Transcript L12
Lecture 12 Convolutions and summary of course Today • Convolutions • Some examples • Summary of course up to this point Remember Phils Problems and your notes = everything http://www.hep.shef.ac.uk/Phil/PHY226.htm Convolutions Imagine that we try to measure a delta function in some way … Despite the fact that the true signal is a spike, our measuring system will always render a signal that is ‘instrumentally limited’ by something often called the ‘resolution function’. Detected True signal signal If the resolution function g(t) is similar to the true signal f(t), the output function c(t) can effectively mask the true signal. True signal Resolution function http://www.jhu.edu/~signals/convolve/index.html Convolved signal Convolutions Deconvolutions We have a problem! We can measure the resolution function (by studying what we believe to be a point source or a sharp line. We can measure the convolution. What we want to know is the true signal! This happens so often that there is a word for it – we want to ‘deconvolve’ our signal. There is however an important result called the ‘Convolution Theorem’ which allows us to gain an insight into the convolution process. The convolution theorem states that:- C (k ) 2 F ( k ) G ( k ) i.e. the FT of a convolution is the product of the FTs of the original functions. We therefore find the FT of the observed signal, c(x), and of the resolution function, g(x), and use the result that If F (k ) C (k ) G ( k ) 2 F (k ) C (k ) G ( k ) 2 in order to find f(x). then taking the inverse transform, f ( x ) 1 2 e ikx C (k ) G (k ) dk Deconvolutions Of course the Convolution theorem is valid for any other pair of Fourier transforms so not only does ….. C (k ) 2 F ( k ) G ( k ) and therefore F ( k ) allowing f(x) to be determined from the FT but also C ( ) f ( x) C (k ) G ( k ) 2 1 2 2 F ( ) G ( ) and therefore F ( ) allowing f(t) to be determined from the FT f (t ) 1 2 F (k )e ikx dk C ( ) G ( ) 2 F ( ) e i t d Example of convolution f (t ) e I have a true signal at between 0 < x < ∞ which I detect using a g (t ) device with a Gaussian resolution function given by at 2 exp 2 2 a What is the frequency distribution of the detected signal function C(ω) given that C ( ) 2 F ( ) G ( ) ? Let’s find F(ω) first for the true signal … F ( ) 1 2 e at e i t F ( ) dt t ( a i ) 1 e 2 a i 0 1 2 e t ( a i ) 1 1 0 a i 2 F ( ) dt 1 2 e 1 a i 2 1 Let’s find G(ω) now for the resolution signal … 2 f (t ) e i t dt t ( a i ) 1 t ( a i ) dt 1 e 2 a i 0 Example of convolution What is the frequency distribution of the detected signal function C(ω) given that C ( ) 2 F ( ) G ( ) ? Let’s find G(ω) now for the resolution signal … g (t ) at 2 exp 2 2 a so G ( ) 1 2 G ( ) at 2 exp 2 2 a 1 2 g (t ) e i t dt i t e dt at 2 We solved this in lecture 10 so let’s go G ( ) exp i t dt 2 straight to the answer 2 2 1 G ( ) exp 2 2a a So if C ( ) F ( ) 2 F ( ) G ( ) then ….. 2 1 1 exp G ( ) 2 a i 2 2a 1 C ( ) 2 and ….. 2 1 exp 2 a i 2 2a 1 1 1 exp 2 a i 2 2a Example of convolution Detection of Dark Matter using NaI(Tl) crystal Sodium iodide crystal Caesium iodide coating PMT (turns tiny light signals into single electrons) The PMT therefore turns a tiny light pulse into an electrical signal which we view on an oscilloscope Dynode stages (amplify single electrons) Example of convolution Detection of Dark Matter using NaI(Tl) crystal Single photoelectron pulse • 10ns pulse duration • 8mV pulse height The rate of light production is convolved with the single electron pulse to yield a net pulse shape Net pulse produced by the crystal following particle interaction • 400ns pulse duration • 1200mV pulse height By deconvolving the pulse shape we can discriminate between events since neutrons and gammas for example give off light at different rates Dirac Delta Function The delta function d(x) has the value zero everywhere except at x = 0 where its value is infinitely large in such a way that its total integral is 1. Formally, for any function f(x) f ( x )d ( x x 0 ) dx f ( x 0 ) d(x – x0) is a spike centred at x = x0 d(x) is a spike centred at x = 0 d ( x ) dx 1 For example x d ( x a ) dx a 2 2 d ( x x 0 ) dx 1 Examples of Fourier Transforms: Diffraction Consider a small slit width ‘a’ illuminated by light of wavelength λ. Minima in diffraction pattern occur at:- a sin Taken from PHY102 Waves and Quanta m Intensity sin( a (sin ) / ) I I0 a (sin ) / where m is an integer. We now can rewrite as 2 a sin I I 0 sinc 2 Examples of Fourier Transforms: Diffraction At the slit, if the light amplitude is f(x), then the light intensity |f(x)|2, will be similar to the ‘top hat’ function from example 1. X domain k K domain 2 a k 2 sin F (k ) ka sinc 2 2 a The diffraction pattern on a distant screen, the intensity at any point has a sinc2 distribution and is given by the Fraunhofer diffraction equation which is related to the Fourier transform squared: |F(k)|2. 2 a sin I ( ) I 0 sinc Convolutions Example in Optics: Diffraction We’ve said that the Fraunhofer diffraction pattern from a single slit is the modulus squared of the Fourier transform of the scattering function. For double slit diffraction, the scattering function at the screen is a convolution of two functions. Here we convolve a top hat function with 2 delta functions so as to yield the characteristic equation representative of light intensity produced from infinitely narrow slits. C (k ) 2 F ( k ) G ( k ) Convolutions Example in Optics: Diffraction Consider two delta functions convolved with the single slit ‘top hat’ function: This example is specially useful because the convolution of a true signal with a delta function is the only time that you can simply express the convolution a a d single ‘top hat’ function f ( x) 1 a x a d d d two delta functions convolution g ( x) d ( x d ) d ( x d ) c( x) 1 for x d a and xd a Let us find the Fourier transforms of f and g, and their moduli squared. From earlier, we know that the FT of a top hat function is a sinc function: F (k ) 2 2 sin ka k so F (k ) 2 2 sin 2 k 2 ka Convolutions Example in Optics: Diffraction a a d d d d For the two delta functions we have:G (k ) So 1 2 | G (k ) | 2 [d ( x d ) d ( x d )] e 2 2 cos kd These are the familiar cos2 fringes we expect for two ‘infinitely narrow’ slits. ikx dx 1 2 [e ikd e ikd ] 2 2 cos kd Convolutions Example in Optics: Diffraction The diffraction pattern observed for the double slits will be the modulus squared of the Fourier transform of the whole diffracting aperture, i.e. the convolution of the delta functions with the top hat function. Hence C (k ) 2 2 2 F ( k ) G ( k ) 2 4 sin 2 2 k 2 ka cos 2 kd We see cos2 fringes modulated by a sinc2 envelope as expected Convolutions We see cos2 fringes modulated by a sinc2 envelope as expected (a) Double slits, separation 2d, infinitely narrow (b) A single slit of finite width 2a (c) Double slits of separation 2d, width 2a Double-slit Interference: Slits of finite width Fringe patterns for double slits of different widths (a) d = 50, a = Slits are 2d apart Slit width is 2a (b) d = 50, a = 5 (c) d = 50, a = 10 Parseval’s Theorem At the end of the Fourier series lectures, we very very briefly met Parseval’s theorem which states that the total energy in a wavefunction is equal to the sum of the energy in each harmonic mode. It can be shown that it is also true to say ….. or F() and f(t) Relating f(x) and F(k), dk | F ( k ) | 2 dx | f ( x ) | 2 d | F ( ) | 2 dt | f ( t ) | 2 For example, when are considering Fraunhofer diffraction, for f(x) and F(k) this formula means the total amount of light forming a diffraction pattern on the screen is equal to the total amount of light passing through the apertures; or for F(w) and f(t), the total amount of light that is recorded by the spectrometer (dispersed according to frequency) is equal to the total amount of light that entered the detector in that time interval. Convolving Two Gaussians Very often we must convolve a true signal and resolution function both of which are Gaussians ….. Remember when we calculated the FT of a Gaussian ? f ( x) a 2 e 2 ax / 2 has FT F (k ) 1 2 e k 2 2a Similarly g ( x) b 2 e f ( x) bx 2 2 has FT G (k ) 1 2 e k 2 2b g ( x) Convolving Two Gaussians So their convolution c(x) has FT C (k ) So C (k ) 2 F (k )G (k ) 1 2 e Therefore C (k ) k 2 2a e 1 2 k 1 2 2 e k 2 1 2a 2 2 2 k 2 2 1 C (k ) so 2b e e where 2 1 1 a e c( x) 2 e x 2 2 2 k 1 1 2 a b 1 b This is just the FT of a single Gaussian c(x) where ….. k 2b Convolving Two Gaussians This is an important result: we have shown that the convolution of two Gaussians characterised by a and b is also a Gaussian and is characterised by . c( x) 2 e x 2 1 2 1 a 1 b The value of is dominated by whichever is the smaller of a or b, and is always smaller then either of them. Since we found that the full width of the Gaussian f(x) is 2 2 a in lecture 9, it is the widest Gaussian that dominates, and the convolution of two Gaussians is always wider than either of the two starting Gaussians. f (x) a 2 x 2 a a 2 2 e ax / 2 Convolving Two Gaussians C (k ) 2 F (k )G (k ) 2 1 2 e k 2 2a 1 2 e k 2 2b Depending on the experiment, physicists and especially astronomers sometimes assume that both the detected signal and the resolution functions are Gaussians and use this relation in order to estimate the true width of their signal. 1 1 1 ab x 2 2 c( x) e 2 a b ab f (x) a 2 x 2 a a 2 2 e ax / 2 c( x) 2 x 2 2 2 e x 2