Similar documents
RADIUS OF CURVATURE AND EVOLUTE OF THE FUNCTION y=f(x)

4.1. Title: data analysis (systems analysis) Annotation of educational discipline: educational discipline includes in itself the mastery of the

( ) = ( ) = {,,, } β ( ), < 1 ( ) + ( ) = ( ) + ( )

Bachelor of Science or Arts Degree Minor Environmental Science Check List

FINANCIAL SERVICES BOARD INSURANCE DEPARTMENT

FINANCIAL SERVICES BOARD INSURANCE DIVISION

Chapter 3 Mathematics of Finance


12: Analysis of Variance. Introduction

SOLUTIONS. Would you like to know the solutions of some of the exercises?? Here you are

NILES COMMUNITY SCHOOLS DUAL ENROLLMENT APPLICATION

3.1. Solving linear equations. Introduction. Prerequisites. Learning Outcomes. Learning Style

Loans Practice. Math 107 Worksheet #23

The PCB Terminal Blocks Model Pitch(mm) Model Pitch Model Pitch Model Pitch DG300 DG360 DG166 DG500 DG301 DG365 DG103 DG500H DG305 DG332K DG128

SMS/Text Message Solution User Guide. How to send bulk SMS messages. Copyright 2013 xrm Consultancy Limited

A step-by-step guide to non-linear regression analysis of experimental data using a Microsoft Excel spreadsheet

CORRELATION ANALYSIS

STATISTICS PROJECT: Hypothesis Testing

Recommend Continued CPS Monitoring. 63 (a) 17 (b) 10 (c) (d) 20 (e) 25 (f) 80. Totals/Marginal

Grade Boundaries. Edexcel International AS/A level

Chapter 4: Nominal and Effective Interest Rates

Financial Mathematics for Actuaries. Chapter 1 Interest Accumulation and Time Value of Money

9-17a Tutorial 9 Practice Review Assignment

FE3A 04 (CFACSD11) Lead a Team to Improve Customer Service

Radicals - Multiply and Divide Radicals

Math Abstract Algebra I Questions for Section 23: Factoring Polynomials over a Field

With compound interest you earn an additional $ ($ $1500).

Solving Compound Interest Problems

2 The Mathematics. of Finance. Copyright Cengage Learning. All rights reserved.

Statistical Models in R

Math Numerical Analysis Homework #2 Solutions

An Introduction to Calculus. Jackie Nicholas

3 3RG78 45 program overview

Salary adjustment method

Lesson 4 Annuities: The Mathematics of Regular Payments

A characterization of trace zero symmetric nonnegative 5x5 matrices

1. Let X and Y be normed spaces and let T B(X, Y ).

On Stock Trading Via Feedback Control When Underlying Stock Returns Are Discontinuous

N C P S E N. y p e W A SV C. Maximum Salary. Minimum Salary. Class Code

Functional Principal Components Analysis with Survey Data

Microeconomic Theory: Basic Math Concepts

CONDUCTOR SHORT-CIRCUIT PROTECTION

Linda Staub & Alexandros Gekenidis

Title Location Date Start Time End Time Description

INFINITE DIMENSIONAL RESTRICTED INVERTIBILITY

Confidence Intervals for the Difference Between Two Means

Review of Basic Options Concepts and Terminology

BUSI 121 Foundations of Real Estate Mathematics

第 9 讲 : 股 票 期 权 定 价 : B-S 模 型 Valuing Stock Options: The Black-Scholes Model

the recursion-tree method

Cost Models for Vehicle Routing Problems Stanford Boulevard, Suite 260 R. H. Smith School of Business

GRAVITATIONAL FIELDS PHYSICS 20 GRAVITATIONAL FORCES. Gravitational Fields (or Acceleration Due to Gravity) Symbol: Definition: Units:

Asymptotics for ruin probabilities in a discrete-time risk model with dependent financial and insurance risks

Constrained optimization.

Terminal Blocks 9080LBA Power Distribution Blocks Copper or Aluminum Wire

ORTHOGONAL POLYNOMIAL CONTRASTS INDIVIDUAL DF COMPARISONS: EQUALLY SPACED TREATMENTS

Algebra 1 Course Information

Black and Scholes - A Review of Option Pricing Model

SOLVING EQUATIONS WITH EXCEL

Duplicating and its Applications in Batch Scheduling

Chapter 22: Borrowings Models

GST-15_Bad Debt. What is 6 Months Bad Debt Relief?

Daten. Masterplan Report Manager. Description. ASEKO GmbH. Version 2.0

Chapter 21: Savings Models

MULTIVARIATE PROBABILITY DISTRIBUTIONS

Reciprocal Cost Allocations for Many Support Departments Using Spreadsheet Matrix Functions

Costs. Accounting Cost{stresses \out of pocket" expenses. Depreciation costs are based on tax laws.

12.S-[F]NPW-02 June All Syllabus Science Dept. College M.Sc. Comp.Sci. & I.T. Ist & IInd Yr

Using Formulas, Functions, and Data Analysis Tools Excel 2010 Tutorial


CHAPTER 5 PREDICTIVE MODELING STUDIES TO DETERMINE THE CONVEYING VELOCITY OF PARTS ON VIBRATORY FEEDER

Biostatistics: DESCRIPTIVE STATISTICS: 2, VARIABILITY

Transcription:

Appliedand Computational NISTIR5916 Mathematics Division ComputingandAppliedMathematicsLaboratory ServiceforSpecialFunctions AProposedSoftwareTest DanielW.Lozier October1996 NationalInstituteofStandardsandTechnology U.S.DEPARTMENTOFCOMMERCE TechnologyAdministration Gaithersburg,MD20899

ThispaperwillappearinnalforminTheQualityofNumericalSoftware: AssessmentandEnhancement,RonaldF.Boisvert,editor,Chapman&Hall, PREPRINT London,1997.ItwaspresentedorallyattheIFIP/TC2/WG2.5WorkingConference7,heldatSt.Catherine'sCollege,Oxford,England,July7{13,1996.

ThisisaproposaltodevelopasoftwaretestserviceattheNationalInstituteof StandardsandTechnologyforuseintestingtheaccuracy,ornumericalprecision, ofmathematicalsoftwareforspecialfunctions.theservicewouldusetheworld ABSTRACT willbeofpracticalutilitytoanyonewhousesspecialfunctionsinphysicsorother runonanetworkofworkstationsattheinstitute.itishopedthatsuchaservice WideWebtoreceivetestrequestsandreturntestresults.Thetestswouldbe applications,andthatitwillstimulatetheinterestofappliedmathematicians whoareinterestedinthecomputationofspecialfunctionsaswellascomputer scientistswhoareinterestedininnovativeusesoftheinternet.theauthor solicitscommentsonanyaspectoftheproposedservice.

1Mathematicalsoftwareisdeeplyembeddedinthecomputingenvironment.Since thisenvironmentisevolvingrapidly,itsimpactonmathematicalsoftwareneeds toberevisitedregularly. Introduction algorithms,particularlyincomputationallinearalgebra.theearlierintroduc- highbandwidthcommunicationlinks,oneresultofwhichisimprovedaccess tionofvectorcomputershadasimilareect.recentadvancesincommunica- tionsandnetworkinghaveledtotheglobalinterconnectionofcomputersby Progressinparallelcomputinghasstimulatedmuchreworkingofnumerical Web.Forexample,electroniccatalogsandrepositoriessuchasxnetlib[8]are toinformationaboutmathematicalsoftwareviatheinternetandworldwide nowconsultedroutinelyforhelpinlocatingandobtainingmathematicalsoftwarepackagesforspecictasks. [4]and[10].However,mathematicalfunctionsseemparticularlyappropriatefor putationofmathematicalfunctions.somereferencescanbecited,forexample demonstratinganewandpotentiallyvaluableuseoftheweb:mathematical softwaretesting.theproblemoftestingisintrinsicallysimplerformathematicalfunctionsthanforotherkindsofnumericalcomputation.theinputand outputeuclideanspaceshavelowdimension.incontrast,numericallinearalgebradealswitheuclideanspacesofhighdimension,andmostothernumerical Vectorandparalleldevelopmentshavehadonlyamodestimpactonthecom- formathematicalfunctionscanbedevisedthatapply,intheory,toallpossible computationsdealwithfunctionspacesofinnitedimension.testprocedures vantages,comparedtocurrenttestingpractice,inusingthewebtotestmath- ematicalsoftware?ouransweristhattestscanbetailoredtosuitaparticular inputs. need,andtheycanbeperformedondemand. Aquestionthatneedsansweringatthispointis:Whatwouldbethead- editorsandsoftwaremaintainersalsoplayarole.thesepeoplehavedirectresponsibilityforthecorrectnessofthesoftware.forcommercialsoftware,high licensefeesarejustiedlargelybythehighcostsassociatedwithsoftwaremain- Toseewhythisisuseful,wedividecurrentpracticeintotwocategories. Suppliertestingisperformedbythesoftwarewriterorprojectteam.Referees, timesindependenttestsareconductedandpublishedinjournalarticlesand institutionalreportsasaguideforprospectiveusers. itleadstoanincreasedcondenceinthecorrectnessofthesoftware.some- groups.usersoftenperformthiskindoftestingfortheirownpurposesbecause tenanceandtesting.independenttestingisperformedbyotherindividualsand testsarenevercomplete,theirresultsmaynotapplydirectlytothenumerical havingbeenperformedatsometimeinthepastinacomputingenvironment thatmaybeverydierentfromtheprevailingone.evenmoreunsettling,since computationofcurrentinterest. Publishedtestsofeitherkindhavea`frozen-in-time'qualityaboutthem, 1

Herewearefollowingtheconventionalterminologyofcallingthetranscendental ticalbecausethetestprogramisunavailable.notableexceptionsarethetest andcomplexarguments,andofcody[6]forspecialfunctionsofrealargument. programsofcodyandwaite[7]andcody[5]forelementaryfunctionsofreal Justthesimplerepetitionofamathematicalfunctiontestisoftenimprac- functionsmetincalculuscourseselementaryfunctionsandthehigherfunctions thatappearinadvancedapplicationsspecialfunctions.thegeneralunavailabilityoftestprogramsisundoubtedlyrelatedtotheconsiderableeortthatis ofrandomnessincluded. someimplementationaldetailsaregiven.theemphasisisonspecialfunctions, programsapplyonlytoabuiltinsetoftestarguments,oftenwithanelement distribution.anotherproblemislackofgenerality.forexample,mosttest requiredtoraisethemtoanacceptablyhighstandardforpublicationorpublic becausethisiswheretheneedisgreatest,buttheservicewillapplyequallywell tailorteststospecicrequirements.thereforeitshouldbeofinterestinboth toelementaryfunctions.theservicewillprovideatoolthatcanbeusedto Inthispaperasoftwaretestserviceforspecialfunctionsisproposedand supplierandindependenttesting. 2Thepurposeoftheproposedsoftwaretestingserviceforspecialfunctionsisto assesstheaccuracy,ornumericalprecision,ofcomputedfunctionvaluesthrough theuseofacomparisonmethod.testvalueswillbecomparedagainstreference Proposal andtechnology.thetestswillbeconductedattheinstituteusingsoftware developedforthepurpose.thetestresultswillbereturnedtotherequesterin valuescomputedinhigherprecisionbyhighlyaccuratealgorithms.testrequestswillbesubmittedtoawebserveratthenationalinstituteofstandards theformofanappropriatedocumentonthewebserver. ReferenceSoftwareThiswillconsistofhighlyaccurateandreliable,butnot Thekeycomponentsoftheservicewillbe referencevaluesofspecialfunctionsoververyextensiveargumentdomains. necessarilyecient,numericalproceduresforgeneratinghigh-precision rithmsbecauseitwillbeembeddedinacomputingenvironmentthatmit- igatesthecomputerarithmeticliabilities(underow,overow,andlimited Thereferencesoftwarewillbeanexcellentrepositoryforadvancedalgo- ComparisonSoftwareThiswillservethepurposeoforchestratingthegenerationofreferencevaluesanddeterminingtheprecisionoftestvalues.The precision)ofconventionalcomputingenvironments. comparisonsoftwarewillutilizeparallelmethodsviathesimpledeviceof domainpartitioning.anappropriatemeasureofprecisionwillbedened intermsofintervalmathematics. 2

CommunicationInterfaceThiswillbeanappropriatelydesignedWebdoc- alternativeapproach,numericalvericationofidentities,isadvocatedandused Thecomparisonmethodwaschosenbecauseofitsconceptualsimplicity.An umentwithassociatedsubdocumentsforacceptingtestrequestsandre- turningtestresultsviatheinternet. byw.j.codyandhiscoworkers.itavoidstheneedforhigherprecisionbutit algorithmusedintheimplementationofthefunction.also,caremustbetaken conclusionsthatcouldariseiftheidentitywerenotentirelyindependentofthe requirescarefulattentioninthechoiceofidentitytoguaranteeagainstincorrect toseparatetheerrorinthefunctionevaluationfromtheerrorintheevaluation oftheidentity.thesecomplicationswillbeavoidedinthetestservicebytaking andcommunication. 3fulladvantageofthetremendouspowerofcurrentcapabilitiesforcomputation forspecialfunctions.notonlymustitbehighlyaccurate,adeniteboundon theerrorineachcomputedreferencevalueisessential.otherwise,noonecan Thereferencesoftwareisattheheartoftheproposedsoftwaretestingsystem ReferenceSoftware becertainoftheresultsofatest.forthisreason,thereferencesoftwareshould bewrittenusingintervaltechniques.anintroductiontointervalcomputations errorbounds. isgiveninthebookbyalefeldandherzberger[2].however,asidefromthe numericalanalyststodevelopintervalalgorithmsthatgeneratetherequired specicmathematicalfunctions.anopportunityandaneedexistsherefor elementaryfunctions,verylittlehasbeenpublishedonintervalalgorithmsfor software.thusitisappropriatetowritethereferencesoftwareinmultiple precision.thefortranpackageofbailey[3]isavailableandapplicableforthis puterarithmeticsystems,bailey'spackagerelievestheneedtobecarefulabout purpose.becauseofitsvastexponentrangeincomparisontoconventionalcom- Theservicemustbeabletotestdouble-precisionaswellassingle-precision underowandoverow.theoccurrenceoftheseconditionscancompletelyinvalidateanotherwisepristinecomputation.thealgorithmswilltakefullyinto considerationstabilityandroundoquestionsbecausethesetoo,ifignored,can destroyacomputation. precision-limitedbecauseitemployspolynomialorrationalapproximationsthat areconstructedwithrespecttoaxedtargetprecision.flexibilityismoreimportantthaneciencyforreferencesoftware.ideally,referencealgorithmswill Highlyecientsoftware,atleastforfunctionsofonevariable,isusually acceptanarbitrarytolerancespecicationsothatthesameprogramscanbe executedinincreasedprecisionwithoutamajoreorttogenerateapproximationcoecientsforthehigherprecision.thismeansthatmethodswillbe 3

constructedfromtaylorexpansions,asymptoticexpansions,dierentialordifferenceequations,integralrepresentations,andotheranalyticalproperties,just asisdoneinmuchexistingsoftwareforfunctionsoftwoormorevariables. Thecomparisonsoftwarehasamathematicalcomponentandacomputersciencecomponent.Themathematicalcomponentisconcernedwithmeasuring ComparisonSoftware 4theerrorintestvalues.Thiscouldbedonesimplywithpointwiseabsoluteor relativeerrorbutanintervalformulationismoreappropriate.thecomputer ues.thisisanaturalapplicationforparallelprocessingwithalooselycoupled sciencecomponentisneededtocollectandprocessthetestandreferenceval- networkofcomputerworkstations. describe,atleastwhenallvariablesarereal.letusconsiderafunction Onlythemathematicalcomponentwillbeconsideredhere.Itiseasyto whererdenotesthesetofrealnumbers.letfbethesetofrealnumbers thatarerepresentableexactlyintheformatofaparticularcomputerarithmetic system,excludinganynonnumericalsymbolicrepresentationssuchas1,0 y=f(x); x2rm; y2r; (1) andnan(not-a-number).thusanapproximatingfunction isdenedbythesoftwaretobetested.ourproblemistomeasuretheerror committedwhen~yistakenasanapproximationtoy. Thepointwiseabsoluteerror,denedforx2Fm,isjustjy?~yj.Because ~y=~f(x); x2fm; ~y2f (2) absoluteerrorisnaturallyassociatedwithxed-pointcomputation,andnot tionf.insteadofrelativeerror=j(~y?y)=yj,weprefertouserelativeprecision oating-point,relativeerrorismoreappropriateexceptnearzerosofthefunc- thisdenitionwasintroducedin[13].sincerp(y;~y)=+o(2),relative precisionandrelativeerrorarenearlythesamewhen~yisagoodapproximation rp(y;~y)=jln(~y=y)jify~y>0; undenedotherwise; (3) (0;1)and(?1;0),respectively. toy.butrelativeprecisionhastheadvantagefordetailederroranalysesthatit isametriconr+andr?,wherethesesymbolsdenotetheopenrealintervals Y=[y`;yu]wherey`;yuaretwoconsecutiveelementsofF.Acriterionthat approximatefunctionvalue~y=~f(x)tosatisfyeither~y=y`or~y=yu.this isappliedsometimesintheconstructionofcomputersoftwareistorequirethe Givenx2Fm,theexactfunctionvaluey=f(x)determinestheinterval 4

willbecalledthecriterionoffullprecision.itcanbeexpressedinadierent way.firstwedenethemachineepsilon wheret+denotesthesuccessoroftinf.thentheapproximatingfunction~f satisesthecriterionoffullprecisionif,andonlyif,rp(y;~y)forallx2fm =max t2frp(t;t+): (4) suchthatf(x)and~f(x)havethesamesign.itiscustomarytoemployfull precision,ornearlyfullprecision,insoftwareforelementaryfunctions. plementationsofoating-pointarithmeticaswellasbysomesoftwareforele- mentaryfunctions,particularlywhensuppliedwithfortrancompilers. Thestrongestpossiblecriterioniscorrectrounding.Requiredby[9]for standardoating-pointarithmeticoperations,itismetbymostup-to-dateim- denoteitspredecessorandsuccessor(wheretheorderingisdenedcomponentwise).ifweregardxasarepresentativeofthemultivariateinterval Thecriterionoffullprecisionisquiterigorous.Ifx2Fm,letx?andx+ thentheuncertaintyinxisreectedintherangeoffasitsargumentsvary unnecessarytorequirefullprecisionin~f(x).infact,thecomputationof~f(x) throughoutx.ifapartialderivativeoffislarge,itcanbearguedthatitis X=[x`;xu]=12[x+x?;x+x+]; (5) tofullprecisionisunwarrantedifitrequiresaninordinateamountofexecution setofallclosedintervalsubsetsofr.alefeldandherzberger[2]showthat,if time.thispenaltyislikelytobeespeciallysevereforspecialfunctions. A=[a`;au]andB=[b`;bu]aretwointervals,thenthefunction Amoreappropriatecriterionofprecisioncanbedened.LetI(R)bethe isametric.also,sinceq([a;a];[b;b])=ja?bj,themetricqgeneralizestheusual metricinr.arithmeticoperationsa+b;a?b;abanda=baredenedin I(R)byoperatingontheendpointsoftheintervals.Theyarecontinuousinthe q(a;b)=maxfja`?b`j;jau?bujg; A;B2I(R) (6) topologyoffi(r);qg.similarly,itispossibletodenecontinuousintervalextensionsofcontinuousrealfunctions.forexample,forthelogarithmicfunction, theintervalextensionln(a)=[lna`;lnau]isdenedandcontinuousoni(r+). Next,wedeneintervalrelativeprecision Thisiseasytocompute,sinceitcanbeshownthat rp(y;~y)=8<:q(lny;ln~y)ify;~y2i(r+); undened rp(?y;?~y)ify;~y2i(r?); otherwise: (7) rp(y;~y)=maxfrp(y`;~y`);rp(yu;~yu)g: 5 (8)

IntervalrelativeprecisionisametriconI(R+)andI(R?),anditgeneralizes pointwiserelativeprecisionsincerp([y;y];[~y;~y])=rp(y;~y). multivariateintervalx=[x`;xu],andassumethefunctionfiscontinuouson X.LetYbetherangeoffonX: Nowconsiderthetestargumentx2Fm,againasarepresentativeofthe Finally,letthetestfunctionvalue~y=~f(x)2Frepresenttheinterval Y=[y`;yu]=f(X)=ff(x)jx`xxug: ~Y=[~y`;~yu]=12[~y+~y?;~y+~y+]: (10) (9) Thenwewillsaythattheapproximatingfunction~fsatisestheintervalcriterionofprecisionif forallxsuchthattherelativeprecisionsaredened.therightsideofthis inequalityprovidesastandardofcomparison.ittakesintoaccountthebehavior offasitsargumentsvarythroughouttheneighborhoodrepresentedbyx.it rp(y;~y)maxf;rp(y`;yu)g (11) establishestheallowablerangeofrelativeerrorsoverthisneighborhood.the leftsidemeasuresthedistancebetweentheallowablerangeoffandtheinterval representedbythetestfunctionvalue.iftheintervalcriterion(11)issatised, satised.inallcaseswhen(11)issatised,asimpleinterpretationintermsof thenthesetintersectiony\~yisnonempty.if~yyory~y,then(11)is pointwiserelativeerrorcanbegiven.thiswillbediscussedinafuturepaper. errorscausedbytruncatinginniteprocesses.thisproblemwillneedtobe estimatedsubstantially.also,itisnecessarytoconstructstrictboundsforall doesnotnecessarilyproducetherange;tothecontrary,therangemaybeover- ofrealfunctions.evaluationofexplicitexpressionsusingintervalarithmetic Afundamentalprobleminintervalmathematicsishowtocomputetherange facedinthedesignandconstructionofreferencesoftwareforthesoftwaretest service. 5function.Foreachtest,thetestrequesterprovidesanargumentsettogether Forthesoftwaretestsystem,anargumentsetisasubsetofthedomainofa withcorrespondingfunctionvaluestothecommunicationinterface.thenthe CommunicationInterface testrequester'sfunctionvalues,andnallythecommunicationinterfacereturns argumentset,thecomparisonsoftwarecomparesthereferencevaluesagainstthe atableorplotoftheintervalrelativeprecisiontothetestrequester. referencesoftwarecomputesthefunctiontohigherprecisionatallpointsinthe sionprocessesbetweenarbitrarybasesareconsideredindetailinmatula[11] processesofdecimal-to-binaryandbinary-to-decimalconversion.baseconver- Acarefuldevelopmentofthesoftwaretestservicerequiresattentiontothe 6

and[12].ap-digit,base-signicancespaceisthesetspofallp-digitnormalizedoating-pointnumbersinthebase,excludingzeroandwithoutregardto size.letspandsqbetwosignicancespaces.theroundingconversionmappingrqfromspintosqisthemappingthatisdenedbyconvertingx2sp intoits-aryexpansiontosucientlyhighprecision,thenroundingittoq base-digits.thetruncationconversionmappingtqisdenedsimilarly.the thenbacktosp.matulaprovedtwotheorems: compositionisanin-and-outconversionmappingwhichmapsspintosq,and Theorem1(BaseConversionTheorem)Ifi6=jforanypositiveintegersi;j,thenthebaseconversionmappingsRq:Sp!SqandTq:Sp!Sq are:1.one-to-oneontotheirrangesifandonlyifq?1p?1; Theorem2(In-and-OutConversionTheorem)Ifi6=jforanypositiveintegersi;j,then 2.ontoifandonlyifp?1q?1. compositionofbaseconversionmappingsispossible.aninterestingkindof 2.RpTq:Sp!Spistheidentityifandonlyifq?12p?1. 1.RpRq:Sp!Spistheidentityifandonlyifq?1>p,and whenthebasesandareintegralpowersofacommonbase.underthe conditionsoftheorem2,rqandtqareone-to-oneontotheirrangesandtheir inversemappingscoincidewithrp:sq!sp. Theconditioni6=jforanypositiveintegersi;jexcludesthetrivialcase arithmeticasdenedin[9].then=2,p=24,=10,andqistobe determinedaccordingtosomecriterion.theroundingconversionmapping Rq10:S24 Asanexample,considerdecimaloutputfromsingle-precisioncomputer 1.one-to-oneontoitsrangeifandonlyifq9; 2.ontoifandonlyifq6; 2!Sq10is exceed6digitsifthecompletesetofq-digitdecimalnumbersistobecovered. similarlyfortq10.thusdecimaloutputprecisionqneednotexceed9digitsif eachinternalnumberistohaveauniquedecimalrepresentation,anditcannot Also,eitherofR24 computerarithmeticf.letsq10bethedecimalsignicancespacewithminimumqsuchthatthenecessaryandsucientconditioninpart1oftheorem2 issatised.finally,letsp0 NowletSpdenotethesignicancespaceassociatedwiththetestrequester's 2Rq10orR24 2Tq10istheidentitymappingifandonlyifq9. 0denotethesignicancespaceassociatedwiththe 7

referencesoftware.weassume0andarepositiveintegralpowersoftwo,and weassumep0issuchthatspsp0 representanargumentsetinbinaryordecimal.ifwechoosebinary,thebase conversionmappingfromsptosp0 ThentheIn-and-OutConversionTheoremallowsustochoosewhetherto 0. thischoiceleadstoprogrammingcomplicationsthatarenotentirelytrivial. Also,conversiontodecimalisnecessaryforhumaninterpretation.Thereforeit conversionmappingsarecorrectlyimplementedinthecomputingenvironmentof seemsthatthedecimalchoiceshouldbeconsidered.assumingthatrounding 0istrivial.Itmustbenoted,however,that thetestrequesterandalsointhecomputingenvironmentusedbythesoftware executionspeed)whetherargumentsetsarerepresentedinbinaryordecimal. testservice,itisimmaterial(exceptpossiblyforpracticalconcernsinvolving values,thatarepassedthroughthecommunicationinterfacefromfmtothe Thesameremarkistrueconcerningothertestdata,suchascomputedfunction referencesoftwareorviceversa. inatinginoneoftwoways: DecimalOriginationThefunctiony=f(x)istobetestedtoobtainageneralimpressionofitsaccuracyoverpartsofitsdomain.Hereknowledge Tosummarize,atestrequesterwillwanttoconsiderargumentsetsasorig- BinaryOriginationThefunctiony=f(x)istobetestedatasetofexactlymachine-representablearguments.Forexample,iffisusedinan Heredecimalrepresentationisstillpermissible,providedtheconditionsof fattheexactargumentsthatariseinaparticularprogramexecution. ofexactbinaryrepresentationsisnotimportant,soitisnaturaltospecify theargumentsetindecimal. applicationprogram,itmightbeusefultohavethecapabilityoftesting argumentsexplicitly.lets:[0;1]m!rmbeamonotonicfunction,where monotonicityisdenedcomponentwise.togeneratejtestarguments,the Fordecimalorigination,useofatestgeneratoravoidstheneedtosupply Theorem2aremet. s(t)=x0(1?t)+xjtandthelogarithmicgenerators(t)=x1?t isused.examplesofunivariatetestgeneratorsaretheequidistantgenerator formula generatorsproducejtestargumentsinthex-interval[x0;xj]withequidistantor xj=s(j=(j+1)); j=1;2;:::;j 0xtJ.These (12) logarithmicspacing.anelementofrandomnessisintroducedbyusingapseudorandomnumbergeneratortoproduceat-sequencet1<t2<:::<tjinstead whenexecutedeitherbythetestrequesterorthesoftwaretestservice. thesoftwaretestservice,andassumingallroundingbaseconversionmappings areimplementedcorrectly,testgeneratorswillproduceidenticalargumentsets ofthet-sequencedenedbytj=j=(j+1).withsucientlycarefulcodingof 8

tospecifyargumentsetsonapropersubmanifoldofthefunctiondomain.let Weconcludethissectionbyintroducingageneralapproachthatcanbeused Df=D(f) 1D(f) 2:::D(f) Thechiefreasonforsuchaprocedureisthatsomeofthevariablesmaybe denotethedomainofafunctionf.ifthedimensionmexceeds1,itmay bedesirabletoholdoneormorevariablesxedforthedurationofthetest. mrm: (13) xedintheapplicationthatgaverisetothetest.astraightforwardapproach m-dimensionaldomain,wherek<m,anditalsocanbeusedtochangethe canbegeneralizedtopermittestingonak-dimensionalsubmanifoldinthe thisunnecessaryspecicationofxedvariables.ithastheadvantagethatit variablesremainingconstantthroughout.however,anotherapproachavoids wouldbetolisttheargumentsetwiththecomponentscorrespondingtoxed coordinatesystem. wehave theonesthatwillvary;theremainingm?kareheldconstant.wesupposethat kisgivensuchthat1km.letbeapermutation,orrearrangement,of (1;2;:::;m),andlet=?1.Denotingthereorderedvariables1;2;:::;m, Firstwereorderthevariablessothattherstkofthemintheneworderare andthetestisappliedtothefunction y=g()=f(x); r=xr; xr=r 2Rk; (r=1;2;:::;m); x2rm; y2r; (14) Dg=D(g) compareeq.(1).thetestrequesterprovidestheintegerk,thepermutation,thexedargumentsk+1;k+2;:::;m,andtheargumentsetinthedomain 1D(g) 2:::D(g) k. (15) Letusconsiderasanexampletheincompletegammafunction (a;z)=z Denea=x1+ix2,z=x3+ix4.Then Z0e?tta?1dt (<a>0): (16) Supposeatestiswantedinwhichx1isheldconstant.Thenk=3,therequired permutationsare=(2;3;4;1),=(4;1;2;3),andthefunctiongisdenedby f(x1;x2;x3;x4)=(x1+ix2;x3+ix4): (17) Alternatively,supposeatestiswantedinwhichthevariablesarerestrictedto realvalues.then==(1;3;2;4)andg(1;2)=(1+i3;2+i4). g(1;2;3)=(4+i1;2+i3): (18) 9

Theproposedsoftwaretestserviceisundergoingactivedevelopmentatthe 6NationalInstituteofStandardsandTechnology.Theinitialemphasisisonthe constructionofthecommunicationinterfaceandassociatedwebdocuments. ConcludingRemarks Thissubstantialprogrammingtaskisbeingaccomplishedwiththeassistanceof M.A.McClainoftheAppliedandComputationalMathematicsDivision. page'forthetestservice.itwillpresentamenuoffunctionsfromwhichthe willfollowtheclassicationthatisusedin[1].specialfunctionsaresubjectto testrequesterwillchoosebyclickingthemouse.initiallyatleast,themenu alternativedenitionsarisingfromvaryingnormalizationcriteria,modication TheenvisionedcommunicationinterfacewillbeaccessedasaWeb`home posedanidenticationprobleminexistingsoftwareforevaluatingandtesting byscalingfunctions,andotherpracticalortheoreticalconsiderations.thishas specialfunctionsbecauseoftheseverelyrestrictedcharactersetusedincomputing.animportantfeatureofwebdocumentsisthattheysupportthefullrange ofmathematicalnotation.thisfeatureisbeingusedtoavoidanyambiguityin theidenticationoffunctionsinoursoftwaretestservice.italsofacilitatesthe possibilityofoeringawiderangeofalternativefunctiondenitionsfortesting. notberelevanttoallusageoftheservice.thereferencesoftwareisessential toallusage,anditisverydemandingtoprovide.itwillrequirealong-term ondemandaswellastoevaluatesoftware.thusthecomparisonsoftwarewill researchanddevelopmenteort.however,symboliccomputingenvironments Thesoftwaretestservicewillbeabletosupplynumericalfunctionvalues existthatsupportnumericalcomputingtoarbitraryprecision.somehaveextensivesupportforspecialfunctions,includingcomputingnumericalvaluesto stricterrorbounds,theyprobablyrepresentthebestcurrentlyavailablesource erencevalues.althoughtheydonotmeetourrequirementsfortheprovisionof highprecision.initiallyatleast,theseenvironmentswillbeusedtosupplyref- usefulinimprovingtheproposedsoftwaretestservice.theauthorwouldbe mostgratefulforreceivingallsuchcomments. paper,readersmayhaveopinions,recommendationsorcriticismsthatwouldbe ofreferencesoftware. Finally,inviewofthenewapproachtotestingthatisintroducedinthis References [1]M.AbramowitzandI.A.Stegun,editors.HandbookofMathematicalFunc- [2]G.AlefeldandJ.Herzberger.IntroductiontoIntervalComputations.AcademicPress,1983. PrintingOce,Washington,DC,1964. tionswithformulas,graphsandmathematicaltables,volume55ofna- tionalbureauofstandardsappliedmathematicsseries.usgovernment 10

[3]D.H.Bailey.Algorithm719:Multiprecisiontranslationandexecutionof [4]R.F.BoisvertandB.V.Saunders.PortablevectorizedsoftwareforBessel functionevaluation.acmtrans.math.software,18:456{469,1992.for FORTRANprograms.ACMTrans.Math.Software,19:288{319,1993. [5]W.J.Cody.Algorithm714.CELEFUNT:Aportabletestpackagefor corrigendumseesamejournalv.19(1993),p.131. [6]W.J.Cody.Algorithm715.SPECFUN:AportableFortranpackageof complexelementaryfunctions.acmtrans.math.software,19:1{21,1993. [7]W.J.CodyandW.Waite.SoftwareManualfortheElementaryFunctions. specialfunctionroutinesandtestdrivers.acmtrans.math.software, 19:22{32,1993. [8]J.Dongarra,T.Rowan,andR.Wade.Softwaredistributionusingxnetlib. PrenticeHall,1980. [9]IEEE.IEEEstandardforbinaryoating-pointarithmetic.ANSI/IEEE Std754-1985,TheInstituteofElectricalandElectronicsEngineers,New ACMTrans.Math.Software,21:79{88,1995. [10]D.W.LozierandF.W.J.Olver.AiryandBesselfunctionsbyparallel SixthSIAMConferenceonParallelProcessingforScienticComputing, integrationofodes.inr.f.sincovecetal.,editors,proceedingsofthe York,1985. [11]D.W.Matula.Baseconversionmappings.InProceedingsofthe1967 SpringJointComputerConference,volume30,pages311{318,1967. Philadelphia,1993. volume2,pages531{538.societyforindustrialandappliedmathematics, [12]D.W.Matula.In-and-outconversions.Comm.ACM,11:47{50,1968. [13]F.W.J.Olver.Anewapproachtoerrorarithmetic.SIAMJ.Numer. Anal.,15:368{393,1978. 11