fundamental estimation and detection limits in linear non
play

Fundamental Estimation and Detection Limits in Linear Non-Gaussian - PowerPoint PPT Presentation

Licentiates Presentation Fundamental Estimation and Detection Limits in Linear Non-Gaussian Systems Gustaf Hendeby Automatic Control Department of Electrical Engineering Linkpings universitet AUTOMATIC CONTROL Gustaf Hendeby


  1. Licentiate’s Presentation Fundamental Estimation and Detection Limits in Linear Non-Gaussian Systems Gustaf Hendeby Automatic Control Department of Electrical Engineering Linköpings universitet AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  2. Motivation Estimation and detection are used everywhere Vital functions rely on it Information is expensive AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  3. Motivation System description e t u t System y t Measure w t x t AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  4. Motivation Noise approximation Gaussian approximation True distribution AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  5. Motivation Noise approximation Gaussian approximation True distribution AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  6. Motivation Noise approximation True distribution Gaussian approximation AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  7. Motivation Noise approximation Gaussian approximation True distribution AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  8. Outline 1. Introduction 2. Noise and Information 3. Estimation Limits 4. Detection Limits 5. Conclusions AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  9. Outline 1. Introduction 2. Noise and Information 3. Estimation Limits 4. Detection Limits 5. Conclusions AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  10. Noise 0.4 0.35 Random effects 0.3 • Measurement noise 0.25 • Process noise p(x) 0.2 More or less informative 0.15 0.1 Description: 0.05 • PDF p ( x ) 0 • Expected value: E ( x ) = µ −4 −3 −2 −1 0 1 2 3 4 x • Variance: var ( x ) = Σ p ( x ) = N ( x ; 0 , 1 ) AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  11. Information and Accuracy y t = θ + e t 2 True 1.8 Gauss Approx. 1.6 1.4 1.2 p(e) 1 0.8 0.6 0.4 0.2 0 −4 −3 −2 −1 0 1 2 3 4 e e t AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  12. Information and Accuracy y t = θ + e t 2 2 True True 1.8 Approx. Based 1.8 Gauss Approx. Meas. Meas. 1.6 1.6 1.4 1.4 1.2 1.2 p( θ |Y t ) p(e) 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 −4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4 θ e θ | Y t e t AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  13. Information and Accuracy y t = θ + e t 2 2 True True 1.8 Approx. Based 1.8 Gauss Approx. Meas. Meas. 1.6 1.6 1.4 1.4 1.2 1.2 p( θ |Y t ) p(e) 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 −4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4 θ e θ | Y t e t AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  14. Information and Accuracy y t = θ + e t 2 2 True True 1.8 Approx. Based 1.8 Gauss Approx. Meas. Meas. 1.6 1.6 1.4 1.4 1.2 1.2 p( θ |Y t ) p(e) 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 −4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4 θ e θ | Y t e t AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  15. Information and Accuracy y t = θ + e t 2 2 True True 1.8 Approx. Based 1.8 Gauss Approx. Meas. Meas. 1.6 1.6 1.4 1.4 1.2 1.2 p( θ |Y t ) p(e) 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 −4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4 θ e θ | Y t e t AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  16. Information and Accuracy y t = θ + e t 2 2 True True 1.8 Approx. Based 1.8 Gauss Approx. Meas. Meas. 1.6 1.6 1.4 1.4 1.2 1.2 p( θ |Y t ) p(e) 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 −4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4 θ e θ | Y t e t AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  17. Definitions: information measures Fisher information (true parameter θ 0 ): � � � ∆ θ I x ( θ ) = − E x θ log p ( x | θ ) � � θ = θ 0 Intrinsic accuracy (true mean µ 0 ): � � ∆ x x log p ( x | µ 0 ) I x = − E x Relative accuracy: Ψ x = var ( x ) I x Kullback-Leibler information: � p ( x ) � � I KL � � p ( · ) , q ( · ) = p ( x ) log dx q ( x ) AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  18. Definitions: information measures Fisher information (true parameter θ 0 ): � � � ∆ θ I x ( θ ) = − E x θ log p ( x | θ ) � � θ = θ 0 Intrinsic accuracy (true mean µ 0 ): � � ∆ x x log p ( x | µ 0 ) I x = − E x Relative accuracy: Ψ x = var ( x ) I x Kullback-Leibler information: � p ( x ) � � I KL � � p ( · ) , q ( · ) = p ( x ) log dx q ( x ) AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  19. Example: intrinsic accuracy Assume y i = θ + e i , e i ∼ N ( µ = 0 , Σ ) , then the intrinsic accuracy is � � ∆ e I e = − E e e log N ( e ; µ , Σ ) 1 e − ( e − µ ) 2 � � ∆ e √ = − E e e log 2 Σ 2 π Σ √ 2 π Σ + ( e − µ ) 2 � �� ∆ e � = E e log e 2 Σ � 1 ( e − µ ) = 1 1 � � � = E e ∇ e = E e Σ = var ( e ) . Σ Σ AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  20. Example: intrinsic accuracy Assume y i = θ + e i , e i ∼ N ( µ = 0 , Σ ) , then the intrinsic accuracy is � � ∆ e I e = − E e e log N ( e ; µ , Σ ) 1 e − ( e − µ ) 2 � � ∆ e √ = − E e e log 2 Σ 2 π Σ √ 2 π Σ + ( e − µ ) 2 � �� ∆ e � = E e log e 2 Σ � 1 ( e − µ ) = 1 1 � � � = E e ∇ e = E e Σ = var ( e ) . Σ Σ AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  21. Example: intrinsic accuracy Assume y i = θ + e i , e i ∼ N ( µ = 0 , Σ ) , then the intrinsic accuracy is � � ∆ e I e = − E e e log N ( e ; µ , Σ ) 1 e − ( e − µ ) 2 � � ∆ e √ = − E e e log 2 Σ 2 π Σ √ 2 π Σ + ( e − µ ) 2 � �� ∆ e � = E e log e 2 Σ � 1 ( e − µ ) = 1 1 � � � = E e ∇ e = E e Σ = var ( e ) . Σ Σ AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  22. Example: intrinsic accuracy Assume y i = θ + e i , e i ∼ N ( µ = 0 , Σ ) , then the intrinsic accuracy is � � ∆ e I e = − E e e log N ( e ; µ , Σ ) 1 e − ( e − µ ) 2 � � ∆ e √ = − E e e log 2 Σ 2 π Σ √ 2 π Σ + ( e − µ ) 2 � �� ∆ e � = E e log e 2 Σ � 1 ( e − µ ) = 1 1 � � � = E e ∇ e = E e Σ = var ( e ) . Σ Σ AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  23. Example: intrinsic accuracy Assume y i = θ + e i , e i ∼ N ( µ = 0 , Σ ) , then the intrinsic accuracy is � � ∆ e I e = − E e e log N ( e ; µ , Σ ) 1 e − ( e − µ ) 2 � � ∆ e √ = − E e e log 2 Σ 2 π Σ √ 2 π Σ + ( e − µ ) 2 � �� ∆ e � = E e log e 2 Σ � 1 ( e − µ ) = 1 1 � � � = E e ∇ e = E e Σ = var ( e ) . Σ Σ AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  24. Example: intrinsic accuracy Assume y i = θ + e i , e i ∼ N ( µ = 0 , Σ ) , then the intrinsic accuracy is � � ∆ e I e = − E e e log N ( e ; µ , Σ ) 1 e − ( e − µ ) 2 � � ∆ e √ = − E e e log 2 Σ 2 π Σ √ 2 π Σ + ( e − µ ) 2 � �� ∆ e � = E e log e 2 Σ � 1 ( e − µ ) = 1 1 � � � = E e ∇ e = E e Σ = var ( e ) . Σ Σ AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  25. Intrinsic Accuracy: outliers p 1 ( x ; ω , k ) = ( 1 − ω ) N ( x ; 0 , Σ )+ ω N ( x ; 0 , k Σ ) Inverse relative accuracy: Ψ − 1 = ( cov ( x ) I x ) − 1 x Σ − 1 : = 1 +( k − 1 ) ω , to get cov ( x ) = 1 Red is informative, blue is not k = 1 yields Gaussian distribution AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

  26. Outline 1. Introduction 2. Noise and Information 3. Estimation Limits 4. Detection Limits 5. Conclusions AUTOMATIC CONTROL Gustaf Hendeby COMMUNICATION SYSTEMS LINKÖPINGS UNIVERSITET Licentiate’s Presentation

Recommend


More recommend