Introduction to Detection Theory


 Augustus Walters
 1 years ago
 Views:
Transcription
1 Introduction to Detection Theory Reading: Ch. 3 in KayII. Notes by Prof. Don Johnson on detection theory, see Ch. 10 in Wasserman. EE 527, Detection and Estimation Theory, # 5 1
2 Introduction to Detection Theory (cont.) We wish to make a decision on a signal of interest using noisy measurements. Statistical tools enable systematic solutions and optimal design. Application areas include: Communications, Radar and sonar, Nondestructive evaluation (NDE) of materials, Biomedicine, etc. EE 527, Detection and Estimation Theory, # 5 2
3 Example: Radar Detection. We wish to decide on the presence or absence of a target. EE 527, Detection and Estimation Theory, # 5 3
4 Introduction to Detection Theory We assume a parametric measurement model p(x θ) [or p(x ; θ), which is the notation that we sometimes use in the classical setting]. In point estimation theory, we estimated the parameter θ Θ given the data x. Suppose now that we choose Θ 0 and Θ 1 that form a partition of the parameter space Θ: Θ 0 Θ 1 = Θ, Θ 0 Θ 1 =. In detection theory, we wish to identify which hypothesis is true (i.e. make the appropriate decision): H 0 : θ Θ 0, null hypothesis H 1 : θ Θ 1, alternative hypothesis. Terminology: If θ can only take two values, Θ = {θ 0, θ 1 }, Θ 0 = {θ 0 }, Θ 1 = {θ 1 } we say that the hypotheses are simple. Otherwise, we say that they are composite. Composite Hypothesis Example: H 0 : θ = 0 versus H 1 : θ (0, ). EE 527, Detection and Estimation Theory, # 5 4
5 The Decision Rule We wish to design a decision rule (function) φ(x) : X (0, 1): φ(x) = { 1, decide H1, 0, decide H 0. which partitions the data space X [i.e. the support of p(x θ)] into two regions: Rule φ(x): X 0 = {x : φ(x) = 0}, X 1 = {x : φ(x) = 1}. Let us define probabilities of false alarm and miss: P FA = E x θ [φ(x) θ] = p(x θ) dx for θ in Θ 0 X 1 P M = E x θ [1 φ(x) θ] = 1 p(x θ) dx X 1 = p(x θ) dx for θ in Θ 1. X 0 Then, the probability of detection (correctly deciding H 1 ) is P D = 1 P M = E x θ [φ(x) θ] = X 1 p(x θ) dx for θ in Θ 1. EE 527, Detection and Estimation Theory, # 5 5
6 Note: P FA and P D /P M are generally functions of the parameter θ (where θ Θ 0 when computing P FA and θ Θ 1 when computing P D /P M ). More Terminology. Statisticians use the following terminology: False alarm Type I error Miss Type II error Probability of detection Power Probability of false alarm Significance level. EE 527, Detection and Estimation Theory, # 5 6
7 Bayesian Decisiontheoretic Detection Theory Recall (a slightly generalized version of) the posterior expected loss: ρ(action x) = L(θ, action) p(θ x) dθ Θ that we introduced in handout # 4 when we discussed Bayesian decision theory. Let us now apply this theory to our easy example discussed here: hypothesis testing, where our action space consists of only two choices. We first assign a loss table: decision rule φ true state Θ 1 Θ 0 x X 1 L(1 1) = 0 L(1 0) x X 0 L(0 1) L(0 0) = 0 with the loss function described by the quantities L(declared true): L(1 0) quantifies loss due to a false alarm, L(0 1) quantifies loss due to a miss, L(1 1) and L(0 0) (losses due to correct decisions) typically set to zero in real life. Here, we adopt zero losses for correct decisions. EE 527, Detection and Estimation Theory, # 5 7
8 Now, our posterior expected loss takes two values: ρ 0 (x) = L(0 1) p(θ x) dθ Θ 1 + L(0 0) p(θ x) dθ } {{ } Θ 0 0 = L(0 1) p(θ x) dθ Θ 1 L(0 1) is constant = L(0 1) p(θ x) dθ Θ } 1 {{ } P [θ Θ 1 x] and, similarly, ρ 1 (x) = L(1 0) p(θ x) dθ Θ 0 L(1 0) is constant = L(1 0) p(θ x) dθ. Θ } 0 {{ } P [θ Θ 0 x] We define the Bayes decision rule as the rule that minimizes the posterior expected loss; this rule corresponds to choosing the dataspace partitioning as follows: X 1 = {x : ρ 1 (x) ρ 0 (x)} EE 527, Detection and Estimation Theory, # 5 8
9 or P [θ Θ 1 x] { }} { { p(θ x) dθ } Θ X 1 = x : 1 L(1 0) L(0 1) p(θ x) dθ Θ } 0 {{ } P [θ Θ 0 x] or, equivalently, upon applying the Bayes rule: X 1 = { x : Θ 1 p(x θ) π(θ) dθ Θ 0 p(x θ) π(θ) dθ 01 loss: For L(1 0) = L(0 1) = 1, we have (1) } L(1 0) L(0 1). (2) decision rule φ true state Θ 1 Θ 0 x X 1 L(1 1) = 0 L(1 0) = 1 x X 0 L(0 1) = 1 L(0 0) = 0 yielding the following Bayes decision rule, called the maximum a posteriori (MAP) rule: X 1 = { x : P [θ Θ 1 x] P [θ Θ 0 x] 1 } or, equivalently, upon applying the Bayes rule: { } Θ X 1 = x : 1 p(x θ) π(θ) dθ Θ 0 p(x θ) π(θ) dθ 1. (4) (3) EE 527, Detection and Estimation Theory, # 5 9
10 Simple hypotheses. Let us specialize (1) to the case of simple hypotheses (Θ 0 = {θ 0 }, Θ 1 = {θ 1 }): X 1 = { x : p(θ 1 x) p(θ 0 x) } {{ } posteriorodds ratio } L(1 0) L(0 1). (5) We can rewrite (5) using the Bayes rule: X 1 = { x : p(x θ 1 ) p(x θ 0 ) } {{ } likelihood ratio } π 0 L(1 0) π 1 L(0 1) (6) where π 0 = π(θ 0 ), π 1 = π(θ 1 ) = 1 π 0 describe the prior probability mass function (pmf) of the binary random variable θ (recall that θ {θ 0, θ 1 }). Hence, for binary simple hypotheses, the prior pmf of θ is the Bernoulli pmf. EE 527, Detection and Estimation Theory, # 5 10
11 Preposterior (Bayes) Risk The preposterior (Bayes) risk for rule φ(x) is E x,θ [loss] = X 1 + X 0 L(1 0) p(x θ)π(θ) dθ dx Θ 0 Θ 1 L(0 1) p(x θ)π(θ) dθ dx. How do we choose the rule φ(x) that minimizes the preposterior EE 527, Detection and Estimation Theory, # 5 11
12 risk? L(1 0) p(x θ)π(θ) dθ dx X 1 Θ 0 + L(0 1) p(x θ)π(θ) dθ dx X 0 Θ 1 = L(1 0) p(x θ)π(θ) dθ dx X 1 Θ 0 L(0 1) p(x θ)π(θ) dθ dx X 1 Θ 1 + L(0 1) p(x θ)π(θ) dθ dx X 0 Θ 1 + L(0 1) p(x θ)π(θ) dθ dx X 1 Θ 1 = const } {{ } not dependent on φ(x) { + L(1 0) p(x θ)π(θ) dθ X 1 Θ 0 } L(0 1) p(x θ)π(θ) dθ dx Θ 1 implying that X 1 should be chosen as { } X 1 : L(1 0) p(x θ)π(θ) dθ L(0 1) p(x θ)π(θ) < 0 Θ 0 Θ 1 EE 527, Detection and Estimation Theory, # 5 12
13 which, as expected, is the same as (2), since minimizing the posterior expected loss minimizing the preposterior risk for every x as showed earlier in handout # loss: For the 01 loss, i.e. L(1 0) = L(0 1) = 1, the preposterior (Bayes) risk for rule φ(x) is E x,θ [loss] = X 1 + X 0 p(x θ)π(θ) dθ dx Θ 0 Θ 1 p(x θ)π(θ) dθ dx (7) which is simply the average error probability, with averaging performed over the joint probability density or mass function (pdf/pmf) or the data x and parameters θ. EE 527, Detection and Estimation Theory, # 5 13
14 Bayesian Decisiontheoretic Detection for Simple Hypotheses The Bayes decision rule for simple hypotheses is (6): Λ(x) } {{ } likelihood ratio = p(x θ 1) p(x θ 0 ) H 1 π 0 L(1 0) π 1 L(0 1) τ (8) see also Ch. 3.7 in KayII. (Recall that Λ(x) is the sufficient statistic for the detection problem, see p. 37 in handout # 1.) Equivalently, log Λ(x) = log[p(x θ 1 )] log[p(x θ 0 )] H 1 log τ τ. Minimum Average Error Probability Detection: In the familiar 01 loss case where L(1 0) = L(0 1) = 1, we know that the preposterior (Bayes) risk is equal to the average error probability, see (7). This average error probability greatly EE 527, Detection and Estimation Theory, # 5 14
15 simplifies in the simple hypothesis testing case: av. error probability = L(1 0) p(x θ } {{ } 0 )π 0 dx X L(0 1) p(x θ } {{ } 1 )π 1 dx X 0 1 = π 0 p(x θ 0 ) dx +π 1 p(x θ 1 ) dx X } 1 X {{ } } 0 {{ } P FA P M where, as before, the averaging is performed over the joint pdf/pmf of the data x and parameters θ, and π 0 = π(θ 0 ), π 1 = π(θ 1 ) = 1 π 0 (the Bernoulli pmf). In this case, our Bayes decision rule simplifies to the MAP rule (as expected, see (5) and Ch. 3.6 in KayII): p(θ 1 x) p(θ 0 x) } {{ } posteriorodds ratio or, equivalently, upon applying the Bayes rule: p(x θ 1 ) p(x θ 0 ) } {{ } likelihood ratio H 1 1 (9) H 1 π 0 π 1. (10) EE 527, Detection and Estimation Theory, # 5 15
16 which is the same as (4), upon substituting the Bernoulli pmf as the prior pmf for θ and (8), upon substituting L(1 0) = L(0 1) = 1. EE 527, Detection and Estimation Theory, # 5 16
17 Bayesian Decisiontheoretic Detection Theory: Handling Nuisance Parameters We apply the same approach as before integrate the nuisance parameters (ϕ, say) out! Therefore, (1) still holds for testing H 0 : θ Θ 0 versus H 1 : θ Θ 1 but p θ x (θ x) is computed as follows: p θ x (θ x) = p θ,ϕ x (θ, ϕ x) dϕ and, therefore, Θ 1 Θ 0 p(θ x) { }} { p θ,ϕ x (θ, ϕ x) dϕ dθ p θ,ϕ x (θ, ϕ x) dϕ } {{ } p(θ x) dθ H 1 L(1 0) L(0 1) (11) EE 527, Detection and Estimation Theory, # 5 17
18 or, equivalently, upon applying the Bayes rule: Θ 1 px θ,ϕ (x θ, ϕ) π θ,ϕ (θ, ϕ) dϕ dθ Θ 0 px θ,ϕ (x θ, ϕ) π θ,ϕ (θ, ϕ) dϕ dθ H 1 L(1 0) L(0 1). (12) Now, if θ and ϕ are independent a priori, i.e. π θ,ϕ (θ, ϕ) = π θ (θ) π ϕ (ϕ) (13) then (12) can be rewritten as Θ 1 π θ (θ) Θ 0 π θ (θ) p(x θ) { }} { p x θ,ϕ (x θ, ϕ) π ϕ (ϕ) dϕ dθ p x θ,ϕ (x θ, ϕ) π ϕ (ϕ) } {{ } p(x θ) dϕ dθ H 1 L(1 0) L(0 1). (14) Simple hypotheses and independent priors for θ and ϕ: Let us specialize (11) to the simple hypotheses (Θ 0 = {θ 0 }, Θ 1 = {θ 1 }): pθ,ϕ x (θ 1, ϕ x) dϕ pθ,ϕ x (θ 0, ϕ x) dϕ H 1 L(1 0) L(0 1). (15) Now, if θ and ϕ are independent a priori, i.e. (13) holds, then EE 527, Detection and Estimation Theory, # 5 18
19 we can rewrite (14) [or (15) using the Bayes rule]: px θ,ϕ (x θ 1, ϕ) π ϕ (ϕ) dϕ px θ,ϕ (x θ 0, ϕ) π ϕ (ϕ) dϕ } {{ } integrated likelihood ratio = same as (6) { }} { p(x θ 1 ) H 1 π 0 L(1 0) p(x θ 0 ) π 1 L(0 1) (16) where π 0 = π θ (θ 0 ), π 1 = π θ (θ 1 ) = 1 π 0. EE 527, Detection and Estimation Theory, # 5 19
20 Chernoff Bound on Average Error Probability for Simple Hypotheses Recall that minimizing the average error probability av. error probability = X 1 + X 0 p(x θ)π(θ) dθ dx Θ 0 Θ 1 p(x θ)π(θ) dθ dx leads to the MAP decision rule: X 1 = { x : Θ 0 p(x θ) π(θ) dθ } p(x θ) π(θ) dθ < 0. Θ 1 In many applications, we may not be able to obtain a simple closedform expression for the minimum average error EE 527, Detection and Estimation Theory, # 5 20
21 probability, but we can bound it as follows: min av. error probability = X1 p(x θ)π(θ) dθ dx Θ 0 + p(x θ)π(θ) dθ dx Θ 1 X 0 using the def. of X 1 = X }{{} data space { } min p(x θ)π(θ) dθ, p(x θ)π(θ) dθ) dx Θ 0 Θ 1 = X X [ = q 0 (x) = q { }} { 1 (x) ] λ [ { }} { ] 1 λ p(x θ)π(θ) dθ p(x θ)π(θ) dθ dx Θ 0 Θ 1 [q 0 (x)] λ [q 1 (x)] 1 λ dx which is the Chernoff bound on the minimum average error probability. Here, we have used the fact that min{a, b} a λ b 1 λ, for 0 λ 1, a, b 0. When x = x 1 x 2. x N EE 527, Detection and Estimation Theory, # 5 21
22 with x 1, x 2,... x N conditionally independent, identically distributed (i.i.d.) given θ and simple hypotheses (i.e. Θ 0 = {θ 0 }, Θ 1 = {θ 1 }) then N q 0 (x) = p(x θ 0 ) π 0 = π 0 p(x n θ 0 ) yielding n=1 N q 1 (x) = p(x θ 1 ) π 1 = π 1 p(x n θ 1 ) n=1 Chernoff bound for N conditionally i.i.d. measurements (given θ) and simple hyp. [ N ] λ N 1 λ = π 0 p(x n θ 0 ) [π 1 p(x n θ 1 )] dx n=1 = π λ 0 π 1 λ 1 = π λ 0 π 1 λ 1 N n=1 n=1 { [p(x n θ 0 )] λ [p(x n θ 1 )] 1 λ dx n } { [p(x 1 θ 0 )] λ [p(x 1 θ 1 )] 1 λ dx 1 } N EE 527, Detection and Estimation Theory, # 5 22
23 or, in other words, 1 N log(min av. error probability) log(πλ 0 π1 1 λ ) + log [p(x 1 θ 0 )] λ [p(x 1 θ 1 )] 1 λ dx 1, λ [0, 1]. If π 0 = π 1 = 1/2 (which is almost always the case of interest when evaluating average error probabilities), we can say that, as N, min av. error probability for N cond. i.i.d. measurements (given θ) and simple hypotheses ( { f(n) exp N min log [p(x 1 θ 0 )] λ [p(x 1 θ 1 )] 1 λ} ) λ [0,1] } {{ } Chernoff information for a single observation where f(n) is a slowlyvarying function compared with the exponential term: lim N log f(n) N = 0. Note that the Chernoff information in the exponent term of the above expression quantifies the asymptotic behavior of the minimum average error probability. We now give a useful result, taken from EE 527, Detection and Estimation Theory, # 5 23
24 K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd ed., San Diego, CA: Academic Press, 1990 for evaluating a class of Chernoff bounds. Lemma 1. Consider p 1 (x) = N (µ 1, Σ 1 ) and p 2 (x) = N (µ 2, Σ 2 ). Then [p 1 (x)] λ [p 2 (x)] 1 λ dx = exp[ g(λ)] where g(λ) = λ (1 λ) 2 (µ 2 µ 1 ) T [λ Σ 1 + (1 λ) Σ 2 ] 1 (µ 2 µ 1 ) [ λ log Σ1 + (1 λ) Σ 2 ] Σ 1 λ Σ 2 1 λ. EE 527, Detection and Estimation Theory, # 5 24
25 Probabilities of False Alarm (P FA ) and Detection (P D ) for Simple Hypotheses 0.2 PDF under H 0 PDF under H 1 Probability Density Function P FA P D τ Test Statistic Comments: Λ(x) { }} { P FA = P [ test } statistic {{ > τ} θ = θ 0 ] X X 1 P D = P [test statistic > τ } {{ } X X 0 θ = θ 1 ]. (i) As the region X 1 shrinks (i.e. τ ), both of the above probabilities shrink towards zero. EE 527, Detection and Estimation Theory, # 5 25
26 (ii) As the region X 1 grows (i.e. τ 0), both of these probabilities grow towards unity. (iii) Observations (i) and (ii) do not imply equality between P FA and P D ; in most cases, as R 1 grows, P D grows more rapidly than P FA (i.e. we better be right more often than we are wrong). (iv) However, the perfect case where our rule is always right and never wrong (P D = 1 and P FA = 0) cannot occur when the conditional pdfs/pmfs p(x θ 0 ) and p(x θ 1 ) overlap. (v) Thus, to increase the detection probability P D, we must also allow for the falsealarm probability P FA to increase. This behavior represents the fundamental tradeoff in hypothesis testing and detection theory and motivates us to introduce a (classical) approach to testing simple hypotheses, pioneered by Neyman and Pearson (to be discussed next). EE 527, Detection and Estimation Theory, # 5 26
27 NeymanPearson Test for Simple Hypotheses Setup: Parametric data models p(x ; θ 0 ), p(x ; θ 1 ), Simple hypothesis testing: H 0 : θ = θ 0 versus H 1 : θ = θ 1. No prior pdf/pmf on θ is available. Goal: Design a test that maximizes the probability of detection P D = P [X X 1 ; θ = θ 0 ] (equivalently, minimizes the miss probability P M ) under the constraint P FA = P [X X 1 ; θ = θ 0 ] = α α. Here, we consider simple hypotheses; classical version of testing composite hypotheses is much more complicated. The Bayesian version of testing composite hypotheses is trivial (as EE 527, Detection and Estimation Theory, # 5 27
28 is everything else Bayesian, at least conceptually) and we have already seen it. Solution. We apply the Lagrangemultiplier approach: maximize L = P D + λ (P FA α ) [ ] = p(x ; θ 1 ) dx + λ p(x; θ 0 ) dx α X 1 X 1 = [p(x ; θ 1 ) λ p(x ; θ 0 )] dx λ α. X 1 To maximize L, set X 1 = {x : p(x ; θ 1 ) λ p(x ; θ 0 ) > 0} = { x : p(x ; θ 1) p(x ; θ 0 ) > λ }. Again, we find the likelihood ratio: Recall our constraint: Λ(x) = p(x ; θ 1) p(x ; θ 0 ). X 1 p(x ; θ 0 ) dx = P FA = α α. EE 527, Detection and Estimation Theory, # 5 28
29 If we increase λ, P FA and P D go down. Similarly, if we decrease λ, P FA and P D go up. Hence, to maximize P D, choose λ so that P FA is as big as possible under the constraint. Two useful ways for determining the threshold that achieves a specified falsealarm rate: Find λ that satisfies or, x : Λ(x)>λ p(x ; θ 0 ) dx = P FA = α expressing in terms of the pdf/pmf of Λ(x) under H 0 : λ p Λ;θ0 (l ; θ 0 ) dl = α. or, perhaps, in terms of a monotonic function of Λ(x), say T (x) = monotonic function(λ(x)). Warning: We have been implicitly assuming that P FA is a continuous function of λ. Some (not insightful) technical adjustments are needed if this is not the case. A way of handling nuisance parameters: We can utilize the integrated (marginal) likelihood ratio (16) under the Neyman Pearson setup as well. EE 527, Detection and Estimation Theory, # 5 29
30 ChernoffStein Lemma for Bounding the Miss Probability in NeymanPearson Tests of Simple Hypotheses Recall the definition of the KullbackLeibler (KL) distance D(p q) from one pmf (p) to another (p): D(p q) = k p k log p k q k. The complete proof of this lemma for the discrete (pmf) case is given in Additional Reading: T.M. Cover and J.A. Thomas, Elements of Information Theory. Second ed., New York: Wiley, EE 527, Detection and Estimation Theory, # 5 30
31 Setup for the ChernoffStein Lemma Assume that x 1, x 2,..., x N are conditionally i.i.d. given θ. We adopt the NeymanPearson framework, i.e. obtain a decision threshold to achieve a fixed P FA. Let us study the asymptotic P M = 1 P D as the number of observations N gets large. To keep P FA constant as N increases, we need to make our decision threshold (γ, say) vary with N, i.e. Now, the miss probability is γ = γ N (P FA ) ) P M = P M (γ) = P M (γ N (P FA ). EE 527, Detection and Estimation Theory, # 5 31
32 ChernoffStein Lemma The ChernoffStein lemma says: lim P FA 0 lim N 1 N log P M = ( ) D p(x n θ 0 ) p(x n θ 1 ) } {{ } KL distance for a single observation where the KL distance between p(x n θ 0 ) and p(x n θ 1 ) ( ) D p(x n θ 0 ) p(x n θ 1 ) discrete (pmf) case = x n p(x n θ 0 ) log [ = E p(xn θ 0 ) log p(x n θ 1 ) ] p(x n θ 0 ) [ p(xn θ 1 ) ] p(x n θ 0 ) does not depend on the observation index n, since x n conditionally i.i.d. given θ. are Equivalently, we can state that [ N + P M f(n) exp N D ( p(x n θ 0 ) p(x n θ 1 ) )] as P FA 0 and N, where f(n) is a slowlyvarying function compared with the exponential term (when P FA 0 and N ). EE 527, Detection and Estimation Theory, # 5 32
33 Detection for Simple Hypotheses: Example Known positive DC level in additive white Gaussian noise (AWGN), Example 3.2 in KayII. Consider where H 0 : x[n] = w[n], n = 1, 2,..., N versus H 1 : x[n] = A + w[n], n = 1, 2,..., N A > 0 is a known constant, w[n] is zeromean white Gaussian noise with known variance σ 2, i.e. w[n] N (0, σ 2 ). The above hypothesistesting formulation is EElike: noise versus signal plus noise. It is similar to the onoff keying scheme in communications, which gives us an idea to rephrase it so that it fits our formulation on p. 4 (for which we have developed all the theory so far). Here is such an EE 527, Detection and Estimation Theory, # 5 33
34 alternative formulation: consider a family of pdfs p(x ; a) = 1 [ (2πσ2 ) exp N 1 2 σ 2 N (x[n] a) 2] (17) n=1 and the following (equivalent) hypotheses: H 0 : a = 0 (off) versus H 1 : a = A (on). Then, the likelihood ratio is Λ(x) = p(x ; a = A) p(x ; a = 0) = 1/(2πσ2 ) N/2 exp[ 1 N 2σ 2 n=1 (x[n] A)2 ] 1/(2πσ 2 ) N/2 exp( 1 N. 2σ 2 n=1 x[n]2 ) Now, take the logarithm and, after simple manipulations, reduce our likelihoodratio test to comparing T (x) = x = 1 N N x[n] n=1 with a threshold γ. [Here, T (x) is a monotonic function of Λ(x).] If T (x) > γ accept H 1 (i.e. reject H 0 ), otherwise accept EE 527, Detection and Estimation Theory, # 5 34
35 H 0 (well, not exactly, we will talk more about this decision on p. 59). The choice of γ depends on the approach that we take. For the Bayes decision rule, γ is a function of π 0 and π 1. For the NeymanPearson test, γ is chosen to achieve (control) a desired P FA. Bayesian decisiontheoretic detection for 01 loss (corresponding to minimizing the average error EE 527, Detection and Estimation Theory, # 5 35
36 probability): log Λ(x) = 1 2 σ σ 2 2 A ( N n=1 N (x[n] A) σ 2 n=1 N (x[n] x[n] } {{ } 0 n=1 ( N n=1 ) x[n] N n=1 (x[n]) 2 H ( 1 π0 ) log π 1 +A) (x[n] + x[n] A) H ( 1 π0 ) log π 1 ) x[n] A 2 N H ( 1 2 σ 2 π0 ) log π 1 AN 2 H 1 σ2 A log ( π0 π 1 ) σ2 since A > 0 and, finally, x H 1 N A log(π 0/π 1 ) + A } {{ 2} γ which, for the practically most interesting case of equiprobable hypotheses π 0 = π 1 = 1 2 (18) simplifies to x H 1 A 2 }{{} γ known as the maximumlikelihood test (i.e. the Bayes decision rule for 01 loss and a priori equiprobable hypotheses is defined as the maximumlikelihood test). This maximumlikelihood detector does not require the knowledge of the noise variance EE 527, Detection and Estimation Theory, # 5 36
37 σ 2 to declare its decision. However, the knowledge of σ 2 is key to assessing the detection performance. Interestingly, these observations will carry over to a few maximumlikelihood tests that we will derive in the future. Assuming (18), we now derive the minimum average error probability. First, note that X a = 0 N (0, σ 2 /N) and X a = A N (A, σ 2 /N). Then min av. error prob. = 1 2 P [X > A } {{ 2 a = 0] } +1 P FA = 1 2 P [ P [ X σ2 /N } {{ } standard normal random var. X A σ2 /N } {{ } standard normal random var. = 1 2 ( N A Q ) } {{ 4 σ 2 } standard normal complementary cdf > A/2 σ2 /N ; a = 0 ] < A/2 A σ2 /N ; a = A ] ( N A Φ 2) } {{ 4 σ 2 } standard normal cdf 2 P [X < A } {{ 2 a = A] } P M ( = Q 1 2 N A 2 σ 2 ) EE 527, Detection and Estimation Theory, # 5 37
38 since Φ( x) = Q(x). The minimum probability of error decreases monotonically with N A 2 /σ 2, which is known as the deflection coefficient. In this case, the Chernoff information is equal to A2 (see Lemma 1): 8 σ 2 { min log λ [0,1] [ λ (1 λ) = min λ [0,1] 2 } [N (0, σ 2 /N)] } {{ } λ [N (A, σ 2 /N)] } {{ } 1 λ dx p 0 (x) p 1 (x) A2 σ 2 ] = A2 8 σ 2 and, therefore, we expect the following asymptotic behavior: min av. error probability N + ) f(n) exp ( N A2 8 σ 2 (19) where f(n) varies slowly with N when N is large, compared with the exponential term: lim N log f(n) N = 0. Indeed, in our example, ( N A Q 2 ) N + 4 σ 2 1 N A 2 2π 4 σ } 2 {{ } f(n) ( exp N ) A2 8 σ 2 EE 527, Detection and Estimation Theory, # 5 38
39 where we have used the asymptotic formula Q(x) x + 1 x 2π exp( x2 /2) (20) given e.g. in equation (2.4) of KayII. EE 527, Detection and Estimation Theory, # 5 39
40 Known DC Level in AWGN: NeymanPearson Approach We now derive the NeymanPearson test for detecting a known DC level in AWGN. Recall that our likelihoodratio test compares with a threshold γ. T (x) = x = 1 N N 1 n=0 x[n] If T (x) > γ, decide H 1 (i.e. reject H 0 ), otherwise decide H 0 (see also the discussion on p. 59). Performance evaluation: Assuming (17), we have T (x) a N (a, σ 2 /N). Therefore, T (X) a = 0 N (0, σ 2 /N), implying ( γ ) P FA = P [T (X) > γ ; a = 0] = Q σ2 /N EE 527, Detection and Estimation Theory, # 5 40
41 and we obtain the decision threshold as follows: γ = σ 2 N Q 1 (P FA ). Now, T (X) a = A N (A, σ 2 /N), implying ( γ A ) P D = P (T (X) > γ a = A) = Q σ2 /N ( ) = Q Q 1 A (P FA ) 2 σ 2 /N ( ) = Q Q 1 (P FA ) NA2 /σ } {{ } 2. = SNR = d 2 Given the falsealarm probability P FA, the detection probability P D depends only on the deflection coefficient: d 2 = NA2 σ 2 = {E [T (X) a = A] E [T (X) a = 0]}2 var[t (X a = 0)] which is also (a reasonable definition for) the signaltonoise ratio (SNR). EE 527, Detection and Estimation Theory, # 5 41
42 Receiver Operating Characteristics (ROC) Comments: P D = Q ( Q 1 (P FA ) d ). As we raise the threshold γ, P FA goes down but so does P D. ROC should be above the 45 line otherwise we can do better by flipping a coin. Performance improves with d 2. EE 527, Detection and Estimation Theory, # 5 42
43 Typical Ways of Depicting the Detection Performance Under the NeymanPearson Setup To analyze the performance of a NeymanPearson detector, we examine two relationships: Between P D and P FA, for a given SNR, called the operating characteristics (ROC). receiver Between P D and SNR, for a given P FA. Here are examples of the two: see Figs. 3.8 and 3.5 in KayII, respectively. EE 527, Detection and Estimation Theory, # 5 43
44 Asymptotic (as N and P FA 0) P D and P M for a Known DC Level in AWGN We apply the ChernoffStein lemma, for which we need to compute the following KL distance: log where ( ) D p(x n a = 0) p(x n a = A) { = E p(xn a=0) log [ p(xn a = A) ]} p(x n a = 0) p(x n a = 0) = N (0, σ 2 ) [ p(xn a = A) ] = (x n A) 2 p(x n a = 0) 2 σ 2 + x2 n 2 σ 2 = 1 2σ 2 (2 A x n A 2 ). Therefore, ( ) D p(x n θ 0 ) p(x n θ 1 ) = A2 2 σ 2 and the ChernoffStein lemma predicts the following behavior of the detection probability as N and P FA 0: ( ) P D 1 f(n) exp N A2 } {{ 2 σ 2 } P M EE 527, Detection and Estimation Theory, # 5 44
45 where f(n) is a slowlyvarying function of N compared with the exponential term. In this case, the exact expression for P M (P D ) is available and consistent with the ChernoffStein lemma: P M = 1 Q (Q 1 (P FA ) ) NA 2 /σ 2 ) = Q( NA2 /σ 2 Q 1 (P FA ) N + = = 1 [ NA 2 /σ 2 Q 1 (P FA )] 2π { exp [ } NA 2 /σ 2 Q 1 (P FA )] 2 /2 1 [ NA 2 /σ 2 Q 1 (P FA )] 2π { exp NA 2 /(2σ 2 ) [Q 1 (P FA )] 2 /2 +Q 1 (P FA ) } NA 2 /σ 2 1 [ NA 2 /σ 2 Q 1 (P FA )] 2π { exp [Q 1 (P FA )] 2 /2 + Q 1 (P FA ) } NA 2 /σ 2 exp[ NA 2 /(2σ 2 )] } {{ } as predicted by the ChernoffStein lemma (21) where we have used the asymptotic formula (20). When P FA EE 527, Detection and Estimation Theory, # 5 45
46 is small and N is large, the first two (green) terms in the above expression make a slowlyvarying function f(n) of N. Note that the exponential term in (21) does not depend on the falsealarm probability P FA (or, equivalently, on the choice of the decision threshold). The ChernoffStein lemma asserts that this is not a coincidence. Comment. For detecting a known DC level in AWGN: The slope of the exponentialdecay term of P M is A 2 /(2σ 2 ) which is different from (larger than, in this case) the slope of the exponential decay of the minimum average error probability: A 2 /(8σ 2 ). EE 527, Detection and Estimation Theory, # 5 46
47 Decentralized NeymanPearson Detection for Simple Hypotheses Consider a sensornetwork scenario depicted by Assumption: Observations made at N spatially distributed sensors (nodes) follow the same marginal probabilistic model: H i : x n p(x n θ i ) where n = 1, 2,..., N and i {0, 1} for binary hypotheses. Each node n makes a hard local decision d n based on its local observation x n and sends it to the headquarters (fusion center), which collects all the local decisions and makes the final global EE 527, Detection and Estimation Theory, # 5 47
48 decision about H 0 or H 1. This structure is clearly suboptimal it is easy to construct a better decision strategy in which each node sends its (quantized, in practice) likelihood ratio to the fusion center, rather than the decision only. However, such a strategy would have a higher communication (energy) cost. We now go back to the decentralized detection problem. Suppose that each node n makes a local decision d n {0, 1}, n = 1, 2,..., N and transmits it to the fusion center. Then, the fusion center makes the global decision based on the likelihood ratio formed from the d n s. The simplest fusion scheme is based on the assumption that the d n s are conditionally independent given θ (which may not always be reasonable, but leads to an easy solution). We can now write p(d n θ 1 ) = P d n D,n (1 P D,n ) 1 d n } {{ } Bernoulli pmf where P D,n Similarly, is the nth sensor s local detection probability. p(d n θ 0 ) = P d n FA,n (1 P FA,n ) 1 d n } {{ } Bernoulli pmf where P FA,n is the nth sensor s local detection falsealarm EE 527, Detection and Estimation Theory, # 5 48
49 probability. Now, log Λ(d) = = N log n=1 N log n=1 [ p(dn θ 1 ) ] p(d n θ 0 [ P d n D,n (1 P D,n ) 1 d n P d n FA,n (1 P FA,n ) 1 d n ] H1 log τ. To further simplify the exposition, we assume that all sensors have identical performance: P D,n = P D, P FA,n = P FA. Define the number of sensors having d n = 1: u 1 = N n=1 d n Then, the loglikelihood ratio becomes log Λ(d) = u 1 log ( PD P FA ) + (N u 1 ) log ( 1 PD 1 P FA ) H1 log τ or u 1 log [ PD (1 P FA ) ] H1 ( 1 PFA ) log τ + N log. (22) P FA (1 P D ) 1 P D EE 527, Detection and Estimation Theory, # 5 49
50 Clearly, each node s local decision d n P D > P FA, which implies is meaningful only if P D (1 P FA ) P FA (1 P D ) > 1 the logarithm of which is therefore positive, and the decision rule (22) further simplifies to u 1 H 1 τ. The NeymanPerson performance analysis of this detector is easy: the random variable U 1 is binomial given θ (i.e. conditional on the hypothesis) and, therefore, P [U 1 = u 1 ] = ( ) N p u 1 (1 p) N u 1 u 1 where p = P FA under H 0 and p = P D under H 1. Hence, the global falsealarm probability is P FA,global = P [U 1 > τ θ 0 ] = N u 1 = τ ( N u 1 ) P u 1 FA (1 P FA ) N u 1. Clearly, P FA,global will be a discontinuous function of τ and therefore, we should choose our P FA,global specification from EE 527, Detection and Estimation Theory, # 5 50
51 the available discrete choices. But, if none of the candidate choices is acceptable, this means that our current system does not satisfy the requirements and a remedial action is needed, e.g. increasing the quantity (N) or improving the quality of sensors (through changing P D and P FA ), or both. EE 527, Detection and Estimation Theory, # 5 51
52 Testing Multiple Hypotheses Suppose now that we choose Θ 0, Θ 1,..., Θ M 1 that form a partition of the parameter space Θ: Θ 0 Θ 1 Θ M 1 = Θ, Θ i Θ j = i j. We wish to distinguish among M > 2 hypotheses, i.e. identify which hypothesis is true: H 0 : θ Θ 0 versus H 1 : θ Θ 1 versus. versus H M 1 : θ Θ M 1 and, consequently, our action space consists of M choices. We design a decision rule φ : X (0, 1,..., M 1): φ(x) = 0, decide H 0, 1,. decide H 1, M 1, decide H M 1 where φ partitions the data space X [i.e. the support of p(x θ)] into M regions: Rule φ: X 0 = {x : φ(x) = 0},..., X M 1 = {x : φ(x) = M 1}. EE 527, Detection and Estimation Theory, # 5 52
53 We specify the loss function using L(i m), where, typically, the losses due to correct decisions are set to zero: L(i i) = 0, i = 0, 1,..., M 1. Here, we adopt zero losses for correct decisions. posterior expected loss takes M values: Now, our ρ m (x) = = M 1 i=0 M 1 i=0 Θ i L(m i) p(θ x) dθ L(m i) Θ i p(θ x) dθ, m = 0, 1,..., M 1. Then, the Bayes decision rule φ is defined via the following dataspace partitioning: X m = {x : ρ m (x) = min ρ l(x)}, m = 0, 1,..., M 1 0 l M 1 or, equivalently, upon applying the Bayes rule X m = = { M 1 } x : m = arg min L(l i) p(x θ) π(θ) dθ 0 l M 1 i=0 Θ i { M 1 x : m = arg min L(l i) 0 l M 1 i=0 Θ i p(x θ) π(θ) dθ }. EE 527, Detection and Estimation Theory, # 5 53
54 The preposterior (Bayes) risk for rule φ(x) is E x,θ [loss] = = M 1 m=0 M 1 m=0 M 1 i=0 X m X m Θ i L(m i) p(x θ)π(θ) dθ dx M 1 L(m i) p(x θ)π(θ) dθ i=0 Θ } {{ i } h m (θ) dx. Then, for an arbitrary h m (x), [ M 1 m=0 ] h m (x) dx X m [ M 1 m=0 X m ] h m (x) dx 0 which verifies that the Bayes decision rule φ minimizes the preposterior (Bayes) risk. EE 527, Detection and Estimation Theory, # 5 54
arxiv:1309.4656v1 [math.st] 18 Sep 2013
The Annals of Statistics 2013, Vol. 41, No. 1, 1716 1741 DOI: 10.1214/13AOS1123 c Institute of Mathematical Statistics, 2013 UNIFORMLY MOST POWERFUL BAYESIAN TESTS arxiv:1309.4656v1 [math.st] 18 Sep 2013
More informationA direct approach to false discovery rates
J. R. Statist. Soc. B (2002) 64, Part 3, pp. 479 498 A direct approach to false discovery rates John D. Storey Stanford University, USA [Received June 2001. Revised December 2001] Summary. Multiplehypothesis
More informationStrong control, conservative point estimation and simultaneous conservative consistency of false discovery rates: a unified approach
J. R. Statist. Soc. B (2004) 66, Part 1, pp. 187 205 Strong control, conservative point estimation and simultaneous conservative consistency of false discovery rates: a unified approach John D. Storey,
More informationOPRE 6201 : 2. Simplex Method
OPRE 6201 : 2. Simplex Method 1 The Graphical Method: An Example Consider the following linear program: Max 4x 1 +3x 2 Subject to: 2x 1 +3x 2 6 (1) 3x 1 +2x 2 3 (2) 2x 2 5 (3) 2x 1 +x 2 4 (4) x 1, x 2
More informationChapter 3 Signal Detection Theory Analysis of Type 1 and Type 2 Data: Metad 0, Response Specific Metad 0, and the Unequal Variance SDT Model
Chapter 3 Signal Detection Theory Analysis of Type 1 and Type 2 Data: Metad 0, Response Specific Metad 0, and the Unequal Variance SDT Model Brian Maniscalco and Hakwan Lau Abstract Previously we have
More informationTHE PROBLEM OF finding localized energy solutions
600 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 3, MARCH 1997 Sparse Signal Reconstruction from Limited Data Using FOCUSS: A Reweighted Minimum Norm Algorithm Irina F. Gorodnitsky, Member, IEEE,
More informationChoosing Multiple Parameters for Support Vector Machines
Machine Learning, 46, 131 159, 2002 c 2002 Kluwer Academic Publishers. Manufactured in The Netherlands. Choosing Multiple Parameters for Support Vector Machines OLIVIER CHAPELLE LIP6, Paris, France olivier.chapelle@lip6.fr
More informationMINITAB ASSISTANT WHITE PAPER
MINITAB ASSISTANT WHITE PAPER This paper explains the research conducted by Minitab statisticians to develop the methods and data checks used in the Assistant in Minitab 17 Statistical Software. OneWay
More informationData Quality Assessment: A Reviewer s Guide EPA QA/G9R
United States Office of Environmental EPA/240/B06/002 Environmental Protection Information Agency Washington, DC 20460 Data Quality Assessment: A Reviewer s Guide EPA QA/G9R FOREWORD This document is
More informationBasics of FloatingPoint Quantization
Chapter 2 Basics of FloatingPoint Quantization Representation of physical quantities in terms of floatingpoint numbers allows one to cover a very wide dynamic range with a relatively small number of
More informationSubspace Pursuit for Compressive Sensing: Closing the Gap Between Performance and Complexity
Subspace Pursuit for Compressive Sensing: Closing the Gap Between Performance and Complexity Wei Dai and Olgica Milenkovic Department of Electrical and Computer Engineering University of Illinois at UrbanaChampaign
More information1 Theory: The General Linear Model
QMIN GLM Theory  1.1 1 Theory: The General Linear Model 1.1 Introduction Before digital computers, statistics textbooks spoke of three procedures regression, the analysis of variance (ANOVA), and the
More informationThe Set Covering Machine
Journal of Machine Learning Research 3 (2002) 723746 Submitted 12/01; Published 12/02 The Set Covering Machine Mario Marchand School of Information Technology and Engineering University of Ottawa Ottawa,
More informationAn Introduction to Regression Analysis
The Inaugural Coase Lecture An Introduction to Regression Analysis Alan O. Sykes * Regression analysis is a statistical tool for the investigation of relationships between variables. Usually, the investigator
More informationHighRate Codes That Are Linear in Space and Time
1804 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL 48, NO 7, JULY 2002 HighRate Codes That Are Linear in Space and Time Babak Hassibi and Bertrand M Hochwald Abstract Multipleantenna systems that operate
More informationPossibilistic InstanceBased Learning
Possibilistic InstanceBased Learning Eyke Hüllermeier Department of Mathematics and Computer Science University of Marburg Germany Abstract A method of instancebased learning is introduced which makes
More informationRegression. Chapter 2. 2.1 Weightspace View
Chapter Regression Supervised learning can be divided into regression and classification problems. Whereas the outputs for classification are discrete class labels, regression is concerned with the prediction
More informationSubmodular Function Maximization
Submodular Function Maximization Andreas Krause (ETH Zurich) Daniel Golovin (Google) Submodularity 1 is a property of set functions with deep theoretical consequences and far reaching applications. At
More informationLoad Shedding for Aggregation Queries over Data Streams
Load Shedding for Aggregation Queries over Data Streams Brian Babcock Mayur Datar Rajeev Motwani Department of Computer Science Stanford University, Stanford, CA 94305 {babcock, datar, rajeev}@cs.stanford.edu
More informationReview of basic statistics and the simplest forecasting model: the sample mean
Review of basic statistics and the simplest forecasting model: the sample mean Robert Nau Fuqua School of Business, Duke University August 2014 Most of what you need to remember about basic statistics
More information8 INTERPRETATION OF SURVEY RESULTS
8 INTERPRETATION OF SURVEY RESULTS 8.1 Introduction This chapter discusses the interpretation of survey results, primarily those of the final status survey. Interpreting a survey s results is most straightforward
More informationChapter 18. Power Laws and RichGetRicher Phenomena. 18.1 Popularity as a Network Phenomenon
From the book Networks, Crowds, and Markets: Reasoning about a Highly Connected World. By David Easley and Jon Kleinberg. Cambridge University Press, 2010. Complete preprint online at http://www.cs.cornell.edu/home/kleinber/networksbook/
More informationCOSAMP: ITERATIVE SIGNAL RECOVERY FROM INCOMPLETE AND INACCURATE SAMPLES
COSAMP: ITERATIVE SIGNAL RECOVERY FROM INCOMPLETE AND INACCURATE SAMPLES D NEEDELL AND J A TROPP Abstract Compressive sampling offers a new paradigm for acquiring signals that are compressible with respect
More informationTWO L 1 BASED NONCONVEX METHODS FOR CONSTRUCTING SPARSE MEAN REVERTING PORTFOLIOS
TWO L 1 BASED NONCONVEX METHODS FOR CONSTRUCTING SPARSE MEAN REVERTING PORTFOLIOS XIAOLONG LONG, KNUT SOLNA, AND JACK XIN Abstract. We study the problem of constructing sparse and fast mean reverting portfolios.
More informationHow to Use Expert Advice
NICOLÒ CESABIANCHI Università di Milano, Milan, Italy YOAV FREUND AT&T Labs, Florham Park, New Jersey DAVID HAUSSLER AND DAVID P. HELMBOLD University of California, Santa Cruz, Santa Cruz, California
More informationNew insights on the meanvariance portfolio selection from de Finetti s suggestions. Flavio Pressacco and Paolo Serafini, Università di Udine
New insights on the meanvariance portfolio selection from de Finetti s suggestions Flavio Pressacco and Paolo Serafini, Università di Udine Abstract: In this paper we offer an alternative approach to
More informationFirms are increasingly relying on software to detect fraud in domains such as security, financial services,
ecision Analysis Vol. 1, No. 3, September 2004, pp. 131 148 issn 15458490 eissn 15458504 04 0103 0131 informs doi 10.1287/deca.1040.0022 2004 INFORMS Configuration of etection Software: A Comparison
More informationA Direct Approach to Data Fusion
A Direct Approach to Data Fusion Zvi Gilula Department of Statistics Hebrew University Robert E. McCulloch Peter E. Rossi Graduate School of Business University of Chicago 1101 E. 58 th Street Chicago,
More informationSteering User Behavior with Badges
Steering User Behavior with Badges Ashton Anderson Daniel Huttenlocher Jon Kleinberg Jure Leskovec Stanford University Cornell University Cornell University Stanford University ashton@cs.stanford.edu {dph,
More informationApproximately Detecting Duplicates for Streaming Data using Stable Bloom Filters
Approximately Detecting Duplicates for Streaming Data using Stable Bloom Filters Fan Deng University of Alberta fandeng@cs.ualberta.ca Davood Rafiei University of Alberta drafiei@cs.ualberta.ca ABSTRACT
More information