This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. ANALYSIS IN THE FREQUENCY DOMAIN Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 1 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. SPECTRAL DENSITY Definition The spectral density of a S.S.P. y (t ) (also called the spectrum of y (t ) ) is defined as: y ( ) y ( ) e j F y ( ). In other words, y ( ) is defined as the Fourier transform F y ( ) of the covariance function. Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 2 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Properties of y ( ) : 1. y ( ) is a real function of the real variable , Imy ( ) 0 2. y ( ) is a positive function, y ( ) 0 3. y ( ) is a even function, y ( ) y ( ) 4. y ( ) is a periodic function with period equal to 2 π y ( ) y ( k 2π) , k Z Observation: as a consequence of 4, we will plot the spectral density in the interval [ , ] . y ( ) Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 3 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Observation y ( ) π is the largest frequency for sinusoidal discrete time signals. Indeed, the minimum period is T 2 which corresponds to 2 2 Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 4 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Inverse transform of y ( ) : y ( ) F y ( ) y ( ) e j y ( ) F bi-univocal relationship 1 1 π y ( ) y ( ) e j d 2π π y ( ) and y ( ) carry the same information on the properties of the process y (t ) (the spectral density is an alternative representation of the process 2nd order properties) Observation: 1 π 1 π j 0 y (0) y ( ) e d y ( )d 2π π 2π π i.e. the process variance is the rescaled area underlying y ( ) . y ( ) π π Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 5 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Spectral density of a white noise (WN) Let us consider e(t ) ~ WN ( , 2 ) . Covariance function: y ( ) 2 e ( ) 0 2 if 0 if 0 Spectral density: e ( ) e ( ) e j e (0)e j 0 e (1)e j e (1)e j e ( 0 ) 2 2 y ( ) -π White Noises have constant and equal to 2 spectral density. Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 6 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. SPECTRAL DENSITY OF S.S.P. GENERATED AS THE OUTPUT OF DIGITAL FILTERS Let the process y (t ) the steady-state output of an asymptotically stable digital filter fed by an S.S.P., i.e. y(t ) F ( z )v(t ) v(t ) y (t ) F (z ) Then, the following formula for the spectral density of y (t ) holds: y ( ) F e y ( ) j 2 v ( ) output spectral density F e j square absolute value of the filter transfer function 2 evaluate for z e j (filter frequency response) v ( ) input spectral density If the input v(t ) is a White Noise with variance 2 , then: y ( ) F e j 2 2 Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 7 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Example (MA(1) process) y(t ) e(t ) c e(t 1) , c (real coefficient) e(t ) ~ WN ( 0 ,1 ) y (0) 12 c 2 1 1 c 2 y (1) 1 c 1 c y ( ) 0 when ..., 3, 4, ... . Covariance function plot y ( ) 1 c2 c0 -5 -4 -3 -2 -1 1 2 3 4 5 c0 1 c2 Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 8 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Spectral density (via the definition) y ( ) y ( ) e j (only y (0), y (1)e j are not null) y (0) y (1)e j y (1)e j 0 1 c 2 ce j e j Euler representation of the exponential e j e j cos( ) j sin( ) cos( ) j sin( ) 2 cos( ) We have y ( ) 1 c 2 2c cos( ) which is: real 0 c even periodic with period 2π Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 9 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Spectral density (computed through the main theorem) y (t ) (1 cz 1 )e(t ) , e(t ) ~ WN ( 0 ,1 ) 2 y ( ) 1 ce j 1 Recalling that: a jb (a jb)(a jb) a 2 b 2 2 1. 1 ce j 1 c cos( ) j sin( ) 1 c cos( ) j sin( ) 2. 1 ce j 1 c cos( ) j sin( ) y ( ) 1 ce j 1 ce j 1 c 2 c(e j e j ) 1 c 2 2c cos( ) Plot of ( ) y y ( ) 1 c 2 c0 1 c2 1 c 2 -π π 2 c0 π 2 π Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 10 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. y (0) 1 c 2 2c 1 c 2 y ( 2 ) 1 c 2 y ( π) 1 c 2 2c (1 c) 2 Let us compute y (0) based on y ( ) 1 π 2 y (0) 1 c 2c cos( ) d 2π π π 1 π 2 1 c d 2c cos( )d 2π π π 1 1 c 2 2csin( ) 1 2π1 c 2 1 c 2 2π 2π Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 11 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. An alternative interpretation of the spectral density (Kinchine-Wiener) Suppose that an S.S.P. y (t ) is filtered through an (ideal) pass-band filter: FPB frequency response y (t ) FP B ~ y (t ) 1 0 FPB pass - band filter ~ y (t ) is the filtered output process Theorem (Kinchine-Wiener): y ( ) lim ~y (0) 0 Spectral density = mean energy of process realizations frequency by frequency. Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 12 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Example (White Noise) A WN process realization is erratic and unpredictable (complete uncorrelation at different time instants) 8 7 6 5 4 3 2 1 0 -1 -2 0 10 20 30 40 50 60 70 80 90 100 The WN spectral density is constant... 2 -π y ( ) ... i.e. WN energy is equally-distributed all over the frequency domain Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 13 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Example (general case) y (t ) 1 e(t ), e WN (0,1) 1 0.9 z 1 4 120 3 100 2 80 1 0 60 -1 40 -2 20 -3 -4 0 20 40 60 80 100 5 0 -4 -2 0 2 4 -2 0 2 4 100 80 60 0 40 20 -5 0 y (t ) 20 40 60 80 100 0 -4 1 e(t ), e WN (0,1) 1 0.9 z 1 Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 14 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Observation Many different representations for an ARMA process 1. “Time-domain” representation: difference equation y(t ) a1 y(t 1) ... am y(t m) c0 e(t ) ... cn e(t n) e(t ) WN ( , 2 ) 2. “Operatorial” representation: transfer function y (t ) C ( z) e(t ) A( z ) e(t ) WN ( , 2 ) 3. “Probabilistic” characterization: mean & covariance function - my - y ( ) 0,1,2,3, 4. “Frequency domain” characterization: mean & spectral density - my - y ( ) (N.B.: neither y ( ) nor y ( ) carry any kind of information about the mean value of the process). Are all the four representations equivalent for a wide-sense process characterization? Yes! Clearly 1 2 , 3 4 , 1,2 3,4 What about 3,4 1,2 ? Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 15 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Let us consider an ARMA process y (t ) C ( z) e(t ) , A( z ) e(t ) WN ( , 2 ) . y (t ) has a rational spectral density: y ( ) C (e j ) 2 A(e j ) 2 2 , where a spectral density is called “rational” if it is a rational function of the variable e j , that is p0 p1e j p2 e j 2 y ( ) q0 q1e j q2 e j 2 Is the reverse true? Yes Theorem. Let y (t ) be a S.S.P. with rational spectral density. Then, there exists a white noise process (t ) with suitable mean and variance and a rational transfer function W (z ) such that: y(t ) W ( z ) (t ) i.e. y (t ) is an ARMA process. Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 16 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Question: is the ARMA representation (i.e. the choice of (t ) and of W (z ) ) unique? NO The same process y(t ) W ( z ) (t ) , (t ) WN ( , 2 ) can be generated according to an infinite number of different ARMA models. ~ ~ ~ I.e. y (t ) can be generated also as y (t ) W ( z ) (t ) , where W ( z ) is a ~ rational transfer function different from W (z ) , and (t ) is a white noise different from (t ) . Case 1. Let be any real number y (t ) W ( z ) (t ) W ( z ) Here, y(t) does not change, we have multiplied W(z) by the identity1!!! 1 ~ ~ (t ) W ( z ) (t ) W ( z ) (t ) Thanks to the rules for the composition of transfer functions where: 1 ~ W ( z ) W ( z ) (it’s still a rational transfer function) ~ (t ) (t ) ~ (it’s still a WN, (t ) WN ( , 2 2 ) )* ~ ~ Hence, y (t ) W ( z ) (t ) is a new ARMA representation of the process y (t ) (note that y (t ) is always the same, it never changed). Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 17 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. ~ * Let us verify that (t ) is actually a white noise ~ ~ ~ E[ (t )] E[ (t )] E[ (t )] ~ ( ) E[( (t ) )( (t ) )] E[( (t ) )( (t ) )] ( ) ~ (clearly, (t ) and (t ) are two different white noises, although strictly correlated) Case 2. Let n be any integer number. Here, y(t) does not change, we have multiplied W(z) by the identity1!!! zn ~ ~ y (t ) W ( z ) (t ) W ( z ) n (t ) z n W ( z ) (t n) W ( z ) (t ) z where: ~ W ( z) z n W ( z) ~ (t ) (t n) Thanks to the rules for the composition of transfer functions and recalling the meaning of the shift operator (it’s still a rational transfer function) (it’s still a white noise)* ~ ~ Hence, y (t ) W ( z ) (t ) is a new ARMA representation of the process y (t ) (note that y (t ) is always the same, it never changed). Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 18 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. ~ * Let us verify that (t ) is actually a white noise ~ E[ (t )] E[ (t n)] (stationarity) ~ ( ) E[( (t n) )( (t n ) )] ( ) (stationarity) ~ (note that (t ) and (t ) are not the same process although they are wide-sense equivalent) Case 3. Let p be any complex number such that p 1. z y (t ) W ( z ) (t ) W ( z ) z where: z p ~ W ( z) W ( z) z p ~ (t ) (t ) p ~ ~ ( t ) W ( z ) (t ) p Here, y(t) does not change, we have multiplied W(z) by the identity1!!! (it’s still a rational transfer function) (plainly, it’s a white noise) ~ ~ Hence, y (t ) W ( z ) (t ) is a new ARMA representation of the process y (t ) (note that y (t ) is always the same, it never changed). ~ ~ (here (t ) and (t ) are the same process, but W (z ) and W ( z ) are different) Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 19 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Case 4. Let q a zero of W (z ) such that q 1. That is: W ( z) W1 ( z) ( z q) Here, y(t) does not change, we have multiplied W(z) by the identity1!!! Then, z 1q y (t ) W ( z ) (t ) W1 ( z ) ( z q) (t ) W1 ( z ) ( z q) (t ) 1 z q z q ~ ~ W1 ( z ) ( z ) ( t ) W ( z ) (t ) 1 z q 1 q where: ~ W ( z ) W1 ( z ) ( z 1q ) Thanks to the rules for the composition of transfer functions (it’s still a rational transfer function) zq ~ (t ) (t ) z 1q is it a white noise? ~ Let us compute the spectral density of (t ) ~ ( ) e j q e j 1 q 2 2 2 e e q e j q 2 j 1q e j 1q j 1 q 2 qe j e j 2 1 q 2 2q cos( ) 2 1 1 j 1 2 j 1 2 e e 1 2 cos( ) q q q q Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 20 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. 1 2 1 cos( ) 2 q q q2 2 q 2 2 1 2 1 2 cos( ) q q ~ The spectral density is constant for all values of , so that (t ) is a white noise! Moreover, 1 1q 1 q E[ (t )] E[ (t )] q q 1 1 1 q 1 q ~ ~ ~ ~ Hence, (t ) WN (q , q 2 ) and y (t ) W ( z ) (t ) is a new ARMA representation of the process y (t ) (note that y (t ) is always the same, it never changed). ~ (note instead that (t ) and (t ) are two different white noises, although strictly correlated) Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 21 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Apart from the four examined ones, there are no other sources of ambiguity in defining an ARMA process. This is expressed by the following theorem which better clarifies the previous one. Theorem (Spectral Factorization) Let y (t ) be a S.S.P. with rational spectral density. Then, there exists an unique white noise process (t ) with suitable mean and variance and an unique rational transfer function W (z ) such that: y(t ) W ( z ) (t ) , and, ( C (z ) and A(z ) are the numerator and denominator of W (z ) ): 1. C (z ) and A(z ) are monic (i.e. the coefficients of the maximum degree terms of C (z ) and A(z ) are equal to 1) 2. C (z ) and A(z ) have null relative degree 3. C (z ) and A(z ) are coprime (i.e. they have no common factors) 4. the absolute value of the poles and the zeroes of W (z ) is less than or equal to 1 (i.e. poles and zeroes are inside the unit circle) When all the four conditions above are satisfied, we will say that y(t ) W ( z ) (t ) is a canonical representation of y (t ) Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 22 This material is protected by copyright and is intended for students’ use only. Sell and distribution are strictly forbidden. Final exam requires integrating this material with teacher explanations and textbooks. Observation. Conditions 1, 2, and 3 remove any ambiguity as due to the process described in Case 1, Case 2, and Case 3, respectively. The first part of Condition 4 assure that W (z ) is asymptotically stable so that y (t ) is well defined, while the second part removes any ambiguity as due to the process described in Case 4. Model Identification and Data Analysis (MIDA) – ANALYSIS IN FREQUENCY DOMAIN 23