Optimal Representation of Piecewise Hölder Smooth Bivariate Functions by the Easy Path Wavelet Transform

Size: px
Start display at page:

Download "Optimal Representation of Piecewise Hölder Smooth Bivariate Functions by the Easy Path Wavelet Transform"

Transcription

1 Optmal Representaton of Pecewse Hölder Smooth Bvarate Functons by the Easy Path Wavelet Transform Gerlnd Plonka 1, Stefane Tenorth 1, and Armn Iske 2 1 Insttute for Numercal and Appled Mathematcs, Unversty of Göttngen, Lotzestr , Göttngen, Germany {plonka,s.tenorth}@math.un-goettngen.de 2 Department of Mathematcs, Unversty of Hamburg, Hamburg, Germany ske@math.un-hamburg.de Abstract The Easy Path Wavelet Transform (EPWT) [25] has recently been proposed by one of the authors as a tool for sparse representatons of bvarate functons from dscrete data, n partcular from mage data. The EPWT s a locally adaptve wavelet transform. It works along pathways through the array of functon values and t explots the local correlatons of the gven data n a smple approprate manner. In ths paper, we am to provde a theoretcal understandng of the performance of the EPWT. In partcular, we derve condtons for the path vectors of the EPWT that need to be met n order to acheve optmal N-term approxmatons for pecewse Hölder smooth functons wth sngulartes along curves. Key words. sparse data representaton, wavelet transform along pathways, N-term approxmaton AMS Subject classfcatons. 41A25, 42C40, 68U10, 94A08 1 Introducton Durng the last few years, there has been an ncreasng nterest n effcent representatons of large hgh-dmensonal data, especally for sgnals. In the one-dmensonal case, wavelets are partcularly effcent to represent pecewse smooth sgnals wth pont sngulartes. In two dmensons, however, tensor product wavelet bases are no longer optmal for the representaton of pecewse smooth functons wth dscontnutes along curves. Wthn the last few years, more sophstcated methods were developed to desgn approxmaton schemes for effcent representatons of two-dmensonal data, n partcular for mages, where correlatons along curves are essentally taken nto account to capture the geometry of the gven data. Curvelets [2, 3], shearlets [12, 13] and drectonlets [34] are examples for non-adaptve hghly redundant functon frames wth strong ansotropc drectonal selectvty. For pecewse Hölder smooth functons of second order wth dscontnutes along C 2 -curves, Candès and Donoho [2] proved that a best approxmaton f N to a gven functon 1

2 f wth N curvelets satsfes the asymptotc bound f f N 2 2 C N 2 (log 2 N) 3, whereas a (tensor product) wavelet expanson leads to asymptotcally only O(N 1 ) [21]. Up to the (log 2 N) 3 factor, ths curvelet approxmaton result s optmal (see [7, Secton 7.4]). A smlar estmate has been acheved by Guo and Labate [12] for shearlet frames. These results, however, are not adaptve wth respect to the assumed regularty of the target functon, and so they cannot be appled to mages of less regularty,.e., mages whch are not at least pecewse C 2 wth dscontnutes along C 2 -curves. Just recently, the N-term approxmaton propertes of compactly supported shearlet frames have been studed also n the 3D case, [19]. Usng two smoothness parameters α and β n (1, 2] controllng the classcal as well as the ansotropc smoothness, nearly optmal approxmaton rates are proven n [19]. In [11], a general framework called parabolc molecules has been presented coverng most of the curvelet- and shearlet-lke constructons and showng that these systems all provde the same approxmaton results. In the bvarate case, for pecewse Hölder smooth functons of order α 2, one may rather adapt the approxmaton scheme to the mage geometry nstead of fxng a bass or a frame beforehand to approxmate f. Durng the last few years, several dfferent approaches were developed for dong so [1, 5, 6, 8, 9, 14, 18, 20, 22, 25, 26, 27, 29, 30, 33]. In [20], for nstance, bandelet orthogonal bases and frames are ntroduced to adapt to the geometrc regularty of the mage. Due to ther constructon, the utlzed bandelets are ansotropc wavelets that are warped along a geometrcal flow to generate orthonormal bases n dfferent bands. LePennec and Mallat [20] showed that ther bandelet dctonary yelds asymptotcally optmal N-term approxmatons, even n more general mage models, where the edges may also be blurred. Further examples for geometry-based mage representatons are the nonlnear edgeadapted (EA) multscale decompostons n [1, 14] (and references theren) based on ENO reconstructons. We remark that the resultng ENO-EA schemes lead to an optmal N- term approxmaton, yeldng f f N 2 2 C N 2 for pecewse C 2 -functons wth dscontnutes along C 2 -curves. Moreover, unlke prevous non-adaptve schemes, the ENO-EA multresoluton technques provde optmal approxmaton results also for BV -spaces and L p spaces, see [1]. In [25], a new locally adaptve dscrete wavelet transform for sparse mage representatons, termed Easy Path Wavelet Transform (EPWT), has been proposed by one of the authors. The EPWT works along pathways through the array of functon values, where t essentally explots the local correlatons of mage values n an approprate manner. We remark that the EPWT s not restrcted to a regular (two-dmensonal) grd of mage pxels, but t can be extended, n a more general settng, to scattered data approxmaton n hgher dmensons. In [26], the EPWT has been appled to data representatons on the sphere. In the mplementaton of the EPWT, one needs to work wth sutable data structures to effcently store the path vectors that need to be accessed durng the performance of the EPWT reconstructon. To reduce the resultng adaptvty costs, we have proposed a hybrd method for smooth mage approxmatons n [27], where an effcent edge representaton by the EPWT s combned wth favorable propertes of the borthogonal tensor product wavelet transform. 2

3 In [28], we have proven optmal N-term approxmaton rates for the EPWT for pecewse Hölder contnuous functons of order α (0, 1]. Our proof n [28] s manly based on an adaptve multresoluton analyss structure, whch s only avalable usng pecewse constant Haar wavelets along the pathways. But the EPWT does not necessarly need to be restrcted to Haar wavelets. In fact, the numercal results n [25, 31] show a much hgher effcency for smoother wavelet bases as e.g. Daubeches D4 flters and borthogonal 7-9 flters. These observatons motvate us to study the N-term approxmaton for Hölder smooth functons of order α > 1 n ths paper. In partcular, we wll derve condtons for the path vectors leadng to optmal N-term approxmaton by the EPWT. More precsely, we present three condtons for the optmal path vectors that mply optmal N-term approxmatons of the form f f N 2 2 C N α (1.1) for the applcaton of the EPWT to pecewse Hölder smooth functons of order α > 1, wth allowng dscontnutes along smooth curves of fnte length. Unfortunately, the three condtons on the path vectors are very dffcult to ensure. We regard these condtons as an dealzed settng whch heurstcally explans the performance of the EPWT wth hgher order wavelets, and whch gve us a hnt on how to construct sutable path vectors. The path constructons of the relaxed EPWT n [25] amed at a good compromse to produce many small wavelet coeffcents, on the one hand, and low costs for path codng, on the other hand. Interestngly enough, our results n ths paper show that the path constructons n [25] yeld already a farly good tradeoff to meet the optmal path vector condtons. In our prevous paper [15], we have appled the EPWT n mage denosng, wth path vectors that approxmatvely satsfy the optmal path vector condtons. Wth usng pecewse constant functons for the approxmaton of a bvarate functon f, the EPWT yelds an adaptve multresoluton analyss when relyng on an adaptve Haar wavelet bass (see [25, 28]). If, however, smoother wavelet bases are utlzed n the EPWT approach, such an nterpretaton s not obvous. In fact, whle Haar wavelets admt a straght forward transfer from one-dmensonal functons along pathways to bvarate Haar-lke functons, we cannot rely on such smple connectons between smooth one-dmensonal wavelets (used by the EPWT) and a bvarate approxmaton of the lowpass functon. Therefore, n ths paper we wll apply a sutable nterpolaton method, by usng polyharmonc splne kernels, to represent the arsng bvarate low-pass functons after each level of the EPWT. One key property of polyharmonc splne nterpolaton s polynomal reproducton of arbtrary order. We wll come back to relevant approxmaton propertes of polyharmonc splnes n Secton 2. The outlne of ths paper s as follows. In Secton 2, we frst ntroduce the utlzed functon model, some ssues on polyharmonc splne nterpolaton, and the EPWT algorthm. Then, n Secton 3, we study the decay of EPWT-wavelet coeffcents, where we wll consder the hghest level of the EPWT n detal. To acheve optmal decay results for the EPWT wavelet coeffcents at all levels, we requre three specfc sde condtons for the path vectors n the EPWT algorthm, the regon condton, the path smoothness condton and the dameter condton, see Secton 3.2. We also show, why the path smoothness condton s very restrctve. In fact, t cannot be met for the usual EPWT as descrbed n [25] for α > 1. Therefore, n ths paper we ntroduce sutable smooth path functons n 3

4 order to derve the path vectors for the EPWT. Fnally, Secton 4 s devoted to the proof of asymptotcally optmal N-term error estmates of the form (1.1) for pecewse Hölder smooth functons. 2 The EPWT and Polyharmonc Splne Interpolaton 2.1 The Functon Model Suppose that Ω [0, 1] 2 approxmates [0, 1] 2 and has suffcently smooth Lpschtz boundary. Further, let F L 2 (Ω) be a pecewse smooth bvarate functon, beng smooth over a fnte set of regons {Ω } 1 K, where each regon Ω has a suffcently smooth Lpschtz boundary Ω of fnte length,.e., the boundary curves have bounded dervatves. Moreover, the set {Ω } 1 K s assumed to be a dsjont partton of Ω, so that K Ω = Ω, =1 where each closure Ω s a connected subset of Ω, for = 1,..., K. More precsely, we assume that F s Hölder smooth of order α > 0 n each regon Ω, 1 K, so that any µ-th dervatve of F on Ω wth µ = α satsfes an estmate of the form F (µ) (x) F (µ) (y) C x y α µ 2 for all x, y Ω. Note that ths assumpton for F s equvalent to the condton that for each x 0 Ω there exsts a bvarate polynomal q α of degree α (usually the Taylor polynomal of F of degree α at x 0 Ω ) satsfyng F (x) q α (x x 0 ) C x x 0 α 2 (2.1) for every x Ω n a neghborhood of x 0, where the constant C > 0 does not depend on x or x 0. But F may be dscontnuous across the boundares between adjacent regons. Note that the Hölder space C α (Ω ) of order α > 0, beng equpped wth the norm F C α (Ω ) := F C α (Ω ) + sup x y µ = α F (µ) (x) F (µ) (y) x y α α 2 concdes wth the Besov space B, (Ω α ), when α s not an nteger. Here, we use the C α (Ω ) norm F C α (Ω ) := sup F (x) + sup µ F (x), x Ω x Ω µ = α see e.g. [4, Chapter 3.2]. Now by quas-unform samplng, the bvarate functon F s assumed to be gven by ts functon values taken at a fnte grd I 2J. For a sutable nteger J > 1, let {F (y)} y I2J be the gven samples of F, where I 2J Ω wth #I 2J = 2 2J, and mn y 2 y 2 2 κ 1 2 J, y 1,y 2 I 2J sup nf x y 2 2 (J+1/2) y I 2J x Ω 4

5 wth a postve constant κ 1 beng ndependent of J. Moreover, let Γ 2J := {y I 2J : y Ω } for 1 K (2.2) be the set of grd ponts that are contaned n the regons Ω, for 1 K. Obvously, K =1 Γ 2J = I 2J, and for the sze #Γ 2J of Γ 2J we have #Γ 2J #I 2J = 2 2J for any 1 K. 2.2 Polyharmonc Splne Interpolaton Next we compute a (pecewse) suffcently smooth approxmaton to F from ts gven samples. To ths end, we construct a sutable nterpolant F 2J satsfyng the nterpolaton condtons F 2J (y) = F (y) for all y I 2J. (2.3) Ths s accomplshed by the applcaton of polyharmonc splne nterpolaton separately n each ndvdual regon Ω, n order to frst obtan, for any ndex, 1 K, a polyharmonc splne nterpolant F 2J satsfyng F 2J (y) = F (y) for all y Γ 2J (2.4) at the nterpolaton ponts Γ 2J Ω. For the requred (global) nterpolant F 2J, we then let K F 2J (x) := F 2J (x)χ Ω (x) for x Ω (2.5) =1 to satsfy the nterpolaton condtons n (2.3), where χ Ω s the characterstc functon of Ω. Recall that polyharmonc splnes, due to Duchon [10], are sutable tools for multvarate nterpolaton from scattered data. For further detals on polyharmonc splnes, especally for relevant aspects concernng ther local approxmaton propertes, we refer to [17, Secton 3.8]. We apply the polyharmonc splne nterpolaton scheme n two dmensons, where the nterpolant F 2J n (2.4) s assumed to have the form F 2J (x) = y Γ 2J c n φ m ( x y 2 ) + p m(x), (2.6) where φ m (r) = r 2m log(r), for m := α, s a fxed polyharmonc splne kernel, and p m P m s a polynomal n the lnear space P m of all bvarate polynomals of degree at most m. Note that, by the form of φ m, the nterpolant F 2J (x) s, for any, n C m. Moreover, by the above choce of m, F 2J (x) s Hölder smooth of order α, and n ths way, we can mantan the local Hölder regularty of F around each lattce pont n Γ 2J. Note that the nterpolant F 2J n (2.6) has #Γ 2J + dm(p m ) parameters,.e., the #Γ 2J coeffcents c n n ts major part and another dm(p m ) = (m + 1)(m + 2)/2 parameters n ts polynomal part. However, the nterpolaton problem (2.4) yelds only #Γ 2J lnear 5

6 condtons on the #Γ 2J + (m + 1)(m + 2)/2 parameters of F 2J. To elmnate addtonal degrees of freedom, we requre another set of (m + 1)(m + 2)/2 lnear constrants on the coeffcents c n, as gven by the vanshng moment condtons c n p(y) = 0 for all p P m. (2.7) y Γ 2J To compute the coeffcents of the polyharmonc splne nterpolant F 2J, ths then amounts to solvng a square lnear system wth #Γ 2J + dm(p m ) equatons, (2.4) and (2.7), for #Γ 2J + dm(p m ) unknowns, gven by the parameters of F 2J. Accordng to the semnal work of Mchell [23] on (condtonally) postve defnte functons, ths square lnear equaton system has a unque soluton, provded that the set of nterpolaton ponts are P m -regular,.e., for p P m we have the mplcaton p(y) = 0 for all p P m = p 0, (2.8) so that every polynomal n P m can unquely be reconstructed from ts values on Γ 2J. In fact, ths comples wth earler results n [10], where the well-posedness of the polyharmonc splne nterpolaton scheme s proven va a varatonal theory concernng (polyharmonc) splnes mnmzng rotaton-nvarant sem-norms n Sobolev spaces. Therefore, we can conclude that the polyharmonc splne nterpolant F 2J n (2.6), satsfyng (2.4) and (2.7), s unque provded that the nterpolaton ponts Γ 2J satsfy the (rather weak) regularty condtons n (2.8). From now we wll tactly assume that the condtons n (2.8) are fulflled. In ths case, we can further conclude that the polyharmonc splne nterpolaton scheme acheves to reconstruct polynomals of degree m. Concernng the local approxmaton order of polyharmonc splne nterpolaton, we refer to [16]. As was proven n [16], the asymptotc estmate F 2J (hy) F (hy) = O(h m+1 ) for h 0 holds for F C m+1, n whch case the polyharmonc splne nterpolaton scheme s sad to have local approxmaton order m + 1, see [16] or [17, Secton 3.8] for detals. We remark that the concept n [16] concernng local approxmaton orders s n contrast to that of global approxmaton orders, where the latter s well-understood for polyharmonc splnes and related radal kernels snce much longer. In the subsequent analyss n ths paper, we requre one specfc approxmaton result for polyharmonc splne nterpolaton concernng ts global approxmaton behavour. We can descrbe ths as follows. From the embeddng theorem for Besov spaces we obtan B α, (Ω ) B α 2,2 (Ω ), see e.g. [32] or [4, page 163]. Snce B α 2,2 (Ω ) s equvalent to the Sobolev space H α (Ω ), see [4, 32], ths allows us to use the estmate F F 2J L 2 (Ω) C F K =1 h α Ω F B α 2,2 (Ω ) (2.9) for the nterpolaton error n Sobolev spaces, as shown n [24], where the fll dstance h Ω := sup x Ω nf y Γ 2J x y 2 2 J+1/2 measures the densty of the nterpolaton ponts n Ω. 6 for 1 K

7 2.3 The EPWT Algorthm Now let us brefly recall the EPWT algorthm from our prevous work [25]. To ths end, let ϕ C β, for β α, be a suffcently smooth, compactly supported, one-dmensonal scalng functon,.e., the nteger translates of ϕ form a Resz bass of the scalng space V 0 := clos L 2span {ϕ( k) : k Z}. Further, let ϕ be a correspondng borthogonal and suffcently smooth scalng functon wth compact support, and let ψ and ψ be a correspondng par of compactly supported wavelet functons. We refer to [4, Chapter 2] for a comprehensve survey on borthogonal scalng functons and wavelet bases and summarze only the notaton needed for the borthogonal wavelet transform. For j, k Z, let ϕ j,k (t) := 2 j/2 ϕ(2 j t k) and ψ j,k (t) := 2 j/2 ψ(2 j t k), lkewse for ϕ and ψ. The functons ϕ, ϕ and ψ, ψ are assumed to satsfy the refnement equatons ϕ(x) = 2 n ϕ(x) = 2 n h n ϕ(2x n) h n ϕ(2x n) ψ(x) = 2 q n ϕ(2x n) n ψ(x) = 2 q n ϕ(2x n) n wth fnte sequences of flter coeffcents (h n ) n Z, ( h n ) n Z and (q n ) n Z, ( q n ) n Z. assumpton, the polynomal reproducton property p m, ϕ j,k ϕ j,k = p m for all j Z, k By s satsfed for any polynomal p m of degree less than or equal to m = α, and so, p m, ψ j,k = 0 for all j, k Z. Wth these assumptons, {ψ j,k : j, k Z} and { ψ j,k : j, k Z} form borthogonal Resz bases of L 2 (R),.e., for each functon f L 2 (R), we have f = f, ψ j,k ψ j,k = f, ψ j,k ψ j,k. j,k Z For any gven unvarate functon f j, j Z, of the form f j (x) = n Z cj (n) ϕ j,n one decomposton step of the dscrete (borthogonal) wavelet transform can be represented n the form f j (x) = f j 1 (x) + g j 1 (x), where f j 1 (x) = n Z j,k Z c j 1 (n) ϕ j 1,n and g j 1 (x) = n Z d j 1 (n) ψ j 1,n wth c j 1 (n) = f j, ϕ j 1,n and d j 1 (n) = f j, ψ j 1,n. (2.10) 7

8 Conversely, one step of the nverse dscrete wavelet transform yelds for gven functons f j 1 and g j 1 the reconstructon f j (x) = n Z c j (n) ϕ j,n wth c j (n) = f j 1, ϕ j,n + g j 1, ϕ j,n. We recall that the EPWT s a wavelet transform that works along path vectors through pont subsets of I 2J. For the characterzaton of sutable path vectors we frst need to ntroduce neghborhoods of ponts n I 2J. For any pont y = (y 1, y 2 ) I 2J, we defne ts neghborhood by N(y) := {x = (x 1, x 2 ) I 2J \ {y} : x y 2 2 J+1/2 }, where x y 2 2 = (x 1 y 1 ) 2 + (x 2 y 2 ) 2. Now the EPWT algorthm s performed as follows. For the applcaton of the 2J-th level of the EPWT we need to fnd a path vector p 2J = (p 2J (n)) 22J 1 n=0 through the pont set I 2J. Ths path vector s a sutable permutaton of all ponts n I 2J, whch can be determned by usng the followng strategy from [25]. Recall that I 2J = K =1 Γ2J, where Γ 2J determnes the lattce ponts n Ω. Start wth one pont p 2J (0) n Γ 2J 1. Now, for a gven n-th component p 2J (n) beng contaned n the set Γ 2J n (2.2), for some {1,..., K}, we choose the next component p 2J (n + 1) of the path vector p 2J, such that p 2J (n + 1) (N(p 2J (n)) Γ 2J ) \ {p 2J (0),..., p 2J (n)}, (2.11).e., p 2J (n + 1) should be a neghbor pont of p 2J (n), lyng n the same set Γ 2J, that has not been used n the path, yet. In stuatons where (N(p 2J (n)) Γ 2J ) \ {p 2J (0),... p 2J (n)} s an empty set, the path s nterrupted, and we need to start a new pathway by choosng the next component p 2J (n+1) from Γ 2J \ {p 2J (0),..., p 2J (n)}. If, however, ths set s also empty, we choose p 2J (n + 1) from the set of remanng ponts I 2J \{p 2J (0),..., p 2J (n)}. For a more detaled descrpton of the path vector constructon we refer to [25]. In partcular, for a sutably chosen path vector p 2J, the number of nterruptons can be bounded by K = C 1 K, where K s the number of regons, and where the constant C 1 does not depend on J but only on the shape of the regons Ω, see [28]. The so obtaned vector p 2J s composed of connected pathways,.e., each par of consecutve components n these pathways s neghborng. Now, we consder the data vector ( c 2J (l) ) 2 2J 1 := ( F 2J ( p 2J (l) )) 2 2J 1 l=0 l=0 and apply one level of a one-dmensonal (perodc) wavelet transform to the functon values of F 2J along the path p 2J. Ths yelds the low-pass vector (c 2J 1 (l)) 22J 1 1 l=0 and the vector of wavelet coeffcents (d 2J 1 (l)) 22J 1 1 l=0 accordng to the formulae n (2.10). Due to the pecewse smoothness of F 2J along the path vector p 2J, t follows that most of the wavelet coeffcents n d 2J 1 are small, whereas the wavelet coeffcents correspondng to an nterrupton of the path (wthn one regon or from one regon to another) may possess sgnfcant ampltudes. We remark already here that, by contrast to the orgnal 8

9 EPWT algorthm descrbed n ths subsecton, for the case of pecewse Hölder smooth mages of order α > 1, we wll ntroduce a smooth path functon p 2J n Secton 3.1. Ths path functon wll determne the components of the path vector p 2J and s assumed to have bounded dervatves almost everywhere. Note that for α > 1, wavelet coeffcents of (arbtrary) large magntude may occur n stuatons where the dervatves of the path functon are not unformly bounded, as e.g. at tght turns of p 2J. As regards the next level of the EPWT, the path vector p 2J yelds a new subset of ponts where Γ 2J 1 Γ 2J 1 := { p 2J (2l) : l = 0,..., 2 2J 1 1 } = K =1 Γ 2J 1, := {p 2J (2l) : p 2J (2l) Γ 2J }. At level j = 2J 1, we frst locate a second connected path vector p 2J 1 = (p 2J 1 (l)) 22J 1 1 l=0 through Γ 2J 1,.e., the entres of p 2J 1 form a permutaton of the ponts n Γ 2J 1. Smlarly as before, we requre that p 2J 1 (n) and p 2J 1 (n + 1) are neghbors lyng n the same pont set Γ 2J 1. Here, p 2J 1 (n) and p 2J 1 (n + 1) are sad to be neghbors,.e., p 2J 1 (n + 1) N(p 2J 1 (n)), ff p 2J 1 (n) p 2J 1 (n + 1) 2 2 J+1. Agan, the number of path nterruptons can be bounded by C 1 K, where C 1 does not depend on J. Then we apply one level of the one-dmensonal wavelet transform to the permuted data vector (c 2J 1 (per 2J 1 (l))) 22J 1 1 l=0, where the permutaton per 2J 1 s defned by per 2J 1 (l) := k ff p 2J 1 (l) = p 2J (2k) for l, k = 0, J 1 1. We obtan the low-pass vector (c 2J 2 (l)) 22J 2 1 l=0 and the vector (d 2J 2 (l)) 22J 2 1 l=0 of wavelet coeffcents. We contnue by teraton over the remanng levels j for j = 2J 2,..., 1, where at any level j we frst construct a path p j = (p j (l)) 2j 1 l=0 through the set Γ j := {p j+1 (2l) : l = 0,..., 2 j 1} = wth applyng smlar strateges as descrbed above. Here, p j (n) and p j (n + 1) are called neghbors, ff p j (n) p j (n + 1) 2 D2 j/2, (2.12) where D 2 s a sutably determned constant (n the above descrpton of p 2J and p 2J 1 we have chosen D = 2). Then we apply the wavelet transform to the permuted vector (c j (per j (l)) 2j 1 l=0, yeldng cj 1 and d j 1. Example. In ths example, we explan the constructon of the path vectors through the remanng data ponts wth the low-pass values by a toy example. To ths end, let Ω = [0, 1) 2 be dvded nto only two regons, Ω 1 and Ω 2. The functon F s assumed to be Hölder contnuous n each of these regons, but may be dscontnuous across the curve separatng the two regons Ω 1 and Ω 2. In our toy example, we have J = 3,.e., an 8 8 mage wth 64 data values. At the hghest level of the EPWT, we choose a path p 6 through the underlyng pont set I 6 = {(2 3 n 1, 2 3 n 2 ) : n 1, n 2 = 0,..., 7} = Γ 6 1 Γ6 2, such that each two consecutve components n the path are neghbors. We frst pck all ponts n Γ 6 1, 9 K =0 Γ j

10 (a) (b) (c) (d) Fgure 1: Path constructon. (a) level sx, (b) level fve, (c) level four, (d) level three. before jumpng to Γ 6 2, see Fgure 1(a). For the path constructon at the next level, we frst determne the set Γ 5 = Γ 5 1 Γ5 2 (contanng only each second component of p6 ), see Fgure 1(b), before we construct a path accordng to the above descrpton. Fgures 1(c) and 1(d) show the sets Γ 4 and Γ 3 along wth ther correspondng path vectors. Besdes, we tred to fnd the path vectors n a way such that the angles formed by the polygonal lne of the path are as large as possble. In fact, the polygonal lnes for p 6 and p 5 do not contan angles beng smaller than 90 degrees, whle the polygonal lne of p 4 possesses two such angles, one n each regon. In fact, locally straght polygonal lnes of the path vectors gve rse to obtan better smoothness bounds for the path functons p j that we wll ntroduce n Secton 3. In ths example, we have p 6 (n + 1) p 6 (n) 2 5 D, p 5 (n + 1) p 5 (n) 2 5 2D, p 4 (n + 1) p 4 (n) 2 4 2D, p 3 (n + 1) p 3 (n) D (wth one path nterrupton at each level for the jump from one regon to the other), so that the path constructon satsfes the above requrements wth D = 5. Ths smple example also llustrates that the path constructon leads at each level to pont sets Γ j wth quas-unformly dstrbuted ponts. See Secton 3.2 for more explanaton about the used quas-unformty. Remark 1. Consderng the above strategy of the EPWT algorthm, t s heurstcally clear that we are able to reduce the number of sgnfcant wavelet coeffcents to a multple of the number K of regons, where the target functon F s smooth. Indeed, only when the path skps from one regon to another, a fnte number of sgnfcant wavelet coeffcents wll occur. Ths s n contrast to the usual tensor product wavelet transform, where the number of sgnfcant wavelet coeffcents s usually related to the total length of the smooth regons boundares and hence depends on the level j of the wavelet transform. 10

11 Remark 2. As shown n [25], the choce of the path vectors n each level s crucal for the performance of the approxmaton by means of the EPWT. In partcular, t s not a good dea to take smply the same path vector p j (l) = p j+1 (2l) at each level, snce one only explots the data correlaton along ths one path and the EPWT s reduced to a onedmensonal transform. In Secton 3.2, we wll determne the so called dameter condton that requres the nequalty (2.12) for neghborng components n the path constructon thereby avodng the above mentoned smple path choce. Our numercal experments wth the EPWT mply that the constant D should be chosen rather small, e.g. D 2, see [31]. On the other hand, a too small D nduces a small number of neghborhood ponts and hence yelds a hgher number of path nterruptons. In Secton 3.2, we wll show that the nequalty (2.12) for path constructon mples a quas-unform dstrbuton of values n Γ j at each level of the EPWT. Remark 3. Note that the mplementaton of the EPWT algorthm, as descrbed above, s rather straghtforward. In the orgnal verson of the EPWT, we frst determne, at each level, a path vector p j, before we apply a wavelet flter bank to the one-dmensonal data vector that s ordered accordng to the path vector. In contrast to that strategy (from our prevous work), we now consder usng a slghtly dfferent approach to obtan the theoretcal estmates of our present paper. To ths end, we derve the path vector p j by frst determnng a pecewse smooth path functon p j, whose constructon s detaled n the followng of ths paper. We also consder slghtly dfferent data vectors c j p(l) n the theoretcal estmates of Sectons 3 and 4, where we wll use L 2 -projecton operators determned by the dual scalng and wavelet functons, ϕ and ψ. Further, we wll employ polyharmonc splne nterpolaton for analytcal purpose. Remark 4. Note that the components of the path vector p j le n I 2J wth contanng 2d entres. Ths s n contrast to the notaton n [25]. Further, unlke n [25], we do no longer consder ndex sets but defne a neghborhood of ponts by the Eucldean dstance. 3 Decay of Wavelet Coeffcents generated by the EPWT Before we turn to the techncal detals, let us frst sketch the basc deas of the proof for optmal N-term approxmatons by the EPWT, where we use slghtly changed notatons. As explaned n the Subsecton 2.2, we consder applyng polyharmonc splne nterpolaton, from gven mage values F (y), separately n the ndvdual domans Ω, = 1,..., K. We assume that F s Hölder smooth of order α on each Ω, so that the polyharmonc splne nterpolant F 2J n (2.5), beng Hölder smooth of order at least α, reconstructs bvarate polynomals of degree m = α. Further, we apply a generalzed noton of a path p gven as a suffcently smooth planar parameter curve. At the 2J-th level of the EPWT, we frst determne a (pecewse) smooth path functon p 2J : [0, 1] Ω, such that each value y I 2J can be approxmated at suffcent accuracy by p 2J ( l ), for some l {0,..., 2 2J 1}. Consecutve values p 2J ( l ) and p 2J ( l+1 ) 2 2J 2 2J 2 2J should approxmate neghborng values n Γ 2J := I 2J lyng n the same set Ω (up to O(K) exceptons). Now, a path vector p 2J R 22J 2 through all values of I 2J s determned by the permutaton gven by the order n whch p 2J (l/2 2J ), l = 0,..., 2 2J 1, approxmates the values n I 2J. 11

12 The path functon p 2J should be constructed such that only at most C 1 K dscontnutes (ncurred by possble transtons from one regon to another) may occur. In ths way, we can bound the number of nterruptons n the path vector p 2J by C 1 K. Next, we consder a one-dmensonal functon f 2J (t) = 22J 1 k=0 c 2J p (k) ϕ 2J,k (t), t [0, 1), whch approxmates the pecewse smooth one-dmensonal and scaled restrcton of F resp. F 2J along the path p 2J satsfyng F ( p 2J (2 2J l) f 2J (2 2J l) 2 Jα. Then, we apply one level of a smooth wavelet transform to f 2J. In ths way, sgnfcant wavelet coeffcents may only occur at a fnte number of locatons on the nterval [0, 1) that correspond to dscontnutes or to tght turns of the path functon p 2J. However, the number of such sgnfcant coeffcents does not depend on 2 J but on the number of regons, K. Therefore, wth the performance of one level of the (perodc) wavelet transform, we wll fnd that most of the wavelet coeffcents of f 2J = f 2J 1 + g 2J 1 occurrng n the wavelet part g 2J 1, are small. At the next j-th level of the EPWT, we consder the set Γ j := { p j+1 (2 j l) : l = 0,..., 2 j 1} = K =1 Γ j. By constructon, we have # Γ j = 2 j. Observe that Γ j are pont sets n Ω. Agan, we then construct a smooth path functon p j that approxmates the values n Γ j n a sutable order and thus determnes the path vector p j whose components are the values n Γ j n the order n whch they are approxmated by p j. To obtan suffcently accurate polyharmonc splne nterpolatons F j to F wth F j (y) = F (y) for y Γ j, and to apply the same arguments as at the hghest level, three specfc condtons on the path functons p j need to be satsfed. We can brefly explan these condtons, termed regon condton, path smoothness condton and dameter condton, as follows (for more detals on these condtons we refer to Subsecton 3.2). Frstly, the regon condton requres that the path functon should prefer to approxmate all values of one regon set Γ j, before proceedng wth the values of the next regon. In other words, the path vector should prefer to traverse the ponts belongng to one regon set Γ j, before jumpng to another regon, where jumps wthn one regon should be avoded. Assumng further that the path functon s smooth wth sutably bounded dervatves (path smoothness condton), we can optmally explot the smoothness of the functon F along the path. Fnally, the dameter condton ensures a quas-unform dstrbuton of remanng ponts n each Γ j, and therefore leads to a suffcently accurate polyharmonc splne nterpolaton F j at each level of the EPWT. We remark that the regon condton and the dameter condton can be forced by usng the strateges for the path constructon as proposed n Subsecton 2.2. The path smoothness condton cannot be satsfed by the orgnal path constructon methods n [25]. But snce the numercal evaluatons contan only a few levels of the EPWT, the path smoothness condton can be forced by avodng small angles n the path vector and by preferrng path snakes, see e.g. examples n [25]. The three path condtons wll allow us to estmate the EPWT wavelet coeffcents smlarly as for one-dmensonal pecewse smooth functons wth a fnte number of sngulartes to fnally obtan an optmal N-term 12

13 approxmaton usng only the N most sgnfcant EPWT wavelet coeffcents for the mage reconstructon. 3.1 The Hghest Level of the EPWT Let us now explan the 2J-th level of the EPWT n detal. The performance of the subsequent levels of the EPWT and the correspondng estmates are then derved n a smlar manner. We consder a suffcently smooth parameter curve p 2J (t), t [0, 1], through Ω approxmatng the values n Γ 2J n a certan order and the correspondng path vector p 2J of length 2 2J wth values n I 2J, where we assume that p 2J (2 2J l) p 2J (l) µ2 J l = 0,..., 2 2J 1, (3.1) and where µ s a fxed small constant (ndependent of J) wth assumng p 2J (t) Ω, for t [l/2 2J, (l + 1)/2 2J ], f p 2J (2 2J l) and p 2J (2 2J (l + 1)) are n Ω. Now, we regard the functon f 2J that s defned by the one-dmensonal restrcton of F 2J along the curve p 2J, f 2J (t) = F 2J ( p 2J (t) ) for t [0, 1]. If the path functon p 2J crosses over from one regon Ω to another, there may occur a dscontnuty of f 2J caused by a dscontnuty of F 2J,.e., there are ndces l and l + 1 n {0, 1,..., 2 2J 1}, where we have a dscontnuty between F 2J ( p 2J (2 2J l)) and F 2J ( p 2J (2 2J (l + 1))),.e., a dscontnuty of f 2J n the subnterval [l/2 2J, (l + 1)/2 2J ]. Wthout loss of generalty, we assume that we have only dscontnutes of f 2J caused by jumps from one regon to another. Recall that the trace theorem for Hölder resp. Besov spaces (see [32]) mples that for F 2J Ω B α, (Ω ), the (scaled) restrcton f 2J (t) along the curve p 2J (t) s agan Hölder smooth of order α n each subnterval of [0, 1], determned by T = {t [0, 1] : p 2J (t) Ω } and wth assumng that the correspondng path vector p 2J has no nterruptons n Ω. In partcular, for any α (0, 1), we fnd the estmate f 2J C α (T ) = sup t 1 t 2 t 1,t 2 T = sup t 1 t 2 t 1,t 2 T F 2J ( p 2J (t 1 )) F 2J ( p 2J (t 2 )) t 1 t 2 α F 2J ( p 2J (t 1 )) F 2J ( p 2J (t 2 )) p 2J (t 1 ) p 2J (t 2 ) α p 2J (t 1 ) p 2J (t 2 ) α 2 2 t 1 t 2 α F 2J C α (Ω ) p 2J α C 1 (T ). For the case α (1, 2), the chan rule yelds f 2J C α (T ) = sup t 1 t 2 t 1,t 2 T sup t 1 t 2 t 1,t 2 T F 2J ( p 2J (t 1 ))( p 2J ) (t 1 ) F 2J ( p 2J (t 2 ))( p 2J ) (t 2 ) t 1 t 2 α 1 ( F 2J ( p 2J (t 1 )) F 2J ( p 2J (t 2 )) 2 p 2J (t 1 ) p 2J (t 2 ) α 1 2 ) + F 2J C 1 (Ω ) ( p2j ) (t 1 ) ( p 2J ) (t 2 ) t 1 t 2 α 1 F 2J C α (Ω ) p 2J α C 1 (T ) + F 2J C 1 (Ω ) p 2J C α (T ). 13 p 2J (t 1 ) p 2J (t 2 ) α 1 2 t 1 t 2 α 1 ( p 2J ) C 1 (T )

14 In general, we have Lemma 3.1 Suppose that the parameter curve p 2J : [0, 1] Ω satsfes n each subnterval T the smoothness condton p 2J C β (T ) C 2 2 Jβ (3.2) for all β max{α, 1} and wth a constant C 2 beng ndependent of J,.e., p j C α (T ) for α > 1 and p j C 1 (T ) for α 1, = 1,..., K, wth strctly bounded dervatves. Then for the one-dmensonal restrcton f 2J (t) = F 2J ( p 2J (t)) along the curve p 2J we have f 2J C α (T ) C2 Jα F 2J C α (Ω ). (3.3) Proof. For α < 2 the asserton already follows from the prevous observatons. Let m := α. Further, we use the abbrevatons F = F 2J, p = p 2J. In the general case the formula by Faà d Bruno mples that D m f 2J (t) := dm f dt 2J (t) = D m (F ( p(t))) s a fnte m lnear combnaton of terms S ν,j (t) := D ν F (x) x= p(t) (D m1 p 1 (t)) b 1 (D m2 p 2 (t)) b 2, where p = p 2J = ( p 1, p 2 ) T, m 1, m 2, b 1, b 2 N wth m 1 + m 2 m, m 1 b 1 + m 2 b 2 = m, and wth ν N 2 0, ν m. For t 1, t 2 T wth t 1 t 2 we now obtan wth (3.2) the estmate S ν,j (t 1 ) S ν,j (t 2 ) t 1 t 2 ( α m ) ( ) (D ν F )( p(t 1 )) (D ν F )( p(t 2 )) p(t1 ) p(t 2 ) α m p(t 1 ) p(t 2 ) α m 2 t 2 1 t 2 ( α m ) + (D ν F )( p(t 2 )) (D m 1 p1 (t 1 )) b 1 (D m 2 p 2 (t 1 )) b 2 (D m 1 p 1 (t 2 )) b 1 (D m 2 p 2 (t 2 )) b 2 t 1 t 2 α m F C α+ ν m (Ω ) p α m C 1 (T ) p b 1 C m 1 (T ) p b 2 C m 2 (T ) + F C ν (Ω ) (D m1 p 1 (t 1 )) b 1 (D m2 p 2 (t 1 )) b 2 ( b 1 p b 2 p b 1 1 C m 1 (T ) p C m 1 +α m (T ) + b 2 p b 1 C m 1 (T ) p b 2 1 C m 2 (T ) p C m 2 +α m (T ) F C α+ ν m (Ω ) 2J(α m) 2 J(m 1b 1 ) 2 J(m 2b 2 ) C m 2 (T ) + F C ν (Ω ) (b 12 J(m 2b 2 +m 1 b 1 m 1 +m 1 +α m) + b 2 2 J(m 1b 1 +m 2 b 2 m 2 +m 2 +α m) ) C ν 2 Jα F C α (Ω ), where we have used the nequalty x k y k k x y max{x, y} k 1 for x, y R, k N. Hence, the asserton follows. Let us assume n the sequel that the path functon p satsfes the smoothness condton (3.2) n Lemma 3.1. Thus, we obtan for the N-th order modulus of smoothness of f 2J the estmate ω N ( f 2J, h) := sup Ñ f 2J h h α f 2J C α (T ) (2 J h) α F 2J C α (Ω ) (3.4) h h wthn the subntervals, where f 2J s smooth,.e., for N = α + 1 and T,h := { t : p 2J (t + kh) Ω, k = 0,..., N }, ) 14

15 see [4]. Next, we consder the L 2 -projecton f 2J := P 2J f 2J of f 2J onto the scalng space V 2J := clos L 2 [0,1)span{ϕ 2J,n : n = 0,..., 2 2J 1}, where ϕ s assumed to be a suffcently smooth scalng functon, see Secton 2.2. Then, f 2J = P 2J f 2J := 2 2J 1 n=0 f 2J, ϕ 2J,n ϕ 2J,n also satsfes a Hölder smoothness condton of order α. Followng along the lnes of [4, Theorem 3.3.3], we obtan n the subntervals T,2 2J the estmate In partcular, f 2J f 2J L (T,2 2J ) = f 2J P 2J f 2J L (T,2 2J ) ω N ( f 2J, 2 2J ) (2 2J ) α f 2J C α (T ) (2 J ) α F 2J C α (Ω ). (3.5) f 2J (2 2J l) f 2J (2 2J l) = F 2J ( p 2J (2 2J l)) f 2J (2 2J l) 2 Jα. In the next step, we decompose the functon f 2J = c 2J p (l)ϕ 2J,l wth c 2J p (l) := f 2J, ϕ 2J,l l nto the low-pass part f 2J 1 and the hgh-pass part g 2J 1. Applyng one level of the onedmensonal wavelet transform to the data set (c 2J p (l)) 22J 1 l=0, we obtan the decomposton f 2J = f 2J 1 + g 2J 1 wth f 2J 1 = 2 2J 1 1 n=0 2 2J 1 1 c 2J 1 p (n) ϕ 2J 1,n and g 2J 1 = n=0 d 2J 1 p (n) ψ 2J 1,n, where c 2J 1 p (n) := f 2J, ϕ 2J 1,n and d 2J 1 p (n) := f 2J, ψ 2J 1,n. From the Hölder smoothness of f 2J, we fnd for t T the representaton f 2J (t) = q α (t t 0 ) + R(t t 0 ) for t 0 {2 2J k : k = 0,..., 2 2J 1} T and t t 0 2 2J, where q α denotes the Taylor polynomal of degree α of f 2J at t 0, and where the remander R satsfes the bound R(t t 0 ) c ϕ (t)2 Jα. Hence, f supp( ψ 2J 1,n ) T for some, the wavelet coeffcents satsfy d 2J 1 p (n) = q α ( t 0 ) + R( t 0 ), ψ 2J 1,n = R( t 0 ), ψ 2J 1,n c ϕ,n 2 Jα ψ 2J 1,n 1 = c ϕ,n 2 ( J+1/2)(α+1), where we have used ψ 2J 1,n 1 = 2 J+1/2 ψ 1. Observe that the constant c ϕ,n n ths nequalty depends on the choce of the wavelet bass but also on the (local) smoothness propertes of f 2J, and hence on the (local) boundedness propertes of the dervatves of the chosen path functon p 2J n Lemma 3.1. Now let Λ 2J 1 be the set of all n {0,..., 2 2J 1 1}, where the above estmate for d 2J 1 (n) s satsfed. Then, the number of the remanng 2 2J 1 #Λ 2J 1 wavelet coeffcents corresponds to the number of postons, where f 2J s dscontnuous (e.g. caused by crossng over of the path functon to another regon) or where dervatves of f 2J are not sutably bounded,.e., the constant c ϕ,n s too large (caused e.g. by tght turns of the path 15

16 functon p 2J ). We assume that ths number s bounded by CK, where K s the number of regons n the orgnal mage F, and where the constant C does not depend on 2 J. Now, we consder the low-pass functon f 2J 1 and reconstruct a bvarate functon F 2J 1 as follows. Takng only the path functon values p 2J (2 2J+1 n), we put Γ 2J 1 := { p 2J (2 2J+1 n) : n = 0,..., 2 2J 1 1, p 2J (2 2J+1 n) Ω } for each = 1,..., K and Γ 2J 1 := K =1 Γ 2J 1. We compute the polyharmonc splne nterpolant K F 2J 1 (x) := c y φ m ( x y 2 ) + p m(x) χ Ω (x), =1 y Γ 2J 1 satsfyng the nterpolaton condtons F 2J 1 ( p 2J (2 2J+1 n) ) = f 2J 1 (2 2J+1 n) for all n = 0,..., 2 2J 1 1. Therefore, F 2J ( p 2J (2 2J+1 n)) F 2J 1 ( p 2J (2 2J+1 n)) = f 2J (2 2J+1 n) f 2J 1 (2 2J+1 n) = f 2J (2 2J+1 n) P 2J 1 f 2J (2 2J+1 n) 2 ( J+1)α = D α (2 J+1/2 ) α, where D = 2, and where the last nequalty agan follows analogously as n (3.5) snce f 2J 1 s the orthogonal projecton of f 2J to V 2J 1 := clos L 2 [0,1)span{ϕ 2J 1,n : n = 0, J 1 1}. The last nequalty mples that F 2J 1 s stll a good approxmaton to F, snce the nterpolaton ponts have changed only slghtly. However, only half of the nterpolaton ponts are left, whch are rregularly dstrbuted n Ω. Moreover, by (3.1) we have max x Ω mn y Γ 2J 1 x y 2 J+1 + µ2 J = D 2 J+1/2. (3.6) Assumng mn y1,y 2 Γ 2J 1 y 1 y 2 D 1 2 J, wth a sutable constant D 1 > 0 ndependent of J, we observe wth (2.9) F 2J F 2J 1 L 2 (Ω ) (2 J+1/2 ) α for all = 1,..., K. Remark. Compared wth the orgnal EPWT algorthm n [25], the condton (3.1) s an mportant relaxaton. Here the scalng and wavelet coeffcents are computed from the functon f 2J (t) that approxmates F 2J ( p 2J (t)), but snce p 2J (2 2J l), l = 0,..., 2 2J 1, do not nterpolate but only approxmate the orgnal grd ponts I 2J, we work wth new functon values F 2J ( p 2J (2 2J l)) of the polyharmonc splne functon F 2J nstead of the gven values F (y), y I 2J. The orgnal EPWT relates to the specal case µ = 0, where we have nterpolaton n (3.1). 16

17 3.2 Condtons for the Path Vectors Before we proceed wth the error estmates for the further levels of the EPWT algorthm, we want to fx specfc sde condtons for the path functons that are requred for our error analyss, and that have been mplctly used already n the estmates at the hghest level n Secton 3.1. The sde condtons are termed (a) regon condton, (b) path smoothness condton, and (b) dameter condton, as stated below. The regon condton and the dameter condton have been smlarly stated already n [28] to prove the N-term approxmaton estmates. The regon condton ensures that at each level of the EPWT, the path functon jumps only C 1 K tmes from one regon Ω to another regon or nsde a regon. The dameter condton ensures that the remanng ponts n Γ j are quas-unformly dstrbuted, such that there s a constant D, not dependng on J or j, satsfyng max mn x y 2 (1 + 2)D2 j/2. (3.7) x Ω y Γ j Further, at each level j of the EPWT, we assume that we can fnd a smooth functon p j : [0, 1] Ω wth unformly bounded dervatves that approxmates the values n Γ j. Ths leads us to the vector p j, whose components from Γ j are sutably ordered, and we have p j (2 j l) p j (l) µ2 j for l = 0,..., 2 j 1. Further, we assume that p j (t) Ω, for t [l/2 j, (l + 1)/2 j ], f p j (2 2J l) and p j (2 2J (l + 1)) are n Ω. Let us ntroduce the three condtons more explctly. (a) Regon condton. At each level j of the EPWT, the path functon p j s chosen, such that t contans only at most C 1 K dscontnutes caused by crossng over from one regon Ω to another regon or by jumpng wthn one regon Ω. (b) Path smoothness condton. The path functon p j satsfes n each subnterval T the smoothness condton p j C β (T ) C 2 2 jβ/2 for all β max{α, 1} and wth a constant C 2 beng ndependent of j,.e., p j C α (T ) for α > 1 and p j C 1 (T ) for α 1, = 1,..., K, wth bounded dervatves. (c) Dameter condton. At each level of the EPWT, we requre for almost all values 2 j l, l = 0,..., 2 j 2, the condton p j (2 j l) p j (2 j (l + 1)) 2 D 2 j/2, (3.8) where D s ndependent of J and j, and where the number of values of p j (2 j l) whch do not satsfy the dameter condton, s bounded by a constant C 3 not dependng on j. Hence, at each level j, consecutve components p j (2 j l) n the path functon should be spatal neghbors. The condtons (a) and (c) can be enforced by the proposed path constructon n Subsecton 2.2, see (2.11) for the regon condton and (2.12) for the dameter condton. Partcularly, the dameter condton ensures a quas-unform dstrbuton of the ponts n Γ j. Ths can be seen nductvely as follows. For j = 2J, the asserton (3.7) s obvous, 17

18 Fgure 2: Smooth path functon n a regon Ω wth smooth boundary. see also (3.6). Generally, assumng (3.7) and (3.8) at level j + 1, t follows that max mn x y 2 max mn x y 2 + max mn y z 2 x Ω y Γ j x Ω y Γ j+1 y Γ j+1 z Γ j (1 + 2)D2 (j+1)/2 + D2 (j+1)/2 = (1 + 2)D 2 j/2. Unfortunately, the gven path smoothness condton s rather restrctve. In Fgure 2, we sketch an example for a choce of a path functon p 2J for a convex doman Ω = Ω 1 wth a smooth boundary wth Hölder smoothness α (1, 2). The constructon of p 2J reles on smooth spral curves that are composed of straght lnes and smooth Bezer curves (or arcs) wth bounded frst dervatve. Observe that n the example of Fgure 2, the curvature of the path functon p 2J does not depend on 2 J but on the curvature of the boundary of Ω 1, whle all ponts n Γ 2J 1 can be well approxmated (e.g. by ther projectons onto the curve p 2J ). Moreover, assumng that the straght lne segments as shown n Fgure 2 are separated at dstance 2 J, the complete path functon p 2J has a length bounded by κ2 J. For a next path functon p 2J 1 one may now thnk about a smooth spral curve composed of straght lne segments n a horzontal drecton (nstead of the vertcal drecton n Fgure 2), and the dstance between two neghborng straght lne segments s now 2 J+1/2, etc. Remarks. 1. To understand the strength of the path smoothness condton (b), we follow the arguments of one of the revewers and show that the path vector p j can only be taken straght along a lne f µ < 1/4. In partcular, f the path functon p 2J s forced to nterpolate the grd ponts of the form (2 J n 1, 2 J n 2 ), n 1, n 2 Z (regular grd), as n the stuaton of the EPWT n [25], the path smoothness condton cannot be satsfed. Indeed, wth (3.1) we fnd the estmate p 2J (l + 1) 2p 2J (l) + p 2J (l 1) p 2J (2 2J (l + 1) 2 p 2J (2 2J (l) + p 2J (2 2J (l 1) + 2 J 4µ. 18

19 But the smoothness condton (3.2) mples that p(2 2J (l + 1) 2 p(2 2J (l) + p(2 2J (l 1) = ( p(2 2J (l + 1) p(2 2J (l)) ( p(2 2J (l) + p(2 2J (l 1)) = (( p 2J ) (t) ( p 2J ) (t 2 2J ) dt I 0 2 2J 2 2J(α 1) p 2J C α (I) = 2 2Jα p 2J C α (I) C 2 2 Jα, where I 0 := [2 2J l, 2 2J (l + 1)] I, and where I denotes an nterval where p 2J s C α smooth. Hence, p 2J (l + 1) 2p 2J (l) + p 2J (l 1) 2 J (C 2 2 J(1 α) + 4µ). For α > 1 the term C2 J(1 α) tends to zero, as J. On the other hand, f the ponts p 2J (l + 1), p 2J (l), p 2J (l 1) (beng samples from the regular grd (2 J n 1, 2 J n 2 ), n 1, n 2 Z) do not le on the same lne, then the norm on the left hand sde s at least 2 J. Hence, for µ < 1/4 the path condtons are ncompatble. 2. The condtons (a) and (c) can slghtly be relaxed the sense that the constants C 1 and C 3 may be allowed to depend polynomally on j (but not exponentally). In fact, consderng the proof of Theorem 4.1 (resp. Theorem 3.3 n [28]), one needs to ensure that a weghted sum of all wavelet coeffcents that are only satsfyng the estmate (3.11) s stll bounded. 3.3 The Further Levels of the EPWT Let us now explan the further levels of the EPWT. These are performed followng along the lnes of the 2J-th level. We start wth the polyharmonc splne nterpolant F j (x) := K c y φ m ( x y 2 ) + p m(x) χ Ω (x) =1 y Γ j that satsfes the nterpolaton condtons F j ( p j+1 (2 j n)) = f j (2 j n) for all n = 0,..., 2 j 1. We fx a sutable path functon p j approxmatng the set Γ j = { p j+1 (2 j n) : n = 0,..., 2 j 1} wth correspondng data values {F j ( p j+1 (2 j n) ) : n = 0,..., 2 j 1}, such that the three path condtons n Secton 3.2 are satsfed. Then we determne the onedmensonal restrcton of F j along p j, f j (t) := F j ( p j (t) ), t [0, 1]. The pecewse Hölder smoothness of f j (t) ensures that ω N ( f j, h) (2 J h) α for T,h := {t : p j (t + kh) Ω, k = 0,..., N}. Consderng the L 2 -projecton P j f j := 2 j 1 n=0 cj p(n) ϕ j,n wth c j p(n) := f j, ϕ j,n of f j onto the scalng space V j := clos L 2 [0,1)span{ϕ j,n : n = 0,..., 2 j 1}, where ϕ s assumed to be a smooth scalng functon as n Subsecton 3.1, we have f j P j f j L (T,2 j ) w N ( f j, 2 J j/2 ) 2 jα/2, 19

20 where we have appled the dameter condton (3.8). We use the decomposton P j f j = l c j p(l) ϕ j,l = f j 1 + g j 1 = n c j 1 p (n) ϕ j 1,n + n d j 1 p (n) ψ j 1,n. Then the Hölder smoothness of P j f j n the ntervals T yelds the Taylor expanson P j f j (t) = q α (t t 0 ) + R(t t 0 ) wth R(t t 0 ) c ϕ (t) D α 2 jα/2 for t, t 0 T, where D s the constant n the dameter condton (3.8), and ths gves the estmate for the wavelet coeffcents correspondng to the regon Ω, d j 1 p (n) = R(t t 0 ), ψ j 1,n c ϕ,n D α 2 jα/2 2 (j 1)/2 ψ 1 c ϕ,n D α 2 (j 1)(α+1)/2, (3.9) where c ϕ,n depends on the local smoothness of P j f j and hence on the dervatve bounds of p j n (3.2). Agan, let Λ j 1 be the set of ndces n from {0,..., 2 j 1 1}, where d j 1 p (n) satsfes the above estmate. Then the number of wavelet coeffcents 2 j 1 #Λ j 1 whch are not satsfyng ths estmate (snce supp ψ j 1,n T for some ) s bounded by a constant ndependent of J and j. Fnally, we obtan the polyharmonc splne nterpolant F j 1 (x) := K c y φ m ( x y 2 ) + p m(x) χ Ω (x), j 1 Γ =1 y Γ j 1 where := { p j (2 j+1 n) : n = 0,..., 2 j 1 1, p j (2 j+1 n) Ω }, Γ j 1 := K =1 Γ j 1, through the nterpolaton condtons F j 1 ( p j (2 j+1 n)) = f j 1 (2 j+1 n) for all n = 0,..., 2 j 1 1. Hence, we obtan the estmate F j ( p j (2 j+1 n) ) F j 1 ( p j (2 j+1 n) ) = f j (2 j+1 n) P j 1 f j (2 j+1 n) 2 jα/2. Let us summarze the above fndngs on the decay of the EPWT wavelet coeffcents n the followng theorem. Theorem 3.2 For j = 2J 1,..., 0, let d j p(l) = f j+1, ψ j,l, l = 0,..., 2 j 1, denote the wavelet coeffcents that are obtaned by applyng the EPWT algorthm to F (accordng to the above defntons n Secton 3), where we assume that F L 2 (Ω) s pecewse Hölder smooth of order α as prescrbed n Subsecton 2.1. Further assume that the path functons p j+1, j = 2J 1,..., 0, n the EPWT algorthm satsfy the regon condton (a), the path smoothness condton (b), and the dameter condton (c) of Subsecton 3.2. Then, for all j = 2J 1,..., 0 and l Λ j, the estmate d j p(l) C D α 2 j(α+1)/2 (3.10) holds, where D > 1 s the constant of the dameter condton (3.8), α s the Hölder exponent of F, and C depends on the utlzed wavelet bass and on the Hölder constant n (2.1). Furthermore, for all l {0,..., 2 j 1} \ Λ j, we obtan the estmate wth some constant C beng ndependent of J and j. d j p(l) C 2 j/2 (3.11) 20

21 Proof. The proof of (3.10) follows drectly from (3.9). Lkewse, for all l {0,..., 2 j 1} \ Λ j,.e., for pont sets that do not satsfy the dameter, the path smoothness, or the regon condton, we observe at least d j p(l) C 2 j/2 = C 2 j/2 snce we can assume that F j s bounded, and hence the above estmate (3.10) holds for α = 0. Thus (3.11) follows. 4 Optmal N-term Approxmaton obtaned from the EPWT Consder now the vector of all EPWT wavelet coeffcents d p = ((d 2J 1 p ) T,..., d 0 p, d 1 wth d j p = (d j p(l)) 2j 1 l=0 for j = 0,..., 2J 1, and wth the mean value d 1 p p ) T = d 1 p (0) := f 0 (0) = 2 2J F 2J (y), y I J together wth the sde nformaton on the path functons p 2J,..., p 1 at each teraton step. Wth ths nformaton the reconstructed mage Frec 2J s unquely recovered, where Frec 2J s the polyharmonc splne nterpolaton satsfyng F 2J rec ( p 2J (2 2J n) ) = f 2J (2 2J n), n = 0,..., 2 2J 1. Indeed, reconsderng the (j + 1)-th level of the EPWT procedure, we observe that the scalng coeffcents c j p(n) = f j+1, ϕ j,n and the wavelet coeffcents d j p = f j+1, ψ j,n determne f j and g j, and hence P j+1 f j+1 unquely. Further, the polyharmonc splne nterpolaton F j+1 s entrely determned by f j+1 (2 (j+1) n) and the sde nformaton about the path functon p j+1 (2 (j+1) n). By the choce of the wavelet bass, t further follows for n = 0,..., 2 j+1 1 that F j+1 ( p j+1 (2 (j+1) n)) F j+1 rec ( p j+1 (2 (j+1) n)) = f j+1 (2 (j+1) n) P j+1 f j+1 (2 (j+1) n) 2 (j+1)α/2, where Frec j+1 s unquely determned by P j+1 f j+1 (2 (j+1) n), n = 0,..., 2 j+1 1 and the sde nformaton about the path p j+1. In order to fnd a sparse approxmaton of the dgtal mage F resp. F 2J, we apply a shrnkage procedure to the EPWT wavelet coeffcents d j p(l), usng the hard threshold functon { x x σ, s σ (x) = for some σ > 0. 0 x < σ, We now study the error of a sparse representaton usng only the N wavelet coeffcents wth largest absolute value for an approxmatve reconstructon of F 2J. For convenence, let SN 2J be the set of ndces (j, l) of the N wavelet coeffcents wth largest absolute value. 21

Report on Image warping

Report on Image warping Report on Image warpng Xuan Ne, Dec. 20, 2004 Ths document summarzed the algorthms of our mage warpng soluton for further study, and there s a detaled descrpton about the mplementaton of these algorthms.

More information

APPENDIX A Some Linear Algebra

APPENDIX A Some Linear Algebra APPENDIX A Some Lnear Algebra The collecton of m, n matrces A.1 Matrces a 1,1,..., a 1,n A = a m,1,..., a m,n wth real elements a,j s denoted by R m,n. If n = 1 then A s called a column vector. Smlarly,

More information

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X Statstcs 1: Probablty Theory II 37 3 EPECTATION OF SEVERAL RANDOM VARIABLES As n Probablty Theory I, the nterest n most stuatons les not on the actual dstrbuton of a random vector, but rather on a number

More information

NUMERICAL DIFFERENTIATION

NUMERICAL DIFFERENTIATION NUMERICAL DIFFERENTIATION 1 Introducton Dfferentaton s a method to compute the rate at whch a dependent output y changes wth respect to the change n the ndependent nput x. Ths rate of change s called the

More information

More metrics on cartesian products

More metrics on cartesian products More metrcs on cartesan products If (X, d ) are metrc spaces for 1 n, then n Secton II4 of the lecture notes we defned three metrcs on X whose underlyng topologes are the product topology The purpose of

More information

Lecture 12: Discrete Laplacian

Lecture 12: Discrete Laplacian Lecture 12: Dscrete Laplacan Scrbe: Tanye Lu Our goal s to come up wth a dscrete verson of Laplacan operator for trangulated surfaces, so that we can use t n practce to solve related problems We are mostly

More information

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems Numercal Analyss by Dr. Anta Pal Assstant Professor Department of Mathematcs Natonal Insttute of Technology Durgapur Durgapur-713209 emal: anta.bue@gmal.com 1 . Chapter 5 Soluton of System of Lnear Equatons

More information

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg prnceton unv. F 17 cos 521: Advanced Algorthm Desgn Lecture 7: LP Dualty Lecturer: Matt Wenberg Scrbe: LP Dualty s an extremely useful tool for analyzng structural propertes of lnear programs. Whle there

More information

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016 U.C. Berkeley CS94: Spectral Methods and Expanders Handout 8 Luca Trevsan February 7, 06 Lecture 8: Spectral Algorthms Wrap-up In whch we talk about even more generalzatons of Cheeger s nequaltes, and

More information

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix Lectures - Week 4 Matrx norms, Condtonng, Vector Spaces, Lnear Independence, Spannng sets and Bass, Null space and Range of a Matrx Matrx Norms Now we turn to assocatng a number to each matrx. We could

More information

MMA and GCMMA two methods for nonlinear optimization

MMA and GCMMA two methods for nonlinear optimization MMA and GCMMA two methods for nonlnear optmzaton Krster Svanberg Optmzaton and Systems Theory, KTH, Stockholm, Sweden. krlle@math.kth.se Ths note descrbes the algorthms used n the author s 2007 mplementatons

More information

2.3 Nilpotent endomorphisms

2.3 Nilpotent endomorphisms s a block dagonal matrx, wth A Mat dm U (C) In fact, we can assume that B = B 1 B k, wth B an ordered bass of U, and that A = [f U ] B, where f U : U U s the restrcton of f to U 40 23 Nlpotent endomorphsms

More information

Affine transformations and convexity

Affine transformations and convexity Affne transformatons and convexty The purpose of ths document s to prove some basc propertes of affne transformatons nvolvng convex sets. Here are a few onlne references for background nformaton: http://math.ucr.edu/

More information

Problem Set 9 Solutions

Problem Set 9 Solutions Desgn and Analyss of Algorthms May 4, 2015 Massachusetts Insttute of Technology 6.046J/18.410J Profs. Erk Demane, Srn Devadas, and Nancy Lynch Problem Set 9 Solutons Problem Set 9 Solutons Ths problem

More information

Numerical Heat and Mass Transfer

Numerical Heat and Mass Transfer Master degree n Mechancal Engneerng Numercal Heat and Mass Transfer 06-Fnte-Dfference Method (One-dmensonal, steady state heat conducton) Fausto Arpno f.arpno@uncas.t Introducton Why we use models and

More information

The Order Relation and Trace Inequalities for. Hermitian Operators

The Order Relation and Trace Inequalities for. Hermitian Operators Internatonal Mathematcal Forum, Vol 3, 08, no, 507-57 HIKARI Ltd, wwwm-hkarcom https://doorg/0988/mf088055 The Order Relaton and Trace Inequaltes for Hermtan Operators Y Huang School of Informaton Scence

More information

Difference Equations

Difference Equations Dfference Equatons c Jan Vrbk 1 Bascs Suppose a sequence of numbers, say a 0,a 1,a,a 3,... s defned by a certan general relatonshp between, say, three consecutve values of the sequence, e.g. a + +3a +1

More information

Lecture 10 Support Vector Machines II

Lecture 10 Support Vector Machines II Lecture 10 Support Vector Machnes II 22 February 2016 Taylor B. Arnold Yale Statstcs STAT 365/665 1/28 Notes: Problem 3 s posted and due ths upcomng Frday There was an early bug n the fake-test data; fxed

More information

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography CSc 6974 and ECSE 6966 Math. Tech. for Vson, Graphcs and Robotcs Lecture 21, Aprl 17, 2006 Estmatng A Plane Homography Overvew We contnue wth a dscusson of the major ssues, usng estmaton of plane projectve

More information

BOUNDEDNESS OF THE RIESZ TRANSFORM WITH MATRIX A 2 WEIGHTS

BOUNDEDNESS OF THE RIESZ TRANSFORM WITH MATRIX A 2 WEIGHTS BOUNDEDNESS OF THE IESZ TANSFOM WITH MATIX A WEIGHTS Introducton Let L = L ( n, be the functon space wth norm (ˆ f L = f(x C dx d < For a d d matrx valued functon W : wth W (x postve sem-defnte for all

More information

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE Analytcal soluton s usually not possble when exctaton vares arbtrarly wth tme or f the system s nonlnear. Such problems can be solved by numercal tmesteppng

More information

20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The first idea is connectedness.

20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The first idea is connectedness. 20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The frst dea s connectedness. Essentally, we want to say that a space cannot be decomposed

More information

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur Module 3 LOSSY IMAGE COMPRESSION SYSTEMS Verson ECE IIT, Kharagpur Lesson 6 Theory of Quantzaton Verson ECE IIT, Kharagpur Instructonal Objectves At the end of ths lesson, the students should be able to:

More information

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0 MODULE 2 Topcs: Lnear ndependence, bass and dmenson We have seen that f n a set of vectors one vector s a lnear combnaton of the remanng vectors n the set then the span of the set s unchanged f that vector

More information

Some modelling aspects for the Matlab implementation of MMA

Some modelling aspects for the Matlab implementation of MMA Some modellng aspects for the Matlab mplementaton of MMA Krster Svanberg krlle@math.kth.se Optmzaton and Systems Theory Department of Mathematcs KTH, SE 10044 Stockholm September 2004 1. Consdered optmzaton

More information

Complete subgraphs in multipartite graphs

Complete subgraphs in multipartite graphs Complete subgraphs n multpartte graphs FLORIAN PFENDER Unverstät Rostock, Insttut für Mathematk D-18057 Rostock, Germany Floran.Pfender@un-rostock.de Abstract Turán s Theorem states that every graph G

More information

= z 20 z n. (k 20) + 4 z k = 4

= z 20 z n. (k 20) + 4 z k = 4 Problem Set #7 solutons 7.2.. (a Fnd the coeffcent of z k n (z + z 5 + z 6 + z 7 + 5, k 20. We use the known seres expanson ( n+l ( z l l z n below: (z + z 5 + z 6 + z 7 + 5 (z 5 ( + z + z 2 + z + 5 5

More information

Errors for Linear Systems

Errors for Linear Systems Errors for Lnear Systems When we solve a lnear system Ax b we often do not know A and b exactly, but have only approxmatons  and ˆb avalable. Then the best thng we can do s to solve ˆx ˆb exactly whch

More information

Feature Selection: Part 1

Feature Selection: Part 1 CSE 546: Machne Learnng Lecture 5 Feature Selecton: Part 1 Instructor: Sham Kakade 1 Regresson n the hgh dmensonal settng How do we learn when the number of features d s greater than the sample sze n?

More information

The Finite Element Method: A Short Introduction

The Finite Element Method: A Short Introduction Te Fnte Element Metod: A Sort ntroducton Wat s FEM? Te Fnte Element Metod (FEM) ntroduced by engneers n late 50 s and 60 s s a numercal tecnque for solvng problems wc are descrbed by Ordnary Dfferental

More information

Module 9. Lecture 6. Duality in Assignment Problems

Module 9. Lecture 6. Duality in Assignment Problems Module 9 1 Lecture 6 Dualty n Assgnment Problems In ths lecture we attempt to answer few other mportant questons posed n earler lecture for (AP) and see how some of them can be explaned through the concept

More information

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009 College of Computer & Informaton Scence Fall 2009 Northeastern Unversty 20 October 2009 CS7880: Algorthmc Power Tools Scrbe: Jan Wen and Laura Poplawsk Lecture Outlne: Prmal-dual schema Network Desgn:

More information

Salmon: Lectures on partial differential equations. Consider the general linear, second-order PDE in the form. ,x 2

Salmon: Lectures on partial differential equations. Consider the general linear, second-order PDE in the form. ,x 2 Salmon: Lectures on partal dfferental equatons 5. Classfcaton of second-order equatons There are general methods for classfyng hgher-order partal dfferental equatons. One s very general (applyng even to

More information

Case A. P k = Ni ( 2L i k 1 ) + (# big cells) 10d 2 P k.

Case A. P k = Ni ( 2L i k 1 ) + (# big cells) 10d 2 P k. THE CELLULAR METHOD In ths lecture, we ntroduce the cellular method as an approach to ncdence geometry theorems lke the Szemeréd-Trotter theorem. The method was ntroduced n the paper Combnatoral complexty

More information

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens THE CHINESE REMAINDER THEOREM KEITH CONRAD We should thank the Chnese for ther wonderful remander theorem. Glenn Stevens 1. Introducton The Chnese remander theorem says we can unquely solve any par of

More information

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal Inner Product Defnton 1 () A Eucldean space s a fnte-dmensonal vector space over the reals R, wth an nner product,. Defnton 2 (Inner Product) An nner product, on a real vector space X s a symmetrc, blnear,

More information

Appendix B. The Finite Difference Scheme

Appendix B. The Finite Difference Scheme 140 APPENDIXES Appendx B. The Fnte Dfference Scheme In ths appendx we present numercal technques whch are used to approxmate solutons of system 3.1 3.3. A comprehensve treatment of theoretcal and mplementaton

More information

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem.

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem. prnceton u. sp 02 cos 598B: algorthms and complexty Lecture 20: Lft and Project, SDP Dualty Lecturer: Sanjeev Arora Scrbe:Yury Makarychev Today we wll study the Lft and Project method. Then we wll prove

More information

ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EQUATION

ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EQUATION Advanced Mathematcal Models & Applcatons Vol.3, No.3, 2018, pp.215-222 ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EUATION

More information

VQ widely used in coding speech, image, and video

VQ widely used in coding speech, image, and video at Scalar quantzers are specal cases of vector quantzers (VQ): they are constraned to look at one sample at a tme (memoryless) VQ does not have such constrant better RD perfomance expected Source codng

More information

Nice plotting of proteins II

Nice plotting of proteins II Nce plottng of protens II Fnal remark regardng effcency: It s possble to wrte the Newton representaton n a way that can be computed effcently, usng smlar bracketng that we made for the frst representaton

More information

Asymptotics of the Solution of a Boundary Value. Problem for One-Characteristic Differential. Equation Degenerating into a Parabolic Equation

Asymptotics of the Solution of a Boundary Value. Problem for One-Characteristic Differential. Equation Degenerating into a Parabolic Equation Nonl. Analyss and Dfferental Equatons, ol., 4, no., 5 - HIKARI Ltd, www.m-har.com http://dx.do.org/.988/nade.4.456 Asymptotcs of the Soluton of a Boundary alue Problem for One-Characterstc Dfferental Equaton

More information

Convexity preserving interpolation by splines of arbitrary degree

Convexity preserving interpolation by splines of arbitrary degree Computer Scence Journal of Moldova, vol.18, no.1(52), 2010 Convexty preservng nterpolaton by splnes of arbtrary degree Igor Verlan Abstract In the present paper an algorthm of C 2 nterpolaton of dscrete

More information

Foundations of Arithmetic

Foundations of Arithmetic Foundatons of Arthmetc Notaton We shall denote the sum and product of numbers n the usual notaton as a 2 + a 2 + a 3 + + a = a, a 1 a 2 a 3 a = a The notaton a b means a dvdes b,.e. ac = b where c s an

More information

Lecture Notes on Linear Regression

Lecture Notes on Linear Regression Lecture Notes on Lnear Regresson Feng L fl@sdueducn Shandong Unversty, Chna Lnear Regresson Problem In regresson problem, we am at predct a contnuous target value gven an nput feature vector We assume

More information

DUE: WEDS FEB 21ST 2018

DUE: WEDS FEB 21ST 2018 HOMEWORK # 1: FINITE DIFFERENCES IN ONE DIMENSION DUE: WEDS FEB 21ST 2018 1. Theory Beam bendng s a classcal engneerng analyss. The tradtonal soluton technque makes smplfyng assumptons such as a constant

More information

NP-Completeness : Proofs

NP-Completeness : Proofs NP-Completeness : Proofs Proof Methods A method to show a decson problem Π NP-complete s as follows. (1) Show Π NP. (2) Choose an NP-complete problem Π. (3) Show Π Π. A method to show an optmzaton problem

More information

Global Sensitivity. Tuesday 20 th February, 2018

Global Sensitivity. Tuesday 20 th February, 2018 Global Senstvty Tuesday 2 th February, 28 ) Local Senstvty Most senstvty analyses [] are based on local estmates of senstvty, typcally by expandng the response n a Taylor seres about some specfc values

More information

Random Walks on Digraphs

Random Walks on Digraphs Random Walks on Dgraphs J. J. P. Veerman October 23, 27 Introducton Let V = {, n} be a vertex set and S a non-negatve row-stochastc matrx (.e. rows sum to ). V and S defne a dgraph G = G(V, S) and a drected

More information

Appendix B. Criterion of Riemann-Stieltjes Integrability

Appendix B. Criterion of Riemann-Stieltjes Integrability Appendx B. Crteron of Remann-Steltes Integrablty Ths note s complementary to [R, Ch. 6] and [T, Sec. 3.5]. The man result of ths note s Theorem B.3, whch provdes the necessary and suffcent condtons for

More information

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty Addtonal Codes usng Fnte Dfference Method Benamn Moll 1 HJB Equaton for Consumpton-Savng Problem Wthout Uncertanty Before consderng the case wth stochastc ncome n http://www.prnceton.edu/~moll/ HACTproect/HACT_Numercal_Appendx.pdf,

More information

PHYS 705: Classical Mechanics. Calculus of Variations II

PHYS 705: Classical Mechanics. Calculus of Variations II 1 PHYS 705: Classcal Mechancs Calculus of Varatons II 2 Calculus of Varatons: Generalzaton (no constrant yet) Suppose now that F depends on several dependent varables : We need to fnd such that has a statonary

More information

CS : Algorithms and Uncertainty Lecture 17 Date: October 26, 2016

CS : Algorithms and Uncertainty Lecture 17 Date: October 26, 2016 CS 29-128: Algorthms and Uncertanty Lecture 17 Date: October 26, 2016 Instructor: Nkhl Bansal Scrbe: Mchael Denns 1 Introducton In ths lecture we wll be lookng nto the secretary problem, and an nterestng

More information

Modeling curves. Graphs: y = ax+b, y = sin(x) Implicit ax + by + c = 0, x 2 +y 2 =r 2 Parametric:

Modeling curves. Graphs: y = ax+b, y = sin(x) Implicit ax + by + c = 0, x 2 +y 2 =r 2 Parametric: Modelng curves Types of Curves Graphs: y = ax+b, y = sn(x) Implct ax + by + c = 0, x 2 +y 2 =r 2 Parametrc: x = ax + bxt x = cos t y = ay + byt y = snt Parametrc are the most common mplct are also used,

More information

FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP

FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP C O L L O Q U I U M M A T H E M A T I C U M VOL. 80 1999 NO. 1 FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP BY FLORIAN K A I N R A T H (GRAZ) Abstract. Let H be a Krull monod wth nfnte class

More information

Polynomial Regression Models

Polynomial Regression Models LINEAR REGRESSION ANALYSIS MODULE XII Lecture - 6 Polynomal Regresson Models Dr. Shalabh Department of Mathematcs and Statstcs Indan Insttute of Technology Kanpur Test of sgnfcance To test the sgnfcance

More information

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space. Lnear, affne, and convex sets and hulls In the sequel, unless otherwse specfed, X wll denote a real vector space. Lnes and segments. Gven two ponts x, y X, we defne xy = {x + t(y x) : t R} = {(1 t)x +

More information

Linear Approximation with Regularization and Moving Least Squares

Linear Approximation with Regularization and Moving Least Squares Lnear Approxmaton wth Regularzaton and Movng Least Squares Igor Grešovn May 007 Revson 4.6 (Revson : March 004). 5 4 3 0.5 3 3.5 4 Contents: Lnear Fttng...4. Weghted Least Squares n Functon Approxmaton...

More information

Notes on Frequency Estimation in Data Streams

Notes on Frequency Estimation in Data Streams Notes on Frequency Estmaton n Data Streams In (one of) the data streamng model(s), the data s a sequence of arrvals a 1, a 2,..., a m of the form a j = (, v) where s the dentty of the tem and belongs to

More information

Maximizing the number of nonnegative subsets

Maximizing the number of nonnegative subsets Maxmzng the number of nonnegatve subsets Noga Alon Hao Huang December 1, 213 Abstract Gven a set of n real numbers, f the sum of elements of every subset of sze larger than k s negatve, what s the maxmum

More information

Geometry of Müntz Spaces

Geometry of Müntz Spaces WDS'12 Proceedngs of Contrbuted Papers, Part I, 31 35, 212. ISBN 978-8-7378-224-5 MATFYZPRESS Geometry of Müntz Spaces P. Petráček Charles Unversty, Faculty of Mathematcs and Physcs, Prague, Czech Republc.

More information

The equation of motion of a dynamical system is given by a set of differential equations. That is (1)

The equation of motion of a dynamical system is given by a set of differential equations. That is (1) Dynamcal Systems Many engneerng and natural systems are dynamcal systems. For example a pendulum s a dynamcal system. State l The state of the dynamcal system specfes t condtons. For a pendulum n the absence

More information

Generalized Linear Methods

Generalized Linear Methods Generalzed Lnear Methods 1 Introducton In the Ensemble Methods the general dea s that usng a combnaton of several weak learner one could make a better learner. More formally, assume that we have a set

More information

Indeterminate pin-jointed frames (trusses)

Indeterminate pin-jointed frames (trusses) Indetermnate pn-jonted frames (trusses) Calculaton of member forces usng force method I. Statcal determnacy. The degree of freedom of any truss can be derved as: w= k d a =, where k s the number of all

More information

DECOUPLING THEORY HW2

DECOUPLING THEORY HW2 8.8 DECOUPLIG THEORY HW2 DOGHAO WAG DATE:OCT. 3 207 Problem We shall start by reformulatng the problem. Denote by δ S n the delta functon that s evenly dstrbuted at the n ) dmensonal unt sphere. As a temporal

More information

THE WEIGHTED WEAK TYPE INEQUALITY FOR THE STRONG MAXIMAL FUNCTION

THE WEIGHTED WEAK TYPE INEQUALITY FOR THE STRONG MAXIMAL FUNCTION THE WEIGHTED WEAK TYPE INEQUALITY FO THE STONG MAXIMAL FUNCTION THEMIS MITSIS Abstract. We prove the natural Fefferman-Sten weak type nequalty for the strong maxmal functon n the plane, under the assumpton

More information

On a direct solver for linear least squares problems

On a direct solver for linear least squares problems ISSN 2066-6594 Ann. Acad. Rom. Sc. Ser. Math. Appl. Vol. 8, No. 2/2016 On a drect solver for lnear least squares problems Constantn Popa Abstract The Null Space (NS) algorthm s a drect solver for lnear

More information

Exercise Solutions to Real Analysis

Exercise Solutions to Real Analysis xercse Solutons to Real Analyss Note: References refer to H. L. Royden, Real Analyss xersze 1. Gven any set A any ɛ > 0, there s an open set O such that A O m O m A + ɛ. Soluton 1. If m A =, then there

More information

a b a In case b 0, a being divisible by b is the same as to say that

a b a In case b 0, a being divisible by b is the same as to say that Secton 6.2 Dvsblty among the ntegers An nteger a ε s dvsble by b ε f there s an nteger c ε such that a = bc. Note that s dvsble by any nteger b, snce = b. On the other hand, a s dvsble by only f a = :

More information

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U) Econ 413 Exam 13 H ANSWERS Settet er nndelt 9 deloppgaver, A,B,C, som alle anbefales å telle lkt for å gøre det ltt lettere å stå. Svar er gtt . Unfortunately, there s a prntng error n the hnt of

More information

COS 521: Advanced Algorithms Game Theory and Linear Programming

COS 521: Advanced Algorithms Game Theory and Linear Programming COS 521: Advanced Algorthms Game Theory and Lnear Programmng Moses Charkar February 27, 2013 In these notes, we ntroduce some basc concepts n game theory and lnear programmng (LP). We show a connecton

More information

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction ECONOMICS 5* -- NOTE (Summary) ECON 5* -- NOTE The Multple Classcal Lnear Regresson Model (CLRM): Specfcaton and Assumptons. Introducton CLRM stands for the Classcal Lnear Regresson Model. The CLRM s also

More information

General viscosity iterative method for a sequence of quasi-nonexpansive mappings

General viscosity iterative method for a sequence of quasi-nonexpansive mappings Avalable onlne at www.tjnsa.com J. Nonlnear Sc. Appl. 9 (2016), 5672 5682 Research Artcle General vscosty teratve method for a sequence of quas-nonexpansve mappngs Cuje Zhang, Ynan Wang College of Scence,

More information

CS 468 Lecture 16: Isometry Invariance and Spectral Techniques

CS 468 Lecture 16: Isometry Invariance and Spectral Techniques CS 468 Lecture 16: Isometry Invarance and Spectral Technques Justn Solomon Scrbe: Evan Gawlk Introducton. In geometry processng, t s often desrable to characterze the shape of an object n a manner that

More information

The Expectation-Maximization Algorithm

The Expectation-Maximization Algorithm The Expectaton-Maxmaton Algorthm Charles Elan elan@cs.ucsd.edu November 16, 2007 Ths chapter explans the EM algorthm at multple levels of generalty. Secton 1 gves the standard hgh-level verson of the algorthm.

More information

Edge Isoperimetric Inequalities

Edge Isoperimetric Inequalities November 7, 2005 Ross M. Rchardson Edge Isopermetrc Inequaltes 1 Four Questons Recall that n the last lecture we looked at the problem of sopermetrc nequaltes n the hypercube, Q n. Our noton of boundary

More information

NON-CENTRAL 7-POINT FORMULA IN THE METHOD OF LINES FOR PARABOLIC AND BURGERS' EQUATIONS

NON-CENTRAL 7-POINT FORMULA IN THE METHOD OF LINES FOR PARABOLIC AND BURGERS' EQUATIONS IJRRAS 8 (3 September 011 www.arpapress.com/volumes/vol8issue3/ijrras_8_3_08.pdf NON-CENTRAL 7-POINT FORMULA IN THE METHOD OF LINES FOR PARABOLIC AND BURGERS' EQUATIONS H.O. Bakodah Dept. of Mathematc

More information

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS SECTION 8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS 493 8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS All the vector spaces you have studed thus far n the text are real vector spaces because the scalars

More information

= = = (a) Use the MATLAB command rref to solve the system. (b) Let A be the coefficient matrix and B be the right-hand side of the system.

= = = (a) Use the MATLAB command rref to solve the system. (b) Let A be the coefficient matrix and B be the right-hand side of the system. Chapter Matlab Exercses Chapter Matlab Exercses. Consder the lnear system of Example n Secton.. x x x y z y y z (a) Use the MATLAB command rref to solve the system. (b) Let A be the coeffcent matrx and

More information

Robust Norm Equivalencies and Preconditioning

Robust Norm Equivalencies and Preconditioning Robust Norm Equvalences and Precondtonng Karl Scherer Insttut für Angewandte Mathematk, Unversty of Bonn, Wegelerstr. 6, 53115 Bonn, Germany Summary. In ths contrbuton we report on work done n contnuaton

More information

Assortment Optimization under MNL

Assortment Optimization under MNL Assortment Optmzaton under MNL Haotan Song Aprl 30, 2017 1 Introducton The assortment optmzaton problem ams to fnd the revenue-maxmzng assortment of products to offer when the prces of products are fxed.

More information

7. Products and matrix elements

7. Products and matrix elements 7. Products and matrx elements 1 7. Products and matrx elements Based on the propertes of group representatons, a number of useful results can be derved. Consder a vector space V wth an nner product ψ

More information

A Quantum Gauss-Bonnet Theorem

A Quantum Gauss-Bonnet Theorem A Quantum Gauss-Bonnet Theorem Tyler Fresen November 13, 2014 Curvature n the plane Let Γ be a smooth curve wth orentaton n R 2, parametrzed by arc length. The curvature k of Γ s ± Γ, where the sgn s postve

More information

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011 Stanford Unversty CS359G: Graph Parttonng and Expanders Handout 4 Luca Trevsan January 3, 0 Lecture 4 In whch we prove the dffcult drecton of Cheeger s nequalty. As n the past lectures, consder an undrected

More information

MA 323 Geometric Modelling Course Notes: Day 13 Bezier Curves & Bernstein Polynomials

MA 323 Geometric Modelling Course Notes: Day 13 Bezier Curves & Bernstein Polynomials MA 323 Geometrc Modellng Course Notes: Day 13 Bezer Curves & Bernsten Polynomals Davd L. Fnn Over the past few days, we have looked at de Casteljau s algorthm for generatng a polynomal curve, and we have

More information

Lecture 2: Gram-Schmidt Vectors and the LLL Algorithm

Lecture 2: Gram-Schmidt Vectors and the LLL Algorithm NYU, Fall 2016 Lattces Mn Course Lecture 2: Gram-Schmdt Vectors and the LLL Algorthm Lecturer: Noah Stephens-Davdowtz 2.1 The Shortest Vector Problem In our last lecture, we consdered short solutons to

More information

DEGREE REDUCTION OF BÉZIER CURVES USING CONSTRAINED CHEBYSHEV POLYNOMIALS OF THE SECOND KIND

DEGREE REDUCTION OF BÉZIER CURVES USING CONSTRAINED CHEBYSHEV POLYNOMIALS OF THE SECOND KIND ANZIAM J. 45(003), 195 05 DEGREE REDUCTION OF BÉZIER CURVES USING CONSTRAINED CHEBYSHEV POLYNOMIALS OF THE SECOND KIND YOUNG JOON AHN 1 (Receved 3 August, 001; revsed 7 June, 00) Abstract In ths paper

More information

EEE 241: Linear Systems

EEE 241: Linear Systems EEE : Lnear Systems Summary #: Backpropagaton BACKPROPAGATION The perceptron rule as well as the Wdrow Hoff learnng were desgned to tran sngle layer networks. They suffer from the same dsadvantage: they

More information

Isogeometric Analysis with Geometrically Continuous Functions on Multi-Patch Geometries. Mario Kapl, Florian Buchegger, Michel Bercovier, Bert Jüttler

Isogeometric Analysis with Geometrically Continuous Functions on Multi-Patch Geometries. Mario Kapl, Florian Buchegger, Michel Bercovier, Bert Jüttler Isogeometrc Analyss wth Geometrcally Contnuous Functons on Mult-Patch Geometres Maro Kapl Floran Buchegger Mchel Bercover Bert Jüttler G+S Report No 35 August 05 Isogeometrc Analyss wth Geometrcally Contnuous

More information

FINITELY-GENERATED MODULES OVER A PRINCIPAL IDEAL DOMAIN

FINITELY-GENERATED MODULES OVER A PRINCIPAL IDEAL DOMAIN FINITELY-GENERTED MODULES OVER PRINCIPL IDEL DOMIN EMMNUEL KOWLSKI Throughout ths note, s a prncpal deal doman. We recall the classfcaton theorem: Theorem 1. Let M be a fntely-generated -module. (1) There

More information

2 More examples with details

2 More examples with details Physcs 129b Lecture 3 Caltech, 01/15/19 2 More examples wth detals 2.3 The permutaton group n = 4 S 4 contans 4! = 24 elements. One s the dentty e. Sx of them are exchange of two objects (, j) ( to j and

More information

Société de Calcul Mathématique SA

Société de Calcul Mathématique SA Socété de Calcul Mathématque SA Outls d'ade à la décson Tools for decson help Probablstc Studes: Normalzng the Hstograms Bernard Beauzamy December, 202 I. General constructon of the hstogram Any probablstc

More information

The Quadratic Trigonometric Bézier Curve with Single Shape Parameter

The Quadratic Trigonometric Bézier Curve with Single Shape Parameter J. Basc. Appl. Sc. Res., (3541-546, 01 01, TextRoad Publcaton ISSN 090-4304 Journal of Basc and Appled Scentfc Research www.textroad.com The Quadratc Trgonometrc Bézer Curve wth Sngle Shape Parameter Uzma

More information

9 Characteristic classes

9 Characteristic classes THEODORE VORONOV DIFFERENTIAL GEOMETRY. Sprng 2009 [under constructon] 9 Characterstc classes 9.1 The frst Chern class of a lne bundle Consder a complex vector bundle E B of rank p. We shall construct

More information

Kernel Methods and SVMs Extension

Kernel Methods and SVMs Extension Kernel Methods and SVMs Extenson The purpose of ths document s to revew materal covered n Machne Learnng 1 Supervsed Learnng regardng support vector machnes (SVMs). Ths document also provdes a general

More information

Chapter 13: Multiple Regression

Chapter 13: Multiple Regression Chapter 13: Multple Regresson 13.1 Developng the multple-regresson Model The general model can be descrbed as: It smplfes for two ndependent varables: The sample ft parameter b 0, b 1, and b are used to

More information

5 The Rational Canonical Form

5 The Rational Canonical Form 5 The Ratonal Canoncal Form Here p s a monc rreducble factor of the mnmum polynomal m T and s not necessarly of degree one Let F p denote the feld constructed earler n the course, consstng of all matrces

More information

The Geometry of Logit and Probit

The Geometry of Logit and Probit The Geometry of Logt and Probt Ths short note s meant as a supplement to Chapters and 3 of Spatal Models of Parlamentary Votng and the notaton and reference to fgures n the text below s to those two chapters.

More information

Grover s Algorithm + Quantum Zeno Effect + Vaidman

Grover s Algorithm + Quantum Zeno Effect + Vaidman Grover s Algorthm + Quantum Zeno Effect + Vadman CS 294-2 Bomb 10/12/04 Fall 2004 Lecture 11 Grover s algorthm Recall that Grover s algorthm for searchng over a space of sze wors as follows: consder the

More information

Comparison of the Population Variance Estimators. of 2-Parameter Exponential Distribution Based on. Multiple Criteria Decision Making Method

Comparison of the Population Variance Estimators. of 2-Parameter Exponential Distribution Based on. Multiple Criteria Decision Making Method Appled Mathematcal Scences, Vol. 7, 0, no. 47, 07-0 HIARI Ltd, www.m-hkar.com Comparson of the Populaton Varance Estmators of -Parameter Exponental Dstrbuton Based on Multple Crtera Decson Makng Method

More information