ebook img

COmpRessed SolvING: Sparse Approximation of PDEs based on Compressed Sensing PDF

187 Pages·2015·4.93 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview COmpRessed SolvING: Sparse Approximation of PDEs based on Compressed Sensing

2.2. | IL MARCHIO, IL LOGOTIPO: LE DECLINAZIONI PolitecnicodiMilano MOX-ModelingandScientificComputing DipartimentodiMatematica DoctoralProgramin MathematicalModelsandMethodsforEngineering-XXVIIICycle COmpRessed SolvING: Sparse Approximation of PDEs based on Compressed Sensing Candidate: SimoneBrugiapaglia AdvisorandTutor: Prof. SimonaPerotto Co-advisor: Prof. StefanoMicheletti TheChairoftheDoctoralProgram: Prof. RobertoLucchetti Academic Year 2015-16 Abstract In this thesis, we deal with a new framework for the numerical approximation ff ofpartialdi erentialequationswhichemploysmainideasandtoolsfromcom- pressedsensinginaPetrov-Galerkinsetting. Thegoalistocomputeans-sparse approximationwithrespecttoatrialbasisofdimensionN (withs(cid:28)N)bypick- ing m(cid:28)N randomly chosen test functions, and to employ sparse optimization techniques to solve the resulting m×N underdetermined linear system. This approach has been named COmpRessedSolvING (in short, CORSING). First, we carry out an extensive numerical assessment of CORSING on ad- ff vection-di usion-reactionequations,bothinaone-andatwo-dimensionalset- ting, showing that the proposed strategy is able to reduce the computational burden associated with a standard Petrov-Galerkin formulation. Successively, we focus on the theoretical analysis of the method. In partic- ular, we prove recovery error estimates both in expectation and in probability, comparingtheerrorassociatedwiththeCORSINGsolutionwiththebests-term approximation error. With this aim, we propose a new theoretical framework based on a variant of the classical inf-sup property for sparse vectors, that is named Restricted Inf-Sup Property, and on the concept of local a-coherence, thatgeneralizesthenotionoflocalcoherencetobilinearformsinHilbertspaces. The recovery results and the corresponding hypotheses are then theoretically ff assessed on one-dimensional advection-di usion-reaction problems, while in the two-dimensional setting the verification is carried out through numerical tests. Finally, a preliminary application of CORSING to three-dimensional advec- ff tion-di usion-reaction equations and to the two-dimensional Stokes problem is also provided. Keywords: partialdifferentialequations,compressedsensing,Petrov-Galerkin formulation, inf-sup property, local coherence, estimates in expectation and probability. 1 Sommario Inquestatesivienepropostounnuovometodoperl’approssimazionenumerica ff di equazioni di erenziali alle derivate parziali, basato sull’applicazione di tec- niche e idee del compressed sensing a discretizzazioni di tipo Petrov-Galerkin. L’obiettivo è quello di calcolare una approssimazione s-sparsa rispetto ad una base trial di dimensione N (con s (cid:28) N), selezionando m (cid:28) N funzioni te- st in maniera randomizzata e, successivamente, risolvere il sistema sottode- terminato ottenuto, di dimensione m×N, tramite tecniche di ottimizzazione sparsa. Questo approccio è stato denominato COmpRessed SolvING (in breve, CORSING). Inprimis,vienecondottaunavastaindaginenumericadelCORSINGsuequa- ff zioniditipodi usione-trasporto-reazionemonodimensionaliebidimensionali, mostrandocomelastrategiapropostasiacapacediridurreilcostocomputazio- nale associato a discretizzazioni di Petrov-Galerkin standard. Successivamente,ilmetodovienestudiatodalpuntodivistateorico. Inpar- ticolare, si dimostrano delle stime di errore in valore atteso e in probabilità, mettendoaconfrontol’erroredellasoluzioneCORSINGel’erroredimigliorap- prossimazione s-sparsa. L’analisi teorica è basata su una variante della classica proprietàdiinf-suppervettorisparsi,denominataproprietàdiinf-supristretta, esulconcettodia-coerenzalocale,chegeneralizzalanozionedicoerenzalocale al caso di forme bilineari su spazi di Hilbert. I risultati teorici e le corrispetti- ff ve ipotesi vengono poi specializzati al caso di equazioni di di usione-traporto- reazionemonodimensionali,mentrenelcasobidimensionaleleipotesivengono verificate numericamente. Infine, risultati preliminari mostrano come il CORSING possa essere appli- ff cato al caso di equazioni di di usione-trasporto-reazione tridimensionali e al problema di Stokes bidimensionale. Parole chiave: equazioni differenziali alle derivate parziali, compressed sen- sing,formulazionediPetrov-Galerkin,proprietàdiinf-sup,coerenzalocale,sti- meinvaloreattesoeinprobabilità. 3 Contents Introduction 9 The COmpRessedSolvING approach . . . . . . . . . . . . . . . . . . . 9 Comparison with existing techniques . . . . . . . . . . . . . . . . . . . 10 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1 Compressedsensing 15 1.1 Three main concepts . . . . . . . . . . . . . . . . . . . . . . . . . 15 1.1.1 Sparsity: what does it mean, exactly? . . . . . . . . . . . . 16 1.1.2 Sensing: the “big soup” . . . . . . . . . . . . . . . . . . . . 18 1.1.3 Recovery: looking for a needle in a haystack . . . . . . . . 20 1.2 Theoretical tastes . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 1.2.1 The Restricted Isometry Property . . . . . . . . . . . . . . 21 1.2.2 The importance of being incoherent . . . . . . . . . . . . 23 1.2.3 Orthogonal Matching Pursuit: “greed is good” . . . . . . 25 1.2.4 Bounded Orthonormal Systems . . . . . . . . . . . . . . . 28 1.2.5 Sampling strategies based on the local coherence . . . . . 35 1.2.6 A guiding example: Haar vs Fourier . . . . . . . . . . . . 36 1.2.7 RIP for generic matrices . . . . . . . . . . . . . . . . . . . 38 2 CORSING:Towardsatheoreticalunderstanding 43 2.1 The Petrov-Galerkin method . . . . . . . . . . . . . . . . . . . . . 43 2.1.1 Weak problems in Hilbert spaces . . . . . . . . . . . . . . 43 2.1.2 From weak problems to linear systems . . . . . . . . . . . 45 2.2 CORSING: COmpRessedSolvING . . . . . . . . . . . . . . . . . . . 48 2.2.1 Description of the methodology . . . . . . . . . . . . . . . 48 ff 2.2.2 Assembling the sti ness matrix . . . . . . . . . . . . . . . 50 2.3 CORSING in action . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 2.3.1 The 1D Poisson problem . . . . . . . . . . . . . . . . . . . 54 ff 2.3.2 A 1D advection-di usion problem . . . . . . . . . . . . . 77 2.4 Extension to the 2D case . . . . . . . . . . . . . . . . . . . . . . . 79 2.4.1 The model 2D Poisson problem . . . . . . . . . . . . . . . 82 2.4.2 A 2D advection-dominated example . . . . . . . . . . . . 85 5 6 CONTENTS 2.4.3 CORSING performance . . . . . . . . . . . . . . . . . . . . 87 2.4.4 Analysis of cost reduction with respect to the full-PG ap- proach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 3 AtheoreticalstudyofCORSING 93 3.1 Formalizing the CORSING procedure . . . . . . . . . . . . . . . . 94 3.1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 3.1.2 Main hypotheses . . . . . . . . . . . . . . . . . . . . . . . 95 3.1.3 The CORSING procedure . . . . . . . . . . . . . . . . . . . 96 3.2 Theoretical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 99 3.2.1 Preliminary results . . . . . . . . . . . . . . . . . . . . . . 99 3.2.2 Non-uniform restricted inf-sup property . . . . . . . . . . 100 3.2.3 Uniform restricted inf-sup property . . . . . . . . . . . . 106 3.2.4 Recovery error analysis under the RISP . . . . . . . . . . . 108 3.2.5 Restricted Isometry Property . . . . . . . . . . . . . . . . 115 3.2.6 Recovery error analysis under the RIP . . . . . . . . . . . 116 3.2.7 Avoiding repetitions during the test selection . . . . . . . 118 ff 3.3 Application to advection-di usion-reaction equations . . . . . . 120 3.3.1 The 1D Poisson equation (HS). . . . . . . . . . . . . . . . 121 3.3.2 The 1D ADR equation (HS) . . . . . . . . . . . . . . . . . 124 3.3.3 The 1D Poisson equation (SH) . . . . . . . . . . . . . . . . 125 3.3.4 The 1D ADR equation (SH) . . . . . . . . . . . . . . . . . 127 3.3.5 The 1D diffusion equation (HS) . . . . . . . . . . . . . . . 127 3.3.6 The 2D Poisson equation (PS) . . . . . . . . . . . . . . . . 133 3.4 Further numerical experiments . . . . . . . . . . . . . . . . . . . 134 3.4.1 Sensitivity analysis of the RISP constant . . . . . . . . . . 134 3.4.2 CORSING validation . . . . . . . . . . . . . . . . . . . . . . 135 3.4.3 Convergence analysis . . . . . . . . . . . . . . . . . . . . . 138 3.4.4 Sensitivity analysis with respect to the Péclet number . . 139 4 FurtherapplicationsofCORSING 143 4.1 The Stokes problem . . . . . . . . . . . . . . . . . . . . . . . . . . 143 4.1.1 Problem setting . . . . . . . . . . . . . . . . . . . . . . . . 144 4.1.2 Petrov-Galerkin discretization . . . . . . . . . . . . . . . . 146 4.1.3 Numerical assessment of full-PG . . . . . . . . . . . . . . . 147 4.1.4 Numerical assessment of CORSING SP . . . . . . . . . . . 149 4.2 Multi-dimensional ADR problems . . . . . . . . . . . . . . . . . . 150 4.2.1 Tensorization . . . . . . . . . . . . . . . . . . . . . . . . . 152 4.2.2 The QS trial and test combination . . . . . . . . . . . . . 153 4.2.3 Local a-coherence upper bound and tensorized random- ization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 4.2.4 Well posedness of full-PG QS for the 2D Poisson problem 158 CONTENTS 7 4.2.5 Numerical results for the 2D case . . . . . . . . . . . . . . 164 4.2.6 Numerical results for the 3D case . . . . . . . . . . . . . . 165 Conclusions 169 Futuredevelopments 171 Acknowledgements 173 Listofacronyms 175 Bibliography 185

Description:
Keywords: partial differential equations, compressed sensing, Petrov- st in maniera randomizzata e, successivamente, risolvere il sistema sottode- .. CS and high-dimensional stochastic parametric PDEs Finally, it is worth men-.
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.