Introduction
Gurariy and Quarta [13 ] initiated in 2004 the study of linear subspaces of continuous functions whose nonzero elements attain their maximum at exactly one point. On a compact nondegenerate interval the dimension is at most one [13 Theorem 6] , whereas two-dimensional examples exist on half-open intervals and on R \mathbb{R} R [13 Proposition 8 and Theorem 9] . For functions on R \mathbb{R} R vanishing at infinity, they also established the sharp bound two [13 Theorem 16] .
In [8 ] , constructions involving embedded spheres were used to obtain finite-dimensional examples on more general domains. The same work established the upper bound m m m on compact subsets of R m \mathbb{R}^m R m , for every positive integer m m m . Corollary 4.4 below generalizes the compact metric example of [8 Example 4.1] and identifies a family of maximally spaceable examples. In 2020, Bernal-González, Cabana-Méndez, Muñoz-Fernández and Seoane-Sepúlveda settled the original question on R \mathbb{R} R by proving that two is the largest possible dimension in C ( R ) C(\mathbb{R}) C ( R ) and in C ( [ a , b ) ) C([a,b)) C ([ a , b )) , for a < b a<b a < b [6 Theorem 3.22] . Their Section 4 establishes the same sharp bound on R \mathbb{R} R when every nonzero function has exactly m m m maximizing points, for a fixed integer m ≥ 2 m \ge2 m ≥ 2 .
For functions vanishing at infinity on R n \mathbb{R}^n R n , with n ≥ 2 n \ge2 n ≥ 2 , the results in [2 Theorems 6 and 7] give admissible subspaces of dimension n n n and exclude dimension n + 2 n+2 n + 2 . Whether dimension n + 1 n+1 n + 1 can occur remained open [2 Conjecture 1] . The same question is also recorded in [5 Conjecture 1.11] . Aron and Seoane-Sepúlveda’s 2025 account [4 Example 2.1] recalls the unique-maximum bounds on compact intervals and on R \mathbb{R} R . General background on lineability and spaceability can be found in the monograph [3 ] and the survey [7 ] .
We examine both the possible dimensions and the extension of prescribed subspaces. All vector spaces and functions are real. Dimensions of vector spaces are Hamel dimensions; covering dimension of a topological space is written dim t o p \dim_{\mathrm{top}} dim top and defined in § 5 . Write N = { 1 , 2 , … } \mathbb{N}=\{1,2,\ldots\} N = { 1 , 2 , … } , ℵ 0 = card ( N ) \aleph_0=\operatorname{card}(\mathbb{N}) ℵ 0 = card ( N ) and c = card ( R ) \mathfrak{c}=\operatorname{card}(\mathbb{R}) c = card ( R ) . For d ∈ N d\in\mathbb{N} d ∈ N , ℓ 2 d \ell_2^d ℓ 2 d denotes R d \mathbb{R}^d R d with its Euclidean norm. For a real function f : X → R f:X\to\mathbb{R} f : X → R on a nonempty set, put
Argmax X f = { x ∈ X : f ( x ) = sup z ∈ X f ( z ) } , Argmin X f = Argmax X ( − f ) . \operatorname{Argmax}_{X} f=\{x\in X:f(x)=\sup_{z\in X}f(z)\},\qquad\operatorname{Argmin}_{X} f=\operatorname{Argmax}_{X}(-f). Argmax X f = { x ∈ X : f ( x ) = z ∈ X sup f ( z )} , Argmin X f = Argmax X ( − f ) .
The supremum is taken in the extended real line. Thus Argmax X f \operatorname{Argmax}_{X}f Argmax X f is empty when f f f is unbounded above or its supremum is not attained; the condition ∣ Argmax X f ∣ = 1 |\operatorname{Argmax}_{X}f|=1 ∣ Argmax X f ∣ = 1 means that f f f attains its maximum at exactly one point. Following the notation of [2 ] , let C ( X ) C(X) C ( X ) be the space of continuous real functions on X X X , with the compact-open topology given by the seminorms f ↦ sup x ∈ H ∣ f ( x ) ∣ f \mapsto\sup_{x\in H} |f(x)| f ↦ sup x ∈ H ∣ f ( x ) ∣ for nonempty compact H ⊂ X H \subset X H ⊂ X . Let C b ( X ) ⊂ C ( X ) C_b(X) \subset C(X) C b ( X ) ⊂ C ( X ) consist of the bounded functions, with norm ∥ f ∥ ∞ = sup x ∈ X ∣ f ( x ) ∣ \|f\|_\infty= \sup_{x\in X} |f(x)| ∥ f ∥ ∞ = sup x ∈ X ∣ f ( x ) ∣ , and define
C ^ ( X ) = { f ∈ C b ( X ) : ∣ Argmax X f ∣ = 1 } . \widehat{C}(X) = \{f \in C_b(X) : |\operatorname{Argmax}_X f| = 1\}. C ( X ) = { f ∈ C b ( X ) : ∣ Argmax X f ∣ = 1 } .
For a Hausdorff space X X X , let C 0 ( X ) C_0(X) C 0 ( X ) be the space of continuous functions f : X → R f : X \to\mathbb{R} f : X → R such that { x : ∣ f ( x ) ∣ ≥ ε } \{x : |f(x)| \ge\varepsilon\} { x : ∣ f ( x ) ∣ ≥ ε } is compact for every ε > 0 \varepsilon> 0 ε > 0 , with the supremum norm. This definition also applies when X X X is not locally compact; in the locally compact Hausdorff case it is the usual Banach space of functions vanishing at infinity. Define
C ^ 0 ( X ) = { f ∈ C 0 ( X ) : ∣ Argmax X f ∣ = 1 } . \widehat{C}_0(X) = \{f \in C_0(X) : |\operatorname{Argmax}_X f| = 1\}. C 0 ( X ) = { f ∈ C 0 ( X ) : ∣ Argmax X f ∣ = 1 } .
For compact Hausdorff K K K , we have C 0 ( K ) = C ( K ) C_0(K) = C(K) C 0 ( K ) = C ( K ) and C ^ 0 ( K ) = C ^ ( K ) \widehat{C}_0(K) = \widehat{C}(K) C 0 ( K ) = C ( K ) . A linear subspace of C ( X ) C(X) C ( X ) whose nonzero members attain unique maxima is called admissible . Although a continuous function with a unique maximum need not be bounded below, every admissible subspace consists of bounded functions, since it contains both f f f and − f -f − f , which attain finite maxima. Its nonzero members also attain unique minima. Thus every admissible subspace of C ( X ) C(X) C ( X ) lies in C ^ ( X ) ∪ { 0 } \widehat{C}(X) \cup\{0\} C ( X ) ∪ { 0 } . We consider these subspaces in C ( X ) C(X) C ( X ) and in C b ( X ) C_b(X) C b ( X ) , with the respective topologies specified above.
The notion of ( α , β ) (\alpha,\beta) ( α , β ) -lineability was introduced in [12 Definition 1.1] . A subset A A A of a real vector space V V V is α \alpha α -lineable if A ∪ { 0 } A\cup\{0\} A ∪ { 0 } contains an α \alpha α -dimensional subspace. For cardinals α ≤ β \alpha\le\beta α ≤ β , it is ( α , β ) (\alpha,\beta) ( α , β ) -lineable if it is α \alpha α -lineable and every subspace E ⊂ A ∪ { 0 } E \subset A\cup\{0\} E ⊂ A ∪ { 0 } of dimension α \alpha α is contained in a subspace W ⊂ A ∪ { 0 } W \subset A \cup\{0\} W ⊂ A ∪ { 0 } of dimension β \beta β . In a topological vector space, ( α , β ) (\alpha,\beta) ( α , β ) -spaceability requires that A A A be α \alpha α -lineable and a closed extension W W W with these properties. The prescribed E E E need not be closed. We include pairs with α = β \alpha= \beta α = β , as in [11 ] . A set is α \alpha α -spaceable if it contains, together with zero, a closed α \alpha α -dimensional subspace; it is maximally spaceable when α \alpha α is the dimension of the ambient space. For normed spaces E 1 E_1 E 1 , E 2 E_2 E 2 and a bounded linear map T : E 1 → E 2 T : E_1 \to E_2 T : E 1 → E 2 , write ∥ T ∥ = sup ∥ f ∥ E 1 ≤ 1 ∥ T f ∥ E 2 \|T\| = \sup_{\|f\|_{E_1}\le1} \|Tf\|_{E_2} ∥ T ∥ = sup ∥ f ∥ E 1 ≤ 1 ∥ T f ∥ E 2 . In 2024, results on ( α , β ) (\alpha,\beta) ( α , β ) -lineability and ( α , β ) (\alpha,\beta) ( α , β ) -spaceability for complements of vector subspaces were obtained in [1 ] , while general obstructions to closed extensions were established in [11 Theorem 2.1] . A related obstruction appears in [2 Theorem 9] ; in Banach spaces its hypotheses require infinite prescribed dimension. The finite prescribed dimensions considered below require a different argument. In particular, failure of ( α , β ) (\alpha,\beta) ( α , β ) -lineability does not mean that no α \alpha α -dimensional subspace extends: one nonextendible α \alpha α -dimensional subspace is enough to disprove the property.
Recall that X X X is countably compact if every countable open cover has a finite subcover. Every continuous map from such a space into a metrizable space has compact image. In particular, C ( X ) = C b ( X ) C(X) = C_b(X) C ( X ) = C b ( X ) for countably compact X X X ; this does not imply C 0 ( X ) = C ( X ) C_0(X) = C(X) C 0 ( X ) = C ( X ) . For a noncompact X X X , the constant function one does not vanish at infinity.
A finite-dimensional admissible subspace can be perturbed, by an arbitrarily small isometric change preserving every maximizing and minimizing point, into an admissible subspace that admits no proper admissible extension. The precise statement and proof appear in Theorem 2.8 . Here maximal by inclusion means that no proper admissible extension exists, not that the subspace has the largest possible dimension.
The perturbation theorem gives the following classification whenever the prescribed dimension is positive and finite.
Theorem A. Let X X X be locally compact Hausdorff with at least two points. For cardinals 1 ≤ α < ℵ 0 1 \le\alpha< \aleph_0 1 ≤ α < ℵ 0 and β ≥ α \beta\ge\alpha β ≥ α ,
C ^ 0 ( X ) is ( α , β ) -lineable ⟺ α = β and C ^ 0 ( X ) is α -lineable . \widehat{C}_0(X)\text{ is }(\alpha,\beta)\text{-lineable}
\Longleftrightarrow
\alpha=\beta\text{ and }\widehat{C}_0(X)\text{ is }\alpha\text{-lineable}. C 0 ( X ) is ( α , β ) -lineable ⟺ α = β and C 0 ( X ) is α -lineable . The same equivalence holds for ( α , β ) (\alpha,\beta) ( α , β ) -spaceability. If instead X X X is countably compact Hausdorff, the same equivalences hold with C ( X ) C(X) C ( X ) in place of C ^ 0 ( X ) \widehat{C}_0(X) C 0 ( X ) , for spaceability in either the supremum norm or the compact-open topology.
On a connected n n n -manifold M M M with n ≥ 2 n \ge2 n ≥ 2 , the largest admissible dimension in C 0 ( M ) C_0(M) C 0 ( M ) has a topological description. By an n n n -manifold we mean a nonempty Hausdorff second-countable space locally homeomorphic to R n \mathbb{R}^{n} R n , without boundary. Write M ≅ N M \cong N M ≅ N for homeomorphism and S n = { a ∈ R n + 1 : ∑ j = 1 n + 1 a j 2 = 1 } S^{n}=\{a\in\mathbb{R}^{n+1}:\sum_{j=1}^{n+1}a_{j}^{2}=1\} S n = { a ∈ R n + 1 : ∑ j = 1 n + 1 a j 2 = 1 } .
Theorem B. Let M M M be a connected n n n -manifold, with n ≥ 2 n\geq2 n ≥ 2 , and let 1 ≤ α ≤ β 1\leq\alpha\leq\beta 1 ≤ α ≤ β be cardinals.
If M ≅ S n M\cong S^{n} M ≅ S n , then C ^ 0 ( M ) \widehat{C}_{0}(M) C 0 ( M ) is ( α , β ) (\alpha,\beta) ( α , β ) -lineable if and only if α = β ≤ n + 1 \alpha=\beta\leq n+1 α = β ≤ n + 1 .
If M ≇ S n M\not\cong S^{n} M ≅ S n , then C ^ 0 ( M ) \widehat{C}_{0}(M) C 0 ( M ) is ( α , β ) (\alpha,\beta) ( α , β ) -lineable if and only if α = β ≤ n \alpha=\beta\leq n α = β ≤ n .
The same equivalences hold for ( α , β ) (\alpha,\beta) ( α , β ) -spaceability.
In particular, C ^ 0 ( R n ) \widehat{C}_{0}(\mathbb{R}^{n}) C 0 ( R n ) is n n n -lineable but not ( n + 1 ) (n+1) ( n + 1 ) -lineable for every n ≥ 2 n\geq2 n ≥ 2 , resolving [2 Conjecture 1] . The case n = 1 n=1 n = 1 is an essential exception: C ^ 0 ( R ) \widehat{C}_{0}(\mathbb{R}) C 0 ( R ) has largest admissible dimension two [13 Theorem 16] . Together with Theorem A, this gives α = β ≤ 2 \alpha=\beta\leq2 α = β ≤ 2 on R \mathbb{R} R .
For a topological space X X X , a zero-set is u − 1 ( { 0 } ) u^{-1}(\{0\}) u − 1 ({ 0 }) for some continuous u : X → [ 0 , 1 ] u:X\to[0,1] u : X → [ 0 , 1 ] . Continuous real functions separate points if, whenever x ≠ y x\neq y x = y , there is a continuous f : X → R f:X\to\mathbb{R} f : X → R with f ( x ) ≠ f ( y ) f(x)\neq f(y) f ( x ) = f ( y ) . Proposition 3.1 characterizes finite admissible dimensions on countably compact spaces with this separation property by zero-sets homeomorphic to spheres. Combined with Theorem A and the infinite-dimensional obstruction in Lemma 4.1, it gives the spaceability criterion in Corollary 4.2. The sphere must be a zero-set; for compact metrizable domains, any embedded sphere has this property.
Corollary 2.9 transfers the finite-dimensional nonextension result from C 0 ( X ) C_{0}(X) C 0 ( X ) to C ( X ) C(X) C ( X ) . On Euclidean domains, we also determine the largest admissible dimension without imposing vanishing at infinity. The upper bound extends to every σ \sigma σ -compact metrizable space of finite covering dimension; on Euclidean domains it is attained.
Theorem C. For every integer n ≥ 1 n\geq1 n ≥ 1 , the set C ^ ( R n ) \widehat{C}(\mathbb{R}^{n}) C ( R n ) is ( n + 1 ) (n+1) ( n + 1 ) -lineable but not ( n + 2 ) (n+2) ( n + 2 ) -lineable. In fact, it contains, together with zero, a linear isometric copy of ℓ 2 n + 1 \ell_{2}^{n+1} ℓ 2 n + 1 . For all cardinals 1 ≤ α ≤ β 1\leq\alpha\leq\beta 1 ≤ α ≤ β , each of the following properties holds if and only if 1 ≤ α = β ≤ n + 1 1\leq\alpha=\beta\leq n+1 1 ≤ α = β ≤ n + 1 :
( α , β ) (\alpha,\beta) ( α , β ) -lineability of C ^ ( R n ) \widehat{C}(\mathbb{R}^{n}) C ( R n ) ;
( α , β ) (\alpha,\beta) ( α , β ) -spaceability in C b ( R n ) C_{b}(\mathbb{R}^{n}) C b ( R n ) with the supremum norm;
( α , β ) (\alpha,\beta) ( α , β ) -spaceability in C ( R n ) C(\mathbb{R}^{n}) C ( R n ) with the compact-open topology.
For n ≥ 2 n\geq2 n ≥ 2 , the largest admissible dimensions in C 0 ( R n ) C_{0}(\mathbb{R}^{n}) C 0 ( R n ) and C ( R n ) C(\mathbb{R}^{n}) C ( R n ) are n n n and n + 1 n+1 n + 1 , respectively. For n = 1 n=1 n = 1 , both dimensions equal two.
Proof of Theorem A
A proper admissible extension forces a continuous section of the evaluation map. A small radial deformation preserves all extrema while preventing such a section.
Convex sets and evaluation maps
The convex-hull arguments are related to [6 Propositions 3.1 and 3.3, Lemma 3.5 and Theorem 3.8] and [8 Lemmas 3.2–3.5] . Continuity of the maximizing-point map on an interval already appears in [13 Lemma 5] . We use the Euclidean inner product ⟨ a , u ⟩ = ∑ j = 1 d a j u j \langle a,u\rangle=\sum_{j=1}^{d}a_{j}u_{j} ⟨ a , u ⟩ = ∑ j = 1 d a j u j and norm ∥ a ∥ = ( ∑ j = 1 d a j 2 ) 1 / 2 \lVert a\rVert=(\sum_{j=1}^{d}a_{j}^{2})^{1/2} ∥ a ∥ = ( ∑ j = 1 d a j 2 ) 1/2 on R d \mathbb{R}^{d} R d . For C ⊂ R d C\subset\mathbb{R}^{d} C ⊂ R d , its Euclidean interior and boundary are written int C \operatorname{int}C int C and ∂ C \partial C ∂ C . We refer to [16 ] for the standard convex-geometric facts used below, including Carathéodory’s theorem, supporting hyperplanes and strict separation.
Lemma 2.1. Let Y Y Y be a nonempty topological space, let d ∈ N d\in\mathbb{N} d ∈ N , and let Ψ : Y → R d \Psi:Y\to\mathbb{R}^{d} Ψ : Y → R d be continuous with compact image. This image condition holds, in particular, when Y Y Y is countably compact. Let C = conv Ψ ( Y ) C=\operatorname{conv}\Psi(Y) C = conv Ψ ( Y ) , where conv \operatorname{conv} conv denotes the set of finite convex combinations. For a ∈ R d ∖ { 0 } a\in\mathbb{R}^{d}\setminus\{0\} a ∈ R d ∖ { 0 } , define
q a : Y → R , q a ( x ) = ⟨ a , Ψ ( x ) ⟩ , q_{a}:Y\to\mathbb{R},\qquad q_{a}(x)=\langle a,\Psi(x)\rangle, q a : Y → R , q a ( x ) = ⟨ a , Ψ ( x )⟩ , and put
H C ( a ) = max u ∈ C ⟨ a , u ⟩ , C a = { u ∈ C : ⟨ a , u ⟩ = H C ( a ) } . H_{C}(a)=\max_{u\in C}\langle a,u\rangle,\qquad C_{a}=\{u\in C:\langle a,u\rangle=H_{C}(a)\}. H C ( a ) = u ∈ C max ⟨ a , u ⟩ , C a = { u ∈ C : ⟨ a , u ⟩ = H C ( a )} . Then the following statements hold.
(1) The set C C C is compact, and
H C ( a ) = max x ∈ Y q a ( x ) , C a = conv Ψ ( Argmax Y q a ) . (1) H_C(a)=\max_{x\in Y}q_a(x),\qquad C_a=\operatorname{conv}\Psi(\operatorname{Argmax}_{Y}q_a). \tag*{(1)} H C ( a ) = x ∈ Y max q a ( x ) , C a = conv Ψ ( Argmax Y q a ) . (1)
(2) If Argmax Y q a = { x a } \operatorname{Argmax}_{Y}q_a=\{x_a\} Argmax Y q a = { x a } , then C a = { Ψ ( x a ) } C_a=\{\Psi(x_a)\} C a = { Ψ ( x a )} and Ψ − 1 ( { Ψ ( x a ) } ) = { x a } \Psi^{-1}(\{\Psi(x_a)\})=\{x_a\} Ψ − 1 ({ Ψ ( x a )}) = { x a } .
Proof. (1) By Carathéodory’s theorem, every point of C C C is a convex combination of at most d + 1 d+1 d + 1 points of Ψ ( Y ) \Psi(Y) Ψ ( Y ) . Thus C C C is the image of the compact space
{ ( λ 1 , … , λ d + 1 ) ∈ [ 0 , 1 ] d + 1 : ∑ j = 1 d + 1 λ j = 1 } × Ψ ( Y ) d + 1 \left\{(\lambda_1,\ldots,\lambda_{d+1})\in[0,1]^{d+1}:\sum_{j=1}^{d+1}\lambda_j=1\right\}\times\Psi(Y)^{d+1} { ( λ 1 , … , λ d + 1 ) ∈ [ 0 , 1 ] d + 1 : j = 1 ∑ d + 1 λ j = 1 } × Ψ ( Y ) d + 1
under the continuous map ( ( λ j ) j = 1 d + 1 , ( x j ) j = 1 d + 1 ) ↦ ∑ j = 1 d + 1 λ j Ψ ( x j ) ((\lambda_j)_{j=1}^{d+1},(x_j)_{j=1}^{d+1})\mapsto\sum_{j=1}^{d+1}\lambda_j\Psi(x_j) (( λ j ) j = 1 d + 1 , ( x j ) j = 1 d + 1 ) ↦ ∑ j = 1 d + 1 λ j Ψ ( x j ) . In particular, C C C is compact and the maximum H C ( a ) H_C(a) H C ( a ) is attained. For any convex combination u = ∑ j = 1 r λ j Ψ ( x j ) u=\sum_{j=1}^{r}\lambda_j\Psi(x_j) u = ∑ j = 1 r λ j Ψ ( x j ) , with r ∈ N r\in\mathbb{N} r ∈ N , λ j ≥ 0 \lambda_j\geq0 λ j ≥ 0 and ∑ j = 1 r λ j = 1 \sum_{j=1}^{r}\lambda_j=1 ∑ j = 1 r λ j = 1 , we have
⟨ a , u ⟩ = ∑ j = 1 r λ j q a ( x j ) ≤ ∑ j = 1 r λ j max Y q a = max Y q a . \langle a,u\rangle=\sum_{j=1}^{r}\lambda_jq_a(x_j)\leq\sum_{j=1}^{r}\lambda_j\max_{Y}q_a=\max_{Y}q_a. ⟨ a , u ⟩ = j = 1 ∑ r λ j q a ( x j ) ≤ j = 1 ∑ r λ j Y max q a = Y max q a .
Since Ψ ( Y ) ⊂ C \Psi(Y)\subset C Ψ ( Y ) ⊂ C , this proves the first identity in (1) .
Write H = H C ( a ) H=H_C(a) H = H C ( a ) . If u ∈ C a u\in C_a u ∈ C a , the same representation gives
0 = H − ⟨ a , u ⟩ = ∑ j = 1 r λ j ( H − q a ( x j ) ) . 0=H-\langle a,u\rangle=\sum_{j=1}^{r}\lambda_j\left(H-q_a(x_j)\right). 0 = H − ⟨ a , u ⟩ = j = 1 ∑ r λ j ( H − q a ( x j ) ) .
Every summand is nonnegative. Therefore q a ( x j ) = H q_a(x_j)=H q a ( x j ) = H whenever λ j > 0 \lambda_j>0 λ j > 0 , so u ∈ conv Ψ ( Argmax Y q a ) u\in\operatorname{conv}\Psi(\operatorname{Argmax}_{Y}q_a) u ∈ conv Ψ ( Argmax Y q a ) . Conversely, if every x j x_j x j with λ j > 0 \lambda_j>0 λ j > 0 belongs to Argmax Y q a \operatorname{Argmax}_{Y}q_a Argmax Y q a , then ⟨ a , u ⟩ = ∑ j = 1 r λ j H = H \langle a,u\rangle=\sum_{j=1}^{r}\lambda_jH=H ⟨ a , u ⟩ = ∑ j = 1 r λ j H = H . This proves C a = conv Ψ ( Argmax Y q a ) C_a=\operatorname{conv}\Psi(\operatorname{Argmax}_{Y}q_a) C a = conv Ψ ( Argmax Y q a ) .
(2) The equality C a = conv Ψ ( Argmax Y q a ) C_a=\operatorname{conv}\Psi(\operatorname{Argmax}_{Y}q_a) C a = conv Ψ ( Argmax Y q a ) gives C a = { Ψ ( x a ) } C_a=\{\Psi(x_a)\} C a = { Ψ ( x a )} . If Ψ ( x ) = Ψ ( x a ) \Psi(x)=\Psi(x_a) Ψ ( x ) = Ψ ( x a ) , then q a ( x ) = ⟨ a , Ψ ( x a ) ⟩ = H q_a(x)=\langle a,\Psi(x_a)\rangle=H q a ( x ) = ⟨ a , Ψ ( x a )⟩ = H . Uniqueness of the maximizer yields x = x a x=x_a x = x a . □
The Minkowski functional provides radial coordinates on a convex body and a parametrization of its boundary by a sphere.
Lemma 2.2. Let C ⊂ R k C\subset\mathbb{R}^{k} C ⊂ R k be compact and convex with nonempty interior, where k ∈ N k\in\mathbb{N} k ∈ N , and let c ∈ int C c\in\operatorname{int}C c ∈ int C . Define the Minkowski functional of C − c C-c C − c by
p : R k → [ 0 , ∞ ) , p ( y ) = inf { t > 0 : y ∈ t ( C − c ) } . p:\mathbb{R}^{k}\to[0,\infty),\qquad p(y)=\inf\{t>0:y\in t(C-c)\}. p : R k → [ 0 , ∞ ) , p ( y ) = inf { t > 0 : y ∈ t ( C − c )} . (1) The function p p p is continuous, p ( y ) = 0 p(y)=0 p ( y ) = 0 exactly when y = 0 y=0 y = 0 , and p ( t y ) = t p ( y ) p(ty)=tp(y) p ( t y ) = tp ( y ) for t ≥ 0 t\geq0 t ≥ 0 . Moreover,
C − c = { y : p ( y ) ≤ 1 } , int ( C − c ) = { y : p ( y ) < 1 } , ∂ ( C − c ) = { y : p ( y ) = 1 } . (2) C-c=\{y:p(y)\leq1\},\qquad\operatorname{int}(C-c)=\{y:p(y)<1\},\qquad\partial(C-c)=\{y:p(y)=1\}. \tag*{(2)} C − c = { y : p ( y ) ≤ 1 } , int ( C − c ) = { y : p ( y ) < 1 } , ∂ ( C − c ) = { y : p ( y ) = 1 } . (2) (2) With S k − 1 = { a ∈ R k : ∥ a ∥ = 1 } S^{k-1}=\{a\in\mathbb{R}^{k}:\lVert a\rVert=1\} S k − 1 = { a ∈ R k : ∥ a ∥ = 1 } , the map
R : S k − 1 → ∂ C , R ( a ) = c + a p ( a ) R:S^{k-1}\to\partial C,\qquad R(a)=c+\frac{a}{p(a)} R : S k − 1 → ∂ C , R ( a ) = c + p ( a ) a is a homeomorphism. Its inverse is R − 1 ( u ) = ( u − c ) / ∥ u − c ∥ R^{-1}(u)=(u-c)/\lVert u-c\rVert R − 1 ( u ) = ( u − c ) / ∥ u − c ∥ .
Proof. (1) Put Q = C − c Q=C-c Q = C − c . Since 0 ∈ int Q 0\in\operatorname{int}Q 0 ∈ int Q and Q Q Q is compact, there are 0 < ρ ≤ R 0 0<\rho\leq R_0 0 < ρ ≤ R 0 such that
{ y : ∥ y ∥ ≤ ρ } ⊂ Q ⊂ { y : ∥ y ∥ ≤ R 0 } . \{y:\lVert y\rVert\leq\rho\}\subset Q\subset\{y:\lVert y\rVert\leq R_0\}. { y : ∥ y ∥ ≤ ρ } ⊂ Q ⊂ { y : ∥ y ∥ ≤ R 0 } .
If y ∈ t Q y\in tQ y ∈ tQ , then ∥ y ∥ ≤ t R 0 \lVert y\rVert\leq tR_0 ∥ y ∥ ≤ t R 0 . Conversely, for y ≠ 0 y\neq0 y = 0 , the vector ρ y / ∥ y ∥ \rho y/\lVert y\rVert ρ y / ∥ y ∥ belongs to Q Q Q . Consequently
∥ y ∥ R 0 ≤ p ( y ) ≤ ∥ y ∥ ρ ( y ∈ R k ) . (3) \frac{\lVert y\rVert}{R_0}\leq p(y)\leq\frac{\lVert y\rVert}{\rho}\qquad(y\in\mathbb{R}^{k}). \tag*{(3)} R 0 ∥ y ∥ ≤ p ( y ) ≤ ρ ∥ y ∥ ( y ∈ R k ) . (3)
For t > 0 t>0 t > 0 , substituting s = t r s=tr s = t r in the defining infimum gives
p ( t y ) = inf { t r : r > 0 , t y ∈ t r Q } = t inf { r > 0 : y ∈ r Q } = t p ( y ) . p(ty)=\inf\{tr:r>0,\;ty\in trQ\}=t\inf\{r>0:y\in rQ\}=tp(y). p ( t y ) = inf { t r : r > 0 , t y ∈ t r Q } = t inf { r > 0 : y ∈ r Q } = tp ( y ) .
For t = 0 t=0 t = 0 , the same identity follows from p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 .
Convexity and 0 ∈ Q 0 \in Q 0 ∈ Q imply s Q ⊂ t Q sQ \subset tQ s Q ⊂ tQ when 0 ≤ s ≤ t 0 \le s \le t 0 ≤ s ≤ t . Hence s > p ( y ) s>p(y) s > p ( y ) implies y ∈ s Q y \in sQ y ∈ s Q : there exists r ∈ ( 0 , s ) r \in(0,s) r ∈ ( 0 , s ) such that y ∈ r Q y \in rQ y ∈ r Q , and r Q ⊂ s Q rQ \subset sQ r Q ⊂ s Q . If also t > p ( z ) t>p(z) t > p ( z ) , convexity gives
y + z s + t = s s + t y s + t s + t z t ∈ Q . \frac{y+z}{s+t}=\frac{s}{s+t}\frac{y}{s}+\frac{t}{s+t}\frac{z}{t}\in Q. s + t y + z = s + t s s y + s + t t t z ∈ Q .
Thus p ( y + z ) ≤ s + t p(y+z)\le s+t p ( y + z ) ≤ s + t . Letting s ↓ p ( y ) s\downarrow p(y) s ↓ p ( y ) and t ↓ p ( z ) t\downarrow p(z) t ↓ p ( z ) proves subadditivity. In particular,
p ( y ) − p ( z ) ≤ p ( y − z ) ≤ ∥ y − z ∥ ρ , p ( z ) − p ( y ) ≤ p ( z − y ) ≤ ∥ y − z ∥ ρ . p(y)-p(z)\le p(y-z)\le\frac{\lVert y-z\rVert}{\rho},\qquad p(z)-p(y)\le p(z-y)\le\frac{\lVert y-z\rVert}{\rho}. p ( y ) − p ( z ) ≤ p ( y − z ) ≤ ρ ∥ y − z ∥ , p ( z ) − p ( y ) ≤ p ( z − y ) ≤ ρ ∥ y − z ∥ .
Therefore ∣ p ( y ) − p ( z ) ∣ ≤ ∥ y − z ∥ / ρ |p(y)-p(z)|\le\lVert y-z\rVert/\rho ∣ p ( y ) − p ( z ) ∣ ≤ ∥ y − z ∥ / ρ , proving continuity. The lower bound in (2.3) proves the assertion about zeros.
If y ∈ Q y\in Q y ∈ Q , then p ( y ) ≤ 1 p(y)\le1 p ( y ) ≤ 1 . If p ( y ) ≤ 1 p(y)\le1 p ( y ) ≤ 1 , then y / ( 1 + 1 / m ) ∈ Q y/(1+1/m)\in Q y / ( 1 + 1/ m ) ∈ Q for every m ∈ N m\in\mathbb{N} m ∈ N ; closedness of Q Q Q gives y ∈ Q y\in Q y ∈ Q after passage to the limit. Next, suppose p ( y ) < 1 p(y)<1 p ( y ) < 1 . Choose s ∈ ( p ( y ) , 1 ) s\in(p(y),1) s ∈ ( p ( y ) , 1 ) and write y = s z y=sz y = sz with z ∈ Q z\in Q z ∈ Q . For ∥ b ∥ < ( 1 − s ) ρ \lVert b\rVert<(1-s)\rho ∥ b ∥ < ( 1 − s ) ρ ,
y + b = s z + ( 1 − s ) b 1 − s ∈ Q , y+b=sz+(1-s)\frac{b}{1-s}\in Q, y + b = sz + ( 1 − s ) 1 − s b ∈ Q ,
because z ∈ Q z\in Q z ∈ Q and b / ( 1 − s ) ∈ Q b/(1-s)\in Q b / ( 1 − s ) ∈ Q . Thus y y y lies in the interior of Q Q Q . Conversely, if 0 ≠ y ∈ int Q 0\ne y\in\operatorname{int}Q 0 = y ∈ int Q , then ( 1 + δ ) y ∈ Q (1+\delta)y\in Q ( 1 + δ ) y ∈ Q for some δ > 0 \delta>0 δ > 0 , whence p ( y ) ≤ ( 1 + δ ) − 1 < 1 p(y)\le(1+\delta)^{-1}<1 p ( y ) ≤ ( 1 + δ ) − 1 < 1 by homogeneity. For y = 0 y=0 y = 0 , this follows from p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 . Since Q Q Q is closed, ∂ Q = Q ∖ int Q \partial Q=Q\setminus\operatorname{int}Q ∂ Q = Q ∖ int Q ; this proves all of (2.2).
(2) For a ∈ S k − 1 a\in S^{k-1} a ∈ S k − 1 , the lower bound in (2.3) gives p ( a ) > 0 p(a)>0 p ( a ) > 0 , and homogeneity gives p ( a / p ( a ) ) = 1 p(a/p(a))=1 p ( a / p ( a )) = 1 . Hence R R R is well defined and continuous. For u ∈ ∂ C u\in\partial C u ∈ ∂ C , let a = ( u − c ) / ∥ u − c ∥ a=(u-c)/\lVert u-c\rVert a = ( u − c ) / ∥ u − c ∥ . Since p ( u − c ) = 1 p(u-c)=1 p ( u − c ) = 1 , we obtain
p ( a ) = 1 ∥ u − c ∥ , R ( a ) = c + ( u − c ) / ∥ u − c ∥ 1 / ∥ u − c ∥ = u . p(a)=\frac{1}{\lVert u-c\rVert},\qquad R(a)=c+\frac{(u-c)/\lVert u-c\rVert}{1/\lVert u-c\rVert}=u. p ( a ) = ∥ u − c ∥ 1 , R ( a ) = c + 1/ ∥ u − c ∥ ( u − c ) / ∥ u − c ∥ = u .
Also ( R ( a ) − c ) / ∥ R ( a ) − c ∥ = a (R(a)-c)/\lVert R(a)-c\rVert=a ( R ( a ) − c ) / ∥ R ( a ) − c ∥ = a , since p ( a ) > 0 p(a)>0 p ( a ) > 0 . The displayed inverse is continuous, so R R R is a homeomorphism. □
For a finite-dimensional admissible subspace of C 0 ( X ) C_0(X) C 0 ( X ) , extension by zero adds an evaluation vector at infinity. The argument below distinguishes whether this vector is an interior or boundary point of the convex hull; the boundary case may produce a segment.
Lemma 2.3. Let X X X be locally compact Hausdorff with at least two points, and let f 1 , … , f k f_1,\ldots,f_k f 1 , … , f k be a basis of a subspace E ⊂ C ^ 0 ( X ) ∪ { 0 } E\subset\widehat{C}_0(X)\cup\{0\} E ⊂ C 0 ( X ) ∪ { 0 } , where k ∈ N k\in\mathbb{N} k ∈ N . Take L = X L=X L = X when X X X is compact, and otherwise take L = X ∪ { ∞ } L=X\cup\{\infty\} L = X ∪ { ∞ } , the one-point compactification, whose neighborhoods of ∞ \infty ∞ are the sets { ∞ } ∪ ( X ∖ H ) \{\infty\}\cup(X\setminus H) { ∞ } ∪ ( X ∖ H ) with H ⊂ X H\subset X H ⊂ X compact. Extend each f j f_j f j by zero at ∞ \infty ∞ . Define
F : L → R k , F ( x ) = ( f 1 ( x ) , … , f k ( x ) ) , C = conv F ( L ) . F:L\to\mathbb{R}^k,\qquad F(x)=(f_1(x),\ldots,f_k(x)),\qquad C=\operatorname{conv}F(L). F : L → R k , F ( x ) = ( f 1 ( x ) , … , f k ( x )) , C = conv F ( L ) . For a ≠ 0 a\ne0 a = 0 , write q a ( x ) = ⟨ a , F ( x ) ⟩ q_a(x)=\langle a,F(x)\rangle q a ( x ) = ⟨ a , F ( x )⟩ on L L L , and use H C ( a ) H_C(a) H C ( a ) and C a C_a C a as in Lemma 2.1.
(1) The set C C C is compact with nonempty interior, and H C ( a ) = max X q a H_C(a)=\max_X q_a H C ( a ) = max X q a . If X X X is noncompact, then 0 ∈ C 0\in C 0 ∈ C and H C ( a ) ≥ 0 H_C(a)\ge0 H C ( a ) ≥ 0 for all a ≠ 0 a\ne0 a = 0 .
(2) If X X X is compact, or if X X X is noncompact and 0 ∈ int C 0\in\operatorname{int}C 0 ∈ int C , then, for Z = F − 1 ( ∂ C ) Z=F^{-1}(\partial C) Z = F − 1 ( ∂ C ) , the restriction F ∣ Z : Z → ∂ C F|_Z:Z\to\partial C F ∣ Z : Z → ∂ C is a homeomorphism and Z ⊂ X Z\subset X Z ⊂ X .
(3) Suppose X X X is noncompact and 0 ∈ ∂ C 0\in\partial C 0 ∈ ∂ C . There is a point x 0 ∈ X x_0\in X x 0 ∈ X which is the unique zero of every nonzero nonpositive function in E E E . Put A = [ 0 , F ( x 0 ) ] = { t F ( x 0 ) : 0 ≤ t ≤ 1 } A=[0,F(x_0)]=\{tF(x_0):0\le t\le1\} A = [ 0 , F ( x 0 )] = { tF ( x 0 ) : 0 ≤ t ≤ 1 } . Then C a = A C_a=A C a = A whenever H C ( a ) = 0 H_C(a)=0 H C ( a ) = 0 , F − 1 ( A ) = { x 0 , ∞ } F^{-1}(A)=\{x_0,\infty\} F − 1 ( A ) = { x 0 , ∞ } , and every point of ∂ C ∖ A \partial C\setminus A ∂ C ∖ A has exactly one preimage, lying in X X X .
Proof. (1) The space L L L is compact Hausdorff, and the coordinate functions of F F F are continuous on L L L . In the noncompact case, their continuity at ∞ \infty ∞ follows from f j ∈ C 0 ( X ) f_j\in C_0(X) f j ∈ C 0 ( X ) and the definition f j ( ∞ ) = 0 f_j(\infty)=0 f j ( ∞ ) = 0 . Thus F F F is continuous, and Lemma 2.1(1) shows that C = conv F ( L ) C=\operatorname{conv}F(L) C = conv F ( L ) is compact. We prove that int C ≠ ∅ \operatorname{int}C\ne\varnothing int C = ∅ . If the affine span of C C C were all of R k \mathbb{R}^k R k , then C C C would contain k + 1 k+1 k + 1 affinely independent points. By convexity, it would contain their convex hull, which has nonempty interior in R k \mathbb{R}^k R k . Consequently, if C C C had empty interior, its affine span would be a proper affine subspace. There would then exist a ∈ R k ∖ { 0 } a \in\mathbb{R}^k \setminus\{0\} a ∈ R k ∖ { 0 } and b ∈ R b \in\mathbb{R} b ∈ R such that
∑ j = 1 k a j f j ( x ) = b ( x ∈ L ) . (4) \sum_{j=1}^{k} a_j f_j(x) = b \qquad(x \in L). \tag*{(4)} j = 1 ∑ k a j f j ( x ) = b ( x ∈ L ) . (4)
If b = 0 b = 0 b = 0 , restriction to X X X contradicts the linear independence of f 1 , … , f k f_1,\ldots,f_k f 1 , … , f k . If b ≠ 0 b \ne0 b = 0 , that restriction is a nonzero constant belonging to E E E . Since X X X has at least two points, such a function does not have a unique maximizer, again a contradiction. Therefore int C ≠ ∅ \operatorname{int} C \ne\varnothing int C = ∅ .
Fix a ∈ R k ∖ { 0 } a \in\mathbb{R}^k \setminus\{0\} a ∈ R k ∖ { 0 } . The restriction q a ∣ X q_a|_X q a ∣ X is nonzero by linear independence, so it attains its maximum at a unique point, which we denote by x a x_a x a . Lemma 2.1 (1) gives H C ( a ) = max L q a H_C(a) = \max_L q_a H C ( a ) = max L q a . If X X X is compact, then L = X L = X L = X , and hence H C ( a ) = max X q a H_C(a) = \max_X q_a H C ( a ) = max X q a .
Suppose now that X X X is noncompact. For every ε > 0 \varepsilon> 0 ε > 0 , the set
K ε = { x ∈ X : ∣ q a ( x ) ∣ ≥ ε } K_\varepsilon= \{x \in X : |q_a(x)| \ge\varepsilon\} K ε = { x ∈ X : ∣ q a ( x ) ∣ ≥ ε }
is compact because q a ∣ X ∈ C 0 ( X ) q_a|_X \in C_0(X) q a ∣ X ∈ C 0 ( X ) . It cannot equal the noncompact space X X X . Choosing x ∈ X ∖ K ε x \in X \setminus K_\varepsilon x ∈ X ∖ K ε gives sup X q a ≥ q a ( x ) > − ε \sup_X q_a \ge q_a(x) > -\varepsilon sup X q a ≥ q a ( x ) > − ε . Since ε > 0 \varepsilon> 0 ε > 0 is arbitrary, sup X q a ≥ 0 \sup_X q_a \ge0 sup X q a ≥ 0 . This supremum is attained at x a x_a x a . Adding the value q a ( ∞ ) = 0 q_a(\infty) = 0 q a ( ∞ ) = 0 therefore does not change it, and we obtain
H C ( a ) = max L q a = max X q a ≥ 0. H_C(a) = \max_L q_a = \max_X q_a \ge0. H C ( a ) = L max q a = X max q a ≥ 0.
Also, F ( ∞ ) = 0 F(\infty) = 0 F ( ∞ ) = 0 , so 0 ∈ C 0 \in C 0 ∈ C .
In the noncompact case, the same argument gives the following relation between the maximizing sets on X X X and on L L L :
Argmax L q a = { { x a } , H C ( a ) > 0 , { x a , ∞ } , H C ( a ) = 0. (5) \operatorname{Argmax}_L q_a =
\begin{cases}
\{x_a\}, & H_C(a) > 0, \\
\{x_a,\infty\}, & H_C(a) = 0.
\end{cases} \tag*{(5)} Argmax L q a = { { x a } , { x a , ∞ } , H C ( a ) > 0 , H C ( a ) = 0. (5)
Indeed, x a x_a x a is the unique maximizer in X X X , and the only additional point of L L L is ∞ \infty ∞ , where q a q_a q a has value zero.
(2) Suppose that X X X is compact, or that X X X is noncompact and 0 ∈ int C 0 \in\operatorname{int} C 0 ∈ int C .
In the compact case, every q a q_a q a , with a ≠ 0 a \ne0 a = 0 , has a unique maximizer on L = X L = X L = X . In the noncompact case, choose ρ > 0 \rho> 0 ρ > 0 such that { u : ∥ u ∥ ≤ ρ } ⊂ C \{u : \|u\| \le\rho\} \subset C { u : ∥ u ∥ ≤ ρ } ⊂ C . Then
H C ( a ) ≥ ⟨ a , ρ a ∥ a ∥ ⟩ = ρ ∥ a ∥ > 0 = q a ( ∞ ) . H_C(a) \ge\left\langle a,\rho\frac{a}{\|a\|}\right\rangle= \rho\|a\| > 0 = q_a(\infty). H C ( a ) ≥ ⟨ a , ρ ∥ a ∥ a ⟩ = ρ ∥ a ∥ > 0 = q a ( ∞ ) .
Thus (2.5) shows that q a q_a q a also has a unique maximizer on L L L , and that this maximizer belongs to X X X . In either case, Lemma 2.1 (2) yields
C a = { F ( x a ) } , F − 1 ( { F ( x a ) } ) = { x a } . (6) C_a = \{F(x_a)\}, \qquad F^{-1}(\{F(x_a)\}) = \{x_a\}. \tag*{(6)} C a = { F ( x a )} , F − 1 ({ F ( x a )}) = { x a } . (6)
Let u ∈ ∂ C u \in\partial C u ∈ ∂ C . The supporting hyperplane theorem gives a ≠ 0 a \ne0 a = 0 such that ⟨ a , v ⟩ ≤ ⟨ a , u ⟩ \langle a,v\rangle\le\langle a,u\rangle ⟨ a , v ⟩ ≤ ⟨ a , u ⟩ for every v ∈ C v \in C v ∈ C . Equivalently, u ∈ C a u \in C_a u ∈ C a . By (2.6), u = F ( x a ) u = F(x_a) u = F ( x a ) and x a x_a x a is its unique preimage in L L L . Consequently, for Z = F − 1 ( ∂ C ) Z = F^{-1}(\partial C) Z = F − 1 ( ∂ C ) , the restriction F ∣ Z : Z → ∂ C F|_Z : Z \to\partial C F ∣ Z : Z → ∂ C is a continuous bijection.
The set Z Z Z is closed in compact L L L , because F F F is continuous and ∂ C \partial C ∂ C is closed. Thus Z Z Z is compact. Since ∂ C \partial C ∂ C is Hausdorff, the continuous bijection F ∣ Z F|_Z F ∣ Z has a continuous inverse. Finally, Z ⊂ X Z \subset X Z ⊂ X : this is immediate when L = X L = X L = X , while in the noncompact case F ( ∞ ) = 0 ∈ int C F(\infty) = 0 \in\operatorname{int} C F ( ∞ ) = 0 ∈ int C implies ∞ ∉ Z \infty\notin Z ∞ ∈ / Z .
(3) Suppose that X X X is noncompact and 0 ∈ ∂ C 0 \in\partial C 0 ∈ ∂ C . Every nonzero nonpositive function in E E E has the same unique zero.
A supporting hyperplane at 0 0 0 gives a 0 ∈ R k ∖ { 0 } a_0 \in\mathbb{R}^k \setminus\{0\} a 0 ∈ R k ∖ { 0 } such that ⟨ a 0 , u ⟩ ≤ 0 \langle a_0,u\rangle\le0 ⟨ a 0 , u ⟩ ≤ 0 for every u ∈ C u \in C u ∈ C . Since 0 ∈ C 0 \in C 0 ∈ C , we have H C ( a 0 ) = 0 H_C(a_0) = 0 H C ( a 0 ) = 0 . Therefore the set
P = { f ∈ E ∖ { 0 } : f ( x ) ≤ 0 for every x ∈ X } \mathcal{P} = \{f \in E \setminus\{0\} : f(x) \le0 \text{ for every } x \in X\} P = { f ∈ E ∖ { 0 } : f ( x ) ≤ 0 for every x ∈ X }
is nonempty, as it contains q a 0 ∣ X q_{a_0}|_X q a 0 ∣ X . By part (1), every f ∈ P f \in\mathcal{P} f ∈ P has maximum zero. Since f f f has a unique maximizer, it has exactly one zero.
Let f , g ∈ P f,g \in\mathcal{P} f , g ∈ P . Their sum is nonzero: if f + g = 0 f+g=0 f + g = 0 , then f = − g f=-g f = − g , and the nonpositivity of both functions would force f = g = 0 f=g=0 f = g = 0 . Hence f + g ∈ P f+g \in\mathcal{P} f + g ∈ P , so there exists x ∈ X x \in X x ∈ X such that ( f + g ) ( x ) = 0 (f+g)(x)=0 ( f + g ) ( x ) = 0 . As f ( x ) ≤ 0 f(x)\le0 f ( x ) ≤ 0 and g ( x ) ≤ 0 g(x)\le0 g ( x ) ≤ 0 , this equality implies f ( x ) = g ( x ) = 0 f(x)=g(x)=0 f ( x ) = g ( x ) = 0 . The unique zeros of f f f and g g g therefore coincide. Fixing one member of P \mathcal{P} P and comparing it with every other member gives a point x 0 ∈ X x_0\in X x 0 ∈ X which is the unique zero of every function in P \mathcal{P} P .
With A = [ 0 , F ( x 0 ) ] A=[0,F(x_0)] A = [ 0 , F ( x 0 )] , we have
C a = A whenever H C ( a ) = 0 , F − 1 ( A ) = { x 0 , ∞ } . C_a=A\quad\text{whenever }H_C(a)=0,\qquad F^{-1}(A)=\{x_0,\infty\}. C a = A whenever H C ( a ) = 0 , F − 1 ( A ) = { x 0 , ∞ } .
Suppose a ≠ 0 a\ne0 a = 0 and H C ( a ) = 0 H_C(a)=0 H C ( a ) = 0 . Then q a ∣ X ∈ P q_a|_X\in\mathcal{P} q a ∣ X ∈ P , so its unique maximizer is x 0 x_0 x 0 . By (5) , Argmax L q a = { x 0 , ∞ } \operatorname{Argmax}_L q_a=\{x_0,\infty\} Argmax L q a = { x 0 , ∞ } . Applying the second equality in (1) and using F ( ∞ ) = 0 F(\infty)=0 F ( ∞ ) = 0 , we obtain
C a = conv F ( { x 0 , ∞ } ) = conv { F ( x 0 ) , 0 } = [ 0 , F ( x 0 ) ] = A . C_a=\operatorname{conv}F(\{x_0,\infty\})=\operatorname{conv}\{F(x_0),0\}=[0,F(x_0)]=A. C a = conv F ({ x 0 , ∞ }) = conv { F ( x 0 ) , 0 } = [ 0 , F ( x 0 )] = A .
In particular, A ⊂ ∂ C A\subset\partial C A ⊂ ∂ C . Indeed, every point of A = C a A=C_a A = C a maximizes the nonzero functional u ↦ ⟨ a , u ⟩ u\mapsto\langle a,u\rangle u ↦ ⟨ a , u ⟩ on C C C , whereas no interior point can do so: a sufficiently small displacement in the direction a a a would remain in C C C and increase the functional.
To determine F − 1 ( A ) F^{-1}(A) F − 1 ( A ) , use the fixed direction a 0 a_0 a 0 chosen above. Since q a 0 ( x 0 ) = 0 q_{a_0}(x_0)=0 q a 0 ( x 0 ) = 0 , the functional u ↦ ⟨ a 0 , u ⟩ u\mapsto\langle a_0,u\rangle u ↦ ⟨ a 0 , u ⟩ vanishes at both endpoints of A A A , and therefore throughout A A A . If x ∈ X x\in X x ∈ X and F ( x ) ∈ A F(x)\in A F ( x ) ∈ A , then q a 0 ( x ) = ⟨ a 0 , F ( x ) ⟩ = 0 q_{a_0}(x)=\langle a_0,F(x)\rangle=0 q a 0 ( x ) = ⟨ a 0 , F ( x )⟩ = 0 . The uniqueness of the zero of q a 0 ∣ X q_{a_0}|_X q a 0 ∣ X gives x = x 0 x=x_0 x = x 0 . Conversely, F ( x 0 ) ∈ A F(x_0)\in A F ( x 0 ) ∈ A and F ( ∞ ) = 0 ∈ A F(\infty)=0\in A F ( ∞ ) = 0 ∈ A . Thus
F − 1 ( A ) = { x 0 , ∞ } . F^{-1}(A)=\{x_0,\infty\}. F − 1 ( A ) = { x 0 , ∞ } .
In particular,
F ( L ) ∩ A = { 0 , F ( x 0 ) } . F(L)\cap A=\{0,F(x_0)\}. F ( L ) ∩ A = { 0 , F ( x 0 )} .
Thus, if F ( x 0 ) ≠ 0 F(x_0)\ne0 F ( x 0 ) = 0 , the points t F ( x 0 ) tF(x_0) tF ( x 0 ) with 0 < t < 1 0<t<1 0 < t < 1 belong to C C C but not to F ( L ) F(L) F ( L ) ; they arise as convex combinations of the vectors 0 = F ( ∞ ) 0=F(\infty) 0 = F ( ∞ ) and F ( x 0 ) F(x_0) F ( x 0 ) .
Finally, let u ∈ ∂ C ∖ A u\in\partial C\setminus A u ∈ ∂ C ∖ A . By the supporting hyperplane theorem, choose a ≠ 0 a\ne0 a = 0 such that ⟨ a , v ⟩ ≤ ⟨ a , u ⟩ \langle a,v\rangle\le\langle a,u\rangle ⟨ a , v ⟩ ≤ ⟨ a , u ⟩ for every v ∈ C v\in C v ∈ C , so that u ∈ C a u\in C_a u ∈ C a . Part (1) gives H C ( a ) ≥ 0 H_C(a)\ge0 H C ( a ) ≥ 0 . If H C ( a ) = 0 H_C(a)=0 H C ( a ) = 0 , the equality just proved would give C a = A C_a=A C a = A , contradicting u ∉ A u\notin A u ∈ / A . Hence H C ( a ) > 0 H_C(a)>0 H C ( a ) > 0 . By (5) , q a q_a q a has exactly one maximizer on L L L , and this point belongs to X X X . Lemma 2.1(2) now shows that u u u has exactly one preimage in L L L , lying in X X X . □
Adding a basis vector adds a coordinate to the evaluation map. The resulting convex body projects onto the original one; continuous sections will be obtained from the endpoints of its vertical fibers.
Lemma 2.4. Let k ∈ N k\in\mathbb{N} k ∈ N and let D ⊂ R k × R D\subset\mathbb{R}^k\times\mathbb{R} D ⊂ R k × R be compact and convex with nonempty interior. Let π : R k × R → R k \pi:\mathbb{R}^k\times\mathbb{R}\to\mathbb{R}^k π : R k × R → R k be the projection π ( u , z ) = u \pi(u,z)=u π ( u , z ) = u , and put C = π ( D ) C=\pi(D) C = π ( D ) . For u ∈ int C u\in\operatorname{int}C u ∈ int C , define
z − ( u ) = min { z : ( u , z ) ∈ D } , z + ( u ) = max { z : ( u , z ) ∈ D } . z_-(u)=\min\{z:(u,z)\in D\},\qquad z_+(u)=\max\{z:(u,z)\in D\}. z − ( u ) = min { z : ( u , z ) ∈ D } , z + ( u ) = max { z : ( u , z ) ∈ D } . (1) The functions z − , z + : int C → R z_-,z_+:\operatorname{int}C\to\mathbb{R} z − , z + : int C → R are continuous, and z − ( u ) < z + ( u ) z_-(u)<z_+(u) z − ( u ) < z + ( u ) with $(u,z_\pm(u))\in\partial D.
(2) Let ℓ : R k × R → R \ell:\mathbb{R}^k\times\mathbb{R}\to\mathbb{R} ℓ : R k × R → R be a nonzero linear functional, ℓ ( u , z ) = ⟨ a , u ⟩ + b z \ell(u,z)=\langle a,u\rangle+bz ℓ ( u , z ) = ⟨ a , u ⟩ + b z , with ℓ ≤ 0 \ell\le0 ℓ ≤ 0 on D D D . Then ℓ ( u , z − ( u ) ) < 0 \ell(u,z_-(u))<0 ℓ ( u , z − ( u )) < 0 when b > 0 b>0 b > 0 , ℓ ( u , z + ( u ) ) < 0 \ell(u,z_+(u))<0 ℓ ( u , z + ( u )) < 0 when b < 0 b<0 b < 0 , and both inequalities hold when b = 0 b=0 b = 0 .
Proof. (1) For each u ∈ C u\in C u ∈ C , the set I u = { z ∈ R : ( u , z ) ∈ D } I_u=\{z\in\mathbb{R}:(u,z)\in D\} I u = { z ∈ R : ( u , z ) ∈ D } is nonempty, compact and convex, hence a closed interval. Its endpoints are z − ( u ) z_-(u) z − ( u ) and z + ( u ) z_+(u) z + ( u ) . For u , v ∈ int C u,v\in\operatorname{int}C u , v ∈ int C and 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , convexity of D D D gives
( ( 1 − t ) u + t v , ( 1 − t ) z − ( u ) + t z − ( v ) ) ∈ D . ((1-t)u+tv,(1-t)z_-(u)+tz_-(v))\in D. (( 1 − t ) u + t v , ( 1 − t ) z − ( u ) + t z − ( v )) ∈ D .
Taking the minimum of I ( 1 − t ) u + t v I_{(1-t)u+tv} I ( 1 − t ) u + t v yields z − ( ( 1 − t ) u + t v ) ≤ ( 1 − t ) z − ( u ) + t z − ( v ) z_-((1-t)u+tv)\le(1-t)z_-(u)+tz_-(v) z − (( 1 − t ) u + t v ) ≤ ( 1 − t ) z − ( u ) + t z − ( v ) . Taking upper endpoints instead gives the reverse inequality for z + z_+ z + . Thus both z − z_- z − and − z + -z_+ − z + are convex. They are bounded because D D D is compact.
Let φ \varphi φ denote either of these two convex functions, and choose M 0 > 0 M_0>0 M 0 > 0 such that ∣ φ ∣ ≤ M 0 |\varphi|\le M_0 ∣ φ ∣ ≤ M 0 on int C \operatorname{int}C int C . Fix u ∈ int C u\in\operatorname{int}C u ∈ int C and choose ρ > 0 \rho>0 ρ > 0 with { v : ∥ v − u ∥ ≤ ρ } ⊂ int C \{v:\|v-u\|\le\rho\}\subset\operatorname{int}C { v : ∥ v − u ∥ ≤ ρ } ⊂ int C . For 0 < η = ∥ v − u ∥ < ρ 0<\eta=\|v-u\|<\rho 0 < η = ∥ v − u ∥ < ρ , put e = ( v − u ) / η e=(v-u)/\eta e = ( v − u ) / η . The identity v = ( 1 − η / ρ ) u + ( η / ρ ) ( u + ρ e ) v=(1-\eta/\rho)u+(\eta/\rho)(u+\rho e) v = ( 1 − η / ρ ) u + ( η / ρ ) ( u + ρ e ) gives
φ ( v ) − φ ( u ) ≤ η ρ ( φ ( u + ρ e ) − φ ( u ) ) ≤ 2 M 0 η ρ . \varphi(v)-\varphi(u)\le\frac{\eta}{\rho}\bigl(\varphi(u+\rho e)-\varphi(u)\bigr)\le\frac{2M_0\eta}{\rho}. φ ( v ) − φ ( u ) ≤ ρ η ( φ ( u + ρ e ) − φ ( u ) ) ≤ ρ 2 M 0 η .
Applying convexity to u = ρ ρ + η v + η ρ + η ( u − ρ e ) u = \frac{\rho}{\rho+\eta}v + \frac{\eta}{\rho+\eta}(u-\rho e) u = ρ + η ρ v + ρ + η η ( u − ρ e ) gives
( ρ + η ) φ ( u ) ≤ ρ φ ( v ) + η φ ( u − ρ e ) , (\rho+\eta)\varphi(u) \leq\rho\varphi(v) + \eta\varphi(u-\rho e), ( ρ + η ) φ ( u ) ≤ ρφ ( v ) + η φ ( u − ρ e ) ,
and therefore
φ ( v ) − φ ( u ) ≥ η ρ ( φ ( u ) − φ ( u − ρ e ) ) ≥ − 2 M 0 η ρ . \varphi(v)-\varphi(u) \geq\frac{\eta}{\rho}\bigl(\varphi(u)-\varphi(u-\rho e)\bigr) \geq-\frac{2M_0\eta}{\rho}. φ ( v ) − φ ( u ) ≥ ρ η ( φ ( u ) − φ ( u − ρ e ) ) ≥ − ρ 2 M 0 η .
These two estimates prove continuity.
To show that z − ( u ) < z + ( u ) z_-(u)<z_+(u) z − ( u ) < z + ( u ) for u ∈ int C u\in\operatorname{int}C u ∈ int C , choose d ∗ ∈ int D d_*\in\operatorname{int}D d ∗ ∈ int D and ϵ > 0 \epsilon>0 ϵ > 0 so that the open ball of radius ϵ \epsilon ϵ about d ∗ d_* d ∗ lies in D D D . Write u ∗ = π ( d ∗ ) u_*=\pi(d_*) u ∗ = π ( d ∗ ) . Since u ′ = ( u − s u ∗ ) / ( 1 − s ) u'=(u-su_*)/(1-s) u ′ = ( u − s u ∗ ) / ( 1 − s ) converges to u ∈ int C u\in\operatorname{int}C u ∈ int C as s ↓ 0 s\downarrow0 s ↓ 0 , choose s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) so small that u ′ ∈ C u'\in C u ′ ∈ C , and then choose d ′ ∈ D d'\in D d ′ ∈ D with π ( d ′ ) = u ′ \pi(d')=u' π ( d ′ ) = u ′ . For d = s d ∗ + ( 1 − s ) d ′ d=sd_*+(1-s)d' d = s d ∗ + ( 1 − s ) d ′ we have π ( d ) = u \pi(d)=u π ( d ) = u . If ∥ b ∥ < s ϵ \lVert b\rVert<s\epsilon ∥ b ∥ < sϵ , then
d + b = s ( d ∗ + b s ) + ( 1 − s ) d ′ ∈ D . d+b=s\left(d_*+\frac{b}{s}\right)+(1-s)d'\in D. d + b = s ( d ∗ + s b ) + ( 1 − s ) d ′ ∈ D .
Writing d = ( u , z ) d=(u,z) d = ( u , z ) , choose 0 < δ < s ϵ 0<\delta<s\epsilon 0 < δ < sϵ . Then ( u , z − δ ) , ( u , z + δ ) ∈ D (u,z-\delta),(u,z+\delta)\in D ( u , z − δ ) , ( u , z + δ ) ∈ D , so z − ( u ) ≤ z − δ < z + δ ≤ z + ( u ) z_-(u)\leq z-\delta<z+\delta\leq z_+(u) z − ( u ) ≤ z − δ < z + δ ≤ z + ( u ) . Neither ( u , z − ( u ) ) (u,z_-(u)) ( u , z − ( u )) nor ( u , z + ( u ) ) (u,z_+(u)) ( u , z + ( u )) can lie in int D \operatorname{int}D int D : an open ball about either point would contain points of D D D with the same first coordinate and a last coordinate beyond the corresponding minimum or maximum. Since D D D is closed, both points belong to ∂ D \partial D ∂ D .
(2) For b > 0 b>0 b > 0 , part (1) gives
ℓ ( u , z − ( u ) ) = ℓ ( u , z + ( u ) ) − b ( z + ( u ) − z − ( u ) ) < 0. \ell(u,z_-(u))=\ell(u,z_+(u))-b\bigl(z_+(u)-z_-(u)\bigr)<0. ℓ ( u , z − ( u )) = ℓ ( u , z + ( u )) − b ( z + ( u ) − z − ( u ) ) < 0.
For b < 0 b<0 b < 0 , it gives
ℓ ( u , z + ( u ) ) = ℓ ( u , z − ( u ) ) + b ( z + ( u ) − z − ( u ) ) < 0. \ell(u,z_+(u))=\ell(u,z_-(u))+b\bigl(z_+(u)-z_-(u)\bigr)<0. ℓ ( u , z + ( u )) = ℓ ( u , z − ( u )) + b ( z + ( u ) − z − ( u ) ) < 0.
If b = 0 b=0 b = 0 , then a ≠ 0 a\neq0 a = 0 and ⟨ a , u ⟩ ≤ 0 \langle a,u\rangle\leq0 ⟨ a , u ⟩ ≤ 0 on C C C . Equality at an interior point u u u is impossible: for sufficiently small t > 0 t>0 t > 0 , the point u + t a ∈ C u+ta\in C u + t a ∈ C would satisfy ⟨ a , u + t a ⟩ = t ∥ a ∥ 2 > 0 \langle a,u+ta\rangle=t\lVert a\rVert^2>0 ⟨ a , u + t a ⟩ = t ∥ a ∥ 2 > 0 . Hence ℓ ( u , z − ( u ) ) = ℓ ( u , z + ( u ) ) = ⟨ a , u ⟩ < 0 \ell(u,z_-(u))=\ell(u,z_+(u))=\langle a,u\rangle<0 ℓ ( u , z − ( u )) = ℓ ( u , z + ( u )) = ⟨ a , u ⟩ < 0 . □
Sections forced by proper extensions
For F : X → R k F:X\to\mathbb{R}^k F : X → R k and U ⊂ R k U\subset\mathbb{R}^k U ⊂ R k , a section of F F F on U U U is a continuous map s : U → X s:U\to X s : U → X satisfying F ( s ( u ) ) = u F(s(u))=u F ( s ( u )) = u for all u ∈ U u\in U u ∈ U . A proper admissible extension yields such a section on int C \operatorname{int}C int C in the countably compact case and on a radial interval in the locally compact case.
Lemma 2.5. Let X X X be countably compact Hausdorff with at least two points, and let f 1 , … , f k f_1,\ldots,f_k f 1 , … , f k be a basis of an admissible subspace of C ( X ) C(X) C ( X ) . Put F = ( f 1 , … , f k ) F=(f_1,\ldots,f_k) F = ( f 1 , … , f k ) and C = conv F ( X ) C=\operatorname{conv}F(X) C = conv F ( X ) . Then C C C is compact with nonempty interior, and, for Z = F − 1 ( ∂ C ) Z=F^{-1}(\partial C) Z = F − 1 ( ∂ C ) , the map F ∣ Z : Z → ∂ C F|_Z:Z\to\partial C F ∣ Z : Z → ∂ C is a homeomorphism. If this subspace has a proper admissible extension in C ( X ) C(X) C ( X ) , there is a continuous s : int C → X s:\operatorname{int}C\to X s : int C → X satisfying F ( s ( u ) ) = u F(s(u))=u F ( s ( u )) = u .
Proof. A continuous image of a countably compact space is countably compact, and a metrizable countably compact space is compact. Hence F ( X ) F(X) F ( X ) is compact, and so is its convex hull by Lemma 2.1. If C C C had empty interior, it would lie in an affine hyperplane: ∑ j = 1 k a j f j = c \sum_{j=1}^k a_jf_j=c ∑ j = 1 k a j f j = c for some a ≠ 0 a\neq0 a = 0 and c ∈ R c\in\mathbb{R} c ∈ R . For c = 0 c=0 c = 0 this contradicts linear independence; for c ≠ 0 c\neq0 c = 0 it gives a nonzero constant in an admissible subspace. Thus int C ≠ ∅ \operatorname{int}C\neq\varnothing int C = ∅ .
For each a ≠ 0 a\neq0 a = 0 , let x a x_a x a be the unique maximizer of ⟨ a , F ⟩ \langle a,F\rangle ⟨ a , F ⟩ . Lemma 2.1 gives
{ y ∈ C : ⟨ a , y ⟩ = H C ( a ) } = { F ( x a ) } , F − 1 ( { F ( x a ) } ) = { x a } . \{y\in C:\langle a,y\rangle=H_C(a)\}=\{F(x_a)\},\qquad F^{-1}(\{F(x_a)\})=\{x_a\}. { y ∈ C : ⟨ a , y ⟩ = H C ( a )} = { F ( x a )} , F − 1 ({ F ( x a )}) = { x a } .
Every boundary point of C C C lies on a supporting hyperplane. Consequently, F ∣ Z F|_Z F ∣ Z is a continuous bijection onto ∂ C \partial C ∂ C . If H H H is closed in Z Z Z , then it is closed in X X X , because Z Z Z is closed in X X X . Thus H H H is countably compact and its metrizable image F ( H ) F(H) F ( H ) is compact, hence closed in ∂ C \partial C ∂ C . This makes F ∣ Z F|_Z F ∣ Z a closed bijection, proving the homeomorphism.
Suppose that g ∉ span { f 1 , … , f k } g\notin\operatorname{span}\{f_1,\ldots,f_k\} g ∈ / span { f 1 , … , f k } gives an admissible extension. Apply the preceding conclusions to Ψ = ( F , g ) : X → R k + 1 \Psi=(F,g):X\to\mathbb{R}^{k+1} Ψ = ( F , g ) : X → R k + 1 and D = conv Ψ ( X ) D=\operatorname{conv}\Psi(X) D = conv Ψ ( X ) . Then D D D is compact with nonempty interior and Ψ ∣ Z D : Z D → ∂ D \Psi|_{Z_D}:Z_D\to\partial D Ψ ∣ Z D : Z D → ∂ D is a homeomorphism, where Z D = Ψ − 1 ( ∂ D ) Z_D=\Psi^{-1}(\partial D) Z D = Ψ − 1 ( ∂ D ) . For π ( u , z ) = u \pi(u,z)=u π ( u , z ) = u , we have π ( D ) = C \pi(D)=C π ( D ) = C . By Lemma 2.4, the map u ↦ ( u , z + ( u ) ) u\mapsto(u,z_{+}(u)) u ↦ ( u , z + ( u )) is continuous from int C \operatorname{int} C int C into ∂ D \partial D ∂ D . Therefore
s ( u ) = ( Ψ ∣ Z D ) − 1 ( u , z + ( u ) ) , F ( s ( u ) ) = π ( Ψ ( s ( u ) ) ) = u s(u)=(\Psi|_{Z_D})^{-1}(u,z_{+}(u)),\qquad F(s(u))=\pi(\Psi(s(u)))=u s ( u ) = ( Ψ ∣ Z D ) − 1 ( u , z + ( u )) , F ( s ( u )) = π ( Ψ ( s ( u ))) = u
defines the required continuous section. □
For functions in C 0 ( X ) C_{0}(X) C 0 ( X ) , the sets { x : ∥ F ( x ) ∥ ≥ δ } \{x:\lVert F(x)\rVert\ge\delta\} { x : ∥ F ( x )∥ ≥ δ } are compact for every δ > 0 \delta>0 δ > 0 . This gives the compactness needed for the radial section.
Lemma 2.6. Let X X X be noncompact locally compact Hausdorff, and let E ⊂ C ^ 0 ( X ) ∪ { 0 } E\subset\widehat{C}_{0}(X)\cup\{0\} E ⊂ C 0 ( X ) ∪ { 0 } have basis f 1 , … , f k f_{1},\ldots,f_{k} f 1 , … , f k , where k ∈ N k\in\mathbb{N} k ∈ N . Let L , F , C L,F,C L , F , C be as in Lemma 2.3. Choose c ∈ int C c\in\operatorname{int} C c ∈ int C , taking c = 0 c=0 c = 0 when 0 ∈ int C 0\in\operatorname{int} C 0 ∈ int C . If E E E has a proper linear extension in C ^ ( X ) ∪ { 0 } \widehat{C}(X)\cup\{0\} C ( X ) ∪ { 0 } , there are y ∈ ∂ ( C − c ) y\in\partial(C-c) y ∈ ∂ ( C − c ) and a continuous map s : ( 0 , 1 ) → X s:(0,1)\to X s : ( 0 , 1 ) → X such that
F ( s ( t ) ) = c + t y ( 0 < t < 1 ) . (7) F(s(t))=c+ty\qquad(0<t<1). \tag*{(7)} F ( s ( t )) = c + t y ( 0 < t < 1 ) . (7)
Proof. The upper boundary of the extended evaluation body can be recovered over interior points of C C C , except possibly along one segment. A radial interval avoiding that segment gives the required section. Choose g ∈ C ( X ) ∖ E g\in C(X)\setminus E g ∈ C ( X ) ∖ E such that W = E + span { g } ⊂ C ^ ( X ) ∪ { 0 } W=E+\operatorname{span}\{g\}\subset\widehat{C}(X)\cup\{0\} W = E + span { g } ⊂ C ( X ) ∪ { 0 } . Both g g g and − g -g − g attain finite maxima, so g g g is bounded:
− max X ( − g ) ≤ g ( x ) ≤ max X g ( x ∈ X ) . -\max_{X}(-g)\le g(x)\le\max_{X}g\qquad(x\in X). − X max ( − g ) ≤ g ( x ) ≤ X max g ( x ∈ X ) .
Define
Ψ : X → R k + 1 , Ψ ( x ) = ( F ( x ) , g ( x ) ) , K = Ψ ( X ) ‾ , D = conv K , \Psi:X\to\mathbb{R}^{k+1},\qquad\Psi(x)=(F(x),g(x)),\qquad K=\overline{\Psi(X)},\qquad D=\operatorname{conv}K, Ψ : X → R k + 1 , Ψ ( x ) = ( F ( x ) , g ( x )) , K = Ψ ( X ) , D = conv K ,
where the closure is taken in R k + 1 \mathbb{R}^{k+1} R k + 1 . The boundedness of F F F and g g g makes K K K compact; hence D D D is compact by Lemma 2.1[1 ] .
(1) We first identify the points belonging to Ψ ( X ∖ H ) ‾ \overline{\Psi(X\setminus H)} Ψ ( X ∖ H ) for every compact subset H ⊂ X H\subset X H ⊂ X . Let K ( X ) \mathcal{K}(X) K ( X ) be the collection of compact subsets of X X X , and put
T = ⋂ H ∈ K ( X ) Ψ ( X ∖ H ) ‾ . (8) T=\bigcap_{H\in\mathcal{K}(X)}\overline{\Psi(X\setminus H)}. \tag*{(8)} T = H ∈ K ( X ) ⋂ Ψ ( X ∖ H ) . (8)
Every set in this intersection is a nonempty compact subset of K K K . For finitely many compact sets H 1 , … , H r H_{1},\ldots,H_{r} H 1 , … , H r , their union is compact and is not all of X X X . Therefore
∅ ≠ Ψ ( X ∖ ⋃ j = 1 r H j ) ⊂ ⋂ j = 1 r Ψ ( X ∖ H j ) ‾ . \varnothing\ne\Psi\left(X\setminus\bigcup_{j=1}^{r}H_{j}\right)\subset\bigcap_{j=1}^{r}\overline{\Psi(X\setminus H_{j})}. ∅ = Ψ ( X ∖ j = 1 ⋃ r H j ) ⊂ j = 1 ⋂ r Ψ ( X ∖ H j ) .
The finite intersection property in compact K K K gives T ≠ ∅ T\ne\varnothing T = ∅ .
For ε > 0 \varepsilon>0 ε > 0 , the set H ε = ⋃ j = 1 k { x : ∣ f j ( x ) ∣ ≥ ε } H_{\varepsilon}=\bigcup_{j=1}^{k}\{x:|f_{j}(x)|\ge\varepsilon\} H ε = ⋃ j = 1 k { x : ∣ f j ( x ) ∣ ≥ ε } is compact. Every point of Ψ ( X ∖ H ε ) ‾ \overline{\Psi(X\setminus H_{\varepsilon})} Ψ ( X ∖ H ε ) has its first k k k coordinates in [ − ε , ε ] [-\varepsilon,\varepsilon] [ − ε , ε ] . Since this holds for every ε > 0 \varepsilon>0 ε > 0 , (2.8) implies T ⊂ { 0 } × R T\subset\{0\}\times\mathbb{R} T ⊂ { 0 } × R . If z ∈ K ∖ T z\in K\setminus T z ∈ K ∖ T , some compact H ⊂ X H\subset X H ⊂ X satisfies z ∉ Ψ ( X ∖ H ) ‾ z\notin\overline{\Psi(X\setminus H)} z ∈ / Ψ ( X ∖ H ) . The decomposition Ψ ( X ) = Ψ ( H ) ∪ Ψ ( X ∖ H ) \Psi(X)=\Psi(H)\cup\Psi(X\setminus H) Ψ ( X ) = Ψ ( H ) ∪ Ψ ( X ∖ H ) and compactness of Ψ ( H ) \Psi(H) Ψ ( H ) give
K = Ψ ( H ) ∪ Ψ ( X ∖ H ) ‾ , K=\Psi(H)\cup\overline{\Psi(X\setminus H)}, K = Ψ ( H ) ∪ Ψ ( X ∖ H ) ,
so z ∈ Ψ ( H ) z\in\Psi(H) z ∈ Ψ ( H ) . We have proved
T ⊂ { 0 } × R , K = Ψ ( X ) ∪ T . (9) T\subset\{0\}\times\mathbb{R},\qquad K=\Psi(X)\cup T. \tag*{(9)} T ⊂ { 0 } × R , K = Ψ ( X ) ∪ T . (9)
This union need not be disjoint. Define
λ = max { z ∈ R : ( 0 , z ) ∈ T } , p + = ( 0 , λ ) ∈ T . \lambda=\max\{z\in\mathbb{R}:(0,z)\in T\},\qquad p_{+}=(0,\lambda)\in T. λ = max { z ∈ R : ( 0 , z ) ∈ T } , p + = ( 0 , λ ) ∈ T .
Thus p + p_{+} p + is the point of T T T with largest last coordinate; λ \lambda λ need not equal sup X g \sup_{X}g sup X g .
The set D D D has nonempty interior. Indeed, if it lay in an affine hyperplane, there would be ( a , b ) ≠ 0 (a,b)\ne0 ( a , b ) = 0 and d ∈ R d\in\mathbb{R} d ∈ R with
∑ j = 1 k a j f j ( x ) + b g ( x ) = d ( x ∈ X ) . (10) \sum_{j=1}^{k}a_{j}f_{j}(x)+bg(x)=d\qquad(x\in X). \tag*{(10)} j = 1 ∑ k a j f j ( x ) + b g ( x ) = d ( x ∈ X ) . (10)
For d = 0 d=0 d = 0 , this contradicts independence of f 1 , … , f k , g f_1,\ldots,f_k,g f 1 , … , f k , g . For d ≠ 0 d\ne0 d = 0 , it puts a nonzero constant in W W W , which is impossible. For the projection π ( u , z ) = u \pi(u,z)=u π ( u , z ) = u , the set π ( K ) \pi(K) π ( K ) is compact and hence closed; since it contains F ( X ) F(X) F ( X ) , it contains F ( X ) ‾ \overline{F(X)} F ( X ) . Conversely, if z ∈ K z\in K z ∈ K , choose x m ∈ X x_m\in X x m ∈ X with Ψ ( x m ) → z \Psi(x_m)\to z Ψ ( x m ) → z ; this is possible because the closure is taken in the metrizable space R k + 1 \mathbb{R}^{k+1} R k + 1 , regardless of whether X X X is metrizable. Then F ( x m ) = π ( Ψ ( x m ) ) → π ( z ) F(x_m)=\pi(\Psi(x_m))\to\pi(z) F ( x m ) = π ( Ψ ( x m )) → π ( z ) , so π ( z ) ∈ F ( X ) ‾ \pi(z)\in\overline{F(X)} π ( z ) ∈ F ( X ) . Thus π ( K ) = F ( X ) ‾ \pi(K)=\overline{F(X)} π ( K ) = F ( X ) . Also F ( L ) F(L) F ( L ) is closed and contains F ( X ) ‾ \overline{F(X)} F ( X ) , while continuity of F F F and density of X X X in L L L give F ( L ) ⊂ F ( X ) ‾ F(L)\subset\overline{F(X)} F ( L ) ⊂ F ( X ) . Hence F ( L ) = F ( X ) ‾ F(L)=\overline{F(X)} F ( L ) = F ( X ) , and consequently
π ( D ) = conv π ( K ) = conv F ( L ) = C . (11) \pi(D)=\operatorname{conv}\pi(K)=\operatorname{conv}F(L)=C. \tag*{(11)} π ( D ) = conv π ( K ) = conv F ( L ) = C . (11)
(2) Fix a direction v = ( a , b ) ∈ R k × R v=(a,b)\in\mathbb{R}^k\times\mathbb{R} v = ( a , b ) ∈ R k × R with b > 0 b>0 b > 0 , and define
q v : X → R , q v ( x ) = ⟨ a , F ( x ) ⟩ + b g ( x ) , H v = H D ( v ) . q_v:X\to\mathbb{R},\qquad q_v(x)=\langle a,F(x)\rangle+bg(x),\qquad H_v=H_D(v). q v : X → R , q v ( x ) = ⟨ a , F ( x )⟩ + b g ( x ) , H v = H D ( v ) .
Here D v = { d ∈ D : ⟨ v , d ⟩ = H v } D_v=\{d\in D:\langle v,d\rangle=H_v\} D v = { d ∈ D : ⟨ v , d ⟩ = H v } , as in Lemma 2.1 . The coefficient b > 0 b>0 b > 0 and independence of the basis imply q v ≠ 0 q_v\ne0 q v = 0 . Thus q v q_v q v has exactly one maximizer x v ∈ X x_v\in X x v ∈ X . The maximum of the linear functional z ↦ ⟨ v , z ⟩ z\mapsto\langle v,z\rangle z ↦ ⟨ v , z ⟩ on K = Ψ ( X ) K=\Psi(X) K = Ψ ( X ) equals sup X q v = max X q v \sup_Xq_v=\max_Xq_v sup X q v = max X q v . It also equals its maximum on D = conv K D=\operatorname{conv}K D = conv K . Since p + ∈ K p_+\in K p + ∈ K , we obtain
H v = max X q v ≥ ⟨ v , p + ⟩ = b λ . H_v=\max_Xq_v\ge\langle v,p_+\rangle=b\lambda. H v = X max q v ≥ ⟨ v , p + ⟩ = bλ .
If H v > b λ H_v>b\lambda H v > bλ , no point of T T T is a maximizer, by (9) . The only maximizing point in K K K is Ψ ( x v ) \Psi(x_v) Ψ ( x v ) . If H v = b λ H_v=b\lambda H v = bλ , a point ( 0 , z ) ∈ T (0,z)\in T ( 0 , z ) ∈ T is maximizing precisely when b z = b λ bz=b\lambda b z = bλ , hence precisely when it is p + p_+ p + . The only maximizing evaluation from X X X is still Ψ ( x v ) \Psi(x_v) Ψ ( x v ) . Applying (1) to the inclusion of K K K into R k + 1 \mathbb{R}^{k+1} R k + 1 gives
D v = { { Ψ ( x v ) } , H v > b λ , [ p + , Ψ ( x v ) ] , H v = b λ . (12) D_v=
\begin{cases}
\{\Psi(x_v)\}, & H_v>b\lambda,\\
[p_+,\Psi(x_v)], & H_v=b\lambda.
\end{cases} \tag*{(12)} D v = { { Ψ ( x v )} , [ p + , Ψ ( x v )] , H v > bλ , H v = bλ . (12)
(3) All maximizing points x v x_v x v arising in the equality case of (12) coincide. To prove this, let v = ( a , b ) v=(a,b) v = ( a , b ) and v ′ = ( a ′ , b ′ ) v'=(a',b') v ′ = ( a ′ , b ′ ) have b , b ′ > 0 b,b'>0 b , b ′ > 0 and H v = b λ H_v=b\lambda H v = bλ , H v ′ = b ′ λ H_{v'}=b'\lambda H v ′ = b ′ λ . The functions q v − b λ q_v-b\lambda q v − bλ and q v ′ − b ′ λ q_{v'}-b'\lambda q v ′ − b ′ λ are nonpositive on X X X . Their sum has supremum zero, since the corresponding affine functional is nonpositive on K K K and vanishes at p + p_+ p + :
sup x ∈ X ( q v ( x ) + q v ′ ( x ) − ( b + b ′ ) λ ) = max z ∈ K ( ⟨ v + v ′ , z ⟩ − ( b + b ′ ) λ ) = 0. \sup_{x\in X}\left(q_v(x)+q_{v'}(x)-(b+b')\lambda\right)
=\max_{z\in K}\left(\langle v+v',z\rangle-(b+b')\lambda\right)
=0. x ∈ X sup ( q v ( x ) + q v ′ ( x ) − ( b + b ′ ) λ ) = z ∈ K max ( ⟨ v + v ′ , z ⟩ − ( b + b ′ ) λ ) = 0.
Since b + b ′ > 0 b+b'>0 b + b ′ > 0 , the function q v + q v ′ q_v+q_{v'} q v + q v ′ is a nonzero member of W W W . It therefore attains its maximum ( b + b ′ ) λ (b+b')\lambda ( b + b ′ ) λ at a point x ∈ X x\in X x ∈ X . At that point the two nonpositive summands have sum zero, so
q v ( x ) − b λ = 0 , q v ′ ( x ) − b ′ λ = 0. q_v(x)-b\lambda=0,\qquad q_{v'}(x)-b'\lambda=0. q v ( x ) − bλ = 0 , q v ′ ( x ) − b ′ λ = 0.
Uniqueness of each maximizer gives x = x v = x v ′ x=x_v=x_{v'} x = x v = x v ′ . If there exists v = ( a , b ) v=(a,b) v = ( a , b ) with b > 0 b>0 b > 0 and H v = b λ H_v=b\lambda H v = bλ , denote the common maximizer by x + x_+ x + and put B + = [ 0 , F ( x + ) ] B_+=[0,F(x_+)] B + = [ 0 , F ( x + )] . If no such v v v exists, put B + = ∅ B_+=\varnothing B + = ∅ .
(4) By Lemma 2.4 and (2.10), the map
Γ : int C → ∂ D , Γ ( u ) = ( u , z + ( u ) ) \Gamma:\operatorname{int}C\to\partial D,\qquad\Gamma(u)=(u,z_+(u)) Γ : int C → ∂ D , Γ ( u ) = ( u , z + ( u ))
is continuous. Put Ω = int C ∖ ( B + ∪ { 0 } ) \Omega=\operatorname{int}C\setminus(B_+\cup\{0\}) Ω = int C ∖ ( B + ∪ { 0 }) . We construct a continuous inverse to Ψ \Psi Ψ along Γ ( Ω ) \Gamma(\Omega) Γ ( Ω ) . Fix u ∈ int C u\in\operatorname{int}C u ∈ int C . By the supporting hyperplane theorem, choose ( a , b ) ≠ 0 (a,b)\ne0 ( a , b ) = 0 such that ( w , z ) ↦ ⟨ a , w ⟩ + b z (w,z)\mapsto\langle a,w\rangle+bz ( w , z ) ↦ ⟨ a , w ⟩ + b z attains its maximum on D D D at Γ ( u ) \Gamma(u) Γ ( u ) . If b < 0 b<0 b < 0 , the point ( u , z − ( u ) ) ∈ D (u,z_-(u))\in D ( u , z − ( u )) ∈ D gives
⟨ a , u ⟩ + b z − ( u ) − ( ⟨ a , u ⟩ + b z + ( u ) ) = − b ( z + ( u ) − z − ( u ) ) > 0 , \langle a,u\rangle+bz_-(u)-\left(\langle a,u\rangle+bz_+(u)\right)
=-b\left(z_+(u)-z_-(u)\right)>0, ⟨ a , u ⟩ + b z − ( u ) − ( ⟨ a , u ⟩ + b z + ( u ) ) = − b ( z + ( u ) − z − ( u ) ) > 0 ,
contradicting the fact that ( u , z + ( u ) ) (u,z_+(u)) ( u , z + ( u )) maximizes ( w , z ) ↦ ⟨ a , w ⟩ + b z (w,z)\mapsto\langle a,w\rangle+bz ( w , z ) ↦ ⟨ a , w ⟩ + b z on D D D . If b = 0 b=0 b = 0 , then a ≠ 0 a\ne0 a = 0 and u u u would maximize w ↦ ⟨ a , w ⟩ w\mapsto\langle a,w\rangle w ↦ ⟨ a , w ⟩ on C C C . This is impossible at an interior point, since u + t a ∈ C u+ta\in C u + t a ∈ C for small t > 0 t>0 t > 0 and increases that value by t ∥ a ∥ 2 t\lVert a\rVert^2 t ∥ a ∥ 2 . Thus b > 0 b>0 b > 0 . For u ∈ Ω u\in\Omega u ∈ Ω , set v = ( a , b ) v=(a,b) v = ( a , b ) for the direction just chosen. The equality H v = b λ H_v=b\lambda H v = bλ is impossible: by (12) , it would imply D v = [ p + , Ψ ( x + ) ] D_v=[p_+,\Psi(x_+)] D v = [ p + , Ψ ( x + )] and hence u = π ( Γ ( u ) ) ∈ π ( D v ) = B + u=\pi(\Gamma(u))\in\pi(D_v)=B_+ u = π ( Γ ( u )) ∈ π ( D v ) = B + , a contradiction. Thus H v > b λ H_v > b\lambda H v > bλ and D v = { Ψ ( x v ) } D_v = \{\Psi(x_v)\} D v = { Ψ ( x v )} by (12) . Since Γ ( u ) ∈ D v \Gamma(u) \in D_v Γ ( u ) ∈ D v , Γ ( u ) = Ψ ( x v ) \Gamma(u) = \Psi(x_v) Γ ( u ) = Ψ ( x v ) . This point has exactly one preimage in X X X , because every preimage would maximize q v q_v q v . Define
σ : Ω → X , Ψ ( σ ( u ) ) = Γ ( u ) , F ( σ ( u ) ) = u . (13) \sigma: \Omega\to X,\qquad\Psi(\sigma(u)) = \Gamma(u),\qquad F(\sigma(u)) = u. \tag*{(13)} σ : Ω → X , Ψ ( σ ( u )) = Γ ( u ) , F ( σ ( u )) = u . (13)
To prove continuity, let Q ⊂ Ω Q \subset\Omega Q ⊂ Ω be a closed Euclidean ball. Since 0 ∉ Q 0 \notin Q 0 ∈ / Q , put δ = min u ∈ Q ∥ u ∥ > 0 \delta= \min_{u\in Q} \lVert u\rVert> 0 δ = min u ∈ Q ∥ u ∥ > 0 . The set Ψ − 1 ( Γ ( Q ) ) \Psi^{-1}(\Gamma(Q)) Ψ − 1 ( Γ ( Q )) is closed in X X X and is contained in
{ x ∈ X : ∥ F ( x ) ∥ ≥ δ } ⊂ ⋃ j = 1 k { x ∈ X : ∣ f j ( x ) ∣ ≥ δ / k } . \{x \in X : \lVert F(x)\rVert\ge\delta\} \subset\bigcup_{j=1}^{k}\{x \in X : |f_j(x)| \ge\delta/\sqrt{k}\}. { x ∈ X : ∥ F ( x )∥ ≥ δ } ⊂ j = 1 ⋃ k { x ∈ X : ∣ f j ( x ) ∣ ≥ δ / k } .
The last inclusion follows because ∣ f j ( x ) ∣ < δ / k |f_j(x)| < \delta/\sqrt{k} ∣ f j ( x ) ∣ < δ / k for all j j j would give ∥ F ( x ) ∥ 2 < k ( δ 2 / k ) = δ 2 \lVert F(x)\rVert^2 < k(\delta^2/k) = \delta^2 ∥ F ( x ) ∥ 2 < k ( δ 2 / k ) = δ 2 . The union on the right is compact, since each f j f_j f j vanishes at infinity. Hence Ψ − 1 ( Γ ( Q ) ) \Psi^{-1}(\Gamma(Q)) Ψ − 1 ( Γ ( Q )) is compact. The restriction of Ψ \Psi Ψ to this set is a continuous bijection onto Γ ( Q ) \Gamma(Q) Γ ( Q ) , and so has continuous inverse. Composing that inverse with Γ ∣ Q \Gamma|_Q Γ ∣ Q proves continuity of σ ∣ Q \sigma|_Q σ ∣ Q . The set Ω \Omega Ω is open; such balls are neighborhoods locally at every point of Ω \Omega Ω . Thus σ \sigma σ is continuous.
(5) We now choose y ∈ ∂ ( C − c ) y \in\partial(C-c) y ∈ ∂ ( C − c ) such that { c + t y : 0 < t < 1 } ⊂ Ω \{c+ty : 0<t<1\} \subset\Omega { c + t y : 0 < t < 1 } ⊂ Ω . Let p : R k → [ 0 , ∞ ) p : \mathbb{R}^{k} \to[0,\infty) p : R k → [ 0 , ∞ ) be the Minkowski functional of C − c C-c C − c from Lemma 2.2. Suppose first that 0 ∈ int C 0 \in\operatorname{int} C 0 ∈ int C , so c = 0 c=0 c = 0 . If B + B_+ B + is nondegenerate, write v + = F ( x + ) ≠ 0 v_+ = F(x_+) \ne0 v + = F ( x + ) = 0 and choose y = − v + / p ( − v + ) ∈ ∂ C y = -v_+/p(-v_+) \in\partial C y = − v + / p ( − v + ) ∈ ∂ C . For 0 < t < 1 0<t<1 0 < t < 1 , the point t y ty t y is nonzero and lies on the ray opposite to B + = [ 0 , v + ] B_+ = [0,v_+] B + = [ 0 , v + ] . Also p ( t y ) = t < 1 p(ty)=t<1 p ( t y ) = t < 1 , so t y ∈ int C ty \in\operatorname{int} C t y ∈ int C . Hence t y ∈ Ω ty \in\Omega t y ∈ Ω . If B + B_+ B + is empty or equals { 0 } \{0\} { 0 } , take any y ∈ ∂ C y \in\partial C y ∈ ∂ C ; the same conclusion follows from p ( t y ) = t p(ty)=t p ( t y ) = t and t y ≠ 0 ty \ne0 t y = 0 . The map s ( t ) = σ ( t y ) s(t)=\sigma(ty) s ( t ) = σ ( t y ) is continuous and satisfies (7) by (13) .
Finally, suppose 0 ∈ ∂ C 0 \in\partial C 0 ∈ ∂ C . By the supporting hyperplane theorem, choose a 0 ≠ 0 a_0 \ne0 a 0 = 0 such that ⟨ a 0 , u ⟩ ≤ 0 \langle a_0,u\rangle\le0 ⟨ a 0 , u ⟩ ≤ 0 for every u ∈ C u \in C u ∈ C , and put f 0 = ∑ j = 1 k ( a 0 ) j f j f_0 = \sum_{j=1}^{k}(a_0)_j f_j f 0 = ∑ j = 1 k ( a 0 ) j f j . Lemma 2.3(3) gives f 0 ≤ 0 f_0 \le0 f 0 ≤ 0 and a unique zero x 0 ∈ X x_0 \in X x 0 ∈ X . For a vector v = ( a , b ) v=(a,b) v = ( a , b ) with b > 0 b>0 b > 0 and H v = b λ H_v=b\lambda H v = bλ , both f 0 f_0 f 0 and q v − b λ q_v-b\lambda q v − bλ are nonpositive. Their sum has supremum zero: its associated affine functional on K K K vanishes at p + p_+ p + . The function f 0 + q v f_0+q_v f 0 + q v is a nonzero member of W W W , because its coefficient of g g g is b > 0 b>0 b > 0 . It attains the maximum b λ b\lambda bλ at some x ∈ X x \in X x ∈ X . Then
0 = f 0 ( x ) + ( q v ( x ) − b λ ) , f 0 ( x ) ≤ 0 , q v ( x ) − b λ ≤ 0 , 0=f_0(x)+(q_v(x)-b\lambda),\qquad f_0(x)\le0,\qquad q_v(x)-b\lambda\le0, 0 = f 0 ( x ) + ( q v ( x ) − bλ ) , f 0 ( x ) ≤ 0 , q v ( x ) − bλ ≤ 0 ,
so f 0 ( x ) = 0 f_0(x)=0 f 0 ( x ) = 0 and q v ( x ) = b λ q_v(x)=b\lambda q v ( x ) = bλ . Thus x = x 0 = x v x=x_0=x_v x = x 0 = x v . Consequently, if such a vector v v v exists, x + = x 0 x_+=x_0 x + = x 0 and
B + = [ 0 , F ( x 0 ) ] ⊂ { u ∈ C : ⟨ a 0 , u ⟩ = 0 } ⊂ ∂ C . B_+=[0,F(x_0)]\subset\{u\in C:\langle a_0,u\rangle=0\}\subset\partial C. B + = [ 0 , F ( x 0 )] ⊂ { u ∈ C : ⟨ a 0 , u ⟩ = 0 } ⊂ ∂ C .
If none exists, B + = ∅ B_+=\varnothing B + = ∅ . In either case, Ω = int C \Omega=\operatorname{int} C Ω = int C because 0 ∈ ∂ C 0\in\partial C 0 ∈ ∂ C . Choose any y ∈ ∂ ( C − c ) y\in\partial(C-c) y ∈ ∂ ( C − c ) . Lemma 2.2 gives p ( c + t y − c ) = t p ( y ) = t < 1 p(c+ty-c)=tp(y)=t<1 p ( c + t y − c ) = tp ( y ) = t < 1 , so c + t y ∈ Ω c+ty\in\Omega c + t y ∈ Ω for 0 < t < 1 0<t<1 0 < t < 1 . Now s ( t ) = σ ( c + t y ) s(t)=\sigma(c+ty) s ( t ) = σ ( c + t y ) is the required map. □
A perturbation preserving all extrema
A continuous scalar function, arbitrarily close to the identity and fixing the endpoints, can fail to admit a continuous right inverse on ( 0 , 1 ) (0,1) ( 0 , 1 ) .
Lemma 2.7. The following statements hold.
For every ε > 0 \varepsilon>0 ε > 0 there is a continuous function w : [ 0 , 1 ] → [ 0 , 1 ] w:[0,1]\to[0,1] w : [ 0 , 1 ] → [ 0 , 1 ] such that
w ( 0 ) = 0 , w ( 1 ) = 1 , 0 < w ( t ) < 1 ( 0 < t < 1 ) , (14) w(0)=0,\qquad w(1)=1,\qquad0<w(t)<1\quad(0<t<1), \tag*{(14)} w ( 0 ) = 0 , w ( 1 ) = 1 , 0 < w ( t ) < 1 ( 0 < t < 1 ) , (14) ∥ w − id [ 0 , 1 ] ∥ ∞ < ε \lVert w-\operatorname{id}_{[0,1]}\rVert_\infty<\varepsilon ∥ w − id [ 0 , 1 ] ∥ ∞ < ε , and w w w is not strictly increasing.
Every continuous w : [ 0 , 1 ] → [ 0 , 1 ] w:[0,1]\to[0,1] w : [ 0 , 1 ] → [ 0 , 1 ] satisfying (14) and admitting a continuous function r : ( 0 , 1 ) → [ 0 , 1 ] r:(0,1)\to[0,1] r : ( 0 , 1 ) → [ 0 , 1 ] with w ( r ( t ) ) = t w(r(t))=t w ( r ( t )) = t for 0 < t < 1 0<t<1 0 < t < 1 is strictly increasing on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
Proof. (1) Choose 0 < τ < min { ε , 1 / 6 } 0<\tau<\min\{\varepsilon,1/6\} 0 < τ < min { ε , 1/6 } and define
w ( t ) = { t , 0 ≤ t ≤ 1 2 , 2 t − 1 2 , 1 2 ≤ t ≤ 1 2 + τ , 1 + 3 τ − t , 1 2 + τ ≤ t ≤ 1 2 + 2 τ , 2 t − 1 2 − 3 τ , 1 2 + 2 τ ≤ t ≤ 1 2 + 3 τ , t , 1 2 + 3 τ ≤ t ≤ 1. w(t)=
\begin{cases}
t, & 0\leq t\leq\frac{1}{2},\\
2t-\frac{1}{2}, & \frac{1}{2}\leq t\leq\frac{1}{2}+\tau,\\
1+3\tau-t, & \frac{1}{2}+\tau\leq t\leq\frac{1}{2}+2\tau,\\
2t-\frac{1}{2}-3\tau, & \frac{1}{2}+2\tau\leq t\leq\frac{1}{2}+3\tau,\\
t, & \frac{1}{2}+3\tau\leq t\leq1.
\end{cases} w ( t ) = ⎩ ⎨ ⎧ t , 2 t − 2 1 , 1 + 3 τ − t , 2 t − 2 1 − 3 τ , t , 0 ≤ t ≤ 2 1 , 2 1 ≤ t ≤ 2 1 + τ , 2 1 + τ ≤ t ≤ 2 1 + 2 τ , 2 1 + 2 τ ≤ t ≤ 2 1 + 3 τ , 2 1 + 3 τ ≤ t ≤ 1.
The adjacent formulas agree at each endpoint. Their values at the four interior breakpoints are, in order, 1 / 2 1/2 1/2 , 1 / 2 + 2 τ 1/2+2\tau 1/2 + 2 τ , 1 / 2 + τ 1/2+\tau 1/2 + τ , and 1 / 2 + 3 τ 1/2+3\tau 1/2 + 3 τ . They lie in ( 0 , 1 ) (0,1) ( 0 , 1 ) , so piecewise affinity proves (14) . On the middle interval, the slope is − 1 -1 − 1 , hence w w w is not strictly increasing. On each interval, w ( t ) − t w(t)-t w ( t ) − t is affine; its values at all breakpoints, including 0 0 0 and 1 1 1 , are 0 , 0 , τ , − τ , 0 , 0 0,0,\tau,-\tau,0,0 0 , 0 , τ , − τ , 0 , 0 . Therefore ∥ w − id [ 0 , 1 ] ∥ ∞ = τ < ε \lVert w-\operatorname{id}_{[0,1]}\rVert_{\infty}=\tau<\varepsilon ∥ w − id [ 0 , 1 ] ∥ ∞ = τ < ε .
(2) The identity w ( r ( t ) ) = t w(r(t))=t w ( r ( t )) = t makes r r r injective, and (14) gives 0 < r ( t ) < 1 0<r(t)<1 0 < r ( t ) < 1 . For 0 < δ < 1 0<\delta<1 0 < δ < 1 , compactness and (14) imply
m δ = min δ ≤ u ≤ 1 w ( u ) > 0 , M δ = max 0 ≤ u ≤ 1 − δ w ( u ) < 1. m_{\delta}=\min_{\delta\leq u\leq1}w(u)>0,\qquad M_{\delta}=\max_{0\leq u\leq1-\delta}w(u)<1. m δ = δ ≤ u ≤ 1 min w ( u ) > 0 , M δ = 0 ≤ u ≤ 1 − δ max w ( u ) < 1.
If 0 < t < m δ 0<t<m_{\delta} 0 < t < m δ , then r ( t ) < δ r(t)<\delta r ( t ) < δ , since otherwise t = w ( r ( t ) ) ≥ m δ t=w(r(t))\geq m_{\delta} t = w ( r ( t )) ≥ m δ . If M δ < t < 1 M_{\delta}<t<1 M δ < t < 1 , then r ( t ) > 1 − δ r(t)>1-\delta r ( t ) > 1 − δ , since otherwise t = w ( r ( t ) ) ≤ M δ t=w(r(t))\leq M_{\delta} t = w ( r ( t )) ≤ M δ . Thus
lim t ↓ 0 r ( t ) = 0 , lim t ↑ 1 r ( t ) = 1. \lim_{t\downarrow0}r(t)=0,\qquad\lim_{t\uparrow1}r(t)=1. t ↓ 0 lim r ( t ) = 0 , t ↑ 1 lim r ( t ) = 1.
A continuous injective real function on an interval is strictly monotone. The two limits force r r r to be increasing, and the intermediate value theorem gives r ( ( 0 , 1 ) ) = ( 0 , 1 ) r((0,1))=(0,1) r (( 0 , 1 )) = ( 0 , 1 ) . For 0 < a < b < 1 0<a<b<1 0 < a < b < 1 , write a = r ( t a ) a=r(t_a) a = r ( t a ) and b = r ( t b ) b=r(t_b) b = r ( t b ) . Then t a < t b t_a<t_b t a < t b , whence w ( a ) = t a < t b = w ( b ) w(a)=t_a<t_b=w(b) w ( a ) = t a < t b = w ( b ) . The endpoint comparisons w ( 0 ) < w ( b ) w(0)<w(b) w ( 0 ) < w ( b ) and w ( a ) < w ( 1 ) w(a)<w(1) w ( a ) < w ( 1 ) follow from (14) . Thus w w w is strictly increasing on [ 0 , 1 ] [0,1] [ 0 , 1 ] . □
The deformation of the evaluation map yields an admissible isometric image of the original subspace with no proper admissible extension. This image need not contain the original subspace, since the definition of ( α , β ) (\alpha,\beta) ( α , β ) -lineability concerns every prescribed α \alpha α -dimensional subspace.
Theorem 2.8. Let X X X be Hausdorff with at least two points, and let E E E be a nonzero finite-dimensional admissible subspace. Suppose either that X X X is locally compact and E ⊂ C 0 ( X ) E\subset C_0(X) E ⊂ C 0 ( X ) , or that X X X is countably compact and E ⊂ C ( X ) E\subset C(X) E ⊂ C ( X ) . For every η > 0 \eta>0 η > 0 there is a linear map S : E → C b ( X ) S:E\to C_b(X) S : E → C b ( X ) with the following properties. In the locally compact case with E ⊂ C 0 ( X ) E\subset C_0(X) E ⊂ C 0 ( X ) , one may also require S ( E ) ⊂ C 0 ( X ) S(E)\subset C_0(X) S ( E ) ⊂ C 0 ( X ) .
(1) For every f ∈ E ∖ { 0 } f\in E\setminus\{0\} f ∈ E ∖ { 0 } ,
max X S f = max X f , min X S f = min X f , \max_X Sf=\max_X f,\qquad\min_X Sf=\min_X f, X max S f = X max f , X min S f = X min f , Argmax X S f = Argmax X f , Argmin X S f = Argmin X f . \operatorname{Argmax}_X Sf=\operatorname{Argmax}_X f,\qquad\operatorname{Argmin}_X Sf=\operatorname{Argmin}_X f. Argmax X S f = Argmax X f , Argmin X S f = Argmin X f . (2) The map S S S is an isometry and ∥ S − I E ∥ < η \lVert S-I_E\rVert<\eta ∥ S − I E ∥ < η , where I E : E → C b ( X ) I_E:E\to C_b(X) I E : E → C b ( X ) is the inclusion and E E E carries the supremum norm.
(3) Every linear subspace W W W satisfying S ( E ) ⊂ W ⊂ C ^ ( X ) ∪ { 0 } S(E)\subset W\subset\widehat{C}(X)\cup\{0\} S ( E ) ⊂ W ⊂ C ( X ) ∪ { 0 } equals S ( E ) S(E) S ( E ) .
Proof. (1) Choose a basis f 1 , … , f k f_1,\ldots,f_k f 1 , … , f k of E E E . If X X X is locally compact and E ⊂ C 0 ( X ) E\subset C_0(X) E ⊂ C 0 ( X ) , use the setup L , F , C L,F,C L , F , C of Lemma 2.3 ; call this case (i). Otherwise use the countably compact hypothesis and set L = X L=X L = X , F = ( f 1 , … , f k ) F=(f_1,\ldots,f_k) F = ( f 1 , … , f k ) and C = conv F ( X ) C=\operatorname{conv}F(X) C = conv F ( X ) ; call this case (ii). Lemma 2.5 applies in case (ii). In both cases F ( L ) F(L) F ( L ) and C C C are compact and int C ≠ ∅ \operatorname{int}C\neq\varnothing int C = ∅ . Choose c ∈ int C c\in\operatorname{int}C c ∈ int C , taking c = 0 c=0 c = 0 in case (i) when X X X is noncompact and 0 ∈ int C 0\in\operatorname{int}C 0 ∈ int C . Let p : R k → [ 0 , ∞ ) p:\mathbb{R}^k\to[0,\infty) p : R k → [ 0 , ∞ ) be the Minkowski functional of C − c C-c C − c from Lemma 2.2 . Choose w w w by Lemma 2.7 (1) so that
ε : = ∥ w − id [ 0 , 1 ] ∥ ∞ < η / 2. (15) \varepsilon:=\lVert w-\operatorname{id}_{[0,1]}\rVert_{\infty}<\eta/2. \tag*{(15)} ε := ∥ w − id [ 0 , 1 ] ∥ ∞ < η /2. (15)
Define
h : L → [ 0 , 1 ] , h ( x ) = p ( F ( x ) − c ) , G : L → R k , G ( x ) = { c + w ( h ( x ) ) h ( x ) ( F ( x ) − c ) , h ( x ) > 0 , c , h ( x ) = 0. (16) \begin{aligned}
h &: L \to[0,1], \qquad h(x)=p(F(x)-c),\\
G &: L \to\mathbb{R}^{k}, \qquad
G(x)=
\begin{cases}
c+\dfrac{w(h(x))}{h(x)}(F(x)-c), & h(x)>0,\\
c, & h(x)=0.
\end{cases} \tag*{(16)}
\end{aligned} h G : L → [ 0 , 1 ] , h ( x ) = p ( F ( x ) − c ) , : L → R k , G ( x ) = ⎩ ⎨ ⎧ c + h ( x ) w ( h ( x )) ( F ( x ) − c ) , c , h ( x ) > 0 , h ( x ) = 0. (16)
The range of h h h lies in [ 0 , 1 ] [0,1] [ 0 , 1 ] by (2) . Continuity of F F F and p p p gives continuity of h h h . When h ( x ) > 0 h(x)>0 h ( x ) > 0 , positive homogeneity gives
p ( F ( x ) − c h ( x ) ) = 1 , p\left(\frac{F(x)-c}{h(x)}\right)=1, p ( h ( x ) F ( x ) − c ) = 1 ,
so ( F ( x ) − c ) / h ( x ) ∈ ∂ ( C − c ) (F(x)-c)/h(x)\in\partial(C-c) ( F ( x ) − c ) / h ( x ) ∈ ∂ ( C − c ) . Choose R 0 > 0 R_0>0 R 0 > 0 with ∥ y ∥ ≤ R 0 \lVert y\rVert\le R_0 ∥ y ∥ ≤ R 0 on C − c C-c C − c . At every point with h ( x ) > 0 h(x)>0 h ( x ) > 0 , we have
∥ G ( x ) − c ∥ = w ( h ( x ) ) ∥ F ( x ) − c h ( x ) ∥ ≤ R 0 w ( h ( x ) ) . \lVert G(x)-c\rVert=w(h(x))\left\lVert\frac{F(x)-c}{h(x)}\right\rVert\le R_0w(h(x)). ∥ G ( x ) − c ∥ = w ( h ( x )) h ( x ) F ( x ) − c ≤ R 0 w ( h ( x )) .
Since w ( 0 ) = 0 w(0)=0 w ( 0 ) = 0 , this bound proves continuity of G G G also where h = 0 h=0 h = 0 . For h > 0 h>0 h > 0 , homogeneity of p p p gives p ( G ( x ) − c ) = w ( h ( x ) ) p ( ( F ( x ) − c ) / h ( x ) ) = w ( h ( x ) ) p(G(x)-c)=w(h(x))p((F(x)-c)/h(x))=w(h(x)) p ( G ( x ) − c ) = w ( h ( x )) p (( F ( x ) − c ) / h ( x )) = w ( h ( x )) . For h = 0 h=0 h = 0 , both sides are zero. Thus
p ( G ( x ) − c ) = w ( h ( x ) ) ( x ∈ L ) . (17) p(G(x)-c)=w(h(x)) \qquad(x\in L). \tag*{(17)} p ( G ( x ) − c ) = w ( h ( x )) ( x ∈ L ) . (17)
In particular, G ( L ) ⊂ C G(L)\subset C G ( L ) ⊂ C . If F ( x ) ∈ ∂ C F(x)\in\partial C F ( x ) ∈ ∂ C , then h ( x ) = 1 h(x)=1 h ( x ) = 1 and w ( h ( x ) ) = 1 w(h(x))=1 w ( h ( x )) = 1 , hence G ( x ) = F ( x ) G(x)=F(x) G ( x ) = F ( x ) . If F ( x ) ∈ int C F(x)\in\operatorname{int} C F ( x ) ∈ int C , then h ( x ) < 1 h(x)<1 h ( x ) < 1 and w ( h ( x ) ) < 1 w(h(x))<1 w ( h ( x )) < 1 , hence G ( x ) ∈ int C G(x)\in\operatorname{int} C G ( x ) ∈ int C . We have proved
F ( x ) ∈ ∂ C ⟹ G ( x ) = F ( x ) , F ( x ) ∈ int C ⟹ G ( x ) ∈ int C . (18) \begin{aligned}
F(x)\in\partial C &\Longrightarrow G(x)=F(x),\\
F(x)\in\operatorname{int} C &\Longrightarrow G(x)\in\operatorname{int} C. \tag*{(18)}
\end{aligned} F ( x ) ∈ ∂ C F ( x ) ∈ int C ⟹ G ( x ) = F ( x ) , ⟹ G ( x ) ∈ int C . (18)
Write G = ( g 1 , … , g k ) G=(g_1,\ldots,g_k) G = ( g 1 , … , g k ) and restrict the coordinates to X X X . These restrictions are continuous and bounded, since G ( L ) ⊂ C G(L)\subset C G ( L ) ⊂ C . In case (ii) this is the required conclusion. In case (i) we also verify that they belong to C 0 ( X ) C_0(X) C 0 ( X ) . This is automatic for compact X X X . For noncompact X X X in case (i), F ( ∞ ) = 0 ∈ C F(\infty)=0\in C F ( ∞ ) = 0 ∈ C . If 0 ∈ int C 0\in\operatorname{int} C 0 ∈ int C , our choice c = 0 c=0 c = 0 gives h ( ∞ ) = 0 h(\infty)=0 h ( ∞ ) = 0 and G ( ∞ ) = 0 G(\infty)=0 G ( ∞ ) = 0 . If 0 ∈ ∂ C 0\in\partial C 0 ∈ ∂ C , the first implication in (18) gives G ( ∞ ) = F ( ∞ ) = 0 G(\infty)=F(\infty)=0 G ( ∞ ) = F ( ∞ ) = 0 . For each j j j and δ > 0 \delta>0 δ > 0 , the set { x ∈ L : ∣ g j ( x ) ∣ ≥ δ } \{x\in L:|g_j(x)|\ge\delta\} { x ∈ L : ∣ g j ( x ) ∣ ≥ δ } is therefore compact in L L L and avoids ∞ \infty ∞ . It is a compact subset of X X X , proving g j ∣ X ∈ C 0 ( X ) g_j|_X\in C_0(X) g j ∣ X ∈ C 0 ( X ) . We use the same symbol g j g_j g j for the restriction.
Define the linear map
S : E → C b ( X ) , S ( ∑ j = 1 k a j f j ) = ∑ j = 1 k a j g j . S:E\to C_b(X), \qquad S\left(\sum_{j=1}^{k}a_jf_j\right)=\sum_{j=1}^{k}a_jg_j. S : E → C b ( X ) , S ( j = 1 ∑ k a j f j ) = j = 1 ∑ k a j g j .
This is well defined because the representation in the basis is unique. Fix a ∈ R k ∖ { 0 } a\in\mathbb{R}^{k}\setminus\{0\} a ∈ R k ∖ { 0 } and put f = ∑ j = 1 k a j f j f=\sum_{j=1}^{k}a_jf_j f = ∑ j = 1 k a j f j . Lemma 2.1 gives H C ( a ) = max X f H_C(a)=\max_X f H C ( a ) = max X f . An interior point of C C C cannot maximize the nonzero functional u ↦ ⟨ a , u ⟩ u\mapsto\langle a,u\rangle u ↦ ⟨ a , u ⟩ , because a small displacement in the direction a a a increases its value while remaining in C C C . On the boundary, G G G agrees with F F F by (18) . Consequently, for every x ∈ X x\in X x ∈ X ,
⟨ a , G ( x ) ⟩ = H C ( a ) ⟺ ⟨ a , F ( x ) ⟩ = H C ( a ) . (19) \langle a,G(x)\rangle=H_C(a)\quad\Longleftrightarrow\quad\langle a,F(x)\rangle=H_C(a). \tag*{(19)} ⟨ a , G ( x )⟩ = H C ( a ) ⟺ ⟨ a , F ( x )⟩ = H C ( a ) . (19)
Moreover, ⟨ a , G ( x ) ⟩ ≤ H C ( a ) \langle a,G(x)\rangle\le H_C(a) ⟨ a , G ( x )⟩ ≤ H C ( a ) since G ( x ) ∈ C G(x)\in C G ( x ) ∈ C , and equality is attained at the original maximizer of f f f . Thus max X S f = max X f \max_X Sf=\max_X f max X S f = max X f and Argmax X S f = Argmax X f \operatorname{Argmax}_X Sf=\operatorname{Argmax}_X f Argmax X S f = Argmax X f . Applying (19) to − a -a − a yields max X ( − S f ) = max X ( − f ) \max_X(-Sf)=\max_X(-f) max X ( − S f ) = max X ( − f ) and equality of their maximizing sets. Since min f = − max ( − f ) \min f=-\max(-f) min f = − max ( − f ) and Argmin f = Argmax ( − f ) \operatorname{Argmin} f=\operatorname{Argmax}(-f) Argmin f = Argmax ( − f ) , this proves all four equalities in part (1).
(2) For f ≠ 0 f\ne0 f = 0 , part (1) gives
∥ S f ∥ ∞ = max { ∣ max X S f ∣ , ∣ min X S f ∣ } = max { ∣ max X f ∣ , ∣ min X f ∣ } = ∥ f ∥ ∞ . \lVert Sf\rVert_\infty=\max\{|\max_X Sf|,|\min_X Sf|\}=\max\{|\max_X f|,|\min_X f|\}=\lVert f\rVert_\infty. ∥ S f ∥ ∞ = max { ∣ X max S f ∣ , ∣ X min S f ∣ } = max { ∣ X max f ∣ , ∣ X min f ∣ } = ∥ f ∥ ∞ .
The identity also holds for f = 0 f=0 f = 0 , so S S S is an isometry. In particular, it is injective and S ( E ) ⊂ C ^ ( X ) ∪ { 0 } S(E)\subset\widehat{C}(X)\cup\{0\} S ( E ) ⊂ C ( X ) ∪ { 0 } . We prove the approximation estimate for f = ∑ j = 1 k a j f j f=\sum_{j=1}^{k}a_jf_j f = ∑ j = 1 k a j f j . For every u ∈ C u\in C u ∈ C , write u = ∑ i = 1 r λ i F ( x i ) u=\sum_{i=1}^{r}\lambda_iF(x_i) u = ∑ i = 1 r λ i F ( x i ) with λ i ≥ 0 \lambda_i\ge0 λ i ≥ 0 and ∑ i = 1 r λ i = 1 \sum_{i=1}^{r}\lambda_i=1 ∑ i = 1 r λ i = 1 . Then
∣ ⟨ a , u ⟩ ∣ = ∣ ∑ i = 1 r λ i f ( x i ) ∣ ≤ ∑ i = 1 r λ i ∥ f ∥ ∞ = ∥ f ∥ ∞ . (20) \lvert\langle a,u\rangle\rvert=\left\lvert\sum_{i=1}^{r}\lambda_i f(x_i)\right\rvert\le\sum_{i=1}^{r}\lambda_i\lVert f\rVert_{\infty}=\lVert f\rVert_{\infty}. \tag*{(20)} ∣⟨ a , u ⟩∣ = i = 1 ∑ r λ i f ( x i ) ≤ i = 1 ∑ r λ i ∥ f ∥ ∞ = ∥ f ∥ ∞ . (20)
In case (i), when X X X is noncompact, this also holds at x i = ∞ x_i=\infty x i = ∞ , where f = 0 f=0 f = 0 . If h ( x ) > 0 h(x)>0 h ( x ) > 0 , put y = ( F ( x ) − c ) / h ( x ) y=(F(x)-c)/h(x) y = ( F ( x ) − c ) / h ( x ) . Then c + y ∈ ∂ C ⊂ C c+y\in\partial C\subset C c + y ∈ ∂ C ⊂ C and c ∈ C c\in C c ∈ C , so (20) gives
∣ ⟨ a , y ⟩ ∣ ≤ ∣ ⟨ a , c + y ⟩ ∣ + ∣ ⟨ a , c ⟩ ∣ ≤ 2 ∥ f ∥ ∞ . \lvert\langle a,y\rangle\rvert\le\lvert\langle a,c+y\rangle\rvert+\lvert\langle a,c\rangle\rvert\le2\lVert f\rVert_{\infty}. ∣⟨ a , y ⟩∣ ≤ ∣⟨ a , c + y ⟩∣ + ∣⟨ a , c ⟩∣ ≤ 2 ∥ f ∥ ∞ .
Subtracting F ( x ) = c + h ( x ) y F(x)=c+h(x)y F ( x ) = c + h ( x ) y from G ( x ) = c + w ( h ( x ) ) y G(x)=c+w(h(x))y G ( x ) = c + w ( h ( x )) y now gives
∣ S f ( x ) − f ( x ) ∣ = ∣ ⟨ a , G ( x ) − F ( x ) ⟩ ∣ = ∣ w ( h ( x ) ) − h ( x ) ∣ ∣ ⟨ a , y ⟩ ∣ ≤ 2 ε ∥ f ∥ ∞ . \begin{aligned}
\lvert Sf(x)-f(x)\rvert&=\lvert\langle a,G(x)-F(x)\rangle\rvert\\
&=\lvert w(h(x))-h(x)\rvert\lvert\langle a,y\rangle\rvert\\
&\le2\varepsilon\lVert f\rVert_{\infty}.
\end{aligned} ∣ S f ( x ) − f ( x )∣ = ∣⟨ a , G ( x ) − F ( x )⟩∣ = ∣ w ( h ( x )) − h ( x )∣ ∣⟨ a , y ⟩∣ ≤ 2 ε ∥ f ∥ ∞ .
If h ( x ) = 0 h(x)=0 h ( x ) = 0 , then F ( x ) = G ( x ) = c F(x)=G(x)=c F ( x ) = G ( x ) = c and the difference is zero. Taking first the supremum over x ∈ X x\in X x ∈ X and then over ∥ f ∥ ∞ ≤ 1 \lVert f\rVert_{\infty}\le1 ∥ f ∥ ∞ ≤ 1 gives
∥ S − I E ∥ = sup ∥ f ∥ ∞ ≤ 1 sup x ∈ X ∣ S f ( x ) − f ( x ) ∣ ≤ 2 ε < η . \lVert S-I_E\rVert=\sup_{\lVert f\rVert_{\infty}\le1}\sup_{x\in X}\lvert Sf(x)-f(x)\rvert\le2\varepsilon<\eta. ∥ S − I E ∥ = ∥ f ∥ ∞ ≤ 1 sup x ∈ X sup ∣ S f ( x ) − f ( x )∣ ≤ 2 ε < η .
(3) We first show that
conv G ( L ) = C . (21) \operatorname{conv}G(L)=C. \tag*{(21)} conv G ( L ) = C . (21)
The set G ( L ) G(L) G ( L ) is compact: in case (i), L L L is compact; in case (ii), L L L is countably compact and G G G has finite-dimensional range. Thus conv G ( L ) \operatorname{conv}G(L) conv G ( L ) is compact. Its inclusion in C C C follows from G ( L ) ⊂ C G(L)\subset C G ( L ) ⊂ C . For each a ≠ 0 a\ne0 a = 0 , the maximum of f = ⟨ a , F ⟩ f=\langle a,F\rangle f = ⟨ a , F ⟩ on X X X equals H C ( a ) H_C(a) H C ( a ) and is attained at a point fixed by the deformation. Hence
max u ∈ conv G ( L ) ⟨ a , u ⟩ = max x ∈ L ⟨ a , G ( x ) ⟩ = H C ( a ) . \max_{u\in\operatorname{conv}G(L)}\langle a,u\rangle=\max_{x\in L}\langle a,G(x)\rangle=H_C(a). u ∈ conv G ( L ) max ⟨ a , u ⟩ = x ∈ L max ⟨ a , G ( x )⟩ = H C ( a ) .
If u 0 ∈ C ∖ conv G ( L ) u_0\in C\setminus\operatorname{conv}G(L) u 0 ∈ C ∖ conv G ( L ) , separation of the point u 0 u_0 u 0 from the compact convex set conv G ( L ) \operatorname{conv}G(L) conv G ( L ) would give a ≠ 0 a\ne0 a = 0 with
H C ( a ) ≥ ⟨ a , u 0 ⟩ > max u ∈ conv G ( L ) ⟨ a , u ⟩ = H C ( a ) , H_C(a)\ge\langle a,u_0\rangle>\max_{u\in\operatorname{conv}G(L)}\langle a,u\rangle=H_C(a), H C ( a ) ≥ ⟨ a , u 0 ⟩ > u ∈ conv G ( L ) max ⟨ a , u ⟩ = H C ( a ) ,
a contradiction. This proves (21) .
Suppose now that S ( E ) S(E) S ( E ) has a proper extension in C ^ ( X ) ∪ { 0 } \widehat{C}(X)\cup\{0\} C ( X ) ∪ { 0 } . In case (i) with noncompact X X X , apply Lemma 2.6 to the basis g 1 , … , g k g_1,\ldots,g_k g 1 , … , g k of S ( E ) S(E) S ( E ) , using (21) and the chosen c c c . This gives y ∈ ∂ ( C − c ) y\in\partial(C-c) y ∈ ∂ ( C − c ) and a continuous s : ( 0 , 1 ) → X s:(0,1)\to X s : ( 0 , 1 ) → X with G ( s ( t ) ) = c + t y G(s(t))=c+ty G ( s ( t )) = c + t y . In all other cases X X X is countably compact, so Lemma 2.5 , applied to the basis g 1 , … , g k g_1,\ldots,g_k g 1 , … , g k , gives a continuous section on int C \operatorname{int}C int C . Composing it with t ↦ c + t y t\mapsto c+ty t ↦ c + t y , for any y ∈ ∂ ( C − c ) y\in\partial(C-c) y ∈ ∂ ( C − c ) , gives the same conclusion. The construction of G G G replaces h ( x ) = p ( F ( x ) − c ) h(x)=p(F(x)-c) h ( x ) = p ( F ( x ) − c ) by w ( h ( x ) ) w(h(x)) w ( h ( x )) . The map s s s just obtained would yield a continuous function r r r satisfying w ( r ( t ) ) = t w(r(t))=t w ( r ( t )) = t for 0 < t < 1 0<t<1 0 < t < 1 . Indeed, define r : ( 0 , 1 ) → [ 0 , 1 ] r:(0,1)\to[0,1] r : ( 0 , 1 ) → [ 0 , 1 ] by r ( t ) = h ( s ( t ) ) r(t)=h(s(t)) r ( t ) = h ( s ( t )) . This map is continuous, and (17) gives
w ( r ( t ) ) = p ( G ( s ( t ) ) − c ) = p ( t y ) = t p ( y ) = t ( 0 < t < 1 ) . w(r(t))=p(G(s(t))-c)=p(ty)=tp(y)=t\qquad(0<t<1). w ( r ( t )) = p ( G ( s ( t )) − c ) = p ( t y ) = tp ( y ) = t ( 0 < t < 1 ) .
Lemma 2.7 (2) would make w w w strictly increasing, contrary to its construction in part (1) of that lemma. Thus no proper extension exists, proving part (3). □ \square □
Proof of Theorem A. Let 1 ≤ α < ℵ 0 1\le\alpha<\aleph_0 1 ≤ α < ℵ 0 and β ≥ α \beta\ge\alpha β ≥ α . If C ^ 0 ( X ) \widehat{C}_0(X) C 0 ( X ) is not α \alpha α -lineable, both ( α , β ) (\alpha,\beta) ( α , β ) -lineability and ( α , β ) (\alpha,\beta) ( α , β ) -spaceability fail by their definitions. Suppose it is α \alpha α -lineable, and choose E ⊂ C ^ 0 ( X ) ∪ { 0 } E\subset\widehat{C}_0(X)\cup\{0\} E ⊂ C 0 ( X ) ∪ { 0 } with dim E = α \dim E=\alpha dim E = α . Theorem 2.8 gives an isometric image S ( E ) S(E) S ( E ) of dimension α \alpha α which has no proper admissible linear extension. If β > α \beta>\alpha β > α , any extension of dimension β \beta β would be proper. Therefore neither ( α , β ) (\alpha,\beta) ( α , β ) -lineability nor ( α , β ) (\alpha,\beta) ( α , β ) -spaceability holds. If β = α \beta=\alpha β = α , every prescribed α \alpha α -dimensional subspace serves as its own extension. It is closed because finite-dimensional subspaces of normed spaces are closed. For countably compact Hausdorff X X X , apply the countably compact case of Theorem 2.8 to an admissible subspace of C ( X ) C(X) C ( X ) . The same argument excludes every β > α \beta> \alpha β > α . On the diagonal, finite-dimensional subspaces are closed in both ambient topologies, since both are Hausdorff topological vector topologies. □ \square □
Corollary 2.9. Let X X X be locally compact Hausdorff with at least two points, and let 1 ≤ α < ℵ 0 1 \le\alpha< \aleph_0 1 ≤ α < ℵ 0 . Assume that C ^ 0 ( X ) \widehat{C}_0(X) C 0 ( X ) is α \alpha α -lineable. For every cardinal β ≥ α \beta\ge\alpha β ≥ α ,
C ^ ( X ) is ( α , β ) -lineable ⟺ α = β . \widehat{C}(X)\text{ is }(\alpha,\beta)\text{-lineable}\quad\Longleftrightarrow\quad\alpha=\beta. C ( X ) is ( α , β ) -lineable ⟺ α = β . The same equivalence holds for ( α , β ) (\alpha,\beta) ( α , β ) -spaceability in C ( X ) C(X) C ( X ) with the compact-open topology and in C b ( X ) C_b(X) C b ( X ) with the supremum norm.
Proof. Choose E ⊂ C ^ 0 ( X ) ∪ { 0 } E \subset\widehat{C}_0(X) \cup\{0\} E ⊂ C 0 ( X ) ∪ { 0 } with dim E = α \dim E=\alpha dim E = α . Theorem 2.8 gives an isometric image S ( E ) S(E) S ( E ) admitting no proper admissible extension in C ( X ) C(X) C ( X ) . This excludes ( α , β ) (\alpha,\beta) ( α , β ) -lineability, and hence both spaceability properties, when β > α \beta>\alpha β > α . For β = α \beta=\alpha β = α , every prescribed subspace is its own extension and is closed in both ambient Hausdorff topological vector spaces, since its dimension is finite. The required α \alpha α -lineability follows from E ⊂ C ^ ( X ) ∪ { 0 } E \subset\widehat{C}(X) \cup\{0\} E ⊂ C ( X ) ∪ { 0 } . □ \square □
Proof of Theorem B
By Theorem A, it remains to determine the admissible dimensions on M M M . For compact M M M , these are characterized by embedded spheres; the noncompact case requires a separate argument at infinity.
Sphere zero-sets and finite-dimensional subspaces
The criterion combines the sphere construction of [8 Theorem 2.1] with the evaluation-map method of [8 Lemmas 3.2–3.5] .
Proposition 3.1. Let X X X be a countably compact space with at least two points whose continuous real functions separate points, and let k ≥ 1 k \ge1 k ≥ 1 be an integer. The following assertions are equivalent:
C ^ ( X ) \widehat{C}(X) C ( X ) is k k k -lineable;
there exists a continuous function u : X → [ 0 , 1 ] u:X\to[0,1] u : X → [ 0 , 1 ] such that u − 1 ( 0 ) u^{-1}(0) u − 1 ( 0 ) is homeomorphic to S k − 1 S^{k-1} S k − 1 .
Whenever these conditions hold, there is a linear isometry T : ℓ 2 k → C b ( X ) T:\ell_2^k\to C_b(X) T : ℓ 2 k → C b ( X ) such that T ( ℓ 2 k ) ∖ { 0 } ⊂ C ^ ( X ) T(\ell_2^k)\setminus\{0\}\subset\widehat{C}(X) T ( ℓ 2 k ) ∖ { 0 } ⊂ C ( X ) , where ℓ 2 k \ell_2^k ℓ 2 k denotes R k \mathbb{R}^k R k with its Euclidean norm. Moreover, the implication (2)⇒ \Rightarrow ⇒ (1) holds whenever continuous real functions on X X X separate points, without countable compactness.
Proof. Assume (1), choose a basis f 1 , … , f k f_1,\ldots,f_k f 1 , … , f k of an admissible subspace, and put F = ( f 1 , … , f k ) F=(f_1,\ldots,f_k) F = ( f 1 , … , f k ) and C = conv F ( X ) C=\operatorname{conv}F(X) C = conv F ( X ) . The separation assumption implies that X X X is Hausdorff. Lemma 2.5 gives a homeomorphism from Z = F − 1 ( ∂ C ) Z=F^{-1}(\partial C) Z = F − 1 ( ∂ C ) onto ∂ C \partial C ∂ C , and Lemma 2.2 gives ∂ C ≅ S k − 1 \partial C\cong S^{k-1} ∂ C ≅ S k − 1 . The continuous function
u ( x ) = min { 1 , dist ( F ( x ) , ∂ C ) } u(x)=\min\{1,\operatorname{dist}(F(x),\partial C)\} u ( x ) = min { 1 , dist ( F ( x ) , ∂ C )} has u − 1 ( 0 ) = Z u^{-1}(0)=Z u − 1 ( 0 ) = Z , proving (2).
Conversely, assume (2) and that continuous real functions separate points of X X X . Put Z = u − 1 ( 0 ) Z=u^{-1}(0) Z = u − 1 ( 0 ) and choose a homeomorphism ϕ : Z → S k − 1 \phi:Z\to S^{k-1} ϕ : Z → S k − 1 . The set Z Z Z is compact. We extend the coordinate functions of ϕ \phi ϕ through an evaluation map into a cube. Let I I I be the set of continuous functions f : X → [ 0 , 1 ] f:X\to[0,1] f : X → [ 0 , 1 ] , and define the evaluation map
e : X ⟶ [ 0 , 1 ] I , e ( x ) = ( f ( x ) ) f ∈ I . e:X\longrightarrow[0,1]^I,\qquad e(x)=(f(x))_{f\in I}. e : X ⟶ [ 0 , 1 ] I , e ( x ) = ( f ( x ) ) f ∈ I . It is injective: a continuous real function distinguishing two points can be composed with the continuous injective function t ↦ 1 2 + 1 π arctan t t\mapsto\frac{1}{2}+\frac{1}{\pi}\arctan t t ↦ 2 1 + π 1 arctan t to give a member of I I I distinguishing them. Only the restriction of e e e to Z Z Z needs to be a topological embedding. Indeed, e ∣ Z e|_Z e ∣ Z is a continuous bijection from a compact space onto a Hausdorff space, hence a homeomorphism. The set e ( Z ) e(Z) e ( Z ) is compact and closed in the compact Hausdorff cube [ 0 , 1 ] I [0,1]^I [ 0 , 1 ] I . By the Tietze extension
theorem on this cube, each ϕ j ∘ ( e ∣ Z ) − 1 \phi_j \circ(e|_Z)^{-1} ϕ j ∘ ( e ∣ Z ) − 1 extends to a continuous A ~ j : [ 0 , 1 ] I → [ − 1 , 1 ] \widetilde{A}_j : [0,1]^I \to[-1,1] A j : [ 0 , 1 ] I → [ − 1 , 1 ] . Then A j = A ~ j ∘ e A_j=\widetilde{A}_j\circ e A j = A j ∘ e extends ϕ j \phi_j ϕ j to X X X . Writing A = ( A 1 , … , A k ) : X → R k A=(A_1,\ldots,A_k):X\to\mathbb{R}^k A = ( A 1 , … , A k ) : X → R k , define
G : X → R k , G ( x ) = ( 1 − u ( x ) ) A ( x ) max { 1 , ∥ A ( x ) ∥ } , G:X\to\mathbb{R}^k,\qquad G(x)=\frac{(1-u(x))A(x)}{\max\{1,\|A(x)\|\}}, G : X → R k , G ( x ) = max { 1 , ∥ A ( x ) ∥ } ( 1 − u ( x )) A ( x ) ,
and define T : R k → C b ( X ) T:\mathbb{R}^k\to C_b(X) T : R k → C b ( X ) by ( T a ) ( x ) = ⟨ a , G ( x ) ⟩ (Ta)(x)=\langle a,G(x)\rangle ( T a ) ( x ) = ⟨ a , G ( x )⟩ for a ∈ R k a\in\mathbb{R}^k a ∈ R k and x ∈ X x\in X x ∈ X . The map G G G is continuous, G ∣ Z = ϕ G|_Z=\phi G ∣ Z = ϕ , and ∥ G ( x ) ∥ ≤ 1 − u ( x ) ≤ 1 \|G(x)\|\leq1-u(x)\leq1 ∥ G ( x ) ∥ ≤ 1 − u ( x ) ≤ 1 for every x ∈ X x\in X x ∈ X . Thus T : ℓ 2 k → C b ( X ) T:\ell_2^k\to C_b(X) T : ℓ 2 k → C b ( X ) is linear and ∥ T a ∥ ∞ ≤ ∥ a ∥ \|Ta\|_\infty\leq\|a\| ∥ T a ∥ ∞ ≤ ∥ a ∥ .
Fix a ≠ 0 a\neq0 a = 0 and let z a = ϕ − 1 ( a / ∥ a ∥ ) z_a=\phi^{-1}(a/\|a\|) z a = ϕ − 1 ( a /∥ a ∥ ) . Then ( T a ) ( z a ) = ∥ a ∥ (Ta)(z_a)=\|a\| ( T a ) ( z a ) = ∥ a ∥ . For x ∉ Z x\notin Z x ∈ / Z , we have u ( x ) > 0 u(x)>0 u ( x ) > 0 and therefore
( T a ) ( x ) ≤ ∣ ⟨ a , G ( x ) ⟩ ∣ ≤ ∥ a ∥ ∥ G ( x ) ∥ ≤ ∥ a ∥ ( 1 − u ( x ) ) < ∥ a ∥ . (Ta)(x)\leq|\langle a,G(x)\rangle|\leq\|a\|\|G(x)\|\leq\|a\|(1-u(x))<\|a\|. ( T a ) ( x ) ≤ ∣ ⟨ a , G ( x )⟩ ∣ ≤ ∥ a ∥∥ G ( x ) ∥ ≤ ∥ a ∥ ( 1 − u ( x )) < ∥ a ∥.
For x ∈ Z x\in Z x ∈ Z , both ϕ ( x ) \phi(x) ϕ ( x ) and a / ∥ a ∥ a/\|a\| a /∥ a ∥ are unit vectors, and
∥ ϕ ( x ) − a ∥ a ∥ ∥ 2 = 2 − 2 ⟨ a , ϕ ( x ) ⟩ ∥ a ∥ . \left\|\phi(x)-\frac{a}{\|a\|}\right\|^2=2-\frac{2\langle a,\phi(x)\rangle}{\|a\|}. ϕ ( x ) − ∥ a ∥ a 2 = 2 − ∥ a ∥ 2 ⟨ a , ϕ ( x )⟩ .
Hence ( T a ) ( x ) = ∥ a ∥ (Ta)(x)=\|a\| ( T a ) ( x ) = ∥ a ∥ holds precisely when ϕ ( x ) = a / ∥ a ∥ \phi(x)=a/\|a\| ϕ ( x ) = a /∥ a ∥ , equivalently when x = z a x=z_a x = z a . Thus T a Ta T a has a unique point of maximum and ∥ T a ∥ ∞ = ∥ a ∥ \|Ta\|_\infty=\|a\| ∥ T a ∥ ∞ = ∥ a ∥ . This proves both (1) and the asserted isometric conclusion. No countable compactness was used in this implication. □
On a compact metrizable space, every embedded sphere is a zero-set, so Proposition 3.1 simplifies as follows.
Corollary 3.2. Let K K K be a compact metrizable space with at least two points, and let k ≥ 1 k\geq1 k ≥ 1 be an integer. Then C ^ ( K ) \widehat{C}(K) C ( K ) is k k k -lineable if and only if K K K contains a subset homeomorphic to S k − 1 S^{k-1} S k − 1 . Whenever these conditions hold, there is a linear isometry T : ℓ 2 k → C ( K ) T:\ell_2^k\to C(K) T : ℓ 2 k → C ( K ) such that T ( ℓ 2 k ) ∖ { 0 } ⊂ C ^ ( K ) T(\ell_2^k)\setminus\{0\}\subset\widehat{C}(K) T ( ℓ 2 k ) ∖ { 0 } ⊂ C ( K ) .
Proof. If C ^ ( K ) \widehat{C}(K) C ( K ) is k k k -lineable, Proposition 3.1 gives a zero-set in K K K homeomorphic to S k − 1 S^{k-1} S k − 1 , which is in particular a subset of K K K with the required property. Conversely, suppose that Z ⊂ K Z\subset K Z ⊂ K is homeomorphic to S k − 1 S^{k-1} S k − 1 . Since the sphere is compact, Z Z Z is compact and therefore closed in K K K . Choose a compatible metric d d d on K K K and define
u ( x ) = min { 1 , d ( x , Z ) } , d ( x , Z ) = inf z ∈ Z d ( x , z ) . u(x)=\min\{1,d(x,Z)\},\qquad d(x,Z)=\inf_{z\in Z}d(x,z). u ( x ) = min { 1 , d ( x , Z )} , d ( x , Z ) = z ∈ Z inf d ( x , z ) . The inequality ∣ d ( x , Z ) − d ( y , Z ) ∣ ≤ d ( x , y ) |d(x,Z)-d(y,Z)|\leq d(x,y) ∣ d ( x , Z ) − d ( y , Z ) ∣ ≤ d ( x , y ) shows that u u u is continuous. Since Z Z Z is closed, u ( x ) = 0 u(x)=0 u ( x ) = 0 holds precisely when x ∈ Z x\in Z x ∈ Z . Thus u − 1 ( 0 ) = Z u^{-1}(0)=Z u − 1 ( 0 ) = Z , and the proposition gives both k k k -lineability and the asserted isometry, with C b ( K ) = C ( K ) C_b(K)=C(K) C b ( K ) = C ( K ) by compactness. □
Example 3.3. Let ω 1 \omega_1 ω 1 be the first uncountable ordinal and equip L = [ 0 , ω 1 ) L=[0,\omega_1) L = [ 0 , ω 1 ) with the order topology.
The ordinal space L L L is normal Hausdorff and countably compact, but not compact. Indeed, every countable subset is contained in a compact ordinal interval [ 0 , β ] [0,\beta] [ 0 , β ] with β < ω 1 \beta<\omega_1 β < ω 1 , whereas the open cover { [ 0 , α ) : 0 < α < ω 1 } \{[0,\alpha):0<\alpha<\omega_1\} {[ 0 , α ) : 0 < α < ω 1 } has no finite subcover.
For each integer k ≥ 1 k\geq1 k ≥ 1 , consider the topological disjoint union
X k = S k − 1 ⊔ L . X_k=S^{k-1}\sqcup L. X k = S k − 1 ⊔ L . Each component is open and closed in X k X_k X k . Since a finite topological disjoint union of normal Hausdorff countably compact spaces has the same properties, X k X_k X k satisfies the hypotheses of Proposition 3.1. However, X k X_k X k is not compact, because its closed subspace L L L is not compact. The function u : X k → [ 0 , 1 ] u:X_k\to[0,1] u : X k → [ 0 , 1 ] defined by u = 0 u=0 u = 0 on S k − 1 S^{k-1} S k − 1 and u = 1 u=1 u = 1 on L L L is continuous and satisfies u − 1 ( 0 ) = S k − 1 u^{-1}(0)=S^{k-1} u − 1 ( 0 ) = S k − 1 . Hence the proposition yields k k k -lineability of C ^ ( X k ) \widehat{C}(X_k) C ( X k ) .
In this example the corresponding linear isometry can also be written explicitly:
T : ℓ 2 k → C b ( X k ) , ( T a ) ( x ) = { ⟨ a , x ⟩ , x ∈ S k − 1 , 0 , x ∈ L . T:\ell_2^k\to C_b(X_k),\qquad(Ta)(x)=
\begin{cases}
\langle a,x\rangle, & x\in S^{k-1},\\
0, & x\in L.
\end{cases} T : ℓ 2 k → C b ( X k ) , ( T a ) ( x ) = { ⟨ a , x ⟩ , 0 , x ∈ S k − 1 , x ∈ L .
The function T a Ta T a is continuous because both components are open, and ∥ T a ∥ ∞ = ∥ a ∥ \lVert Ta\rVert_\infty=\lVert a\rVert ∥ T a ∥ ∞ = ∥ a ∥ . For a ≠ 0 a\ne0 a = 0 , its maximum is ∥ a ∥ > 0 \lVert a\rVert>0 ∥ a ∥ > 0 and is attained only at a / ∥ a ∥ ∈ S k − 1 a/\lVert a\rVert\in S^{k-1} a / ∥ a ∥ ∈ S k − 1 . Indeed, on the sphere,
∥ a ∥ − ⟨ a , x ⟩ = ∥ a ∥ 2 ∥ x − a ∥ a ∥ ∥ 2 , \lVert a\rVert-\langle a,x\rangle=\frac{\lVert a\rVert}{2}\left\lVert x-\frac{a}{\lVert a\rVert}\right\rVert^2, ∥ a ∥ − ⟨ a , x ⟩ = 2 ∥ a ∥ x − ∥ a ∥ a 2 ,
whereas T a = 0 Ta=0 T a = 0 on L L L . Thus T ( ℓ 2 k ) ∖ { 0 } ⊂ C ^ ( X k ) T(\ell_2^k)\setminus\{0\}\subset\widehat{C}(X_k) T ( ℓ 2 k ) ∖ { 0 } ⊂ C ( X k ) .
Remark 3.4. The countable compactness assumption in Proposition 3.1 cannot simply be omitted: the implication (1)⇒ \Rightarrow ⇒ (2) fails even for X = R X=\mathbb{R} X = R and k = 2 k=2 k = 2 . By [13 Theorem 16] , C ^ 0 ( R ) \widehat{C}_0(\mathbb{R}) C 0 ( R ) is 2-lineable, so C ^ ( R ) \widehat{C}(\mathbb{R}) C ( R ) is also 2-lineable. However, R \mathbb{R} R contains no subset homeomorphic to S 1 S^1 S 1 . Indeed, such a subset would be compact and connected, hence a nondegenerate closed interval. Removing an interior point disconnects that interval, whereas removing any single point from S 1 S^1 S 1 leaves a connected space. Therefore R \mathbb{R} R contains no zero-set homeomorphic to S 1 S^1 S 1 , and the equivalence in Proposition 3.1 does not extend to all normal Hausdorff spaces.
Local lower bounds and countably compact manifolds
A coordinate chart gives the lower bound, with all functions supported in a common compact set.
Lemma 3.5. Let X X X be Hausdorff and suppose that some nonempty open subset of X X X is homeomorphic to an open subset of R n \mathbb{R}^n R n , where n ≥ 1 n\ge1 n ≥ 1 . There exist a compact set H ⊂ X H\subset X H ⊂ X and a linear isometry T : ℓ 2 n → C 0 ( X ) T:\ell_2^n\to C_0(X) T : ℓ 2 n → C 0 ( X ) such that every T a Ta T a vanishes outside H H H and every T a Ta T a with a ≠ 0 a\ne0 a = 0 has a unique point of maximum. In particular, C ^ 0 ( X ) \widehat{C}_0(X) C 0 ( X ) is n n n -lineable.
Proof. Choose a homeomorphism χ : U → O \chi:U\to O χ : U → O , where U ⊂ X U\subset X U ⊂ X and O ⊂ R n O\subset\mathbb{R}^n O ⊂ R n are nonempty open sets. Choose a ∈ O a\in O a ∈ O and r > 0 r>0 r > 0 such that the closed Euclidean ball of radius r r r about a a a is contained in O O O . Set V = χ − 1 ( { u : ∥ u − a ∥ < r } ) V=\chi^{-1}(\{u:\lVert u-a\rVert<r\}) V = χ − 1 ({ u : ∥ u − a ∥ < r }) . The inverse image of the corresponding closed ball is compact in X X X . It is closed because X X X is Hausdorff, and contains V V V , so V ‾ \overline{V} V is compact. The map
θ : V → R n , θ ( x ) = χ ( x ) − a r − ∥ χ ( x ) − a ∥ \theta:V\to\mathbb{R}^n,\qquad\theta(x)=\frac{\chi(x)-a}{r-\lVert\chi(x)-a\rVert} θ : V → R n , θ ( x ) = r − ∥ χ ( x ) − a ∥ χ ( x ) − a
is a homeomorphism: the inverse sends y ∈ R n y\in\mathbb{R}^n y ∈ R n to χ − 1 ( a + r y / ( 1 + ∥ y ∥ ) ) \chi^{-1}(a+ry/(1+\lVert y\rVert)) χ − 1 ( a + r y / ( 1 + ∥ y ∥)) . For 1 ≤ j ≤ n 1\le j\le n 1 ≤ j ≤ n , define
f j : X → R , f j ( x ) = { 2 θ j ( x ) 1 + ∥ θ ( x ) ∥ 2 , x ∈ V , 0 , x ∉ V . f_j:X\to\mathbb{R},\qquad f_j(x)=
\begin{cases}
\dfrac{2\theta_j(x)}{1+\lVert\theta(x)\rVert^2}, & x\in V,\\
0, & x\notin V.
\end{cases} f j : X → R , f j ( x ) = ⎩ ⎨ ⎧ 1 + ∥ θ ( x ) ∥ 2 2 θ j ( x ) , 0 , x ∈ V , x ∈ / V .
We verify continuity at points outside V V V . For R > 0 R>0 R > 0 , the set K R = θ − 1 ( { u : ∥ u ∥ ≤ R } ) K_R=\theta^{-1}(\{u:\lVert u\rVert\le R\}) K R = θ − 1 ({ u : ∥ u ∥ ≤ R }) is a compact subset of V V V , and hence is closed in X X X . If x ∈ V ∖ K R x\in V\setminus K_R x ∈ V ∖ K R , then
∣ f j ( x ) ∣ ≤ 2 ∥ θ ( x ) ∥ 1 + ∥ θ ( x ) ∥ 2 ≤ 2 ∥ θ ( x ) ∥ < 2 R . \lvert f_j(x)\rvert\le\frac{2\lVert\theta(x)\rVert}{1+\lVert\theta(x)\rVert^2}\le\frac{2}{\lVert\theta(x)\rVert}<\frac{2}{R}. ∣ f j ( x )∣ ≤ 1 + ∥ θ ( x ) ∥ 2 2 ∥ θ ( x )∥ ≤ ∥ θ ( x )∥ 2 < R 2 .
Given ε > 0 \varepsilon>0 ε > 0 , choose R > 2 / ε R>2/\varepsilon R > 2/ ε . The open set X ∖ K R X\setminus K_R X ∖ K R contains X ∖ V X\setminus V X ∖ V , and ∣ f j ∣ < ε \lvert f_j\rvert<\varepsilon ∣ f j ∣ < ε throughout it. Since f j = 0 f_j=0 f j = 0 on X ∖ V X\setminus V X ∖ V , this proves continuity there. Continuity in V V V follows from the formula. Moreover, f j f_j f j vanishes outside V ‾ \overline{V} V . For ε > 0 \varepsilon>0 ε > 0 , the set { x : ∣ f j ( x ) ∣ ≥ ε } \{x:\lvert f_j(x)\rvert\ge\varepsilon\} { x : ∣ f j ( x )∣ ≥ ε } is a closed subset of compact V ‾ \overline{V} V ; hence f j ∈ C 0 ( X ) f_j\in C_0(X) f j ∈ C 0 ( X ) .
Fix b = ( b 1 , … , b n ) ≠ 0 b=(b_1,\ldots,b_n)\ne0 b = ( b 1 , … , b n ) = 0 and put f b = ∑ j = 1 n b j f j f_b=\sum_{j=1}^n b_jf_j f b = ∑ j = 1 n b j f j . For x ∈ V x\in V x ∈ V , write u = θ ( x ) u=\theta(x) u = θ ( x ) . Then
∥ b ∥ − f b ( x ) = ∥ b ∥ − 2 ⟨ b , u ⟩ 1 + ∥ u ∥ 2 = ∥ b ∥ 1 + ∥ u ∥ 2 ∥ u − b ∥ b ∥ ∥ 2 ≥ 0. (22) \lVert b\rVert-f_b(x)=\lVert b\rVert-\frac{2\langle b,u\rangle}{1+\lVert u\rVert^2}
=\frac{\lVert b\rVert}{1+\lVert u\rVert^2}\left\lVert u-\frac{b}{\lVert b\rVert}\right\rVert^2\ge0. \tag*{(22)} ∥ b ∥ − f b ( x ) = ∥ b ∥ − 1 + ∥ u ∥ 2 2 ⟨ b , u ⟩ = 1 + ∥ u ∥ 2 ∥ b ∥ u − ∥ b ∥ b 2 ≥ 0. (22)
Equality holds precisely when x = θ − 1 ( b / ∥ b ∥ ) x=\theta^{-1}(b/\lVert b\rVert) x = θ − 1 ( b / ∥ b ∥) . Outside V V V we have f b ( x ) = 0 < ∥ b ∥ f_b(x)=0<\lVert b\rVert f b ( x ) = 0 < ∥ b ∥ , so this is the unique point of maximum on X X X . Moreover,
∣ f b ( x ) ∣ ≤ ∥ b ∥ 2 ∥ u ∥ 1 + ∥ u ∥ 2 ≤ ∥ b ∥ ( x ∈ V ) , ∥ f b ∥ ∞ = ∥ b ∥ . \lvert f_b(x)\rvert\le\lVert b\rVert\frac{2\lVert u\rVert}{1+\lVert u\rVert^2}\le\lVert b\rVert\quad(x\in V),\qquad\lVert f_b\rVert_\infty=\lVert b\rVert. ∣ f b ( x )∣ ≤ ∥ b ∥ 1 + ∥ u ∥ 2 2 ∥ u ∥ ≤ ∥ b ∥ ( x ∈ V ) , ∥ f b ∥ ∞ = ∥ b ∥ .
Thus T b = f b Tb=f_b T b = f b defines the asserted linear isometry. Every T b Tb T b vanishes outside the compact set H = V ‾ H=\overline{V} H = V .
The chart hypothesis in Lemma 3.5 is not necessary: S n − 1 S^{n-1} S n − 1 supports an admissible copy of ℓ 2 n \ell_2^n ℓ 2 n by the coordinate functions, although it has no nonempty open subset homeomorphic to an open subset of R n \mathbb{R}^n R n . For n ≥ 2 n \ge2 n ≥ 2 this follows from invariance of dimension, and for n = 1 n = 1 n = 1 from the discreteness of S 0 S^0 S 0 . The guaranteed dimension n n n cannot be increased in general: for n ≥ 2 n \ge2 n ≥ 2 , Theorem B gives the exact bound n n n on R n \mathbb{R}^n R n ; for n = 1 n = 1 n = 1 , Corollary 3.2 gives the exact bound one on [ 0 , 1 ] [0,1] [ 0 , 1 ] : it contains S 0 S^0 S 0 , but no subset homeomorphic to S 1 S^1 S 1 , since every nondegenerate compact connected subset of an interval has a cut point, whereas S 1 S^1 S 1 has none.
On a countably compact manifold, the evaluation map yields an embedded sphere. We use invariance of domain in the following form: every continuous injection from an open subset of R n \mathbb{R}^n R n into R n \mathbb{R}^n R n has open image. Consequently, a continuous injective map between topological n n n -manifolds without boundary is open when restricted to coordinate neighborhoods [14 Theorem 2B.3] .
Lemma 3.6. Let M M M be a connected countably compact Hausdorff space locally homeomorphic to R n \mathbb{R}^n R n , where n ≥ 1 n \ge1 n ≥ 1 . No second-countability assumption is imposed in this lemma.
(1) Every linear subspace of C ^ ( M ) ∪ { 0 } \widehat{C}(M) \cup\{0\} C ( M ) ∪ { 0 } has dimension at most n + 1 n+1 n + 1 .
(2) The set C ^ ( M ) \widehat{C}(M) C ( M ) is ( n + 1 ) (n+1) ( n + 1 ) -lineable if and only if M ≅ S n M \cong S^n M ≅ S n . If M ≇ S n M \not\cong S^n M ≅ S n , its largest admissible dimension is n n n .
Proof. (1) Suppose an admissible subspace had dimension at least n + 2 n+2 n + 2 . Selecting n + 2 n+2 n + 2 linearly independent members and applying Lemmas 2.5 and 2.2 would give an embedding e : S n + 1 → M e:S^{n+1}\to M e : S n + 1 → M . Choose a coordinate neighborhood U U U in S n + 1 S^{n+1} S n + 1 and a coordinate neighborhood V V V in M M M such that e ( U ) ⊂ V e(U)\subset V e ( U ) ⊂ V . Let χ : U → O ⊂ R n + 1 \chi:U\to O\subset\mathbb{R}^{n+1} χ : U → O ⊂ R n + 1 and ψ : V → P ⊂ R n \psi:V\to P\subset\mathbb{R}^n ψ : V → P ⊂ R n be the corresponding charts. Then
h = ψ ∘ e ∘ χ − 1 : O → R n h=\psi\circ e\circ\chi^{-1}:O\to\mathbb{R}^n h = ψ ∘ e ∘ χ − 1 : O → R n
is a continuous injection from a nonempty open subset of R n + 1 \mathbb{R}^{n+1} R n + 1 . Composing it with i : R n → R n + 1 i:\mathbb{R}^n\to\mathbb{R}^{n+1} i : R n → R n + 1 , i ( u ) = ( u , 0 ) i(u)=(u,0) i ( u ) = ( u , 0 ) , gives a continuous injection i ∘ h : O → R n + 1 i\circ h:O\to\mathbb{R}^{n+1} i ∘ h : O → R n + 1 . Invariance of domain [14 Theorem 2B.3] makes its image open in R n + 1 \mathbb{R}^{n+1} R n + 1 . But that image lies in R n × { 0 } \mathbb{R}^n\times\{0\} R n × { 0 } , which has empty interior, a contradiction. This also excludes infinite-dimensional admissible subspaces, since any such subspace contains n + 2 n+2 n + 2 independent members.
(2) If C ^ ( M ) \widehat{C}(M) C ( M ) is ( n + 1 ) (n+1) ( n + 1 ) -lineable, Lemmas 2.5 and 2.2 give an embedding e : S n → M e:S^n\to M e : S n → M . In coordinate charts, invariance of domain makes e e e an open map. Thus e ( S n ) e(S^n) e ( S n ) is open in M M M . It is also closed, since it is compact and M M M is Hausdorff. Connectedness of M M M and nonemptiness of e ( S n ) e(S^n) e ( S n ) therefore give e ( S n ) = M e(S^n)=M e ( S n ) = M . Thus M ≅ S n M\cong S^n M ≅ S n . Conversely, if M ≅ S n M\cong S^n M ≅ S n , the zero-set M M M itself, defined by the constant zero function, satisfies Proposition 3.1 with k = n + 1 k=n+1 k = n + 1 . Hence C ^ ( M ) \widehat{C}(M) C ( M ) is ( n + 1 ) (n+1) ( n + 1 ) -lineable. If M M M is not homeomorphic to S n S^n S n , part (1) and the exclusion of dimension n + 1 n+1 n + 1 give the upper bound n n n ; Lemma 3.5 gives the matching lower bound.
The upper bound on noncompact manifolds
A second-countable countably compact space is compact, so the preceding lemma leaves the noncompact case open. For functions in C 0 ( M ) C_0(M) C 0 ( M ) , extension by zero to the one-point compactification may introduce a second maximizer. The boundary case of Lemma 2.3 accounts for this obstruction.
Lemma 3.7. Let M M M be a connected noncompact Hausdorff second-countable n n n -manifold without boundary, where n ≥ 2 n\ge2 n ≥ 2 . Every linear subspace of C ^ 0 ( M ) ∪ { 0 } \widehat{C}_0(M)\cup\{0\} C 0 ( M ) ∪ { 0 } has dimension at most n n n .
Proof. Suppose, towards a contradiction, that there exist linearly independent functions f 1 , … , f n + 1 ∈ C 0 ( M ) f_1,\ldots,f_{n+1}\in C_0(M) f 1 , … , f n + 1 ∈ C 0 ( M ) whose span is contained in C ^ 0 ( M ) ∪ { 0 } \widehat{C}_0(M)\cup\{0\} C 0 ( M ) ∪ { 0 } . Let L = M ∪ { ∞ } L=M\cup\{\infty\} L = M ∪ { ∞ } be the one-point compactification of M M M , extend each f j f_j f j continuously to L L L by setting f j ( ∞ ) = 0 f_j(\infty)=0 f j ( ∞ ) = 0 , and define
F : L ⟶ R n + 1 , F ( x ) = ( f 1 ( x ) , … , f n + 1 ( x ) ) , C = conv F ( L ) . F:L\longrightarrow\mathbb{R}^{n+1},\qquad F(x)=(f_1(x),\ldots,f_{n+1}(x)),\qquad C=\operatorname{conv}F(L). F : L ⟶ R n + 1 , F ( x ) = ( f 1 ( x ) , … , f n + 1 ( x )) , C = conv F ( L ) .
By Lemma 2.3 (1), the set C C C is compact, 0 = F ( ∞ ) ∈ C 0=F(\infty)\in C 0 = F ( ∞ ) ∈ C , and int C ≠ ∅ \operatorname{int}C\ne\varnothing int C = ∅ . Lemma 2.2 (2) therefore gives ∂ C ≅ S n \partial C\cong S^n ∂ C ≅ S n .
Suppose, for a contradiction, that 0 ∈ int C 0 \in\operatorname{int} C 0 ∈ int C . In this case, Lemma 2.3(2) gives a homeomorphism F ∣ F − 1 ( ∂ C ) : F − 1 ( ∂ C ) → ∂ C F|_{F^{-1}(\partial C)}: F^{-1}(\partial C) \to\partial C F ∣ F − 1 ( ∂ C ) : F − 1 ( ∂ C ) → ∂ C . Since F ( ∞ ) = 0 ∉ ∂ C F(\infty)=0 \notin\partial C F ( ∞ ) = 0 ∈ / ∂ C , its inverse defines an embedding of S n S^n S n into M M M . By invariance of domain in coordinate charts, the image of this embedding is open in M M M . It is also compact and hence closed, because M M M is Hausdorff. Connectedness of M M M then implies that the image is all of M M M , contradicting noncompactness. Thus 0 ∈ ∂ C 0 \in\partial C 0 ∈ ∂ C .
Lemma 2.3(3) now provides x 0 ∈ M x_0 \in M x 0 ∈ M such that, writing v 0 = F ( x 0 ) v_0=F(x_0) v 0 = F ( x 0 ) and A = [ 0 , v 0 ] ⊂ ∂ C A=[0,v_0]\subset\partial C A = [ 0 , v 0 ] ⊂ ∂ C , every point of ∂ C ∖ A \partial C\setminus A ∂ C ∖ A has exactly one preimage in M M M , and
F − 1 ( A ) = { x 0 , ∞ } . F^{-1}(A)=\{x_0,\infty\}. F − 1 ( A ) = { x 0 , ∞ } .
We claim that v 0 = 0 v_0=0 v 0 = 0 . The set A A A has empty interior relative to ∂ C \partial C ∂ C . This is immediate if v 0 = 0 v_0=0 v 0 = 0 . If v 0 ≠ 0 v_0\ne0 v 0 = 0 and A A A contained a nonempty relatively open set, a coordinate chart of the n n n -manifold ∂ C \partial C ∂ C inside that set would give a continuous injection from a nonempty open subset of R n \mathbb{R}^n R n into the line R v 0 \mathbb{R}v_0 R v 0 . Identifying this line with R \mathbb{R} R and then including it as a coordinate line in R n \mathbb{R}^n R n , invariance of domain would make the resulting image open in R n \mathbb{R}^n R n . This is impossible, since n ≥ 2 n\ge2 n ≥ 2 and a line has empty interior in R n \mathbb{R}^n R n .
Since A A A is closed and has empty relative interior, ∂ C ∖ A \partial C\setminus A ∂ C ∖ A is dense in ∂ C \partial C ∂ C . Moreover, F ( L ) F(L) F ( L ) is compact, hence closed in R n + 1 \mathbb{R}^{n+1} R n + 1 , and contains ∂ C ∖ A \partial C\setminus A ∂ C ∖ A . Therefore ∂ C ⊂ F ( L ) \partial C\subset F(L) ∂ C ⊂ F ( L ) . If v 0 ≠ 0 v_0\ne0 v 0 = 0 , the point v 0 / 2 ∈ A v_0/2\in A v 0 /2 ∈ A would consequently have a preimage in L L L . Such a preimage would belong to F − 1 ( A ) = { x 0 , ∞ } F^{-1}(A)=\{x_0,\infty\} F − 1 ( A ) = { x 0 , ∞ } , although F ( x 0 ) = v 0 ≠ v 0 / 2 F(x_0)=v_0\ne v_0/2 F ( x 0 ) = v 0 = v 0 /2 and F ( ∞ ) = 0 ≠ v 0 / 2 F(\infty)=0\ne v_0/2 F ( ∞ ) = 0 = v 0 /2 . This contradiction proves
F ( x 0 ) = 0 , A = { 0 } , F − 1 ( { 0 } ) = { x 0 , ∞ } . (23) F(x_0)=0,\qquad A=\{0\},\qquad F^{-1}(\{0\})=\{x_0,\infty\}. \tag*{(23)} F ( x 0 ) = 0 , A = { 0 } , F − 1 ({ 0 }) = { x 0 , ∞ } . (23)
For each u ∈ ∂ C ∖ { 0 } u\in\partial C\setminus\{0\} u ∈ ∂ C ∖ { 0 } , denote its unique preimage by s ( u ) s(u) s ( u ) . This defines an injective map
s : ∂ C ∖ { 0 } ⟶ M ∖ { x 0 } , F ( s ( u ) ) = u . s:\partial C\setminus\{0\}\longrightarrow M\setminus\{x_0\},\qquad F(s(u))=u. s : ∂ C ∖ { 0 } ⟶ M ∖ { x 0 } , F ( s ( u )) = u .
We verify that s s s is continuous. Every point of ∂ C ∖ { 0 } \partial C\setminus\{0\} ∂ C ∖ { 0 } has a compact neighborhood H H H contained in this punctured boundary. The set F − 1 ( H ) F^{-1}(H) F − 1 ( H ) is compact in L L L and avoids both x 0 x_0 x 0 and ∞ \infty ∞ . The restriction F : F − 1 ( H ) → H F:F^{-1}(H)\to H F : F − 1 ( H ) → H is a continuous bijection from a compact space to a Hausdorff space, so its inverse s ∣ H s|_H s ∣ H is continuous. Since the interiors of these neighborhoods cover ∂ C ∖ { 0 } \partial C\setminus\{0\} ∂ C ∖ { 0 } , the map s s s is continuous.
Put U = s ( ∂ C ∖ { 0 } ) U=s(\partial C\setminus\{0\}) U = s ( ∂ C ∖ { 0 }) . Because ∂ C ≅ S n \partial C\cong S^n ∂ C ≅ S n , we have ∂ C ∖ { 0 } ≅ R n \partial C\setminus\{0\}\cong\mathbb{R}^n ∂ C ∖ { 0 } ≅ R n . Invariance of domain in coordinate charts shows that U U U is open in M M M and that s s s is a homeomorphism onto U U U . Furthermore,
U ‾ M ∖ U ⊂ { x 0 } . (24) \overline{U}^{M}\setminus U\subset\{x_0\}. \tag*{(24)} U M ∖ U ⊂ { x 0 } . (24)
Indeed, if x ∈ U ‾ M ∖ U x\in\overline{U}^{M}\setminus U x ∈ U M ∖ U , continuity of F F F and closedness of ∂ C \partial C ∂ C imply F ( x ) ∈ ∂ C F(x)\in\partial C F ( x ) ∈ ∂ C . If F ( x ) ≠ 0 F(x)\ne0 F ( x ) = 0 , uniqueness of its preimage would give x = s ( F ( x ) ) ∈ U x=s(F(x))\in U x = s ( F ( x )) ∈ U , a contradiction. Thus F ( x ) = 0 F(x)=0 F ( x ) = 0 , and (3.2) gives x = x 0 x=x_0 x = x 0 . It follows that U U U is both open and closed in M ∖ { x 0 } M\setminus\{x_0\} M ∖ { x 0 } .
The space M ∖ { x 0 } M\setminus\{x_0\} M ∖ { x 0 } is connected. Its connected components are open, since it is locally path connected, and closed in M ∖ { x 0 } M\setminus\{x_0\} M ∖ { x 0 } . If a component V V V did not have x 0 x_0 x 0 in its closure in M M M , then V V V would also be closed in M M M . As V V V is open in M M M , this would contradict connectedness of M M M . Thus every component has x 0 x_0 x 0 in its closure. Choose an open coordinate ball B 0 B_0 B 0 centered at x 0 x_0 x 0 . The set B 0 ∖ { x 0 } B_0\setminus\{x_0\} B 0 ∖ { x 0 } is path connected because it is homeomorphic to a punctured open ball in R n \mathbb{R}^n R n and n ≥ 2 n\ge2 n ≥ 2 . It meets every component of M ∖ { x 0 } M\setminus\{x_0\} M ∖ { x 0 } . Since a connected set lies in a single connected component, there can be only one such component. Consequently, the nonempty relatively open and closed set U U U equals M ∖ { x 0 } M\setminus\{x_0\} M ∖ { x 0 } , and hence
M ∖ { x 0 } = U ≅ R n . (25) M\setminus\{x_0\}=U\cong\mathbb{R}^n. \tag*{(25)} M ∖ { x 0 } = U ≅ R n . (25)
Finally, we show directly that this forces M M M to be compact. Let h : M ∖ { x 0 } → R n h:M\setminus\{x_0\}\to\mathbb{R}^n h : M ∖ { x 0 } → R n be a homeomorphism. Choose a compact coordinate ball B B B with x 0 x_0 x 0 in its interior, and put V = int M B V=\operatorname{int}_M B V = int M B . Since ∂ M B = B ∖ V \partial_M B=B\setminus V ∂ M B = B ∖ V is compact and avoids x 0 x_0 x 0 , there exists R > 0 R>0 R > 0 such that h ( ∂ M B ) ⊂ { y : ∥ y ∥ < R } h(\partial_M B)\subset\{y:\lVert y\rVert<R\} h ( ∂ M B ) ⊂ { y : ∥ y ∥ < R } . Define
D = { y ∈ R n : ∥ y ∥ > R } , K = h − 1 ( { y ∈ R n : ∥ y ∥ ≤ R } ) . D=\{y\in\mathbb{R}^n:\lVert y\rVert>R\},\qquad K=h^{-1}(\{y\in\mathbb{R}^n:\lVert y\rVert\le R\}). D = { y ∈ R n : ∥ y ∥ > R } , K = h − 1 ({ y ∈ R n : ∥ y ∥ ≤ R }) .
The set K K K is compact in M M M , hence closed, and does not contain x 0 x_0 x 0 . Thus V ∖ K V \setminus K V ∖ K is an open neighborhood of x 0 x_0 x 0 . Since x 0 x_0 x 0 is not isolated, this neighborhood contains a point x ≠ x 0 x \ne x_0 x = x 0 , for which h ( x ) ∈ D ∩ h ( V ∖ { x 0 } ) h(x) \in D \cap h(V \setminus\{x_0\}) h ( x ) ∈ D ∩ h ( V ∖ { x 0 }) .
The set D D D is path connected because n ≥ 2 n \ge2 n ≥ 2 : its radial projection onto the sphere of radius R + 1 R+1 R + 1 deforms it to a path-connected sphere. It is also disjoint from h ( ∂ M B ) h(\partial M_B) h ( ∂ M B ) . Hence
D = ( D ∩ h ( V ∖ { x 0 } ) ) ∪ ( D ∩ h ( M ∖ B ) ) D = \bigl(D \cap h(V \setminus\{x_0\})\bigr) \cup\bigl(D \cap h(M \setminus B)\bigr) D = ( D ∩ h ( V ∖ { x 0 }) ) ∪ ( D ∩ h ( M ∖ B ) )
is a union of two disjoint relatively open sets. The first is nonempty, so the second is empty. Therefore M ∖ B ⊂ K M \setminus B \subset K M ∖ B ⊂ K , and consequently M = B ∪ K M = B \cup K M = B ∪ K . Both B B B and K K K are compact, making M M M compact. This contradicts the hypothesis.
We have excluded admissible subspaces of dimension n + 1 n+1 n + 1 . Any subspace of larger finite or infinite dimension contains n + 1 n+1 n + 1 linearly independent functions and is therefore excluded as well. Thus every linear subspace of C ^ 0 ( M ) ∪ { 0 } \widehat{C}_0(M) \cup\{0\} C 0 ( M ) ∪ { 0 } has dimension at most n n n . The restriction n ≥ 2 n \ge2 n ≥ 2 was used to show that A A A has empty relative interior in ∂ C \partial C ∂ C , that a coordinate ball punctured at its center is connected, and that the exterior of a Euclidean ball is connected. These are precisely the steps unavailable in dimension one. □
Proof of Theorem B. If M ≅ S n M \cong S^n M ≅ S n , Lemma 3.6 gives largest admissible dimension n + 1 n+1 n + 1 . For compact M ≇ S n M \not\cong S^n M ≅ S n , the same lemma gives dimension n n n . For noncompact M M M , Lemma 3.5 supplies dimension n n n , and Lemma 3.7 excludes larger dimensions. Thus the possible positive dimensions are precisely those stated in the theorem, and all are finite. Theorem A gives both extension properties exactly when α = β \alpha=\beta α = β lies in the corresponding range. □
Spaceability on countably compact domains
Closed infinite-dimensional admissible subspaces can exist even when every prescribed infinite-dimensional spaceability property fails. The obstruction comes from a sequence converging to a nonzero constant.
Infinite prescribed dimensions
For the supremum norm, the following construction is the obstruction from [11 Theorem 2.1] . Its uniform convergence also gives the compact-open conclusion.
Lemma 4.1. Let X X X be a topological space with at least two points, and let α ≥ ℵ 0 \alpha\ge\aleph_0 α ≥ ℵ 0 be a cardinal. If C ^ ( X ) \widehat{C}(X) C ( X ) is α \alpha α -lineable, there is an α \alpha α -dimensional subspace V ⊂ C ^ ( X ) ∪ { 0 } V \subset\widehat{C}(X) \cup\{0\} V ⊂ C ( X ) ∪ { 0 } for which no closed subspace of C ( X ) C(X) C ( X ) containing V V V is contained in C ^ ( X ) ∪ { 0 } \widehat{C}(X) \cup\{0\} C ( X ) ∪ { 0 } . The same conclusion holds for closed extensions in C b ( X ) C_b(X) C b ( X ) . Consequently, for every ℵ 0 ≤ α ≤ β \aleph_0 \le\alpha\le\beta ℵ 0 ≤ α ≤ β , the set C ^ ( X ) \widehat{C}(X) C ( X ) is not ( α , β ) (\alpha,\beta) ( α , β ) -spaceable in either ambient topology.
Proof. Choose an α \alpha α -dimensional subspace E ⊂ C ^ ( X ) ∪ { 0 } E \subset\widehat{C}(X) \cup\{0\} E ⊂ C ( X ) ∪ { 0 } and a Hamel basis ( f i ) i ∈ I (f_i)_{i \in I} ( f i ) i ∈ I , where card ( I ) = α \operatorname{card}(I)=\alpha card ( I ) = α . Every f i f_i f i is nonzero and bounded, so replacing it by f i / ∥ f i ∥ ∞ f_i/\lVert f_i\rVert_\infty f i / ∥ f i ∥ ∞ gives a basis with ∥ f i ∥ ∞ = 1 \lVert f_i\rVert_\infty=1 ∥ f i ∥ ∞ = 1 for all i i i . Let 1 : X → R \mathbf{1}:X \to\mathbb{R} 1 : X → R be the constant function with value one. A nonzero constant has at least two maximizing points on X X X , so
E ∩ span { 1 } = { 0 } . (26) E \cap\operatorname{span}\{\mathbf{1}\}=\{0\}. \tag*{(26)} E ∩ span { 1 } = { 0 } . (26)
Since I I I is infinite, choose distinct indices i m ∈ I i_m \in I i m ∈ I for m ∈ N m \in\mathbb{N} m ∈ N . Define
t i = { 1 / m , i = i m for some m ∈ N , 1 , i ∉ { i m : m ∈ N } , φ i : X → R , φ i = 1 + t i f i , t_i =
\begin{cases}
1/m, & i=i_m \text{ for some }m \in\mathbb{N},\\
1, & i \notin\{i_m:m \in\mathbb{N}\},
\end{cases}
\qquad
\varphi_i:X \to\mathbb{R},
\qquad
\varphi_i=1+t_i f_i, t i = { 1/ m , 1 , i = i m for some m ∈ N , i ∈ / { i m : m ∈ N } , φ i : X → R , φ i = 1 + t i f i ,
and put V = span { φ i : i ∈ I } V=\operatorname{span}\{\varphi_i:i \in I\} V = span { φ i : i ∈ I } . We have dim V = α \dim V=\alpha dim V = α . Let J ⊂ I J \subset I J ⊂ I be finite, and suppose ∑ i ∈ J a i φ i = 0 \sum_{i \in J}a_i\varphi_i=0 ∑ i ∈ J a i φ i = 0 . Expanding the definition gives
∑ i ∈ J a i t i f i = − ( ∑ i ∈ J a i ) 1. \sum_{i \in J} a_i t_i f_i=-\left(\sum_{i \in J}a_i\right)\mathbf{1}. i ∈ J ∑ a i t i f i = − ( i ∈ J ∑ a i ) 1 .
The left side lies in E E E and the right side in span { 1 } \operatorname{span}\{1\} span { 1 } . By (26) , both sides are zero. Independence of the f i f_i f i gives a i t i = 0 a_i t_i=0 a i t i = 0 for every i ∈ J i\in J i ∈ J . Since t i > 0 t_i>0 t i > 0 , all coefficients a i a_i a i vanish. Thus ( φ i ) i ∈ I (\varphi_i)_{i\in I} ( φ i ) i ∈ I is independent, and dim V = card ( I ) = α \dim V=\operatorname{card}(I)=\alpha dim V = card ( I ) = α .
Next, take a nontrivial finite coefficient family ( a i ) i ∈ J (a_i)_{i\in J} ( a i ) i ∈ J and put
f = ∑ i ∈ J a i t i f i , c = ∑ i ∈ J a i . f=\sum_{i\in J}a_i t_i f_i,\qquad c=\sum_{i\in J}a_i. f = i ∈ J ∑ a i t i f i , c = i ∈ J ∑ a i .
Independence of the f i f_i f i and positivity of t i t_i t i give f ≠ 0 f\ne0 f = 0 . The corresponding element of V V V is f + c 1 f+c1 f + c 1 . For every x ∈ X x\in X x ∈ X ,
( f + c 1 ) ( x ) = sup y ∈ X ( f ( y ) + c ) ⟺ f ( x ) + c = sup y ∈ X f ( y ) + c (f+c1)(x)=\sup_{y\in X}(f(y)+c)\quad\Longleftrightarrow\quad f(x)+c=\sup_{y\in X}f(y)+c ( f + c 1 ) ( x ) = y ∈ X sup ( f ( y ) + c ) ⟺ f ( x ) + c = y ∈ X sup f ( y ) + c
⟺ f ( x ) = sup y ∈ X f ( y ) . \Longleftrightarrow\quad f(x)=\sup_{y\in X}f(y). ⟺ f ( x ) = y ∈ X sup f ( y ) .
Hence Argmax X ( f + c 1 ) = Argmax X f \operatorname{Argmax}_X(f+c1)=\operatorname{Argmax}_X f Argmax X ( f + c 1 ) = Argmax X f , which is a singleton. Thus V ⊂ C ^ ( X ) ∪ { 0 } V\subset\widehat{C}(X)\cup\{0\} V ⊂ C ( X ) ∪ { 0 } .
Finally,
∥ φ i m − 1 ∥ ∞ = t i m ∥ f i m ∥ ∞ = 1 m ⟶ 0. (27) \|\varphi_{i_m}-1\|_\infty=t_{i_m}\|f_{i_m}\|_\infty=\frac{1}{m}\longrightarrow0. \tag*{(27)} ∥ φ i m − 1 ∥ ∞ = t i m ∥ f i m ∥ ∞ = m 1 ⟶ 0. (27)
For each nonempty compact H ⊂ X H\subset X H ⊂ X , the defining seminorm of the compact-open topology satisfies
sup x ∈ H ∣ φ i m ( x ) − 1 ∣ ≤ ∥ φ i m − 1 ∥ ∞ = 1 m . \sup_{x\in H}|\varphi_{i_m}(x)-1|\le\|\varphi_{i_m}-1\|_\infty=\frac{1}{m}. x ∈ H sup ∣ φ i m ( x ) − 1∣ ≤ ∥ φ i m − 1 ∥ ∞ = m 1 .
Thus φ i m → 1 \varphi_{i_m}\to1 φ i m → 1 both uniformly and in the compact-open topology. Any subspace closed in either ambient topology and containing V V V must therefore contain 1 1 1 . But 1 ≠ 0 1\ne0 1 = 0 and Argmax X 1 = X \operatorname{Argmax}_X1=X Argmax X 1 = X has at least two points. Such a closed subspace cannot be contained in C ^ ( X ) ∪ { 0 } \widehat{C}(X)\cup\{0\} C ( X ) ∪ { 0 } . If α \alpha α -lineability fails, ( α , β ) (\alpha,\beta) ( α , β ) -spaceability fails by definition. Otherwise, the subspace just constructed excludes it for every β ≥ α \beta\ge\alpha β ≥ α , proving the final assertion. □ \square □
The restriction to infinite α \alpha α in this lemma is essential. Every finite-dimensional admissible subspace is already closed in both ambient topologies and is therefore a closed admissible extension of itself. The compact-open topology is Hausdorff on C ( X ) C(X) C ( X ) for any topological space X X X , since its seminorms include evaluation at each singleton. Thus no separation axiom or local compactness is needed for the infinite-dimensional obstruction.
Corollary 4.2. Let X X X be a countably compact space with at least two points whose continuous real functions separate points, and let 1 ≤ α ≤ β 1\le\alpha\le\beta 1 ≤ α ≤ β be cardinals. Then C ^ ( X ) \widehat{C}(X) C ( X ) is ( α , β ) (\alpha,\beta) ( α , β ) -spaceable if and only if there is k ∈ N k\in\mathbb{N} k ∈ N such that α = β = k \alpha=\beta=k α = β = k and X X X contains a zero-set homeomorphic to S k − 1 S^{k-1} S k − 1 . The equivalence holds in C b ( X ) C_b(X) C b ( X ) with the supremum norm and in C ( X ) C(X) C ( X ) with the compact-open topology.
Proof. The hypothesis on continuous functions implies that X X X is Hausdorff. For finite α \alpha α , the countably compact case of Theorem A says that C ^ ( X ) \widehat{C}(X) C ( X ) is ( α , β ) (\alpha,\beta) ( α , β ) -spaceable precisely when α = β \alpha=\beta α = β and C ^ ( X ) \widehat{C}(X) C ( X ) is α \alpha α -lineable, in either topology. Writing k = α k=\alpha k = α , Proposition 3.1 identifies this last condition with the existence of the required zero-set. For infinite α \alpha α , Lemma 4.1 excludes ( α , β ) (\alpha,\beta) ( α , β ) -spaceability in both topologies. □ \square □
Closed admissible subspaces from Banach spaces
The Hilbert-space example of [8 Example 4.1] extends to reflexive strictly convex Banach spaces: the weak unit ball is compact, and every nonzero continuous linear functional has a unique maximizer. Recall that a norm is strictly convex if ∥ ( x + y ) / 2 ∥ < 1 \|(x+y)/2\|<1 ∥ ( x + y ) /2∥ < 1 whenever x ≠ y x\ne y x = y and ∥ x ∥ = ∥ y ∥ = 1 \|x\|=\|y\|=1 ∥ x ∥ = ∥ y ∥ = 1 .
Proposition 4.3. Let E E E be a nonzero real Banach space with continuous dual E ∗ E^* E ∗ , and put B E = { x ∈ E : ∥ x ∥ ≤ 1 } B_E=\{x\in E:\|x\|\le1\} B E = { x ∈ E : ∥ x ∥ ≤ 1 } . Let K = ( B E , σ ( E , E ∗ ) ) K=(B_E,\sigma(E,E^*)) K = ( B E , σ ( E , E ∗ )) , where σ ( E , E ∗ ) \sigma(E,E^*) σ ( E , E ∗ ) is the weak topology, that is, the coarsest topology on E E E making every functional in E ∗ E^* E ∗ continuous, restricted to B E B_E B E . The following are equivalent.
E E E is reflexive and strictly convex.
(2) K K K is compact and every nonzero f ∈ E ∗ f \in E^* f ∈ E ∗ has a unique point of maximum on K K K . When these conditions hold, K K K is compact Hausdorff and
J : E ∗ ⟶ C ( K ) , ( J f ) ( x ) = f ( x ) , J:E^* \longrightarrow C(K), \qquad(Jf)(x)=f(x), J : E ∗ ⟶ C ( K ) , ( J f ) ( x ) = f ( x ) ,
is a linear isometry with closed image satisfying J ( E ∗ ) ∖ { 0 } ⊂ C ^ ( K ) J(E^*)\setminus\{0\}\subset\widehat{C}(K) J ( E ∗ ) ∖ { 0 } ⊂ C ( K ) . If E E E is separable, then K K K is metrizable.
Proof. The weak topology is Hausdorff because continuous linear functionals separate points by the Hahn–Banach theorem. Kakutani’s theorem [9 Theorem 3.17] states that B E B_E B E is weakly compact if and only if E E E is reflexive.
Assume (1). Every f ∈ E ∗ f\in E^* f ∈ E ∗ is continuous on K K K by the definition of the weak topology. Compactness and symmetry of B E B_E B E imply
max x ∈ B E f ( x ) = sup x ∈ B E ∣ f ( x ) ∣ = ∥ f ∥ . \max_{x\in B_E} f(x)=\sup_{x\in B_E}|f(x)|=\|f\|. x ∈ B E max f ( x ) = x ∈ B E sup ∣ f ( x ) ∣ = ∥ f ∥.
If f ≠ 0 f\ne0 f = 0 and f ( x ) = ∥ f ∥ f(x)=\|f\| f ( x ) = ∥ f ∥ for some x ∈ B E x\in B_E x ∈ B E , then ∥ f ∥ = f ( x ) ≤ ∥ f ∥ ∥ x ∥ \|f\|=f(x)\le\|f\|\|x\| ∥ f ∥ = f ( x ) ≤ ∥ f ∥∥ x ∥ forces ∥ x ∥ = 1 \|x\|=1 ∥ x ∥ = 1 . If distinct x , y ∈ B E x,y\in B_E x , y ∈ B E both maximized f f f , strict convexity would give
∥ f ∥ = f ( x + y 2 ) ≤ ∥ f ∥ ∥ x + y 2 ∥ < ∥ f ∥ , \|f\|=f\left(\frac{x+y}{2}\right)\le\|f\|\left\|\frac{x+y}{2}\right\|<\|f\|, ∥ f ∥ = f ( 2 x + y ) ≤ ∥ f ∥ 2 x + y < ∥ f ∥ ,
a contradiction. This proves (2).
Conversely, (2) and Kakutani’s theorem give reflexivity. If the norm were not strictly convex, there would exist distinct x , y ∈ E x,y\in E x , y ∈ E with ∥ x ∥ = ∥ y ∥ = ∥ ( x + y ) / 2 ∥ = 1 \|x\|=\|y\|=\|(x+y)/2\|=1 ∥ x ∥ = ∥ y ∥ = ∥ ( x + y ) /2∥ = 1 . By Hahn–Banach, some f ∈ E ∗ f\in E^* f ∈ E ∗ satisfies ∥ f ∥ = 1 \|f\|=1 ∥ f ∥ = 1 and f ( ( x + y ) / 2 ) = 1 f((x+y)/2)=1 f (( x + y ) /2 ) = 1 . Since f ( x ) , f ( y ) ≤ 1 f(x),f(y)\le1 f ( x ) , f ( y ) ≤ 1 and ( f ( x ) + f ( y ) ) / 2 = 1 (f(x)+f(y))/2=1 ( f ( x ) + f ( y )) /2 = 1 , both f ( x ) f(x) f ( x ) and f ( y ) f(y) f ( y ) equal one. They are distinct maximizing points, contradicting (2). Thus (1) follows.
The formula defining J J J is linear, and ∥ J f ∥ ∞ = sup x ∈ B E ∣ f ( x ) ∣ = ∥ f ∥ \|Jf\|_\infty=\sup_{x\in B_E}|f(x)|=\|f\| ∥ J f ∥ ∞ = sup x ∈ B E ∣ f ( x ) ∣ = ∥ f ∥ . The image is closed because E ∗ E^* E ∗ is complete and J J J is an isometry. Finally, if E E E is separable and reflexive, then E ∗ E^* E ∗ is separable [9 Corollary 3.27] , and the weak topology on B E B_E B E is metrizable by [9 Theorem 3.29] . □
Both assumptions are necessary for the weak unit-ball construction. For example, E = ( R 2 , ∥ ⋅ ∥ ∞ ) E=(\mathbb{R}^2,\|\cdot\|_\infty) E = ( R 2 , ∥ ⋅ ∥ ∞ ) is reflexive but not strictly convex, and f ( x 1 , x 2 ) = x 1 f(x_1,x_2)=x_1 f ( x 1 , x 2 ) = x 1 is maximized at every point of { 1 } × [ − 1 , 1 ] \{1\}\times[-1,1] { 1 } × [ − 1 , 1 ] .
For 1 ≤ r < ∞ 1\le r<\infty 1 ≤ r < ∞ , let ℓ r \ell_r ℓ r denote the Banach space of real sequences a = ( a j ) j ∈ N a=(a_j)_{j\in\mathbb{N}} a = ( a j ) j ∈ N satisfying ∑ j = 1 ∞ ∣ a j ∣ r < ∞ \sum_{j=1}^{\infty}|a_j|^r<\infty ∑ j = 1 ∞ ∣ a j ∣ r < ∞ , with norm ∥ a ∥ r = ( ∑ j = 1 ∞ ∣ a j ∣ r ) 1 / r \|a\|_r=\left(\sum_{j=1}^{\infty}|a_j|^r\right)^{1/r} ∥ a ∥ r = ( ∑ j = 1 ∞ ∣ a j ∣ r ) 1/ r . For E = ℓ q E=\ell_q E = ℓ q , this gives compact metric domains with closed admissible copies of ℓ p \ell_p ℓ p .
Corollary 4.4. For each 1 < p < ∞ 1<p<\infty 1 < p < ∞ , put q = p / ( p − 1 ) q=p/(p-1) q = p / ( p − 1 ) and K q = ( B ℓ q , σ ( ℓ q , ℓ p ) ) K_q=(B_{\ell_q},\sigma(\ell_q,\ell_p)) K q = ( B ℓ q , σ ( ℓ q , ℓ p )) . Then K q K_q K q is compact metrizable, and
J p : ℓ p ⟶ C ( K q ) , ( J p a ) ( x ) = ∑ j = 1 ∞ a j x j , J_p:\ell_p\longrightarrow C(K_q),\qquad(J_pa)(x)=\sum_{j=1}^{\infty}a_jx_j, J p : ℓ p ⟶ C ( K q ) , ( J p a ) ( x ) = j = 1 ∑ ∞ a j x j , is a linear isometry with closed admissible image. In particular, C ^ ( K q ) \widehat{C}(K_q) C ( K q ) is maximally spaceable.
For 1 ≤ α < ℵ 0 1\le\alpha<\aleph_0 1 ≤ α < ℵ 0 and β ≥ α \beta\ge\alpha β ≥ α ,
C ^ ( K q ) is ( α , β ) -lineable ⟺ α = β . \widehat{C}(K_q)\text{ is }(\alpha,\beta)\text{-lineable}\quad\Longleftrightarrow\quad\alpha=\beta. C ( K q ) is ( α , β ) -lineable ⟺ α = β . For all cardinals 1 ≤ α ≤ β 1\le\alpha\le\beta 1 ≤ α ≤ β ,
C ^ ( K q ) is ( α , β ) -spaceable ⟺ α = β < ℵ 0 . \widehat{C}(K_q)\text{ is }(\alpha,\beta)\text{-spaceable}\quad\Longleftrightarrow\quad\alpha=\beta<\aleph_0. C ( K q ) is ( α , β ) -spaceable ⟺ α = β < ℵ 0 . For a ≠ 0 a\ne0 a = 0 , the unique maximizing point is
x j a = sgn ( a j ) ∣ a j ∣ p − 1 ∥ a ∥ p p − 1 ( j ∈ N ) , x_j^a=\frac{\operatorname{sgn}(a_j)|a_j|^{p-1}}{\|a\|_p^{p-1}}\qquad(j\in\mathbb{N}), x j a = ∥ a ∥ p p − 1 sgn ( a j ) ∣ a j ∣ p − 1 ( j ∈ N ) , where $\operatorname{sgn}(0)=0.
Proof. The space ℓ q \ell_q ℓ q is separable, reflexive and strictly convex, and its dual is isometrically ℓ p \ell_p ℓ p . These standard properties of the L p L^p L p spaces apply to counting measure; see [9 Chapter 4] . Proposition 4.3 therefore gives compactness, metrizability and the asserted isometry. Hölder’s inequality gives absolute convergence of the defining series. Since ( p − 1 ) q = p (p-1)q=p ( p − 1 ) q = p ,
∑ j = 1 ∞ ∣ x j a ∣ q = ∑ j = 1 ∞ ∣ a j ∣ p ∥ a ∥ p q = 1 , ( J p a ) ( x a ) = ∑ j = 1 ∞ ∣ a j ∣ p ∥ a ∥ p p − 1 = ∥ a ∥ p . \sum_{j=1}^{\infty}|x_j^a|^q=\frac{\sum_{j=1}^{\infty}|a_j|^p}{\|a\|_p^q}=1,\qquad(J_pa)(x^a)=\frac{\sum_{j=1}^{\infty}|a_j|^p}{\|a\|_p^{p-1}}=\|a\|_p. j = 1 ∑ ∞ ∣ x j a ∣ q = ∥ a ∥ p q ∑ j = 1 ∞ ∣ a j ∣ p = 1 , ( J p a ) ( x a ) = ∥ a ∥ p p − 1 ∑ j = 1 ∞ ∣ a j ∣ p = ∥ a ∥ p .
The proposition ensures uniqueness, so this is the maximizing point.
We verify the dimension assertion explicitly. The cardinality of ℓ p ⊂ R N \ell_p\subset\mathbb{R}^{\mathbb{N}} ℓ p ⊂ R N is at most c \mathfrak{c} c . For every 0 < t < 1 0<t<1 0 < t < 1 , the sequence a t = ( t j ) j ≥ 1 a^t=(t^j)_{j\geq1} a t = ( t j ) j ≥ 1 lies in ℓ p \ell_p ℓ p , since ∑ j ≥ 1 t j p = t p / ( 1 − t p ) < ∞ \sum_{j\geq1}t^{jp}=t^p/(1-t^p)<\infty ∑ j ≥ 1 t j p = t p / ( 1 − t p ) < ∞ . For distinct t 1 , … , t r ∈ ( 0 , 1 ) t_1,\ldots,t_r\in(0,1) t 1 , … , t r ∈ ( 0 , 1 ) , a relation ∑ i = 1 r c i a t i = 0 \sum_{i=1}^r c_i a^{t_i}=0 ∑ i = 1 r c i a t i = 0 gives ∑ i = 1 r c i t i j = 0 \sum_{i=1}^r c_i t_i^j=0 ∑ i = 1 r c i t i j = 0 for 1 ≤ j ≤ r 1\leq j\leq r 1 ≤ j ≤ r . The determinant
det ( t i j ) j , i = 1 r = ( ∏ i = 1 r t i ) ∏ 1 ≤ i < ℓ ≤ r ( t ℓ − t i ) \det(t_i^j)_{\substack{j,i=1}}^r=\left(\prod_{i=1}^r t_i\right)\prod_{1\leq i<\ell\leq r}(t_\ell-t_i) det ( t i j ) j , i = 1 r = ( i = 1 ∏ r t i ) 1 ≤ i < ℓ ≤ r ∏ ( t ℓ − t i )
is nonzero. Hence all c i c_i c i vanish and dim ℓ p = c \dim\ell_p=\mathfrak{c} dim ℓ p = c . Choose a countable dense subset D D D of the compact metric space K q K_q K q . Continuous functions are determined by their restrictions to D D D , so
c = dim J p ( ℓ p ) ≤ dim C ( K q ) ≤ card C ( K q ) ≤ card ( R D ) ≤ c . \mathfrak{c}=\dim J_p(\ell_p)\leq\dim C(K_q)\leq\operatorname{card}C(K_q)\leq\operatorname{card}(\mathbb{R}^D)\leq\mathfrak{c}. c = dim J p ( ℓ p ) ≤ dim C ( K q ) ≤ card C ( K q ) ≤ card ( R D ) ≤ c .
Thus the closed image is maximally dimensional. It also supplies admissible subspaces of every positive finite dimension. Since K q K_q K q is compact, Theorem A yields the finite-dimensional extension classification. Lemma 4.1 excludes every infinite prescribed dimension for spaceability, proving the remaining assertion.
The range 1 < p < ∞ 1<p<\infty 1 < p < ∞ is sharp for the specific dual-ball model in this corollary. At p = 1 p=1 p = 1 , the corresponding domain would be B ℓ ∞ B_{\ell_\infty} B ℓ ∞ with its weak topology; it is not compact, and the coordinate functional x ↦ x 1 x\mapsto x_1 x ↦ x 1 has many maximizers. Even changing to the weak-star topology, which makes that ball compact, does not restore uniqueness. At p = ∞ p=\infty p = ∞ , the domain would be B ℓ 1 B_{\ell_1} B ℓ 1 , which is not weakly compact. Moreover, a = ( 1 − 1 / ( j + 1 ) ) j ≥ 1 ∈ ℓ ∞ a=(1-1/(j+1))_{j\geq1}\in\ell_\infty a = ( 1 − 1/ ( j + 1 ) ) j ≥ 1 ∈ ℓ ∞ defines a functional with supremum one on B ℓ 1 B_{\ell_1} B ℓ 1 and no maximizing point: for x ≠ 0 x\ne0 x = 0 ,
∑ j = 1 ∞ a j x j ≤ ∑ j = 1 ∞ a j ∣ x j ∣ < ∑ j = 1 ∞ ∣ x j ∣ ≤ 1 , ⟨ a , e j ⟩ ⟶ 1 , \sum_{j=1}^{\infty}a_jx_j\leq\sum_{j=1}^{\infty}a_j|x_j|<\sum_{j=1}^{\infty}|x_j|\leq1,\qquad\langle a,e_j\rangle\longrightarrow1, j = 1 ∑ ∞ a j x j ≤ j = 1 ∑ ∞ a j ∣ x j ∣ < j = 1 ∑ ∞ ∣ x j ∣ ≤ 1 , ⟨ a , e j ⟩ ⟶ 1 ,
while the value at x = 0 x=0 x = 0 is zero.
Proof of Theorem C
Each direction on the coefficient sphere determines the unique maximizer of the corresponding function. Directions with the same maximizer form a closed equivalence relation by boundedness of the evaluation vectors and uniqueness of the maximizing points. Its equivalence classes are contractible. The Vietoris–Begle theorem identifies the cohomology of this quotient with that of the sphere; the covering dimension of the domain bounds the dimension of the quotient.
Topological definitions and the mapping theorem
A space is σ \sigma σ -compact if it is a countable union of compact subsets. By replacing the first r r r compact sets by their union, we may always arrange a representation X = ⋃ r ≥ 1 K r X=\bigcup_{r\geq1}K_r X = ⋃ r ≥ 1 K r with K r ⊂ K r + 1 K_r\subset K_{r+1} K r ⊂ K r + 1 compact. For a metrizable space Y Y Y , write dim t o p Y \dim_{\mathrm{top}}Y dim top Y for its covering dimension. The inequality dim t o p Y ≤ m \dim_{\mathrm{top}}Y\leq m dim top Y ≤ m means that every finite open cover has an open refinement of multiplicity at most m + 1 m+1 m + 1 : no point belongs to more than m + 1 m+1 m + 1 members of the refinement. Here a cover V \mathcal{V} V refines U \mathcal{U} U if each member of V \mathcal{V} V is contained in a member of U \mathcal{U} U . This is the usual covering dimension [10 Section 3.1] ; the subscript distinguishes it from Hamel dimension. We use monotonicity on metrizable subspaces and the countable closed sum theorem [10, Theorems 4.1.3, 4.1.7 and 4.1.9}:
Y = ⋃ r ≥ 1 Y r , Y r closed in Y , dim t o p Y r ≤ m ⟹ dim t o p Y ≤ m . (28) Y=\bigcup_{r\geq1}Y_r,\qquad Y_r\text{ closed in }Y,\quad\dim_{\mathrm{top}}Y_r\leq m\quad\Longrightarrow\quad\dim_{\mathrm{top}}Y\leq m. \tag*{(28)} Y = r ≥ 1 ⋃ Y r , Y r closed in Y , dim top Y r ≤ m ⟹ dim top Y ≤ m . (28)
Also dim t o p R n = n \dim_{\mathrm{top}} \mathbb{R}^{n}=n dim top R n = n [10 Theorem 1.8.2] .
We use Čech cohomology with integer coefficients. For compact metrizable Y Y Y , it can be computed as
H ˇ j ( Y ; Z ) = lim U H j ( N ( U ) ; Z ) , (29) \check{H}^{j}(Y;\mathbb{Z})=\lim_{\mathcal{U}} H^{j}(N(\mathcal{U});\mathbb{Z}), \tag*{(29)} H ˇ j ( Y ; Z ) = U lim H j ( N ( U ) ; Z ) , (29)
where U \mathcal{U} U ranges over finite open covers ordered by refinement and N ( U ) N(\mathcal{U}) N ( U ) is the nerve of the cover. Its vertices are the members of U \mathcal{U} U , and a finite family spans a simplex exactly when its intersection is nonempty. Refinements induce the homomorphisms in this direct limit; continuous maps induce cohomology homomorphisms contravariantly. See [17 Sections 6.7–6.9] . A nonempty compact contractible space is acyclic, meaning that its reduced Čech cohomology groups vanish in every degree.
We use the following precise form of the Vietoris–Begle mapping theorem [17 Theorem 6.9.15, p. 344] : if P P P , Q Q Q are compact metrizable spaces and q : P → Q q:P\to Q q : P → Q is continuous and onto, with every fiber nonempty and contractible, then
q ∗ : H ˇ j ( Q ; Z ) ⟶ H ˇ j ( P ; Z ) (30) q^{*}:\check{H}^{j}(Q;\mathbb{Z})\longrightarrow\check{H}^{j}(P;\mathbb{Z}) \tag*{(30)} q ∗ : H ˇ j ( Q ; Z ) ⟶ H ˇ j ( P ; Z ) (30)
is an isomorphism for every j ≥ 0 j\geq0 j ≥ 0 . Indeed, q q q is closed by compactness, and contractibility makes its fibers acyclic in every degree, so the finite-degree theorem applies with an arbitrarily large degree cutoff. Spanier formulates the theorem in Alexander cohomology, which agrees with Čech cohomology for compact metrizable spaces.
Two computations will be used. For r ≥ 1 r\geq1 r ≥ 1 ,
H ˇ r ( S r ; Z ) ≅ Z . (31) \check{H}^{r}(S^{r};\mathbb{Z})\cong\mathbb{Z}. \tag*{(31)} H ˇ r ( S r ; Z ) ≅ Z . (31)
This follows from the agreement with simplicial cohomology on finite polyhedra and the usual sphere computation; see [14 Section 3.1] . On the other hand, for compact metrizable Y Y Y ,
dim t o p Y ≤ m ⟹ H ˇ j ( Y ; Z ) = 0 ( j > m ) . (32) \dim_{\mathrm{top}}Y\leq m\quad\Longrightarrow\quad\check{H}^{j}(Y;\mathbb{Z})=0\quad(j>m). \tag*{(32)} dim top Y ≤ m ⟹ H ˇ j ( Y ; Z ) = 0 ( j > m ) . (32)
To verify this directly, refine any finite open cover to one of multiplicity at most m + 1 m+1 m + 1 and take a finite subcover of the refinement. Its nerve has no simplices of dimension greater than m m m , so its cochain groups, and hence its cohomology groups, vanish in degrees greater than m m m . Every class from the original cover therefore becomes zero in such a refinement, proving (5.5) by the definition of the direct limit.
An upper bound on σ \sigma σ -compact domains
Proposition 5.1. Let X X X be a σ \sigma σ -compact metrizable space with at least two points and dim t o p X ≤ m < ∞ \dim_{\mathrm{top}}X\leq m<\infty dim top X ≤ m < ∞ , where m m m is a nonnegative integer. Every admissible linear subspace of C ( X ) C(X) C ( X ) has Hamel dimension at most m + 1 m+1 m + 1 .
Proof. It suffices to bound a finite-dimensional admissible subspace E = span { f 1 , … , f d } E=\operatorname{span}\{f_{1},\ldots,f_{d}\} E = span { f 1 , … , f d } with d ≥ 2 d\geq2 d ≥ 2 ; if a larger finite or infinite dimension existed, it would contain such a subspace with d = m + 2 d=m+2 d = m + 2 . For m a t h b f a ∈ R d mathbf{a}\in\mathbb{R}^{d} ma t hb f a ∈ R d , put
f a = ∑ i = 1 d a i f i , F ( x ) = ( f 1 ( x ) , … , f d ( x ) ) , H ( a ) = sup x ∈ X ⟨ a , F ( x ) ⟩ . f_{\mathbf{a}}=\sum_{i=1}^{d}a_{i}f_{i},\qquad F(x)=(f_{1}(x),\ldots,f_{d}(x)),\qquad H(\mathbf{a})=\sup_{x\in X}\langle\mathbf{a},F(x)\rangle. f a = i = 1 ∑ d a i f i , F ( x ) = ( f 1 ( x ) , … , f d ( x )) , H ( a ) = x ∈ X sup ⟨ a , F ( x )⟩ .
Every f i f_{i} f i is bounded, since both f i f_{i} f i and − f i -f_{i} − f i attain a finite maximum. Consequently F ( X ) F(X) F ( X ) is bounded and H H H is finite, with
∣ H ( a ) − H ( b ) ∣ ≤ sup x ∈ X ∣ f a ( x ) − f b ( x ) ∣ ≤ ∑ i = 1 d ∣ a i − b i ∣ ∥ f i ∥ ∞ . (33) \lvert H(\mathbf{a})-H(\mathbf{b})\rvert\leq\sup_{x\in X}\lvert f_{\mathbf{a}}(x)-f_{\mathbf{b}}(x)\rvert\leq\sum_{i=1}^{d}\lvert a_{i}-b_{i}\rvert\lVert f_{i}\rVert_{\infty}. \tag*{(33)} ∣ H ( a ) − H ( b )∣ ≤ x ∈ X sup ∣ f a ( x ) − f b ( x )∣ ≤ i = 1 ∑ d ∣ a i − b i ∣ ∥ f i ∥ ∞ . (33)
In particular, H H H is continuous. It also satisfies H ( t a ) = t H ( a ) H(t\mathbf{a})=tH(\mathbf{a}) H ( t a ) = t H ( a ) for t ≥ 0 t\geq0 t ≥ 0 and H ( a + b ) ≤ H ( a ) + H ( b ) H(\mathbf{a}+\mathbf{b})\leq H(\mathbf{a})+H(\mathbf{b}) H ( a + b ) ≤ H ( a ) + H ( b ) , by taking suprema in the corresponding pointwise identities and inequality. For a ≠ 0 \mathbf{a}\ne0 a = 0 , let x a x_{\mathbf{a}} x a be the unique maximizer of f a f_{\mathbf{a}} f a . Linear independence ensures f a ≠ 0 f_{\mathbf{a}}\ne0 f a = 0 . No nonzero constant belongs to E E E , since X X X has at least two points. We will construct a compact metrizable quotient Q Q Q of S d − 1 S^{d-1} S d − 1 such that
H ˇ d − 1 ( Q ; Z ) ≅ Z , dim t o p Q ≤ m . \check{H}^{d-1}(Q;\mathbb{Z}) \cong\mathbb{Z}, \qquad\dim_{\mathrm{top}} Q \le m. H ˇ d − 1 ( Q ; Z ) ≅ Z , dim top Q ≤ m .
By (5.5), these properties force d − 1 ≤ m d-1 \le m d − 1 ≤ m .
On S = S d − 1 S=S^{d-1} S = S d − 1 define a ∼ b a\sim b a ∼ b if x a = x b x_a=x_b x a = x b , and let R = { ( a , b ) ∈ S × S : a ∼ b } R=\{(a,b)\in S\times S:a\sim b\} R = {( a , b ) ∈ S × S : a ∼ b } . The relation R R R is closed. Suppose ( a j , b j ) ∈ R (a_j,b_j)\in R ( a j , b j ) ∈ R and ( a j , b j ) → ( a , b ) (a_j,b_j)\to(a,b) ( a j , b j ) → ( a , b ) . Set x j = x a j = x b j x_j=x_{a_j}=x_{b_j} x j = x a j = x b j . Boundedness of F ( X ) F(X) F ( X ) permits a subsequence for which F ( x j ) → v ∈ R d F(x_j)\to v\in\mathbb{R}^d F ( x j ) → v ∈ R d . Although v v v need not belong to F ( X ) F(X) F ( X ) , continuity of H H H gives
⟨ a , v ⟩ = lim j H ( a j ) = H ( a ) , ⟨ b , v ⟩ = lim j H ( b j ) = H ( b ) . (34) \langle a,v\rangle=\lim_{j}H(a_j)=H(a),\qquad\langle b,v\rangle=\lim_{j}H(b_j)=H(b). \tag*{(34)} ⟨ a , v ⟩ = j lim H ( a j ) = H ( a ) , ⟨ b , v ⟩ = j lim H ( b j ) = H ( b ) . (34)
We have a + b ≠ 0 a+b\ne0 a + b = 0 . Otherwise b = − a b=-a b = − a and H ( a ) + H ( − a ) = 0 H(a)+H(-a)=0 H ( a ) + H ( − a ) = 0 , so sup X f a = inf X f a \sup_X f_a=\inf_X f_a sup X f a = inf X f a . Then f a f_a f a would be a nonzero constant, a contradiction. Furthermore, since ⟨ a + b , F ( x j ) ⟩ ≤ H ( a + b ) \langle a+b,F(x_j)\rangle\le H(a+b) ⟨ a + b , F ( x j )⟩ ≤ H ( a + b ) , we obtain
H ( a + b ) ≤ H ( a ) + H ( b ) = ⟨ a + b , v ⟩ = lim j ⟨ a + b , F ( x j ) ⟩ ≤ H ( a + b ) . H(a+b)\le H(a)+H(b)=\langle a+b,v\rangle=\lim_{j}\langle a+b,F(x_j)\rangle\le H(a+b). H ( a + b ) ≤ H ( a ) + H ( b ) = ⟨ a + b , v ⟩ = j lim ⟨ a + b , F ( x j )⟩ ≤ H ( a + b ) .
Thus H ( a + b ) = H ( a ) + H ( b ) H(a+b)=H(a)+H(b) H ( a + b ) = H ( a ) + H ( b ) . At the maximizer z = x a + b z=x_{a+b} z = x a + b ,
0 = H ( a ) + H ( b ) − f a + b ( z ) = ( H ( a ) − f a ( z ) ) + ( H ( b ) − f b ( z ) ) . 0=H(a)+H(b)-f_{a+b}(z)=(H(a)-f_a(z))+(H(b)-f_b(z)). 0 = H ( a ) + H ( b ) − f a + b ( z ) = ( H ( a ) − f a ( z )) + ( H ( b ) − f b ( z )) .
Both summands are nonnegative, so both are zero. Hence z = x a = x b z=x_a=x_b z = x a = x b , proving ( a , b ) ∈ R (a,b)\in R ( a , b ) ∈ R . The sequential argument proves closedness because S × S S\times S S × S is metrizable.
Let Q = S / ∼ Q=S/\sim Q = S / ∼ be the set of equivalence classes and let q : S → Q q:S\to Q q : S → Q send a direction to its class. We give Q Q Q the quotient topology: O ⊂ Q O\subset Q O ⊂ Q is open exactly when q − 1 ( O ) q^{-1}(O) q − 1 ( O ) is open in S S S . We verify that Q Q Q is compact metrizable. It is compact as a continuous image of S S S . If A ⊂ S A\subset S A ⊂ S is closed, then
q − 1 ( q ( A ) ) = { a ∈ S : there is b ∈ A with ( a , b ) ∈ R } q^{-1}(q(A))=\{a\in S:\text{there is }b\in A\text{ with }(a,b)\in R\} q − 1 ( q ( A )) = { a ∈ S : there is b ∈ A with ( a , b ) ∈ R }
is the projection of the compact set R ∩ ( S × A ) R\cap(S\times A) R ∩ ( S × A ) , and is therefore closed. By the definition of the quotient topology, q ( A ) q(A) q ( A ) is closed in Q Q Q . Thus q q q is closed. Each class is a closed subset of S S S , hence compact. For two distinct classes, choose disjoint open neighborhoods U , V U,V U , V of these compact subsets in S S S . The open sets Q ∖ q ( S ∖ U ) Q\setminus q(S\setminus U) Q ∖ q ( S ∖ U ) and Q ∖ q ( S ∖ V ) Q\setminus q(S\setminus V) Q ∖ q ( S ∖ V ) contain the respective classes and are disjoint. Hence Q Q Q is Hausdorff.
Take a countable base of S S S and let B \mathcal{B} B consist of its finite unions. The countable collection
{ Q ∖ q ( S ∖ U ) : U ∈ B } \{Q\setminus q(S\setminus U):U\in\mathcal{B}\} { Q ∖ q ( S ∖ U ) : U ∈ B }
is a base of Q Q Q : if ξ ∈ O \xi\in O ξ ∈ O with O O O open, the compact class q − 1 ( { ξ } ) q^{-1}(\{\xi\}) q − 1 ({ ξ }) is covered by finitely many basic open subsets contained in q − 1 ( O ) q^{-1}(O) q − 1 ( O ) ; their union U U U gives ξ ∈ Q ∖ q ( S ∖ U ) ⊂ O \xi\in Q\setminus q(S\setminus U)\subset O ξ ∈ Q ∖ q ( S ∖ U ) ⊂ O . Thus Q Q Q is second countable. A compact Hausdorff space is regular, so the Urysohn metrization theorem [15 Theorem 34.1, p. 215] makes Q Q Q metrizable.
Each fiber of q q q is contractible. Fix a class A A A and a 0 ∈ A a_0\in A a 0 ∈ A , and write x = x a 0 x=x_{a_0} x = x a 0 . The class cannot contain antipodal points: if a , − a ∈ A a,-a\in A a , − a ∈ A , then x x x maximizes both f a f_a f a and − f a -f_a − f a , making f a f_a f a constant. Therefore
Φ : A × [ 0 , 1 ] → S , Φ ( a , t ) = ( 1 − t ) a + t a 0 ∥ ( 1 − t ) a + t a 0 ∥ (35) \Phi:A\times[0,1]\to S,\qquad\Phi(a,t)=\frac{(1-t)a+ta_0}{\|(1-t)a+ta_0\|} \tag*{(35)} Φ : A × [ 0 , 1 ] → S , Φ ( a , t ) = ∥ ( 1 − t ) a + t a 0 ∥ ( 1 − t ) a + t a 0 (35)
is well defined and continuous. Indeed, a zero denominator with 0 < t < 1 0<t<1 0 < t < 1 would give a = − a 0 a=-a_0 a = − a 0 (and t = 1 / 2 t=1/2 t = 1/2 ), which was excluded. For y ≠ x y\ne x y = x , both f a ( x ) − f a ( y ) f_a(x)-f_a(y) f a ( x ) − f a ( y ) and f a 0 ( x ) − f a 0 ( y ) f_{a_0}(x)-f_{a_0}(y) f a 0 ( x ) − f a 0 ( y ) are strictly positive. Consequently
f ( 1 − t ) a + t a 0 ( x ) − f ( 1 − t ) a + t a 0 ( y ) = ( 1 − t ) ( f a ( x ) − f a ( y ) ) + t ( f a 0 ( x ) − f a 0 ( y ) ) > 0. f_{(1-t)a+ta_0}(x)-f_{(1-t)a+ta_0}(y)=(1-t)(f_a(x)-f_a(y))+t(f_{a_0}(x)-f_{a_0}(y))>0. f ( 1 − t ) a + t a 0 ( x ) − f ( 1 − t ) a + t a 0 ( y ) = ( 1 − t ) ( f a ( x ) − f a ( y )) + t ( f a 0 ( x ) − f a 0 ( y )) > 0.
Dividing by the positive denominator in (5.8) shows that x Φ ( a , t ) = x x_{\Phi(a,t)}=x x Φ ( a , t ) = x . Thus Φ \Phi Φ takes values in A A A , with Φ ( a , 0 ) = a \Phi(a,0)=a Φ ( a , 0 ) = a and Φ ( a , 1 ) = a 0 \Phi(a,1)=a_0 Φ ( a , 1 ) = a 0 . Every fiber is therefore contractible. Equations (5.3) and (5.4) yield
H ˇ d − 1 ( Q ; Z ) ≅ H ˇ d − 1 ( S d − 1 ; Z ) ≅ Z . (36) \check{H}^{d-1}(Q;\mathbb{Z})\cong\check{H}^{d-1}(S^{d-1};\mathbb{Z})\cong\mathbb{Z}. \tag*{(36)} H ˇ d − 1 ( Q ; Z ) ≅ H ˇ d − 1 ( S d − 1 ; Z ) ≅ Z . (36)
It remains to bound dim t o p Q \dim_{\mathrm{top}} Q dim top Q from the domain X X X . Choose an increasing compact cover X = ⋃ r ≥ 1 K r X=\bigcup_{r\geq1}K_r X = ⋃ r ≥ 1 K r and put
T r = { a ∈ S : x a ∈ K r } , Q r = q ( T r ) . T_r=\{a\in S:x_a\in K_r\},\qquad Q_r=q(T_r). T r = { a ∈ S : x a ∈ K r } , Q r = q ( T r ) .
The set T r T_r T r is closed in S S S . To check this, let a j ∈ T r a_j\in T_r a j ∈ T r converge to a a a . A subsequence satisfies x a j → x ∈ K r x_{a_j}\to x\in K_r x a j → x ∈ K r , and continuity of F F F and (5.6) give f a ( x ) = lim j f a j ( x a j ) = lim j H ( a j ) = H ( a ) f_a(x)=\lim_j f_{a_j}(x_{a_j})=\lim_j H(a_j)=H(a) f a ( x ) = lim j f a j ( x a j ) = lim j H ( a j ) = H ( a ) . Uniqueness implies x = x a ∈ K r x=x_a\in K_r x = x a ∈ K r . Hence T r T_r T r is compact and Q r Q_r Q r is compact, thus closed in Q Q Q .
The map m r : T r → K r m_r:T_r\to K_r m r : T r → K r , m r ( a ) = x a m_r(a)=x_a m r ( a ) = x a , is continuous. Indeed, for any sequence a j → a a_j\to a a j → a in T r T_r T r , every convergent subsequence of the points x a j x_{a_j} x a j has limit x a x_a x a , by the same calculation. If x a j x_{a_j} x a j did not converge to x a x_a x a , compactness of K r K_r K r would give a subsequence staying outside some neighborhood of x a x_a x a and a further convergent subsequence, a contradiction.
The map q ∣ T r : T r → Q r q|_{T_r}:T_r\to Q_r q ∣ T r : T r → Q r is a continuous closed surjection, since its domain is compact and its codomain is Hausdorff. It is therefore a quotient map. Since m r m_r m r is constant precisely on its equivalence classes, it induces a continuous bijection
m ‾ r : Q r → m r ( T r ) ⊂ K r , m ‾ r ( q ( a ) ) = x a . \overline{m}_r:Q_r\to m_r(T_r)\subset K_r,\qquad\overline{m}_r(q(a))=x_a. m r : Q r → m r ( T r ) ⊂ K r , m r ( q ( a )) = x a .
For injectivity, x a = x b x_a=x_b x a = x b implies a ∼ b a\sim b a ∼ b and q ( a ) = q ( b ) q(a)=q(b) q ( a ) = q ( b ) ; surjectivity follows from the definition of m r ( T r ) m_r(T_r) m r ( T r ) . Its domain is compact and its codomain Hausdorff, so it is a homeomorphism. By monotonicity, dim t o p Q r = dim t o p m r ( T r ) ≤ m \dim_{\mathrm{top}}Q_r=\dim_{\mathrm{top}}m_r(T_r)\leq m dim top Q r = dim top m r ( T r ) ≤ m . Every maximizer belongs to some K r K_r K r , so S = ⋃ r T r S=\bigcup_r T_r S = ⋃ r T r and Q = ⋃ r Q r Q=\bigcup_r Q_r Q = ⋃ r Q r . The countable closed sum theorem (5.1) now gives dim t o p Q ≤ m \dim_{\mathrm{top}}Q\leq m dim top Q ≤ m . If d − 1 > m d-1>m d − 1 > m , (5.5) contradicts (5.9). Therefore d ≤ m + 1 d\leq m+1 d ≤ m + 1 . □
The maximizing-point map need not be continuous on the whole coefficient sphere. Restricting to the compact sets T r T_r T r is what allows the countable closed sum theorem to be applied to Q Q Q .
An isometric construction attaining the Euclidean bound
The lower-bound construction follows the sphere-extension method of [2 Theorem 5] and [8 Theorem 2.1] . A closed subset of R n \mathbb{R}^n R n admitting a continuous bijection onto S n S^n S n gives the points at which the maxima are attained.
Lemma 5.2. For every n ≥ 1 n\geq1 n ≥ 1 , there is a linear isometry T : ℓ 2 n + 1 → C b ( R n ) T:\ell_2^{n+1}\to C_b(\mathbb{R}^n) T : ℓ 2 n + 1 → C b ( R n ) whose nonzero image functions attain their maximum at exactly one point.
Proof. We first construct, by induction, a closed proper subset A n ⊂ R n A_n\subset\mathbb{R}^n A n ⊂ R n and a continuous bijection u n : A n → S n u_n:A_n\to S^n u n : A n → S n . For n = 1 n=1 n = 1 , set
A 1 = [ 0 , ∞ ) , u 1 ( t ) = ( cos 2 π t 1 + t , sin 2 π t 1 + t ) . A_1=[0,\infty),\qquad u_1(t)=\left(\cos\frac{2\pi t}{1+t},\sin\frac{2\pi t}{1+t}\right). A 1 = [ 0 , ∞ ) , u 1 ( t ) = ( cos 1 + t 2 π t , sin 1 + t 2 π t ) .
The function t ↦ t / ( 1 + t ) t\mapsto t/(1+t) t ↦ t / ( 1 + t ) maps [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) bijectively onto [ 0 , 1 ) [0,1) [ 0 , 1 ) , with inverse s ↦ s / ( 1 − s ) s\mapsto s/(1-s) s ↦ s / ( 1 − s ) , and the usual angular parametrization maps [ 0 , 1 ) [0,1) [ 0 , 1 ) bijectively onto S 1 S^1 S 1 . Thus u 1 u_1 u 1 is continuous and bijective. Only a continuous bijection is needed: the inverse is not required to be continuous. In fact, no u n u_n u n constructed below is a homeomorphism, since A n A_n A n is noncompact whereas S n S^n S n is compact.
For n ≥ 2 n\geq2 n ≥ 2 , assume A n − 1 A_{n-1} A n − 1 and u n − 1 u_{n-1} u n − 1 have been constructed. The product P = A n − 1 × R P=A_{n-1}\times\mathbb{R} P = A n − 1 × R is closed and proper in R n \mathbb{R}^n R n . Its complement is nonempty and open, so choose distinct points p + , p − ∉ P p_+,p_-\notin P p + , p − ∈ / P and put
A n = P ∪ { p + , p − } , u n ( x , t ) = ( u n − 1 ( x ) 1 + t 2 , t 1 + t 2 ) , u n ( p ± ) = ( 0 , … , 0 , ± 1 ) . A_n=P\cup\{p_+,p_-\},\qquad
u_n(x,t)=\left(\frac{u_{n-1}(x)}{\sqrt{1+t^2}},\frac{t}{\sqrt{1+t^2}}\right),\qquad
u_n(p_\pm)=(0,\ldots,0,\pm1). A n = P ∪ { p + , p − } , u n ( x , t ) = ( 1 + t 2 u n − 1 ( x ) , 1 + t 2 t ) , u n ( p ± ) = ( 0 , … , 0 , ± 1 ) .
The set A n A_n A n is closed and remains proper, because removing two points from the nonempty open set R n ∖ P \mathbb{R}^n\setminus P R n ∖ P does not exhaust it. The two added points are isolated in A n A_n A n , and P P P is open and closed in A n A_n A n , so the displayed formulas define a continuous map. On P P P ,
∥ u n ( x , t ) ∥ 2 = ∥ u n − 1 ( x ) ∥ 2 + t 2 1 + t 2 = 1. \lVert u_n(x,t)\rVert^2=\frac{\lVert u_{n-1}(x)\rVert^2+t^2}{1+t^2}=1. ∥ u n ( x , t ) ∥ 2 = 1 + t 2 ∥ u n − 1 ( x ) ∥ 2 + t 2 = 1.
For a point ( v , s ) ∈ S n (v,s) \in S^n ( v , s ) ∈ S n with ∣ s ∣ < 1 |s| < 1 ∣ s ∣ < 1 , its unique preimage in P P P is determined by
t = s 1 − s 2 , x = u n − 1 − 1 ( v 1 − s 2 ) . t=\frac{s}{\sqrt{1-s^2}},\qquad x=u_{n-1}^{-1}\left(\frac{v}{\sqrt{1-s^2}}\right). t = 1 − s 2 s , x = u n − 1 − 1 ( 1 − s 2 v ) .
Indeed, ∥ v ∥ 2 = 1 − s 2 \|v\|^2=1-s^2 ∥ v ∥ 2 = 1 − s 2 , and substitution gives t / 1 + t 2 = s t/\sqrt{1+t^2}=s t / 1 + t 2 = s and u n − 1 ( x ) / 1 + t 2 = v u_{n-1}(x)/\sqrt{1+t^2}=v u n − 1 ( x ) / 1 + t 2 = v . Neither pole is obtained on P P P , since ∣ t ∣ / 1 + t 2 < 1 |t|/\sqrt{1+t^2}<1 ∣ t ∣/ 1 + t 2 < 1 . The two added points supply precisely these poles, proving bijectivity.
Write u n = ( u n , 1 , … , u n , n + 1 ) u_n=(u_{n,1},\ldots,u_{n,n+1}) u n = ( u n , 1 , … , u n , n + 1 ) . The Tietze extension theorem [15 Theorem 35.1, p. 219] applies to the closed subset A n A_n A n of the normal space R n \mathbb{R}^n R n . It gives continuous B j : R n → [ − 1 , 1 ] B_j:\mathbb{R}^n\to[-1,1] B j : R n → [ − 1 , 1 ] extending u n , j u_{n,j} u n , j . Put B = ( B 1 , … , B n + 1 ) : R n → R n + 1 B=(B_1,\ldots,B_{n+1}):\mathbb{R}^n\to\mathbb{R}^{n+1} B = ( B 1 , … , B n + 1 ) : R n → R n + 1 and define G : R n → R n + 1 G:\mathbb{R}^n\to\mathbb{R}^{n+1} G : R n → R n + 1 by
G ( x ) = B ( x ) ( 1 + dist ( x , A n ) ) max { 1 , ∥ B ( x ) ∥ } ( x ∈ R n ) . (37) G(x)=\frac{B(x)}{(1+\operatorname{dist}(x,A_n))\max\{1,\|B(x)\|\}}\qquad(x\in\mathbb{R}^n). \tag*{(37)} G ( x ) = ( 1 + dist ( x , A n )) max { 1 , ∥ B ( x ) ∥ } B ( x ) ( x ∈ R n ) . (37)
The distance function is continuous, since ∣ dist ( x , A n ) − dist ( y , A n ) ∣ ≤ ∥ x − y ∥ |\operatorname{dist}(x,A_n)-\operatorname{dist}(y,A_n)|\le\|x-y\| ∣ dist ( x , A n ) − dist ( y , A n ) ∣ ≤ ∥ x − y ∥ . The denominator in (5.10) is positive, so G G G is continuous. For x ∈ A n x\in A_n x ∈ A n , we have B ( x ) = u n ( x ) B(x)=u_n(x) B ( x ) = u n ( x ) , ∥ B ( x ) ∥ = 1 \|B(x)\|=1 ∥ B ( x ) ∥ = 1 and dist ( x , A n ) = 0 \operatorname{dist}(x,A_n)=0 dist ( x , A n ) = 0 ; hence G ( x ) = u n ( x ) G(x)=u_n(x) G ( x ) = u n ( x ) . For x ∉ A n x\notin A_n x ∈ / A n , closedness gives dist ( x , A n ) > 0 \operatorname{dist}(x,A_n)>0 dist ( x , A n ) > 0 and
∥ G ( x ) ∥ = ∥ B ( x ) ∥ ( 1 + dist ( x , A n ) ) max { 1 , ∥ B ( x ) ∥ } ≤ 1 1 + dist ( x , A n ) < 1. \|G(x)\|=\frac{\|B(x)\|}{(1+\operatorname{dist}(x,A_n))\max\{1,\|B(x)\|\}}\le\frac{1}{1+\operatorname{dist}(x,A_n)}<1. ∥ G ( x ) ∥ = ( 1 + dist ( x , A n )) max { 1 , ∥ B ( x ) ∥ } ∥ B ( x ) ∥ ≤ 1 + dist ( x , A n ) 1 < 1.
Thus G G G takes every value of S n S^n S n at exactly one point, and every other value lies in the open unit ball.
Define ( T a ) ( x ) = ⟨ a , G ( x ) ⟩ (Ta)(x)=\langle a,G(x)\rangle ( T a ) ( x ) = ⟨ a , G ( x )⟩ for a ∈ R n + 1 a\in\mathbb{R}^{n+1} a ∈ R n + 1 . This is a linear map into C b ( R n ) C_b(\mathbb{R}^n) C b ( R n ) , with ∣ ( T a ) ( x ) ∣ ≤ ∥ a ∥ ∥ G ( x ) ∥ ≤ ∥ a ∥ |(Ta)(x)|\le\|a\|\|G(x)\|\le\|a\| ∣ ( T a ) ( x ) ∣ ≤ ∥ a ∥∥ G ( x ) ∥ ≤ ∥ a ∥ . For a ≠ 0 a\ne0 a = 0 , put x a = u n − 1 ( a / ∥ a ∥ ) x_a=u_n^{-1}(a/\|a\|) x a = u n − 1 ( a /∥ a ∥ ) . Then ( T a ) ( x a ) = ∥ a ∥ (Ta)(x_a)=\|a\| ( T a ) ( x a ) = ∥ a ∥ . If x ∉ A n x\notin A_n x ∈ / A n , the strict norm estimate gives ( T a ) ( x ) < ∥ a ∥ (Ta)(x)<\|a\| ( T a ) ( x ) < ∥ a ∥ . If x ∈ A n x\in A_n x ∈ A n , both u n ( x ) u_n(x) u n ( x ) and a / ∥ a ∥ a/\|a\| a /∥ a ∥ are unit vectors, and
∥ u n ( x ) − a ∥ a ∥ ∥ 2 = 2 − 2 ⟨ a , u n ( x ) ⟩ ∥ a ∥ . \left\|u_n(x)-\frac{a}{\|a\|}\right\|^2=2-\frac{2\langle a,u_n(x)\rangle}{\|a\|}. u n ( x ) − ∥ a ∥ a 2 = 2 − ∥ a ∥ 2 ⟨ a , u n ( x )⟩ .
Consequently ( T a ) ( x ) = ∥ a ∥ (Ta)(x)=\|a\| ( T a ) ( x ) = ∥ a ∥ if and only if u n ( x ) = a / ∥ a ∥ u_n(x)=a/\|a\| u n ( x ) = a /∥ a ∥ , equivalently x = x a x=x_a x = x a . We have proved both uniqueness of the maximum and ∥ T a ∥ ∞ = ∥ a ∥ \|Ta\|_\infty=\|a\| ∥ T a ∥ ∞ = ∥ a ∥ . Thus T T T is an isometry and its image has dimension n + 1 n+1 n + 1 . □
Proof of Theorem C. Proposition 5.1, applied to R n \mathbb{R}^n R n , bounds every admissible dimension by n + 1 n+1 n + 1 , while Lemma 5.2 realizes dimension n + 1 n+1 n + 1 . This proves the first assertion, including exclusion of infinite dimensions.
Let 1 ≤ α ≤ β 1\le\alpha\le\beta 1 ≤ α ≤ β be cardinals. If α > n + 1 \alpha>n+1 α > n + 1 , ordinary α \alpha α -lineability fails, so all three extension properties in the statement fail by definition. If α = β ≤ n + 1 \alpha=\beta\le n+1 α = β ≤ n + 1 , subspaces of the construction supply α \alpha α -lineability. Every prescribed α \alpha α -dimensional admissible subspace is its own extension and is closed in both ambient Hausdorff topological vector spaces, proving all three properties.
Suppose α < β \alpha<\beta α < β and α ≤ n + 1 \alpha\le n+1 α ≤ n + 1 . If α = n + 1 \alpha=n+1 α = n + 1 , no admissible subspace of dimension β \beta β exists, by the upper bound, so no such extension can exist. If 1 ≤ α ≤ n 1\le\alpha\le n 1 ≤ α ≤ n , Lemma 3.5 supplies an α \alpha α -dimensional admissible subspace of C 0 ( R n ) C_0(\mathbb{R}^n) C 0 ( R n ) . Corollary 2.9 then excludes ( α , β ) (\alpha,\beta) ( α , β ) -lineability in C ^ ( R n ) \widehat{C}(\mathbb{R}^n) C ( R n ) and hence excludes both spaceability properties. This argument also covers n = 1 n=1 n = 1 and α = 1 \alpha=1 α = 1 . □
Funding
The work of the second author was supported by the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq, Brazil), under grants 406457/2023-9 (CNPq/MCTI No. 10/2023), 403964/2024-5 (MCTI/CNPq No. 16/2024) and 305807/2025-0.
Declarations
Competing interests. The authors declare no competing interests.
Data availability. No datasets were generated or analyzed during the current study.
Use of generative AI. ChatGPT (OpenAI) was used for exploratory calculations, consistency checks, assistance with the organization of arguments and language editing. The mathematical arguments and final text were reviewed by the authors, who take responsibility for the content.
References
Departamento de Ciências Exatas e Tecnologia da Informação, Universidade Federal Rural do Semi-Árido, Rio Grande do Norte, Brazil Email address: anderson.barbosa@ufersa.edu.br
Departamento de Matemática, Universidade Federal da Paraíba, João Pessoa, Brazil Email address: daniel.pellegrino@academico.ufpb.br
Coordenação do Curso de Matemática — Bacharelado, Universidade Federal do Maranhão, São Luís, Brazil Email address: anselmo.junior@ufma.br