In mathematics, particularly in functional analysis, a seminorm is like a norm but need not be positive definite. Seminorms are intimately connected with convex sets: every seminorm is the Minkowski functional of some absorbing disk and, conversely, the Minkowski functional of any such set is a seminorm.

A topological vector space is locally convex if and only if its topology is induced by a family of seminorms.

Definition

Let <math>X</math> be a vector space over either the real numbers <math>\R</math> or the complex numbers <math>\Complex.</math>

A real-valued function <math>p : X \to \R</math> is called a if it satisfies the following two conditions:

  1. Subadditivity/Triangle inequality: <math>p(x + y) \leq p(x) + p(y)</math> for all <math>x, y \in X.</math>
  2. Absolute homogeneity: <math>p(s x) =|s|p(x)</math> for all <math>x \in X</math> and all scalars <math>s.</math>

These two conditions imply that <math>p(0) = 0</math> and that every seminorm <math>p</math> also has the following property:

<ol start=3>

<li>Nonnegativity: <math>p(x) \geq 0</math> for all <math>x \in X.</math></li>

</ol>

Some authors include non-negativity as part of the definition of "seminorm" (and also sometimes of "norm"), although this is not necessary since it follows from the other two properties.

By definition, a norm on <math>X</math> is a seminorm that also separates points, meaning that it has the following additional property:

<ol start=4>

<li>Positive definite/Positive/: whenever <math>x \in X</math> satisfies <math>p(x) = 0,</math> then <math>x = 0.</math></li>

</ol>

A is a pair <math>(X, p)</math> consisting of a vector space <math>X</math> and a seminorm <math>p</math> on <math>X.</math> If the seminorm <math>p</math> is also a norm then the seminormed space <math>(X, p)</math> is called a .

Since absolute homogeneity implies positive homogeneity, every seminorm is a type of function called a sublinear function. A map <math>p : X \to \R</math> is called a if it is subadditive and positive homogeneous. Unlike a seminorm, a sublinear function is necessarily nonnegative. Sublinear functions are often encountered in the context of the Hahn–Banach theorem.

A real-valued function <math>p : X \to \R</math> is a seminorm if and only if it is a sublinear and balanced function.

Examples

<ul>

<li>The on <math>X,</math> which refers to the constant <math>0</math> map on <math>X,</math> induces the indiscrete topology on <math>X.</math></li>

<li>Let <math>\mu</math> be a measure on a space <math>\Omega</math>. For an arbitrary constant <math>c \geq 1</math>, let <math>X</math> be the set of all functions <math>f: \Omega \rightarrow \mathbb{R}</math> for which

<math display="block">\lVert f \rVert_c := \left( \int_{\Omega}| f |^c \, d\mu \right)^{1/c}</math>

exists and is finite. It can be shown that <math>X</math> is a vector space, and the functional <math>\lVert \cdot \rVert_c</math> is a seminorm on <math>X</math>. However, it is not always a norm (e.g. if <math>\Omega = \mathbb{R}</math> and <math>\mu</math> is the Lebesgue measure) because <math>\lVert h \rVert_c = 0</math> does not always imply <math>h = 0</math>. To make <math>\lVert \cdot \rVert_c</math> a norm, quotient <math>X</math> by the closed subspace of functions <math>h</math> with <math>\lVert h \rVert_c = 0</math>. The resulting space, <math>L^c(\mu)</math>, has a norm induced by <math>\lVert \cdot \rVert_c</math>.</li>

<li>If <math>f</math> is any linear form on a vector space then its absolute value <math>|f|,</math> defined by <math>x \mapsto |f(x)|,</math> is a seminorm.</li>

<li>A sublinear function <math>f : X \to \R</math> on a real vector space <math>X</math> is a seminorm if and only if it is a , meaning that <math>f(-x) = f(x)</math> for all <math>x \in X.</math></li>

<li>Every real-valued sublinear function <math>f : X \to \R</math> on a real vector space <math>X</math> induces a seminorm <math>p : X \to \R</math> defined by <math>p(x) := \max \{f(x), f(-x)\}.</math></li>

<li>Any finite sum of seminorms is a seminorm. The restriction of a seminorm (respectively, norm) to a vector subspace is once again a seminorm (respectively, norm).</li>

<li>If <math>p : X \to \R</math> and <math>q : Y \to \R</math> are seminorms (respectively, norms) on <math>X</math> and <math>Y</math> then the map <math>r : X \times Y \to \R</math> defined by <math>r(x, y) = p(x) + q(y)</math> is a seminorm (respectively, a norm) on <math>X \times Y.</math> In particular, the maps on <math>X \times Y</math> defined by <math>(x, y) \mapsto p(x)</math> and <math>(x, y) \mapsto q(y)</math> are both seminorms on <math>X \times Y.</math></li>

<li>If <math>p</math> and <math>q</math> are seminorms on <math>X</math> then so are

<math display="block">(p \vee q)(x) = \max \{p(x), q(x)\}</math> and <math display="block">(p \wedge q)(x) := \inf \{p(y) + q(z) : x = y + z \text{ with } y, z \in X\}</math>

where <math>p \wedge q \leq p</math> and <math>p \wedge q \leq q.</math>

</li>

<li>The space of seminorms on <math>X</math> is generally not a distributive lattice with respect to the above operations. For example, over <math>\R^2</math>, <math>p(x, y) := \max(|x|, |y|), q(x, y) := 2|x|, r(x, y) := 2|y| </math> are such that

<math display="block">((p \vee q) \wedge (p \vee r)) (x, y) = \inf \{\max(2|x_1|, |y_1|) + \max(|x_2|, 2|y_2|) : x = x_1 + x_2 \text{ and } y = y_1 + y_2\}</math> while <math>(p \vee q \wedge r) (x, y) := \max(|x|, |y|)</math></li>

<li>If <math>L : X \to Y</math> is a linear map and <math>q : Y \to \R</math> is a seminorm on <math>Y,</math> then <math>q \circ L : X \to \R</math> is a seminorm on <math>X.</math> The seminorm <math>q \circ L</math> will be a norm on <math>X</math> if and only if <math>L</math> is injective and the restriction <math>q\big\vert_{L(X)}</math> is a norm on <math>L(X).</math></li>

</ul>

Minkowski functionals and seminorms

Seminorms on a vector space <math>X</math> are intimately tied, via Minkowski functionals, to subsets of <math>X</math> that are convex, balanced, and absorbing. Given such a subset <math>D</math> of <math>X,</math> the Minkowski functional of <math>D</math> is a seminorm. Conversely, given a seminorm <math>p</math> on <math>X,</math> the sets<math>\{x \in X : p(x) < 1\}</math> and <math>\{x \in X : p(x) \leq 1\}</math> are convex, balanced, and absorbing and furthermore, the Minkowski functional of these two sets (as well as of any set lying "in between them") is <math>p.</math>

Algebraic properties

Every seminorm is a sublinear function, and thus satisfies all properties of a sublinear function, including convexity, <math>p(0) = 0,</math> and for all vectors <math>x, y \in X</math>:

the reverse triangle inequality:

<math display=block>|p(x) - p(y)| \leq p(x - y)</math>

and also

<math display=inline>0 \leq \max \{p(x), p(-x)\}</math> and <math>p(x) - p(y) \leq p(x - y).</math>

For any vector <math>x \in X</math> and positive real <math>r > 0:</math>

<math display=block>x + \{y \in X : p(y) < r\} = \{y \in X : p(x - y) < r\}</math>

and furthermore, <math>\{x \in X : p(x) < r\}</math> is an absorbing disk in <math>X.</math>

If <math>p</math> is a sublinear function on a real vector space <math>X</math> then there exists a linear functional <math>f</math> on <math>X</math> such that <math>f \leq p</math> and furthermore, for any linear functional <math>g</math> on <math>X,</math> <math>g \leq p</math> on <math>X</math> if and only if <math>g^{-1}(1) \cap \{x \in X : p(x) < 1\} = \varnothing.</math>

Other properties of seminorms

Every seminorm is a balanced function.

A seminorm <math>p</math> is a norm on <math>X</math> if and only if <math>\{x \in X : p(x) < 1\}</math> does not contain a non-trivial vector subspace.

If <math>p : X \to [0, \infty)</math> is a seminorm on <math>X</math> then <math>\ker p := p^{-1}(0)</math> is a vector subspace of <math>X</math> and for every <math>x \in X,</math> <math>p</math> is constant on the set <math>x + \ker p = \{x + k : p(k) = 0\}</math> and equal to <math>p(x).</math>

Furthermore, for any real <math>r > 0,</math>

<math display="block">r \{x \in X : p(x) < 1\} = \{x \in X : p(x) < r\} = \left\{x \in X : \tfrac{1}{r} p(x) < 1 \right\}.</math>

If <math>D</math> is a set satisfying <math>\{x \in X : p(x) < 1\} \subseteq D \subseteq \{x \in X : p(x) \leq 1\}</math> then <math>D</math> is absorbing in <math>X</math> and <math>p = p_D</math> where <math>p_D</math> denotes the Minkowski functional associated with <math>D</math> (that is, the gauge of <math>D</math>). In particular, if <math>D</math> is as above and <math>q</math> is any seminorm on <math>X,</math> then <math>q = p</math> if and only if <math>\{x \in X : q(x) < 1\} \subseteq D \subseteq \{x \in X : q(x) \leq\}.</math>

If <math>(X, \|\,\cdot\,\|)</math> is a normed space and <math>x, y \in X</math> then <math>\|x - y\| = \|x - z\| + \|z - y\|</math> for all <math>z</math> in the interval <math>[x, y].</math>

Every norm is a convex function and consequently, finding a global maximum of a norm-based objective function is sometimes tractable.

Relationship to other norm-like concepts

Let <math>p : X \to \R</math> be a non-negative function. The following are equivalent:

<ol>

<li><math>p</math> is a seminorm.</li>

<li><math>p</math> is a convex <math>F</math>-seminorm.</li>

<li><math>p</math> is a convex balanced G-seminorm.</li>

</ol>

If any of the above conditions hold, then the following are equivalent:

<ol>

<li><math>p</math> is a norm;</li>

<li><math>\{x \in X : p(x) < 1\}</math> does not contain a non-trivial vector subspace.</li>

<li>There exists a norm on <math>X,</math> with respect to which, <math>\{x \in X : p(x) < 1\}</math> is bounded.</li>

</ol>

If <math>p</math> is a sublinear function on a real vector space <math>X</math> then the following are equivalent:

<ol>

<li><math>p</math> is a linear functional;</li>

<li><math>p(x) + p(-x) \leq 0 \text{ for every } x \in X</math>;</li>

<li><math>p(x) + p(-x) = 0 \text{ for every } x \in X</math>;</li>

</ol>

Inequalities involving seminorms

If <math>p, q : X \to [0, \infty)</math> are seminorms on <math>X</math> then:

<ul>

<li><math>p \leq q</math> if and only if <math>q(x) \leq 1</math> implies <math>p(x) \leq 1.</math></li>

<li>If <math>a > 0</math> and <math>b > 0</math> are such that <math>p(x) < a</math> implies <math>q(x) \leq b,</math> then <math>a q(x) \leq b p(x)</math> for all <math>x \in X.</math> </li>

<li>Suppose <math>a</math> and <math>b</math> are positive real numbers and <math>q, p_1, \ldots, p_n</math> are seminorms on <math>X</math> such that for every <math>x \in X,</math> if <math>\max \{p_1(x), \ldots, p_n(x)\} < a</math> then <math>q(x) < b.</math> Then <math>a q \leq b \left(p_1 + \cdots + p_n\right).</math></li>

<li>If <math>X</math> is a vector space over the reals and <math>f</math> is a non-zero linear functional on <math>X,</math> then <math>f \leq p</math> if and only if <math>\varnothing = f^{-1}(1) \cap \{x \in X : p(x) < 1\}.</math></li>

</ul>

If <math>p</math> is a seminorm on <math>X</math> and <math>f</math> is a linear functional on <math>X</math> then:

<ul>

<li><math>|f| \leq p</math> on <math>X</math> if and only if <math>\operatorname{Re} f \leq p</math> on <math>X</math> (see footnote for proof).</li>

<li><math>f \leq p</math> on <math>X</math> if and only if <math>f^{-1}(1) \cap \{x \in X : p(x) < 1 = \varnothing\}.</math></li>

<li>If <math>a > 0</math> and <math>b > 0</math> are such that <math>p(x) < a</math> implies <math>f(x) \neq b,</math> then <math>a |f(x)| \leq b p(x)</math> for all <math>x \in X.</math></li>

</ul>

Hahn–Banach theorem for seminorms

Seminorms offer a particularly clean formulation of the Hahn–Banach theorem:

:If <math>M</math> is a vector subspace of a seminormed space <math>(X, p)</math> and if <math>f</math> is a continuous linear functional on <math>M,</math> then <math>f</math> may be extended to a continuous linear functional <math>F</math> on <math>X</math> that has the same norm as <math>f.</math>

A similar extension property also holds for seminorms:

:Proof: Let <math>S</math> be the convex hull of <math>\{m \in M : p(m) \leq 1\} \cup \{x \in X : q(x) \leq 1\}.</math> Then <math>S</math> is an absorbing disk in <math>X</math> and so the Minkowski functional <math>P</math> of <math>S</math> is a seminorm on <math>X.</math> This seminorm satisfies <math>p = P</math> on <math>M</math> and <math>P \leq q</math> on <math>X.</math> <math>\blacksquare</math>

Topologies of seminormed spaces

Pseudometrics and the induced topology

A seminorm <math>p</math> on <math>X</math> induces a topology, called the , via the canonical translation-invariant pseudometric <math>d_p : X \times X \to \R</math>; <math>d_p(x, y) := p(x - y) = p(y - x).</math>

This topology is Hausdorff if and only if <math>d_p</math> is a metric, which occurs if and only if <math>p</math> is a norm.

This topology makes <math>X</math> into a locally convex pseudometrizable topological vector space that has a bounded neighborhood of the origin and a neighborhood basis at the origin consisting of the following open balls (or the closed balls) centered at the origin:

<math display=block>\{x \in X : p(x) < r\} \quad \text{ or } \quad \{x \in X : p(x) \leq r\}</math>

as <math>r > 0</math> ranges over the positive reals.

Every seminormed space <math>(X, p)</math> should be assumed to be endowed with this topology unless indicated otherwise. A topological vector space whose topology is induced by some seminorm is called .

Equivalently, every vector space <math>X</math> with seminorm <math>p</math> induces a vector space quotient <math>X / W,</math> where <math>W</math> is the subspace of <math>X</math> consisting of all vectors <math>x \in X</math> with <math>p(x) = 0.</math> Then <math>X / W</math> carries a norm defined by <math>p(x + W) = p(x).</math> The resulting topology, pulled back to <math>X,</math> is precisely the topology induced by <math>p.</math>

Any seminorm-induced topology makes <math>X</math> locally convex, as follows. If <math>p</math> is a seminorm on <math>X</math> and <math>r \in \R,</math> call the set <math>\{x \in X : p(x) < r\}</math> the ; likewise the closed ball of radius <math>r</math> is <math>\{x \in X : p(x) \leq r\}.</math> The set of all open (resp. closed) <math>p</math>-balls at the origin forms a neighborhood basis of convex balanced sets that are open (resp. closed) in the <math>p</math>-topology on <math>X.</math>

Stronger, weaker, and equivalent seminorms

The notions of stronger and weaker seminorms are akin to the notions of stronger and weaker norms. If <math>p</math> and <math>q</math> are seminorms on <math>X,</math> then we say that <math>q</math> is than <math>p</math> and that <math>p</math> is than <math>q</math> if any of the following equivalent conditions holds:

  1. The topology on <math>X</math> induced by <math>q</math> is finer than the topology induced by <math>p.</math>
  2. If <math>x_{\bull} = \left(x_i\right)_{i=1}^{\infty}</math> is a sequence in <math>X,</math> then <math>q\left(x_{\bull}\right) := \left(q\left(x_i\right)\right)_{i=1}^{\infty} \to 0</math> in <math>\R</math> implies <math>p\left(x_{\bull}\right) \to 0</math> in <math>\R.</math>
  3. If <math>x_{\bull} = \left(x_i\right)_{i \in I}</math> is a net in <math>X,</math> then <math>q\left(x_{\bull}\right) := \left(q\left(x_i\right)\right)_{i \in I} \to 0</math> in <math>\R</math> implies <math>p\left(x_{\bull}\right) \to 0</math> in <math>\R.</math>
  4. <math>p</math> is bounded on <math>\{x \in X : q(x) < 1\}.</math>
  5. If <math>\inf{} \{q(x) : p(x) = 1, x \in X\} = 0</math> then <math>p(x) = 0</math> for all <math>x \in X.</math>
  6. There exists a real <math>K > 0</math> such that <math>p \leq K q</math> on <math>X.</math>

The seminorms <math>p</math> and <math>q</math> are called if they are both weaker (or both stronger) than each other. This happens if they satisfy any of the following conditions:

<ol>

<li>The topology on <math>X</math> induced by <math>q</math> is the same as the topology induced by <math>p.</math></li>

<li><math>q</math> is stronger than <math>p</math> and <math>p</math> is stronger than <math>q.</math></li>

<li>If <math>x_{\bull} = \left(x_i\right)_{i=1}^{\infty}</math> is a sequence in <math>X</math> then <math>q\left(x_{\bull}\right) := \left(q\left(x_i\right)\right)_{i=1}^{\infty} \to 0</math> if and only if <math>p\left(x_{\bull}\right) \to 0.</math></li>

<li>There exist positive real numbers <math>r > 0</math> and <math>R > 0</math> such that <math>r q \leq p \leq R q.</math></li>

</ol>

Normability and seminormability

A topological vector space (TVS) is said to be a (respectively, a ) if its topology is induced by a single seminorm (resp. a single norm).

A TVS is normable if and only if it is seminormable and Hausdorff or equivalently, if and only if it is seminormable and T<sub>1</sub> (because a TVS is Hausdorff if and only if it is a T<sub>1</sub> space).

A is a topological vector space that possesses a bounded neighborhood of the origin.

Normability of topological vector spaces is characterized by Kolmogorov's normability criterion.

A TVS is seminormable if and only if it has a convex bounded neighborhood of the origin.

Thus a locally convex TVS is seminormable if and only if it has a non-empty bounded open set.

A TVS is normable if and only if it is a T<sub>1</sub> space and admits a bounded convex neighborhood of the origin.

If <math>X</math> is a Hausdorff locally convex TVS then the following are equivalent:

<ol>

<li><math>X</math> is normable.</li>

<li><math>X</math> is seminormable.</li>

<li><math>X</math> has a bounded neighborhood of the origin.</li>

<li>The strong dual <math>X^{\prime}_b</math> of <math>X</math> is normable.</li>

<li>The strong dual <math>X^{\prime}_b</math> of <math>X</math> is metrizable.</li>

</ol>

Furthermore, <math>X</math> is finite dimensional if and only if <math>X^{\prime}_{\sigma}</math> is normable (here <math>X^{\prime}_{\sigma}</math> denotes <math>X^{\prime}</math> endowed with the weak-* topology).

The product of infinitely many seminormable space is again seminormable if and only if all but finitely many of these spaces trivial (that is, 0-dimensional).

Topological properties

<ul>

<li>If <math>X</math> is a TVS and <math>p</math> is a continuous seminorm on <math>X,</math> then the closure of <math>\{x \in X : p(x) < r\}</math> in <math>X</math> is equal to <math>\{x \in X : p(x) \leq r\}.</math></li>

<li>The closure of <math>\{0\}</math> in a locally convex space <math>X</math> whose topology is defined by a family of continuous seminorms <math>\mathcal{P}</math> is equal to <math>\bigcap_{p \in \mathcal{P p^{-1}(0).</math></li>

<li>A subset <math>S</math> in a seminormed space <math>(X, p)</math> is bounded if and only if <math>p(S)</math> is bounded.</li>

<li>If <math>(X, p)</math> is a seminormed space then the locally convex topology that <math>p</math> induces on <math>X</math> makes <math>X</math> into a pseudometrizable TVS with a canonical pseudometric given by <math>d(x, y) := p(x - y)</math> for all <math>x, y \in X.</math></li>

<li>The product of infinitely many seminormable spaces is again seminormable if and only if all but finitely many of these spaces are trivial (that is, 0-dimensional).</li>

</ul>

Continuity of seminorms

If <math>p</math> is a seminorm on a topological vector space <math>X,</math> then the following are equivalent:

<ol>

<li><math>p</math> is continuous.</li>

<li><math>p</math> is continuous at 0;</li>

<li><math>\{x \in X : p(x) < 1\}</math> is open in <math>X</math>;</li>

<li><math>\{x \in X : p(x) \leq 1\}</math> is closed neighborhood of 0 in <math>X</math>;</li>

<li><math>p</math> is uniformly continuous on <math>X</math>;</li>

<li>There exists a continuous seminorm <math>q</math> on <math>X</math> such that <math>p \leq q.</math></li>

</ol>

In particular, if <math>(X, p)</math> is a seminormed space then a seminorm <math>q</math> on <math>X</math> is continuous if and only if <math>q</math> is dominated by a positive scalar multiple of <math>p.</math>

If <math>X</math> is a real TVS, <math>f</math> is a linear functional on <math>X,</math> and <math>p</math> is a continuous seminorm (or more generally, a sublinear function) on <math>X,</math> then <math>f \leq p</math> on <math>X</math> implies that <math>f</math> is continuous.

Continuity of linear maps

If <math>F : (X, p) \to (Y, q)</math> is a map between seminormed spaces then let

<math display="block">\|F\|_{p,q} := \sup \{q(F(x)) : p(x) \leq 1, x \in X\}.</math>

If <math>F : (X, p) \to (Y, q)</math> is a linear map between seminormed spaces then the following are equivalent:

<ol>

<li><math>F</math> is continuous;</li>

<li><math>\|F\|_{p,q} < \infty</math>;</li>

<li>There exists a real <math>K \geq 0</math> such that <math>p \leq K q</math>;

  • In this case, <math>\|F\|_{p,q} \leq K.</math></li>

</ol>

If <math>F</math> is continuous then <math>q(F(x)) \leq \|F\|_{p,q} p(x)</math> for all <math>x \in X.</math>

The space of all continuous linear maps <math>F : (X, p) \to (Y, q)</math> between seminormed spaces is itself a seminormed space under the seminorm <math>\|F\|_{p,q}.</math>

This seminorm is a norm if <math>q</math> is a norm.

Generalizations

The concept of in composition algebras does share the usual properties of a norm.

A composition algebra <math>(A, *, N)</math> consists of an algebra over a field <math>A,</math> an involution <math>\,*,</math> and a quadratic form <math>N,</math> which is called the "norm". In several cases <math>N</math> is an isotropic quadratic form so that <math>A</math> has at least one null vector, contrary to the separation of points required for the usual norm discussed in this article.

An or a is a seminorm <math>p : X \to \R</math> that also satisfies <math>p(x + y) \leq \max \{p(x), p(y)\} \text{ for all } x, y \in X.</math>

Weakening subadditivity: Quasi-seminorms

A map <math>p : X \to \R</math> is called a if it is (absolutely) homogeneous and there exists some <math>b \leq 1</math> such that <math>p(x + y) \leq b p(p(x) + p(y)) \text{ for all } x, y \in X.</math>

The smallest value of <math>b</math> for which this holds is called the

A quasi-seminorm that separates points is called a on <math>X.</math>

Weakening homogeneity - <math>k</math>-seminorms

A map <math>p : X \to \R</math> is called a if it is subadditive and there exists a <math>k</math> such that <math>0 < k \leq 1</math> and for all <math>x \in X</math> and scalars <math>s,</math><math display="block">p(s x) = |s|^k p(x)</math> A <math>k</math>-seminorm that separates points is called a on <math>X.</math>

We have the following relationship between quasi-seminorms and <math>k</math>-seminorms:

See also

Notes

Proofs

References

  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!---->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • <!-- -->
  • Sublinear functions
  • The sandwich theorem for sublinear and super linear functionals