*********************************************************************** * * * PNEW - A BUNDLE-NEWTON ALGORITHM FOR NONSMOOTH * * UNCONSTRAINED OPTIMIZATION. * * * *********************************************************************** 1. Introduction: ---------------- The double-precision FORTRAN 77 basic subroutine PNEW is designed to find a close approximation to a local minimum of a nonlinear nonsmooth function F(X) with simple bounds on variables and general linear constraints. Here X is a vector of N variables and F(X), is assumed to be a locally Lipschitz continuous function. Simple bounds are assumed in the form X(I) unbounded if IX(I) = 0, XL(I) <= X(I) if IX(I) = 1, X(I) <= XU(I) if IX(I) = 2, XL(I) <= X(I) <= XU(I) if IX(I) = 3, XL(I) = X(I) = XU(I) if IX(I) = 5, where 1 <= I <= N. General linear constraints are assumed in the form C(I) unbounded if IC(I) = 0, CL(I) <= C(I) if IC(I) = 1, C(I) <= CU(I) if IC(I) = 2, CL(I) <= C(I) <= CU(I) if IC(I) = 3, CL(I) = C(I) = CU(I) if IC(I) = 5, where C(I) = A_I * X, 1 <= I <= NC, are linear functions. To simplify user's work, three additional easy to use subroutines are added. They call the basic general subroutine PNEW: PNEWU - unconstrained nonsmooth optimization, PNEWS - nonsmooth optimization with simple bounds, PNEWL - nonsmooth optimization with simple bounds and general linear constraints. All subroutines contain a description of formal parameters and extensive comments. Furthermore, two test programs TNEWU and TNEWL are included, which contain several test problems (see [4]). These test programs serve as examples for using the subroutines, verify their correctness and demonstrate their efficiency. In this short guide, we describe all subroutines which can be called from the user's program. A detailed description of methods is given in [2] and [3]. In the description of formal parameters, we introduce a type of the argument that specifies whether the argument must have a value defined on entry to the subroutine (I), whether it is a value which will be returned (O), or both (U), or whether it is an auxiliary value (A). Note that the arguments of the type I can be changed on output under some circumstances, especially if improper input values were given. Besides formal parameters, we can use a COMMON /STAT/ block containing statistical information. This block, used in each subroutine, has the following form: COMMON /STAT/ NDECF,NRES,NRED,NREM,NADD,NIT,NFV,NFG,NFH The arguments have the following meaning: Argument Type Significance ---------------------------------------------------------------------- NDECF O Positive INTEGER variable that indicates the number of matrix decompositions. NRES O Positive INTEGER variable that indicates the number of restarts. NRED O Positive INTEGER variable that indicates the number of reductions. NREM O Positive INTEGER variable that indicates the number of constraint deletions during the QP solutions. NADD O Positive INTEGER variable that indicates the number of constraint additions during the QP solutions. NIT O Positive INTEGER variable that indicates the number of iterations. NFV O Positive INTEGER variable that indicates the number of function evaluations. NFG O Positive INTEGER variable that specifies the number of subgradient evaluations. NFH O Positive INTEGER variable that specifies the number of Hessian evaluations. 2. Subroutines PNEWU, PNEWS, PNEWL: ----------------------------------- The calling sequences are CALL PNEWU(NF,NA,X,IPAR,RPAR,F,GMAX,IHES,IPRNT,ITERM) CALL PNEWS(NF,NA,NB,X,IX,XL,XU,IPAR,RPAR,F,GMAX,IHES, & IPRNT,TERM) CALL PNEWL(NF,NA,NB,NC,X,IX,XL,XU,CF,IC,CL,CU,CG,IPAR, & RPAR,F,GMAX,IHES,IPRNT,ITERM) The arguments have the following meaning. Argument Type Significance ---------------------------------------------------------------------- NF I Positive INTEGER variable that specifies the number of variables of the objective function. NA I Nonnegative INTEGER variable that specifies the maximum bundle dimension. The choice NA=0 causes that the default value NA=NF+3 will be taken. NB I Nonnegative INTEGER variable that specifies whether the simple bounds are suppressed (NB=0) or accepted (NB>0). NC I Nonnegative INTEGER variable that specifies the number of linear constraints; if NC=0 the linear constraints are suppressed. X(NF) U On input, DOUBLE PRECISION vector with the initial estimate to the solution. On output, the approximation to the minimum. IX(NF) I On input (significant only if NB>0) INTEGER vector containing the simple bounds types: IX(I)=0 - the variable X(I) is unbounded, IX(I)=1 - the lower bound X(I) >= XL(I), IX(I)=2 - the upper bound X(I) <= XU(I), IX(I)=3 - the two side bound XL(I) <= X(I) <= XU(I), IX(I)=5 - the variable X(I) is fixed (given by its initial estimate). XL(NF) I DOUBLE PRECISION vector with lower bounds for variables (significant only if NB>0). XU(NF) I DOUBLE PRECISION vector with upper bounds for variables (significant only if NB>0). CF(NC) A DOUBLE PRECISION vector which contains values of constraint functions (only if NC>0). IC(NC) I On input (significant only if NC>0) INTEGER vector which contains constraint types: IC(K)=0 - the constraint CF(K) is not used, IC(K)=1 - the lower constraint CF(K) >= CL(K), IC(K)=2 - the upper constraint CF(K) <= CU(K), IC(K)=3 - the two side constraint CL(K) <= CF(K) <= CU(K), IC(K)=5 - the equality constraint CF(K) = CL(K). CL(NC) I DOUBLE PRECISION vector with lower bounds for constraint functions (significant only if NC>0). CU(NC) I DOUBLE PRECISION vector with upper bounds for constraint functions (significant only if NC>0). CG(NF*NC) I DOUBLE PRECISION matrix whose columns are normals of the linear constraints (significant only if NC>0). IPAR(5) A INTEGER parameters: IPAR(1)=MIT, IPAR(2)=MFV, IPAR(3)-NONE, IPAR(4)=MTESX, IPAR(5)=MTESF. Parameters MIT, MFV, MTESX, MTESF are described in Section 3 together with other parameters of the subroutine PBUN. RPAR(6) A DOUBLE PRECISION parameters: RPAR(1)=XMAX, RPAR(2)=TOLX, RPAR(3)=TOLF, RPAR(4)=TOLB, RPAR(5)=TOLG, RPAR(6)=ETA. Parameters XMAX, TOLX, TOLF, TOLB, TOLG, ETA are described in Section 3 together with other parameters of the subroutine PBUN. F O DOUBLE PRECISION value of the objective function at the solution X. GMAX O DOUBLE PRECISION maximum absolute value of a partial derivative of the Lagrangian function. IHES I INTEGER variable that specifies a way for computing second derivatives: IHES=0 - numerical computation, IHES=1 - analytical computation by the user supplied subroutine HES. IPRNT I INTEGER variable that specifies PRINT: IPRNT= 0 - print is suppressed, IPRNT= 1 - basic print of final results, IPRNT=-1 - extended print of final results, IPRNT= 2 - basic print of intermediate and final results, IPRNT=-2 - extended print of intermediate and final results. ITERM O INTEGER variable that indicates the cause of termination: ITERM= 1 - if |X - XO| was less than or equal to TOLX in MTESX subsequent iterations, ITERM= 2 - if |F - FO| was less than or equal to TOLF in MTESF subsequent iterations, ITERM= 3 - if F is less than or equal to TOLB, ITERM= 4 - if GMAX is less than or equal to TOLG, ITERM=11 - if NFV exceeded MFV, ITERM=12 - if NIT exceeded MIT, ITERM< 0 - if the method failed. If ITERM=-6, then the termination criterion has not been satisfied, but the point obtained is usually acceptable. The subroutines PNEWU, PNEWS, PNEWL require user supplied subroutines FUNDER and HES that defines the objective function, its subgradient and has the form SUBROUTINE FUNDER(NF,X,F,G) SUBROUTINE HES(NF,X,H) The arguments of user supplied subroutines have the following meaning. Argument Type Significance ---------------------------------------------------------------------- NF I Positive INTEGER variable that specifies the number of variables of the objective function. X(NF) I DOUBLE PRECISION an estimate to the solution. F O DOUBLE PRECISION value of the objective function at the point X. G(NF) O DOUBLE PRECISION subgradient of the objective function at the point X. H(NH) O DOUBLE PRECISION matrix containing the second order information at the point X (NH is equal to NF*(NF+1)/2). 3. Subroutine PNEW: ------------------- This general subroutine is called from all the subroutines described in Section 2. The calling sequence is CALL PNEW(NF,NA,NB,NC,X,IX,XL,XU,CF,IC,CL,CU,CG,AF,IA,AFD,AG, & IAA,AR,AZ,G,H,HF,AH,S,SO,XO,GO,XMAX,TOLX,TOLF,TOLB,TOLG,ETA, & GMAX,F,MIT,MFV,MTESX,MTESF,IPRNT,IHES,ITERM). The arguments NF, NA, NB, NC, X, IX, XL, XU, CF, IC, CL, CU, CG, GMAX, F, IHES, IPRNT, ITERM, have the same meaning as in Section 2. Other arguments have the following meaning M is equal to NF*(NF+1)/2): Argument Type Significance --------------------------------------------------------------------- AF(5*NA) A DOUBLE PRECISION vector of bundle function values. IA(NA) A INTEGER vector containing types of bundle functions. AFD(NA) A DOUBLE PRECISION vector of bundle function increments. AG(NF*NA) A DOUBLE PRECISION matrix whose columns are bundle subgradients. IAA(NA) A INTEGER vector containing indices of active functions. AR(NAR) A DOUBLE PRECISION matrix containing triangular decomposition of the orthogonal projection kernel (NAR is equal to (NF+1)*(NF+2)/2). AZ(NF+1) A DOUBLE PRECISION vector of Lagrange multipliers. G(NF) A DOUBLE PRECISION subgradient of the objective function. H(NH) A DOUBLE PRECISION aggregate Hessian matrix (NH is equal to NF*(NF+1)/2). HF(NH) A DOUBLE PRECISION Hessian matrix of the objective function. AH(NA*NH) A DOUBLE PRECISION Bundle of Hessian matrices. S(NF+1) A DOUBLE PRECISION direction vector. SO(NF) A DOUBLE PRECISION auxiliary vector. XO(NF) A DOUBLE PRECISION vector which contains increments of variables. GO(NF+1) A DOUBLE PRECISION gradient of the Lagrangian function. XMAX I DOUBLE PRECISION maximum stepsize; the choice XMAX=0 causes that the default value 1.0D3 will be taken. TOLX I DOUBLE PRECISION tolerance for the change of the coordinate vector X; the choice TOLX=0 causes that the default value TOLX=1.0D-16 will be taken. TOLF I DOUBLE PRECISION tolerance for the change of function values; the choice TOLF=0 causes that the default value TOLF=1.0D-8 will be taken. TOLB I DOUBLE PRECISION minimum acceptable function value; the choice TOLB=0 causes that the default value TOLB=-1.0D60 will be taken. TOLG I DOUBLE PRECISION tolerance for the termination criterion; the choice TOLG=0 causes that the default value TOLG=1.0D-6 will be taken. ETA I DOUBLE PRECISION distance measure parameter. MIT I INTEGER variable that specifies the maximum number of iterations; the choice MIT=0 causes that the default value 200 will be taken. MFV I INTEGER variable that specifies the maximum number of function evaluations; the choice |MFV|=0 causes that the default value 500 will be taken. MTESX I INTEGER variable that specifies the maximum number of iterations with changes of the coordinate vector X smaller than TOLX; the choice MTESX=0 causes that the default value MTESX=20 will be taken. MTESF I INTEGER variable that specifies the maximum number of iterations with changes of function values smaller than TOLF; the choice MTESF=0 causes that the default value MTESF=2 will be taken. The subroutine PNEW has a modular structure. The following list contains its most important subroutines: PF1HS1 - Numerical computation of the Hessian matrix. PDDBQ2 - Determination of the descent direction using quadratic programming subroutine and bundle updating. PLQDF1 - Dual range space method for solving a quadratic programming subproblem with linear constraints (see [1]). PS1L05 - Line search using function values and derivatives. The subroutine PNEW requires the user supplied subroutine FUNDER which is described in Section 2. 4. Subroutine PLQDF1: --------------------- Since the dual range space method for special quadratic programming subproblems arising in bundle type nonsmooth optimization can be used separately in many applications (e.g. in minimax optimization), we describe the subroutine PLQDF1 in more details. The calling sequence is CALL PLQDF1(NF,NA,NC,X,IX,XL,XU,AF,AFD,IA,IAA,AG,AR,AZ, & CF,IC,CL,CU,CG,G,H,S,MFP,KBF,KBC,IDECF,ETA0,ETA2,ETA9, & EPS7,EPS9,XNORM,UMAX,GMAX,N,ITERQ) The arguments NF, NA, NC, X, IX, XL, XU, AF, CF, IC, CL, CU, CG have the same meaning as in Section 2 (only with the difference that the arguments X and AF are of the type (I), i.e. they must have a value defined on entry to PLQDF1 and they are not changed). The arguments AFD, IA, IAA, AG, AR, AZ have the same meaning as in Section 3 (only with the difference that the arguments AFD, IAA, AR, AZ are of the type (O), i.e. their values can be used subsequently). Other arguments have the following meaning: Argument Type Significance ---------------------------------------------------------------------- G(NF+1) O DOUBLE PRECISION gradient of the Lagrangian function. H(NH) U DOUBLE PRECISION Choleski decomposition of the approximate Hessian matrix (NH is equal to NF*(NF+1)/2). S(NF+1) O DOUBLE PRECISION direction vector. MFP I INTEGER variable that specifies the type of the computed point. MFP=1 - computation is terminated whenever an arbitrary feasible point is found, MFP=2 - computation is terminated whenever an optimum feasible point is found, MFP=3 - computation starts from the previously reached point and is terminated whenever an optimum feasible point is found. KBF I INTEGER variable that specifies simple bounds on variables. KBF=0 - simple bounds are suppressed, KBF=1 - one sided simple bounds, KBF=2 - two sided simple bounds. KBC I INTEGER variable that specifies general linear constraints. KBC=0 - linear constraints are suppressed, KBC=1 - one sided linear constraints, KBC=2 - two sided linear constraints. IDECF U INTEGER variable that specifies the type of matrix decomposition. IDECF= 0 - no decomposition, IDECF= 1 - Choleski decomposition, IDECF= 9 - inversion, IDECF=10 - diagonal matrix. ETA0 I DOUBLE PRECISION machine precision (the recommended value is 1.0D-15. ETA2 I DOUBLE PRECISION tolerance for positive definiteness in the Choleski decomposition. ETA9 I DOUBLE PRECISION maximum floating point number. EPS7 I DOUBLE PRECISION tolerance for linear independence of constraints (the recommended value is 1.0D-10). EPS9 I DOUBLE PRECISION tolerance for the definition of active constraints (the recommended value is 1.0D-8). XNORM O DOUBLE PRECISION value of the linearized minimax function. UMAX O DOUBLE PRECISION maximum absolute value of the negative Lagrange multiplier. GMAX O DOUBLE PRECISION infinity norm of the gradient of the Lagrangian function. N O INTEGER dimension of a manifold defined by active constraints. ITERQ O INTEGER variable that indicates the type of the computed feasible point. ITERQ= 1 - an arbitrary feasible point was found, ITERQ= 2 - the optimum feasible point was found, ITERQ=-1 - an arbitrary feasible point does not exist, ITERQ=-2 - the optimum feasible point does not exist. 5. Form of printed results: --------------------------- The form of printed results is specified by the parameter IPRNT as is described in Section 3. Here we demonstrate individual forms of printed results by the simple use of the program TNEWU described in the next section (with NEXT=16). If we set IPRNT=1, then the printed results will have the form NIT= 12 NFV= 14 NFG= 14 F= -.84140833D+00 G= .6734D-06 ITERM= 4 If we set IPRNT=-1, then the printed results will have the form EXIT FROM PNEW : NIT= 12 NFV= 14 NFG= 14 F= -.84140833D+00 G= .6734D-06 ITERM= 4 X= -.1262566D+00 -.3437830D-01 -.6857198D-02 .2636066D-01 .6729492D-01 -.2783995D+00 .7421866D-01 .1385240D+00 .8403122D-01 .3858031D-01 If we set IPRNT=2, then the printed results will have the form ENTRY TO PNEW : NIT= 0 NFV= 1 NFG= 1 F= .00000000D+00 G= .1000D+61 NIT= 1 NFV= 3 NFG= 3 F= .53370664D+04 G= .1200D+05 NIT= 2 NFV= 4 NFG= 4 F= .66499712D+02 G= .2610D+02 NIT= 3 NFV= 5 NFG= 5 F= .33934270D+02 G= .3771D+02 NIT= 4 NFV= 6 NFG= 6 F= .12040341D+01 G= .3214D+01 NIT= 5 NFV= 7 NFG= 7 F= .51324695D+00 G= .1459D+01 NIT= 6 NFV= 8 NFG= 8 F= -.76915236D+00 G= .2347D+01 NIT= 7 NFV= 9 NFG= 9 F= -.83859100D+00 G= .2683D+00 NIT= 8 NFV= 10 NFG= 10 F= -.84140726D+00 G= .2491D-02 NIT= 9 NFV= 11 NFG= 11 F= -.84140726D+00 G= .3213D-03 NIT= 10 NFV= 12 NFG= 12 F= -.84140726D+00 G= .4862D-03 NIT= 11 NFV= 13 NFG= 13 F= -.84140726D+00 G= .5265D-05 NIT= 12 NFV= 14 NFG= 14 F= -.84140833D+00 G= .6734D-06 EXIT FROM PNEW : NIT= 12 NFV= 14 NFG= 14 F= -.84140833D+00 G= .6734D-06 ITERM= 4 If we set IPRNT=-2, then the printed results will have the form ENTRY TO PNEW : NIT= 0 NFV= 1 NFG= 1 F= .00000000D+00 G= .1000D+61 NIT= 1 NFV= 3 NFG= 3 F= .53370664D+04 G= .1200D+05 NIT= 2 NFV= 4 NFG= 4 F= .66499712D+02 G= .2610D+02 NIT= 3 NFV= 5 NFG= 5 F= .33934270D+02 G= .3771D+02 NIT= 4 NFV= 6 NFG= 6 F= .12040341D+01 G= .3214D+01 NIT= 5 NFV= 7 NFG= 7 F= .51324695D+00 G= .1459D+01 NIT= 6 NFV= 8 NFG= 8 F= -.76915236D+00 G= .2347D+01 NIT= 7 NFV= 9 NFG= 9 F= -.83859100D+00 G= .2683D+00 NIT= 8 NFV= 10 NFG= 10 F= -.84140726D+00 G= .2491D-02 NIT= 9 NFV= 11 NFG= 11 F= -.84140726D+00 G= .3213D-03 NIT= 10 NFV= 12 NFG= 12 F= -.84140726D+00 G= .4862D-03 NIT= 11 NFV= 13 NFG= 13 F= -.84140726D+00 G= .5265D-05 NIT= 12 NFV= 14 NFG= 14 F= -.84140833D+00 G= .6734D-06 EXIT FROM PNEW : NIT= 12 NFV= 14 NFG= 14 F= -.84140833D+00 G= .6734D-06 ITERM= 4 X= -.1262566D+00 -.3437830D-01 -.6857198D-02 .2636066D-01 .6729492D-01 -.2783995D+00 .7421866D-01 .1385240D+00 .8403122D-01 .3858031D-01 6. Verification of the subroutines: ----------------------------------- Subroutine PNEWU can be verified and tested using the program TNEWU. This program calls the subroutines TIUD19 (initiation), TFFU19 (function evaluation) and TFGU19 (subgradient evaluation) containing 20 academic test problems with at most 50 variables [4]. The results obtained by the program TNEWU on a PC computer with Microsoft Power Station Fortran compiler have the following form. NIT= 58 NFV= 59 NFG= 59 F= 0.134104250E-18 G= 0.895E-05 ITERM= 2 NIT= 7 NFV= 8 NFG= 8 F= 0.167657006E-10 G= 0.579E-05 ITERM= 4 NIT= 15 NFV= 17 NFG= 17 F= 1.95222450 G= 0.304E-03 ITERM= 4 NIT= 10 NFV= 11 NFG= 11 F= 2.00000682 G= 0.216E-04 ITERM= 4 NIT= 14 NFV= 15 NFG= 15 F= -2.99999999 G= 0.676E-04 ITERM= 4 NIT= 3 NFV= 5 NFG= 5 F= 7.20000000 G= 0.000E+00 ITERM= 4 NIT= 16 NFV= 17 NFG= 17 F= -1.41421356 G= 0.565E-07 ITERM= 4 NIT= 11 NFV= 13 NFG= 13 F= -1.00000000 G= 0.416E-07 ITERM= 4 NIT= 10 NFV= 11 NFG= 11 F= -1.00000000 G= 0.456E-06 ITERM= 4 NIT= 24 NFV= 25 NFG= 25 F= -7.99999997 G= 0.322E-02 ITERM= 4 NIT= 13 NFV= 15 NFG= 15 F= -44.0000000 G= 0.421E-05 ITERM= 4 NIT= 7 NFV= 8 NFG= 8 F= 22.6001727 G= 0.126E-02 ITERM= 4 NIT= 22 NFV= 24 NFG= 24 F= -32.3486790 G= 0.341E-02 ITERM= 4 NIT= 83 NFV= 84 NFG= 84 F= -2.91970018 G= 0.108E-02 ITERM= 4 NIT= 116 NFV= 123 NFG= 123 F= 0.559813085 G= 0.240E-05 ITERM= 4 NIT= 12 NFV= 14 NFG= 14 F=-0.841408335 G= 0.673E-06 ITERM= 4 NIT= 68 NFV= 72 NFG= 72 F= 9.78577208 G= 0.107E-03 ITERM= 4 NIT= 40 NFV= 42 NFG= 42 F= 16.7038553 G= 0.178E+00 ITERM= 4 NIT= 36 NFV= 37 NFG= 37 F= 0.383737024E-08 G= 0.609E-08 ITERM= 2 NIT= 24 NFV= 25 NFG= 25 F= 0.452894273E-08 G= 0.100E-07 ITERM= 2 NITER = 589 NFVAL = 625 NSUCC = 20 TIME= 0:00:00.05 The rows corresponding to individual test problems contain the number of iterations NIT, the number of function evaluations NFV, the number of subgradient evaluations NFG, the final value of the objective function F, the value of the criterion for the termination G and the cause of termination ITERM. Subroutine PNEWL can be verified and tested using the program TNEWL. This program calls the subroutines TIUD22 (initiation), TAFU22 (function evaluation), TAGU22 (subgradient evaluation) containing 15 academic test problems with at most 20 variables [4]. The results obtained by the program TNEWL on a PC computer with Microsoft Power Station Fortran compiler have the following form. NIT= 6 NFV= 7 NFG= 7 F=-0.389659516 G= 0.165E-07 ITERM= 4 NIT= 2 NFV= 11 NFG= 11 F=-0.330357143 G= 0.111E-15 ITERM= 4 NIT= 37 NFV= 38 NFG= 38 F=-0.448910785 G= 0.134E-06 ITERM= 4 NIT= 9 NFV= 10 NFG= 10 F=-0.429280609 G= 0.311E-04 ITERM= 4 NIT= 28 NFV= 29 NFG= 29 F= -1.85961870 G= 0.741E-07 ITERM= 4 NIT= 9 NFV= 10 NFG= 10 F= 0.101830889 G= 0.114E-06 ITERM= 4 NIT= 69 NFV= 70 NFG= 70 F= 0.813571432E-12 G= 0.737E-05 ITERM= 4 NIT= 15 NFV= 16 NFG= 16 F= 24.3062091 G= 0.130E-06 ITERM= 4 NIT= 43 NFV= 45 NFG= 45 F= 133.728305 G= 0.409E-03 ITERM= 4 NIT= 94 NFV= 98 NFG= 98 F= 0.506947996 G= 0.407E-06 ITERM= 4 NIT= 1148 NFV= 1192 NFG= 1192 F= 0.282340278E-03 G= 0.315E-01 ITERM= 2 NIT= 1239 NFV= 1241 NFG= 1241 F= -1768.80243 G= 0.205E-02 ITERM= 2 NIT= 71 NFV= 80 NFG= 80 F= 1227.22608 G= 0.635E-02 ITERM= 2 NIT= 54 NFV= 55 NFG= 55 F= 7049.24803 G= 0.949E-04 ITERM= 2 NIT= 1382 NFV= 1385 NFG= 1385 F= 174.869642 G= 0.322E-01 ITERM= 2 NITER = 4206 NFVAL = 4287 NSUCC = 15 TIME= 0:00:00.41 References: ----------- \bibitem{luv} LukĻan L., Vl\v cek J. A Bundle-Newton Method for Nonsmooth Unconstrained Minimization. To appear in Mathematical Programming. [1] Luksan L.: Dual Method for Solving a Special Problem of Quadratic Programming as a Subproblem at Linearly Constrained Nonlinear Minimax Approximation. Kybernetika 20 (1984) 445-457. [2] Luksan L., Vlcek J.: A Bundle-Newton Method for Nonsmooth Unconstrained Minimization. Mathematical Programming. [3] Luksan L., Vlcek J.: NDA: Algorithms for Nondifferentiable Optimization. Research Report V-797, Institute of Computer Science, Academy of Sciences of the Czech Republic, Prague, Czech Republic, 2000. [4] Luksan L., Vlcek J.: Subroutines for Testing Nonsmooth Unconstrained and Linearly Constrained Optimization Problems. Research Report V-798, Institute of Computer Science, Academy of Sciences of the Czech Republic, Prague, Czech Republic, 2000.