If fX is Continuous Seminorm Then F is Continuous
Seminorm
A seminorm defined on a Banach space X is a nonnegative real-valued functional |x| defined on H that satisfies the properties |αx| = |α| |x| and |x + y| ⩽ |x| + |y|.
From: Pure and Applied Mathematics , 1977
Numerical Analysis of Wavelet Methods
Albert Cohen , in Studies in Mathematics and Its Applications, 2003
Remark 3.2.1
The Besov semi-norm is often instead defined in its integral form
(3.2.14)
i.e. if q < +∞. The equivalence between the above expression and its discrete analog of (3.2.13) is easily derived from the monotonicity of the modulus of smoothness (ωn (f, t) p ≤ ωn (f, t') p if t < t'). Also note that with constants independent of J ≥ 0 such that diam(Ω) ≤ 2−J .
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S0168202403800061
Topological Vector Spaces
Henri Bourlès , in Fundamentals of Advanced Mathematics 2, 2018
3.3.2 Semi-norms
(I) A semi-norm on E is a mapping that satisfies the following conditions for all x, y ∈ E and all :
-
(N 1 ) p (0) = 0.
-
(N 2 ) p (λx) = | λ | p (x).
-
(N 3 ) p (x + y) ≥ p (x) + p (y) ("triangle inequality").
A semi-norm p is said to be a norm if it also satisfies the condition
-
(N 4 ) p (x) = 0 ⇒ x = 0.
Any semi-norm p is a convex mapping (exercise) and determines a pseudometric d :(x, y) ↦ p (x − y). If p is a norm, then d is a metric. Let p be a semi-norm. For all α > 0, write
The sets B p (α) and B p c (α) are both disked and absorbing, and the following conditions are equivalent (exercise): (i) p is continuous on E; (ii) p is continuous at 0; (iii) p is uniformly continuous; (iii) for all α > 0, B p (α) is open in E (so is an open neighborhood of 0); (iv) for all α > 0, B p c (α) is a neighborhood of 0. Given a disked and absorbing set A ⊂ E, define p A (x) = inf {ρ : ρ > 0, x ∈ ρA} for all x ∈ E. Then p A is a semi-norm on E (exercise), called the gauge (or Minkowski functional) of A; furthermore, B p A (1) ⊂ A ⊂ B p A c (1). If A is open (resp. closed) in E, then p A is continuous (resp. lower semi-continuous, cf. Definition 2.31) and A = B p A (1) (resp. A = B p A c (1)) (exercise*: cf. [BKI 81], Chapter II, section 2.10, Proposition 23). Conversely, suppose that p A is lower semi-continuous; then and ([P1], section 1.1.2 (VI)); but p A − 1(]1, ∞[) is open, so p A − 1([0, 1]) is closed.
(II) A topological vector space is said to be semi-normed if its topology is determined by a semi-norm p, that is, if {B p (α): α > 0} is a fundamental system of neighborhoods of 0 for (any such space is Hausdorff if and only if p is a norm).
Let E be a vector space on and suppose that p 1, p 2 are two semi-norms on E. These semi-norms are said to be equivalent if they determine the same topology on E. The proof of claim (1) of the next theorem is an exercise. Claim (2) follows from Lemma 3.6:
Theorem 3.17
- 1)
-
The two semi-norms p 1, p 2 on E are equivalent if and only if there exist real numbers α > 0, β > 0 such that αp 1 ≤ p 2 ≤ βp 1.
- 2)
-
On , all norms are equivalent.
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/B9781785482496500031
Sobolev Spaces
In Pure and Applied Mathematics, 2003
6.29 Domains of Finite Width
Consider the problem of determining for what domains Ω in R n is the seminorm
actually a norm on W 0 m,p (Ω) equivalent to the standard norm
This problem is closely related to the problem of determining for which unbounded domains Ω the imbedding W 0 m,p (Ω) → Lp (Ω) is compact because both problems depend on estimates for the Lp norm of a function in terms of Lp estimates for its derivatives.
We can easily show the equivalence of the above seminorm and norm for a domain of finite width, that is, a domain in Rn that lies between two parallel planes of dimension (n − 1). In particular, this is true for any bounded domain.
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S0079816903800089
Sobolev Spaces
In Pure and Applied Mathematics, 2003
4.29(Seminorms)
For 1 ≤ p < ∞ and for integers j, 0 ≤ j ≤ m, we introduce functionals |·| j·p on W m·p (Ω) as follows:
Clearly |u|0·p = |u|0,p = |u| p is the norm on Lp (Ω) and
If j ≥ 1, we call |·| j, p a seminorm . It has all the properties of a norm except that |u| j, p = 0 need not imply u = 0 in Wm,p (Ω). For example, u may be a nonzero constant function if Ω has finite volume. Under certain circumstances which we begin to investigate in Paragraph 6.29, |·| m·p is a norm on (Ω) equivalent to the usual norm |·| m,p . In particular, this is so if Ω is bounded.
For now we will confine our attention to these seminorms as they apply to functions in (ℝ n ).
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S0079816903800065
Topics in Multivariate Approximation and Interpolation
Angela Kunoth , in Studies in Computational Mathematics, 2006
5 Multilevel Regularization
Next we extend the least squares fitting strategy to include into the objective functional a Sobolev seminorm with positive smoothness index α as a regularizing term in the least squares functional (1) to enforce a smooth approximation. This classical approach for α ∈ {1,2} was studied, e.g., in [23] for general splines, [55] for splines with free knots, [34] for hierarchical splines, [37] for splines on triangulations, or [32] in a wavelet reformulation of the spline problem. Here we choose more generally so that we consider
(14)
where the not necessary integral α > 0 parameterizes the smoothing and v > 0 balances the fidelity of the data and the smoothness. In view of the statements made after (5), the wavelet formulation allows to represent functions in Sobolev norms in terms of sequence norms of weighted coefficients from their wavelet expansions. Accordingly, following [11], we replace by the corresponding sequence norm on the right hand side of (14),
(15)
The minimization of (14) using this sequence norm is now equivalent to the solution of the normal equations
(16)
Here R is a diagonal matrix with entries
(17)
if J is the highest level of approximation and the sizes of the identity matrices depend on the cardinality of Λ j. This representation of the smoothing term in the wavelet context has some interesting properties. First, the different dyadic scales are decoupled which means that the weak decoupling between levels in M is maintained, and the effect of the regularization term boils down to a penalization of the higher frequencies. Thus, the parameter v controls the balance between fidelity and regularization while α controls the relative penalization across scales. Second, since (5) holds for the whole scale of fractional Sobolev spaces {H α (Ω)} α ∈ [0,γ), one has easy access to the entire scale which reduces to a simple diagonal scaling in the wavelet framework by means of the diagonal matrix R. Third, the formulation is independent of the wavelet family. Of course, the smoothness of the wavelets imposes a limit on the maximal value of α for which (15) holds.
We have now two parameters in (14), the smoothness index α and the weight parameter v to balance the data approximation and the smoothness part which can be used to adjust as best as possible to the given data. An example of an application from photogrammetry with fixed α = 4 and v = 0.01 is provided in Figs. 4 and 5. It should be mentioned that for the linear pre-wavelets we employ the norm equivalence (15) does not hold for this high value of α. Nevertheless, the reconstruction in Fig. 5 gives the best result for this case.
Fig. 4. Left: Original data, a 3D point set consisting of 330.000 points of an industrial site (taken by Leica Cyrax 2500, Prof. Staiger, GH Essen). Middle: Vertical view of a section of the original data. Right: Sampling geometry of this section.
Fig. 5. Approximation of the data from Fig. 4 by wavelet reconstruction with regularization for v = 0.01, α = 4 and thresholding parameter ε = 10− 3. Left: Reconstruction for J = 4. Middle: Wavelet coefficients of type (1, 1). Right: Reconstruction for J = 6. The final approximation contains #Λ6 = 2623 coefficients as opposed to 16384 coefficients for the full grid.
Standard cross validation techniques start out with computing the weight parameter v for a fixed choice of α, which requires to solve an additional system of normal equations. Below we once more exploit the multilevel framework provided by wavelets and introduce a multilevel version of the cross validation procedure. This scheme turns out to be both relatively inexpensive from a computational point of view as well as adjusting well to smoothness and to localization effects.
Let us first describe the classical situation where the smoothness parameter α is fixed. Different ways to select an appropriate regularization parameter v in (14) were proposed, where one of the most popular choices are cross validation methods, see, e.g., [60]. Particularly, in the Generalized Cross Validation (GCV) method one computes v as the minimizer of the GCV potential with respect to the Euclidean norm || • ||, that is,
(18)
where the influence matrix H(v) is defined as H(v) := A(M + vR)− 1 AT. The influence matrix is never computed explicitly, since only the application of H(v) to a vector needs to be carried out. This in turn requires the solution of a linear system with matrix M + vR which is of a similar type as in (16). The trace in the denominator of (18) is usually estimated stochastically as tr(B) = u TB u where u is a random vector with entries 1 and − 1 and B is a quadratic symmetric matrix, see [41]. Consequently, the evaluation of GCV (v) for each v requires the solution of two linear systems of the type (16). The total computation time then depends on the number of GCV evaluations which are needed to find the minimum of GCV (v). In [63] an automatized computation of a GCV involving up to 15 evaluations of the functional is reported, together with the CPU time which is needed in the solution of a system of normal equations with penalization for a given v. We have displayed in [7] a series of bivariate examples where we have applied these GCV techniques and where the solution of each linear system costed less than one second.
The GCV technique yields an optimal v in the sense that it minimizes the variance under the assumption that the data has been corrupted by white noise, see, e.g., [59]. This theory developed for splines immediately carries over to the wavelet case[5]. Moreover, as shown in [7] the underlying heuristics can also be applied successfully to other data fitting approaches which require some regularization such as for data sets with a highly varying point distribution or for data with holes in the domain where typically overfitting artifacts occur.
The main issue in [7] is the following multilevel version of the GCV algorithm used within the context of the adaptive coarse-to-fine Algorithm. The idea is to simultaneously use the GCV with the hierarchical growth of the wavelet tree as the levels become higher. We have observed above that the wavelet representation penalizes the different dyadic levels separately with the same weight 22αj . In view of the structure of the diagonal matrix R given in (17), we prescribe instead a diagonal penalizing matrix with the same values v j ≥ 0 for all entries of each level j, where the vj are computed independently following a GCV criterion. Now at each level j the normal equations attain the form
(19)
where the diagonal matrix Rj is defined component–wise as for some set of scalars . These scalars are computed inductively at each level following a GCV criterion. At level j 0, we define the influence matrix
and is obtained by minimization of the corresponding GCV potential (18). At any subsequent level j the influence matrix is defined as
(20)
to have vj as only variable since the from previous levels j 0 ≤ j′ < j are already computed. vj is then determined by minimizing (18). Recall that the influence matrices are never computed explicitly as they only need to be applied to a vector. The new penalizing term can no more be interpreted as stemming from a smoothing term since no relation between the penalizing parameters is prescribed. Nevertheless, this approach offers some interesting advantages. It is easily built into the coarse–to–fine growth of the tree. One can attain higher flexibility for the smoothing effect. Moreover, overfitting artifacts are typically localized in scale which enables us to disentangle the different scales. Last, the method is much cheaper as the GCV method with both α and v as free parameters from a computational point of view.
As illustration, we apply the multilevel GCV algorithm to the data set in Fig. 6 obtained from [49] which corresponds to a bathymetrical study of part of the sea floor of the Dominican Republic. As can be seen from the left image, measurements of the sea floor depth are irregularly distributed, forming lines, clusters and holes. The central plot shows a visualization of the depth of the full set of data points using piecewise linear interpolation. For the classical GCV, there is no obvious way to predict which α may give a good reconstruction. In contrast, our multilevel GCV algorithm gives a sufficiently good reconstruction in the right plot.
Fig. 6. Left: Sampling geometry with N = 10732 points. Middle: Piecewise linear interpolation. Right: Reconstruction with multilevel GCV.
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S1570579X06800066
FN-Topologies and Group-Valued Measures
Hans Weber , in Handbook of Measure Theory, 2002
PROOF
(a) Since it is enough to prove that (μ n ) is uniformly exhaustive with respect to any continuous seminorm on G. we may assume that G = (G, ) is a seminormed group. Then the semivariation of μ n is bounded by Corollary 2.7. Therefore there are numbers ∈ n > 0 such that η:= is a bounded function. It follows that η is an exhaustive submeasure and μ n is η-continuous for every n. Suppose that (μn)n is not uniformly exhaustive. Then there is an ∈ > 0, a disjoint sequence (an ) n ∈ in R and a subsequence (vn ) n ∈ of (μ n ) n ∈ such that for every n ∈ . By Lemma 10.2, (an )n ∈ has a subsequence (bn )n ∈ , such that the restriction of η on S := σ({bn : n ∈ }) is σ-order continuous. Applying Theorem 10.1 to one obtains that ( )n ∈ is -equicontinuous. Since η(bn ) → 0, it follows that sup m → 0(n → +∞), a contradiction to ≥ ∈.
(b) follows from (a) and Corollary 7.2.
(c) Let u be the supremum of the μ n -topologies, n ∈ . Then u is σ-order continuous if μ n ,n∈N are σ-additive (see Proposition 4.2). By (b), μ is μ-continuous, hence σ-additive.
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/B9780444502636500178
Functional Analysis and its Applications
Wieslaw Żelazko , in North-Holland Mathematics Studies, 2004
Theorem 2
Let A be a complex (resp. real) topological division algebra with continuous inverse and possessing a continuous p-homogeneous seminorm , 0 < p ≤ 1. Then A is topologically isomorphic to C (resp. to either of R, C, H).
The requirement that the operation x → x −1 is necessary for validity of the above result, but in the case of an F-algebra it is satisfied automatically. It follows from the following result proved by Arens in [3] and by Banach in [10] in the separable case, and in [29] in the general case.
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S0304020804801804
Real Reductive Groups I
In Pure and Applied Mathematics, 1988
Theorem
Let (π, H) be a finitely generated, admissible, Hilbert representation of G. Set V = HK and Λ = Λ V . There is a positive constant d such that if μ ∈(H ∞ )′K then there exists a continuous semi-norm, σμ, on H∞ with the property that
for v ∈ H∞ and a ∈ Cl(A +).
Let μ ∈ (H ∞ )′K . 4.3.3 implies that μ = σ(w) with w ∈ HK . Lemmas 2.A.2.2 and 2.A.2.3 imply that there exists δ ∈ a* and C > 0 such that if x, y ∈ H then
This clearly implies that if μ ∈ (H ∞ )′K then there exists, σ′ μ , a continuous semi-norm on H ∞ such that
- (1)
-
The idea of the proof is to show that if δ(Hj ) > Λ(Hj ) then we can replace δ in (1) by δ – mαj with m = min{1/2, μ(Hj ) – Λ(Hj )} at the cost of possibly changing the semi-norm σ′ μ and putting in a term (1 + log ‖a‖) d .
Let α ∈ Δ0. Set F = Δ0 – {α}. If α = αj then set H = Hj . Then aF = R H. Set at = exp(tH). If a ∈ Cl(A +) then a can be written uniquely in the form a = a′at with a = exp(Σ xkHk ), xk ≥ 0, xj = 0 and t ≥ 0.
Let q be the canonical projection of V ˜ onto V ˜/n F V ˜.
- (2)
-
If q(μ) = 0 then there exists a continuous semi-norm t′ μ on H ∞ such that |μ(π(a)v)| ≤ a δ–α γ′ μ (v), for a ∈ Cl(A +) and v ∈ H ∞.
Let X 1,…, Xp be a basis of n F consisting of root vectors for a corresponding to the roots β 1,…, βp respectively. Our assumption implies that μ = Σ Xkμk with μk ∈ V ˜. Hence
(2) now follows from 3.8.6(1).
Let for z ∈ C, (V ˜/n F )V ˜) z denote the generalized eigenspace for H with eigenvalue z. Let Pz be the projection of (V ˜/n F V ˜) onto (V ˜/n F V ˜) z corresponding to the H-weight space decomposition. Let μ ∈ V ˜. Then q(μ) = Σ Pzq(μ). Let μz ∈ V ˜ be such that q(μz ) = Pzq(μ). Then μ – Σ μz ∈ n F V ˜. We now estimate μz (π(a)v) for each z. Set μz = v. Let be a basis for U(a F )q(v). We assume that . Let vk ∈ V ˜ be such that for k ≥ 2. Now
and B = [bkn ] has the property that
- (3)
-
We also note that
- (4)
-
Let a′ ∈ Cl(A +) be such that (a′) α = 1. We set
and
Then
- (5)
-
This implies that
- (6)
-
We now estimate the terms in (6). (1) implies that
- (7)
-
‖F(0, a′; v)‖ ≤ (a′) δ β(v) with β a continuous semi-norm on H∞ . (2) implies that
- (8)
-
- (9)
-
This follows immediately from (3). These estimates imply that if t ≥ 0 then
for some continuous semi-norm β on H ∞ and some positive constant C. We observe that (1 + s) p e –ɛs is bounded by a constant C for ɛ > 0 and s ≥ 0. We therefore have
- (10)
-
Here C is a positive constant and β is a continuous semi-norm on H∞ .
There are now two cases.
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S0079816908628238
Real Reductive Groups I
In Pure and Applied Mathematics, 1988
5.1.2 Proposition
Let (π, H) be a Hilbert representation of G. If HK is tempered then (π, H) satisfies the weak inequality. If HK is rapidly decreasing then (π, H) satisfies the strong inequality.
Let V = HK . Theorem 4.3.5 implies that there exists d > 0 such that if w ∈ V then there exists a continuous semi-norm, σ w , depending on w such that |〈π(a)v, w〉| ≤ σ w (v)(1 + log ‖a‖) d a Λ (Λ = Λ V ) for all a ∈ Cl(A +). Let w 1, …, wp be a basis of the span of Kw. Let σ(v) = sup k ∈ K Σ σ w (kv). π(k)w = Σgj (k)Wj with each gj a continuous function on K. 〈π(k 1 ak 2)v, w〉 = 〈π(ak 2)v, π(k 1)–1 w〉 = Σ conj(gj ((k 1)–1))〈π(ak 2)v, w〉. It follows that
- (1)
-
If w ∈ V and v ∈ H∞ then |〈π(k 1 ak 2)v, w〉| ≤ σ(v)(1 + log ‖a‖) d aΛ
for a ∈ Cl(A +) and k 1, k 2 ∈ K.
We now prove the result. Suppose that V is tempered. Then αΛ ≤ a −ρ for all a ∈ Cl(A +). Now, a –ρ ≤ Ξ(a) for all a ∈ Cl(A +) (Theorem 4.5.3). Since Ξ(k 1 gk 2) = Ξ(g) for all k 1, k 2∈ K, the first assertion now follows immediately from (1).
If μ ∈ + a* then for each r > 0 there exists a positive constant Cr such that, a –μ ≤ Cr (1 + log ‖a‖)–r for a ∈ Cl(A +). Hence, the second assertion is also a direct consequence of (1).
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S007981690862824X
Harmonic Analysis Tools for Solving the Incompressible Navier–Stokes Equations
Marco Cannone , in Handbook of Mathematical Fluid Dynamics, 2005
2.5.3. Le Jan-Sznitman spaces
Recently, Le Jan and Sznitman [137,138] considered the space of tempered distributions f whose Fourier transform verifies
(116)
Now, if in the previous expression we consider instead of , we obtain the (semi)-norm of a homogeneous Sobolev space. This is not the case: the functions whose Fourier transform is bounded define the pseudo-measure space of Kahane. In other words, a function f belongs to the space introduced by Le Jan and Sznitman if and only if , Δ being the Laplacian (in three dimensions). A simple calculation (see [48]) shows that condition (116) is written, in the dyadic decomposition Δ j of Littlewood and Paley in the form and defines in this way "the homogeneous Besov space" .
Let us note that this quantity is not a norm, unless we work in S′ modulo polynomials, as we did in Section 2.2 in the case of homogeneous Besov spaces (for example, if f is a constant or, more generally a polynomial of degree 1, it is easy to see that ). Another possibility to avoid this technical point is to ask that . In other words, the Banach functional space relevant to our study is defined by
(117)
A generalization of this functional space was recently introduced in the paper by Bhattacharya, Chen, Dobson, Guenther, Orum, Ossiander, Thomann and Waymire (see [8]).
Read full chapter
URL:
https://www.sciencedirect.com/science/article/pii/S1874579205800060
quezadasumpeormses92.blogspot.com
Source: https://www.sciencedirect.com/topics/mathematics/seminorm
0 Response to "If fX is Continuous Seminorm Then F is Continuous"
Enregistrer un commentaire