Covering dimension and finite-dimensional embeddings

Try

A finite set, an interval and a square differ in a way that open-set structure can detect. Covering dimension measures how much overlap is unavoidable when an open cover is refined into smaller open pieces.

This lesson restricts its elementary dimension arguments to compact metric or explicitly stated compact Hausdorff spaces. It does not infer dimension from a drawing or from the number of coordinates used by one particular representation.

Predict

Predict whether embedding a circle in the plane means its topological dimension is two. It does not. A circle is one-dimensional, although it cannot embed into the real line.

Predict what a zero-dimensional compact space might look like. It need not be finite or discrete: the Cantor set has a basis of clopen cylinder sets. Small separated pieces can be organized at every scale without isolating individual points.

Experiment

State the convention for finite covers.

Dimension statement	Refinement requirement
dim X≤0	Finite open refinement with no point in two members
dim X≤1	Finite open refinement with no point in three members
dim X≤n	Finite open refinement of multiplicity at most n+1
dim X=n	The least such nonnegative integer for nonempty X

A refinement covers X and each of its members lies inside a member of the original cover. Multiplicity counts how many refining sets contain one point.

Understand

For a compact space with a clopen basis, refine any finite open cover by clopen neighborhoods and use compactness to choose finitely many. Form their Boolean atoms by intersecting each chosen set or its complement. The nonempty atoms are pairwise disjoint clopen sets covering X, and each lies in one of the original chosen neighborhoods. Thus the covering dimension is zero for a nonempty such space. For the usual middle-thirds Cantor set, fixing the first n ternary digits, each zero or two, selects its intersection with one of finitely many closed construction intervals. The gaps make that cylinder relatively open as well as closed. Cylinder diameters are 3^(−n), so they refine every relative neighborhood and form a clopen basis. This establishes the Cantor-set example.

For [0,1], a Lebesgue number δ for a finite open cover lets us use a sufficiently fine chain of relative open intervals. Choose grid spacing h with 6h/5<δ and radius 3h/5 around the grid points. These intervals cover, each lies in a cover member, and no point belongs to three because their centers are h apart. Hence dimension is at most one.

Dimension zero is impossible for a nontrivial connected compact Hausdorff space. Cover it by the complements of two distinct singletons. A multiplicity-one finite open refinement is a disjoint open cover; its member containing the first point misses the second and is clopen. That contradicts connectedness. Therefore the interval has dimension exactly one.

The compact-metric form of the Menger–Nöbeling theorem says that a compact metric space with covering dimension at most n embeds into R^(2n+1). The construction below proves this form using small covers, general position and Baire category. It does not assert the broader noncompact theorem or that the dimension bound is minimal.

From finite separation to a compact-space embedding

Two disjoint edges in a plane can cross even when all four vertices are different. Move one edge slightly out of the plane: its top projection still crosses, but the actual convex hulls separate. This is the finite geometric mechanism behind the target dimension.

Lift two crossing edges out of the same plane

Height of vertices C and D 00.1250.51 Weight s of B in (1−s)A+sB 00.250.50.751 Weight t of D in (1−t)C+tD 00.250.50.751

Predict whether the crossing in the top projection disappears when the actual three-dimensional edges separate. Coordinates are dimensionless. Both convex combinations use disjoint vertex sets.

A=(−1,0,0), B=(1,0,0), C=(0,−1,0), D=(0,1,0). Affine determinant: 0; the four vertices are affinely dependent.

Selected points: (0, 0, 0) and (0, 0, 0). Actual distance: 0.000000. Exact minimum over all continuous weights s,t∈[0,1]: 0.

The actual convex combinations coincide. The squared distance is (2s−1)²+(2t−1)²+h², so the minimum occurs at s=t=1/2. It is proved over the full weight square, not inferred from the control grid.

Reset

This finite example illustrates the disjoint-convex-hull step. It does not compute an embedding of an arbitrary compact space or replace the Baire proof. Controls are temporary and award no mastery.

The MIT 18.901 calendar explicitly places Euclidean embedding after Baire spaces and dimension theory. We now assemble those mechanisms. Let X be a nonempty compact metric space, let n≥0, and assume every finite open cover has a finite open refinement of multiplicity at most n+1. Set m=2n+1. An embedding means a homeomorphism onto its image, not merely an injective formula.

Complete proof: general position, dense separating maps and Baire

1. Nearby vertices in general position

Any finite list of proposed vertices in Rᵐ can be perturbed by arbitrarily small amounts so that every subset of at most m+1 vertices is affinely independent. Choose the new vertices successively. For the next vertex, avoid the affine hull of each subset of at most m earlier vertices. By the inductive independence property each such hull is a proper affine subspace, and there are only finitely many to avoid.

Here is why every ball contains a point outside those subspaces. Enlarge each proper affine subspace to a hyperplane with nonzero normal. Choose a direction v not orthogonal to any of these finitely many normals: vectors v(t)=(1,t,…,tᵐ⁻¹) suffice, since each dot product is a nonzero polynomial with finitely many roots. A line through the center of the ball in that direction meets each hyperplane in at most one point. Choose a sufficiently small line parameter avoiding those finitely many values. This proves the perturbation lemma, including m=1.

If two disjoint subsets of these vertices have at most m+1 vertices in total, their convex hulls are disjoint. Otherwise an equality of convex combinations gives a nonzero affine dependence: coefficients on one subset are nonnegative and sum to one, coefficients on the other are their negatives and sum to minus one. Their total is zero. Affine independence rules this out. Zero coefficients can be discarded.

2. Approximate any continuous map by a map separating distant pairs

Fix ε>0, a continuous f:X→Rᵐ and a desired approximation error η>0. Uniform continuity and compactness give a finite open cover whose sets have diameter less than ε and oscillation of f less than η/3. Refine it by a finite open cover V₁,…,Vᵣ of multiplicity at most n+1. Discard empty members. Each Vᵢ still has the same diameter and oscillation bounds, and choose pᵢ∈Vᵢ.

Construct continuous weights without invoking an unproved partition-of-unity theorem. Put wᵢ(x)=distance from x to X∖Vᵢ when that complement is nonempty, and wᵢ=1 when Vᵢ=X. These functions are continuous and positive exactly on Vᵢ. Their sum is positive because the sets cover X. Normalize to λᵢ=wᵢ/Σwⱼ; thus λᵢ≥0, Σλᵢ=1, and λᵢ(x)>0 only when x∈Vᵢ.

Perturb f(pᵢ) to general-position vertices vᵢ within η/3 and define g(x)=Σλᵢ(x)vᵢ. This is continuous. Every active term at x has both |vᵢ−f(pᵢ)|<η/3 and |f(pᵢ)−f(x)|<η/3, so the convex weighted estimate gives |g(x)−f(x)|<2η/3. Hence the uniform approximation error is less than η.

If d(x,y)≥ε, no Vᵢ contains both x and y, because its diameter is less than ε. Their active vertex sets are therefore disjoint. Each has at most n+1 vertices, so their union has at most 2n+2=m+1. The convex-hull lemma proves g(x)≠g(y). We have approximated f arbitrarily closely while separating all pairs at distance at least ε, not merely a finite sample of pairs.

3. Open dense conditions in a complete function space

Let C(X,Rᵐ) carry the uniform norm. It is complete: a uniformly Cauchy sequence has coordinatewise real limits, the same Cauchy estimate proves uniform convergence, and a uniform limit of continuous maps is continuous. Compactness makes each continuous map bounded, so the norm is finite. This reuses the uniform-Cauchy mechanism.

Let Uε consist of continuous maps separating every pair with d(x,y)≥ε. The approximation construction proves Uε is dense. It is also open. The pair set Kε consisting of those (x,y) is closed in compact X×X, hence compact. If it is empty, Uε is the whole function space. Otherwise, for g∈Uε the continuous positive function |g(x)−g(y)| has a positive minimum μ on Kε. Any map h within μ/3 of g in uniform norm has |h(x)−h(y)|≥μ−2μ/3>0 on Kε. Thus a uniform ball around g stays inside Uε.

4. One map separates every distinct pair

By the proved Baire theorem, the intersection of U₁, U₁⁄₂, U₁⁄₃, and so on is dense in the nonempty complete function space, hence nonempty. Choose F in this intersection. If x≠y, choose k with 1/k≤d(x,y). Membership in U₁⁄ₖ gives F(x)≠F(y). Thus F is continuous and injective.

Finally, a continuous bijection from compact X to its image in the Hausdorff space Rᵐ is a homeomorphism: every closed subset of X is compact, its image is compact and hence closed in the image space, so the inverse is continuous. Therefore F is an embedding. The image is compact, hence closed in Rᵐ. The empty space has its unique empty embedding separately.

The infinite conclusion comes from Baire and compactness, not from executing finitely many perturbations. A map satisfying one Uε can still identify nearby points. Conversely, plain injectivity does not automatically imply an embedding for noncompact domains: the usual angular map from [0,2π) onto the circle is continuous and bijective, but points approaching 2π have images approaching the image of zero, so the inverse is not continuous there.

A noncompact theorem with a different target

Can distances to named landmarks encode a point even when compactness is unavailable? A single landmark on the real line cannot distinguish −1 and 1. Distances to a dense sequence can. We now prove that every nonempty separable metric space (X,d) embeds in the countable cube Q=[0,1]ᴺ with its product topology. “Separable” means that X has a countable dense subset. Choose a sequence a₁,a₂,… whose values form such a subset; repeat points if the subset is finite.

Define the nth coordinate φₙ(x)=d(x,aₙ)/(1+d(x,aₙ)), and F(x)=(φ₁(x),φ₂(x),…). Every coordinate lies in [0,1), even when d is unbounded. Give Q the metric

ρ(u,v)=Σ from n=1 to infinity of 2⁻ⁿ|uₙ−vₙ|.

The series converges because each coordinate difference is at most 1. Positivity and symmetry follow termwise; a zero sum forces equality of every coordinate, and summing coordinate triangle inequalities proves the triangle inequality. The tail after coordinate N is at most 2⁻ᴺ.

This metric induces the product topology. A small ρ-ball controls any specified finite set of coordinates because |uₙ−vₙ|≤2ⁿρ(u,v). Conversely, to fit inside a ρ-ball of radius ε, first choose N with 2⁻ᴺ<ε/2, then require all first N coordinate differences to be less than ε/2. Their weighted sum is less than ε/2; the remaining tail is less than ε/2. These are finite-coordinate neighborhoods, precisely the product topology mechanism.

Continuity. The scalar map h(r)=r/(1+r) satisfies |h(r)−h(s)|=|r−s|/((1+r)(1+s))≤|r−s| for r,s≥0. The distance triangle inequality gives |d(x,aₙ)−d(y,aₙ)|≤d(x,y). Therefore ρ(F(x),F(y))≤d(x,y)Σ2⁻ⁿ=d(x,y). In particular F is continuous.

Injectivity. For x≠y set δ=d(x,y)>0. Density supplies aₙ with d(x,aₙ)<δ/3. Then d(y,aₙ)>2δ/3 by the triangle inequality. Since h is strictly increasing, their nth coordinates differ. This proves separation for every pair, not just the displayed choices.

Inverse continuity without compactness. Fix x and ε>0. Choose aₙ with d(x,aₙ)<ε/4. The cube neighborhood V defined by uₙ<h(ε/2) contains F(x). If F(y) belongs to V, strict monotonicity gives d(y,aₙ)<ε/2, hence d(y,x)<3ε/4<ε. Thus the inverse from F(X) to X is continuous at F(x). This replaces the compact-to-Hausdorff inverse argument; injectivity alone would not suffice. The empty space embeds uniquely as the empty subset.

One distance can hide a reflection

Predict whether the first distance coordinate distinguishes −1 from 1. Add the anchor 1, then explain what changed. These are the first five anchors of a dense rational enumeration of R, not the entire embedding.

Point x -2-1012416 Point y -2-1012416 Number N of shown coordinates 12345

• n=1, anchor 0: φₙ(x)=0.500000, φₙ(y)=0.500000; weight 0.5; contribution 0.000000.

Original distance |x−y|=2. Product-metric partial sum Sₙ=0.000000. Omitted tail is at most 0.500000. Therefore the full product distance lies between 0.000000 and 0.500000. These are bounds, not a computed infinite sum.

Displayed coordinates coincide for distinct points. Additional anchors are necessary; finite agreement does not refute the infinite embedding.

Check the prediction and an escaping sequence

At anchor 0, −1 and 1 both give 1/2. At anchor 1 they give 2/3 and 0. For x=m tending to positive infinity, each fixed anchor coordinate tends to 1. The limit (1,1,…) lies in the cube but is not an image point. Noncompact embedded images need not be closed. A finite selection of m values illustrates the limit, while the proof uses all m.

Reset coordinates

Temporary practice with dimensionless values. No infinite embedding or covering dimension is computed by this control.

The theorem preserves topology, not the original metric: all cube distances are at most 1, while R has arbitrarily large distances. It proves no finite-dimensional target bound. For R, choose any dense rational enumeration beginning 0,1,−1,2,−2. As x=m tends to infinity, every fixed φₙ(m) tends to 1. The tail estimate proves convergence in ρ to (1,1,…), which is outside F(R), since every coordinate of every image point is strictly less than 1. The image is therefore not closed. There is no contradiction with compactness of the cube: a subspace of a compact space need not be closed or compact.

The Hilbert-cube universality statement in Zielinski’s research paper, Preliminaries (page 3) provides a source comparison. The distance-coordinate construction above is supplied in full. It does not prove the noncompact finite-dimensional Menger–Nöbeling theorem, or the lower bound for the dimension of Euclidean cubes.

Changed-case proof practice

Let X=(0,1) with its usual distance. Explain how a dense rational sequence in X produces an embedding. Does its image have to be closed in Q? Use xₘ=1/m for m≥2 to construct a cube limit, and prove that no point of X has that image. Then decide whether the general density proof guarantees that the first hundred coordinates suffice, and whether this particular interval admits a stronger finite-coordinate argument.

Feedback and proof criteria

For each anchor aₙ in (0,1), the coordinate limit is aₙ/(1+aₙ). Finite-coordinate convergence plus the uniform tail bound proves convergence in ρ. If an x in (0,1) had these coordinates, strict monotonicity would imply |x−aₙ|=aₙ for every n. Density gives an anchor with 0<aₙ<x/3, for which |x−aₙ|>2x/3>aₙ, a contradiction. The limit is outside the image. The general density argument supplies no universal first-hundred guarantee: an enumeration may even repeat the same anchor for its first hundred terms. However, for this interval, any two distinct anchors already separate every pair. Indeed, if distinct x,y have equal distances to an anchor a, then a=(x+y)/2; two distinct anchors cannot both equal that midpoint. Distinguish this special one-dimensional argument from the general theorem. A complete embedding answer still includes the product-tail argument and inverse continuity, not merely injectivity or a formula for F.

A finite-dimensional noncompact embedding with a complete proof

A locally finite simplicial complex has finitely many simplices incident to each vertex. Give its realization the usual topology obtained by gluing closed simplices along common faces. We prove: a countable locally finite complex of dimension at most n admits a proper piecewise-linear embedding into R^(2n+1). Proper means that the inverse image of every compact set is compact. This is a noncompact theorem, but its locally finite simplicial hypothesis is stronger than merely being a separable metric space of covering dimension n.

Enumerate the vertices v₁,v₂,… and put m=2n+1. Choose their images pⱼ successively with first coordinate in (j,j+1/2), avoiding every affine span of at most m previous images. Only finitely many proper affine subspaces are forbidden at each step; the line-avoidance argument already proved above shows that their union cannot fill this open strip. Thus every set of at most m+1 images is affinely independent. For finitely many vertices use the same construction with a finite list.

Extend affinely over each simplex: F(Σaⱼvⱼ)=Σaⱼpⱼ for nonnegative barycentric coefficients summing to one. Definitions agree on shared faces, and the gluing topology makes F continuous. If F(x)=F(y), the union of their minimal supporting simplices has at most 2n+2=m+1 vertices. Subtract the two barycentric expressions. Affine independence forces every coefficient difference to vanish, so x=y, including the case of overlapping simplices. This is why considering only disjoint edges would not have been enough.

For properness fix a compact target set contained in the slab with first coordinate at most R. If every vertex of a simplex has index greater than R, its entire affine image has first coordinate greater than R. Hence F’s inverse image of that compact set lies in the union of simplices incident to vertices with index at most R. There are finitely many such vertices and local finiteness makes that union a finite union of closed compact simplices. The inverse image is closed in this finite union, hence compact.

Finally F has continuous inverse on its image. To check sequentially, if F(xₖ)→F(x), the convergent sequence together with its limit is compact. Properness puts all xₖ and x in a compact inverse image. Every convergent subsequence has limit x by continuity and injectivity. If xₖ did not converge to x, a subsequence outside a neighborhood of x would have a further convergent subsequence, a contradiction. Locally finite countable realizations are metrizable: locally each point meets a finite complex, and the usual weak realization is second countable and regular; equivalently one can use its standard piecewise-simplex metric. The same argument can be phrased with compact closed sets, avoiding sequences: a proper map from this locally compact Hausdorff domain into Euclidean space is closed, as a compact target neighborhood reduces closure to the compact case. Therefore the inverse is continuous.

Changed case. Apply the theorem to the doubly infinite chain of edges with vertices indexed by the integers. Why can placing consecutive vertices at 1,1/2,1/3,… in the line fail to be proper? What in our proof prevents a distant part of the chain from accumulating at a finite point?

Feedback: separate existence, injection and properness

An ordinary geometric realization of the chain is R, so it embeds properly in R and also in the theorem's sufficient target R³. A map sending one ray's successive vertices to 1,1/2,1/3,… has an infinite tail inside a compact interval; its inverse image is noncompact, even if the ray map itself is injective. Our index-dependent first-coordinate strips force every bounded target set to meet only simplices incident to a finite set of vertices. Local finiteness is essential to turning that finite vertex set into finitely many simplices. A countably infinite star with infinitely many edges at one center does not satisfy that hypothesis.

This proof does not replace the full Menger–Nöbeling theorem for arbitrary separable metric spaces. It closes the explicit noncompact locally finite polyhedral case and identifies the stronger theorem rather than claiming every space has a locally finite triangulation.

Why an n-cube really has dimension n

Here is a complete lower-bound route through the finite Sperner parity argument. All triangulations used below are finite, and n≥1; the point cube has dimension zero separately.

Sperner lemma. Label vertices of a triangulated n-simplex with 0,…,n so that a vertex on a face can use only a label of a vertex of that face. An odd number of small n-simplices have all n+1 labels. For n=0 the assertion is immediate. Inductively count incidences of small n-simplices with their facets bearing exactly labels 0,…,n−1. An interior facet is counted twice. A boundary facet of this label type can lie only in the original face opposite vertex n, where induction gives an odd number. Within a small n-simplex the number of these facets is one if all n+1 labels occur, two if just 0,…,n−1 occur with one repeated, and zero otherwise. Reducing both counts modulo two proves the assertion. No geometric sketch substitutes for the boundary-label restriction.

Brouwer fixed point consequence. A continuous self-map f of the standard simplex has a fixed point. Otherwise, at every vertex v of arbitrarily fine subdivisions choose a label i with vᵢ>fᵢ(v); some such coordinate exists because both coordinate sums are one and equality in every coordinate would be a fixed point. A zero coordinate cannot receive that label, so Sperner applies. Select a fully labeled small simplex from each subdivision. Compactness and mesh tending to zero give a common subsequential limit x for all its labeled vertices. Continuity yields xᵢ≥fᵢ(x) for every i. Their equal sums force equality in every coordinate, contradiction. The theorem passes to a cube by a homeomorphism: send an interior center to an interior center and each radial segment linearly to the corresponding radial segment in the other convex body, after identifying their ambient n-dimensional affine spaces. The boundary-distance functions are positive and continuous, so this radial map and its inverse are continuous.

No retraction. There is no continuous map R from the centered cube [−1,1]ⁿ to its boundary equal to the identity on that boundary. If it existed, the self-map x↦−R(x) would have a fixed point. That point lies on the boundary, where R(x)=x; hence x=−x=0, which is not on the boundary.

For the upper dimension bound, triangulate the cube into simplices of sufficiently small diameter, and take their vertex open stars. In a triangulation a point belongs to the open stars of exactly the vertices of its minimal supporting simplex, at most n+1. A star’s diameter is at most twice the mesh, because its points lie in simplices sharing its vertex. Arbitrarily fine subdivisions therefore make the stars refine any given open cover, using a Lebesgue number. The finite open-star cover proves dim[0,1]ⁿ≤n.

For the lower bound, suppose every finite open cover had an open refinement of multiplicity at most n. Begin with a finite cover of sup-norm diameter less than ε, where 0<ε<1/4. A subordinate finite refinement V₁,…,Vᵣ would have the same diameter bound. Pick pᵢ∈Vᵢ and construct the continuous distance weights λᵢ used earlier, with positive weights only inside Vᵢ. The map F(x)=Σλᵢ(x)pᵢ satisfies ||F(x)−x||∞<ε. At most n coefficients are nonzero at each x, so its image lies in a finite union of simplices of dimension at most n−1. Such a union cannot contain an open ball: each simplex lies in a proper affine hyperplane and the finite line-avoidance argument applies. Choose y in (2ε,1−2ε)ⁿ outside this image.

Radially project from y to the boundary of the cube. This defines a continuous map r_y on the cube minus y. For a boundary point x, the segment from x to F(x) cannot meet y: one coordinate of x is 0 or 1 and that coordinate stays within ε of the corresponding face, whereas y is more than 2ε from every face. Consequently H(x,t)=r_y((1−t)x+tF(x)) is a homotopy on the boundary from its identity to g(x)=r_y(F(x)). The latter extends to the whole cube because F avoids y.

Such an extension plus this homotopy would give a forbidden retraction. To see this explicitly, recenter/rescale to [−1,1]ⁿ and write x=ru with ||u||∞=1. On r≤1/2 use g(2x). On 1/2≤r≤1 use H(u,2−2r). At r=1/2 both equal g(u), and at r=1 the value is u=x. These formulas are continuous and land on the boundary. The contradiction proves the lower bound and hence dim[0,1]ⁿ=n.

Proof transfer. Why does the lower-bound argument fail if there may be n+1 active weights? Because their convex hull may be n-dimensional and fill an open set, so the omitted interior point y need not exist. Why is an injective picture of the cube insufficient? Covering dimension concerns every cover and a lower bound on refinement multiplicity; it is not the coordinate count of one picture.

Primary comparison: Karasev, Covering dimension using toric varieties, inspected 30 October 2026, treats the Lebesgue covering theorem through a different method. The finite parity, fixed-point and refinement argument above is provided explicitly; no toric geometry is a prerequisite for it.

Proper embeddings beyond simplicial complexes

There is a substantially broader noncompact theorem that still admits a proof from this course’s tools. Let X be a locally compact, second-countable Hausdorff space, and suppose every compact subspace has covering dimension at most n. Then X has a proper embedding into R^(2n+1). Proper means inverse images of compact sets are compact. This includes spaces that have not been triangulated. The MIT supplementary Notes I, page 2 directs the reader to the noncompact extension in Munkres §50, exercise 7. The argument here is an original assembly of the already proved compact approximation, metrization, extension and Baire tools.

Proof: a proper starting map, bounded perturbations and compact tests

Exhaustion and a proper height. Local compactness and Hausdorff separation give a neighborhood of each point with compact closure. Second countability gives a countable such cover: first choose, for each point, a basis member inside a relatively compact neighborhood, then discard repetitions. Enumerate their compact closures C₁,C₂,…. Construct compact Kⱼ recursively so that C₁∪⋯∪Cⱼ⊂Kⱼ and Kⱼ⊂int Kⱼ₊₁. For the inductive step, cover Kⱼ by finitely many relatively compact open neighborhoods and take the union of their closures together with C₁,…,Cⱼ₊₁. These Kⱼ exhaust X, and every compact subset lies in some Kⱼ: their interiors form an increasing open cover.

The space is regular Hausdorff by local compactness and hence metrizable by the already proved second-countable metrization theorem. For each j the disjoint closed sets Kⱼ and X∖int Kⱼ₊₁ admit a continuous ψⱼ:X→[0,1], zero on the first and one on the second, by the metric Urysohn lemma. Put h=Σⱼψⱼ. This is locally a finite sum: a neighborhood contained in int Kᴺ has ψⱼ=0 for every j≥N. Outside Kᴺ the first N−1 summands equal one, so h≥N−1. Thus each sublevel set h⁻¹([0,R]) is closed and contained in a compact Kᴺ for N−1>R. It is compact, and h is proper.

A complete space of proper maps. Write m=2n+1 and f₀=(h,0,…,0). Let B=C_b(X,Rᵐ), the bounded continuous functions with supremum norm. Uniform Cauchy sequences converge uniformly to a bounded continuous limit, so B is complete even though X need not be compact. The affine space A=f₀+B inherits this complete metric. Every f∈A is proper: if |f(x)|≤R then h(x)=|f₀(x)|≤R+||f−f₀||∞, a compact sublevel bound. The inverse image of any compact target set is closed inside that compact sublevel.

Open dense compact tests. For positive integers N,k let Uᴺ,ₖ be those f∈A that separate every pair x,y∈Kᴺ with d(x,y)≥1/k. These sets are open by exactly the positive-minimum argument on the compact pair set used in the compact proof. They are dense as well. Given f∈A and η>0, apply the compact general-position approximation to f restricted to Kᴺ, obtaining g:Kᴺ→Rᵐ that separates the required pairs and satisfies ||g−f||∞<η/(2√m). Extend each coordinate of g−f from the closed set Kᴺ to X by the bounded Tietze theorem, with absolute value at most η/(2√m). The vector extension e is bounded and has ||e||∞≤η/2. Therefore f+e belongs to A, is within η of f, and agrees with g on Kᴺ. It belongs to Uᴺ,ₖ. No extension of an injective map is being claimed: only its small coordinate correction is extended, and separation is required on this one compact set.

Take all tests, then check the inverse. Baire gives F in the countable intersection of the Uᴺ,ₖ. Any distinct pair lies in one Kᴺ and has positive distance, so one test separates it. Hence F is injective, continuous and proper. To see inverse continuity without using compactness of all X, suppose F(xᵢ)→F(x). The convergent sequence together with its limit is a compact target set; properness puts all xᵢ and x in a compact subset of the metrizable X. Any subsequence not converging to x would have a further convergent subsequence with limit y≠x; continuity would give F(y)=F(x), contradicting injectivity. Thus xᵢ→x. Since the image is metric, this proves inverse continuity. The same compact-preimage argument shows F(X) is closed in Rᵐ. This proves the theorem.

Changed-case proof task. Let X be the disjoint union of countably many compact metrizable spaces Xⱼ of dimension at most n. Show that X satisfies the theorem without constructing a common triangulation. A compact subset of X meets only finitely many components (use the component open cover). Its intersections with those components are closed compact subspaces; restricting a finite cover and adjoining the complement proves that a closed subspace cannot have larger covering dimension. Refine separately on the finitely many disjoint components, preserving multiplicity. Thus every compact subset has dimension at most n. Each component has a countable basis, their countable union is a basis for X, and compact components give local compactness. The theorem yields a proper embedding. A picture that places all components in balls accumulating at an omitted finite point is not proper, even if it is injective.

Why the remaining general theorem is genuinely different. The rationals with their usual metric have a clopen basis and are zero-dimensional, but are not locally compact: every neighborhood contains a rational sequence converging to an irrational point and so cannot have compact closure in Q. They already embed in R by inclusion, but not properly as a closed subset of R, since a closed subset of R is locally compact. The general separable-metric Menger–Nöbeling theorem promises an embedding, not a proper embedding. Neither exhaustion by compact neighborhoods nor the proper-map proof above applies to arbitrary separable metric spaces.

Apply

Prove the circle cannot embed in R. A continuous embedding would have compact connected image, a nondegenerate closed interval. Removing a point mapping to an interior interval point disconnects that image, while removing any one circle point leaves a connected space. The restricted homeomorphism would preserve connectedness, giving a contradiction.

Repair “The theorem’s target dimension is minimal.” The bound 2n+1 is general and sufficient. A circle embeds in R², even though the general n=1 bound supplies R³.

Continue

The compact-metric finite-dimensional theorem and the noncompact separable-metric countable-cube theorem are proved above. The cube dimension proof, the proper finite-dimensional embedding for countable locally finite complexes, and the broader proper embedding for locally compact second-countable Hausdorff spaces with the stated compact-subspace dimension bound are now supplied. The full noncompact Menger–Nöbeling theorem for arbitrary separable metric spaces remains a separately identified dependency. Optional depth: explicitly construct the finite clopen atom refinement for a small cylinder cover of the Cantor set, then distinguish that finite calculation from the proof for every finite cover.