Minimum Mean-Square Error (MMSE) and Linear MMSE (LMMSE) Estimation
|
|
- Beverly Mills
- 5 years ago
- Views:
Transcription
1 Minimum Mean-Square Error (MMSE) and Linear MMSE (LMMSE) Estimation Outline: MMSE estimation, Linear MMSE (LMMSE) estimation, Geometric formulation of LMMSE estimation and orthogonality principle. Reading: Chapter 12 in Kay-I. EE 527, Detection and Estimation Theory, # 4b 1
2 MMSE Estimation Consider the following problem: A signal Θ = θ is transmitted through a noisy channel, modeled using the conditional pdf f X Θ (x θ), which is the likelihood function of θ. We observe X = x. The signal Θ has known prior (marginal) pdf f Θ (θ) = π(θ) which summarizes our knowledge about Θ before (i.e. prior to) collecting X = x. We wish to estimate Θ using the observation X = x: θ = θ(x) = g(x). We choose g(x) to minimize the Bayesian (preposterior) mean-square error: BMSE = E Θ,X {[ θ(x) Θ] 2 } = E Θ,X {[g(x) Θ] 2 }. Here, θ(x) that achieve the minimum BMSE are called minimum MSE (MMSE) estimates of Θ. θ(x) may not be unique. EE 527, Detection and Estimation Theory, # 4b 2
3 A Reminder: MMSE Estimation Theorem 1. The MMSE estimate of Θ (based on the observation X = x) is given by θ MMSE = g(x) = E Θ X (Θ x). (1) The minimum BMSE (i.e. the BMSE of θmmse (x) = E Θ X [Θ x]) is MBMSE = E X [var Θ X (Θ X)] = E Θ (Θ 2 ) E X {[E Θ X (Θ X)] 2 }. (2) Lemma 1. We first show that is achieved for min b E Θ [(b Θ) 2 ] = var Θ (Θ) b = E Θ (Θ). Therefore, in absence of any observations, the MMSE estimate EE 527, Detection and Estimation Theory, # 4b 3
4 of Θ is equal to the mean of the (prior, marginal) pdf of Θ: E Θ [(Θ b) 2 ] = E Θ [(Θ E Θ (Θ) + E Θ (Θ) b) 2 ] { = E Θ [Θ E Θ (Θ)] 2 + [E Θ (Θ) b] 2 } +2 [E Θ (Θ) b] [Θ E Θ (Θ)] = E Θ [(Θ E Θ [Θ]) 2 ] + (E Θ [Θ] b) 2 +2 [E Θ (Θ) b] E Θ [Θ E Θ (Θ)] }{{} 0 E Θ {[Θ E Θ (Θ)] 2 } with equality if and only if b = E Θ (Θ). Proof. (Theorem 1) We now consider our MMSE estimation problem, write BMSE of an estimator g(x) as BMSE = E Θ,X {[Θ g(x)] 2 } iter. exp. = E X ( E Θ X {[Θ g(x)] 2 X} }{{} ρ( b θ X), see handout # 4 ) and use Lemma 1 to conclude that, for each X = x, the posterior expected squared loss ρ( θ X) = E Θ X {[Θ g(x)] 2 x} EE 527, Detection and Estimation Theory, # 4b 4
5 is minimized for Thus, BMSE is minimized for g(x) = E Θ X (Θ x). g(x) = E Θ X (Θ X). We now find the minimum BMSE: MBMSE = E Θ,X {[Θ E Θ X (Θ X)] 2 } Comments: iter. exp. = E X [E Θ X {[Θ E Θ X (Θ X)] 2 X}] = E X [var Θ X (Θ X)]. (3) E X [ θ MMSE (X)] = E Θ (Θ) unbiased on average. (4) However, θmmse (X) is practically never unbiased in the classical sense: E X Θ [ θ MMSE (X) θ] θ in general. (5) You will show (5) in a HW assignment. EE 527, Detection and Estimation Theory, # 4b 5
6 For independent Θ and X, the MMSE estimate of Θ is θ MMSE (X) = E Θ (Θ). The estimation error E = θ MMSE (X) Θ (6) and the MMSE estimate θ MMSE (X) are orthogonal: E Θ,X [E θ MMSE (X)] = E Θ,X {[ θ MMSE (X) Θ] θ MMSE (X)} iter. exp. = E X {E Θ X ([ θ MMSE (X) Θ] θ MMSE (X) X)} = E X { θ MMSE (X) E Θ X [Θ θ MMSE (X) X]} = 0 since θ MMSE (X) = E Θ X [Θ X]. It is clear from this derivation that the estimation error E in (6) is orthogonal to any function g(x) of X: E Θ,X {[Θ θ MMSE (X)] g(x)} = E X {E Θ X ([Θ θ MMSE (X)] g(x) X)} = E X {g(x) E Θ X [Θ θ MMSE (X) X]} = 0. The law of conditional variances [(5) in handout # 0b] EE 527, Detection and Estimation Theory, # 4b 6
7 implies var Θ (Θ) = E X [var Θ X (Θ X)] }{{} MBMSE, see (3) +var X ( E Θ X [Θ X] ) }{{} bθ MMSE (X), see (1) i.e. the sum of the minimum BMSE for estimating Θ and variance of the MMSE estimate of Θ is equal to the (marginal, prior) variance of Θ. EE 527, Detection and Estimation Theory, # 4b 7
8 Additive Gaussian Noise Channel Consider a communication channel with input Θ N (µ Θ, τ 2 Θ) and noise W N (0, σ 2 ) where Θ and W are independent and the measurement X is modeled as X = Θ + W. (7) Find the MMSE estimate of Θ based on X and the resulting minimum BMSE (MBMSE), i.e. E Θ X (Θ X) and E X [var Θ X (Θ X)], see (1) and (2). Note: We have already considered this problem in handout # 4. We revisit it here with focus on MMSE estimation and finding MBMSE. Solution: From (7), we have: f X Θ (x θ) = N (x θ, σ 2 ). EE 527, Detection and Estimation Theory, # 4b 8
9 We now find f Θ X (θ x) using the Bayes rule: f Θ X (θ x) f Θ (θ) f X Θ (x θ) exp [ 1 2 τ 2 (θ µ Θ ) 2] exp [ 1 (x θ)2] Θ 2 σ2 [ exp 1 2 ( 1 τ Θ σ 2) θ2 + ( 1 τ 2 µ Θ + 1 ] Θ σ 2 x) θ = N implying that ( θ 1 x + 1 σ 2 τ 2 Θ µ Θ σ 2 τ Θ 2 ) ( 1, σ ) 1 τ 2 Θ θ MMSE (X) = E Θ X (Θ X) = var Θ X (Θ X) = and, consequently, ( 1 σ τ 2 Θ 1 X + 1 µ σ 2 τ Θ 2 Θ 1 σ τ 2 Θ (8) ) 1 (9) MBMSE = E X [var Θ X (Θ X)] = ( 1 σ τ 2 Θ ) 1. (10) Note: In the above example, the MMSE estimate is a linear (more precisely, constant + linear = affine) function of the EE 527, Detection and Estimation Theory, # 4b 9
10 observation X. This is not always the case, e.g. for f Θ X (θ x) = { x e x θ x > 0, θ 0 0 otherwise we obtain E Θ X (Θ X) = 1/X. Here is another example. Computing the MMSE estimator: Another example. EE 527, Detection and Estimation Theory, # 4b 10
11 EE 527, Detection and Estimation Theory, # 4b 11
12 Gaussian Linear Model (Theorem 10.3 in Kay-I) Theorem 2. Consider the linear model: X = H θ + W H is a known matrix, and W N (0, C W ) Θ N (µ Θ, C Θ ) where W and Θ are independent and C W, µ Θ, and C Θ are known hyperparameters. Then, the posterior pdf f Θ X (θ x) is Gaussian: f Θ X (θ x) = N (H T C 1 W ( θ (H T C 1 W H + C 1 Θ ) 1 (H T C 1 W x + C 1 Θ µ Θ ), H + C 1 Θ ) 1). (11) EE 527, Detection and Estimation Theory, # 4b 12
13 Proof. f Θ X (θ x) f X Θ (x θ) π(θ) exp[ 1 2 (x H θ)t C 1 W (x H θ)] exp[ 1 2 (θ µ Θ) T C 1 Θ (θ µ Θ )] exp( 1 2 θt H T C 1 W exp( 1 2 θt C 1 Θ = exp[ 1 2 θt (H T C 1 W ( N (H T C 1 W (H T C 1 W Hθ + x T C 1 W H θ) θ + µ T Θ C 1 Θ θ) H + C 1 Θ ) θ + (x T C 1 W H + µ T Θ C 1 Θ ) θ] H + C 1 Θ ) 1 (H T C 1 W x + C 1 Θ µ Θ ), H + C 1 Θ ) 1). Comments: DC-level estimation in AWGN with known variance introduced on p. 17 of handout # 4 is a special case of this result, see also Example 10.2 in Kay-I. EE 527, Detection and Estimation Theory, # 4b 13
14 Examine the posterior mean: E Θ X (θ x) = ( H T C 1 W H }{{} likelihood precision (H T C 1 W x }{{} data-dependent term + C }{{} 1 Θ ) 1 prior precision + C 1 Θ µ }{{} θ prior-dependent term Noninformative (flat) prior on θ and white noise. Consider the Jeffreys noninformative (flat) prior pdf for θ: ). π(θ) 1 (C 1 θ = 0) and white noise: C W = σ 2 I }{{} identity matrix. Then, f Θ X (θ x) in (11) simplifies to f Θ X (θ x) = N ( θ bθ LS (x) {}}{ (H T H) 1 H T x, σ 2 (H T H) 1). Prediction: We now practice prediction for this model. Say we wish to predict a X coming from the following model: X = h T θ + W EE 527, Detection and Estimation Theory, # 4b 14
15 where W N (0, σ 2 ) is independent from W, implying that X and X are conditionally independent given Θ = θ and, therefore, f X Θ,X(x θ, x) = f X Θ(x θ) = N (x h T θ, σ 2 ). Then, our posterior predictive pdf is [along the lines of (10)] f X X(x x) = f X Θ(x θ) f }{{} Θ X (θ x) }{{} N (x h T θ,σ2 ) N (θ b θ(x),cpost ) dθ where θ(x) = (H T C 1 W C post = (H T C 1 W implying H + C 1 Θ ) 1 (H T Cw 1 x + C 1 Θ µ Θ ) H + C 1 Θ ) 1 f X X(x x) = N (h T θ(x), h T C post h + σ 2 ). EE 527, Detection and Estimation Theory, # 4b 15
16 Linear MMSE (LMMSE) Estimation For exact MMSE estimation, we need to know the joint pdf (or joint pmf) f Θ,X (θ, x), typically specified through the prior (marginal) pdf/pmf f Θ (θ) and conditional pdf/pmf f X Θ (x θ), which together yield the joint pdf (or pmf, or combined pdf/pmf) f Θ,X (θ, x) = f X Θ (x θ) f Θ (θ). This information may not be available. We typically have estimates of the first and second moments of the signal and the observation, i.e. of the means, variances, and covariance between Θ and X. This information is generally not sufficient for MMSE estimation of Θ, but is sufficient for linear MMSE (LMMSE) estimation of Θ, i.e. for finding estimates of the form: θ = θ(x) = a X + b (12) that minimize BMSE: BMSE = E Θ,X {[Θ θ(x)] 2 }. The minimization is with respect to a and b. EE 527, Detection and Estimation Theory, # 4b 16
17 Note: Even though it is more appropriate to refer to this estimator as affine MMSE estimator, linear MMSE estimator is the most widely used name for it. In most applications, we consider zero-mean X and Θ; then, our estimator is indeed linear, see Theorem 3 below. Theorem 3. The LMMSE estimate of Θ is θ(x) = cov Θ,X(Θ, X) σ 2 X }{{} a opt and its BMSE is given by [X E X (X)] + E Θ (Θ) = ρ Θ,X σ Θ X E X (X) σ X + E Θ (Θ) (13) MBMSE linear = cov Θ,X (Θ θ(x), Θ) (14) = σ 2 Θ cov2 Θ,X(Θ, X) σ 2 X = (1 ρ 2 Θ,X) σ 2 Θ. (15) Here, cov Θ,X (Θ, X) = E Θ,X [(Θ µ Θ ) (X µ X )] = E Θ,X (Θ X) E Θ (Θ) E X (X) EE 527, Detection and Estimation Theory, # 4b 17
18 is introduced on p. 4 of handout # 0b, var Θ (Θ) = cov Θ (Θ, Θ) = σ 2 Θ var X (X) = σ 2 X and ρ Θ,X is the correlation coefficient between Θ and X, defined as ρ Θ,X = cov Θ,X (Θ, X) varθ (Θ) var X (X) = cov Θ,X(Θ, X) σ Θ σ X where σ Θ = σ 2 Θ and σ X = σ 2 X are the (marginal) standard deviations of Θ and X. Proof. Suppose first that the constant a has already been chosen. Then, choosing the constant b to minimize the BMSE E Θ,X [(Θ a X b) 2 ] is equivalent to finding b that minimizes E Ξ [(Ξ b) 2 ], where Ξ = Θ a X. This problem is solved in Lemma 1, and the optimal b is b = E Ξ (Ξ), i.e. b = E Ξ (Ξ) = E Θ,X (Θ a X) = E Θ (Θ) a E X (X). (16) EE 527, Detection and Estimation Theory, # 4b 18
19 Substituting (16) into E Θ,X [(Θ a X }{{} Ξ b) 2 ] yields: E Ξ {[Ξ E Ξ (Ξ)] 2 } = var Ξ (Ξ) = var Θ,X (Θ a X) (17) = σ 2 Θ + a 2 σ 2 X 2 a cov Θ,X (Θ, X) (18) which is easy to minimize with respect to a. In particular, differentiating (18) with respect to a and setting the result to zero yields 2 a σ 2 X 2 cov Θ,X (Θ, X) = 0 i.e. a cov Θ,X (X, X) cov Θ,X (Θ, X) = 0 and, finally, cov Θ,X (a X Θ, X) = 0 which is the famous orthogonality principle. Clearly, the optimal a is a opt = cov Θ,X(Θ, X) σ 2 X (19) and (13) follows. We summarize the orthogonality principle: EE 527, Detection and Estimation Theory, # 4b 19
20 or, equivalently, cov Θ,X (a opt X Θ, X) = 0 (20) cov Θ,X ( θ(x) }{{} LMMSE est. of Θ based on X ) Θ, X = 0. (21) Substituting (19) into (17) yields MMSE linear = cov Θ,X (Θ a opt X, Θ a opt X) }{{} var Θ,X (Θ a opt X) = cov Θ,X (Θ a opt X, Θ) a opt cov Θ,X (Θ a opt X, X) }{{} 0, by (20) = σ 2 Θ cov2 Θ,X(Θ, X) σ 2 X and (15) follows. By completing the squares, it is easy to check EE 527, Detection and Estimation Theory, # 4b 20
21 that, for any a R, var Θ,X (Θ a X) = var Θ,X (Θ a X + a opt X a opt X) ) = var Θ,X ((Θ a opt X) (a a opt ) X = var Θ,X (Θ a opt X) }{{} MBMSE linear +(a a opt ) 2 var X (X) 2 (a a opt ) cov Θ,X (Θ a opt X, X) }{{} 0, by (20) = σ 2 Θ [cov Θ,X(Θ, X)] 2 σ 2 X }{{} see (15) +σ 2 X (a a opt ) 2 (22) which proves MMSE optimality of (19). Comments: E X [ θ(x)] = E Θ (Θ) also true for the MMSE estimate, see (4). If ρ Θ,X = 0, i.e. Θ and X are uncorrelated, then θ(x) = E Θ (Θ) = constant i.e. LMMSE estimation ignores the observation X. EE 527, Detection and Estimation Theory, # 4b 21
22 If ρ Θ,X = ±1, i.e. Θ E Θ (Θ) and X E X (X) are linearly dependent with probability one, then the LMMSE estimate is perfect. EE 527, Detection and Estimation Theory, # 4b 22
23 LMMSE vs. MMSE In general, the LMMSE estimate is not as good as the MMSE estimate. Example: Suppose that X U( 1, 1) uniform pdf, see the table of distributions and Θ = X 2. The MMSE estimate of Θ based on X is θ(x) = E Θ X (Θ X) = X 2 which is perfect. To find the LMMSE estimate of Θ based on X, we need E X (X) = 0 E Θ (Θ) = 1 1 x dx = 1 3 cov Θ,X (Θ, X) = E Θ,X (Θ X) 0 = 0 Θ and X uncorr. EE 527, Detection and Estimation Theory, # 4b 23
24 yielding the LMMSE estimate θ(x) = E Θ (Θ) = 1 3 i.e. the observation X completely determines Θ. is totally ignored even though it An class of random signals for which the MMSE estimate is linear is the class of jointly Gaussian random signals, e.g. Θ and X in the additive Gaussian noise channel example on p. 8. EE 527, Detection and Estimation Theory, # 4b 24
25 Linear MMSE Estimation: A Geometric Formulation We first introduce some background: A vector space V (e.g. the common Euclidean space) consists of a set of vectors that is closed under two operations: vector addition: if v 1, v 2 V, then v 1 + v 2 V and scalar multiplication: if a R and v V, then a v V. An inner product, (e.g. scalar product product in Euclidean spaces), is an operation u v satisfying commutativity: u v = v u, linearity: (a u + b v) w = a u w + b v w, and the inner product of any vector with itself is non-negative: u u 0, and u u = 0 if and only if u = 0. The norm of u is defined as u = u u. u and v are orthogonal (written u v) if and only if u v = 0. EE 527, Detection and Estimation Theory, # 4b 25
26 A vector space with an inner product is called an innerproduct space. Example: Euclidean space with the scalar product. How about a vector space for random variables? Consider random variables X and Y as vectors in an innerproduct space V that contains all RVs defined over the same probability space, with vector addition: V 1 + V 2 V, scalar multiplication: a V V, inner product: V 1 V 2 = cov V 1,V 2 (V 1, V 2 ) (check that it is a legitimate inner product), the norm of V : V = var V (V 2 ) = σ V. Hence, EE 527, Detection and Estimation Theory, # 4b 26
27 inner product cov X,Y (X, Y ) norm of X σ X norm of Y σ Y cos ϕ ρ X,Y. The linear MMSE estimation problem can be recast in the above geometric framework after substituting the optimal b from (16) into E Θ,X {[Θ a X b] 2 }, yielding var Θ,X (Θ a X) = Θ a X 2. We wish to minimize this variance with respect to a. Clearly, Θ a X 2 is minimized if (θ a X) X EE 527, Detection and Estimation Theory, # 4b 27
28 i.e. if cov Θ,X (Θ a X, X) = 0 and, consequently, the MMSE-optimal linear term a is a opt = cov Θ,X(Θ, X) var X (X) = cov Θ,X(Θ, X) σ 2. X EE 527, Detection and Estimation Theory, # 4b 28
29 To summarize: Orthogonality principle: (Θ a opt X) X i.e. see (20). cov Θ,X (Θ a opt X, X) = 0 (23) EE 527, Detection and Estimation Theory, # 4b 29
30 Additive White Noise Channel Consider again the communication channel example on p. 8, with input Θ having mean µ Θ and variance τ 2 Θ and noise W having mean zero and variance σ 2, where Θ and W are independent and the measurement X is X = Θ + W. Find the LMMSE estimate of Θ based on X and the resulting BMSE (MBMSE linear ). We need E Θ (Θ) = µ Θ E X (X) = E X,W (Θ + W ) = E Θ (Θ) + E W (W ) = µ Θ and cov Θ,X (Θ, X) = cov Θ,W (Θ, Θ + W ) Θ and W uncorr. = τ 2 Θ cov X (X) = cov X,W (Θ + W, Θ + W ) Θ and W uncorr. = τ 2 Θ + σ 2. EE 527, Detection and Estimation Theory, # 4b 30
31 The LMMSE estimate of X is θ(x) = cov Θ,X(Θ, X) σ 2 X [X E X (X)] + E X (X) = = = τ 2 Θ τ 2 Θ + σ 2 (X µ X) + µ X τ 2 Θ τ 2 Θ + σ 2 X + 1 σ 2 X + 1 τ 2 Θ µ Θ 1 σ τ 2 Θ σ2 τ 2 Θ + σ 2 µ Θ which is the same as the MMSE estimate in (8). EE 527, Detection and Estimation Theory, # 4b 31
32 Example: Estimating the Bias of a Coin Suppose that (prior) pdf of heads Θ of a coin is f Θ (θ) = U(θ 0, 1) = i (0,1) (θ). We flip this coin N times and record the number of heads X. Then, if the coin flips are independent, identically distributed (i.i.d.), the conditional pdf of X given Θ = θ is ( ) N f X Θ (x θ) = θ x (1 θ) N x = Bin(x N, θ) binomial pdf. x (24) Find the MMSE and LMMSE estimates of Θ based on X. MMSE: f Θ X (θ x) f Θ (θ) f X Θ (x θ) i (0,1) (θ) θ x (1 θ) N x = Beta(θ x + 1, N x + 1) see the table of distributions. Now, the MMSE estimate of Θ is θ MMSE (x) = E Θ X (Θ X = x) = x + 1 N + 2. EE 527, Detection and Estimation Theory, # 4b 32
33 LMMSE: We need µ Θ = E Θ (Θ) = 1 2 mean of uniform(0, 1) pdf µ X = E Θ,X (X) iter. exp. = E Θ [E X Θ (X Θ)] = E Θ ( N}{{} Θ ) = 1 2 N mean of binomial pdf in (24) and σ 2 X cond. var. = E Θ { var X Θ (X Θ) }{{} var of binomial in (24) = E Θ [N Θ (1 Θ)] + var Θ (N Θ) = N E Θ [Θ (1 Θ)] + N 2 var Θ (Θ) = N ( ) + N } + var Θ { E X Θ [X Θ] } }{{} mean of binomial in (24) = N (N + 2) 12 cov Θ,X (Θ, X) = E Θ,X (Θ X) µ Θ µ X iterated exp. = E Θ {E X Θ [Θ X Θ]} N 4 = E Θ [Θ E X Θ (X Θ)] N 4 = E Θ{Θ N}{{} Θ } N 4 mean of binomial in (24) = N 3 N 4 = N 12. EE 527, Detection and Estimation Theory, # 4b 33
34 Now, θ(x) = cov Θ,X(Θ, X) σ 2 (X µ X ) + µ Θ X N/12 = N (N + 2)/12 (X 1 2 N) = X + 1 N + 2. In this example, the MMSE and LMMSE estimates of θ are the same. EE 527, Detection and Estimation Theory, # 4b 34
35 Linear MMSE Estimation: the Vector Case (FIR Wiener Filter) Consider the signal of interest with prior knowledge described by the pdf Θ f Θ (θ) and an N-dimensional random vector X representing the observations. The MMSE estimate of X is the conditional expectation E Θ X (Θ X) which may be difficult to find in practice, since it requires knowledge of the joint distribution of Θ and X. The linear MMSE estimate of X is easier to find, since it depends only on the means, variances, and covariances of the random variables and vectors involved. EE 527, Detection and Estimation Theory, # 4b 35
36 Linear MMSE Estimation via the Orthogonality Principle We wish to find an N 1 vector a and a constant b such that θ(x) = a T X + b = N a i X i + b i=1 minimizes the BMSE BMSE = E Θ,X {[Θ θ(x)] 2 } where X[0] X =.... X[N 1] Suppose first that the constant vector a has already been chosen. Then, choosing the constant b to minimize the BMSE BMSE = E Θ,X [(Θ a T X b) 2 ] is equivalent to finding b that minimizes E Ξ [(Ξ b) 2 ] where Ξ = Θ a T X. This problem is solved in Lemma 1 and the optimal b is b = E Ξ (Ξ) = E Θ,X (Θ a T X) = E Θ (Θ) a T E X (X). (25) EE 527, Detection and Estimation Theory, # 4b 36
37 We view Θ, X[0],..., X[N 1] as vectors in an inner-product space. The linear MMSE estimation problem can be cast into our geometric framework after substituting the optimal b in (25) into BMSE = E Θ,X {[Θ θ(x)] 2 }, yielding var Θ,X (Θ a T X) = Θ a T X 2. (26) We minimize this variance with respect to a. Clearly, Θ a T X 2 is minimized if a is chosen to satisfy the orthogonality principle: (Θ a T X) subspace V N spanned by X[0], X[1],..., X[N 1] EE 527, Detection and Estimation Theory, # 4b 37
38 or, equivalently, cov Θ,X (Θ a T X, X[n]) = 0 n = 0, 1,..., N 1 (27) which gives the following set of equations: or N 1 l=0 ( N 1 cov Θ,X[n] (Θ, X[n]) cov X l=0 ) a l X[l], X[n] = 0 cov X[n],X[l] (X[n], X[l]) a l = cov X[n],Θ (X[n], Θ). (28) Define the crosscovariance vector between X and Θ and covariance matrix of X as cov X[0],X (X[0], Θ) σ X,Θ = cov X,Θ (X, Θ) = cov X[1],Θ (X[1], Θ). cov X[N 1],Θ (X[N 1], Θ) and Σ X = cov X (X) and use these definitions to compactly write (28): Σ X a = σ X,Θ. EE 527, Detection and Estimation Theory, # 4b 38
39 If Σ X is a positive definite matrix, we can solve for a: a opt = Σ 1 X σ X,Θ (29) and, finally, the LMMSE estimate of Θ is [using (25)] θ(x) = a T opt X + E Θ (Θ) a T opt E X (X) = σ T X,ΘΣ 1 X }{{} a T opt [X E X (X)] + E Θ (Θ). (30) Compare this result to the scalar case in (13): θ(x) = cov Θ,X(Θ, X) σ 2 X }{{} a opt [X E X (X)] + E Θ (Θ). We now find the minimum BMSE of our LMMSE estimator: substitute (29) into (26), yielding MBMSE linear = cov Θ,X (Θ a T opt X, Θ a T opt X) = cov Θ,X (Θ a T opt X, Θ) cov Θ,X (Θ a T opt X, a T opt X) = cov Θ,X (Θ a T opt X, Θ) cov Θ,X (Θ a T opt X, X) }{{} = 0, see (27) a opt = cov Θ,X (Θ a T opt X, Θ) (31) EE 527, Detection and Estimation Theory, # 4b 39
40 which can also be written as and further simplified: MBMSE linear = cov Θ,X (Θ θ(x), Θ) (32) MBMSE linear = σ 2 X a T opt cov X,Θ (X, Θ) }{{} σ X,Θ see (29) = σ 2 Θ σ T X,Θ Σ 1 X σ X,Θ. (33) Compare this result to the scalar case in (15): MBMSE linear = σ 2 Θ cov2 Θ,X(Θ, X) σ 2. X EE 527, Detection and Estimation Theory, # 4b 40
41 Example: Additive Noise Channel Again: Θ N (µ Θ, τ 2 Θ) where µ Θ and τ 2 Θ are known hyperparameters. We collect multiple observations X[n], modeled as X[n] = Θ + W [n] n = 0, 1,..., N 1 where W [n] are zero-mean uncorrelated RVs with known variance σ 2. We also know that Θ and W [n] are uncorrelated for all n. Find the LMMSE estimate of Θ based on X = Find also the minimum BMSE. X[0]... X[N 1]. By the orthogonality principle (27), we have: ( N 1 cov Θ,X[n] (Θ, X[n]) cov X l=0 ) a l X[l], X[n] = 0 EE 527, Detection and Estimation Theory, # 4b 41
42 for n = 0, 1,..., N 1. Here, cov Θ,X[n] (Θ, X[n]) = cov Θ,W [n] (Θ + W [n], Θ) = cov Θ (Θ, Θ) + cov W [n],x (W [n], Θ) = var Θ (Θ) = τ 2 Θ (34) cov X[l],X[n] (X[l], X[n]) = cov Θ,W (Θ + W [l], Θ + W [n]) { τ = Θ, 2 l n τ Θ 2 + σ 2 (35), l = n and, therefore, τ 2 Θ = (τ Θ 2 + σ 2 ) a 0 + τ 2 Θ a τ 2 Θ a N 1 τ 2 Θ = τ 2 Θ a 0 + (τ Θ 2 + σ 2 ) a τ 2 Θ a N 1 τ 2 Θ = τ 2 Θ a 0 + τ 2 Θ a (τ Θ 2 + σ 2 ) a N 1. Now, by symmetry, a opt,1 = a opt,2 =... = a opt,n = τ 2 Θ N τ 2 Θ + σ 2 EE 527, Detection and Estimation Theory, # 4b 42
43 yielding θ(x) = = = τ 2 Θ N τ 2 Θ + σ 2 (X[n] µ Θ ) + µ Θ N 1 n=0 τ 2 ( N 1 Θ N τ 2 Θ + σ 2 N σ 2 X + 1 τ 2 Θ µ Θ N + 1 σ 2 τ Θ 2 n=0 ) X[n] + σ 2 N τ 2 Θ + σ 2 µ Θ (36) where X = 1 N N 1 n=0 X[n]. The minimum average MSE of our LMMSE estimator follows EE 527, Detection and Estimation Theory, # 4b 43
44 by using (31): MBMSE linear = cov Θ,X (Θ θ(x), Θ) = τ 2 Θ cov Θ,X ( θ(x), Θ) ( = τ 2 τ 2 ( N 1 Θ Θ cov Θ,X N τ 2 Θ + σ 2 see (36) see (34) = τ 2 Θ = τ 2 Θ = = τ 2 Θ N τ 2 Θ + σ 2 N 1 n=0 τ 2 Θ N τ 2 Θ + σ 2 N τ 2 Θ τ 2 Θ σ 2 N τ 2 Θ + σ 2 ( N σ τ 2 Θ ) 1 n=0 ) ) X[n], Θ cov X[n],Θ (X[n], Θ) which is the same as τ 2 N in (15) of handout # 4. EE 527, Detection and Estimation Theory, # 4b 44
Nonlife Actuarial Models. Chapter 7 Bühlmann Credibility
Nonlife Actuarial Models Chapter 7 Bühlmann Credibility Learning Objectives 1. Basic framework of Bühlmann credibility 2. Variance decomposition 3. Expected value of the process variance 4. Variance of
More informationBayesian Methods: Naïve Bayes
Bayesian Methods: Naïve Bayes Nicholas Ruozzi University of Texas at Dallas based on the slides of Vibhav Gogate Last Time Parameter learning Learning the parameter of a simple coin flipping model Prior
More informationMixture Models & EM. Nicholas Ruozzi University of Texas at Dallas. based on the slides of Vibhav Gogate
Mixture Models & EM Nicholas Ruozzi University of Texas at Dallas based on the slides of Vibhav Gogate Previously We looked at -means and hierarchical clustering as mechanisms for unsupervised learning
More informationSpecial Topics: Data Science
Special Topics: Data Science L Linear Methods for Prediction Dr. Vidhyasaharan Sethu School of Electrical Engineering & Telecommunications University of New South Wales Sydney, Australia V. Sethu 1 Topics
More informationThe Simple Linear Regression Model ECONOMETRICS (ECON 360) BEN VAN KAMMEN, PHD
The Simple Linear Regression Model ECONOMETRICS (ECON 360) BEN VAN KAMMEN, PHD Outline Definition. Deriving the Estimates. Properties of the Estimates. Units of Measurement and Functional Form. Expected
More informationStat 139 Homework 3 Solutions, Spring 2015
Stat 39 Homework 3 Solutions, Spring 05 Problem. Let i Nµ, σ ) for i,..., n, and j Nµ, σ ) for j,..., n. Also, assume that all observations are independent from each other. In Unit 4, we learned that the
More informationMidterm Exam 1, section 2. Thursday, September hour, 15 minutes
San Francisco State University Michael Bar ECON 312 Fall 2018 Midterm Exam 1, section 2 Thursday, September 27 1 hour, 15 minutes Name: Instructions 1. This is closed book, closed notes exam. 2. You can
More informationA COURSE OUTLINE (September 2001)
189-265A COURSE OUTLINE (September 2001) 1 Topic I. Line integrals: 2 1 2 weeks 1.1 Parametric curves Review of parametrization for lines and circles. Paths and curves. Differentiation and integration
More informationPolynomial DC decompositions
Polynomial DC decompositions Georgina Hall Princeton, ORFE Joint work with Amir Ali Ahmadi Princeton, ORFE 7/31/16 DIMACS Distance geometry workshop 1 Difference of convex (dc) programming Problems of
More informationBuilding an NFL performance metric
Building an NFL performance metric Seonghyun Paik (spaik1@stanford.edu) December 16, 2016 I. Introduction In current pro sports, many statistical methods are applied to evaluate player s performance and
More informationOn the convergence of fitting algorithms in computer vision
On the convergence of fitting algorithms in computer vision N. Chernov Department of Mathematics University of Alabama at Birmingham Birmingham, AL 35294 chernov@math.uab.edu Abstract We investigate several
More informationPrediction Market and Parimutuel Mechanism
Prediction Market and Parimutuel Mechanism Yinyu Ye MS&E and ICME Stanford University Joint work with Agrawal, Peters, So and Wang Math. of Ranking, AIM, 2 Outline World-Cup Betting Example Market for
More informationCalculation of Trail Usage from Counter Data
1. Introduction 1 Calculation of Trail Usage from Counter Data 1/17/17 Stephen Martin, Ph.D. Automatic counters are used on trails to measure how many people are using the trail. A fundamental question
More informationImperfectly Shared Randomness in Communication
Imperfectly Shared Randomness in Communication Madhu Sudan Harvard Joint work with Clément Canonne (Columbia), Venkatesan Guruswami (CMU) and Raghu Meka (UCLA). 11/16/2016 UofT: ISR in Communication 1
More informationFinding your feet: modelling the batting abilities of cricketers using Gaussian processes
Finding your feet: modelling the batting abilities of cricketers using Gaussian processes Oliver Stevenson & Brendon Brewer PhD candidate, Department of Statistics, University of Auckland o.stevenson@auckland.ac.nz
More informationBusiness and housing market cycles in the euro area: a multivariate unobserved component approach
Business and housing market cycles in the euro area: a multivariate unobserved component approach Laurent Ferrara (a) and Siem Jan Koopman (b) http://staff.feweb.vu.nl/koopman (a) Banque de France (b)
More informationSection I: Multiple Choice Select the best answer for each problem.
Inference for Linear Regression Review Section I: Multiple Choice Select the best answer for each problem. 1. Which of the following is NOT one of the conditions that must be satisfied in order to perform
More informationThe Intrinsic Value of a Batted Ball Technical Details
The Intrinsic Value of a Batted Ball Technical Details Glenn Healey, EECS Department University of California, Irvine, CA 9617 Given a set of observed batted balls and their outcomes, we develop a method
More informationChapter 20. Planning Accelerated Life Tests. William Q. Meeker and Luis A. Escobar Iowa State University and Louisiana State University
Chapter 20 Planning Accelerated Life Tests William Q. Meeker and Luis A. Escobar Iowa State University and Louisiana State University Copyright 1998-2008 W. Q. Meeker and L. A. Escobar. Based on the authors
More informationChapter 12 Practice Test
Chapter 12 Practice Test 1. Which of the following is not one of the conditions that must be satisfied in order to perform inference about the slope of a least-squares regression line? (a) For each value
More informationLecture 5. Optimisation. Regularisation
Lecture 5. Optimisation. Regularisation COMP90051 Statistical Machine Learning Semester 2, 2017 Lecturer: Andrey Kan Copyright: University of Melbourne Iterative optimisation Loss functions Coordinate
More informationIntroduction to Pattern Recognition
Introduction to Pattern Recognition Jason Corso SUNY at Buffalo 12 January 2009 J. Corso (SUNY at Buffalo) Introduction to Pattern Recognition 12 January 2009 1 / 28 Pattern Recognition By Example Example:
More informationWhich On-Base Percentage Shows. the Highest True Ability of a. Baseball Player?
Which On-Base Percentage Shows the Highest True Ability of a Baseball Player? January 31, 2018 Abstract This paper looks at the true on-base ability of a baseball player given their on-base percentage.
More information2 When Some or All Labels are Missing: The EM Algorithm
CS769 Spring Advanced Natural Language Processing The EM Algorithm Lecturer: Xiaojin Zhu jerryzhu@cs.wisc.edu Given labeled examples (x, y ),..., (x l, y l ), one can build a classifier. If in addition
More informationName May 3, 2007 Math Probability and Statistics
Name May 3, 2007 Math 341 - Probability and Statistics Long Exam IV Instructions: Please include all relevant work to get full credit. Encircle your final answers. 1. An article in Professional Geographer
More informationISyE 6414 Regression Analysis
ISyE 6414 Regression Analysis Lecture 2: More Simple linear Regression: R-squared (coefficient of variation/determination) Correlation analysis: Pearson s correlation Spearman s rank correlation Variable
More informationFirst-Server Advantage in Tennis Matches
First-Server Advantage in Tennis Matches Iain MacPhee and Jonathan Rougier Department of Mathematical Sciences University of Durham, U.K. Abstract We show that the advantage that can accrue to the server
More informationGLMM standardisation of the commercial abalone CPUE for Zones A-D over the period
GLMM standardisation of the commercial abalone for Zones A-D over the period 1980 2015 Anabela Brandão and Doug S. Butterworth Marine Resource Assessment & Management Group (MARAM) Department of Mathematics
More informationi) Linear programming
Problem statement Sailco Corporation must determine how many sailboats should be produced during each of the next four quarters (one quarter = three months). The demand during each of the next four quarters
More informationLooking at Spacings to Assess Streakiness
Looking at Spacings to Assess Streakiness Jim Albert Department of Mathematics and Statistics Bowling Green State University September 2013 September 2013 1 / 43 The Problem Collect hitting data for all
More informationLogistic Regression. Hongning Wang
Logistic Regression Hongning Wang CS@UVa Today s lecture Logistic regression model A discriminative classification model Two different perspectives to derive the model Parameter estimation CS@UVa CS 6501:
More informationIntroduction to Pattern Recognition
Introduction to Pattern Recognition Jason Corso SUNY at Buffalo 19 January 2011 J. Corso (SUNY at Buffalo) Introduction to Pattern Recognition 19 January 2011 1 / 32 Examples of Pattern Recognition in
More informationAccounting for the Evolution of U.S. Wage Inequality
Accounting for the Evolution of U.S. Wage Inequality Lutz Hendricks UNC Preliminary May 24, 2013 1/42 Motivation Wage inequality has increased in the U.S. since the 1970s standard deviation of log wages,
More informationEquation 1: F spring = kx. Where F is the force of the spring, k is the spring constant and x is the displacement of the spring. Equation 2: F = mg
1 Introduction Relationship between Spring Constant and Length of Bungee Cord In this experiment, we aimed to model the behavior of the bungee cord that will be used in the Bungee Challenge. Specifically,
More informationThe notion of independence in the theory of evidence: An algebraic study
The notion of independence in the theory of evidence: An algebraic study Fabio Cuzzolin a a Perception group INRIA Rhône-Alpes 655, avenue de l Europe 38334 SAINT ISMIER CEDEX, France Abstract In this
More informationRobust specification testing in regression: the FRESET test and autocorrelated disturbances
Robust specification testing in regression: the FRESET test and autocorrelated disturbances Linda F. DeBenedictis and David E. A. Giles * Policy and Research Division, Ministry of Human Resources, 614
More informationModelling and Simulation of Environmental Disturbances
Modelling and Simulation of Environmental Disturbances (Module 5) Dr Tristan Perez Centre for Complex Dynamic Systems and Control (CDSC) Prof. Thor I Fossen Department of Engineering Cybernetics 18/09/2007
More informationDistancei = BrandAi + 2 BrandBi + 3 BrandCi + i
. Suppose that the United States Golf Associate (USGA) wants to compare the mean distances traveled by four brands of golf balls when struck by a driver. A completely randomized design is employed with
More informationECO 745: Theory of International Economics. Jack Rossbach Fall Lecture 6
ECO 745: Theory of International Economics Jack Rossbach Fall 2015 - Lecture 6 Review We ve covered several models of trade, but the empirics have been mixed Difficulties identifying goods with a technological
More informationGerald D. Anderson. Education Technical Specialist
Gerald D. Anderson Education Technical Specialist The factors which influence selection of equipment for a liquid level control loop interact significantly. Analyses of these factors and their interactions
More informationExistence of Nash Equilibria
Existence of Nash Equilibria Before we can prove the existence, we need to remind you of the fixed point theorem: Kakutani s Fixed Point Theorem: Consider X R n a compact convex set and a function f: X
More informationProjecting Three-Point Percentages for the NBA Draft
Projecting Three-Point Percentages for the NBA Draft Hilary Sun hsun3@stanford.edu Jerold Yu jeroldyu@stanford.edu December 16, 2017 Roland Centeno rcenteno@stanford.edu 1 Introduction As NBA teams have
More informationCOMPLEX NUMBERS. Powers of j. Consider the quadratic equation; It has no solutions in the real number system since. j 1 ie. j 2 1
COMPLEX NUMBERS Consider the quadratic equation; x 0 It has no solutions in the real number system since x or x j j ie. j Similarly x 6 0 gives x 6 or x 6 j Powers of j j j j j. j j j j.j jj j j 5 j j
More informationDevelopment of Decision Support Tools to Assess Pedestrian and Bicycle Safety: Development of Safety Performance Function
Development of Decision Support Tools to Assess Pedestrian and Bicycle Safety: Development of Safety Performance Function Valerian Kwigizile, Jun Oh, Ron Van Houten, & Keneth Kwayu INTRODUCTION 2 OVERVIEW
More informationBayes, Oracle Bayes, and Empirical Bayes
Bayes, Oracle Bayes, and Empirical Bayes Empirical Bayes Inference Robbins (1951) Robbins (1956) Compound Decision Procedures Empirical Bayes Question How does empirical Bayes relate to Bayesian and frequentist
More informationLecture 10. Support Vector Machines (cont.)
Lecture 10. Support Vector Machines (cont.) COMP90051 Statistical Machine Learning Semester 2, 2017 Lecturer: Andrey Kan Copyright: University of Melbourne This lecture Soft margin SVM Intuition and problem
More informationDOE Golfer Experiment
DOE Golfer Experiment A Design of Experiments Report Travis Anderson Jake Munger Deshun Xu 11/11/2008 INTRODUCTION We used Response Surface Methodology to optimize a golf putter. A face centered Central
More informationAlgebra I: A Fresh Approach. By Christy Walters
Algebra I: A Fresh Approach By Christy Walters 2005 A+ Education Services All rights reserved. No part of this publication may be reproduced, distributed, stored in a retrieval system, or transmitted,
More informationSTANDARD SCORES AND THE NORMAL DISTRIBUTION
STANDARD SCORES AND THE NORMAL DISTRIBUTION REVIEW 1.MEASURES OF CENTRAL TENDENCY A.MEAN B.MEDIAN C.MODE 2.MEASURES OF DISPERSIONS OR VARIABILITY A.RANGE B.DEVIATION FROM THE MEAN C.VARIANCE D.STANDARD
More informationCT4510: Computer Graphics. Transformation BOCHANG MOON
CT4510: Computer Graphics Transformation BOCHANG MOON 2D Translation Transformations such as rotation and scale can be represented using a matrix M ee. gg., MM = SSSS xx = mm 11 xx + mm 12 yy yy = mm 21
More informationPREDICTING the outcomes of sporting events
CS 229 FINAL PROJECT, AUTUMN 2014 1 Predicting National Basketball Association Winners Jasper Lin, Logan Short, and Vishnu Sundaresan Abstract We used National Basketball Associations box scores from 1991-1998
More informationMathematics of Pari-Mutuel Wagering
Millersville University of Pennsylvania April 17, 2014 Project Objectives Model the horse racing process to predict the outcome of a race. Use the win and exacta betting pools to estimate probabilities
More informationAnalysis of Shear Lag in Steel Angle Connectors
University of New Hampshire University of New Hampshire Scholars' Repository Honors Theses and Capstones Student Scholarship Spring 2013 Analysis of Shear Lag in Steel Angle Connectors Benjamin Sawyer
More informationThe next criteria will apply to partial tournaments. Consider the following example:
Criteria for Assessing a Ranking Method Final Report: Undergraduate Research Assistantship Summer 2003 Gordon Davis: dagojr@email.arizona.edu Advisor: Dr. Russel Carlson One of the many questions that
More informationSPD Pressure Sensor Families
DATASHEET SPD Pressure Sensor Families 1/7 Introduction to the principals of Smart Pressure Devices 1 Basic principles Pressure sensors are based on the principle of bending a membrane caused by the pressure
More informationProbabilistic Modelling in Multi-Competitor Games. Christopher J. Farmer
Probabilistic Modelling in Multi-Competitor Games Christopher J. Farmer Master of Science Division of Informatics University of Edinburgh 2003 Abstract Previous studies in to predictive modelling of sports
More informationThe Reliability of Intrinsic Batted Ball Statistics Appendix
The Reliability of ntrinsic Batted Ball Statistics Appendix Glenn Healey, EECS Department University of California, rvine, CA 92617 Given information about batted balls for a set of players, we review
More informationTwo Machine Learning Approaches to Understand the NBA Data
Two Machine Learning Approaches to Understand the NBA Data Panagiotis Lolas December 14, 2017 1 Introduction In this project, I consider applications of machine learning in the analysis of nba data. To
More informationMath 4. Unit 1: Conic Sections Lesson 1.1: What Is a Conic Section?
Unit 1: Conic Sections Lesson 1.1: What Is a Conic Section? 1.1.1: Study - What is a Conic Section? Duration: 50 min 1.1.2: Quiz - What is a Conic Section? Duration: 25 min / 18 Lesson 1.2: Geometry of
More informationBézier Curves and Splines
CS-C3100 Computer Graphics Bézier Curves and Splines Majority of slides from Frédo Durand vectorportal.com CS-C3100 Fall 2016 Lehtinen Before We Begin Anything on your mind concerning Assignment 1? Linux
More informationAlgebra I: A Fresh Approach. By Christy Walters
Algebra I: A Fresh Approach By Christy Walters 2016 A+ Education Services All rights reserved. No part of this publication may be reproduced, distributed, stored in a retrieval system, or transmitted,
More informationA Class of Regression Estimator with Cum-Dual Ratio Estimator as Intercept
International Journal of Probability and Statistics 015, 4(): 4-50 DOI: 10.593/j.ijps.015040.0 A Class of Regression Estimator with Cum-Dual Ratio Estimator as Intercept F. B. Adebola 1, N. A. Adegoke
More informationOPTIMIZATION OF A WAVE CANCELLATION MULTIHULL SHIP USING CFD TOOLS
OPTIMIZATION OF A AVE CANCELLATION MULTIHULL SHIP USING CFD TOOLS C. Yang, R. Löhner and O. Soto School of Computational Sciences, George Mason University Fairfax VA 030-4444, USA ABSTRACT A simple CFD
More informationBiomechanics and Models of Locomotion
Physics-Based Models for People Tracking: Biomechanics and Models of Locomotion Marcus Brubaker 1 Leonid Sigal 1,2 David J Fleet 1 1 University of Toronto 2 Disney Research, Pittsburgh Biomechanics Biomechanics
More informationMassey Method. Introduction. The Process
Massey Method Introduction Massey s Method, also referred to as the Point Spread Method, is a rating method created by mathematics professor Kenneth Massey. It is currently used to determine which teams
More informationSimulating Major League Baseball Games
ABSTRACT Paper 2875-2018 Simulating Major League Baseball Games Justin Long, Slippery Rock University; Brad Schweitzer, Slippery Rock University; Christy Crute Ph.D, Slippery Rock University The game of
More information1. A tendency to roll or heel when turning (a known and typically constant disturbance) 2. Motion induced by surface waves of certain frequencies.
Department of Mechanical Engineering Massachusetts Institute of Technology 2.14 Analysis and Design of Feedback Control Systems Fall 2004 October 21, 2004 Case Study on Ship Roll Control Problem Statement:
More informationEstimating Paratransit Demand Forecasting Models Using ACS Disability and Income Data
Estimating Paratransit Demand Forecasting Models Using ACS Disability and Income Data Presenter: Daniel Rodríguez Román University of Puerto Rico, Mayagüez Co-author: Sarah V. Hernandez University of Arkansas,
More informationHow Many Iterations Are Enough?
T ECOLOTE R ESEARCH, I NC. Bridging Engineering and Economics Since 1973 How Many Are Enough? Alfred Smith 2008 Joint SCEA/ISPA Annual Conference and Training Workshop Pacific Palms Conference Resort,
More informationBayesian Optimized Random Forest for Movement Classification with Smartphones
Bayesian Optimized Random Forest for Movement Classification with Smartphones 1 2 3 4 Anonymous Author(s) Affiliation Address email 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
More informationLesson 14: Modeling Relationships with a Line
Exploratory Activity: Line of Best Fit Revisited 1. Use the link http://illuminations.nctm.org/activity.aspx?id=4186 to explore how the line of best fit changes depending on your data set. A. Enter any
More informationNUMB3RS Activity: Walkabout. Episode: Mind Games
Teacher Page 1 NUMB3RS Activity: Walkabout Topic: Random Walks & Grade Level: 8-12 Objective: Learn about random walks and use them to compute an approximation to. Materials: TI-83/84 Plus calculator or
More informationClassroom Tips and Techniques: The Partial-Fraction Decomposition. Robert J. Lopez Emeritus Professor of Mathematics and Maple Fellow Maplesoft
Classroom Tips and Techniques: The Partial-Fraction Decomposition Robert J. Lopez Emeritus Professor of Mathematics and Maple Fellow Maplesoft Introduction Students typically meet the algebraic technique
More informationCourse 495: Advanced Statistical Machine Learning/Pattern Recognition
Course 495: Advanced Statistical Machine Learning/Pattern Recognition Lectures: Stefanos Zafeiriou Goal (Lectures): To present modern statistical machine learning/pattern recognition algorithms. The course
More informationBesides the reported poor performance of the candidates there were a number of mistakes observed on the assessment tool itself outlined as follows:
MATHEMATICS (309/1) REPORT The 2013 Mathematics (309/1) paper was of average standard. The paper covered a wide range of the syllabus. It was neither gender bias nor culture bias. It did not have language
More informationknn & Naïve Bayes Hongning Wang
knn & Naïve Bayes Hongning Wang CS@UVa Today s lecture Instance-based classifiers k nearest neighbors Non-parametric learning algorithm Model-based classifiers Naïve Bayes classifier A generative model
More informationQuantitative Methods for Economics Tutorial 6. Katherine Eyal
Quantitative Methods for Economics Tutorial 6 Katherine Eyal TUTORIAL 6 13 September 2010 ECO3021S Part A: Problems 1. (a) In 1857, the German statistician Ernst Engel formulated his famous law: Households
More informationSTAT/MATH 395 PROBABILITY II
STAT/MATH 395 PROBABILITY II Quick review on Discrete Random Variables Néhémy Lim University of Washington Winter 2017 Example Pick 5 toppings from a total of 15. Give the sample space Ω of the experiment
More informationAnabela Brandão and Doug S. Butterworth
Obtaining a standardised CPUE series for toothfish (Dissostichus eleginoides) in the Prince Edward Islands EEZ calibrated to incorporate both longline and trotline data over the period 1997-2013 Anabela
More informationEE 364B: Wind Farm Layout Optimization via Sequential Convex Programming
EE 364B: Wind Farm Layout Optimization via Sequential Convex Programming Jinkyoo Park 1 Introduction In a wind farm, the wakes formed by upstream wind turbines decrease the power outputs of downstream
More informationTitle: 4-Way-Stop Wait-Time Prediction Group members (1): David Held
Title: 4-Way-Stop Wait-Time Prediction Group members (1): David Held As part of my research in Sebastian Thrun's autonomous driving team, my goal is to predict the wait-time for a car at a 4-way intersection.
More informationPGA Tour Scores as a Gaussian Random Variable
PGA Tour Scores as a Gaussian Random Variable Robert D. Grober Departments of Applied Physics and Physics Yale University, New Haven, CT 06520 Abstract In this paper it is demonstrated that the scoring
More informationRiverboat Simulator Activity
Riverboat Simulator Activity Purpose: The purpose of this activity is to analyze the relationship between the two vector components of motion for a river boat as it travels across a river in the presence
More informationCombining Experimental and Non-Experimental Design in Causal Inference
Combining Experimental and Non-Experimental Design in Causal Inference Kari Lock Morgan Department of Statistics Penn State University Rao Prize Conference May 12 th, 2017 A Tribute to Don Design trumps
More informationMarch Madness A Math Primer
University of Kentucky Feb 24, 2016 NCAA Tournament Background Single elimination tournament for Men s and Women s NCAA Div. 1 basketball held each March, referred to as March Madness. 64 teams make tournament
More informationAnalysis of Variance. Copyright 2014 Pearson Education, Inc.
Analysis of Variance 12-1 Learning Outcomes Outcome 1. Understand the basic logic of analysis of variance. Outcome 2. Perform a hypothesis test for a single-factor design using analysis of variance manually
More informationSupport Vector Machines: Optimization of Decision Making. Christopher Katinas March 10, 2016
Support Vector Machines: Optimization of Decision Making Christopher Katinas March 10, 2016 Overview Background of Support Vector Machines Segregation Functions/Problem Statement Methodology Training/Testing
More informationHow To Win The Ashes A Statistician s Guide
How To Win The Ashes A Statistician s Guide Kevin Wang Sydney Universtiy Mathematics Society Last Modified: August 21, 2015 Motivation How come SUMS doesn t do statistics talk very often? Motivation How
More informationMATH 118 Chapter 5 Sample Exam By: Maan Omran
MATH 118 Chapter 5 Sample Exam By: Maan Omran Problem 1-4 refer to the following table: X P Product a 0.2 d 0 0.1 e 1 b 0.4 2 c? 5 0.2? E(X) = 1.7 1. The value of a in the above table is [A] 0.1 [B] 0.2
More informationQuantitative Literacy: Thinking Between the Lines
Quantitative Literacy: Thinking Between the Lines Crauder, Noell, Evans, Johnson Chapter 6: Statistics 2013 W. H. Freeman and Company 1 Chapter 6: Statistics Lesson Plan Data summary and presentation:
More informationNovel empirical correlations for estimation of bubble point pressure, saturated viscosity and gas solubility of crude oils
86 Pet.Sci.(29)6:86-9 DOI 1.17/s12182-9-16-x Novel empirical correlations for estimation of bubble point pressure, saturated viscosity and gas solubility of crude oils Ehsan Khamehchi 1, Fariborz Rashidi
More informationBridge Decomposition of Restriction Measures Tom Alberts joint work with Hugo Duminil (ENS) with lots of help from Wendelin Werner University of
Bridge Decomposition of Restriction Measures Tom Alberts joint work with Hugo Duminil (ENS) with lots of help from Wendelin Werner University of Toronto Self-Avoiding Walk Bridges Self-Avoiding Walk Bridges
More informationy ) s x x )(y i (x i r = 1 n 1 s y Statistics Lecture 7 Exploring Data , y 2 ,y n (x 1 ),,(x n ),(x 2 ,y 1 How two variables vary together
Statistics 111 - Lecture 7 Exploring Data Numerical Summaries for Relationships between Variables Administrative Notes Homework 1 due in recitation: Friday, Feb. 5 Homework 2 now posted on course website:
More information3D Inversion in GM-SYS 3D Modelling
3D Inversion in GM-SYS 3D Modelling GM-SYS 3D provides a wide range of inversion options including inversion for both layer structure and physical properties for gravity and magnetic data. There is an
More informationOptimizing Cyclist Parking in a Closed System
Optimizing Cyclist Parking in a Closed System Letu Qingge, Killian Smith Gianforte School of Computing, Montana State University, Bozeman, MT 59717, USA Abstract. In this paper, we consider the two different
More informationLecture 16: Chapter 7, Section 2 Binomial Random Variables
Lecture 16: Chapter 7, Section 2 Binomial Random Variables!Definition!What if Events are Dependent?!Center, Spread, Shape of Counts, Proportions!Normal Approximation Cengage Learning Elementary Statistics:
More informationStudent Outcomes. Lesson Notes. Classwork. Discussion (20 minutes)
Student Outcomes Students explain a proof of the converse of the Pythagorean Theorem. Students apply the theorem and its converse to solve problems. Lesson Notes Students had their first experience with
More informationQueue analysis for the toll station of the Öresund fixed link. Pontus Matstoms *
Queue analysis for the toll station of the Öresund fixed link Pontus Matstoms * Abstract A new simulation model for queue and capacity analysis of a toll station is presented. The model and its software
More informationExcel Solver Case: Beach Town Lifeguard Scheduling
130 Gebauer/Matthews: MIS 213 Hands-on Tutorials and Cases, Spring 2015 Excel Solver Case: Beach Town Lifeguard Scheduling Purpose: Optimization under constraints. A. GETTING STARTED All Excel projects
More informationBy Lawrence D. Brown University of Pennsylvania
The Annals of Applied Statistics 2008, Vol. 2, No. 1, 113 152 DOI: 10.1214/07-AOAS138 c Institute of Mathematical Statistics, 2008 IN-SEASON PREDICTION OF BATTING AVERAGES: A FIELD TEST OF EMPIRICAL BAYES
More information