piece exactly as the cover by balls. The same proof therefore applies verbatim, with half-balls substituted for balls at boundary charts, and \(\pi_1(M)\) is countable for a manifold with boundary as well.

Smooth Structures on Manifolds with Boundary

Everything so far has been topological. To do calculus on a manifold with boundary we repeat the construction that produced smooth manifolds from topological ones. We single out a maximal atlas whose transition maps are diffeomorphisms. There is exactly one new ingredient. Transition maps between boundary charts are maps between open subsets of the half-space \(\mathbb{H}^n\), not of \(\mathbb{R}^n\), and the Euclidean definition of smoothness we already have applies only to maps on open subsets of \(\mathbb{R}^n\). A subset of \(\mathbb{H}^n\) that meets the boundary hyperplane is not open in \(\mathbb{R}^n\), so before anything else we must say what it means for a map defined on such a set to be smooth.

Smoothness on Subsets of the Half-Space

The difficulty is that partial derivatives at a boundary point would have to be one-sided in the last variable, and a one-sided derivative is a weaker object than a genuine derivative. It sees only the values of the function on one side. The standard remedy declares a map smooth on a half-space subset exactly when it is the restriction of an honestly smooth map defined on a full Euclidean neighborhood.

At points of \(U\) lying in \(\operatorname{Int} \mathbb{H}^n\) the condition is automatic and adds nothing. Such a point already has a full neighborhood inside \(U\), and \(F\) is its own extension there. The definition has content only at boundary points of \(U\), where it demands that \(F\) extend smoothly across the hyperplane to one side. A commonly given alternative description is that \(F\) is smooth on \(U\) precisely when all partial derivatives of all orders exist throughout \(U\) and extend continuously up to the boundary. The one-sided derivatives at boundary points then match the limits of the interior derivatives. That the two descriptions are equivalent is a nontrivial extension theorem, which we neither prove nor use. We take the extension form as primary because it makes the chain rule available without one-sided bookkeeping.

That the extension requirement is a genuine restriction, not a formality satisfied by every continuous-looking function, is shown by a single example. Consider on the half-disk \(\mathbb{B}^2 \cap \mathbb{H}^2\) the function \[ g(x, y) = \sqrt{y}, \quad y \ge 0. \] It is continuous on its domain and infinitely differentiable at every interior point. But it is not smooth in the sense just defined: its partial derivative with respect to \(y\), \[ \frac{\partial g}{\partial y} = \frac{1}{2\sqrt{y}}, \] diverges as \(y \to 0^+\), so no continuous extension of the \(y\)-derivative to the boundary exists, and a fortiori no smooth extension of \(g\) to an open neighborhood of any boundary point in \(\mathbb{R}^2\) can exist. Such an extension would have a finite \(y\)-derivative at the boundary, which the blow-up forbids. Smoothness up to the boundary is therefore a real constraint, exactly the one needed to keep the chain rule and the Jacobian well-behaved at the edge.

Smooth Manifolds with Boundary

With smoothness defined on half-space subsets, the compatibility of charts is defined exactly as before. A transition map between two charts of a manifold with boundary is a homeomorphism between open subsets of model spaces, each of which is either \(\mathbb{R}^n\) or \(\mathbb{H}^n\). The half-space notion of smoothness applies to it in every case, so we may ask whether it is a diffeomorphism in that sense, smooth with smooth inverse.

This is the smooth structure of the boundaryless theory with two adjustments and no others. A chart may now map into \(\mathbb{H}^n\) as well as \(\mathbb{R}^n\), and the smooth-compatibility condition is read with the half-space notion of smoothness in place of the purely Euclidean one. Suppose no chart meets the boundary, so that every transition is a map between open subsets of \(\mathbb{R}^n\). The two readings then coincide and the definition reduces exactly to the boundaryless one. A smooth manifold without boundary is thus the special case of a smooth manifold with boundary in which \(\partial M = \varnothing\).

Smooth Charts and Half-Balls

The shape-based refinements of a chart specialize to the boundary setting in the obvious way, mirroring the smooth coordinate balls of the boundaryless theory.

Regular coordinate balls came from a packaging device: a ball sitting with compact closure inside a strictly larger smooth chart. That packaging has a boundary version, in which the larger chart is itself a half-ball.

The basis result of the boundaryless theory persists as well.

The proof is the smooth refinement of the topological basis result of the previous section, adapted exactly as the boundaryless basis-of-regular-balls argument was adapted from its topological predecessor. Every chart in the construction is required to lie in the smooth structure, and at boundary points the half-ball model replaces the ball. We will not rewrite that argument here, since it differs from the one already given only by this substitution.

Open Submanifolds with Boundary

One further construction of the boundaryless theory carries over, and it is the one the pages ahead will lean on most heavily. Passing to an open subset costs nothing, and it costs nothing at the boundary either.

That \(U\) is a topological \(n\)-manifold with boundary is immediate. Hausdorffness and second countability are inherited by subspaces, and a point of \(U\) has, inside \(U\), a neighborhood modeled on an open subset of \(\mathbb{R}^n\) or of \(\mathbb{H}^n\), namely the intersection with \(U\) of any chart domain around it.

The rest of the verification is the one already carried out in the boundaryless case. Every point of \(U\) lies in the domain of some smooth chart \((W, \varphi)\) for \(M\), and the restriction of that chart to \(W \cap U\) is smoothly compatible with every chart of \(\mathcal{A}\), since each restricted transition map is an original transition map restricted to an open subdomain. By maximality of \(\mathcal{A}\) the restricted chart belongs to \(\mathcal{A}\), and since \(W \cap U \subseteq U\) it belongs to \(\mathcal{A}_U\). So the charts of \(\mathcal{A}_U\) cover \(U\), and any two of them are smoothly compatible because they already lie in \(\mathcal{A}\). The single new feature is that a restricted chart may be a boundary chart, whose transitions are read in the half-space sense; that reading is stable under restriction to an open subdomain exactly as the Euclidean one is, since a smooth extension across the hyperplane restricts to a smooth extension.

The identity \(\partial U = U \cap \partial M\) needs nothing beyond the description of \(\mathcal{A}_U\). If \(p \in U\) is a boundary point of \(M\), some chart of \(\mathcal{A}\) carries \(p\) into \(\partial \mathbb{H}^n\), and restricting that chart to its intersection with \(U\) yields a chart of \(\mathcal{A}_U\) that does the same, so \(p\) is a boundary point of \(U\). Conversely every chart of \(\mathcal{A}_U\) is a chart of \(\mathcal{A}\), so a boundary point of \(U\) is a boundary point of \(M\). Shrinking a domain therefore neither manufactures nor destroys boundary, and this is what makes open subsets a usable supply of smaller manifolds with boundary: an argument that is local by nature may be run inside a single chart domain without leaving the category.

Product and Smooth Invariance

Two results close the theory. The first records how the boundary behaves under products. The second is the smooth counterpart of the invariance theorem stated earlier, and it is the technical high point of the page. Where the topological invariance required the machinery of algebraic topology, the smooth invariance falls to the chain rule and a single fact about smooth maps with invertible derivative.

Products

Taking products of manifolds with boundary requires care, and the care is exactly the issue of corners. The product of two closed half-lines is a closed quarter-plane, whose edge is not a smooth hypersurface but a pair of rays meeting at a right angle. To stay within the category of manifolds with boundary, one must restrict products so that at most one factor has a boundary.

The single-boundary-factor restriction is essential, not a convenience. The product of two manifolds with boundary has, along the locus where both boundaries meet, points whose neighborhoods are modeled on a product of two half-spaces \(\mathbb{H}^n \times \mathbb{H}^m\) rather than on a single half-space. The obstruction is not topological. The squaring map \(z \mapsto z^2\) of the complex plane carries the closed quarter-plane homeomorphically onto the closed upper half-plane, so a corner can be flattened away by a homeomorphism. That map has vanishing derivative at the origin, and no diffeomorphism flattens a corner. We do not prove that last fact here, since it belongs to the theory of the objects that accommodate corners, the strictly larger category of manifolds with corners, whose local model is the orthant \(\{x : x^1 \ge 0, \ldots, x^k \ge 0\}\). Their theory is developed much later in the manifold series, and it is the proper setting for products such as the cube \([0,1]^n\) or the constrained joint-angle spaces noted at the start of this page.

Smooth Invariance of the Boundary

We come to the theorem that gives the boundary of a smooth manifold its meaning. The interior and boundary points were defined existentially, a point counting as a boundary point if some boundary chart sends it to \(\partial \mathbb{H}^n\). The topological invariance theorem guarantees that the labels are well-defined, and it was stated without proof because its topological version needs singular homology. For smooth manifolds the same conclusion is available by elementary means, and we prove it in full.

The contrast with the topological invariance theorem stated earlier is worth drawing explicitly, because it illustrates a recurring theme. A smooth structure, once present, makes available tools far more elementary than those the bare topology demands. The topological statement says that no homeomorphism can identify a neighborhood of an interior point with a neighborhood of a boundary point. It resists point-set methods and is proved through the local homology groups of singular homology theory. The smooth statement just proved needs none of that. The transition map is differentiable, its derivative is forced to be invertible by the chain rule, and invertibility of the derivative is exactly the hypothesis under which a smooth map is open. Two routes of utterly different cost thus reach the same geometric fact, that interior and boundary are intrinsic, and the smooth route is the cheap one. This is one of the first concrete dividends of the smooth structure built over the course of these pages, and it is the foundation on which the boundary integral of the next stage of the theory will rest.