Manifolds with Boundary

Half-Space and the Need for Boundary

Every manifold we have constructed so far — spheres, projective spaces, tori, the Grassmannians, and the topological manifolds of the opening pages — has been locally modeled on \(\mathbb{R}^n\), with no edges and no rim. 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 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.

The reason to develop this enlarged notion now, as the closing topic of the topological and smooth foundations, is that the boundary is precisely where the most important integral identities of geometry live. The classical theorems of vector calculus — Green's theorem in the plane, the divergence theorem of Gauss, the curl theorem of Stokes — all relate an integral over a region to an integral over the edge of that region. 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 — that requires the machinery of differential forms, taken up later in the series — but 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 — the Dirichlet and Neumann problems, free-boundary problems in the calculus of variations — are posed on regions whose edges carry the prescribed data, and the integration theory we developed on the measure-theoretic side has its natural geometric counterpart on manifolds with boundary. In machine learning and robotics the same structure recurs: a configuration space subject to constraints — joint angle limits on an articulated arm, 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, in which a three-degree-of-freedom arm is parametrized by a triple of joint angles, 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, placing it in the slightly larger class of manifolds with corners, which the manifold series treats 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, \[ \operatorname{Int} \mathbb{H}^n = \bigl\{ (x^1, \ldots, x^n) : x^n > 0 \bigr\}, \] 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 \(\{x^n = 0\}\); it is naturally identified with \(\mathbb{R}^{n-1}\) by discarding the final coordinate, and under this identification it is itself a copy of \(\mathbb{R}^{n-1}\), a \((n-1)\)-dimensional Euclidean space sitting inside \(\mathbb{R}^n\). The degenerate case \(n = 0\) follows the convention above: \(\mathbb{H}^0 = \{0\}\) is a single point with \(\operatorname{Int} \mathbb{H}^0 = \{0\}\) and \(\partial \mathbb{H}^0 = \varnothing\).

A point of caution about the symbols \(\operatorname{Int}\) and \(\partial\), to which we return 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. This is the description we use as a local model. It is a separate matter — and the central subtlety of the next section — whether the interior and boundary of an abstract manifold, defined intrinsically without reference to any ambient space, 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 — a statement we will be able to make precise once the boundary is defined. The deliberate effect of the definition is to enlarge the class of admissible spaces, not to replace 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. An open subset of \(\mathbb{H}^n\) that misses the hyperplane \(\partial \mathbb{H}^n\) is, by definition of the subspace topology, an open subset of \(\mathbb{R}^n\) as well; a chart with such an image is therefore both expressible as an interior chart and harmlessly regardable as a boundary chart whose image avoids the edge. The classification into "interior" and "boundary" is genuinely informative only when the image of a boundary chart actually meets \(\partial \mathbb{H}^n\). 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 — "lies 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. That this never happens — that the two classes are genuinely intrinsic — is a theorem.

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 that lies beyond what point-set methods can reach. Its standard proof uses the machinery of singular homology — specifically, the local homology groups that distinguish a point of \(\operatorname{Int} \mathbb{H}^n\) from a point of \(\partial \mathbb{H}^n\) — which 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, and the contrast between the two proofs — one requiring homology, one requiring only the chain rule — 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 — the points that touch both \(X\) and its complement. The manifold boundary just defined is an intrinsic notion, depending 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 itself illustrates the agreement and the closed unit ball 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 an accident of the ambient dimension matching the manifold dimension. 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 equals its own topological boundary, all of it. The manifold boundary is the notion this page concerns; 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 (b) 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 — the existence of well-behaved bases, local compactness, paracompactness, connectivity — all survive the introduction of a boundary. 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; 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 (a) refines the earlier basis of precompact coordinate balls by allowing half-balls in the cover; the single-chart construction in \(\mathbb{R}^n\) is replaced, at boundary charts, by the identical construction inside \(\mathbb{H}^n\), and the countable assembly is unchanged. Statement (b) follows from (a) exactly as before, since local compactness needs only one precompact neighborhood at each point, and half-balls are precompact. Statement (c), paracompactness, is deduced from the basis of (a) and local compactness by the same compact-exhaustion argument, which never referred to the absence of boundary. Statements (d) and (e) are the connectivity properties: coordinate half-balls, like coordinate balls, are path-connected (a half-ball is convex, hence path-connected, and path-connectedness transfers under the chart homeomorphism), so the manifold again has a basis of path-connected open sets, and the component count is governed by second-countability just as before. Statement (f), 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. An equivalent and frequently more practical 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 matching limits of interior derivatives. The two formulations agree; 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 \(\overline{\mathbb{B}}^2 \cap \mathbb{H}^2\) the function \[ g(x, y) = \sqrt{y}, \qquad y \ge 0. \] 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, contradicting the blow-up. 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, and we may ask whether it is a diffeomorphism — smooth with smooth inverse — in that sense.

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. When no chart meets the boundary — when every transition is a map between open subsets of \(\mathbb{R}^n\) — the two readings coincide and the definition reduces exactly to the boundaryless one, so a smooth manifold without boundary is 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.

The same packaging that produced regular coordinate balls — a ball sitting with compact closure inside a strictly larger smooth chart — has a boundary version, in which the larger chart is itself a half-ball.

Finally, the basis result of the boundaryless theory persists. Just as every smooth manifold has a countable basis of regular coordinate balls, every smooth manifold with boundary has a countable basis consisting of regular coordinate balls at interior points and regular coordinate half-balls at boundary points. 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.

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. Such a product is not homeomorphic to an open subset of any \(\mathbb{H}^{n+m}\) near the corner: it has a corner there, a feature no half-space model possesses. The objects that accommodate corners form 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 is a boundary point if some boundary chart sends it to \(\partial \mathbb{H}^n\) — and the topological invariance theorem, which guarantees the labels are well-defined, 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 — that no homeomorphism can identify a neighborhood of an interior point with a neighborhood of a boundary point — 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. The same geometric fact — interior and boundary are intrinsic — is reached by two routes of utterly different cost, 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.