Noise Enhancement for Weighted Sum of Type I and II Error Probabilities with Constraints

: In this paper, the noise-enhanced detection problem is investigated for the binary hypothesis-testing. The optimal additive noise is determined according to a criterion proposed by DeGroot and Schervish (2011), which aims to minimize the weighted sum of type I and II error probabilities under constraints on type I and II error probabilities. Based on a generic composite hypothesis-testing formulation, the optimal additive noise is obtained. The sufﬁcient conditions are also deduced to verify whether the usage of the additive noise can or cannot improve the detectability of a given detector. In addition, some additional results are obtained according to the speciﬁcity of the binary hypothesis-testing, and an algorithm is developed for ﬁnding the corresponding optimal noise. Finally, numerical examples are given to verify the theoretical results and proofs of the main theorems are presented in the Appendix.


Introduction
In the binary hypothesis testing problem, there are usually a null hypothesis H 0 and an alternative hypothesis H 1 , and the objective of testing is to be determine truthfulness of them based on the observation data and a decision rule.Due to the presence of noise, the decision result obviously cannot be absolutely correct.Generally, two erroneous decisions may occur in the signal detection: type I error that rejects a true null hypothesis and type II error that accepts a false null hypothesis [1].
In the classical statistical theory, the Neyman-Pearson criterion is usually applied to obtain a decision rule that minimizes the type II error probability β with a constraint on the type I error probability α.However, the minimum β may not always correspond to the optimal decision result.For instance, in the example 1 of [2], a binary hypothesis testing is designed to determine the mean of a normal data: the mean equals to −1 under H 0 and equals to 1 under H 1 .Under the constraint that the type I error probability is fixed to 0.05, the type II error probability is decreased from 0.0091 to 0.00000026 when the data size increases from 20 to 100, whereas the rejection region of the null hypothesis is changed from (0.1, +∞) to (−0.51, +∞).In such case, more information brings a worse decision result even a smaller type II error probability is achieved with the fixed type I error probability.Similarly, the decision rule that minimizes the type I error probability for a fixed type II error probability may not perform well.Therefore, it could not be appropriate to simply minimize one of the two error probabilities in practice.The ideal case is to develop a decision criterion that minimizes the two types of error probabilities simultaneously, but it is almost impossible in practical applications.
In order to obtain a better decision result to balance the type I and II error probabilities, DeGroot and Schervish [1] proposed a criterion to minimize a weighted sum of type I and II error probabilities, i.e., min where φ represents the decision rule, c 1 and c 2 are the weight coefficients corresponding to α and β, respectively, and c 1 , c 2 > 0. Furthermore, DeGroot

•
Formulation of the optimization problem for minimizing the noise modified weighted sum of type I and II error probabilities under the constraints on the two error probabilities is presented.

•
Derivations of the optimal noise that minimizes the weighted sum and sufficient conditions for improvability and nonimprovability for a general composite hypothesis testing problem are provided.

•
Analysis of the characteristics of the optimal additive noise that minimizes the weighted sum for a simple hypothesis testing problem is studied and the corresponding algorithm to solve the optimization problem is developed.

•
Numerical results are presented to verify the theoretical results and to demonstrate the superior performance of the proposed detector.
The remainder of this paper is organized as follows: in Section 2, a noise modified composite hypothesis testing problem is formulated first for minimizing the weighted sum of type I and II error probabilities under different constraints.Then the sufficient conditions for improvability and Entropy 2017, 19, 276 3 of 22 nonimprovability are given and the optimal additive noise is derived.In Section 3, additional theoretical results are analyzed for a simple hypothesis testing problem.Finally, simulation results are shown in Section 4 and conclusions are made in Section 5.
Notation: Lower-case bold letters denote vectors, with υ i denoting the i-th element of υ; θ denotes the value of parameter Θ; f υ (υ|θ) denotes the pdf of υ for a given parameter value Θ = θ; Ω i denotes the set of all possible parameter values of Θ under H i ; δ(•) denotes the Dirac function; ∩, ∪ and ∅ denote intersection, union and null set, respectively; * , (•) T , , E{•}, min, max and arg denote convolution, transpose, integral, expectation, minimum, maximum and argument operators, respectively; inf{•} and sup{•} denote the infimum and supremum operators, respectively; ∑ means summation; ∇ and H denote the respective gradient and Hessian operators.

Problem Formulation
Consider the following binary composite hypothesis testing problem: where x ∈ R N is the observation vector, H 0 and H 1 are the null and the alternative hypothesizes, respectively, θ denotes the value of parameter Θ, f x (x|θ) represents the pdf of x for a given parameter value Θ = θ.The parameter Θ has multiple possible values under each hypothesis and denote the pdf of any parameter value Θ = θ under H 0 and H 1 by 0 (θ) and 1 (θ).In addition, Ω 0 and Ω 1 denote the respective sets of all possible values of Θ under H 0 and H 1 .It is true that Ω 0 ∩ Ω 1 = ∅ and the union of them forms the parameter space Ω, i.e., Ω = Ω 0 ∪ Ω 1 .Without loss of generality, a decision rule (detector) is considered as: where Γ 0 and Γ 1 form the observation space Γ.Actually, the detector chooses In order to investigate the performance of the detector achieved via an additive noise, a noise modified observation y is obtained by adding an independent additive noise n to the original observation x, i.e., y = x + n.For a given parameter value Θ = θ, the pdf of y is calculated by the convolution of the pdfs of x and n, given by: where p n (n) denotes the pdf of n.For a fixed detector, the noise modified type I and II error probabilities of the detector for given parameter values now is expressed as: Correspondingly, the average noise modified type I and II error probabilities are calculated by: From ( 6) and ( 7), the weighted sum of the two types of average error probabilities is obtained as: where c 1 and c 2 are the weights assigned for the type I and II error probabilities, which can be predefined according to the actual situations.For example, if the prior probabilities are known, the value of c 1 and c 2 equal the prior probabilities corresponding to H 0 and H 1 , respectively.Besides, the values of c 1 and c 2 can also be determined based on the expected decision results.In this work, the aim is to find the optimal independent additive noise, which minimizes the weighted sum of the average error probabilities under the constraints on the maximum type I and II error probabilities for different parameter values.The optimization problem can be formulated as below: where α o and β o are the upper limits for the type I and II error probabilities, respectively.In order to explicitly express the optimization problem described in ( 9) and (10), substituting (3) into (4) produces: where It should be noted that A θ (n) can be viewed as the type I error probability obtained by adding a constant vector n to x for θ ∈ Ω 0 .Therefore, α x (φ; θ) = A θ (0) = Γ φ(x) f x (x|θ)d x denotes the type I error probability for the original observation x.
Similarly, β y (φ; θ) in ( 5) can be expressed as: where The B θ (n) can be treated as the type II error probability obtained by adding a constant vector n to x for θ ∈ Ω 1 and With (11) and ( 13), (8) becomes: Accordingly, Er(n) is the weighted sum of two types of average error probabilities achieved by adding a constant vector n to the original observation x.Naturally, Er x = Er(0) denotes the weighted sum of type I and II average error probabilities for the original observation x.
Combined (11), ( 13) and ( 15), the optimization problem in ( 9) and (10) now is: subject to   achieved simultaneously by adding any additive noise.In addition, even if the conditions in Theorem 1 are not satisfied, the detector can also be non-improvable.This implies the sufficient conditions for improvability need to be addressed.The sufficient conditions for improvability are discussed now.Suppose that A θ (x) (∀θ ∈ Ω 0 ), B θ (x) (∀θ ∈ Ω 1 ) and Er(x) are second-order continuously differentiable around x = 0.In order to facilitate the subsequent analysis, six auxiliary functions are predefined as follows based on the first and second partial derivatives of A θ (x), B θ (x) and Er(x) with respect to the elements of x.The first three auxiliary functions a θ (x, g) and er (1) (x, g) are defined as the weight sums of the first partial derivatives of A θ (x), B θ (x) and Er(x), respectively, based on the coefficient vector g.Specifically: Entropy 2017, 19, 276 where g is a N-dimensional column vector, g T is the transposition of g, x i and g i are the i-th elements of x and g, respectively.In addition, ∇ denotes the gradient operator, thereby ∇A θ (x) (∇B θ (x),∇Er(x)) is a N-dimensional column vector with i-th element ∂A θ (x)/∂x i (∂B θ (x)/∂x i , ∂E r (x)/∂x i ), i = 1, . . ., N.
The last three auxiliary functions a (2) θ (x, g) and er (2) (x, g) are defined as the weight sums of the second partial derivatives of A θ (x), B θ (x) and Er(x) based on the N × N coefficient matrix gg T , i.e., where H denote the Hessian operator, , where i, j = 1, . . ., N.
Theorem 2. Suppose that Λ 0 and Λ 1 are the sets of all possible values of θ that maximize A θ (0) and B θ (0), respectively, B θ (0).The detector is improvable, if there exists a N-dimensional column vector g that satisfies one of the following conditions for all θ 0 ∈ Λ 0 and θ 1 ∈ Λ 1 : (1) er (1) (x, g) x=0 < 0 , a The proof is presented in Appendix B. Theorem 2 indicates that under the condition (1), (2) or (3), there always exist noises that decrease the weighted sum of average error probabilities under the constraints on the type I and II error probabilities.In addition, alternative sufficient conditions for improvability can be obtained by defining the following two functions, and they are: where I(t) and S(t) are the minimum weighted sum of two types of average error probabilities and the maximum type II error probability for a given maximum type I error probability obtained via adding a constant vector, respectively.If there is a t 0 ≤ α o such that I(t 0 ) ≤ Er(0) and S(0) ≤ β o , the detector is improvable.More specifically, there exists a constant vector n 0 that satisfies max Er(n 0 ) ≤ Er(0) and max θ∈Ω 0 B θ (n 0 ) ≤ β o simultaneously.However, in most cases, the solution of the optimization problem in (17) and ( 18) is not a constant vector.A more practical sufficient condition for improvability is shown in Theorem 3. are second-order continuously differentiable around t = α, and I ( α) < 0 and S ( α) < 0 hold at the same time, then the detector is improvable.
The proof is given in Appendix C. Additionally, the following functions J(ε) and G(ε) are defined: A similar conclusion to the Theorem 3 can be made as well, provided in Corollary 1.
The proof is similar to that of Theorem 3 and it is omitted here.

Optimal Additive Noise
In general, it is difficult to solve the optimization problem in ( 17) and ( 18) directly, because the solution is obtained based on the search over all possible additive noises.Hence, in order to reduce the computational complexity, one can utilize Parzen window density estimation to obtain an approximate solution.Actually, the pdf of the optimal additive noise can be approximated by: where η l ≥ 0 and ∑ L l=1 η l = 1, while ϑ l (•) represents the window function that satisfies ϑ l (x) ≥ 0 for any x and ϑ l (x)dx = 0 for l = 1, . . ., L. The window function can be a cosine window, rectangular window, or Gauss window function.With (29), the optimization problem is simplified to obtain the parameter values corresponding to each window function.In such cases, global optimization algorithms can be applied such as Particle swarm optimization (PSO), Ant colony algorithm (ACA), and Genetic algorithm (GA) [26][27][28].
If the numbers of parameter values in Ω 0 and Ω 1 are finite, the optimal additive noise for ( 17) and ( 18) is a randomization of no more than M + K constant vectors.In this case, Ω 0 and Ω 1 can be expressed by Ω 0 = {θ 01 , θ 02 , . . . ,θ 0M } and Ω 1 = {θ 11 , θ 12 , . . . ,θ 1K }, where M and K are finite positive integers.The Theorem 4 states this claim.Theorem 4. Suppose that each component in the optimal additive noise is finite, namely n i ∈ [a i , b i ] for i = 1, . . ., N, where a i and b i are two finite values.If A θ 0i (•) and B θ 1i (•) are continuous functions, the pdf of the optimal additive noise for the optimization problem in ( 17) and ( 18) can be expressed as: where η l ≥ 0 and The proof is similar to that of Theorem 4 in [17] and Theorem 3 in [13], and omitted here.In some special cases, the optimal additive can be solved directly based on the characteristics of I(t) (H(ε)).For example, let Er min = min

the optimal additive
Entropy 2017, 19, 276 8 of 22 noise is a constant vector with pdf of p n (n) = δ(n − n m ).In addition, equality of max

Noise Enhanced Simple Hypothesis Testing
In this section, the noise enhanced binary simple hypothesis testing problem is considered, which is a special case of the optimization problem in ( 9) and (10).Therefore, the conclusions obtained in Section 2 are also applicable in this section.Furthermore, due to the specificity of simple binary hypothesis testing problem, some additional results are also obtained.

Problem Formulation
When Ω i = {θ i }, i = 0, 1, the composite binary hypothesis testing problem described in ( 1) is simplified to a simple binary hypothesis testing problem.In this case, the probability of θ i under H i equals to 1, i.e., i (θ) = 1 for i = 0, 1.Therefore, the corresponding noise modified type I and II error probabilities is rewritten as: where f 0 (•) and f 1 (•) represent the pdfs of x under H 0 and H 1 , respectively, and A 0 (n) and B 1 (n) are: Correspondingly, the weighted sum of noise modified type I and II error probabilities is calculated by: where As a result, the optimization problem in ( 9) and (10) becomes: Based on the definitions in (33) and (34), A 0 (n) and B 1 (n) are viewed as the noise modified type I and II error probabilities obtained by adding a constant vector noise.Furthermore, A 0 (0) and B 1 (0) are the original type I and II error probabilities, respectively.

Algorithm for the Optimal Additive Noise
According to the Theorem 4 in Section 2.3, the optimal additive noise for the optimization problem in (37) and ( 38 (1 − η)δ(n − n 2 ).In order to find the values of η, n 1 and n 2 , we first divide each constant vector n into four disjoint sets according to the relationships of A 0 (n) and α o , B 1 (n) and β o .To be specific, the four disjoint sets are Then, we calculate the minimum Er(n), the corresponding set of all possible values of n is denoted by It should be noted that n ∈ Q e is the optimal additive noise that minimizes the weighted sum without constraints.
It is obvious that Q 1 , Q 2 and Q 3 do not exist if all the elements of Q e belong to Q 4 .In other words, if Q e ⊂ Q 4 , there is no additive noise that satisfies E{Er(n)} < Er(0) under the constraints of E{A 0 (n)} ≤ α o and E{B 1 (n)} ≤ β o .Therefore, if the detector is improvable, the elements of Q e must come from Q 1 , Q 2 and/or Q 3 .Theorem 5 is now provided to find the values of η, n 1 and n 2 . ( and Er ( The corresponding proofs are provided in Appendix D. From ( 3) and (4) in Theorem 5, under the constraints on E{A 0 (n)} ≤ α o and E{B 1 (n)} ≤ β o , the solution of the optimization problem in (37) and ( 38) is identical with the additive noise that minimizes ).In such cases, the optimal solution can be obtained easily by referring the algorithm provided in [14].

Numerical Results
In this section, a binary hypothesis testing problem is studied to verify the theoretical analysis, and it is: where x ∈ R is an observation, Θ is a constant or random variable, and v is the background noise with pdf p v (•).From (39), the pdf of x under H 0 is f 0 (x) = p v (x), and the pdf of x under H 1 for a given parameter value Θ = θ is denoted by , where p θ (•) represents the pdf of Θ = θ.A noise modified observation y is obtained via adding an additive independent noise n to the observation x, i.e., y = x + n.If the additive noise n is a constant vector, the pdf of y under H 0 is calculated as f 0 (y) = f 0 (x − n), and the pdf of y under In addition, a linear-quadratic detector is utilized here, given by:

Rayleigh Distribution Background Noise
Suppose that Θ = s is a constant, the problem shown in (39) represents a simple binary hypothesis testing problem.Here, we set d 0 = d 2 = 0 and d 1 = 1, then the detector becomes It is assumed that the background noise v obeys the mixture of Rayleigh distributions with zero-means such that , where m i ≥ 0 for i = 1, . . ., M, ∑ M i=1 m i = 1, and In the simulations, the variances of all the Rayleigh components are assumed to be the same, i.e., σ i = σ for i = 1, . . ., M. In addition, the parameters are specified as 4 and m i = 0.25 for i = 1, . . ., 4. From (33) and (34), the noise modified type I error probability A 0 (n) and type II error probability B 1 (n) obtained by adding a constant vector n is calculated as: where ) and c 2 = α x /(α x + β x ), the noise modified weighted sum of the two types of error probabilities obtained via adding a constant vector is From Section 3.2, the pdf of the optimal additive noise that minimizes weighted sum of type I and II error probabilities is denoted by , under the two constraints that α y ≤ α x and β y ≤ β x .Moreover, the optimal additive noise for the case without any constraints is a constant vector.
Figure 1 plots the minimum noise modified weighted sums of type I and II error probabilities obtained under no constraint and two constraints that α y ≤ α x and β y ≤ β x , and the original weighted sum without adding any noise for different values of σ when s = 3 and γ = s/2.When σ ≤ σ 1 , there is no noise that decreases the weighted sum.With the increase of σ, noise exhibits a positive effect on the detection performance.To be specific, when σ 1 < σ < σ 2 , the weighted sum can be decreased by adding a constant vector for the no constraint case.When σ > σ 2 , the weighted sum can be decreased adding the noise under two constraints.The noise modified weighted sum obtained without any constraints is less than or equal to that obtained under the two constraints, and the difference between them first decreases to zero for σ 3 < σ < σ 4 and then gradually increases when σ > σ 4 .In addition, once σ exceeds a certain value, no noise exists that can decrease the weighted sum for any cases.
weighted sum can be decreased adding the noise under two constraints.The noise modified weighted sum obtained without any constraints is less than or equal to that obtained under the two constraints, and the difference between them first decreases to zero for  Figure 2 shows the type I and II error probabilities corresponding to the weighted sum in Figure 1.From both Figures 1 and 2, it is observed that one of the noise modified Type I and II error probabilities performs worse than the original one for the no constraint case.Therefore, though the noise modified weighted sum obtained with no constraint is less than that obtained under the two constraints, the corresponding noise is actually not suitable to add to the observation.Furthermore, when the minimum value of the noise modified weighted sum is obtained under the two constraints, the corresponding type II probability equals to the original one and the type I probability achieves the minimum for 2 , the minimum values of the noise modified weighted sum obtained under no constraint is equal to that obtained under two constraints, and the corresponding type I and II error probabilities are the same, which also agrees with part (2) in Theorem 5.In order to further illustrate the results in Figures 1  and 2, Table 1 provides the optimal additive noises added for the two different cases.Figure 2 shows the type I and II error probabilities corresponding to the weighted sum in Figure 1.From both Figures 1 and 2, it is observed that one of the noise modified Type I and II error probabilities performs worse than the original one for the no constraint case.Therefore, though the noise modified weighted sum obtained with no constraint is less than that obtained under the two constraints, the corresponding noise is actually not suitable to add to the observation.Furthermore, when the minimum value of the noise modified weighted sum is obtained under the two constraints, the corresponding type II probability equals to the original one and the type I probability achieves the minimum for σ 2 < σ < σ 3 .Conversely, when σ > σ 4 , the corresponding type I probability equals to the original one and the type II probability achieves the minimum.The results are consistent with part (3) and part (4) in Theorem 5. Especially, for σ 3 < σ < σ 4 , the minimum values of the noise modified weighted sum obtained under no constraint is equal to that obtained under two constraints, and the corresponding type I and II error probabilities are the same, which also agrees with part (2) in Theorem 5.In order to further illustrate the results in Figures 1 and 2, Table 1 provides the optimal additive noises added for the two different cases.    2 s γ = .The corresponding type I and II error probabilities are depicted in Figure 4a,b, respectively.It is seen in Figure 3, the improvement of the weighted sum obtained by adding noise first increases and then decreases with the increase of s, and finally they all converge to the same value.The differences for the cases with and without constraints are very small in most cases.In the small interval of s, i.e., 1 2 ( , ) s s s ∈ , the difference even decreases to zero.On the other  Figure 3 depicts the minimum noise modified weighted sums of the type I and II error probabilities versus s for the cases of no constraint and two constraints, and the original weighted sum, when σ = 1 and γ = s/2.The corresponding type I and II error probabilities are depicted in Figure 4a,b, respectively.It is seen in Figure 3, the improvement of the weighted sum obtained by adding noise first increases and then decreases with the increase of s, and finally they all converge to the same value.The differences for the cases with and without constraints are very small in most cases.In the small interval of s, i.e., s ∈ (s 1 , s 2 ), the difference even decreases to zero.On the other hand, the noise modified type I error probability obtained under no constraint is significantly greater than the original one for s < s 1 , while the corresponding type II error probability is less than that obtained under the two constraints.The situation, however, is reversed for s 2 < s < s 3 .When s > s 3 , there is no noise that decreases the weighted sum under the two constraints, while the weighted sum is still decreased by adding a constant vector for no constraint case.When s > s 4 , the weighted sum cannot be decreased by adding any noise for all the cases.Furthermore, Table 2 shows the optimal additive noises that minimize the weighted sum under the cases of no and two constraints.Figure 5 shows the minimum noise modified weighted sums of type I and II error probabilities versus γ for the cases of no constraint and two constraints, and the original weighted sum, when 1 σ = and 3 s = .The corresponding type I and II error probabilities are depicted in Figure 6a,b, respectively.As illustrated in Figure 5, when γ is close to zero, the original weighted sum x Er approaches to zero.In such case, no additive noise exists to decrease the weighted sum.For the case of two constraints, the improvement of the weighted sum first increases for  Figure 5 shows the minimum noise modified weighted sums of type I and II error probabilities versus γ for the cases of no constraint and two constraints, and the original weighted sum, when σ = 1 and s = 3.The corresponding type I and II error probabilities are depicted in Figure 6a,b, respectively.As illustrated in Figure 5, when γ is close to zero, the original weighted sum Er x approaches to zero.In such case, no additive noise exists to decrease the weighted sum.For the case of two constraints, the improvement of the weighted sum first increases for γ 1 < γ < γ 2 and then decreases for γ 2 < γ < γ 3 , and no improvement can be obtained when γ > γ 2 .On the other hand, the minimum noise modified weighted sum obtained under no constraint is smaller than that obtained under the two constraints for γ 1 < γ < γ 3 , and the difference between them first increases and then decreases for both γ 1 < γ < γ 2 and γ 2 < γ < γ 3 .When γ > γ 3 , there still exists a constant vector that decreases the weighted sum, but it may be not a suitable noise in the practical application according to the type II probability depicted in Figure 6b.Furthermore, in order to study the results illustrated in Figures 5 and 6, Table 3 shows the optimal additive noises that minimize the weighted sum for the cases of no and two constraints.> , there still exists a constant vector that decreases the weighted sum, but it may be not a suitable noise in the practical application according to the type II probability depicted in Figure 6b.Furthermore, in order to study the results illustrated in Figures 5 and 6, Table 3 shows the optimal additive noises that minimize the weighted sum for the cases of no and two constraints.Suppose that Θ is a random variable with following pdf: Therefore, we have Ω 0 = {0} and Ω 1 = {s, −s}.In the simulations, we set d 0 = 1, d 1 = 0, d 2 = −s 2 /4 and γ = 0, the detector is expressed as: Moreover, we assume that v is a zero-mean symmetric Gaussian mixture noise with pdf of , where m i ≥ 0, ∑ M i=1 m i = 1 and: Let M = 4 and the mean values of the symmetric Gaussian components are set as [0.05 0.52 −0.52 −0.05] with corresponding weights [0.35 0.15 0.15 0.35].In addition, the variances of Gaussian components are the same, i.e., σ i = σ for i = 0, . . ., 4. According to ( 12) and ( 14), the noise modified type I error probability obtained by adding a constant vector n to x is calculated by: and the corresponding type II error probabilities for Θ = s and −s are respectively calculated as: where exp(− t 2 2σ 2 )dt.Accordingly: Therefore, the original type I and type II error probabilities for Θ = s and −s are α x (φ; 0) = A 0 (0), β x (φ; s) = B s (0) and β x (φ; −s) = B −s (0), respectively.
Due to the symmetry property of v, one obtains B s (0) = B −s (0).In this case, the original average type II error probability is The noise modified weighted sum of type I and average type II error probabilities corresponding to the constant vector is expressed by Er The values of c 1 and c 2 are still specified as β x /(α x + β x ) and α x /(α x + β x ), respectively.From Theorem 4 in Section 2.3, the optimal additive noise that minimizes the weighted sum is a randomization with a pdf of , where η i ≥ 0 for i = 1, . . ., 3, and ∑ 3 i=1 η i = 1. Figure 7 shows the detection performance of the original detector and the noise enhanced detector that minimizes the weighted sum of type I and average type II error probabilities under the constraints that α y (φ; 0) ≤ α o and max From Figure 7, the original weighted sums, type I error probabilities, and type II error probabilities for Θ = s and −s increase as σ decreases towards zero.In Figure 7a, when σ is close to zero, the weighted sum can be decreased significantly.With the increase of σ, the improvement obtained by adding noise is reduced gradually to zero.In other words, the phenomenon of noise-enhanced detection performance cannot occur when σ exceeds a certain value.In Figure 7b, the noise modified type I error probability stays at 0.1500 for σ < 0.07 and then increases gradually to equal to the original type I error probability.Moreover, the noise modified type II error probabilities for Θ = s corresponding to the minimum weighted sum increases from zero to that of original detector, shown in Figure 7c, while the type II error probabilities for Θ = −s of the noise enhanced detector is equal to that of the original detector all the time.In fact, the type II error probability for Θ = s also reaches the minimums under the constraints that α y (φ; 0) ≤ α o and max θ∈Ω 1 β y (φ; θ) ≤ β o in this example.In addition, Table 4 offers the optimal additive noises that minimize the weighted sum for different values of σ to explain the results in Figure 7.It should be noted that the optimal noise is not unique.y.From Figure 7, the original weighted sums, type I error probabilities, and type II error probabilities for s Θ = and s − increase as σ decreases towards zero.In Figure 7a, when σ is close to zero, the weighted sum can be decreased significantly.With the increase of σ , the improvement obtained by adding noise is reduced gradually to zero.In other words, the phenomenon of noise-enhanced detection performance cannot occur when σ exceeds a certain value.In Figure 7b, the noise modified type I error probability stays at 0.1500 for 0.07 σ < and then increases gradually to equal to the original type I error probability.Moreover, the noise modified type II error probabilities for s Θ = corresponding to the minimum weighted sum increases from zero to that of original detector, shown in Figure 7c, while the type II error probabilities for s Θ = − of the noise enhanced detector is equal to that of the original detector all the time.In fact, the type II error probability for s Θ = also reaches the minimums under the constraints that ( ;0)  4 offers the optimal additive noises that minimize the weighted sum for different values of σ to explain the results in Figure 7.It should be noted that the optimal noise is not unique.Figure 8a demonstrates the weighted sums of type I and average type II error probabilities of the original detector and the noise enhanced detector versus s, where σ = 0.08 and ρ = 0.6.The corresponding type I error probability and type II error probabilities for Θ = s and −s are depicted in Figure 8b-d, respectively.From Figure 8a, the weighted sum cannot be decreased under the constraints on different error probabilities for s < s 1 and s < s 2 .Conversely, there exists additive noise under the constraints that reduces the weighted sum for s 1 < s < s 2 , and the corresponding improvement first increases and then decreases with the increase of s.Comparing Figure 8b with Figure 8a, it is noted that the change of the noise modified type I error probability is similar to that of the noise modified weighted sum.In Figure 8c, the noise modified type II error probability for Θ = s first decreases to the minimum and then increases as s increases, while the type II error probability for Θ = −s of the noise modified detector is always equal to that of the original detector, shown in Figure 8d.In addition, in order to further illustrate the results in Figure 8, Table 5 shows the optimal noises that minimize the weighted sum under the case of two constraints.

Conclusions
In this paper, a noise-enhanced detection problem has been investigated for a general composite hypothesis testing.Under the constraints of type I and II error probabilities, the minimization of the weighted sum of average type I and II error probabilities has been explored by adding an additive independent noise.The sufficient conditions for improvability of the weighted sum are provided, and a simple algorithm to search the optimal noise is developed.Then some additional theoretical results are made based on the specificity of the binary simple hypothesis testing problem.The studies on different noise distributions confirm the theoretical analysis that the optimal additive noise indeed minimizes the weighted sum under certain conditions.To be noted that, theoretical results can also be extended to a broad class of noise enhanced optimization problems under two inequality constraints such as the minimization of Bayes risk under the different constraints of condition risks for a binary hypothesis testing problem.

θ∈Ω 1 β
x (φ; θ) be the respective maximum type I and II error probabilities without adding any noise, and suppose that α ≤ α o , β ≤ β o and S( α) = β.If I(t) and S(t) ) is a randomization of most two constant vectors with the pdf p opt n (n) = ηδ(n − n 1 ) + Entropy 2017, 19, 276 9 of 22 then Er y opt is obtained when η = η 2 , and the corresponding E{A 0 (n)} achieves the minimum and E{B 1 (n)} = β o .(4) If Q e ⊂ Q 3 , then Er y opt is achieved when η = η 1 , and the corresponding E{A 0 (n)} = α o and E{B 1 (n)} reaches the minimum.
) where d 0 , d 1 and d 2 are detector parameters, and γ denotes the detection threshold.In the numerical examples, α o = α x and β o = β x , where α x and β x are the original type I and II error probabilities, respectively.Entropy 2017, 19, 276 10 of 22

.
In addition, once σ exceeds a certain value, no noise exists that can decrease the weighted sum for any cases.

Figure 1 .
Figure 1.The minimum noise modified weighted sums of the type I and II error probabilities obtained under no constraint and two constraints, and the original weighted sum for different σ when 3 s = and equals to the original one and the type II probability achieves the minimum.The results are consistent with part (3) and part (4) in Theorem 5. Especially, for

3 Figure 1 .
Figure 1.The minimum noise modified weighted sums of the type I and II error probabilities obtained under no constraint and two constraints, and the original weighted sum for different σ when s = 3 and γ = s/2.

Figure 2 .
Figure 2. The type I (a) and II (b) error probabilities corresponding to the weighted sum in Figure 1.

Figure 3
Figure 3 depicts the minimum noise modified weighted sums of the type I and II error probabilities versus s for the cases of no constraint and two constraints, and the original weighted sum, when 1 σ = and

4 Figure 2 .
Figure 2. The type I (a) and II (b) error probabilities corresponding to the weighted sum in Figure 1.

Entropy 2017, 19 , 278 13 of 23 Figure 3 .Figure 3 .
Figure 3.The minimum noise modified weighted sums of the type I and II error probabilities obtained under no constraint and two constraints, and the original weighted sum for different s when 1 σ = and

Figure 3 .
Figure 3.The minimum noise modified weighted sums of the type I and II error probabilities obtained under no constraint and two constraints, and the original weighted sum for different s when 1 σ = and

Figure 4 .
Figure 4.The type I and II error probabilities corresponding to the weighted sum in Figure 3 are shown in (a) and (b), respectively.

Figure 4 .
Figure 4.The type I and II error probabilities corresponding to the weighted sum in Figure 3 are shown in (a) and (b), respectively.

1 3
the other hand, the minimum noise modified weighted sum obtained under no constraint is smaller than that obtained under the two constraints for

Figure 5 .
Figure 5.The minimum noise modified weighted sums of the type I and II error probabilities obtained under no constraint and two constraints, and the original weighted sum for different γ when 1 σ = and 3 s = .

Figure 5 .
Figure 5.The minimum noise modified weighted sums of the type I and II error probabilities obtained under no constraint and two constraints, and the original weighted sum for different γ when σ = 1 and s = 3.

Figure 5 .Figure 6 .
Figure 5.The minimum noise modified weighted sums of the type I and II error probabilities obtained under no constraint and two constraints, and the original weighted sum for different γ when 1 σ = and 3 s = .

2 Figure 6 .
Figure 6.The type I and II error probabilities corresponding to the weighted sum in Figure 5 are shown in (a) and (b), respectively.

θ∈Ω 1 β
y (φ; θ) ≤ β o , for different values of σ where s = 1 and ρ = 0.6.The minimum achievable noise modified weighted sum is plotted in Figure7a, and the corresponding type I error probability and type II error probabilities for Θ = s and −s are depicted in Figure7b-d, respectively.

Figure 7 .
Figure 7.The weighted sums, type I error probabilities, and type II error probabilities for s Θ = . In addition, Table

Figure 7 .
Figure 7.The weighted sums, type I error probabilities, and type II error probabilities for Θ = s and −s of the original detector and the noise enhanced detector for different σ where s = 1 and ρ = 0.6 shown in (a), (b), (c) and (d), respectively

Figure 8 .
Figure 8.The weighted sums, type I error probabilities, and type II error probabilities for s Θ =

Figure 8a demonstrates theFigure 8 .
Figure 8a demonstrates the weighted sums of type I and average type II error probabilities of the original detector and the noise enhanced detector versus s, where

Table 1 .
The optimal additive noises that minimize the weighted sum under two constraints and no constraint for various σ where

Table 1 .
The optimal additive noises that minimize the weighted sum under two constraints and no constraint for various σ where s = 3 and γ = s/2.

Table 2 .
The optimal additive noises that minimize the weighted sum under two constraints and no constraint for various s where σ = 1 and γ = s/2.

Table 2 .
The optimal additive noises that minimize the weighted sum under two constraints and no constraint for various s where

Table 3 .
The optimal additive noises that minimize the weighted sum under two constraints and no constraint for various γ where

Table 3 .
The optimal additive noises that minimize the weighted sum under two constraints and no constraint for various γ where σ = 1 and s = 3.

Table 4 .
The optimal additive noises that minimize the weighted sum under two constraints for various σ where s = 1 and ρ = 0.6.

Table 4 .
The optimal additive noises that minimize the weighted sum under two constraints for various σ where

Table 5 .
The optimal additive noises that minimize the weighted sum under two constraints for various s where σ = 0.08 and ρ = 0.6.