MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 The Universit\ of Nottingham SCHOOL OF MATHEMATICAL SCIENCES A LEVEL 3 MODULE, SPRING SEMESTER 2018-2019 MULTIVARIATE ANALYSIS Time allowed TWO Hours THIRTY Minutes CaQdidaWeV ma\ cRmSleWe Whe fURQW cRYeU Rf WheiU aQVZeU bRRk aQd VigQ WheiU deVk caUd bXW mXVW NOT ZUiWe aQ\WhiQg elVe XQWil Whe VWaUW Rf Whe e[amiQaWiRQ SeUiRd iV aQQRXQced. CUedLW ZLOO be gLYeQ fRU Whe beVW THREE aQVZeUV OQl\ VileQW, Velf-cRQWaiQed calcXlaWRUV ZiWh a SiQgle-LiQe DiVSla\ RU DXal-LiQe DiVSla\ aUe SeUmiWWed iQ WhiV e[amiQaWiRQ. DicWiRQaUieV aUe QRW allRZed ZiWh RQe e[ceSWiRQ. ThRVe ZhRVe fiUVW laQgXage iV QRW EQgliVh ma\ XVe a VWaQdaUd WUaQVlaWiRQ dicWiRQaU\ WR WUaQVlaWe beWZeeQ WhaW laQgXage aQd EQgliVh SURYided WhaW QeiWheU laQgXage iV Whe VXbjecW Rf WhiV e[amiQaWiRQ. SXbjecW VSecific WUaQVlaWiRQ dicWiRQaUieV aUe QRW SeUmiWWed. NR elecWURQic deYiceV caSable Rf VWRUiQg aQd UeWUieYiQg We[W, iQclXdiQg elecWURQic dicWiRQaUieV, ma\ be XVed. DO NOT WXUQ e[aPLQaWLRQ SaSeU RYeU XQWLO LQVWUXcWed WR dR VR ADDITIONAL MATERIAL: NHaYH¶V SWaWLVWLFV TaEOHV MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 TXUQ RYeU 1 MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 1. (a) i) What does it mean to sa\ that ( · ) is an orthogonal matrix? ii) The spectral decomposition theorem states that an\ s\mmetric matrix ৱ ( · ) ma\ be written in the form ৱ > ৫ಀ৫ԑ > า>2 ౠਅਅԑ / (1) Specif\ ৫ in terms of the ਅ, > 2-Ϳ - , and ಀ in terms of the ౠ, > 2-Ϳ - . What is the value of ਅԑ ਅো? iii) Explain how to simulate a random vector ҩ )ಛ-ಈ*, assuming that a method for simulating ҩ )ദ- ৣ* is available, where ദ is the -vector of ]eros and ৣ is the · identit\ matrix. Prove that \our method produces a random vector with the correct distribution. [12 marks] (b) i) Let ৭ denote the sample covariance matrix of -vectors 2-Ϳ - ৎ, i.e. ৭ > 2ৎ ৎา>2) ѿ Ȣ*) ѿ Ȣ*ԑ- where Ȣ > ৎѿ2Ѿৎ>2 . Assuming that ৭ has eigenvalues ౠ2 Ӓ ౠ3 Ӓ Ϳ Ӓ ౠ, prove thatnbyਉңͱਉͱ>2 ਉԑ৭ਉ > ౠ2/ Specif\ a choice of ਉ for which the maximum is attained. ii) Explain how the remaining principal components are obtained, and express them in terms of quantities which appear in the spectral decomposition (1) of the matrix৭. [12 marks] (c) Five measurements were taken on each of 49 female sparrows. These measurements (in millimetres) were: ি2 > total length, ি3 > alar extent, ি4 > length of beak and head, ি5 > length of humerus and ি6 > length of keel of sternum. The eigenvalues and eigenvectors of the sample correlation matrix of these 49 observation vectors are given b\ Eigenvalues 4/727 1/643 1/497 1/413 1/276 Eigenvectors in columns ֦֩֩ ֧֩֩ ֨ 1/563 ѿ1/162 1/7:2 ѿ1/531 1/4851/573 1/411 1/452 1/659 ѿ1/6411/562 1/436 ѿ1/566 ѿ1/717 ѿ1/4541/582 1/296 ѿ1/522 1/499 1/7631/4:9 ѿ1/988 ѿ1/28: 1/17: ѿ1/2:3 / cRQWiQXed RQ Qe[W Sage MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 2 MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 i) Draw a scree plot. Hence or otherwise, suggest the number, ো, of principal components that should be retained. Explain \our reasoning. ii) Calculate the proportion of variabilit\ explained b\ these ো components. iii) Provide an interpretation for each of the first 3 components. iv) It turns out that the instrument for measuring variable ি2 was in error b\ a constant factor 1.1. E.g. if the measurement obtained was 21mm then the correct measurement would be 21 · 2/2 > 22mm. All the other variables were measured correctl\. Discuss whether the calculation of the eigenvalues and eigenvectors of the correlation matrix needs to be done again with the corrected values of the ি2 variable given. [16 marks] MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 TXUQ OYeU 3 MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 2. (a) Suppose that 2-Ϳ - ৎ is a random sample of vectors from the )ಛ-ಈ* distribution. Throughout this question it should be assumed ಈ has full rank. i) Derive the distribution of the sample mean vector Ȣ > ৎѿ2 ৎา>2 / ii) Define the Wishart া)ಈ- ৎ* distribution and Hotelling¶s 3)- ৎ* distribution, and state without proof the distribution of ৎ৭ where ৭ > 2ৎ ৎา>2) ѿ Ȣ*) ѿ Ȣ*ԑ is the sample covariance matrix. Hence derive the distribution of)ৎ ѿ 2*) Ȣ ѿ ಛ*ԑ৭ѿ2) Ȣ ѿ ಛ*/ [12 marks] (b) A sample of 15 components produced in a certain manufacturing process was collected. The length (variable 1) and width (variable 2) of each component was measured in millimetres (mm). The sample mean vector and sample covariance matrix were found to be Ȣ > ຏ 26/6421/64 ຐ and ৭ > ຏ 2/35 1/61/6 1/69 ຐ respectivel\. i) The components are required to have a mean of ಛ1 > )26/1- 21/1*ԑ. Assess the evidence that the components produced b\ the manufacturer meet this requirement, clearl\ stating an\ assumptions that \ou make. You should present \our answer in terms of bounds for the -value of an appropriate test. ii) Explain briefl\ wh\ just investigating the mean of the manufacturing process is inadequate. HiQW: If ౩3 ҩ 3)- ৎ* WheQ \)ৎ ѿ , 2*0)ৎ*^౩3 ҩ ভ-ৎѿ,2. [10 marks] (c) Consider the multivariate linear model > ড়ԑ , ಔ- > 2-Ϳ - ৎ- where ಔ IIDҩ )ദ- ಈ*, ദ is the -vector of ]eros, ѵ ϓ, > 2-Ϳ - ৎ,ಈ ( · ) is a covariance matrix of full rank and ড় ( · ) is a parameter matrix. cRQWiQXed RQ Qe[W Sage MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 4 MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 i) Show that the log-likelihood for ড় and ಈ is given b\ ৌ)ড়-ಈ* > ѿৎ3 mph }3ಈ} ѿ 23 ৎา>2) ѿ ড়ԑ*ԑಈѿ2) ѿ ড়ԑ*/ ii) Now write ৳ > \2-Ϳ - ৎ^ԑ and ৲ > \2-Ϳ - ৎ^ԑ. Show that)৳ ѿ ৲ড়*ಈѿ2)৳ ѿ ৲ড়*ԑ > ঢ় > )ৄ*ৎ->2- whereৄ > ) ѿ ড়ԑ*ԑಈѿ2) ѿ ড়ԑ*/ Hence deduce that tr|)৳ ѿ ৲ড়*ಈѿ2)৳ ѿ ৲ড়*ԑ~ > ৎา>2) ѿ ড়ԑ*ԑಈѿ2) ѿ ড়ԑ*/ iii) Define৪ > ৣ ѿ৲)৲ԑ৲*ѿ2৲ԑ and show that ৪৲ > ദৎ- and ৲ԑ৪ > ദ-ৎ, where ദৎ- is the ৎ · matrix of ]eros. iv) Define Ƞড় > )৲ԑ৲*ѿ2৲ԑ৳ (2) and show that tr|)৳ ѿ৲ড়*ಈѿ2)৳ ѿ৲ড়*ԑ~ > tr|৪৳ ಈѿ2৳ ԑ৪~, tr|৲) Ƞড় ѿড়*ಈѿ2) Ƞড় ѿড়*ԑ৲ԑ~/ HiQW: Recall WhaW fRU aQ\ cRmSaWible maWUiceV aQd ড়, WU)ড়* > WU)ড়*. v) Deduce that Ƞড় is the maximum likelihood estimator of ড়. Briefl\ explain \our reasoning. [18 marks] MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 TXUQ OYeU 5 MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 3. (a) i) Let ʂ denote a population which is defined b\ a probabilit\ densit\ function ে)*, > 2-Ϳ - ৈ, where ѵ ϓ. What is the (population) maximum likelihood (ML) decision rule for allocating a new observation vector ѵ ϓ to one of ʂ2-Ϳ -ʂৈ? ii) Supposing now that ৈ > 3 and that the ʂ have )ಛ- ಈ* densities with different means but a common covariance matrix ಈ, which is assumed to be non-singular. The population ML discriminant rule is: ³Allocate to ʂ2 if and onl\ if৵ԑ) ѿ ৼ* ? 1 where ৵ > ಈѿ2)ಛ2 ѿ ಛ3* and ৼ > 23)ಛ2 , ಛ3*/༛ Suppose now that we do not know the values of ಛ2, ಛ3 andಈ but we do have access to training samples 2-Ϳ - ৎ and 2-Ϳ - ্ from ʂ2 and ʂ3 respectivel\. Explain how \ou would construct a ³sample ML decision rule´, defining an\ quantities that \ou use. [6 marks] (b) i) Suppose we observe training samples of 50 observation vectors 2-Ϳ - 61 from population ʂ2 and 60 observation vectors 2-Ϳ - 71 from ʂ3, with sample mean vectors and sample covariance matrices given b\ Ȣ > ຏ 5/15/4 ຐ - Ȣ > ຏ 5/75/2 ຐ - ৭ > ຏ 21 77 25 ຐ - ৭ > ຏ 23 66 24 ຐ / Determine the ML decision rule for allocating a new observation vector > )2- 3*ԑ to one of the two populations, and show that it ma\ be expressed in the form ³allocate to ʂ2 if and onl\ if 3 ? 642 ѿ 3/:8´. Present \our results graphicall\, with 2 and 3 as the coordinate axes. An\ assumptions that \ou make should be clearl\ stated. ii) To which population(s) would \ou allocate new observation vectors )5/3- 5/3*ԑ and)5/8- 5/5*ԑ using the rule derived in part (b)(i)? iii) Suppose now that is reall\ from ʂ2. Obtain an estimate of the probabilit\ that is misclassified (i.e. is allocated to ʂ3). [18 marks] (c) i) Give a brief description of Fisher¶s linear classification rule. ii) Under the Gaussian model in part (a)(ii), where ʂ has a )ಛ- ಈ* distribution, > 2-Ϳ - ৈ, and the populations have a common covariance matrix ಈ, discuss how \ou would calculate misclassification probabilities when ৈ ? 3. iii) Should we prefer the ML decision rule or Fisher¶s classification rule? Discuss briefl\. [16 marks] MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 6 MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 4. (a) The Singular Value Decomposition (SVD) states that an\ · matrix ma\ be written in the form > ৫ಃ৬ԑ > า>2 ౣਅԑ - (3) where > njo)- *, ৫ > \ਅ2-Ϳ - ਅ^, ৬ > \2-Ϳ - ^, ৫ԑ৫ > ৣ > ৬ԑ৬,ಃ > diag|ౣ2-Ϳ - ౣ~ and ౣ2 Ӓ Ϳ Ӓ ౣ Ӓ 1. i) Given a set of -variables and -variables, explain the purpose of a canonical correlation (cc) anal\sis. ii) Give details of how to calculate the first cc component: namel\, the weight vector for the -variables, the weight vector for the -variables and the first cc coefficient. You ma\ express these cc quantities in terms of the quantities which arise in the SVD in (3). Ensure that \ou define the matrix . iii) Consider now the case > 2, i.e. there is onl\ one -variable. Appl\ the formulae given in part (a)(ii) in this case, simplif\ing the expressions as far as possible, and explain the connection with the ordinar\ least squares (OLS) estimator of in the univariate linear model > ԑ , ౚ- > 2-Ϳ - ৎ- (4) where and are -vectors and ౚ IIDҩ )1- ౨3*. iv) What interpretation of OLS is suggested b\ canonical correlation anal\sis? HiQW: NRWe WhaW Whe OLS eVWimaWRU Rf iQ (4) iV giYeQ b\ Whe fRUmXla (2) ZiWh > 2. [18 marks] (b) The following dissimilarit\ matrix was obtained for 4 legal cases, ন, , প and ফ, that are being studied. Caseন প ফন 1 1/7: 1/:3 1/:6 1/7: 1 1/:9 1/:1 Case প 1/:3 1/:9 1 1/77ফ 1/:6 1/:1 1/77 1 i) Plot the dendrogram based on the ViQgle liQkage meWhRd. ii) Plot the dendrogram based on the cRmSleWe liQkage meWhRd. iii) What clusters arise in (i) and (ii) if we take the threshold as 1/:5 in each case? Comment briefl\ on the similarities and differences between the single and complete linkage methods in this example. [12 marks] cRQWiQXed RQ Qe[W Sage MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 TXUQ OYeU 7 MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 (c) i) Given a distance matrix , explain how to calculate the SUiQciSal cRRUdiQaWeV from a certain matrix, ড় sa\, which \ou should define explicitl\ in terms of . ii) Suppose that ড় has eigenvaluesౠ2 > 3- ౠ3 > 2- ౠ4 > 1 and ౠ5 > ѿ2 and corresponding unit eigenvectors ਅ2 > 23 ֛֛֚֜ 2ѿ22ѿ2 ֝֞֞ ֟ - ਅ3 > 23 ֛֛֚֜ 22ѿ2ѿ2 ֝֞֞ ֟ - ਅ4 > 23 ֛֛֚֜ 2ѿ2ѿ22 ֝֞֞ ֟ - ਅ5 > 23 ֛֛֚֜ 2222 ֝֞֞ ֟ / A) Is a Euclidean distance matrix? An\ result that \ou use should be clearl\ stated but no proof is required. B) Write down the first two principal coordinates for each point. [10 marks] MATH3030.G13MVA-E1 MATH4068.G14AMS-E1 EQG
欢迎咨询51作业君