Research Journal of Mathematics and Statistics 4(1): 1-5, 2012 ISSN: 2040-7505 © Maxwell Scientific Organization, 2012 Submitted: October 12, 2011 Accepted: November 18, 2011 Published: February 25, 2012 On D-Optimality Criterion of Non-Overlapping and Overlapping Segmentation of the Response Surfaces 1 1 T.A. Ugbe and 2P.E Chigbu Department of Mathematics/Statistics and Computer Science, University of Calabar, Calabar, Cross River State, Nigeria 2 Departments of Statistics, University of Nigeria, Nsukka, Enugu State, Nigeria Abstract: The aim of this study was to find out the Segment which generates a design that is D-optimal(has minimum variance). This was achieved by partitioning the Response Surface into two equal segments for Nonoverlapping and Overlapping Segments and then select support points which formed the design matrices for the segments. Further more, the determinant of the information Matrices of the designs were compared for both Non-overlapping and Overlapping Segments using the Unbiased response function and the biased response function respectively. It was found that design 1 is D-optimal and also G-optimal (by equivalence theorem given in theorem 1),that is the Non-Overlapping Segmentation of the Response Surface forms a design which is D- and G-optimal for the first order unbiased response function and second order biased response function, respectively. In other words, the Overlapping Segmentation of the Response Surfaces form a design that was not D-optimal. Key words: Design matrix, determinant, information matrix, optimality, segmentation, support points ‘‘Prediction Criterion’’.An experimental design which is optimal with respect to an estimation Criterion like the D-optimality is the one that maximizes Parameter information by minimizing Variabilities of the parameters. A design which is optimal with respect to a prediction Criterion like Qoptimality maximizes information about a response Surface by focusing on the Prediction qualities of the fitted model. D-optimality is the most important and popular design Criterion in the life applications, which is introduced by Wald (1943), put the emphasis on the quality of the parameter estimates. D-optimality Criterion is also known as determinant Criterion and is defined as; INTRODUCTION Optimal designs are experimental designs that are generated based on a particular Optimality Criterion and are generally Optimal only for a Specific Statistical Model. The proper Meaning of ‘Optimal’ depends on the situation, and can include: Most effective, Minimum Variance, Minimum bias, etc. Kiefer and Wolfowitz (1959) developed a theoretical frame work for optimal design Criteria by expressing a design as a probability Measure representing the allocation of observations at any point in the design space: Their theoretical approach to design optimality and their introduction of the D- and E Optimality Criteria for the linear regression model laid the ground work for other design Criteria A-, F-, G- and Q- Optimality Criteria etc. Each design optimality Criterion addresses a Specific goal in the experiment to be performed or achieves a Specific property in the final fitted regression model. Many of these design Criteria were originally developed for the homogeneous Variance Linear Models, but most have been adopted for use in a non-linear and non-homogeneous Variance Situation as well. The basic idea underlying design optimality theory is that statistical inference about quantities of interest can be improved by Max det (X/X) = Min det (X/X)G1 xi, i = 1, …, n xi, i = 1, …, n Which means maximizing the determinant of the information matrix or equivalently, minimizing the determinant of the inverse of the information matrix. The aim of D-optimality is essentially a parameter estimation Criterion. This was called lately, D-optimality by Kiefer and Wolfowitz (1959). It is the most well studied problem which is Seen in the Literature by Kiefer (1959), Fedorov (1972), Silvey (1980), Pazman (1986), Atkinson and Donev (1992), Pukelsheim (1993), Mandal (2000) and Bamanga and Asiribo (2006) etc. The objective of this paper is to partition the Response Surface into 2 Segments, obtain design matrices from each ‘‘Optimally’’ selecting levels of the control Variables. In general, a design Optimality Criterion can be characterized as an ‘‘Estimation Criterion’’ or Corresponding Author: T.A. Ugbe, Department of Mathematics/Statistics and Computer Science, University of Calabar, Calabar, Cross River State, Nigeria, Tel.: 08160980759 1 Res. J. Math Stat., 4(1): 1-5, 2012 segment, find out which segment is D-optimal using Nonoverlapping and overlapping Segments for both unbiased and biased response functions respectively. X2 (0,1) (1,1) MATERIALS AND METHODS This study was carried out in the University of Nigeria,Nsukka,Enugu State,Nigeria.The essence of this work was to explore the Response Surface and segment it into Non-overlapping and Overlapping segments for unbiased and biased models, and then find out the segment that is D-optimal. Segmentation is the partitioning of the Response ~ Surface [(Experimental Space), X ] , into Subspaces called Segments. Support points are selected from each Segment which give rise to design matrices of the Kth segments: Onukogu and Chigbu (2002). The Segments used here are Non-overlapping and Overlapping and are Partitioned into two equal Segments. Illustrative description of Segmentation of the Response Surfaces are shown in the Fig. 1 and 2. (0,1/2) S1 (1,1/2) S2 X1 (0,0) (1/2,0) Fig. 1: Segmentation{(Non-overlapping);describing the partitioning of the response surface into two segments, S1 and S2} X2 (0,2) Equivalence of D- and G-optimality criteria: In design of experiment, the main tool for checking the optimality of a candidate design are equivalence theorems, Kiefer (1974), Pukelsheim and Torsney (1991) show that the Dand G-optimality Criteria are equivalent if Fe-2 equal to a constant. In other words, a D-optimal design is also minimax and on the other hand a minimax design is Doptimal: Fedorov (1972). The equivalence theorems have been stated and proved by many authors, Kiefer and Wolfowitz (1960), Kiefer (1974), Pukelsheim (1980) and Pazman (1986). The statement of the theorem as given: (0,1) (0,0) Theorem1: Given M(.*) to be non-singular and Fe = 1, a design is D-optimal if it is G-optimal, that is: (1.2) (1/2,2) (2.2) (3/2,2) (3/2,1) (1/2,1) S1 S12 S21 (1,0) (1/2,0) S2 (2,0) (3/2,0) X1 -2 Max x ∈s ( x ) det M (.*) = Fig. 2 : Segmentation{(overlapping);describing the partitioning of the response surface into two segments,S1 and S2 such that S1 overlaps into S2,and is written as S21 and S2 overlaps into S1,and is witten as S12} – det M(.) Max = x ∈s ( x ) X/ M -1(.*)X = Min{ X2 2 7/4 3/2 X/ M G1(.*) X}, S(x)< X~ , M (.),M (.*) m M,n×m Max x ∈s ( x ) 5/4 Max X/M-1(.*)X = n, n is the number of linearly independent parameters in the model, that is the rank of M(.*) . 1* 3/4 1/2 RESULTS AND DISCUSSION * * * * * S1 * * S2 * * 1/4 To consider D-optimality criterion of Nonoverlapping and overlapping segmentation of responses surfaces, we first of all consider the response functions (a) First order unbiased response function (Linear) is given by: (0,0) 1/4 1/2 3/4 1 5/4 3/2 7/4 2 X1 Fig. 3: Non-overlapping segmentation of the response surfaces {showing different support points, marked with asterisk which are picked to form design design matrices from S1 and S2, respectively} f (x1,x2) = a00 +a10x1 +a22x2 2 Res. J. Math Stat., 4(1): 1-5, 2012 X2 2 7/4 3/2 X2 2 7/4 3/2 * * ** 5/4 S1 3/2 ** 3/4 S21 1/2 1/4 1/2 * 3/4 1 5/4 * S2 * * 1/4 * 0 * 1/2 * 1/4 3/2 7/4 * 3/4 * S12 * * S1 1 S2 * * 5/4 * 1 * 2 * * (0) X1 Fig. 4: Overlapping segmentation of the response surface, {showing different support points marked with asterisk which are picked to form design matrices from S12 and S21, respectively} 1/4 1/2 3/4 5/4 1 3/2 7/4 2 X1 Fig. 5: Non-overlapping segmentation of the response surface {same as Fig. 3} SA = {x1, x2; 3/4 #x1 #1, 0 #x2 #2} SB = {x1, x2; 9/8 #x1 #3/2,1/4 #x2 #2} We define the segments (from Fig. 3) by: where SA = S12 (overlapping portion of S2 into S1 ) SB = S21 (overlapping portion of S1 into S2 ) S1 = {x1, x2; 0 #x1 #2, 0 # x2 #2} S2 = {x1, x2; 0 #x1 #2, 1/2 #x2 # 2} The design matrices are: The design matrices are: ⎛ ⎜1 ⎜ ⎜1 ⎜ ⎜1 X A = ⎜⎜ 1 ⎜ ⎜1 ⎜ ⎜ ⎜1 ⎝ ⎛ 1 0 0⎞ ⎛ 1 2 0⎞ ⎜ ⎟ ⎟ ⎜ 1 7 1 1 2 ⎜ ⎟ ⎜1 2 4⎟ ⎜1 1 1⎟ ⎟ ⎜ 3 2 2 ⎟ , X = ⎜ 1 2 1⎟ X1 = ⎜ 2 ⎜ 1 21 2⎟ ⎜ 1 2 1⎟ ⎜ ⎟ ⎜ 7 1⎟ ⎜ 1 0 1⎟ ⎜1 4 2⎟ ⎟ ⎜ ⎜ ⎟ ⎝ 1 2 2⎠ ⎝ 1 21 23 ⎠ ) . 12708 0.9583⎞ ⎛1 ⎟ ⎜ 1 . . . X B/ X B = ⎜ 12708 16328 12083 ⎟ N ⎟ ⎜ 0.9583 12083 . . 12396 ⎠ ⎝ ( 15417 13750 . . ⎛1 ⎞ ⎜ ⎟ 1 / 0.8854 19792 . . X 2 X 2 = ⎜15417 ⎟ N ⎜13750 ⎟ 19792 2 2188 . . . ⎝ ⎠ ( 7 8 3 4 ⎛1 ⎞ . 0.8750 11662 ⎜ ⎟ 1 / . X A X A ) = ⎜ 0.8750 0.7760 10521 ( ⎟ N ⎜ ⎟ . . . 10521 17500 ⎝ 11667 ⎠ 0.3353 1 ⎛1 ⎞ ⎜ ⎟ 1 / X 1 X 1 = ⎜ 0.3333 01667 . 0.4167⎟ N ⎜1 ⎟ . 0.4167 14167 ⎝ ⎠ ( 9 1⎞ ⎛ ⎜1 ⎟ ⎞ 8 2⎟ ⎜ 0⎟ 5 ⎜ ⎟ ⎟ ⎜1 4 1 ⎟ 1⎟ ⎜ ⎟ 3⎟ ⎜ 1 9 3⎟ ⎟ ⎜ 8 4⎟ 2⎟ , XB = ⎜ ⎟ 5 2⎟ ⎜1 2⎟ ⎟ 4 ⎜ ⎟ 1 ⎟⎟ ⎜ 11 5 ⎟ ⎜1 ⎟ ⎟ 8 4⎟ 3⎟ ⎜ ⎜ 2 1⎟ 2⎠ ⎜1 ⎟ ⎝ 2 4⎠ 3 4 1 7 8 1 ) M (ς N 2 ) = ⎛ 2.0000 18750 . 2.3750 ⎞ ⎜ ⎟ 1 1 M (ζ N 1 ) = ( X 1/ X 1 ) + X 2/ X 2 ) = ⎜ 18750 . 3.0521 2.3959⎟ ( N N ⎜ ⎟ ⎝ 2.3750 2.3959 3.6355⎠ ) 2.1458 2.1250 ⎞ ⎛2 ⎜ ⎟ 1 1 X A/ X A + X B/ X B = ⎜ 2.1458 2.4089 2.2604⎟ N N ⎜ 2.1250 2.2604 2.9896⎟ ⎝ ⎠ ( ) ( ) where M(.N2) is the information matrix of overlapping segmentation. where M(.N1) is the information matrix of the Nonoverlapping segmentation det M(.N2) = 0.1552 |det M(.N1)>det M(.N2) det M (.N1) = 2.053 Thus by comparison det M (.N1) is maximized, this implies that .N1 is D-optimal. By Equivalence theorem stated in theorem 1, the design .N1 is also G-optimal. We define the segments (from Fig.4) by: 3 Res. J. Math Stat., 4(1): 1-5, 2012 Therefore the Non-overlapping segmentation of the response surface generates or forms a design which is Dand-G-optimal for a first order unbiased model. (b) Second-order biased response function is given by: X2 2 15/8 7/4 13/8 3/2 11/8 5/4 9/8 1 7/8 3/4 5/8 1/2 3/8 1/4 1/8 f(x1,x2) = a00+a10x1+a20x2+a12x1x2+a11x12+a22x22 We define the segments (from Fig. 5) by: S1 = {x1,x2;0 # x1 #7/8, 0 # x2 # 2} S2 = {x1 , x2;9/8 #x1 #2, 1/2 #x2 #2} The design matrices are: * * 1 2 0 3 4 1 4 7 8 3 4 1 0 3 2 1 2 2 1 4 ⎛1 3 2 ⎜ ⎜1 2 ⎜ ⎜ 1 74 X2 = ⎜ 9 ⎜1 8 ⎜ 5 ⎜1 4 ⎜ ⎝ 1 13 8 1 2 0 9 8 1 8 14 8 3 16 1 4 0 9 16 1 16 49 64 9 16 1 2 ⎤ 1⎥ 0⎥ 9⎥ ⎥ 4⎥ 1⎥ 4⎥ ⎥ 4⎥ 1⎥ ⎥ 16 ⎦ 3 2 S12 * * S1 1/4 1/2 S21 3/4 S2 5/4 3/2 7/4 1 2 X1 Fig. 6: Overlapping segmentation of the response Surfaces {same as Fig. (4)} 9 4 1 2 4 4 3 2 21 8 7 4 63 32 49 16 81 64 15 8 75 32 25 16 13 16 * * 0 ⎡ ⎢1 ⎢1 ⎢ ⎢1 ⎢ X1 = ⎢ 1 ⎢ ⎢ ⎢1 ⎢ ⎢1 ⎣ * 169 64 ⎛2 ⎜ ⎜ 2.0625 ⎜ 2.3125 M (ς N 1 ) = ⎜ ⎜ 2.8307 ⎜ 2.8307 ⎜ ⎝ 3.6068 1⎞ ⎟ 4 ⎟ ⎟ 9 4 ⎟ ⎟ 49 16 ⎟ 225 ⎟ 64 ⎟ 1 ⎟ 4 ⎠ 2.0625 2.3125 2.8229 2.8307 3.6068 ⎞ ⎟ 2.8307 2.8229 4.0176 4.3428 4.5801⎟ ⎟ 2.8229 3.6068 4.5801 4.0175 61611 . ⎟ 4.0176 4.5801 6.6331 6.3207 7.9921 ⎟ 4.3428 4.0175 6.3207 7.1238 6.6331 ⎟ ⎟ 4.5801 61611 7.9921 6.6331 10.9987⎠ . det M (.N1) = 0.0021 We define the segments (from Fig. 6)by: SA = {1/2 #x1 #7/8, 1/4 #x2 #3/2} SA = {1#x1# 3/2,1#x2#2} where: SA = S12 (overlapping portion of S2 into S1) SB=S21 (overlapping portion of S1 into S2) The design matrices are: 0.5208 0.8750 ⎛1 ⎜ 0 5208 0.3672 0.6146 . ⎜ ⎜ 0.8750 0.6146 12604 . 1 ( X / X )= ⎜ N 1 1 ⎜ 0.6146 0.4661 0.9661 ⎜ 0.3672 0.2757 0.4661 ⎜ 0.9661 2.0859 . ⎝ 12604 0.6146 0.3672 12604 . ⎞ ⎟ 0.4661 0.2757 0.9661 ⎟ 0.9661 0.4661 2.0859 ⎟ ⎟ 0.7715 0.3685 16790 . ⎟ 0.3685 0.2142 0.7715⎟ ⎟ 16790 0.7715 3.6882 ⎠ . ⎡1 ⎢ ⎢1 ⎢1 XA = ⎢ ⎢1 ⎢1 ⎢ ⎢⎣1 . . 15417 14375 2.2083 2.4635 2.3464 ⎞ ⎛1 ⎜ ⎟ . . 2.4635 2.2083 35514 4.0671 3.6139 ⎟ ⎜ 15417 ⎜ 14375 . . 2.2083 2.3464 3.6139 35514 4.0752⎟ 1 ⎟ ( X / X )= ⎜ . . 3.6139 58617 5.9522 6.3131 ⎟ N 2 2 ⎜ 2.2083 35514 ⎜ 2.4635 4.0671 35514 ⎟ . . 5.9522 6.9096 58617 ⎜ ⎟ . . . . . . 2 3464 3 6139 4 0752 6 3131 58617 7 3106 ⎝ ⎠ ⎛1 ⎜ ⎜1 ⎜ ⎜1 XB = ⎜ 1 ⎜ ⎜1 ⎜ ⎝1 1 1 M (ς N 1 ) = ( X 1/ X 1 ) + ( X 2/ X 2 ) N N 4 1 2 3 2 3 4 1 4 3 4 3 4 9 16 9 16 7 8 7 8 49 64 49 64 5 8 1 4 5 32 25 64 3 4 1 3 4 9 16 1 2 5 4 5 8 1 4 1 1 1 9 8 9 8 81 64 5 4 3 2 25 6 11 8 2 121 64 3 2 7 4 9 4 5 4 13 8 25 16 ⎤ ⎥ ⎥ 49 ⎥ 64 1 ⎥ 16 ⎥ 1⎥ ⎥ 25 16 ⎥ ⎦ 9 4 9 4 1⎞ ⎟ ⎟ 9 ⎟ 4 ⎟ 4 ⎟ ⎟ 49 16 ⎟ ⎟ 169 64 ⎠ 81 64 Res. J. Math Stat., 4(1): 1-5, 2012 ⎛1 ⎜ ⎜ 0.6667 ⎜ 0.9375 1 ( X A/ X A ) = ⎜ 0.6016 N ⎜ ⎜ 0.4635 ⎜ . ⎝ 10339 ⎛ 1 ⎜ . ⎜ 125 ⎜ 15 . 1 ( X / X )= ⎜ . N B B ⎜ 15807 ⎜ 15885 . ⎜ ⎝ 2.3698 M (ς N 2 ) = ( . 0.6667 0.9375 0.6016 0.4635 10339 ⎞ ⎟ 0.4635 0.6016 0.4066 0.3346 0.6312⎟ ⎟ . . 0.6016 10339 0.6312 0.4066 12393 ⎟ 0.4066 0.6312 0.4071 0.2882 0.7211⎟ 0.3346 0.4066 0.2882 0.2494 0.4071⎟ ⎟ . . 0.6312 12393 0.7211 0.4071 15684 ⎠ 125 . 15885 . 15 . 15807 . 19245 2.3698 2.4176 2.5042 . 2.0316 2.4176 2.9344 2.6519 2.0508 2.5042 2.6519 2.6869 31051 3.9082 38289 . . ) ( 1 1 X A/ X A + X B/ X B N N ⎛ 2.0000 ⎜ . ⎜ 19167 ⎜ 2.4375 M (ς N 2 ) = ⎜ ⎜ 2.1823 ⎜ 2.0521 ⎜ ⎜ 3.4036 ⎝ 15885 . 19245 2.0316 2.0508 . 4.1161 REFERENCES Atkinson, A.C. and A.N. Donev, 1992. Optimum Experimental Design. Oxford Science Publication. Bamanga, M.A. and O.E. Asiribo, 2006. The Doptimality Criterion used in determination of new Optimum NPK Fertilizer rate for application on hybrid Maize in Nigeria. Int. J. Numer. Math, 1(2): Fedorov, V.V., 1972. Theory of Optimal Experiments. Acad. Press, New York. Kiefer, J., 1959. Optimum experimental designs. J. Roy. Stat.Soc. B, 21: 272-304. Kiefer, J. and J. Wolfowitz, 1959. Optimum designs in regressional problems. Ann. Math. Stat., 30: 271-274. Kiefer, J., 1974. General Equivalence theory for designs (Approximate theory). Ann. Stat., 2: 849-879. Kiefer, J. and J. Wolfowitz, 1960. The Equivalence of two Extremum Problem. Can. J. Math., 12: 363-366. Mandal, S., 2000. Construction of optimizing distributions with applications in estimation and optimal design. Ph.D. Thesis, University of Glasgow. Onukogu, I.B. and P.E. Chigbu, 2002. Super Convergent Line Series (in Optimal Design of Experiment and Mathematical Programming): AP Express Publishers, Nsukka. Pazman, A., 1986. Foundations of Optimum Experimental Design: D-Reidal Publishing Company, Boston. Pukelsheim, F., 1980. On linear regression designs with maximum information. J. Stat. Plan. Inference, 4: 339-364. Pukelsheim, F. and B. Torsney, 1991. Optimal Weights for experimental designs on linearly independent Support points. Ann. Stat, 19: 1614-1625. Pukelsheim, F., 1993. Optimal Design of Experiments. Wiley, New York. Silvey, S.D., 1980. Optimal Design. Chapmal and Hall, Country. Wald, A., 1943. On the efficient Design of Statistical investigations. Ann. Math. Stat., 14: 134-140. 2.3698⎞ ⎟ 31051 . ⎟ ⎟ 31051 . ⎟ 38289 . ⎟ 4.1161⎟ ⎟ 6.6694⎠ ) 19167 2.4375 2.1823 2.0521 3.4036 ⎞ . ⎟ 2.0521 2.5260 2.4382 2.3854 3.7367 ⎟ ⎟ 2.5260 3.4036 3.0488 2.9108 51475 . ⎟ 2.4382 3.6488 3.3416 2.9401 4.5500 ⎟ 2.3854 2.9108 2.9401 2.9364 4.5232 ⎟⎟ 3.7363 51475 4.5500 4.5232 8.2378⎟⎠ . det M(.N12) = 0.00001406 |det M(.N1)> det M(.N2) In the same manner .N1 is D-optimal since det M(.N1) is maximized, again by equivalence theorem it is also Goptimal. Therefore, the Non-overlapping segmentation of the response surface generate a design which is D- and Goptimal for a second order biased model. CONCLUSION From the foregoing, we conclude that .N1 is Doptimal and also G-optimal (by equivalence theorem given in theorem 1). That is, the Non-overlapping segmentation of the response surface forms a design which is D- and G-optimal for a first order unbiased response function and second order biased response function, respectively. 5