Sindbad~EG File Manager

Current Path : /usr/home/beeson/public_html/michaelbeeson/research/papers/
Upload File :
Current File : /usr/home/beeson/public_html/michaelbeeson/research/papers/nsappt.dvi

����;� TeX output 2001.01.22:2036������l�����'���&��d�d�D��tG�G�cmr17�Using�7tNonstandard�Analysis�to�Ensure����W?�the�7tCorrectness�of�Sym��qb�s�olic�Computations��������û��X�Qcmr12�Mic��rhael��Beeson������g3�Departmen��rt��of�Mathematics�and�Computer�Science���������San��Jose�State�Univ��rersit�y���������San��Jose,�California�95192,�USA�������g�email:�8�b�S�eeson@mathcs.sjsu.edu������Ofj����Jan��ruary��22,�2001��!7Q���!J�.t�:		cmbx9�Abstract���!��e��-o���		cmr9�Nonstandard�%ianalysis�has�b�A�een�put�to�use�in�a�theorem-pro��9v�er,�Ufwhere�%iit����Xassists���in�the�analysis�of�form��9ulae�in�v�olving�limits.���The�theorem-pro�v�er�in����Xquestion��"is�used�in�the�computer�program��/�j��		cmti9�Mathp��ert��to�ensure�the�correct-����Xness�ɨof�calculations�in�calculus.�7Although�non-standard�analysis�is�widely����Xview��9ed���as�non-constructiv�e,���it�can�alternately�b�A�e�view�ed�as�a�metho�A�d�of����Xreducing�*�logical�manipulation�(of�form��9ulae�with�quan�tiers)�to�computa-����Xtion��<(with�rewrite�rules).��hW��:�e�giv��9e�a�logical�theory�of�nonstandard�analysis����Xwhic��9h���is�implemen�ted�in��Mathp��ert����.�H�W��:�e�describ�A�e�a�pro�cedure�for�\elim-����Xination�aof�innitesimals"�(also�implemen��9ted�in��Mathp��ert�)�and�pro�v�e�its����Xcorrectness.�� ����?�1��N�ffcmbx12�1��WL�In���tro�s3duction�����?�K�`y

cmr10�The�tgeneral�con���text�of�this�researc�h�is�the�use�of�logic�to�ensure�the�correctness����?of�<�computational�steps.�(Man���y�computational�steps�ha�v�e�side�conditions�that����?m���ust��b�Ge�satised�when�they�are�applied.���Curren�t�computational�systems�gen-����?erally�6`ignore�these�side�conditions,�<�with�the�consequence�that�their�answ���ers�are����?unreliable.�eF��*�or�0�example,�g�in�Macsyma�one�can�en���ter��
�b>

cmmi10�a�4��=�0,�then�0�divide�b�Goth����?sides�}of�the�equation�b���y��a��with�the�command�%�=a�.���Macsyma�will�reply�1�	J=�0,����?b�Gecause���it�has�mindlessly�applied�the�rules��a=a�=�=�1���and�0�=a�=�=�0.�G�More���exam-����?ples��can�b�Ge�found�in�Beeson[�5��],�,�Harrison�and�Th�����Gery[�17��
],�and�a�larger�n���um�b�Ger����?of�K�examples�can�b�Ge�found�in�Stoutemey���er.[�26��
]�Con�trary�to�the�rst�impressions����?of��some�observ���ers,���these�problems�cannot�b�Ge�corrected�b�y�simply�adding�some����?simple�UUc���hec�ks�to�algebraic�computation�systems.����NA���computation��whic���h�app�Gears�to�b�e�a�sequence�of�mathematical�form���ulae,����?one���p�Ger�line,��eac���h�one�obtained�b�y�transforming�the�previous�one�according������1����*�l�����'������?�to�ʗa�computation�rule,���really�has�a�more�in���tricate�logical�structure.�юEac�h�line����?dep�Gends��"on�a�list�of��;�':

cmti10�assumptions�,�	whic���h�are�usually�not�written.�!0Some�of�the����?computation��Lrules�generate�more�assumptions,��and�some�use�the�assumptions����?to��qv���erify�their�side�conditions.�S&As�the�computation�pro�Gceeds,��more�assumptions����?are�V�generated;�W*but�also�the�assumptions�themselv���es�ma�y�simplify�and�ev�en�dis-����?app�Gear��if�they�simplify�to��8��<x

cmtt10�true�.�4�It�can�also�happ�en�that�the�simplication�of����?assumptions���rev���eals�a�con�tradiction�among�the�assumptions,��so�that�what�has����?b�Geen�UUderiv���ed�is�a��non���se��}'quitur�.����NThe�B�side�conditions�of�computational�rules�are�sometimes�simple,�}�suc���h�as����?denominators�Ib�Geing�nonzero,�-�or�things�inside�square�ro�ots�b�eing�nonnegativ���e.����?This��so�Gon�leads�to�questions�of�the�domain�of�expressions:�Вit�is�clear�that�an����?adequate��pcomputational�system�m���ust�b�Ge�able�to�calculate�the�domain�of�ev�ery����?expression.����NThe�E�sp�Gecic�imp�etus�for�this�w���ork�arose�in�the�dev�elopmen�t�of��Mathp��}'ert��ز,����?whic���h�E�is�a�computerized�en�vironmen�t�for�learning�algebra,���trigonometry��*�,�and����?calculus.�@F��*�or���an�in���tro�Gductory�description�of�this�system,���see�[�6��].��Mathp��}'ert��w���as����?designed���to�supp�Gort�a�logical�apparatus�sucien���t�to�ensure�correctness�of�the����?answ���ers�/�it�deliv�ers.��Algebra�and�trigonometry��*�,�f�if�w�e�omit�indexed�sums,�f�can����?b�Ge��1supp�orted�more�or�less�on�the�basis�so�far�describ�ed.�@]Sp�ecically��*�,��h�Mathp��}'ert����?�uses�c3an��infer-r��}'efute-assume��algorithm[�5��,��7��	c4].��bWhen�a�rule�of�computation�re-����?quires�<
a�side�condition,�u��Mathp��}'ert��rst�attempts�to�infer�that�condition�(from����?the��<curren���t�assumptions�and�kno�wn,��tsimple,�axioms);�Dif��<that�fails�it�attempts�to����?refute��the�condition.�CXIf�the�refutation�succeeds�it�refuses�to�apply�the�op�Geration.����?(This���stops�the�MA���CSYMA���example�ab�Go�v�e.)�SMIf�neither�inference�nor�refutation����?succeeds,�ѓit���assumes�the�condition.�:�In�this�w���a�y���the�list�of�assumptions�can�gro���w.����NCalculus�C`in���tro�Gduces�an�additional�logical�problem,�z+not�required�in�elemen�tary����?algebra�gZand�trigonometry��*�,�k�whose�solution�with�the�aid�of�nonstandard�analysis����?is�cthe�sub��8ject�of�this�pap�Ger.���The�additional�problem�is�the�treatmen���t�of�b�ound����?v��q�ariables.�q�Normally���in�logic�w���e�think�of�v�ariables�b�Gound�b���y�quan�tiers.�q�In����?calculus,���ho���w�ev�er,�v��q�ariables�fare�b�Gound�b���y�the�limit�op�erator,���the�denite�in���tegral����?op�Gerator,�u�and�=�indexed�sums.����^��ٓ�Rcmr7�1������Consider,�for�example,�the�problem�of�determining����?the�UUdomain�of������������}�	0e�rcmmi7�n������ܼ�����u

cmex10�X�����ܯ��k�+B�=1�����9�x�����k��됵:���捑?�This��pdomain�m���ust�come�out�to�b�Ge��true�{the�function�is�dened�for�all��x�.�FW��*�e����?m���ust���rst�of�all�inform�the�system�that�index�v��q�ariables�are�alw�a�ys�in�tegers.�0�That����?b�Geing��\done,��^the�main�problem�stands�out:��ֵx���^��k����is�dened�only�for��x����
!",�

cmsy10�6�=�0����_��k��e>����?�0.���Mathp��}'ert�{n�handles�this�problem�as�follo���ws:���When�doing�computation,���the����?expression��Dtree�is�tra���v�ersed��Ddepth-rst.��Some�computation�rules�are�applied��?�X-�ff��v�	J=�����"5��-:��Aa�cmr6�1����L��|{Ycmr8�In��general,��~b�<round�v��ariables�can�b�e�reduced�to�functional�op�erations�and�lam�Îb�da�abstrac-��	��tion,��as�ԙis�w�Îell-understo�<ro�d�ԙb�y�exp�<rerts.���The�additional�problem�is�not�merely�the�existence�of���additional���binding�op�<rerators,��
but�the�rules�for�mixing�these�op�erators�with�sym�Îb�olic�compu-���tation��Xcorrectly��J�.�������2�����l�����'������?�on��the�w���a�y��do�wn�(that�is,�O�b�Gefore�their�argumen�ts�are�simplied)�and�some����?on���the�w���a�y���up�(after�their�argumen���ts�are�simplied).�	k�When�it�passes�in�to����?an���indexed�sum,����Mathp��}'ert��generates�the��temp�or�ary��assumption�that�the�index����?v��q�ariable��Yis�b�Get���w�een��Ythe�lo���w�er��Yand�upp�er�limits.��In�the�example,��[that�w���ould�b�e����?the�h9assumption�1�擸��k���^�Ew�k�7*���n�.��sThe�h9assumption�1�擸��n�h9�w���ould�also�b�Ge�generated����?(if�D�it�could�not�b�Ge�inferred�or�refuted),�H"but�that�is�not�a�temp�orary�assumption����?and�Ênot�the�p�Goin���t�of�in�terest�here.��hThe�temp�Gorary�assumption�is�a�v��q�ailable�for����?use��Rwhile�computation�is�pro�Gceeding�inside�the�indexed�sum.��In�particular,�it����?is�]�a���v��q�ailable�during�the�analysis�of�the�domain�of��x���^��k��됲,�`"whic�h�therefore�comes�out����?ev���erywhere��Idened.�T�When�computation�exits�the�indexed�sum�after�tra�v�ersal����?of�UUthe�en���tire�expression�tree,�the�temp�Gorary�assumption�is�\disc�harged".����NLogically���sp�Geaking:���eac���h�line�of�computation�can�b�e�view���ed�as�a�\sequen�t"����?���)��A�,�"where�U�A��is�the�visible�line�and��is�the�list�of�curren���t�assumptions.�WrThe����?list�tf�can�gro���w�from�line�to�line�as�computation�rules�generate�assumptions;���it����?can���also�shrink�if�assumptions�simplify��*�.��But�during�computation,���while�w���e�are����?xed�Ւat�a�single�line,���it�can�also�gro���w�and�shrink�as�computation�go�Ges�deep�er����?in���to�yband�then�out�of�v��q�ariable-binding�functors�on�the�righ�t�side.���This�dynamic����?qualit���y�UUis�essen�tial�for�the�prop�Ger�handling�of�b�ound�v��q�ariables.��!č��?�2��WL�The�ffProblem,�its�Solution,�and�Examples������?���N�cmbx12�2.1��]�The��Problem��uT��?�The�*sc���heme�describ�Ged�ab�o���v�e�*seems�adequate�for�the�treatmen���t�of�denite�in-����?tegrals���and�indexed�sums.�k-But�calculus�requires�the�treatmen���t�of�limit�terms����?as���w���ell,��and�it�is�far�from�ob�vious�what�temp�Gorary�assumptions�m�ust�b�Ge�made����?when�UUen���tering�a�limit�term.�q�Consider��Example���1�:���|�����N�lim��-������x�O!�cmsy7�!�0�������*����^���������<$���պ(�sin����x���^��2���պ(�w�fe%U�	(֍�	7�x������K��+�����<$��d1��l�w�fe�	(֍�x�8���1���������A���^��������?�In���order�to�w���ork�correctly�with�the�\limitand"�(the�form�ula�inside�the�limit),����?w���e�NYm�ust�b�Ge�able�to�infer�that��x�f�6�=�0�and�that��x��6�=�1.�\�Of�course,���w���e�cannot����?generate���an���y�assumption�in�v�olving��x��that�will�app�Gear�outside�the�scop�e�of�the����?limit�UUop�Gerator.����NHere���is��Example��2�.�lYIn�calculating�the�deriv��q�ativ���e�of�������p����2����fe��3荵x����^)�from�the�limit�de-����?nition���of�deriv��q�ativ���e,���w�e���reac�h�a�step�in�whic�h�w�e�w�an�t�to�rewrite�(�����p���UW���fe����x�8�+��h���� H�)���^��2��S�as����?�x�mײ+��h�.�`(Here��ǵh��is�the�b�Gound�v��q�ariable.)�The�rule�(������p���UW����feB��;��y����
�O�)���^��2����=�K��y���is�only�v��q�alid�when����?0�����y�[ٲ,��Rand�Ϲw���e�cannot�generate�the�assumption�0�����x��x�+��h�,�b�Gecause�Ϲit�in���v�olv�es����?the�!}b�Gound�v��q�ariable.�`�The�correct�assumption�to�generate�is�0���<�x�.�Ho���w�!}can�this����?desired�UUresult�b�Ge�reac���hed,�b�y�a�general-purp�Gose�algorithm?����NThe��eproblem�is�this:�e�w���e�m�ust�ensure�that�when�computation�is�p�Gerformed����?inside�+�a�limit�term,�4#the�logical�inferences�p�Gerformed�in�v���erifying�side�conditions������3������l�����'������?�are�37correct,�:
and�an���y�assumptions�generated�b�y�op�Gerations�p�erformed�inside�the����?limit�UUterm�do�not�in���v�olv�e�UUthe�b�Gound�limit�v��q�ariable.���6���?�2.2��]�A��solution�using�nonstandard�analysis��uT��?�A�V�w���ork��q�able�V�solution�to�this�problem�w�as�describ�Ged�six�y�ears�ago.[�5��]�It�amoun�ts����?to���in���tro�Gducing�second-order�logic�in�to�the�computation�system,�R�for�example����?a��second-order�functor��true���E�ff&f��ǫin���E�ff&f��nbhd�,�D�whic���h��tak�es�a�prop�Gosition�and�p�oin���t�as����?argumen���t���and�is�true�if�the�prop�Gosition�is�true�in�a�neigh�b�Gorho�o�d���of�the�p�Goin�t.����?Although�UUthis�solution�w���ork�ed,�UUit�w���as�complicated.����NThis��Mpap�Ger�describ�es�another,�Omore�elegan���t�solution,�using�nonstandard�anl-����?ysis.���In�kHessence,���it�is�a�w���a�y�kHto�replace�second-order�inferences�b���y�rst-order����?computation:�xQcalls�X�to�the�limit-calculator.�{�This�sa���v�ed�X�sev�eral�thousand�lines�of����?complicated�UUco�Gde�in��Mathp��}'ert��ز.����^��2������N�The��idea�is�to�treat�limit�terms�in�exactly�the�same�w���a�y��as�denite�in���tegrals����?and��indexed�sums�are�treated:�oto�mak���e�a�temp�Gorary�(rst-order)�assumption����?when��en���tering�the�term�and�disc�harge�it�on�exit,��and�let�computation�and�infer-����?ence��3b�Goth�pro�ceed�normally�while�inside�the�term.�saThe�assumption�is�ob���vious����?once�dynonstandard�analysis�is�considered:��when�en���tering��lim���H
���x�!�a��$J�f���(�x�),�hAmak�e�the����?temp�Gorary�S�assumption�that��x�❸��a�S�is�innitesimal,���and��x�or�6�=��a�.�m�F��*�or�S�one-sided����?limits�ܴfrom�the�righ���t,���assume��a���<�x�ܴ�instead�of��x����6�=��a�,�and�ܴfor�limits�from�the����?left,�UUassume��x��<�a�.����NWhen�pcomputation�completes�its�tra���v�ersal�pof�the�limitand,�an���y�assumptions����?whic���h���ha�v�e�b�Geen�generated�in�v�olving�the�limit�v��q�ariable,���and�not�already�elimi-����?nated��Sb���y�simplication,��m�ust�b�Ge�eliminated:�DFw�e�cannot�generate�an�assumption����?in���v�olving�+0a�b�Gound�v��q�ariable�outside�its�scop�e.�c�If�this�can�b�e�done�b���y�making�ad-����?ditional���assumptions,��$it�is�done.�?�If�it�can't,�then�the�computation�or�inference����?w���e�UUw�ere�attempting�has�failed.����NW��*�e�һsupply�the�system�with�b�Goth�axioms�and�computational�rules�for�dealing����?with�Ginnitesimals�and�innite�nonstandard�n���um�b�Gers.�GThe�Gfundamen�tal�con-����?cepts�! in�our�theory�are��x����T͍�������+3�����=�����
UN�y�[ٲ,�+�whic���h�means�that��x��v���y�|��is�! innitesimal�(or�zero),����?and��[��UX[�x�]�],�<�the�standard�part�of��x�.���Since�w���e�w�ork�in�the�logic�of�partial�terms,����?�2�"V

cmbx10�LPT�,�@%there�is�no�problem�ab�Gout�[��UX[�x�]�]�@%b�eing�undened�for�innite�n���um�b�ers.�j�The����?axioms��are�stated�using�the�predicate����T͍������+3����=�����
�pand�the�function�sym���b�Gol�[��UX[��]�].�^�The��com-����?putation�+2rules�are�directed�to���w�ards�+2the�goal�of�b�Geing�able�to�compute�[��UX[�t�]�]�+2for����?as�}@man���y�terms��t��as�p�Gossible.��Here�\compute"�means�to�eliminate�nonstandard����?v��q�ariables.����NF��*�or�iCexample,�n>reconsider�the�computation�of�the�deriv��q�ativ���e�of�������p���������fe��3荵x����u��,�where�w���e��?���ff��v�	J=�����"5��-:�2����LܻThe�Geect�of�this�c�Îhange�(made�in�June,�%C1992)�w�as�in�visible�to�the�user.���Logic�is�in�the��	��bac�Îkground�זin��"#�f�cmti8�Mathp���ert��[�.���Studen�ts�of�calculus�are�not�taugh�t�to�w�orry�ab�<rout�logical�niceties.���But�Z�since��Mathp���ert��[�,�s@unlik�Îe�its�answ�er-only�cousins,�s@pro�<rduces�only�correct�results,�it�m�Îust�tend���to�\Slogical�matters�b�<rehind�the�scenes.�\The�c�Îhange�to�nonstandard�analysis�w�as�th�us�doubly���in�Îvisible:�`�logic�"4is�visible�only�if�y�ou�explicitly�c�ho�<rose��2�@�cmbx8�View��	Assumptions�,�ukand�b�esides,���nonstandard��Xanalysis�pro�<rduces�the�same�results�as�second-order�inference.�������4����*ޠl�����'������?�had�=�to�rewrite�(�����p���UW���fe����x�8�+��h���� H�)���^��2���4�as��x�	��+��h�.�i�This�=�no���w�tak�es�place�under�the�assumption����?that�<�h��is�innitesimal�(and�not�zero).�i�When��Mathp��}'ert��tries�to�v���erify�0�����x��+��h�,����?it�.�will�compute�the�standard�part�of��x��ֲ+��h�.вto�b�Ge��x�,�6�and�then�simplify�0�����x��ֲ+��h����?�to�Oo0���<�x�-�_��(0�����x�-�^��0�����h�).�o�The�system�then�simplies�this�to�0��<�x�.�o�If�w���e�had����?b�Geen�UUinside�a�one-sided�limit�from�the�righ���t,�it�w�ould�ha�v�e�simplied�to�0�����x�.�������?�2.3��]�Examples��uT��?�One�ynreferee�ask���ed�for�more�examples�correctly�w�ork�ed�b�y��Mathp��}'ert�.��The�p�Goin�t����?to���b�Ge�emphasized�here�is�that��Mathp��}'ert��will�correctly�w���ork��every��example�of�a����?limit�b�problem�found�in�calculus�textb�Go�oks.��)A���t�b�least,�f(it�has�w�ork�ed�h�undreds�of����?suc���h�A%examples�correctly��*�.�k
The�relev��q�an�t�p�Goin�t�is�not�that�it�arriv�es�at�the�correct����?answ���er�Q{for�the�limit�(whic�h�it�do�Ges),�R@but�that�it�(1)�analyzes�the�domain�of�the����?expression�d�in���v�olving�the�limit�correctly��*�,�h�(2)�mak�es�no�unnecessary�assumptions����?while���calculating�the�limit,���(3)�mak���es�all�necessary�assumptions�to�ensure�cor-����?rectness,��\and���(4)�mak���es�no�incorrect�assumptions,�suc���h�as�assumptions�in�v�olving����?the���limit�v��q�ariable�made�outside�the�scop�Ge�of�the�v�ariable.���These�claims�can�b�Ge����?v���eried���b�y�c�ho�Gosing��Assumptions��from�the��View��men�u�of��Mathp��}'ert��in�order����?to��insp�Gect�the�assumption�list�as�a�limit�computation�progresses.�O�Here�are�some����?sp�Gecic�UUexamples.���������Tlim��-���ڏ��x�!�1�����1�=x����h:�Example�TB3����˰��?Mathp��}'ert�_G�correctly�sho���ws�an�empt�y�assumption�list�after�analyzing�the�domain����?of�UUthis�expression.���������lim��-���ׁǴx�!�1����骫�tan�����x����h:�Example�TB4������?�Again�4��Mathp��}'ert��sho���ws�an�empt�y�assumption�list,�l�ev�en�though�the�expression����?for��xthe�domain�of��tan����x��is��x�Y��6�=�(2�n�s��+�1)��[�=�2,�Áwhic���h��xin�v�olv�es�a�new�existen�tially����?quan���tied�UUv��q�ariable��n��of�t�yp�Ge��int�,�generated�b�y�the�domain�analyzer.���}�������lim��-���ѕ��x�!�2�������<$����;�x���^��2���S��8�1����ɟw�feER�	������$��p���UW��$��fe����x�8���1�����������h:�Example�TB5�������?�First,�a�no�_nassumption�is�generated�b���y�the�requiremen�t�that�things�inside�square����?ro�Gots�fb�e�nonnegativ���e.��Second,�jBwhen�the�expression�����$��p����i��$��fe����x�8���1������"�d��$��p���+���$��fe����x�8���1����EV�is�simplied����?to��A�x��|���1,��;the�side�condition��x����1������0�is�prop�Gerly�inferred,��;b�ecause�this�tak���es����?place��inside�a�limit�as��x��approac���hes�2.��OThe�side�condition�reduces�to���1>�
��1,����?where�UU��^ϲis�a�nonstandard�v��q�ariable�whose�distance�to�2�is�innitesimal.�����������lim�������x�!�1������Zcmr5�+��������<$����x���^��2���S��8�1���Ó�w�feER�	������$��p���UW��$��fe����x�8���1�����������h:�Example�TB6�������?�No���w�}�the�limit�p�Goin�t�is�at�the�zero�of�the�denominator,���but�b�Gecause�it�is�a�one-����?sided�UUlimit,�ev���erything�go�Ges�as�with�the�previous�example.���}�������lim��-���ѕ��x�!�1�������<$����;�x���^��2���S��8�1����ɟw�feER�	������$��p���UW��$��fe����x�8���1�����������h:�Example�TB7���������5����;Πl�����'������?�This��is�the�same�as�the�previous�example�except�it�is�a�t���w�o-sided��limit.���The����?domain�analyzer�can�determine�the�expression�is�undened,�-?since�the�domain����?of�����$��p����K��$��fe�������BZ��8�1����'E;for���a�nonstandard����n�whose�distance�to�1�is�innitesimal�will�reduce����?to�q�the�prop�Gosition��������1,�x�whic���h�will�reduce�to�to��false��using�the�algorithm�of����?this�UUpap�Ger.����N�Mathp��}'ert�mu�includes�man���y�mathematical�op�Gerations�with�side�conditions,�s}and����?one���can�giv���e�arbitrarily�complicated�examples,��but�these�mathematically�simple����?examples�UUshould�suce�to�illustrate�the�logical�p�Goin���ts�in�v�olv�ed.���6���?�2.4��]�Correctness��of�the�Algorithm��uT��?�The�(ab�Go���v�e�examples�illustrate�the�use�of�nonstandard�analysis�in�assisting�with����?inferences�N�and�reductions�made�inside�the�scop�Ge�of�a�limit.�o�The�pap�er�will�giv���e����?an��.algorithm�making�this�metho�Gd�systematic.�G�In�particular,��jthe�algorithm�m���ust����?eliminate��a�nonstandard�v��q�ariable�in���tro�Gduced�in�this�w�a�y��*�.��Certain�steps�of�the����?algorithm��Preplace�a�form���ula�suc�h�as����3<�๲1�b�y�a�form�ula�(for�example��false�)����?whic���h�d�is�not�logically�equiv��q�alen�t.���Therefore�a�correctness�pro�Gof�is�called�for.����?Theorem��3�of�this�pap�Ger�pro���vides�suc�h�a�correctness�pro�Gof.��SThe�pro�of�is�based����?on�љ\in���terv��q�al�seman�tics":�jOa�form�ula���(��	z�)�in�v�olving�the�nonstandard�v��q�ariable������?�is���in���terpreted�to�mean�that���(�x�)�is�true�for�all��x��in�a�suitable�(punctured)����?neigh���b�Gorho�o�d��`of�the�limit�p�Goin�t.��(The�neigh�b�Gorho�o�d��`will�b�Ge�one-sided�for�a����?one-sided�UUlimit.)��!č��?�3��WL�Logic�ffwith�t���yp�s3es�and�partial�terms�����?�The��main�purp�Gose�of�this�pap�er�is�to�describ�e�an�algorithm�for�\innitesimal����?elimination"�p_and�pro���v�e�p_its�correctness.���Ho���w�ev�er,�w!it�p_is�not�so�easy�ev���en�to��state����?�the�B:correctness�theorem,�Fas�that�necessarily�in���v�olv�es�B:a�formal�logical�theory�for����?calculus.�:Since��the�algorithm�in���v�olv�es��nonstandard�analysis,���w���e�also�need�a����?formal�V�theory�for�nonstandard�analysis.�v�In�this�section�w���e�form�ulate�a�suitable����?rst-order�o�theory�of�standard�calculus.��@This�requires�some�reform���ulations�of����?traditional�zSrst-order�logic,���to�deal�with�partial�terms�suc���h�as�������p���Ϫ����fe��3荵x�����Ʋ,�and�to�deal����?with��athe�t���yp�Ge-em�b�edding��aproblem.�,!Readers�who�do�not�wish�to�c���hec�k��athe�details����?of�UUthe�correctness�pro�Gof�ma���y�b�e�able�to�skip�this�section.����NT��*�o���form���ulate�a�theory�of�real�n�um�b�Gers�with�square�ro�ot�in�ordinary�predicate����?logic,���w���e��w�ould�ha�v�e�to�allo�w�����s0�p���
���s0�fe㏟�Ѝ�(�������]Y�1)�as�a�term.��It�is�unnatural�to�w�ork�with������?����p���GUW����fe��3荵x����Q��when�
�x��is�negativ���e�(when�w�e�do�not�ha�v�e�complex�n�um�b�Gers�in�mind).��It����?is��9b�Getter�to�use�a�Logic�of�P���artial�T��*�erms�(�LPT�)��whic�h�tak�es�\�t��is�dened"�as����?a��primitiv���e�concept.���A���suitable�logic��LPT��has�b�Geen�form���ulated[�3��].���One�of�its����?rules���for�form���ula�formation�is�that�if��t��is�a�term,���then��t��θ#��is�a�form�ula,���whic�h����?means�UU�t��is�dened.����NThe���t���yp�Ge-em�b�edding���problem�already�arises�when�w���e�w�an�t�to�discuss�b�Goth����?in���tegers��and�real�n�um�b�Gers,��and�ha�v�e�3�b�Ge�b�oth�an�in���teger�and�a�real�n�um�b�Ger.������6����J��l�����'������?�T��*�o��understand�the�problem�clearly�,�ޑconsider�that�w���e�will�w�an�t�t�w�o�t�yp�Ges,�ޑ�int����?�(for�q�in���tegers)�and��real�?��(for�real�n�um�b�Gers).��In�a�notation�to�b�e�fully�explained����?b�Gelo���w,�V�w�e�#Awrite��8�x:��int���8�to�quan���tify�o�v�er�in�tegers,�V�and��x�K�:���real��A��for�#Aa�form�ula����?expressing�,2that��x��is�a�real�n���um�b�Ger.��_W��*�e�,2can�then�state�the�comm���utativit�y�,2of����?adding�fin���tegers�and�reals,��and�the�comm�utativit�y�of�adding�reals�and�reals,��and����?w���e�UUshould�b�Ge�able�to�deriv�e�the�former�from�the�latter.����NSp�Gecically��*�,��w���e���should�b�e�able�to�deriv���e��8�n:��int����8�x:��real��?�(�n��d�+��x���=��x��d�+��n�)����?from�t:�8�x:��real��?�8�y�[�:��real���(�x�v��+��y�"�=���y�҆�+��x�),��@making�use�of�the�t���yp�Ge-em�b�edding�t:axiom����?�8�n���:���int����(�n��:���real���).����NThere�{=are�t���w�o�{=traditional�metho�Gds�to�solv���e�the�t�yp�Ge-em�b�edding�{=problem:����?The�Q�rst�is�to�use�t���w�o�Q�unary�predicates��N��(�x�)�and��R�Dz(�x�),���with�the�t���yp�Ge-em�b�edding����?axioms�ז�N��(�x�)��.�!��R�Dz(�x�).���This�has�the�disadv��q�an���tage�that�all�quan�tiers�m�ust�b�Ge����?explicitly��relativized�to�these�unary�predicates.���The�second�metho�Gd�is�to�use����?a�t���w�o-sorted�predicate�calculus,�ɐwith�one�sort�for�in�tegers�and�one�for�reals.����?This�"@has�the�disadv��q�an���tage�that�there�m�ust�b�Ge�an�explicit�t�yp�Ge-em�b�edding�"@func-����?tion�%�that�con���v�erts�%�the�in���teger�3�to�the�real�n�um�b�Ger�3.��When�writing�pap�ers,����?as�N]opp�Gosed�to�computer�programs,���one�can�mak���e�some�con�v�en�tions�that�slur����?the��Edistinctions�b�Get���w�een��Ethese�t���w�o��Eapproac�hes,��but�to�b�Ge�precise�one�needs�an����?apparatus�=�that�com���bines�the�b�Gest�of�these�t�w�o�approac�hes.�+.In�this�approac�h����?a�j7v��q�ariable�is�lab�Gelled�with�its�t���yp�e�(as�in�t���w�o-sorted�j7predicate�calculus),��pbut����?predicate�k�and�function�sym���b�Gols�are�not�required�to�tak�e�exactly�one�t�yp�Ge�as����?argumen���t.�Q�Th�us��}the�form���ula��x��<�y�PV�is��}w�ell-formed�regardless�of�whether��x��and��y����?�ha���v�e�c(t�yp�Ge��real��or��integer�.��BI�c%do�not�kno�w�of�an�y�men�tion�of�suc�h�a�v�ersion�of����?predicate�"�calculus�in�the�literature,�,�but�it�is�a�straigh���tforw�ard�"�matter�to�reduce����?it���to�kno���wn�v�ersions,�*�and�w�e�will�use�it.�q�T��*�o�illustrate�the�legal�syn�tax:��if�w�e����?ha���v�e�_�a�t���yp�Ge��int�,�b1then��8�x:��int����8�y�[�:��int�����(�x�?��+��y�4�=��=�y����+��x�)�_�w�ould�b�Ge�a�legal�form�ula.����?W��*�e��use�the�notation��x:��int����to�indicate�t���yp�Ge�information�attac�hed�to�a�v��q�ariable.����?On�I�the�other�hand��x���:���int����will�I�b�Ge�written�for�a�form���ula.�m�Ocially�this�should����?b�Ge����x:��int����:����int���,�	but�w���e�will�nev�er�write�it�out�lik�e�this.�R.In�an�implemen�tation,����?some�{�bits�of�(or�accessible�from)�the�ob��8ject�represen���ting��x��will�b�Ge�set�aside�for����?t���yp�Ge��information,��and�the�notation��x:��int����is�mean�t�to�denote�that�those�bits����?con���tain�<the�co�Gde�for�the�t�yp�Ge���int��0�.��|The�use�of�t�w�o�dieren�t�sym�b�Gols��x:��int�����?�(whic���h��is�a�v��q�ariable)�and��x���:���int�����(whic�h��is�a�form�ula)�helps�in�writing�a�precise����?grammar�UUb�Gelo���w.����NThis��7sho���ws�the�t�yp�Ge�information�explicitly��*�.�;�When�writing�on�pap�er,�ӣw���e�sho�w����?the�dbt���yp�Ge�lab�els�on�v��q�ariables�only�when�they�are�b�ound,���as�in��8�n:��int����8�x:��real���(�x�V��+����?�n�#ɲ=��n�*��+��x�),��but��'the�t���yp�Ge�lab�els�are�nev���erthless�ocially�presen�t�with�eac�h����?o�Gccurrence�J�of�the�v��q�ariable.��W��*�e�also�allo���w�`t�ypings'�to�b�Ge�used�as�atomic�form�ulae.����?Th���us�_�8�n:��int�����(�n��i�:���real���]�)�is�a�legal�form���ula,�a�whic�h�can�b�Ge�used�to�express�a�t�yp�Ge-����?em���b�Gedding.����NW��*�e���set�out�here�the�syn���tax�and�rules�of�rst-order�logic�with�partial�terms����?and��t���yp�Ged�v��q�ariables.��\This�v�ersion�of�rst-order�logic�giv�es�a�go�Go�d��solution�of����?the�UUt���yp�Ge-em�b�edding�UUproblem�and�of�the�undened-terms�problem.������7����Y��l�����'������S?��variable::=�
�identifier.type����S?�term::=�?�variable�|�constant�|�functor(termlist)����S?�termlist::=�?�term�|�term,termlist����S?�atomic���E�ff&f��ǫformula::=���term�?��#��|�������true�?�|�false�|�������term�?�:�
�type�|�������term�?��=��term�|�������term�?�infix���E�ff&f��ǫrelation�term�|�������pfunctor(termlist)����S?�formula::=�
�atomic���E�ff&f��ǫformula�?�|�����?�formula�?��^��formula�|�����?�formula�?��_��formula�|�����?�formula�?��!��formula�|�����?ٸ:�?��formula�|�����?ٸ8�?��variable�(formula)�����?ٸ9�?��variable�(formula)����S?�formula���E�ff&f��ǫlist::=�
�formula�?�|�formula,formulalist�|�<empty>����S?�sequent::=�
�formulalist�?��)��formula����S?�classical���E�ff&f��ǫsequent::=�?�formulalist��)��formulalist����N�Note���that�these�rules�mak���e��x:A��a�v��q�ariable,���not��x��standing�alone.�$�Ho�w�ev�er,����?in��\writing�on�pap�Ger,���w���e�suppress�the�t�yping�information�and�just�write��x�,���except����?where�d-quan���tiers�bind�v��q�ariables.�!eW��*�e�form�ulate�the�logic�using�Gen�tzen�sequen�ts.����?The�գrules�for�prop�Gositional�calculus�are�the�usual�Gen���tzen�rules�(see�e.g.�G7system����?G1�UUin�Kleene[�19��
]).����NT��*�o��"sp�Gecify�an���y�particular�system�in�this�logic,��Vw�e�m�ust�add�grammar�rules����?for����functor�,���pfunctor�,��constant�,�and��type�.�*	This�is�the�same�as�in�the�usual����?presen���tations���of�rst-order�logic,���where�a�particular�rst-order�theory�m�ust�b�Ge����?sp�Gecied��Bb���y�its�function�sym�b�Gols�and�constan�ts.���Here�w�e�m�ust�also�giv�e�t�yp�Ges����?and�Ek�pfunctors�,��pthat�is,�sym���b�Gols�used�for�atomic�prop�ositions.�B	W��*�e�also�adopt����?the�>usual�con���v�en�tion�>regarding�arities�of�functors:�ݚeac���h�functor�comes�with�a����?sp�Gecied�O&arit���y��*�,�Pbso�that�not�just�an�y��termlist��can�b�Ge�put�in�to�just�an�y�functor.����?The���grammar�ab�Go���v�e���do�es�not�enforce�suc���h�a�restriction,�نbut�all�readers�will����?kno���w�Ԏho�w�to�mak�e�it�do�so,��\at�some�cost�in�readabilit�y��*�.��rT�o�x�the�ideas,��\one����?ma���y�UUtak�e�the�rules����N�type�?�::-�real�|�int����Ninfix���E�ff&f��ǫrelation�?�::-�<�|����|��6�=��|�=����Nfunctor�?�::-�ln,�sqrt,���:���:�:�����?�and�UUha���v�e�no�rule�at�all�for��pfunctor�.����NThe�Gen���tzen�rules�for�v��q�ariables�and�quan�tiers�require�mo�Gdication,�A:b�oth����?b�Gecause�}�of�the�use�of��LPT�}ܲand�b�ecause�of�the�use�of�t���yp�ed�v��q�ariables.��~In�these����?rules,�����_and��are�lists�of�form���ulae,��A��is�a�t���yp�Ge,�and��t��is�a�term.��Substitution����?of�UUa�term�for�a�v��q�ariable�is�indicated�b���y�a�slash.�q�Here�are�the�rules:������8����	kY�l�����'�����C܍�����<$����β�;���b���:��A��)���[�b:A=x�]����Οw�feS,d�	(֍�
����)�8�x:A�(��)�������������<$���A7�;���b���:��A;��[�b:A=x�]��)�� ���A7�w�fe^}��	(֍�
���;����9�x:A�(��)���)�� �������������<$����o��;���t���:��A;��[�t=x:A�]��)�� ����o�w�fe]!"�	(֍�\m��;����8�x:A�(��)���)�� �������������<$����$����)���[�t=x:A�]���)��t��:��A����$�w�feyA��	(֍�����;�������)�9�x:A�(��)���������NThe�D�rst�t���w�o�D�of�these�rules�are�sub��8ject�to�the�usual�restriction�on�v��q�ariables,����?that�|��b��is�not�free�in�the�conclusion�of�the�rule,��Vthat�is,�the�part�b�Gelo���w�the�line.����?These��'rules�are�the�in���tuitionistic�rules;��the�classical�rules�are�obtained�just�b�y����?allo���wing���more�form�ulae�on�the�righ�t�of��)�.�E�All�theories�used�in�this�pap�Ger�will����?b�Ge�UUclassical.����NThe�UUuse�of�constan���t�prop�Gositions��true��and��false��requires�the�axioms���������)���true�������Eafalse���R�;�������)������?�and�UUa�reform���ulation�of�the�rule�for�in�tro�Gducing�negation:���j������<$���Ea�;������)���false����Ea�w�fe:u>�	(֍������)�:�������MЍ�?�With�@/this�rule,�Djw���e�can�mak�e�sure�the�succeden�t�(righ�t�hand�side�of��)�)�is�nev�er����?empt���y��*�.����NThe�UUequalit���y�axioms�in��LPT��tak�e�the�form��������x���=��x�8�^��(�x���=��y�"�!��y��=��x�)�������(�t���#�_�s��#!��t��=��s�)�8�^���(�t�)���!���(�s�)����NThe���relationship�b�Get���w�een���the�t���w�o���sym�b�ols���w�e�ha�v�e�used�for�t�yping�is�giv�en����?in�UUthe�follo���wing�\t�yping�axioms":������8�x:A�(�x���:��A�)�����z�x:B�G��:���A��!�9�z�p�:A�(�x��=��z��)����?T��*�o���illustrate�the�need�for�the�second�t���yping�axiom,���consider�trying�to�pro�v�e����?�8�x:A�(�x�#�:���int��:�!��x��:���real���),��]when���giv���en�only�the�axiom��8�x:��int����(�x�#�:���real���).��vW��*�e����?ha���v�e�$*to�argue�as�follo���ws:�qLet��x:A��b�Ge�giv�en.��FSupp�Gose��x�β:���int��².�Then�$*b���y�the������9����
v��l�����'������?�second�wt���yping�axiom,�u�x��is�equal�to�some�in�teger��z�p��,�uand�b�y�the�t�yp�Ge-em�b�edding����?of�UUin���tegers�in�to�reals,��z���is�also�a�real;�hence�b�y�equalit�y�axioms��x���:���real����.����NThere��are�some�rules�required�b���y�the�logic�of�partial�terms.�wqThe�rst�one����?sa���ys��pthat�there�are�no�undened�ob��8jects.�1&Undenedness�is�a�prop�Gert�y�that�terms����?can��p�Gossess,�4not�a�prop�ert���y�that�ob��8jects�b�eing�sp�ok���en�ab�out�can�p�ossess.�/�All����?ob��8jects�UUexist.�����Զ>���)��x:A��#����N�The�;�v���ersion�of��LPT�;ٲalready�in�the�literature[�3��]�requires��strictness�:�efor�eac�h����?�n�-ary��@function�sym���b�Gol��f�ϲand�eac�h�sequence�of�terms��t����i��TL�,�and�eac�h��i�����n�,�w�e��@ha�v�e�����M#�f���(�t����1��|s�;����:�:�:����;���t����n��q~�)���#!��t����i��d�#����?�W��*�e�s�shall�k���eep�this�condition�in�the�presen�t�form�ulation,�{though�one�w�ould�lik�e����?to��w���eak�en�it,�J�b�Gecause�it�w�as�essen�tial�to�the�completeness�pro�Gof�for��LPT�.�In����?particular,��>it���implies�that�the�meaning�of��t�!��:��A����is�\�t��is�dened�and�of�t���yp�Ge��A�".����?Strictness�UUalso�applies�to�relation�sym���b�Gols�and�equalit�y:����Ŕ2�x���=��y�"�!��x��#�^�y��#����?�and�����L��R�Dz(�x;���y�[ٲ)���!��x��#�^�y�"�#����?�for�UUeac���h�atomic�relation�sym�b�Gol��R�i�(inx�or�not).����NNote��(that�in�the�t���yp�Ged�v�ersion�of��LPT���presen�ted�here,�ethe�denedness�sym-����?b�Gol�-�do�es�not�en���ter�in�to�the�quan�tier�rules.����^��3���x�The�un�t�yp�Ged�v�ersion[�3��]�has�the����?rule��C܍�����<$���	����)���[�t=x�]���)��t��#���	͟w�fef�f�	(֍�����;�������)�9�x�(��)������MЍ�?in�UUplace�of�the�rule�used�here:��5�������<$������)���[�t=x��:��A�]���)��t��:��A����w�fe~��	(֍�����;�������)�9�x��:��A�(��)������MЍ�?Y��*�ou�6�migh���t�w�onder�ho�w�the�denedness�sym�b�Gol�will�ev�er�en�ter�in�to�pro�Gofs�then.����?There�!Care�at�least�t���w�o�!Cw�a�ys:�	�from�the�existence�axiom,�T>and�from�non-logical����?axioms.�q�F��*�or�UUexample,�y���ou�migh�t�ha�v�e�the�nonlogical�axiom������Äw��O�p�����Ο�O�fe���㍵x�����X�#!���(����O�p���UW��O�fe���㍵x����s�)�����2��C��=��x��?��f�ff��v�	J=�����"5��-:�3����LܻSince��1w�Îe�did�not�sp�<recify�the�grammar�rules�for���C�scmtt8�type�,��there�ma�y�really�b�<re�sev�eral�\t�yp�<red��	��v�Îersions"��bof�our�theory��J�.���If�w�e�further�require�that�there�are�only�constan�t�t�yp�<res,���as�is�the�case���in�3|�Mathp���ert�,�S�then�the�consistency�of�the�t�Îyp�<red�v�ersion�follo�ws�from�that�of�the�un�t�yp�<red�v�ersion���in�~+the�same�w�Îa�y�~+as�for�ordinary�logic,���b�Îy�in�terpreting�the�t�yp�<res�as�predicates.��A�~completeness���theorem��Xw�Îould�require�additional�w�ork,�but�completeness�is�not�used�in�this�pap�<rer.��������10����9�l�����'������?�with�UUthe�aid�of�whic���h�y�ou�could�deduce������3�x���:���real��?�^�8�x�<��0��)�:����Op���UW��O�fe���㍵x����Ӌ�#����?�(Of�
course�the�pro�Gof�will�ha���v�e�
to�use�some�other�non-logical�axioms�enabling�us����?to�UUpro���v�e�that�squares�are�nev�er�negativ�e.)����NAs�*�an�example�to�clarify�the�rules,�3and�demonstrate�ho���w�t�yp�Ge�em�b�Gedding�is����?handled,�UUconsider�the�comm���utativit�y�UUof�addition�in�the�form�����샸8�x:��real��?�8�y�[�:��int�����(�x�8�+��y�"�=���y����+��x�)����?This�F�is�deriv��q�able�from�the�t���yp�Ge�em�b�Gedding��8�x:��int�����(�x���:���real����)�F�and�the�comm�uta-����?tivit���y�UUla�w�for�reals:�����L��8�x:��real��?�8�y�[�:��real����(�x�8�+��y�"�=���y����+��x�)����?b���y�Ba�straigh�tforw�ard�formalization�of�the�informal�argumen�t:�բLet��x:��real��6�and����?�y�[�:��int��㽲b�Ge�#�giv���en;���then�b�y�the�t�yp�Ge�em�b�Gedding�w�e�ha�v�e��x�)�:���real��B�and�#Ƶy�{�:���real���,����?hence�UU�x�8�+��y�"�=���y����+��x�.�q�Note�that�the�second�t���yping�axiom�w�as�not�used.����N�R��}'emark���:�>9One��8referee�ask���ed�ab�Gout�the�seman�tics�of�this�system,��for�example,����?what���is�the�truth�v��q�alue�of�����$��p����$��fe�����1����NƸ6�=�t�0?��hThe�answ���er�is,�טthat�this�form�ula�has�no����?truth�v��q�alue�b�Gecause�of�the�undened�constituen���t�form�ula.�uA��denition�of�the����?seman���tics��Bw�ould�b�Ge�sup�er
uous,��}b�ecause�of�the�reduction�of��LPT��'�to�ordinary����?logic���in�[�4��].�PMF��*�or�a�discussion�of�in���terpretations�of��LPT��,�and�a�completeness����?theorem��for�this�logic�see�the�exercises�on�page�99�of�Beeson[�3��],��or�see�Beeson[�4��].��!č��?�4��WL�An�ffaxiomatization�of�standard�analysis�����?�In�\this�section�w���e�rst�giv�e�a�\standard"�theory�adequate�for�formalizing�calcu-����?lus.�f�W��*�e�3�will�k���eep�things�simple�b�y�using�just�t�w�o�basic�t�yp�Ges:�`�in�teger�(abbrevi-����?ated���int�)�and��real�.�Q�Here��integer��means�negativ���e�or�nonnegativ�e�in�teger.�Q�The����?basic�*�relation�sym���b�Gols�are�=,�3)�<�,���,�and�*��6�=.�c�The�basic�functions�are��x��q�+��y�[ٲ,�3)�x����y��,����?�x���^��y��·�,�j�x=y�[ٲ,���x�.���W��*�e�f�also�x�once�and�for�all�a�set��F��'�of�unary�functions�to�b�Ge�rep-����?resen���ted,���for���example,�square�ro�Got�and�natural�log.���Mathp��}'ert��actually�includes����?man���y��Dmore�functions.��It�is�imp�Gortan�t,��ho�w�ev�er,�that��Dthe�functions�dened�b�y����?terms�UUin�this�language�form�a�set�closed�under�dieren���tiation.����^��4����?�X-�ff��v�	J=�����"5��-:�4����LܻHence�in��Mathp���ert��[�,�X�when�w�Îe�w�an�ted�to�include�the�gamma�function,�X�w�e�also�had�to��	��include���the�p�<rolygamma�function,���whic�Îh�is�needed�for�the�deriv��ativ�es�of�the�gamma�function.���The���deriv��ativ�Îe�of���K�cmsy8�j��2cmmi8�x�j��is��j�x�j�=x�;��)w�e�don't�mean�to�imply�that�all�the�functions�are�dieren�tiable���ev�Îerywhere.���Sp�<recically��J�,�~Ew�e�hassume�that�for�eac�Îh�unary�function�sym�b�<rol��f�JF�in�the�theory�there���is�'ha�term�whic�Îh�can�b�<re�pro�v�ed�to�denote�the�deriv��ativ�e�of��f�	/�except�at�isolated�p�<roin�ts�of�the���domain��Xof��f��ǻ.��������11�����w�l�����'�������?�4.1��]�Nonlogical��axioms��uT��?�The�UUnonlogical�axioms�of�the�theory�of�standard�calculus�are�as�follo���ws:���8��N(1)�UUThe�sc���hema�of�mathematical�induction����N(2)�UUThe�axioms�of�ordered�exp�Gonen���tial�elds����N(3)�UUDedekind's�sc���hema����N(4)�UUAxioms�ab�Gout�the�sp�ecic�functions�in��F����?�Dedekind's�!psc���hema�expresses,�Tvfor�eac�h�term��t�(�x�)�of�the�theory��*�,�Tvthat�if��t�(�x�)�is����?dened�֊and�uniformly�con���tin�uous�֊on�the�in���terv��q�al��a�����x����b�֊�and�if��a��<�b�;K�^��t�(�a�)���<����?�0�B��^��t�(�b�)�߭�>��0,�g�then�d�9�x��:���real��ߡ�(�t�(�x�)�=�0�B��^��a�߭���x�B��^��x�߭���b�).��The�usual���-����denition����?of�.runiform�con���tin�uit�y�.ris�to�b�Ge�used.��The�use�of�uniform�instead�of�p�oin���t�wise����?con���tin�uit�y�UUis�of�no�signicance.����NUnder��i(2)�w���e�ha�v�e�included�not�only�the�ordered�eld�axioms,�Smbut�also����?axioms�\�ab�Gout�exp�onen���tiation,���including�b�oth�the�usual�algebraic�axioms�and����?axioms�UUrelating�exp�Gonen���tiation�and�order.����NUnder�UU(4)�w���e�include�the�follo�wing�axioms�for�square�ro�Gots�and�logarithms��b������x�����0��!��(����O�p���UW��O�fe���㍵x����s�)�����2��C��=��x�)�����s��x��<��0��!�:����Op���UW��O�fe���㍵x����Ӌ�#��nT���N�x��>��0��!��e������ln��ܟ���x���ز=��x�������x�����0��!�:�����ln�����x��#���썍�?�4.2��]�Puiseux��series��uT��?�A�#�Puiseux�#�series�is�lik���e�a�p�Go�w�er�series,�W�except�the�exp�Gonen�ts�are�m�ultiples�of����?some�V�xed�fraction,���rather�than�in���tegers.��Also,�the�series�is�allo���w�ed�V�to�b�Gegin�with����?a��nite�n���um�b�Ger��of�terms�with�negativ���e�exp�onen���ts.��'Puiseux�series�are�dened����?and�\discussed,���for�example,�on�page�98�of�Siegel[�25��
],�but�the�preceding�brief����?description��9should�b�Ge�adequate�to�understand�this�pap�er.��uIt�is�necessary�that����?the�INelemen���tary�prop�Gerties�of�Puiseux�series�b�e�pro���v��q�able�in��T�c��;�MQfor�example,�K�that����?the��!limit�as��x��approac���hes�zero�of�a�Puiseux�series�in��x��is�equal�to�the�limit�of�its����?leading��^term.�JuF��*�or�this�it�suces�that�T��Apro���v�e��^the�fundamen���tal�prop�Gerties�ab�out����?whic���h��wrational�p�Go�w�ers�of��x��dominate�whic�h�other�rational�p�Go�w�ers�for�small��x�.����?These��>results�can�b�Ge�deriv���ed�from�the�axioms�of�exp�onen���tial�ordered�elds.�KThe����?nonlogical��Aaxioms�for�eac���h�function�included�in�the�theory�should�b�Ge�sucien�t����?to�=�justify�the�calculation�of�Puiseux�series�for�that�function�where�suc���h�a�series����?exists.���썍�?�4.3��]�Some��extensions�of�the�basic�theory��uT��?�This�:Tsection�describ�Ges�three�in���teresting�extensions�of�the�theory��*�,�?�whic�h�for�sim-����?plicit���y�Uha�v�e�b�Geen�omitted�from�the�basic�syn�tax,��but�to�whic�h�the�results�of�the����?pap�Ger�UUcan�b�e�extended.�������12����
�1�l�����'������N�T��*�raditional�W�rst-order�logic�do�Ges�not�allo���w�the�formation�of�terms�with�b�ound����?v��q�ariables,��msuc���h��has��lim���d����x�!�c��#�K�f���(�x�)�or�denite�in�tegrals.��It�is�not�dicult�to�form�u-����?late��the�syn���tax�required�for�allo�wing�suc�h�terms,�:tand�ev�en,�:tusing�the�logic�of����?partial�b3terms,�ekto�allo���w�suc�h�terms�to�b�Ge�undened�sometimes.��b(Since�freshmen����?are���routinely�required�to�use�this�syn���tax,�"�it�had�b�Getter�not�b�e�to�o�dicult�to����?form���ulate.)����NThe�psystem�implemen���ted�in��Mathp��}'ert��has,���in�addition�to�the�t�yp�Ges��int�,����real�,����?and����complex�,��Jthe�t���yp�Ge��notnumber�,�whose�mem���b�Gers�are�denoted�b�y�constan�ts��1�,����?�1�,�q��complexinfinity�,��bounded���E�ff&f��ǫoscillations�,�and�8��unbounded���E�ff&f��oscillations�,����?whic���h�V)can�b�Ge�used�as�v��q�alues�of�limit�terms.�tCThese�terms�are�not�treated�as�un-����?dened,��but��*as�dened�v��q�alues�whic���h�are�not�real�(or�complex)�n�um�b�Gers.�aGThey����?are��vital�for�the�prop�Ger�treatmen���t�of�limits�at�innit�y�and�innite�limits.��
The����?additional�e�t���yp�Ge��notnumber��will�not�b�e�dealt�with�in�this�pap�er,�j but�it�presen���ts����?no�Qdiculties;��the�treatmen���t�of�nonstandard�analysis�b�Gelo�w�extends�to�a�theory����?with�UUthis�t���yp�Ge.����NThe�jsimple�t���yp�Ge�theory�giv�en�here�is�adequate�for�freshman�calculus.��Ho�w-����?ev���er,���the��'logical�framew�ork�used�extends�easily�to�the�t�yp�Ge�theories�prom�ulgated����?in��Ea�series�of�pap�Gers�o���v�er��Ethe�last�couple�of�decades�b���y�S.�F��*�eferman.�ęA��(recen�t����?form���ulation���can�b�Ge�found�in�F��*�eferman[�16��
].��The�results�giv�en�in�this�pap�Ger�will����?extend�(7to�this�setting;�7Aand�the�algorithms�discussed�b�Gelo���w�ma�y�ha�v�e�ev�en�more����?in���teresting��applications�at�higher�t�yp�Ges.�]�In�particular�F��*�eferman's�theories�allo�w����?an�UUin���teresting�mixture�of�t�yp�Ged�and�un�t�yp�Ged�v��q�ariables.��!č��?�5��WL�An�ffaxiomatization�of�nonstandard�analysis�����?�Giv���en��na�t�yp�Ged�language�suc�h�as�has�b�Geen�describ�ed,���w���e�add�a�new�t�yp�Ge�corre-����?sp�Gonding���to�eac���h�old�t�yp�Ge.�>�In�our�case,��[w�e�add�the�t�yp�Ges�of��extende��}'d�{inte�ger����and����?�extende��}'d��mr�e�al�,��"whic���h�aUw�e�abbreviate�as��eint�?��and��ereal�.� rIn�tuitiv�ely��*�,��"the�nonnega-����?tiv���e�1�extended�in�tegers�are�a�nonstandard�mo�Gdel�of�(at�least)�P�eano�Arithmetic,����?and�d�the�extended�reals�are�(at�least)�a�non-Arc���himedean�real-closed�eld�in-����?cluding��ythe�nonstandard�in���tegers.�T�W��*�e�refer�to��int�?��and��real��as�the�\standard"����?t���yp�Ges�UUand��eint�?��and��ereal��as�the�\extended"�t���yp�Ges.����NThe��t���w�o�pieces�of�syn�tax�that�enable�us�sp�Gecically�to�deal�with�nonstandard����?concepts�<�are�as�follo���ws:��a�binary�relation��x����T͍�������+3�����=�����
UN�y�[ٲ,�t�whose�in�tended�meaning�is�\�x����y����?�is��uinnitesimal�(or�zero)",��oand�a�unary�function�sym���b�Gol�[��UX[�x�]�]��ufor�the�standard�part����?of�UU�x�.����NThe����standar��}'d��fr�agment��of�our�theory�consists�of�the�form���ulae�and�terms����?built��Yup�without�men���tioning�the��extende��}'d��	typ�es��Y�ereal�?��and��eint�,���and�without����?men���tioning����T͍��_����+3���_�=�������and�_[��UX[��]�].�~�In�particular,�0!a��standar��}'d�4�formula��is�one�con���taining�no����?v��q�ariables�UU(b�Gound�or�free)�of�t���yp�e��eint�?��or��ereal�,�and�not�con���taining����T͍������+3����=�����q�or�[��UX[��]�].����NW��*�e���shall�list�the�nonlogical�axioms�of�our�theory�precisely�.�?uThe�rst�group����?of�x�nonlogical�axioms�are�the�t���yp�Ge-em�b�edding�x�axioms.��jThey�sa���y�that�an��int�?��is����?b�Goth�UUa��real�?��and�an��eint�,�and�ev���erything�is�an��ereal�.�������13�����z�l�����'����������x���:���int���$�!��x��:���real����͍����x���:���int���$�!��x��:���eint�����Թյx���:���ereal�����N�The�UUnext�axioms�concern�the�congruence�relation:������i��x����T͍�������+3�����=�����
UN�y�"�!���y����T͍������+3����=�����
�'�x������x����T͍�������+3�����=�����
UN�y����^�8�y����T͍��"����+3���"�=�����
�'�z�7��!���x����T͍������+3����=������z�����=x����T͍�������+3�����=�����
UN�x����N�The�9�next�axioms�concern�the�standard�part.��They�sa���y�that�the�standard����?part�M|is�a�partial�function�mapping�extended�reals�to�reals,�Oextended�in���tegers�to����?in���tegers,��whic�h��is�the�iden���tit�y��on�the�standard�reals�and�in���tegers,�and�resp�Gects����?the��congruence�relation.�_LMoreo���v�er,�(�t�w�o��n�um�b�Gers�are�congruen�t�if�they�ha�v�e�the����?same��pstandard�part.�?{(In���tuitiv�ely��*�,�ܞthe��pstandard�part�is�a�partial�function�b�Gecause����?it���is�not�required�to�b�Ge�dened�on�\innite"�extended�reals,���but�only�on�those����?reals�UUwhic���h�are�innitesimally�close�to�a�standard�real.)��������x����T͍�������+3�����=�����
UN�y����^�8�[��UX[�x�]�]���#!��[��UX[�x�]�]��=�[��UX[�y�[ٲ]�]����âb[��UX[�x�]�]��=�[��UX[�y�[ٲ]�]���!��x����T͍������+3����=�����
UN�y�����[�x���:���ereal��?�^�8�[��UX[�x�]�]���#!��[��UX[�x�]�]��:���real������[��x���:���eint��?�^�8�[��UX[�x�]�]���#!��[��UX[�x�]�]��:���int������4r�[��UX[�x�]�]���#!��x����T͍������+3����=�����
UN[��UX[�x�]�]������d�x���:���int���$�!��[��UX[�x�]�]��=��x�����Mfx���:���real���!�!��[��UX[�x�]�]��=��x����N�These���axioms�do�not�prev���en�t���congruence�from�b�Geing�\to�o�small".��-The�fol-����?lo���wing���axiom�sa�ys�that�the�p�Gositiv�e�n�um�b�Gers�congruen�t�to�zero�are�exactly�those����?less���then�the�recipro�Gcals�of�all�the�standard�in���tegers.��rThis�is�called�the��axiom����?of���innitesimals�.�������x����T͍�������+3�����=�����
UN�y�"�$��8�n:��int����(�n�>��0��!�j�x�8���y�[ٸj���<��1�=n�)����NNote��>that�the�v���ersion�of�the�logic�of�partial�terms�w�e�use�requires�all�pred-����?icates���to�b�Ge�\strict".�2�In�particular��x�26�=��y���!��x��#�,���so���the�previous�axioms�imply����?[��UX[�x�]�]�UUis�dened�whenev���er�the�t�yp�Ge�of��x��is�a�standard�t�yp�Ge.����NThe���last�axiom�is�needed�to�mak���e�sure�the�extended�t�yp�Ges�really�are�dieren�t����?from�UUthe�standard�t���yp�Ges:������D˸9�x:��eint��?�8�n:��int����(�x��>�n�)����N�Exer��}'cise��[�:�q�Pro���v�e�UU[��UX[�x�8�+��y�[ٲ]�]��=�[�[�x�]�]�8�+�[�[�y�[ٲ]�].�������14�������l�����'�������?�5.1��]�T���ransference��uT��?�The�˛�tr��}'ansfer��of�a�standard�form���ula��A��is�another�form�ula��A���^������,��'obtained�b�y�c�hang-����?ing�the�t���yp�Ge�of�b�ound�v��q�ariables�from��int�?��to��eint��and�from��real��to��ereal�.�UW(If����?there��w���ere�more�t�yp�Ges�in�our�theory��*�,�Jthey�w�ould�come�in�pairs,�Jstandard�and����?extended,�Zand�%�the�transfer�w���ould�replace�eac�h�standard�t�yp�Ge�b�y�its�extended����?coun���terpart.)����NGiv���en�<)a�theory��T����in�the�standard�fragmen�t�of�our�language,�u�w�e�form�the����?�nonstandar��}'d�;�version�of�T�,�
whic���h�w�e�call��T��c���^�����s�,�9�b�y�adding�to��T�o��all�the�form�ulae����?�A���^������$�l�A��W�for��A��a�standard�form���ula�of��T�c��.���An�y�free�v��q�ariables�in��A��just�pla�y�the����?role�UUof�xed�(standard)�parameters.��UQ���?�5.2��b@Nonstandard��T���reatmen��t�of�Limits��uT��?�Let��صT�5g�b�Ge�an�ordinary�mathematical�theory�form���ulated�in�the�standard�fragmen�t����?of�<�the�language�discussed�ab�Go���v�e.�(W��*�e�<�supp�ose�that��T��O�con���tains�the�sc�hema�of����?mathematical�C�induction,�GWand�at�least�the�axioms�of�the�theory�of�ordered�elds����?(in���whic���h�ev�ery�p�Gositiv�e�elemen�t�has�a�square�ro�Got).�&_This�is�enough�to�dene����?absolute��Bv��q�alues,��}and�hence�to�state�the�usual���-��%�denition�of�limit,�and�pro���v�e����?the�UUbasic�theorems�ab�Gout�that�concept.����NIf�9�T����is�form���ulated�so�as�to�explicitly�include�limit�terms,�>�w�e�supp�Gose�that��T����?�also�Euincludes�a�dening�axiom�sc���hema�making�the�v��q�alue�of�limit�terms�equal�to����?what�*�y���ou�get�from�the���-���ɲdenition.�c�Sp�Gecically��*�,�3cfor�eac�h�term��t��of��T�c��,�3cw�e�ha�v�e��+F����Telim����Sٴx�!�c���d�+�t���=��y�"�$�8�:��real��?�9��Ҫ:��real���8�x:��real���(0���<��j�x�8���c�j���<��'��!�j�t�8���y�[ٸj���<��)��
��?In���this�form���ula�the�v��q�ariable��y�IJhas�t�yp�Ge���real���ܲ,��Qnot���ereal��#*ٲ.�r�Since�w�e�ha�v�e�not����?form���ulated�e��T��P�with�explicit�limit�terms,�i�the�meaning�of�this�form�ula�is�that�the����?left-hand�UUside�is�an�abbreviation�(at�the�meta-lev���el)�for�the�righ�t-hand�side.����NW��*�e�twill�no���w�giv�e�the�nonstandard�c�haracterization�of�limits�implemen�ted�in����?�Mathp��}'ert�'�its�precise�logical�statemen���t.�bVThis�is,�0Eof�course,�not�new:�Z�it�is�just�the����?usual�
~nonstandard�c���haracterization�of�limits.[�23��
]�The�new�p�Goin�ts�ab�Gout�it�made����?here��iare�(1)�this�c���haracterization�can�b�Ge�formalized�in�the�theory��T��c���^�����s�,��nand�(2)����?�T��c���^����	QȲcan�UUb�Ge,�and�has�b�een,�computationally�implemen���ted.����N�R��}'emark���:�x�It�X�should�b�Ge�noted�that�our�theory�is�m���uc�h�X�w�eak�er�than�the�usual����?theories��of�nonstandard�analysis,��3in�that�there�are�no�v��q�ariables�for�real-v�alued����?functions,�؛or��mev���en�sequences�of�reals.�=�All�functions�are�giv�en�b�y�explicit�terms�of����?the�:~theory��*�.�h�This�is�not�imp�Gortan���t�to�the�metho�ds�illustrated�here,�?�whic���h�w�ould����?apply��Bto�an���y�natural�mathematical�theory��*�.�7lIt�do�Ges�mean�that�there�is�something����?to��b�Ge�c���hec�k�ed,���as��the�usual�pro�of�in���v�olv�es��using�the�transference�principle�in�the����?form��8that�sa���ys�sequences�dened�on�the�standard�in�tegers�extend�to�sequences����?dened�UUon�the�extended�in���tegers.��w���?�Theorem.����Ap�0J

cmsl10�AIf�0�T�0�con���tains�the�axioms�of�ordered�elds�and�the�sc�hema�of�math-����?ematical���induction,���then���lim���n1���x�!�c��#���f���(�x�)��=��y��y�Ais���equiv��q�alen���t�in��T��c���^����	��Ato��8��	z:��ereal��?�(�����T͍��)d����+3���)d�=��������?�c�8�^���В�6�=���c��!��f���(��	z�)����T͍������+3����=�����
UN�y�[ٲ)�A.��������15�����}�l�����'������?�Pr��}'o�of���:�b�Supp�Gose����lim����}���x�!�c��#�ϵf���(�x�)��=��y�[ٲ.�ێThen���let������T͍��������+3������=������c�,��but������6�=���c�.�W��*�e�m���ust�sho�w����?�f���(��	z�)����T͍�������+3�����=�����
UN�y�[ٲ.�q�W��*�e�UUha���v�e�b�y�the�denition�of�limit���Í�f՟�8�:��real��?�9��Ҫ:��real���8�x:��real���(0���<��j�x�8���c�j���<��'��!�j�f���(�x�)�8���y�[ٸj���<��)����?Let�Qٵn��b�Ge�a�standard�in���teger,�R�and�tak�e�����=�1�=n�.�p�W��*�e�Q�th�us�ha�v�e�a�standard�real������?�suc���h�UUthat�����8`�8�x:��real��?�(0���<��j�x�8���c�j���<��'��!�j�f���(�x�)�8���y�[ٸj���<��1�=n�)�:���ۍ�?�Note�0�that�transference�is�applicable�since��y��s�is�standard.��By�transference,�gkw���e����?ha���v�e������8�x:��ereal���(0���<��j�x�8���c�j���<��'��!�j�f���(�x�)�8���y�[ٸj���<��)����?T��*�ak���e�UU�x���=���	z�.�q�W�e�UUha���v�e�0���<��j��BZ��8�c�j�.�q�Hence���Í���ȸj��BZ��8�c�j���<��'��!�j�f���(��	z�)����y�[ٸj��<��1�=n:����?�But���since������T͍��ո���+3���ղ=�����
�N�c�,��Hw���e�ha�v�e��j��aA��Wǵc�j�[�<��䔲b�y���the�axiom�of�innitesimals.���Therefore����?�j�f���(��	z�)�8���y�[ٸj���<��1�=n�.�q�Hence,�UUb���y�the�axiom�of�innitesimals,��f��(��	z�)����T͍�������+3�����=�����
UN�y�[ٲ.����NNo���w��Vfor�the�other�direction.�c�W��*�e�pro�Gceed�b�y�con�tradiction.����^��5���
�?�Supp�Gose��x����T͍���ĸ���+3����IJ=��������?�c���!��f���(�x�)����T͍������+3����=�����
UN�y�[ٲ,�UUand�supp�Gose�that��f��(�x�)�do�Ges�not�approac���h��y��.�as��x���!��c�.�q�Then����S��9�:��real��?�(���>��0�8�^�8�k�P�:��int����9�x:��real���(0���<��j�x�8���c�j���<��1�=k��w�^�8�j�f���(�x�)����y�[ٸj���>��)����?Fix�UUsuc���h�a�(standard)���.�q�Then�applying�transference,����|�8�k�P�:��eint��?�9�x:��ereal���(�j�x�8���c�j���<��1�=k��w�^�8�j�f���(�x�)����y�[ٸj���>��)����?Cho�Gose�)�an�innite�in���teger��k�P��,�2iand�let��x��b�e�suc���h�that��j�x�ᓸ��c�j���<��1�=k�2*�^��j�f���(�x�)����y�[ٸj���>����?�.���Then�g�b���y�the�axiom�of�innitesimals,��9�x����T͍���G����+3����G�=�����笵c�,�so�g�w�e�ha�v�e��f���(�x�)����T͍���G����+3����G�=�����笵y�[ٲ,��9whic�h����?con���tradicts�UU�j�f���(�x�)�8���y�[ٸj���>��.������?�5.3��]�Con��v�erting��nonstandard�pro�`ofs�to�standard�pro�ofs��uT��?�What�n�w���e�need�to�kno�w�is�that�if�a�standard�theorem�is�pro�v�ed�using�nonstandard����?metho�Gds���it�is�true.�,In�order�to�quote�the�w���ell-kno�wn���fact�that�ev���ery�theorem����?with�5a�nonstandard�pro�Gof�also�has�a�standard�pro�of,�mw���e�need�to�connect�our����?theory����T��c���^�����
�with�some�theory�in�the�literature,�
�for�whic���h�that�conserv��q�ation�result����?has��b�Geen�pro���v�ed.����^��6���ҫ�There��are�sev���eral�axiomatizations�of�nonstandard�analysis�in����?the���literature.�_�Robinson�himself�b�Gegan�with�a�t���yp�e-theoretic�v���ersion�and�later����?ga���v�e�y�a�set-theoretic�v���ersion[�24��
].�߱Kreisel[�20��]�y�ga�v�e�one�as�a�preliminary�to�his����?theorem.�mMore�GJrecen���tly��*�,�JE.�Nelson�dev�elop�Ged�his�theory��IST�.[�21��
,��22��
GL]�The�most��?��N�ff��v�	J=�����"5��-:�5����LܻClassical�n#logic�seems�necessary�here,��Ub�<recause�the�congruence�form�Îulation�of�limits�do�es��	��not��Xcon�Îtain�the�computational�information�ab�<rout�ho�w�to�compute���"$�as�a�function�of���.��	�>�����"5��-:�6����LܻSuc�Îh�@a�theorem�w�as�rst�pro�v�ed�b�y�Kreisel.[�20���]�The�original�pro�<rof�is�highly�mo�del-theoretic���in���nature,���and�giv�Îes�no�hin�t�of�ho�w�to�transform�a�nonstandard�pro�<rof�explicitly�in�to�a�\corre-���sp�<ronding"��Xstandard�pro�of.��������16�����q�l�����'������?�b�Geautiful���pro�of�of�suc���h�a�conserv��q�ation�result�is�found�in�Nelson's�axiomatization,����?so�UUw���e�will�refer�to�that�v�ersion.����NNelson's�&�theory�is�formed�from�a�theory�of�standard�analysis�b���y�adding�a����?new��Ounary�predicate��standar��}'d��(and�three�axioms,�܄whic���h�w�e�do�not�discuss�here).����?The�(�main�dierence�b�Get���w�een�(��T��c���^����	%%�and��IST�(��is�not�in�the�nonstandard�part�at�all,����?but���in�the�fact�that��IST����is�based�on�set�theory��*�,�Uwhile��T��c���^�����G�is�a�t���yp�Ged�theory�.�R�It�is����?w���ell�&�kno�wn,�0*ho�w�ev�er,�ho�w�&�to�in���terpret�t�yp�Ges�in�set�theory��*�,�0*so�that�eac�h�form�ula����?�A��D�of��T� Ӳhas�an�in���terpretation�in�Zermelo�set�theory��Z���.�?Our�rst�task�is�to�extend����?that�UUin���terpretation�to�an�in�terpretation�of��T��c���^����	QȲin��IST�.����NFirst,�I��iwill��sho���w�ho�w�to�dene�the�fundamen�tal�concepts��x����T͍�������+3�����=�����
UN�y�O[�and��y�"�=��[��UX[�x�]�]����?of�mu�T��c���^����	i�in�terms�of��IST�.�Nelson�denes�\innitesimal"�to�mean�p�Gositiv���e�but�less����?than�O�all�standard�reals,�P�and�then�he�denes��x����T͍�������+3�����=�����
UN�y����if��x�-����y��is�O�innitesimal.�o�One����?can�Hnthen�dene�the�standard�part�of��x��as�that�(necessarily�unique)�standard����?n���um�b�Ger�,N�y��'�suc�h�that��x����T͍�������+3�����=�����
UN�y�[ٲ,�4�if�suc�h�a��y��'�exists.�dThe�fundamen�tal�t�yp�Ges��ereal�and����?�real�,���eint�?��and��R�int�,�also�ha���v�e�natural�in�terpretations�in��IST�,�enabling�us�to����?translate�a�atomic�form���ulae�of�the�form��t���:��A�.� �T�yp�Ges�a�are�simply�in�terpreted�as�sets.����?The���easiest�w���a�y���to�handle�the�denedness�sym���b�Gol�is�to�c�ho�Gose�some�otherwise����?irrelev��q�an���t�l�set��junk��and�mak�e�it�the�v��q�alue�of�all�undened�terms.�$,Of�course,���junk����?�cannot���b�Ge�a�(standard�or�nonstandard)�real�n���um�b�er.�>.Then���the�in���terpretation����?of�d׵x���#��is�just��x��6�=���junk��=Z�.��MThese�rules�determine�a�natural�translation�of��T��c���^����	aJ�in���to����?�IST�.����NNelson�X�pro���v�ed�the�conserv��q�ation�result�that�if��A��is�an�in�ternal�form�ula�pro�v-����?able��in��IST�,�then��A��has�a�pro�Gof�in�the�standard�set�theory�on�whic���h��IST���is����?based.����^��7����:�Applying�UUNelson's�theorem,�w���e�obtain�the�desired�result:����?�Theorem.����ALet���A��Ab�Ge�a�standard�theorem�of��T��c���^�����s�A.�V�Then�the�translation�of��A��Ain���to����?�IST�UU�Ahas�a�standard�pro�Gof�in�classical�Zermelo�set�theory��*�.����N�Pr��}'o�of���:���Let�z;�A��b�Ge�a�standard�theorem�of��T��c���^�����s�.��yThen�the�translation�of��A��in���to����?�IST�,�has�,Na�pro�Gof�in��IST�.�By�Nelson's�theorem,�bit�has�a�pro�of�in�Zermelo�set����?theory��*�.��!č��?�6��WL�An�ffinnitesimal-elimination�algorithm�����?�This��section�con���tains�a�precise�denition�of�the�innitesimal-elimination�algo-����?rithm�o�used�b���y��Mathp��}'ert��ز.�%BThe�purp�Gose�of�the�algorithm�is�to�eliminate�a�nonstan-����?dard�5v��q�ariable�whic���h�w�as�in�tro�Gduced�when�computation�en�tered�a�limit�term,�;�so����?to�aoeac���h�nonstandard�v��q�ariable���	z:��ereal���`�there�corresp�Gonds�an�assumption������T͍��侸���+3���侲=�����
� �x�,����?where�>��x��is�a�standard�v��q�ariable.�.If�the�limit�is�one-sided,�ythere�is�also�an�as-����?sumption����x�S�<���$�(for�limits�from�the�righ���t)�or���] <�x��(for�limits�from�the�left);����?otherwise�԰there�is�an�assumption��x��Y�6�=���	z�.���Th���us�԰eac�h�nonstandard�v��q�ariable�has��?�X-�ff��v�	J=�����"5��-:�7����LܻMoreo�Îv�er,�t�the�T�pro�<rof�is�fully�constructiv�Îe,�unlik�e�T�the�original�pro�<rof�of�Kreisel's�theorem.��	��Nelson��3giv�Îes�an�algorithm�for�con�v�erting�nonstandard�pro�<rofs�in�to�standard�pro�<rofs.�ƿThe�algo-���rithm��Xis�reminiscen�Ît�of�the�G����odel��Diale���ctic�a��X�in�terpretation.��������17������l�����'������?�a���\domain�inequalit���y"�asso�Gciated�to�it.��As�a�matter�of�notation,��w�e�will�use������?�for�/a�(the)�nonstandard�v��q�ariable�asso�Gciated�to�a�limit�as��x��approac���hes��c�,�<fand����?w���e�*?will�use��h��for�the�corresp�Gonding�innitesimal�quan�tit�y����-��⳵c�.�ckStrictly�sp�Geak-����?ing�0�\innitesimal�elimination"�is�only�accurate�when��c�5�=�0;���generally�0�w���e�are����?eliminating�UUa�nonstandard�v��q�ariable���	z�.����NThe���algorithm�to�b�Ge�dened�here�eliminates�only��one��innitesimal.�5�F��*�orm���ulas����?in���v�olving���nested�(iterated)�limits�can�therefore�not�b�Ge�treated�b���y�this�metho�d.����?Whether�(pthis�limitation�can�b�Ge�remo���v�ed�(pb�y�more�careful�b�Go�okk���eeping,�]6w�e�do����?not��kno���w.��The�usefulness�of�the�algorithm�dep�Gends�on�computational�co�de�for����?computing�URlimits,�whic���h�in�turn�relies�on�Puiseux�series�computations,�and�the����?computation�g�of�nested�limits�is�notoriously�more�dicult�than�simple�limits.����?Probably�J�the�dicult���y�of�computation�is�the�main�obstacle�to�extending�the����?metho�Gd,�UUrather�than�an���y�inheren�t�logical�diculties.���6���?�6.1��]�In��tro�`duction��to�the�Algorithm��uT��?�The��Xmetho�Gd�can�b�e�partly�describ�ed�as�a�small�collection�of�rewrite�rules,��Yfor����?reducing�L�logical�form���ulae�to�simpler�logical�form�ulae.�n�One�hop�Ges�that�v��q�ariables����?of���t���yp�Ge��ereal�?��will�b�e�eliminated�b���y�the�rewriting,�֡so�that�a�logical�form�ula����?generated�2�inside�the�scop�Ge�of�a�b�ound�limit�v��q�ariable�(and�hence�in���v�olving�2�an����?innitesimal)�Frwill�b�Ge�reduced�to�a�standard�form���ula�not�in�v�olving�the�b�Gound����?v��q�ariable.�E�The���rules�in���v�olv�e���the�decomp�Gosition�of�an�extended�real�in���to�its�\stan-����?dard�W6part"�and�its�\nonstandard�part",���analogous�to�the�decomp�Gosition�of�a����?complex�'�n���um�b�Ger�in�to�real�and�imaginary�part.�b�W��*�e�ha�v�e�already�in�tro�Gduced�the����?notation�./[��UX[�x�]�]�for�the�standard�part;���let�us�dene���NS�����(�x�)�0�=��x��p���[��UX[�x�]�].��UHere�are����?the�A�t���w�o�most�imp�Gortan�t�rules.�k-(These�are�logical�equiv��q�alences,�E|but�they�will�b�Ge����?used�UUas�rewrite�rules�from�left�to�righ���t.)����N�x��<�y�"�!��[��UX[�x�]�]���<��[��UX[�y�[ٲ]�]�8�_��([��UX[�x�]�]��=�[��UX[�y��]�]�8�^���NS���i�(�x�)���<���NS��*��(�y��)����N�x�����y�"�!��[��UX[�x�]�]���<��[��UX[�y�[ٲ]�]�8�_��([��UX[�x�]�]��=�[��UX[�y��]�]�8�^���NS���i�(�x�)������NS��*��(�y��)����?These�;�rules�do�not�y���et�constitute�the�denition�of�an�algorithm.�iRF��*�or�one�thing,����?w���e���ha�v�en't�said�ho�w�to�apply�the�rules,��or�ho�w�to�compute�the�standard�and����?nonstandard���parts�in���v�olv�ed.��But���more�imp�Gortan���tly��*�,��these�rules�alone�w�on't�do����?the���job.��iConsider�the�example�of�computing�the�deriv��q�ativ���e�of�������p���+7����fe��3荵x�����S�.�Under�the����?limit,��zw���e���ha�v�e�to�simplify������p���
�ڟ��fe����x�8�+��h����%a�#�.�0�This�comes�to�0�����x��>�+��h�.�0�The�standard�part����?of�h2�x�Es�+��h��is��x��and�the�nonstandard�part�is��h�,�l�so�the�rules�ab�Go���v�e�simplify�this�to����?0���<�x�̣�_��(0��=��x�̣�^��0�����h�).�5What��5�Mathp��}'ert��do�Ges�when�simplifying�0�����h��5�is�to�searc���h����?for��sthe�domain�inequalit���y�in�the�list�of�assumptions�(that�is,���in�the�an�teceden�t����?of��the�curren���t�sequen�t).���If�this�inequalit�y�is��h���6�=�0,�E�sho�wing��that�the�limit�it����?came��ofrom�w���as�a�t�w�o-sided�limit,�jthen��Mathp��}'ert��simplies�0�����h��o�to��false�.�P{This����?pro�Gcedure�ȶcannot�b�e�describ�ed�as�a�rewrite�rule,��ev���en�a�rewrite�rule�with�side����?condition.�EHIt���do�Ges�ha���v�e���the�desired�result,��at�least�in�this�example,�b�Gecause�then����?0��c�<�x��1�_��(�x��c�=�0��1�^��0��c���h�)��Osimplies�to�0��c�<�x�.��The��Ononstandard�v��q�ariable��h��has����?b�Geen���eliminated�correctly��*�.�oNote�that�if�w���e�had�b�een�calculating�the�one-sided�������18�����l�����'������?�limit�S�from�the�righ���t,�TA0�����h�S��w�ould�ha�v�e�simplied�to��true��b�y�the�ordinary�rules����?used���b���y��Mathp��}'ert��ز,��using�the�domain�inequalit�y�0�Mg�<�h�,��so���w�e�w�ould�ha�v�e�come����?out�UUwith�0���<�x�8�_��0��=��x�,�UUwhic���h�simplies�to�0�����x�.����NThe�e.third�rule�used�in�the�innitesimal�elimination�algorithm�is�this:��zIf�the����?domain���inequalit���y�for�the�nonstandard�v��q�ariable����[�is���8Ѹ6�=�/W�c�,���then���<�/Wc�,��������c�,����?����>��c�,���and�t�����c�t�all�rewrite�to��false�.��This�rule�ma���y�app�Gear�problematic,����?since��#it�certainly�sho���ws�that�the�\rewrite"�in�question�do�Ges�not�preserv�e�a�direct����?logical��equiv��q�alence.��LHo���w�ev�er,�0�in�the�informal�\in�terv��q�al�seman�tics"�motiv��q�ating����?the�i�algorithm,�o the�form���ula���(��	z�)�means�that����is�true�in�a�neigh�b�Gorho�o�d�i�of�the����?limit��p�Goin���t��c�,�Ҽand�in�terpreted�this�w�a�y��*�,�Ҽit�mak�es�sense�to�rewrite���В>��c��as��false�,����?etc.����NThe�}fstandard�part�of��x�,��kdenoted�[��UX[�x�]�]�}fab�Go���v�e,�is�}fabstractly�dened,�so�the�al-����?gorithm��Vfor�innitesimal�elimination�can't�b�Ge�considered�dened�b���y�the�ab�o���v�e����?three�3rules�un���til�w�e�giv�e�metho�Gds�for�computing�the�standard�and�nonstandard����?part.�B�A�ȏcomputational�Ȳpro�Gcedure��stdpart��(implemen���ted�in��Mathp��}'ert�)�attempts����?to�9�compute�[��UX[�t�]�]�9�for�an�explicitly�giv���en�term��t�.�h�This�function�computes�the�stan-����?dard��part�of�a�v��q�ariable��x��of�t���yp�Ge��ereal�?��or��eint��b���y�lo�Goking�in�the�list�of�curren�t����?assumptions�(�for�an�assumption��x����T͍�������+3�����=�����
UN�t�,�1�and�returning��stdpart(�t�)�.�b�The�standard����?part�5�of�an�expression�not�in���v�olving�5�an�y�v��q�ariables�of�t�yp�Ge��ereal�?��or��eint��is�just����?the��Einput�expression.��On�comp�Gound�expressions�con���taining�nonstandard�v��q�ari-����?ables,�Tk�stdpart�T1�uses�leading�terms�of�Puiseux�series�in�a�w���a�y�T1describ�Ged�in�more����?detail�?uat�the�end�of�this�section.�j}In�applying�the�t���w�o�?urewrite�rules�listed�ab�Go���v�e,����?the�UUfunction��stdpart�(�x�)�is�used�to�implemen���t�[��UX[�x�]�].����NW��*�e�N
assume�that��stdpart��has�b�Geen�correctly�programmed,�O�in�the�sense�that����?it�UUsatises�the�follo���wing�prop�Gerties:����N(1)�]�If��stdpart��returns�successfully�an�answ���er��q����on�input��t��then��q��is�a�standard����?term��and��T��c���^����
_�pro���v�es���;���I�͞�)���q����T͍��`�����+3���`��=�����,o�t�,�C�where��is�the�list�of�assumptions�curren���t����?when�V�stdpart��is�called�and��I��is�the�list�of�assumptions�ab�Gout�the�nonstandard����?v��q�ariable(s)�UUin���v�olv�ed�in��t�.����N(2)���stdpart��will�alw���a�ys��terminate,���either�in�success�or�failure.�-�W��*�e�emphasize����?that��6�stdpart��will�no�doubt�fail�to�compute�a�correct�answ���er�on�a�sucien�tly����?complicated���input,�but�it�will�not�go�in���to�innite�regress.�SFIt�will�succeed�or�giv�e����?up.����N(3)�UUIf��stdpart��terminates�successfully�on�input��t�,�then��T���pro���v�es��������0lim�����Ф�x�!�c���ȏ��t�[�x=�	z�]��=���stdpart��'\��(�t�)���Ǎ�NSimilarly��*�,�~�there�Iis�a�computational�pro�Gcedure��nonstdpart��whic���h�attempts�to����?compute���(an�appro���ximation�to)�the�non-standard�part�of�a�term.�H�Lik�e��stdpart�,����?it�6will�alw���a�ys�6terminate,�nDbut�ma���y�fail�to�compute�the�answ�er.�When�it�do�Ges����?terminate�!�successfully��*�,�U w���e�ha�v�e���nonstdpart��;>��(�t�)����T͍��%����+3���%�=�������h�NS��b�(�t�),�U pro�v��q�ably�in��T��c���^�����s�.�׮Of����?course�H�this�prop�Gert���y�w�ould�b�Ge�satised�b�y�taking���nonstdpart��:e��(�t�)��=�0,�Kobut�H�suc�h����?a�.�w���eak�algorithm�for���nonstdpart��;K�(�t�)�w�ould�not�b�Ge�in�teresting.��The�essen�tial�������19������l�����'������?�prop�Gert���y�?�of���nonstdpart��:\��(�t�)�is�that,�Dwhen�it�terminates�successfully��*�,�it�correctly����?determines�UUthe�sign�of��t�8����stdpart�(�t�)��6c�in�UUa�neigh���b�Gorho�o�d�UUof�the�limit�p�Goin�t.����^��8������N�The��Malgorithm�emplo���y�ed��Malso�mak���es�use�of��simplic��}'ation�.���Simplication�is����?applied�a+b�Goth�to�mathematical�expressions�and�to�logical�form���ulae.��IY��*�ou�ma�y����?think�
�of�simplication�as�something�lik���e�normalizing�a�term�using�rewrite�rules,����?but�v�some�of�the�rules�used�are�more�general�than�rewrite�rules.�ի�infer��op�Gerates����?b���y��simplifying�its�argumen�t,�>hop�Gefully�to��true�,�while��refute��tries�to�simplify�to����?�false�.����NThe�texact�nature�of��infer��and��refute��are�not�germane:��w���e�need�just�a�few����?prop�Gerties�UUof�them.�q�Namely��*�,��Wf��N(1)�K�W��*�e�assume�that��infer��and��refute��alw���a�ys�K�terminate�(whether�successful����?or��enot),��iand�that�if���infer���	�(��)�succeeds,�then����is�pro���v��q�able.�	�More�precisely��*�,��T��c���^�������?�pro���v�es��M�;�������)���,��where��is�the�list�of�assumptions�curren���t�when��infer��is�called,����?and����is�the�list�(often�empt���y��*�,���but�not�alw�a�ys)�of�new�assumptions�generated����?during�طthe�execution�of��infer�.���As�a�matter�of�notation,���w���e�will�alw�a�ys�use�����?for�UUa�list�of�assumptions�generated�during�computation.��Wf��N(2)��RW��*�e�also�assume�that�if��t��simplies�to��q�+�(for�terms��t��and��q�[ٲ)�then��T��c���^�����Ųpro���v�es����?���)��t��=��q�[ٲ.�GSimilarly��*�,���if��A���and�� �1�are�prop�Gositions,�and����simplies�to�� �[ٲ,�then��T��c���^�������?�pro���v�es�V��T�)����$�� �[ٲ.�tIn�practice�this�h���yp�Gothesis�is�satised,�VBb�ecause�the�rules�of����?simplication�^come�directly�from�the�mathematical�axioms�of��T�c��,�`Btogether�with����?simple�UUrules�of�prop�Gositional�calculus.����N(3)��W��*�e�assume�that��infer��is�able�to�pro���v�e��true�inequalities�of�the�form����?�ah���^��u��n�>���0�[Munder�assumptions�of�one�of�the�forms��h�>��0,�\̵h��6�=�0,��h�<��0,�where�[M�u��is����?a�UUrational�n���um�b�Ger.����NW��*�e�{>th���us�ha�v�e,���apparen�tly��*�,�sev�eral�{>algorithms�to�consider:���the�innitesimal-����?elimination���algorithm���v��\ral���;��the�simplication�algorithm;��infer�,��Nand��refute�.���In����?realit���y�KFthese�algorithms�are�not�all�separate:�l�the�innitesimal-elimination�algo-����?rithm�l9and�the�ordinary�mathematical�and�logical�simplications�(whic���h�satisfy����?(1)�l@ab�Go���v�e)�are�b�Goth�incorp�orated�in���to�a�single�algorithm���v��\ral���V�,���and��infer��and����?�refute��ֲb�Goth�w���ork�b�y�calling���v��\ral���.���J�infer����(��)�succeeds�if���v��\ral���(��)�<A=���true��<5�,�-6and�����?�refute��\��(��)�}�succeeds�if���v��\ral�����(��)��Z=���false���K�.��vIn�}�addition�there�are�the�algorithms����?�stdpart�UU�and��nonstdpart�,�whic���h�in�v�olv�e�a�m�utual�recursion�with���v��\ral���k�.����NT��*�o�Qpa���v�oid�confusion:�i�\simplication"�is�not�dened�b�y�a�m�utual�recursion����?with�����v��\ral��(��.�GwThe���w���ord�\simplication"�in�this�pap�Ger�refers�to�logical�and�math-����?ematical���steps�whic���h�ha�v�e�nothing�to�do�with�innitesimal�elimination,�$fand����?whic���h���preserv�e�ordinary�mathematical�and�logical�equiv��q�alence.�V�All�these�steps����?are�
~mathematically�simple�and�represen���t�equiv��q�alences�that�can�b�Ge�stated�and����?pro���v�ed���in��T��c���^�����s�.�R�Ho���w�ev�er,�
Fit���is�allo���w�ed���that�there�b�Ge�simplication�steps�in���v�olving����?the��snonstandard�v��q�ariable.�J#F��*�or�example,��{if��c�?K�=�0��sso�that�the�nonstandard�v�ari-����?able�~��'��represen���ts�an�innitesimal,�P�w�e�migh�t�allo�w�simplifying��sin��D��'��to���	z�.��BNo��?����ff��v�	J=�����"5��-:�8����LܻThis���do�<res�not�imply�that�only�functions�whic�Îh�do�not�oscillate�near�a�limit�p�oin�Ît�can�b�e��	��correctly��handled.�A'Some�suc�Îh�functions�can�b�<re�handled�correctly��J�,�0but�not�b�y�the�part�of�the���algorithm��Xcalling��nonstdpart�.��������20����נl�����'������?�assumption�ʦab�Gout�suc���h�simplications�is�made�for�purp�oses�of�our�correctness����?pro�Gof��&except�the�assumption�of�correctness,��prop�ert���y�(2)�ab�o���v�e.�?;In��&particular����?the���correctness�pro�Gof�do�es�not�require�a�commitmen���t�to�whether�suc�h�simpli-����?cations�>�are�actually�used�or�not.�.�It�should�b�Ge�noted�that�the�simplication����?metho�Gds���ma���y�use�the�curren�t�assumptions�,�tfor�example�b�y�simplifying����to����?�true��X�if����o�Gccurs�in�the�list�.�QF��*�or�complete�precision�w���e�should�sa�y�\�t��simplies����?to���q���with�curren���t�assumptions�"�instead�of�just�\�t��simplies�to��q�[ٲ",���but��will����?alw���a�ys�UUb�Ge�clear�from�the�con���text.����NThe���correctness�pro�Gof�for�our�innitesimal-elimination�algorithm�do�es�re-����?quire���more�detailed�information�ab�Gout�ho���w���v��\ral��fZ�w�orks�on�inequalities��ah���^��u��o��>�5��0.����?Sp�Gecically��*�,��(3)�}�is�ne�when��a��and��u��are�sp�ecic�n���um�b�ers,��but�}�in�general�it����?should�UUb�Ge�strengthened�to�the�follo���wing:����N(4)�����v��\ral���L�will���terminate�on��ah���^��u��
3��>��5�0�when�one�of�the�assumptions��h�<��0,����?0���<�h�,�>�or�8�h��6�=�0�is�in�(or�immediate�from����^��9����_�)�,�>�with�the�(correct)�result��true��or����?�false�,���p�Gossibly��{generating�additional�assumptions�in�the�pro�cess,���if�additional����?assumptions���are�required�to�decide�the�signs�of��a��and��u��and�whether��u��is�o�Gdd����?or�UUnot�(see�b�Gelo���w).�q�Here��a��and��u��are�arbitrary�standard�expressions.����^��10������N�Although��K(4)�is�listed�as�an�assumption�here,��HI��6will�explain�ho���w�the�actual����?algorithms�R)emplo���y�ed�b�y��Mathp��}'ert��are�constructed�to�satisfy�this�assumption.����?These��]details�are�irrelev��q�an���t�to�the�innitesimal�elimination�algorithm�except����?that�UUthey�justify�(4),�whic���h�is�required�for�the�correctness�pro�Gof.����NThe��Edetails�in�question�concern�the�situation�in�whic���h�the�exp�Gonen�t��u��and/or����?the��co�Gecien���t��a��are�not�sp�ecic�rational�n���um�b�ers,�3but��expressions�con���taining����?other�>v��q�ariables.�U�Observ���e�that�t�w�o�things�are�relev��q�an�t�to�determining�the�sign�of����?�ah���^��u��p��under�6Mone�of�the�assumptions��h��>��0,�<��h�<��0,�or�6M�h��6�=�0.�goNamely:�bCIs��u��an�o�Gdd����?fractional���exp�Gonen���t�(that�is,���an�o�dd�in���teger�or�a�quotien�t�with�ev�en�denominator����?and���o�Gdd�n���umerator)?��aWhat�is�the�sign�of��a�?�Observ���e�further�that�if��u��is�o�Gdd����?and��the�assumption�is��h�@I�6�=�0,��:then��the�sign�of��a��is�irrelev��q�an���t,�b�Gecause��ah���^��u���n�will����?tak���e�B�b�Goth�signs,�}�while�if��u��is�ev�en,�}�then�the�sign�of��a��determines�the�sign�of����?�ah���^��u��:b�.���Therefore�\L�Mathp��}'ert��rst�applies�the��infer�,��
�refute�,��assume�\L�metho�Gd�to����?the���question�whether��u��is�o�Gdd.���If�this�can�neither�b�e�inferred�nor�refuted,���it�is����?assumed,�L�and�no�assumption�need�b�Ge�made�ab�out�the�sign�of��a�.��If�it�can�b�e����?refuted,���ho���w�ev�er,�then���the��infer�,��refute�,��assume��metho�Gd�is�applied�to��a��>��0.����?In�UUan���y�ev�en�t,�then,���v��\ral���will�terminate�on��ah���^��u��z�>���0,�establishing�(4).��?��f�ff��v�	J=�����"5��-:�9����LܻThe��Upro�Îvision�ab�<rout�\immediate�from"�is�necessary�b�ecause�if�the�limit�p�oin�Ît�is�not�zero,��	��then��the�domain�inequalit�Îy�whic�h�is�actually�in��has�the�form,���for�example,����h�6�=��c���for�a���nonstandard���v��ariable���,���and��h�\t�=���؍���c�;�ѭso��׾h�\t�6�=�0�is�not�literally�in�,���but�w�Îe�still�need���v��t�al���ݻto���terminate.��	�>�����w���-:�10����LܻIf��/the�expression��a��is�sucien�Îtly�complicated�that��infer��and��refute��cannot�determine�its���sign,��Pbut�w�its�sign�is�nev�Îertheless�actually�determined�b�y�the�assumptions�in�,��Pan�inconsisten�t���assumption�u�ma�Îy�b�<re�added,��in�whic�h�case�the�conclusion�of�the�correctness�theorem�will�b�<re���v��acuously��Xtrue.��������21����#�l�����'�������?�6.2��]�Precise��Denition�of�the�Algorithm��uT��?�The�K�algorithm�in�question,��<here�denoted���v��\ral�����,�tak���es�three�inputs:�^lA�Khlist��of����?(standard)��&assumptions,�ޚt���w�o�assumptions��I���ab�Gout�the�nonstandard�v��q�ariable������?�(namely��*�,��������T͍��U����+3���U�=�����gڵc��ڲand�a�domain�inequalit���y),�and�a�logical�form���ula����(whic�h�ma�y����?con���tain����	z�,���but�no�other�nonstandard�v��q�ariables�or�sym�b�Gols)�to�b�e�simplied.����?The�6�pro�Gcedure���v��\ral�����will�alw���a�ys�6�terminate,�<�either�with�success�or�failure.�g�In�case����?it�k
succeeds,��{it�pro�Gduces�t���w�o�k
outputs:��7the�simplied�form���v��\ral���#�(��)�of���,�whic���h����?for�X�full�precision�m���ust�b�Ge�written�as���v��\ral���	�(�;���I���;���),���and�X�a�(p�ossibly�empt���y)�list����?of��Enew�assumptions�.�<�T��*�o�b�Ge�precise,��w���e�w�ould�ha�v�e�to�write�something�lik�e�����?�v��\ral��M7�(�;���I���;���)�:��prop��!�for�-the�simplied�prop�Gosition,�$iand���v��\ral��OC�(�;���I��;���)�:��assumption��7��for����?,��to��Xindicate�that�the�return�v��q�alue�is�a�pair�consisting�of�a�prop�Gosition�and�a����?list���of�assumptions.���This�w���ould�b�Ge�hop�elessly�unreadable,��}so�instead�w���e�write�����?�v��\ral��M7�(��)�~�for�the�simplied�prop�Gosition�and��for�the�list�of�generated�assumptions.����NMoreo���v�er,�)�w�e��ha�v�e�to�distinguish�t�w�o�kinds�of�success:�Vfwhen�only�rules�that����?preserv���e��logical�equiv��q�alence�ha�v�e�b�Geen�used,�rw�e�sa�y��v��\ral��succeeds�\p�Gositiv�ely".����?Otherwise�odsuccess�is�\negativ���e".���The�exact�denition�of�the�algorithm�m�ust����?indicate��Vat�eac���h�return�whether�the�return�is�with�failure,��with�p�Gositiv�e�success,����?or�Ӌnegativ���e�success.��iThe�w�ords�\p�Gositiv�e"�and�\negativ�e"�ha�v�e�no�in�trinsic����?signifance�UUhere;�w���e�migh�t�as�w�ell�ha�v�e�used�\red"�and�\blue".����NThe�UUdenition�of�the�algorithm���v��\ral���follo���ws:����N�Step�g11.�"ƲWhen��:��v��\ral����op�Gerates�:�on�an���y�input�form�ula���,�tiit�ma�y�rst�p�Gerform����?some���ordinary�logical�simplications,��or�mathematical�simplications�on�the����?terms���in�the�form���ula.�b�The�only�relev��q�an�t�prop�Gert�y�of�these�simplications�is����?that��xthey�preserv���e�logical�equiv��q�alence�under�the�assumptions��;����,��where�����?is���a�(p�Gossibly�empt���y)�list�of�new�assumptions�generated�b�y�the�simplications����?p�Gerformed.�i�If��Ysome�simplication�is�p�erformed,�'Ywith�result�� �Y2�dieren���t�from���,����?return����v��\ral�����(� �[ٲ).�*�If��no�simplications�can�b�Ge�applied,��Qand����con���tains�only�standard����?v��q�ariables,�i
then��e�v��\ral��L�terminates�ewith�p�Gositiv���e�success,�returning����unc���hanged.��If����?no�ʒsimplications�can�b�Ge�applied�and����con���tains�the�nonstandard�v��q�ariable,���go����?to�FNstep�2�if����is�a�disjunction�or�a�conjunction.�l�If�it�is�an�inequalit���y�or�equalit�y����?in�g�whic���h�b�Goth�sides�con�tain���	z�,�lgo�to�step�3.��xIf�it�is�an�equalit�y�or�inequalit�y�in����?whic���h�z�exactly�one�side�con�tains���	z�,��but�that�side�is�not�exactly����,��go�to�step�4.����?If��one�side�is�exactly���(f�and�the�other�side�is��c�,�)�go�to�step�5.�_�If�one�side�is����and����?the�UUother�is�not��c�,�go�to�step�6.����N�R��}'emark��:�>[�On��|input�con���taining�only�standard�v��q�ariables,��v��\ral��do�Ges�nothing�but����?ordinary�|�simplication.��W��*�e�assume�that���v��\ral��0�terminates�on�input�con���taining����?only���standard�v��q�ariables.�t�(This�w���ould�b�Ge�guaran�teed,�Lfor�example,�if�all�the����?simplications��shortened�the�form���ula.�VLIn�practice�some�simplications�lengthen����?the�ysform���ula,��nbut�pro�ving�the�termination�of�ordinary�simplication�pro�Gcedures�is����?not��Ywhat�this�pap�Ger�is�ab�out.)�MtThe�simplications�to�b�e�applied�include�the�la���w����?rewriting��u�A��M�!��B�S�as��B�
f�_���:�A�,��and�the�la���ws�pushing�negation�in�w�ards,��the�la�w����?that�O�\
attens"�disjunctions�whose�argumen���ts�include�a�disjunction�(w�e�allo�w����?disjunctions�4of�man���y�argumen�ts;��for�example��A����_��(�B��_��C���)�4
attens�to��A����_��B��_��C���,�������22����3�l�����'������?�whic���h�2Rmeans��_�(�A;���B��q;�C���)),�9Sthe�2Rla�w�that�
attens�conjunctions,�9Sand�the�la�ws�that����?rewrite��negations�of�inequalities�as�other�inequalities.����^��11�����Our�base�system�has����?no��tatomic�form���ulae�other�than�inequalities�and�equalities;�_�in�general�(if�the����?system���is�enlarged)�other�atomic�form���ulae�migh�t�app�Gear�negated,��Abut��v��\ral��will����?not�j�p�Gerform�an���ything�but�pure�mathematical�simplication�on�suc�h�form�ulae.�#�In����?subsequen���t��,steps�of�the�algorithm,���it�ma�y�therefore�b�Ge�assumed�that�implication����?and�UUnegation�do�not�o�Gccur.����N�Step���2.���On��Gdisjunctions�and�conjunctions,���if���v��\ral��U��do�Ges�not�p�erform�some����?immediate�UUsimplications,�it�is�called�on�eac���h�argumen�t�of�the�input�in�turn:�������v��\ral���J$�(�P����1���S�^��8�:���:�:��g�^�8�P����n��q~�)���$���v��\ral���.�(�P����1��|s�)��^���:���:�:��g�^���v��\ral��o��(�P����n��q~�)�������v��\ral���J$�(�P����1���S�_��8�:���:�:��g�_�8�P����n��q~�)���$���v��\ral���.�(�P����1��|s�)��_���:���:�:��g�_���v��\ral��o��(�P����n��q~�)����?These���form���ulae�are�written�with��$��instead�of�=,���b�Gecause�the�actual�denition�of����?the�{`algorithm�calls�for�computing�the�righ���t-hand�side,���and�then�p�Gerforming�some����?simplications,���suc���h��ras�dropping�duplicate�conjuncts�or�disjuncts�and�com�bining����?certain���inequalities.���The�exact�nature�of�these�simplications�is�not�imp�Gortan���t����?to�\nthe�correctness�pro�Gof�as�long�as�they�preserv���e�(pro�v��q�able)�equiv�alence.��If�an���y����?of��the�recursiv���e�calls���v��\ral��
�(�P����k��됲)�fails,�%�then�return�with�failure.�̯In�the�case�of����?conjunction,��if���all�the�recursiv���e�calls�return�with�p�Gositiv�e�success,��then�return����?return�UUwith�p�Gositiv���e�success,�else�return�with�negativ�e�success.����NDisjunction�{is�treated�dieren���tly��*�.��If�all�the�recursiv�e�calls�return�with�p�Gosi-����?tiv���e���success,���return�with�p�Gositiv�e�success.�cBIf�exactly�one�returns�with�negativ�e����?success,��Yreturn���with�negativ���e�success.�@�If�one�or�more�fails,�return�unsuccessfully��*�.����?If��more�than�one�returns�with�negativ���e�success,��and�one�of�those�returning�with����?negativ���e��vsuccess�is�a�conjunction,���then�use�the�distributiv�e�la�w�on�the�original����?input�MWto�mo���v�e�MWthe�disjunction�in,�N�and�call��v��\ral��recursiv���ely�on�the�result.�oThis�is����?referred�UUto�b�Gelo���w�as�the�\forced�distribution"�clause.����NSince�Ȅdisjunctions�ha���v�e�Ȅb�Geen�
attened�and�negations�ha���v�e�Ȅb�een�pushed�in,����?the�eonly�other�p�Gossibilit���y�is�t�w�o�or�more�negativ�e�successes�on�atomic�form�ulae.����?In���this�case,��>w���e�return�unsuccessfully��*�.�\�This�is�referred�to�b�Gelo�w�as�the�\forced����?failure���clause".��(The�purp�Gose�of�this�clause�should�b�e�made�clear�in�Remark�1����?follo���wing�UUTheorem�3�b�Gelo�w.)����NThe�6�recursiv���e�calls�to���v��\ral���
�on�the�righ�t�can�generate�assumptions�.��The����?question�_�arises�of�ho���w�these�are�handled.��
The�simplest�thing�is�just�to�tak�e����?their�Emunion�and�return�that�as�the��from�the�call�on�the�left.�BIn�actualit���y��*�,����?though,��3the��*calls�on�the�righ���t�are�not�made�in�parallel,�but�sequen���tially��*�,�and�the����?assumptions�s�generated�b���y�the�rst�call�will�b�Ge�part�of�the��for�the�second�call,����?and���so�on.�8dIn�other�w���ords,��only�one�list�of�curren�t�assumptions�is�main�tained,����?and�UUan���y�new�assumptions�generated�are�added�to�that�list.��?�X-�ff��v�	J=�����w���-:�11����LܻThese���la�Îws�include�rewriting��:�(�a���<�b�)���as��b�����a����and�sev�eral�similar�la�ws.�E�W��J�e�treat��a���6�=��b��	���as��Xan�inequalit�Îy�sym�b�<rol,�not�a�negation.��������23����EĠl�����'������N�Step��L3.�7��If���the�input����is�an�inequalit���y��r�5<��s�,�ɬand�b�Goth�sides�con�tain���	z�,�ɬreturn�����?�v��\ral��M7�(0���<�s�8���r�G�).����^��12�������N�The�W�part�of���v��\ral���`�in�whic���h�nonstandard�analysis�is�used�comes�in�to�pla�y�only����?when��9��v��\ral���R�is�9�called�on�an�equalit���y�or�inequalit�y�con�taining�a�nonstandard�v��q�ariable.����?Recall��]that�w���e�ha�v�e�assumed�there�is�only�one�nonstandard�v��q�ariable,��call�it���	z�,����?in�*the�input.�cZInequalities�could�b�Ge�of�the�form��r�5<��s�,�2��r����s�,�2�or�*�r��6�=��s�,�2�and�*input����?�r�ū�=�~��s��i�is�also�allo���w�ed.��F��*�or��isimplicit�y�w�e�consider�here�only��r�ū<�~�s�;��sother�forms����?are�UUtreated�similarly��*�.����N�Step��~4�.�If��hthe�input�is��r�k�<�$�s��and�exactly�one�side�con���tains���	z�,��mmak�e�v��q�arious����?purely��mathematical�simplications�(ab�Gout�whic���h�w�e�need�to�kno�w�only�that����?they��preserv���e�mathematical�equiv��q�alence),���and�then�solv�e�the�inequalit�y�for�the����?nonstandard���v��q�ariable����R�if�p�Gossible.�R(F��*�or�example,��92���<��F�1�will�b�ecome����<����?�1�=�2.)����^��13����3�If���the�result�of�these�computations�is�� �[ٲ,��return���v��\ral��3��(� ��).�TMIf�the�inequalit���y����?cannot�UUb�Ge�solv���ed�for���	z�,�go�to�step�6.����N�Step�/5.�D�If���the�input�is�an�inequalit���y�con�taining�a�single�nonstandard�v��q�ariable����?��	z�,��and�ؾthe�domain�inequalit���y�for����8�is���В�6�=���c��(indicating�that����arose�from�a�t���w�o-����?sided��Llimit�as����Ʋapproac���hes��c�),�Jand�the�input�inequalit�y�has�one�of�the�forms����?��6<<�,�c�,�����>�c�,��������c�,��������c�,������=��c�,���return��T�false�.�(�If�the�input�is���6<�6�=��c�,���return����?�true�.����^��14���挲In�H�these�cases,�Kl�v��\ral��returns�with�negativ���e�success.�m�(All�negativ�e-success����?returns��originate�here,�#Hand�are�passed�up���w�ards��b�y�returns�from�recursiv�e�calls.)����?If�UUneither�side�is��c�,�go�to�step�6.����N�Step��}6�.�6�(W��*�e���get�here�either�from�step�5�with�an�inequalit���y�solv�ed�for����.�whose����?other��?side�is�not��c�,��yor�from�step�4�with�an�inequalit���y�that�could�not�b�Ge�solv�ed����?for��ݵ�	z�.)��`If�the�input����is��r��m<�kPs��with��r����not�zero,��set����to�the�simplied�form�of����?0�i#�<�s�y����r�G�.��vOnce���the�input�has�the�form�0�i#�<�s�,���compute����S����=���stdpart��'�Ų(�s�).�W��*�e����?ha���v�e�W7assumed�that�this�computation�will�alw���a�ys�W7terminate,�W�either�in�success�or����?failure.��If��it�fails,���then���v��\ral��%�(0��M�<�s�)��terminates�unsuccessfully��*�,�returning�0��M�<�s�.����?This�{is�the�only�step�in�the�algorithm�(b�Gesides�the�forced�failure�clause�in�step����?2)���that�can�cause�an�unsuccessful�return�of���v��\ral��ך�.�STIn�this�case,��Othe�nonstandard����?v��q�ariable�UUwill�not�ha���v�e�UUb�Geen�eliminated.����NIf�>-�S�Ѻ�is�successfully�computed,�B�w���e�try�to�determine�its�sign.�jW��*�e�assume�that����?the���computation�of�the�standard�part�has�already�simplied�the�result�so�that����?there���is�no�p�Goin���t�in�trying�to�infer��S�DŲ=��80;�(if�that�w�ould�succeed,��S�u\�is�already����?zero.�bXIn�'case��S����is�zero,�0Kw���e�do�not�pro�Gceed�immediately�to�compute�the�Puiseux����?series�I�using��nonstdpart�,���b�Gecause�in�some�cases�of�in���terest,�the�result�w���ould��?�M3�ff��v�	J=�����w���-:�12����LܻAs�>*a�practical�matter,�X_simplication�do�<res�b�etter�at�pro�Îving�quan�tities�are�p�<rositiv�e�than��	��at��;pro�Îving�one�quan�tit�y�exceeds�another,��tb�<recause�it�is�easier�to�nd�cancellations�and�certain���other��Xsimplications�that�w�Îa�y��J�.��	�>�����w���-:�13����LܻThis���step�will�mak�Îe�further�assumptions�(add�to�the�list�)�if�need�b�<re.�M_F��J�or�example,����q�I{���<��1��Lwhere��q�8ǻis�a�v��ariable�ma�Îy�reduce�to�����<��1�=q��generating��Lthe�assumption�0����<�q�I{�,���if��Lthis���can��Xb�<re�neither�inferred�nor�refuted.�������w���-:�14����LܻIn�N�the�case�of�a�one-sided�limit�from�the�righ�Ît,�l�w�e�N�w�ould�return��true��on��c�&m<��,�lҾc�����,�����\t�=��c�,��Xand��false��on���\t<�c��X�and���\t���c�.��������24����U�l�����'������?�b�Ge�ܫfailure.��W��*�e�rst�call��infer��0����<�s�.����^��15������If�ܫthis�succeeds,���w���e�return��true�,�with����?p�Gositiv���e�UUsuccess.�q�If�not,�go�to�step�9.����NIf����S�>��is�not�zero,��call��infer��on�the�inequalit���y�0���<�S����.�9If���it�succeeds,�then�return����?�true�,�UUwith�p�Gositiv���e�success.����^��16���ꭲIf�not,�go�to�step�7.����N�Step�<�7�.�R#Call���g�infer���(�S�Z�<���0).�If��gthis�succeeds,�	dthen�return��false�,�with�p�Gositiv���e����?success;�UUelse�go�to�step�8.����N�Step��8�.�?�Call����infer��0�9g�6�=��S����.�(It���migh���t�b�Ge�p�ossible�to�infer�this�without�b�eing����?able�k�to�determine�the�sign�of��S����,�q for�example�if�0��'�6�=��S���is�k�a�curren���t�assumption.)����?In���that�case�w���e�add�0�b��<�S�F*�to���the�list��of�new�assumptions�and�return��true�.����?If�UUthe�call�to��infer��do�Ges�not�succeed,�return���,���
�0���<�S��m�_�8�(�S�Z��=�0��^���v��\ral��o��(0��<�s����S����)�:����?�If��the�recursiv���e�call�to��v��\ral��fails,���the�return�is�unsuccessful;�
�otherwise�the�success����?is�UUp�Gositiv���e�or�negativ�e,�the�same�as�is�returned�from�the�recursiv�e�call.���qǟ�^��1������N�Step��]9�.�x)Call����nonstdpart��on��s�.�This�results�in�attempting�to�compute�the����?leading���term�of�the�Puiseux�series�for��s�.��If�this�computation�do�Ges�not�succeed,�����?�v��\ral��M7�(0��E�<�s�)���terminates�with�failure.�NNo���w�supp�Gose�that�it�succeeds.�The�series����?in�F�question�is,�Ib���y�denition�of�the�algorithm��nonstdpart�,�computed�in�p�Go���w�ers����?of����h�\x�=���~��t��c�,��^where��c��is�the�standard�part�of���	z�.����^��2���	��It�is�imp�Gortan���t�to�note�that����?computation�j7of�the�series�m���ust�include�simplifying�the�co�Gecien�ts,��>b�ecause�j7there����?can��b�Ge�cancellations.�Q�Consider,��for�example,��h�wa����sin��i'�h�,�in��whic���h�the�linear�terms����?cancel��out,��and�the�actual�leading�term�is��h���^��3��|s�=�6.���The�leading�term�can�ha���v�e��a����?sym���b�Golic��Zp�o�w�er;�	]the��Zexp�onen�t��Zdo�es�not�ha���v�e�to�b�Ge�a�sp�ecic�rational�n���um�b�Ger.����?Since��simplication�is�not�(and�cannot�b�Ge)�p�erfect,���ho���w�ev�er,�the��leading�term�as����?calculated��nmigh���t�still�really�b�Ge�zero.��W��*�e�therefore�call�it�the�apparen�t�leading����?term,�UUas�the�true�leading�term�w���ould�b�Ge�farther�out�in�the�series.����^��3����?�l?�ff��v�	J=�����w���-:�15����LܻThis��lma�Îy�succeed,���ev�en�in�case��s��has�no�Puiseux�series.��|F��J�or�example,����Mathp���ert��can�infer��	��that��X�e���-:�q�%cmsy6��1�;�cmmi6�=x���k�is�p�<rositiv�Îe.��	�>�����w���-:�16����LܻThis��part�of�the�algorithm�enables�us,��for�example,�to�simplify��h�z�<��1��to��true��inside�a���limit��as��h����!��0.�{�Th�Îus��w�e�will�correctly�generate�no�assumption�when�cancelling��h�����1�inside���suc�Îh��Xa�limit.�������"5��-:�1����LܻIn�gnthe�recursiv�Îe�call�to��v��t�al�,�}jthe�standard�part�of��s�����S���will�gncome�out�zero�in�step�6,�and�w�Îe���will��go�to�step�9.���It�ma�Îy�help�the�reader�to�consider�the�example�of�Section�6.1,��in�whic�h�w�e���ha�Îv�e��Ethe�input�0��T���x����+��h�.�)�In��Ethis�case�the�standard�part�is��x�,���and�b�<roth�the�attempt�to�infer���0�}��<�x��M�and�the�attempt�to�infer��x�}�<��0��Mwill�fail,��Jso�w�Îe�will�return�0�}��<�x���_��(�x�}��=�0���^���v��t�al��
�(0�}��<�h�).���This��Xwill�turn�out�to�b�<re�0�\t�<�x��_��(�x�\t�=�0��^���false��#��),��Xwhic�Îh�simplies�to�0�\t�<�x�.�������"5��-:�2����LܻIt��is�necessary��nonstdpart��m�Îust�\kno�w"�ab�<rout�whic�h�p�<roin�t�to�compute�the�Puiseux���series.�=�There��}seems�to�b�<re�no�w�Îa�y��}to�ensure�this�in�general�except�to�mak�Îe�the�restriction���(whic�Îh���w�e�ha�v�e�made)�that�there�is�only�one�nonstandard�v��ariable,��!arising�from�a�limit�term���and��Xhence�uniquely�asso�<rciated�to�a�p�oin�Ît��c��ab�out�whic�Îh�w�e�are�to�compute�series.�������"5��-:�3����LܻAn�\example�without�extra�v��ariables�in�whic�Îh�the�apparen�t�leading�term�is�actually�zero���w�Îould��8b�<re�complicated,�ǥb�ecause��simplify��will�correctly�handle�simple�examples�lik�Îe��h��P����sin��
���h�,���but���suc�Îh�examples�surely�do�exist.�k\The�situation�of�an�apparen�t�leading�term�whose�sign���cannot�0�b�<re�determined�will�of�course�o�ccur�if�extra�v��ariables�are�presen�Ît�whose�signs�are�not���determined���b�Îy�the�curren�t�assumptions,���for�example�0���<�ah����where��a��is�a�v��ariable�and�no���assumption��Xis�made�ab�<rout�the�sign�of��a�.��������25����g��l�����'������N�Let�˒us�call�the�apparen���t�leading�term�of�the�Puiseux�series��ah���^��u��:b�.�C�Then�return�����?�v��\ral��M7�(0���<�ah���^��u��:b�).����NThat��completes�the�denition�of���v��\ral���"�.�7�Note�that�the�steps�in�the�computation����?b���y��{��v��\ral��.��are�{�of�t�w�o�kinds:���(i)�ordinary�mathematical�and�logical�reductions�that����?preserv���e�iological�equiv��q�alence,�nvand�(ii)�innitesimal-elimination�steps.��The�latter����?ha���v�e�,Kb�Geen�sp�ecied�with�complete�precision.���The�former�ha���v�e�,Knot�b�een�fully����?sp�Gecied;�<�instead�0�sp�ecic�assumptions�ha���v�e�b�Geen�listed�whic�h�they�m�ust�satisfy����?in�k)order�to�mak���e�the�correctness�pro�Gof�w�ork.��EThis�means�that�more�simplica-����?tions��of�kind�(i)�can�b�Ge�added�to�the�algorithm�without�ha���ving�to�rew�ork�the����?correctness�UUpro�Gof.����NIn�7�particular,�p-w���e�ha�v�e�discussed�a�theory�with�nonlogical�axioms�only�for����?natural�L�log�and�square�ro�Got�functions,�N�thereb���y�a�v�oiding�the�need�to�discuss�the����?formalization�Spof�other�functions.��Our�theorem�has�b�Geen�form���ulated�in�suc�h�a�w�a�y����?that���it�immediately�extends�to�a�theory�including�sine,���as�so�Gon�as�the�diligen���t����?reader�2denes�that�function�b���y�suitable�non-logical�axioms�in��T�t��(for�example����?�y��[ٟ�^��00��	*>�=����y�>�and��?�y�[ٲ(0)�=�0�and��y��[ٟ�^��0��*�(0)�=�1)�and�pro���v�es�the�T��*�a�ylor�series�expansion����?of�^6the�sine�function�at�an�arbitrary�cen���ter�on�the�basis�of�those�axioms.��jT��*�o����?co���v�er�g7the�theory�actually�used�in��Mathp��}'ert��ز,�k�this�w���ould�ha�v�e�to�b�Ge�done�for�the����?trigonometric,�8�in���v�erse�1�trigonometric,�h�yp�Gerb�olic,�in���v�erse�1�h�yp�Gerb�olic,�Bessel�1�and����?other�UUsp�Gecial�functions.��!č��?�7��WL�Examples�����?�In��this�section�w���e�trace�the�algorithm�on�t�w�o�illustrativ�e�examples.��
It�is�not����?logically���necessary�to�read�this�section�to�follo���w�the�rest�of�the�pap�Ger.�ϞW��*�e����?con���tin�ue�UUthe�n���um�b�Gering�UUof�examples�from�section�2.����N�Example���8�[�:��:Calculate�t�the�deriv��q�ativ���e�of�����s0�p���
t���s0�few ��Ѝ�f���(�x�)���� �,�|]using�the�limit�denition�of����?deriv��q�ativ���e,�Cwhere��>�f�Ͳis�a�p�Golynomial.�PkW��*�e�will�ha�v�e�to�compute���v��\ral��(T�(0�����f���(�x�p��+��h�))����?for����h����T͍��բ����+3���բ�=�����rb0.�X�Step�1�will�simplify��f���,� =if�it�is�not�already�simplied,�and�pass�the����?inequalit���y���to�Step�4.�S�Here�some�mathematical�simplications�will�b�Ge�made;��Min����?practice�x��Mathp��}'ert��incorp�Gorates�non���trivial�algorithms�for�p�olynomial�inequali-����?ties,��but��they�are�irrelev��q�an���t�here.��Let�us�assume�that�the�resulting�inequalit�y����?can��$still�not�b�Ge�solv���ed�exactly��*�,��.e.g.�D�if��f�ᳲis�irreducible�with�sev�eral�real�ro�Gots.�D�F��*�or����?illustrativ���e���purp�Goses�let�us�neglect�the�fact�that�only�rather�complicated�p�oly-����?nomial��inequalities�w���ould�really�remain�unsolv�ed,�lso�that�w�e�can�think�ab�Gout����?simple���examples�lik���e��f���(�x�)��w=��x���^��2�������1���or��x���^��3���+���x�.�*�The�inequalit���y�will�then�b�Ge����?passed�*�to�Step�6,�3owhic���h�requires�computing�the�standard�part�of��f���(�x�� �+��h�).�c�The����?standard���part�of��f���(�x�8�+��h�)���is��f��(�x�).�F�If��infer��can�sho���w�that�0���<�f��(�x�),��that���will�end����?the���computation;���for�example�if��f���(�x�)�u=��x���^��2����+�~v1.���Assuming���the�eorts�in�Steps����?7�"eand�8�to�determine�the�sign�of��f���(�x�)�also�fail�(as�they�will�in�practice�for�a����?p�Golynomial���that�has�passed�Step�4),���w���e�come�to�Step�9,�in�whic���h��nonstdpart����?�will���b�Ge�called.��HIt�will�compute�the�apparen���t�leading�term�of�the�Puiseux�series����?in�[r�h��of��f���(�x�<�+��h�).�� The�[rrst�candidate�for�that�will�b�Ge��hf����^��0���Ȳ(�x�).�� If��infer��can�sho���w�������26����z��l�����'������?�that�r�0��n�<�f�����^��0���Ȳ(�x�),�zVas�in�the�example��x���^��3����+�L��x�,��v��\ral��will�return��false��on�0��n�<�hf�����^��0���Ȳ(�x�),����?and�UUthe�top-lev���el�call�to��v��\ral��will�compute�����wt0���<�f���(�x�)�8�_��(�f��(�x�)��=�0�8�^���v��\ral��o��(0�����hf������0���Ȳ(�x�))����?whic���h�UUin�the�example��x���^��3���S�+�8�x��w�ould�simplify�(using��x���^��3���S�+�8�x���=��x�(�x���^��2���+�8�1))�UUto������X0���<�x�8�_���v��\ral��o��(0���<�h�)����?and���nally�to�0�{̵<�x�.��
In���the�example��f���(�x�)�{�=��x���^��2�������(�1,����infer����will�not�b�Ge�able�to����?determine�UUthe�sign�of��f�����^��0���Ȳ(�x�).�q�The�computation�will�come�to������ 0���<�x�����2���S��8�1��_���v��\ral��o��(0���<��2�hx�)����?and�:[�v��\ral��will�ha���v�e�:[to�solv���e�the�inequalit�y�0�Dʵ<��2�hx�:[�for��h�.� �T��*�o�do�this�without����?making��<assumptions�ab�Gout�the�sign�of��x�,��t�Mathp��}'ert��returns�(0���<�x�į�^��0���<�h�)�į�_��(�x��<����?�0�`ȸ^��h�*�<��0).�%]When��1�v��\ral��is�called�recursiv���ely�on�this�input,��(the�computation�hits����?the�UUforced�distribution�clause�in�step�2.�q�After�distribution�w���e�get����T��(0���<�x�8�_��x��<��0)�8�^��(0���<�h�8�_��x��<��0)�8�^��(0���<�x�8�^��h��<��0)�8�^��(0���<�h�8�_��h��<��0)����?The���last�conjunct�simplies�to��true�,� }in�view�of��h���6�=�0,�in����Mathp��}'ert�,�although����?this�r�w���ould�not�b�Ge�guaran�teed�b�y�the�general�assumptions�w�e�ha�v�e�made�ab�Gout����?simplication.�\�With�ror�without�this�simplication,�#�v��\ral��applied�to�the�displa���y�ed����?form���ula�_�returns��false��with�negativ�e�success.��[Therefore�the�original�call�to��v��\ral����?�returns��P0��c�<�x���^��2��|���1�1.��This�is�the�correct�condition�for�the�dieren���tiabilit�y�of������?�����p���GUW�����feln�UU��x���r�2���S��8�1����b��.����N�Example���9�[�:�q�Consider�UUthe�one-sided�limit���䍍����lim��-���̣��x�!�3+���������-��p�������-��feln�	�S��x���r�2���S��8�9������?In�]�analyzing�the�domain�of�this�expression,��F�Mathp��}'ert��will�generate�the�expression����?���	z��^��2����m��9�����0,�where��ĵ��>�is�a�nonstandard�v��q�ariable�sub��8ject�to�the�assumptions������T͍��В����+3���В�=�����
^�3����?and�+���В>���3.�c�A���t�Step�1,�3�the�inequalit�y����	z��^��2��kc���v�9�����0�+�will�simplify�to�the�disjunction����?3�k������0�_�z���u)���3.��This��will�reac���h�Step�2,���whic�h�will�call���v��\ral���B�separately�on�the����?t���w�o�}�inequalities�3�
������R�and�}ص����3.��PLet�us�consider������3�rst.��PThis�will�b�Ge����?sen���t�*�to�Step�6,�3@where�it�will�b�Ge�put�in�the�form�0�����3�㬸���	z�.�c�Since�*�the�standard����?part���of�����is�3,���the�standard�part�of�the�righ���t�side�is���6,�and�so�w���e�try�to�infer����?0���<���6,�Jwhic���h���fails,�and�w���e�go�to�Step�7.�zIn�Step�7,�w���e�try�to�infer���6���<��0,����?whic���h�UUsucceeds,�so���v��\ral���returns��false��with�p�Gositiv�e�success.����NNo���w��consider�the�other�inequalit�y��*�,�#D3�٫����	z�.�`This��reac�hes�Step�6�and�is�put����?in��Dthe�form�0�������������3.��Since��Dthe�standard�part�of���龲is�3,�w���e�get�zero�for�the����?standard�xopart,��5so�the�attempted�inferences�in�Steps�6,�7,�and�8�all�fail,�and�w���e����?arriv���e�at�Step�9.��This�calls��nonstdpart��on����2�����3.�This�sets��h���=����2�����3�and����?asks��for�the�leading�term�of�the�Pusieux�series�in��h��of���
ϸ�U�3.���That�leading�������27�����0�l�����'������?�term�3is�of�course�just��h��itself;�l"that�is,�=��a��޲=��u��=�1.��aNo���w��3�v��\ral��FI�(�h�>��0)�3is�called.����?The�F�computation�of���v��\ral��}Ҳ(�h��>��0)�F�reac���hes�Step�5,�I�and�returns��true��with�negativ�e����?success,��since��the�domain�inequalit���y�is�3���<��	z�.�E�If��this�had�b�Geen�a�t�w�o-sided�limit,����?or�!a�one-sided�limit�from�the�left,�'�the�return�v��q�alue�w���ould�ha�v�e�b�Geen��false��with����?negativ���e�UUsuccess.����NReturning�!no���w�to�the�original�call�to���v��\ral��EX�on�the�disjunction,��w�e�ha�v�e�for�the����?results�Soon�the�disjuncts,����false��with�p�Gositiv���e�success�and��true��with�negativ�e����?success.��oThe��7rules�in�Step�2�cause�us�to�return��true��with�negativ���e�success�as����?the��onal�answ���er.��The�meaning�of�this�result�is�that�the�limitand�is�dened�in����?the�'�vicinit���y�of�the�limit�p�Goin�t.�b�If�the�limit�had�b�Geen�t�w�o-sided,�0�the�return�v��q�alue����?w���ould�a!ha�v�e�b�Geen��false�,�dre
ecting�the�fact�that�the�limitand�is�dened�only�on����?one�UUside�of�the�limit�p�Goin���t.�� �x���?�8��WL�A�ffcorrectness�pro�s3of�����?�Our�_�aim�here�is�to�form���ulate�and�pro�v�e�a�theorem�to�the�eect�that�the�pro�Gce-����?dure���describ�Ged�ab�o���v�e���is�correct.�aVThis�theorem�can�b�e�summarized�as�follo���ws:����?Although���v��\ral��A�do�Ges�not�preserv���e�strict�logical�equiv��q�alence,�0�it�preserv�es�equiv��q�a-����?lence��Fin�\neigh���b�Gorho�o�d��Fseman�tics".��Sp�ecically��*�,�a��Fform�ula�in�v�olving����is�true����?in�C�\neigh���b�Gorho�o�d�seman���tics"�if�it�is�true�for�all�real��x��in�a�neigh�b�Gorho�o�d�of�the����?limit�Фp�Goin���t��c�.�E�The�neigh�b�Gorho�o�d�Фin�question�m�ust�b�Ge�a�punctured�neigh�b�Gorho�o�d����?and�UUm���ust�b�Ge�one-sided�if�the�original�limit�w�as�one-sided.����NT��*�o�c�form���ulate�suc�h�a�theorem�precisely��*�,�gDlet���nbhd��c��(�;���c�)�b�Ge�the�(standard)�for-����?m���ula�i~expressing�that����is�true�in�some�punctured�neigh�b�Gorho�o�d�i~of��c��as�a�function����?of�UU�h�,�namely�(in�the�case�of�a�t���w�o-sided�UUlimit)��V㍑G%�9�a:��real��?�9�b:��real����(�a��<�c�8�^��c��<�b�8�^�8�x:��real���(�a��<�x�8�^��x��<�b�8�^��x���6�=��c��!���[�x=�	z�]))����?W��*�e�TKwill�not�treat�the�case�of�one-sided�limits�explicitly�,�T�as�the�details�are�quite����?similar���to�the�treatmen���t�of�t�w�o-sided�limits.��gSince��c��will�b�Ge�clear�from�the����?con���text�UUw�e�will�often�write�simply���nbhd��UI�(��)�instead�of���nbhd���(�;���c�).����NThis�y�seman���tics�dep�Gends�on�eac�h�nonstandard�v��q�ariable����S�b�Geing�clearly�asso-����?ciated�\�with�a�limit�p�Goin���t��c��and�either�a�one-sided�limit�or�a�t�w�o-sided�limit.����?F��*�or�vhthis�reason,�~�our�elimination�pro�Gcedure�assumes�that�there�is�only�one�non-����?standard�"v��q�ariable�and�it�is�so�asso�Gciated.���This�will�b�e�adequate�for�sym���b�olic����?computation�UUnot�in���v�olving�UUnested�limit�terms.���?�Theorem.����A(Elimination��of�an�innitesimal)�Let����Ab�Ge�a�list�of�standard�as-����?sumptions,���and��blet��I�PD�Ab�Ge�the�list�of�t���w�o��bassumptions,���#��:����ereal�����T͍�� t�����+3��� t��=�����+V��c���:���real��x�A,�and����?a�J�\domain�inequalit���y",�L�either���В<��c�A,���>��c�A,�or�J���6�=���c�A.�nMLet����Ab�Ge�a�quan���tier-free����?form���ula,���standard��except�that�it�ma�y�con�tain���	z�A.�HZThen���v��\ral��"�(��)��=���v��\ral���.�(�;���I���;���)���Awill����?terminate.�^�If����v��\ral��RԲ(��)���Aterminates�with�success,�'Clet����Ab�Ge�the�list�of�an���y�additional����?assumptions�UUgenerated�during�the�computation�of���v��\ral���k�(��)�A.�q�Then:�����N�(i)�UU�and���v��\ral���k�(��)�are�b�Goth�standard.�������28�����D�l�����'������N�(ii)�o�T��c���^����	kz�pro���v�es��;�������)���nbhd����(��)��$���v��\ral��)�(��)�where��nbhd�(giv���en�precisely�ab�Go�v�e)����?expresses�UUthat����is�true�in�a�punctured�neigh���b�Gorho�o�d�UUof��c�.����N(iii)�j��;�������)���nbhd���(��)��$���v��\ral��"�(��)�is�pro���v��q�able�in�classical�set�theory��*�,�pGand�hence����?true.����N�R��}'emark���1�[�:��<Consider�c�the�example�input�������w�c���_��c�����	z�.��vIf�c�step�1�of�the����?algorithm��fails�to�simplify�this�input�to��true�,�so�that�it�w���ould�b�Ge�passed�to�step����?2,�y�then�rm�v��\ral��will�b�Ge�forced�to�return�unsuccessfully��*�,�b�Gecause�it�will�succeed�neg-����?ativ���ely�/�on�eac�h�(atomic)�disjunct.�e8Since��v��\ral��returns��false��on�eac�h�disjunction,����?without�#�the�\forced�failure"�clause�w���e�w�ould�get�the�incorrect�answ�er��false��on����?this�Ғinput;�0the�answ���er�is�incorrect�b�Gecause�the�input�is�a�theorem.��~Of�course,����?if��hthe�logical�simplier�in�step�1�of�the�algorithm�do�Ges�an���ything�at�all,��it�is����?lik���ely��to�catc�h�this�example�b�Gefore�it�is�forced�to�fail,��but�there�will�certainly����?b�Ge��imore�complex�examples�whic���h�escap�e�the�simplier�and�pro�duce�disjuncts�of����?this�UUkind.����N�R��}'emark��2�[�:��oThe��)theorem�do�Ges�not�place�a�restriction�on�the�syn���tactical����?form�>�of���.�-�Ho���w�ev�er,�x�the�>�theorem�only�states�what�happ�Gens�if���v��\ral��u��(��)�can�b�e����?successfully��computed.��The�\forced�failure"�clause�in�Step�1�of�the�algorithm����?th���us��)has�a�similar�eect�to�a�restriction�on�the�syn�tax�of���.�.The�follo�wing�section����?of��the�pap�Ger�will�sho���w�that�(under�additional�assumptions)�the�forced�failure����?clause���is�nev���er�used;��<this�is�similar�to�remo�ving�a�syn�tactical�restriction.���This����?form���ulation���is�b�Getter�b�ecause�it�clearly�separates�the�questions�of�the�correctness����?of�����v��\ral��L��and���the�applicabilit���y�of���v��\ral���Ͳ.��F��*�or�example,��the�forced�failure�clause�isn't����?used�\�on�the�example���܄���
�c�=��_��c�����e��b�Gecause�\�it�gets�simplied�to��true��in�Step�1����?of�UUthe�algorithm.����N�R��}'emark��3�[�:���Consider��\the�example�arising�from�the�limit�computation�of����?the��deriv��q�ativ���e�of�������p���Y7����fe��3荵x����S�.�}jWhen�this�w�as�informally�considered�(Section�6.1),�/�the����?analysis��3of�0��4���x���+��h��3�led�to�the�form���ula�0��4�<�x���_��(0��4=��x���^��0��4���h�).�obAlthough����?this�{2is�a�disjunction,���only�one�disjunct�con���tains�the�nonstandard�v��q�ariable��h�,�so����?it��meets�the�restriction.�As�it�turns�out,��Ethough,�when��the�algorithm�is�traced����?on��this�example,�@this�disjunction�is�nev���er�passed�as�input�to��v��\ral��an�yw�a�y��*�.��See����?the�UUfo�Gotnotes�in�the�pro�of.����N�Pr��}'o�of���:���The�|�rst�part�of�the�theorem�concerns�the�termination�of���v��\ral����.��OThis����?is��apro���v�ed�b�y�a�straigh�tforw�ard�induction�on�the�computation�of���v��\ral���w�,���using�the����?assumptions�
�that��infer�,�;��refute�,��stdpart�,��nonstdpart�,�and�
��simplify��alw���a�ys����?terminate.�.The��Bonly�clause�that�can�cause�failure�of���v��\ral��K��is�the�failure�to�compute����?a���Puiseux�series.��	W��*�e�no���w�turn�to�the�pro�Gof�of�(i),��(ii),�and���(iii)�under�the����?assumption�UUthat���v��\ral���k�(��)�returns�successfully��*�.����NAccording���to�Nelson's�theorem,��[(ii)�implies�(iii).����^��4�����W��*�e�turn�to�the�v���erication����?of��(i)�and�(ii).�-�There�are�three�separate�cases�to�the�pro�Gof,���according�to�the�sense��?��f�ff��v�	J=�����"5��-:�4����LܻW��J�e�aw�Îould�lik�e�to�strengthen�(iii)�to�pro�v��abilit�y�in��T�.:�.�jHThat�can�probably�b�<re�done,���if��T��	���includes�e{the�t�Îyp�<re�constructor��A�L��!��B�d��,���b�y�e{adapting�the�pro�<rof�of�Nelson's�theorem,���but�if��T����do�<res�bnot�con�Îtain�the�constructor��A��/�!��B�d��,�$then�bthe�p�ossibilit�Îy�of�claiming�pro�v��abilit�y�in��T�.��in���(3)��Xis�an�op�<ren�question.��������29�����!�l�����'������?�of��6the�domain�inequalit���y�in��I�p�(that�is,���whether�it�came�from�a�left�limit,�righ���t����?limit,��or��Ct���w�o-sided�limit).�@�W��*�e�deal�explicitly�only�with�the�t�w�o-sided�case;���the����?t���w�o�UUone-sided�cases�are�similar.����NBefore�tgiving�the�pro�Gof�of�(ii)�in�detail,�&;w���e�p�oin���t�to�a�k�ey�step.�^'By�standard����?tec���hniques�UUof�nonstandard�analysis,���nbhd��UI�(��)�is�equiv��q�alen�t�in��T��c���^����	QȲto������r�8��	z:��ereal���(�����T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!���)�:����?�Therefore�UUthe�follo���wing�is�equiv��q�alen�t�to�(ii)�in��T��c���^�����s�:����~�|�I���;�����;�����)�8��	z:��ereal��?�(�����T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!���)��$���v��\ral���.�(��)����?The�B�heart�of�the�pro�Gof�is�the�follo���wing�simple�observ��q�ation:�h�in�case�the�last�step����?is��the�problematic�reduction�of��c��<��ߓ�to���false�,��it�w���orks,�b�Gecause��8��	z:��ereal��?�(�����T͍��В����+3���В�=��������?�c�8�^���В�6�=���c��$��c�<��	z�)�UUis�equiv��q�alen���t�to��false��in��T��c���^�����s�.����NIn�7Rorder�to�handle�the�maxim���um�v��q�ariet�y�of�input�form�ulae,�=Sw�e�ha�v�e�to�k�eep����?trac���k���of�whether��v��\ral��returns�with�p�Gositiv�e�or�negativ�e�success.�J�The�imp�Gortan�t����?p�Goin���t��Jab�out�p�ositiv���e�success�is�this:���w�e�will�pro�v�e�that�if���v��\ral��2`�(��)�returns�with����?p�Gositiv���e�UUsuccess,�then������m��8��	z:��ereal���(�����T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!��(��v��\ral��7�(��)��$���))�����1(�iv�[ٲ)�����NF��*�or�f'example,��j[�v��\ral���q�(���<��!c�D�+�1)�is�going�to�b�Ge��true�,�j[with�p�ositiv���e�success,�j[and����?for�UUan���y������T͍��В����+3���В�=�����
^ȵc�,�w�e�will�ha�v�e���В<��c�8�+�1.����NW��*�e�)�will�pro���v�e�)�(ii)�b���y�induction�on�the�n�um�b�Ger�of�steps�of�kinds�(1)�to�(9)�in����?the�ݥcomputation�of���v��\ral����(��),���sim���ultaneously�pro�ving�(iv)�in�case���v��\ral����(��)�returns����?with�
�p�Gositiv���e�success.�Y�Note�that�(iv)�implies�(ii).�The�basis�case�(immediate�re-����?turn)��Bcorresp�Gonds�to�step�1�in�the�denition�of���v��\ral���X�.�@�In�that�case�no�nonstandard����?v��q�ariables�O�are�in���v�olv�ed,�P�and�O�w�e�ha�v�e�assumed�that���v��\ral���6�terminates�and�preserv�es����?logical���equiv��q�alence�in�the�sense�that��;�����Uz�)����$���v��\ral�����(��)���is�pro���v�able�in�T.�But����?since���no�nonstandard�v��q�ariables�are�presen���t,�����nbhd����(��)�A=�$������is�a�logical�trivialit�y��*�,����?hence�UUpro���v��q�able�in�T.�Then�(ii)�follo�ws�immediately��*�.����NThere��are�man���y�cases�in�the�induction�step,�1�corresp�Gonding�to�steps�2-9�of����?the���denition�of���v��\ral�����.���W��*�e�tak���e�these�cases�one�b�y�one.���In�eac�h�case�w�e�rep�Geat����?the�,�relev��q�an���t�clause�of�the�denition,�4�so�that�the�reader�will�not�need�t�w�o�copies����?of�UUthe�pap�Ger�to�follo���w�the�pro�of.����NCase�3�(2).�f�First�consider�conjunctions.�F��*�or�notational�simplicit���y�w�e�consider����?a���conjunction�of�t���w�o���argumen�ts.�)�The�computation�rule�implies�(using�the�fact����?that�UUsimplication�preserv���es�pro�v��q�able�equiv�alence)�������e�v��\ral����{�(�A�8�^��B��q�)���$���v��\ral���.�(�A�)�8�^���v��\ral��o��(�B��)�:����?�By�UUinduction�h���yp�Gothesis�w�e�ha�v�e��������v��\ral���!�(�A�)���$�8��	z�(�����T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!��A�)�������30�����=�l�����'��������S��v��\ral����ܲ(�B��q�)���$�8��	z�(�����T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!��B��)����?Hence�������Y�v��\ral����o�(�A�8�^��B��q�)���$�8��	z�(�����T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!��A�8�^��B��)����?in�UUview�of�the�logical�iden���tit�y������m�8��	zP��o�^�8�8��Q���$�8��	z�(�P��^�8�Q�)�:����?�If��+��v��\ral��bβ(�A��̸^��B��q�)�+�returns�with�p�Gositiv���e�success,�aQthat�is�b�ecause�b�oth���v��\ral��bβ(�A�)�and�����?�v��\ral��M7�(�B��q�)�Mdo�so,�N�in�whic���h�case�w�e�establish�(iv)�easily:�m�for������T͍��В����+3���В�=�����
^ȵc�(Z�^���В�6�=���c�,�N�w�e�Mha�v�e����������v��\ral���!�(�A�8�^��B��q�)�����ۀ�$���������v��\ral������(�A�)�8�^���v��\ral��o��(�B��q�)��������ۀ�$�������A�8�^��B������?�That�UUcompletes�the�argumen���t�for�conjunction.����NNo���w�UUconsider�disjunctions.�q�Since�the�iden�tit�y������m�8��	zP��o�_�8�8��Q���$�8��	z�(�P��_�8�Q�)�:����?�is�.ein�general�not�v��q�alid,�d�w���e�cannot�carry�out�a�similar�pro�Gof�for�arbitrary�dis-����?junctions.�NHo���w�ev�er,��xin���the�denition�of��v��\ral��w���e�imp�Gosed�the�restriction�that�at����?most�:one�of�the�recursiv���e�calls�returns�with�negativ�e�success.��w(Otherwise�the����?forced���distribution�clause�calls��v��\ral��recursiv���ely�on�a�logically�equiv��q�alen�t�form�ula;����?in��cthat�case�there�is�nothing�to�pro���v�e.)�$�Therefore��c(since�for�simplicit���y�w�e�are����?sho���wing��the�pro�Gof�only�for�t�w�o�disjuncts)�w�e�can�assume�that���v��\ral��9��(�A�)�returns����?with�UUp�Gositiv���e�success,�so�that�b�y�(iv)�w�e�ha�v�e�����o�8��	z�(�����T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!��(��v��\ral��7�(�A�)��$��A�))����?Then�UUto�complete�the�pro�Gof�w���e�need�only�the�v��q�alid�iden�tit�y�������P��o�_�8�8��	zQ���$�8���(�P��o�_�8�Q�)����?in�1�whic���h��P��z�do�Ges�not�con�tain���	z�.��(Here��P��z�is���v��\ral��i�(�A�).)�T��*�o�mak���e�the�argumen�t����?clear�]let�us�write��8��	z::A��to�abbreviate��8��:��ereal���(�����T͍��В����+3���В�=�����
^ȵc���^���В�6�=���c��!��A�).�_ Then�]w���e����?ha���v�e��������C��v��\ral���z��(�A�8�_��B��q�)�������n�$��������o�v��\ral���+��(�A�)�8�_���v��\ral��o��(�B��q�)����������n�$�������o8��	z::A�8�_�8��::B����������n�$�������o8��	z::��v��\ral��7�(�A�)�8�_�8��::B����������n�$��������o�v��\ral���+��(�A�)�8�_�8��	z::B����������n�$�������o8��	z::�(��v��\ral��7�(�A�)�8�_��B��q�)����������n�$�������o8��	z�(�A�8�_��B��q�)���������31���� ��l�����'������N�If��'�v��\ral��?=�(�A����_��B��q�)�'returns�with�p�Gositiv���e�success,��that�is�b�ecause�b�oth���v��\ral��?=�(�A�)�and�����?�v��\ral��M7�(�B��q�)�Mdo�so,�N�in�whic���h�case�w�e�establish�(iv)�easily:�m�for������T͍��В����+3���В�=�����
^ȵc�(Z�^���В�6�=���c�,�N�w�e�Mha�v�e����������v��\ral���!�(�A�8�_��B��q�)�����ۀ�$���������v��\ral������(�A�)�8�_���v��\ral��o��(�B��q�)��������ۀ�$�������A�8�_��B������N�Case��
(3).��If�the�input����is�an�inequalit���y��r�g�<� �s�,��wand��r��'�is�not�zero,�and�b�Goth����?sides�UUcon���tain���	z�,�return���v��\ral���k�(0���<�s�8���r�G�).����NArgue�UUas�follo���ws�in��T��c���^�����s�:��������r���)���v��\ral���.�(�r�5<�s�)������f:�$�������f;�v��\ral���Q�(0���<�s�8���r�G�))���������f:�$�������f;�nbhd��f/�(0���<�s�8���r�G�)���������f:�$�������f;�nbhd��f/�(�r�5<��s�)������?(b���y��
denition�of���v��\ral��� �).�-The�latter�is�equiv��q�alen�t�(b�y�induction�h�yp�Gothesis�pro�v��q�ably����?in��h�T�c��)�to���nbhd���\�(0�Cߵ<�s�7����r�G�).��Since���h�v��\ral��
~�(�r���<�C�s�)��$���v��\ral��z��(0��<�s�7����r��)),�2�w���e��hha�v�e�����?�v��\ral��M7�(�r�5<��s�)��$���nbhd����(0��<�s��'���r�G�).�^SEviden���tly����nbhd���(0���<�s��'���r��)���$���nbhd����(�r�5<�s�),�&�so��w���e����?ha���v�e��UU�v��\ral���k�(�r�5<��s�)��$���nbhd����(�r�<�s�)�UUas�desired.����NIf�UUthe�return�is�with�p�Gositiv���e�success,�then�(iv)�is�immediate:�������.����)���v��\ral���.�(�r�5<�s�)������ʩ�$�������ʪ�v��\ral����(0���<�s�8���r�G�)���������ʩ�$������ʪ�0���<�s�8���r���������ʩ�$������ʪ�r�5<��s�������N�Case�EU(4).�A�If�the�input�is��r��4<�Ws��and�exactly�one�side�con���tains���	z�,��Up�Gerform����?some�Wmathematical�simplications�that�preserv���e�logical�equiv��q�alence�and�then����?solv���e��the�inequalit�y�for�the�nonstandard�v��q�ariable����.�if�p�Gossible.�O�(F��*�or�example,����?2��W�<�N!�1��[will�b�Gecome���<�N!�1�=�2.)�d�If�the�result�of�these�computations�is�� �[ٲ,���return�����?�v��\ral��M7�(� �[ٲ).����NSolving��tan�inequalit���y�can�generate�an�assumption,�ӻso�the��returned�from�����?�v��\ral��M7�(�r�l�<�%{s�)���can�b�Ge�nonempt���y��*�.��Ho�w�ev�er,��a���solv�ed�inequalit�y�is�equiv��q�alen�t�to�the����?original��	inequalit���y�under�the�generated�assumptions�(if�an�y).�@�That�is,���T��c���^�����|�pro�v�es�����3M�;�������)�� �"�$��r�5<�s:����?�By�UUinduction�h���yp�Gothesis,��T��c���^����	QȲpro�v�es�����,4���)���v��\ral���.�(� �[ٲ)��$���nbhd����(� ��)�:����?�Therefore�UU�T��c���^����	QȲpro���v�es�����48�;�������)���v��\ral���.�(� �[ٲ)��$���nbhd����(�r�5<�s�)�������32����!��l�����'������?�But�UUsince���v��\ral���k�(�r�5<��s�)�=���v��\ral���.�(� �[ٲ),�UU�T��c���^����	QȲpro���v�es��UV���<<�;�������)���v��\ral���.�(�r�5<�s�)��$���nbhd����(�r�<�s�)����?whic���h�UUis�(ii)�in�this�case.����NIf��jthe�return�is�with�p�Gositiv���e�success,��then�w�e�pro�v�e�(iv)�as�follo�ws:�A�in��T��c���^�����ݲw�e����?ha���v�e�UUb�y�induction�h�yp�Gothesis��UV������;�������)�� �"�$���v��\ral���.�(� �[ٲ)����?Hence�����o��;�������)���v��\ral���.�(0��<�r�G�)��$�� ��fg��?�Hence�����O��;�������)���v��\ral���.�(0��<�r�G�)��$��0��<�r����N�Case��M(5).�ǰIf�the�input�inequalit���y�con�tains�a�single�nonstandard�v��q�ariable���	z�,����?and��;the�domain�inequalit���y�for������is���T�6�=�J��c��(indicating�that����arose�from�a�t���w�o-����?sided��Llimit�as����Ʋapproac���hes��c�),�Jand�the�input�inequalit�y�has�one�of�the�forms����?��k<�a�c�,��,��>�c�,��,�����c�,��,�����c�,��,return���false�.���If�the�input�is���k�6�=��c�,��,return��true�.����?Since�H�the�return�in�this�case�is�alw���a�ys�H�with�negativ���e�success,�KUw�e�do�not�ha�v�e�to����?establish�UU(iv).����NLet���us�tak���e�the�input�forms�one�at�a�time.�2:Supp�Gose����is���В<��c�.�Then���nbhd�����(��)����?sa���ys���that�for�all��x��in�some�punctured�t�w�o-sided�neigh�b�Gorho�o�d���of��c�,�>�x��<�c�,�whic���h����?is�refutable�since�eac���h�suc�h�neigh�b�Gorho�o�d�includes�some�p�Goin�ts�larger�than��c�.����?Hence�����nbhd���̲(��)�MG�$���v��\ral���]�(��),���since���b�Goth�are�equiv��q�alen���t�to��false�.�cPSimilarly�if����is����?�c��<��	z�,��Zsince���eac���h�punctured�t�w�o-sided�neigh�b�Gorho�o�d���of��c��con�tains�some�p�Goin�ts����?less��}than��c�.�	%?The�non-strict�inequalities�are�treated�the�same�w���a�y��*�.�Finally�,����?consider�*{the�input���В�6�=���c�.�cF��*�or�this���,��3
�nbhd��3�(��)�sa���ys�that�for�all��x��in�a�punctured����?neigh���b�Gorho�o�d�"of��c�,�O�x��6�=��c�.��.Since�"it�is�a��punctur��}'e�d�"�neigh�b�Gorho�o�d,�Othis�"is�true.����?Hence��UU�v��\ral���k�(�;���c�)���$���UU�as�desired,�since�in�this�case���v��\ral���(�;���c�)�is��true�.���NCase��	(6).�TIf�the�input����is��r�5<��s��with��r�C&�not�zero,�
�set����to�the�simplied�form����?of�UU0���<�s�8���r�G�.�q�If�the�input�has�the�form�0���<�s�,�compute��S�Z��=���stdpart��'\��(�s�).����NThe���replacemen���t�of��r�o�<�(�s��b�y�an�equiv��q�alen�t�form�preserv�es�pro�v��q�able�equiv�a-����?lence,�A�so�=w���e�ma�y�assume�w�e�are�already�in�the�case�of�input�0���<�s�.�i�In�=this�part����?of�3�the�pro�Gof�w���e�are�assuming�that���v��\ral�����do�es�not�fail,�:~so�w���e�can�assume�that��S��U�is����?successfully��Gcomputed.���If��S�Բis�zero�w���e�go�to�step�9,��so�under�this�case�w�e�ma�y����?assume��*�S�.�6�=��{0.��HThe�denition�of���v��\ral���j�then�sa���ys�to�call��infer��on�the�inequalit�y����?0���<�S����.�fIf�2Mit�succeeds,�9Oreturn��true�,�with�p�Gositiv���e�success.�fIf�it�do�es�not�succeed,����?w���e�
4are�in�case�(7),�7lso�under�this�case,�w���e�ma�y�assume��infer��returns��true��on����?0���<�S����.����NSince��.the�success�is�p�Gositiv���e,���w�e��.are�required�to�establish�not�only�(ii),�but����?(iv).�H�Since��^(iv)�implies�(ii),��it�will�suce�to�pro���v�e��^(iv).�W��*�e�m���ust�sho�w�in��T��c���^�������?�that�������;�������)������T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!��(��v��\ral��7�(0��<�s�)��$��0��<�s�)�������33����"�p�l�����'������?�Since��UU�v��\ral���k�(0���<�s�)�UUis��true�,�this�b�Goils�do���wn�to�sho�wing��Y!���E�;�������)������T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!��0��<�s����?�Using���the�nonstandard�denition�of�limit�(Theorem�1),�
}it�w���ould�suce�to�pro�v�e�����x��;�������)���ߤ�lim�����x�!�c����j�s�[�x=�	z�]��>��0��J荑?W��*�e�Riha���v�e�assumed�that�the�algorithm��stdpart��has�the�prop�Gert�y�that��T��c���^����	Nܲpro�v�es�������բlim�������x�!�c����|h�s�[�x=�	z�]��=���stdpart��'\��(�s�)����?so�UUit�will�suce�to�sho���w�that��T��c���^����	QȲpro�v�es������R�;�������)��0��<���stdpart��'\��(�s�)����?That�UUis,�since��S�Z��=����stdpart��'\��(�S����),����̓��;�������)��0��<�S�T:����?�W��*�e�Dxkno���w�that��infer��returns��true��on�0���<�S����.�l(W�e�Dxha���v�e�assumed�(prop�Gert�y�(1)�of����?�infer�)�UUthat��infer��is�sound,�so��T��c���^����	QȲpro���v�es�UU�;�������)��0��<�S����,�as�desired.���A��NCase�m(7).��When�the�input�has�the�form�0���<�s�,�K�and��m�stdpart��(��(�s�)�=��S����,�call�����?�infer��V��(�S��<����0).��4If��ythis�succeeds,��Bthen�return��false�,�with�p�Gositiv���e�success.��4Else����?go�UUto�step�8.����NW��*�e��ma���y�supp�Gose�then�that���infer��*3�(�S��<��x�0)�succeeds.��vW�e�ha���v�e��to�pro���v�e��not����?only��(ii),���but�(iv),�b�Gecause�the�success�is�p�ositiv���e.�M�W��*�e�are�required�then�to�sho�w����?that�UU�T��c���^����	QȲpro���v�es��������;�������)������T͍��В����+3���В�=�����
^ȵc�8�^���В�6�=���c��!�:�(0��<�s�)�:���ፑ?�As�UUin�the�previous�case,�it�w���ould�suce�to�kno�w��T��c���^����	QȲpro�v�es��Y!���x��;�������)���ߤ�lim�����x�!�c����j�s�[�x=�	z�]��<��0����?But�UUw���e�kno�w������%�;�������)���ߤ�lim�����x�!�c����j�s�[�x=�	z�]�=��S������?�and��T��c���^����
��pro���v�es��;�������)��S���<��0�b�Gecause��infer��succeeds�on��S���<��0.��The�desired����?conclusion�UUfollo���ws�immediately��*�.���A��NCase��(8).���If�the�input�is�0��L�<�s���and���stdpart��(f��(�s�)��L=��S����,��call���infer��0��L�6�=��S��.���(It����?migh���t�ݳb�Ge�p�ossible�to�infer�this�without�b�eing�able�to�determine�the�sign�of��S����,����?for��example�if�0���6�=��S����is��a�curren���t�assumption.)�UIn�that�case�w�e�add�0���<�S����to��the����?list�8�of�new�assumptions�and�return��true�,�=�with�p�Gositiv���e�success.�hIf�the�call�to����?�infer�UU�do�Ges�not�succeed,�return��Y!���
�0���<�S��m�_�8�(�S�Z��=�0��^���v��\ral��o��(0��<�s����S����)�:��������34����#��l�����'������?�If��the�recursiv���e�call�to��v��\ral��fails,���the�return�is�unsuccessful;�
�otherwise�the�success����?is�UUp�Gositiv���e�or�negativ�e,�the�same�as�is�returned�from�the�recursiv�e�call.����NFirst���consider�the�case�in�whic���h���infer����(0��ʸ6�=��S����)���succeeds.�n�T��*�o�establish�(iv),����?w���e�UUha�v�e�to�pro�v�e�����O�;����0���<�S�Z��)��0��<�S��$���true���
9��?�whic���h��7is�a�trivialit�y��*�.�7o(The�in�teresting�thing�here�is�that�hop�Gefully�this�will�not����?create�con���tradictory�assumptions�in�the�an�teceden�t,�Obut�since�in�general�that����?cannot��b�Ge�en���tirely�a�v�oided,��there�is�nothing�to�pro�v�e�ab�Gout�it.��Note�that�this����?case��8comes�up�only�when�there�are�v��q�ariables�or�complicated�expressions�whose����?order�UUrelation�to��c��cannot�b�Ge�determined�b���y�the�system.)����NNo���w���supp�Gose�that�the�call�to���infer�����(0�/e�6�=��S����)���do�es�not�succeed.�-�Supp�ose�rst����?that�the�recursiv���e�call�to��v��\ral��returns�with�p�Gositiv�e�success.���Then�w�e�ha�v�e�to����?establish�UU(iv).�q�By�induction�h���yp�Gothesis,�w�e�ha�v�e��k_����4���)���v��\ral���.�(0��<�s�8���S����)���$��0��<�s�8���S����?�Hence�������
����)���v��\ral���.�(0��<�s�)�������x�$�������y�S�Z��=��0�8�^���v��\ral��o��(0��<�s����S����)����������x�$�������y�S�Z��=��0�8�^��0��<�s����S����������x�$�������y�0���<�s����k_��?�whic���h�UUpro�v�es�(iv).����NNo���w�$w�e�pro�v�e�(ii).�U�Supp�Gose��infer��0���6�=��S����succeeds.�U�W��*�e�ha���v�e�to�sho�w�that��T��c���^�������?�pro���v�es���������)��0��<�S�Z��$���nbhd����(0��<�s�)�:��
9��?�As�UUin�case�(6),�w���e�ha�v�e��T��c���^����	QȲpro�v�es���������)���ߤ�lim�����x�!�c����j�s�[�x=�	z�]�=��S���_��?�and�UUhence,�under�the�assumption�0���<�S����,�UUw���e�ha�v�e���nbhd��UI�(�s��<�c�)�UUas�desired.����NIf�UUthe�attempt�to�infer�0���6�=��S���do�Ges�UUnot�succeed,�w���e�are�to�return������0���<�S��m�_�8�(�S�Z��=�0��^���v��\ral��o��(0��<�s����S����))�:����?�By�UUprop�Gert���y�(3)�of��stdpart������̵S�Z��=���ߤlim������x�!�c����j�s�[�x=�	z�]��]&��?from�UUwhic���h�it�follo�ws�that���������nbhd����ݲ(0���<�s�)��$��0��<�S��m�_�8�(�S�Z��=�0��^���nbhd��8Բ(0��<�s�))����?or�UUequiv��q�alen���tly�����~j��nbhd���jIJ(0���<�s�)��$��0��<�S��m�_�8�(�S�Z��=�0��^���nbhd��8Բ(0��<�s����S����))�������35����$��l�����'������?�But�UUb���y�induction�h�yp�Gothesis,�w�e�ha�v�e��������)���nbhd����(0��<�s�8���S����)���$���v��\ral���.�(0��<�s�8���S��)�:����?�Therefore����wv_���)���nbhd����(0��<�s�)��$��0��<�S��m�_�8�(�S�Z��=�0��^���v��\ral��o��(0��<�s����S����)�:����?�In�UUview�of�the�denition�of��v��\ral��in�this�case,�that�is�nothing�but�����N���)���nbhd����(0��<�s�)��$���v��\ral���.�(0��<�s�)�;����?�whic���h�UUw�as�what�w�e�had�to�pro�v�e.����NCase�BK(9).�8�The�input�0�R�<�s�BK�has���stdpart��(��(�s�)�R=�0.�Call�BK�nonstdpart��on��s�.����?This��results�in�attempting�to�compute�the�leading�term�of�the�Puiseux�series����?for��s�.�YSince�w���e�ha�v�e�assumed���v��\ral��B*�(0���<�s�)�returns�successfully��*�,��w�e�kno�w�that�this����?computation�f�succeeds.�"CThe�series�in�question�is�computed�in�p�Go���w�ers�f�of��h���=���e?��[ŵc�,����?where�_ӵc��is�the�standard�part�of���	z�.��BLet�us�call�the�apparen���t�leading�term�of�the����?Puiseux��series��ah���^��u��:b�.�,aThen���v��\ral���5�(0���<�ah���^��u���)��will�b�Ge�called.�,aIf�this�is�a�t���w�o-sided��limit,����?a�`�domain�inequalit���y�equiv��q�alen�t�to��h���6�=�0�`�is�in�the�assumption�list�.���Therefore,����?b���y��	assumption�(4)�in�Section�6.1,�%�the�return�v��q�alue�from���v��\ral��3�(0���<�ah���^��u��:b�)��	will�b�Ge����?�true�ҽ�or��false��(p�Gossibly�this�computation�adds�an�assumption�ab�out�the�sign����?of���a��to�the�previously-presen���t�assumptions,��but�since�this�is�the�recursiv�e�call,����?b���y�UUinduction�h�yp�Gothesis�that�assumption�is�already�in�the�list�).����NConsider�Ƙrst�the�case�in�whic���h�the�return�v��q�alue�is��true�.�ŒBy�induction����?h���yp�Gothesis�UUw�e�ha�v�e������r���)���nbhd����(0��<�ah�����u��:b�)����?Since�UU�T���pro���v�es�the�elemen�tary�prop�Gerties�of�Puiseux�series,�w�e�ha�v�e����Ʋ����)���nbhd����(0��<�s�)����?whic���h�UUestablishes�(ii).����NIn�UUcase�the�success�w���as�p�Gositiv�e,�w�e�ha�v�e�b�y�induction�h�yp�Gothesis����������)��h����T͍������+3����=�����
UN0�8�^��h���6�=�0��!��(��true����$��0��<�ah�����u��:b�)����?and�UUhence����������)��(�h����T͍������+3����=�����
UN0�8�^��h���6�=�0��!��0��<�ah�����u��:b�)����?Again�<6using�the�fact�that��T��Ųpro���v�es�<6the�elemen���tary�prop�Gerties�of�Puiseux�series,����?w���e�UUha�v�e�����������)��h����T͍������+3����=�����
UN0�8�^��h���6�=�0��!��0��<�s����?�establishing�UU(iv)�in�case���v��\ral���k�(0���<�ah���^��u��:b�)�=���true����.�������36����%�l�����'������N�No���w��consider�the�case���v��\ral���,�(0�W��<�ah���^��u��:b�)�=���false�����.�vBy��induction�h�yp�Gothesis�w�e����?ha���v�e�UU�T��c���^����	QȲpro�v�es�����a���)���false���!�$���nbhd����(0��<�ah�����u��:b�)��6D��?from�UUwhic���h,�b�y�the�elemen�tary�prop�Gerties�of�Puiseux�series,�w�e�ha�v�e���ƍ��˦���)���false���!�$���nbhd����(0��<�s�)����?establishing�UU(ii)�for�this�case.����NIf�UUthe�success�is�p�Gositiv���e,�w�e�ha�v�e�����c����)���false���!�$��0��<�ah�����u�����?�whence�UUw���e�obtain�����˦���)���false���!�$���nbhd����(0��<�s�)��6D��?establishing�UU(iv)�for�this�case.��&Q`���?�9��WL�When�ffwill�v���al�succeed?�����?�The�._ab�Go���v�e�theorem�con�tains�as�a�h�yp�Gothesis�the�assumption�that���v��\ral��eu�(��)�termi-����?nates��Mwith�success.�`�This�is�the�result�of�practical�in���terest:���it�guaran�tees�that����?w���e��get�no�wrong�answ�ers�from�inferences�made�b�y��Mathp��}'ert��during�limit�cal-����?culations.���Ho���w�ev�er,�rit�lFis�also�of�in���terest�to�determine�when�the�algorithm�do�Ges����?succeed,���or���when�it�do�Ges�succeed�and�do�es�not�generate�a�con���tradictory�set�of����?assumptions.��The�din���terest�of�this�question�is�b�Goth�theoretical�and�practical.����?The���practical�application�is�this:��^if�the�algorithm�succeeds,���but�generates�con-����?tradictory���assumptions,��asoft���w�are�using�this�algorithm�to�guard�against�incorrect����?steps��ma���y�still�tak�e�an�incorrect�step.��A�t�least�in�that�situation�w�e�will�kno�w����?it,���pro���vided���that�w�e�recognize�the�assumptions�as�con�tradictory;���but�of�course����?it�$w���ould�b�Ge�b�etter�to�pro���v�e�$that�under�certain�conditions�this�do�es�not�happ�en.����?That�UUis�what�w���e�shall�do�in�this�section.����NOriginally�Z�I�Z?had�hop�Ged�that�\all�expressions�in�the�form���ula�ha�v�e�Puiseux����?series��sthat�the�system�can�compute"�should�suce�for�success,��;and�the�deter-����?minabilit���y�
of�the�signs�of�the�leading�co�Gecien�ts�in�the�Puiseux�series�should����?b�Ge�Y+sucien���t�for�success�without�con�tradictory�assumptions.����^��5������This�turns�out�to����?b�Ge�/dtrue,�e�but�only�if�one�mak���es�stronger�assumptions�ab�out�the�simplication����?algorithm�<�than�are�necessary�for�the�correctness�pro�Gof.�i�There�are�t���w�o�<�places�in����?the��algorithm�for��v��\ral��that�can�pro�Gduce�failure:�)�the�failure�to�compute�a�Puiseux����?series,��Xand��Xthe�forced�failure�clause�in�step�2.�=�The�question�w���e�tak�e�up�in�this����?section�6is�whether�the�forced-failure�clause�has�an���y�real�eect:�N8is�there�an�input����?����on�whic���h��v��\ral��will�actually�use�the�forced-failure�clause?��The�answ�er�is:��RY��*�es,����?if�UUsimplication�is�w���eak;�No,�if�simplication�is�strong�enough.��?�Ѱ�ff��v�	J=�����"5��-:�5����LܻThis�Hnsucien�Ît�condition�migh�t�b�<re�made�ev�en�stronger,�e4as�for�example��v��t�al��will�w�ork�on��	��0�\t�<�e���-:��1�=h����and��Xev�Îen�on�0�\t�<�he���-:��1�=h���������37����&��l�����'������N�Recall��that�our�elimination�algorithm�con���tains�some�steps�of�\simplication"����?whic���h��Rha�v�e�not�b�Geen�completely�sp�ecied;��Sthe�correctness�pro�of�dep�ends�only�on����?the�I�assumption�that�these�steps�preserv���e�logical�equiv��q�alence,�L	and�hence�applies����?to���v��q�arious�v���ersions�of�the�algorithm,��with�\w�eak"�or�\strong"�simplication����?steps���allo���w�ed.��W��*�e�shall�sho�w�that�the�answ�er�to�the�question�is�not�the�same����?for�UUev���ery�v�ersion�of�the�algorithm�co�v�ered�b�y�the�correctness�theorem.����NOne���ob���vious�condition�whic�h�will�guaran�tee�a�v�oidance�of�the�forced-failure����?clause��is�that�the�input����should�b�Ge�a�conjunction�of�equalities�and�inequalities.����?Indeed,���a��disjunction�of�suc���h�things�can�b�Ge�allo�w�ed,���pro�vided�at�most�one�of�the����?disjuncts��con���tains�the�nonstandard�v��q�ariable���	z�.�=This�w�ould�not�allo�w���В����c��?�_��c������?��	z�;�'�but���this�form���ula�and�man�y�others�will�b�Ge�allo�w�ed�if�w�e�allo�w����to�b�Ge�suc�h����?that�]�after�Ysimplic��}'ation��it�has�the�form�just�stated.�\uF��*�or�example,�")��В����c���_��c��������?�will�K^simplify�to��true��in�step�1.�S�(This�is�not�guaran���teed�b�y�an�y�assumptions����?w���e�}ha�v�e�stated�ab�Gout�the�simplication�algorithm,���but�it�w�ould�b�Ge�true�of�an�y����?reasonable�:simplier,�s8and�certainly�is�true�of��Mathp��}'ert�.)��Another�example�is����?���<���c������1��_��c��+�1��<��	z�.���This��has����x�in�t���w�o��disjuncts,��and�will�not�simplify�(or����?at��least�migh���t�not�simplify��*�,�'if�simplication�is�w�eak).�^�But�when���v��\ral��n"�is�called�on����?the���disjuncts,���b�Goth�return��true��with�p�ositiv���e�success,���so�indeed���v��\ral����turns�out����?to�F�return��true��with�p�Gositiv���e�success�on�this�form�ula�to�Go.�l�This�is�the�reason�for����?distinguishing�UUt���w�o�kinds�of�success.����NThe���example�at�the�end�of�Section�6�is�an�indication�that�without�more����?assumptions��Bab�Gout�simplication,�Žthe�forced�failure�clause�can�b�e�encoun���tered����?in��natural�problems.�_�There�w���e�needed�to�kno�w�that�simplication�could�reduce����?�c���<���X�_��޵��
<�c��Ҳto��true��in�the�presence�of�the�assumption��c����6�=���	z�,��qwhic���h���is�not����?guaran���teed�pb�y�the�h�yp�Gotheses�of�the�correctness�theorem.���Ho�w�ev�er,�v�ev�en�more����?assumptions�URare�needed,�USas�the�follo���wing�discussion�will�sho�w.�q�T��*�o�construct�an����?example���that�w���ould�encoun�ter�the�forced-failure�clause,���w�e�w�an�t�a�disjunction����?�A��$�_��B��.�suc���h�K�that���v��\ral��ΐ�succeeds,��Wbut�with�negativ�e�success,��Won�b�Goth��A��and��B��q�.����?F��*�or�Dsimplicit���y�let�us�tak�e��c�
P�=�0,�G�and�Dtry�to�nd��A��in�the�form�0�
P�<�u�D�and��B����?�in��mthe�form�0����<�v�[ٲ.�W��*�e��mm���ust�mak�e�sure�that�0����<�u��D�_��0����<�v�=F�do�Ges��mnot�simplify����?as��*in�the�example�0�hz�<����_�ym�0��<����	z�,��_whic���h��*simplies�to��true��when�0�hz�6�=������is��*in����?.��W��*�e��try�to�construct��u��and��v�>߲so�that�0��>�<�u��U�_��0��>�<�v��is��pro���v��q�able�in��T��c���^�����s�,�rbut����?the��_system�can't�simplify�it�to��true�,�&!and�indeed��v��\ral��reduces�eac���h�disjunct�to����?�false�.�fNBut�2��v��\ral��will�fail�unless�it�can�compute�the�Puiseux�series�of��u��and��v�[ٲ,�9�so����?they��can't�b�Ge�arbitrarily�complicated.�U�Y��*�ou�migh���t�try�0���<���	z��^��1�=�3��-d�_�� �0��<�����	z��^��1�=�5���D�,��but����?�Mathp��}'ert��l�will�simplify�this�to�0���<����_����0��<������and��lthen�to��true�.�RIn�general,����?the�sconstruction�of�a�coun���terexample�along�these�lines�w�ould�dep�Gend�on�the����?Puiseux�ӻseries�computations�b�Geing�more�capable�than�the�simplication�rules.����?F��*�or�UUexample,�one�could�try������&0���<���sin���޵�BZ�_�8�0��<�������arctan��#������?�If���the�system�w���ere�able�to�compute�the�Puiseux�series,���but�not�to�simplify�the����?inequalities���0�#��<���sin��M���y�and�0��<�������arctan��#���	z�,��w���e�w�ould�get�a�coun�terexample.�������38����'i�l�����'������?�Nothing�in�the�assumptions�used�in�the�pro�Gof�of�the�innitesimal�elimination����?theorem�p�w���ould�prev�en�t�this�b�Geing�the�case.��This�sho�ws�that�if�simplication�is����?to�Go�Q,w���eak�to�replace�innitesimal�terms�with�the�leading�terms�of�their�Puiseux����?series,�UUthe�forced�failure�clause�can�b�Ge�encoun���tered.����NNo���w��sI��Swill�sho�w�that�if�simplication�is�strong,��the�forced�failure�clause�is����?nev���er�5used.��hSupp�Gose�that�simplication�is�allo�w�ed�to�compute�Puiseux�series����?in�7�h�@��=����d����c��and�reduce�inequalities�b���y�replacing�eac�h�side�with�the�leading����?term���of�its�Puiseux�series.�n�If�the�signs�of�the�co�Geen���ts�can�b�e�determined,����?eac���h���inequalit�y�(or�equalit�y)�con�taining���w�on�whic�h��v��\ral��returns�successfully�will����?simplify�to�an�inequalit���y���В<��c�,�)(�c�<��	z�,�or�to��true��or��false�,�or�to�a�form���ula�not����?con���taining�k���	z�,�qNunder�the�assumptions��c����T͍���d����+3����d�=�����
���;�����޸6�=��d�c�.���This�set�of�four�inequalities����?is���closed�under�conjunction�and�disjunction,��so�an���y�disjunction�con�taining�more����?than��Pone�inequalit���y�in�v�olving����ʲcan�b�Ge�simplied�to�one�in�v�olving�only�one�suc�h����?inequalit���y��*�.����NIf���the�signs�of�the�co�Gecien���ts�cannot�b�e�determined,���as�in�the�example�at�the����?end���of�Section�6,��w���e�will�get�a�disjunction�of�conjunctions�of�one�or�the�forms����?�P��ڸ^��K�c���<��	z�,��or��w�P��^��K���<�c�,��or��wjust��P�c��,�where��P�N�do�Ges�not�con���tain���	z�.�1-Either�the����?equalit���y-solving�F�co�Gde�itself�or�the�forced�distribution�clause�in�step�2�will�cause����?the��0distributiv���e�la�w�to�b�Ge�used�on�suc�h�a�disjunction.�1The�result�is�a�conjunction����?of�ڦdisjunctions.�H�If�these�disjunctions�con���tain�more�than�one�inequalit�y�in�v�olving����?��	z�,��they���can�b�Ge�com���bined�in�to�one�as�ab�Go�v�e.��Hence�the�form�ev�en�tually�passed����?to��'�v��\ral��will�b�Ge�a�conjunction�of�disjunctions�of�the�forms��P���^�B���В<��c�,��ʵP��^��c��<��	z�,���or����?�P�c��,��where�#�P�o��do�Ges�not�con���tain���	z�.�YaSince�these�disjunctions�ha�v�e�only�one�disjunct����?in���v�olving�Oϵ�	z�,�P�the�computation�of��v��\ral��cannot�encoun���ter�the�forced�failure�clause.�� �󍍑?�10��_fdRelated�ffW���fork�����?�This�טis�not�the�rst�use�of�nonstandard�analysis�in�automatic�deduction;��Bal-����?lan���t�yne�N�and�Bledso�Ge�ha���v�e�N�exp�erimen�ted�N�with�it[�1��,��2��	N�].�^F��*�rom�the�viewp�oin���t�of����?automated��,deduction,���I��
think�that�the�metho�Gd�has�m���uc�h��,more�p�oten���tial�than����?has�G�b�Geen�exploited�so�far�in��Mathp��}'ert��to�supp�ort�education�in�calculus.�H�The����?Univ���ersit�y�<�of�T��*�exas�group�under�Bledso�Ge's�direction[�9��,��10��
<�]�righ���tfully�considered����?it��an�imp�Gortan���t�ac�hiev�emen�t�when�their�pro�v�er�could�automatically�nd�a�pro�Gof����?that�’the�sum�of�t���w�o�’con�tin�uous�functions�is�con�tin�uous�(using�the�standard����?epsilon-delta�M�denition�of�con���tin�uit�y).�Z�In�M�nonstandard�analysis,���this�theorem����?b�Gecomes�UUa�trivialit���y��*�.����^��6����:�Let��h��b�e�an�innitesimal,�and�compute:�����]�2(�f�Lo�+�8�g�[ٲ)(�x��+��h�)��=��f���(�x�8�+��h�)�+��g��(�x��+��h�)����T͍�������+3�����=�����
UN�f���(�x�)�+��g��(�x�)��=�(�f�Lo�+�8�g��)(�x�)����NThe�@�essen���tial�reason�wh�y�nonstandard�analysis�w�as�helpful�in��Mathp��}'ert��and����?in�I�the�ab�Go���v�e�I�example,���is�that�t���w�o�I�alternating�quan���tiers�are�reduced�to�a��?�ș�ff��v�	J=�����"5��-:�6����LܻThis��Vobserv��ation�is�in�no�w�Îa�y��Vmean�t�to�detract�from�the�ac�hiev�emen�t�of�the�UT��Gpro�v�er�in��	��nding�::a�standard�pro�<rof.��lThis�same�researc�Îh�group�pioneered�the�use�of�non-standard�analysis���in��Xautomated�deduction.��������39����(,��l�����'������?�quan���tier-free��form�ulation.���Once�w�e�ha�v�e�a�quan�tier-free�form�ulation,�Bpro�Gof����?searc���h��can�b�Ge�reduced�to�rewrite-rule�st�yle�computation,���with�tremendous�gains����?in���eciency��*�.���This�is�true�for�computerized�inference,��and�probably�for�h���uman����?inference�UUas�w���ell.����NThe�~�idea�that�side�conditions�needed�for�op�Gerations�can�b�e�added�to�the����?assumption��8list�has�also�b�Geen�used�in�W��*�u's�metho�d�for�pro�ofs�in�geometry��*�,���as����?implemen���ted�UUb�y�Chou[�13��
].����NOther���systems�allo���wing�the�use�of�partial�terms�ha�v�e�b�Geen�implemen�ted,����?for��aexample�IMPS[�14��
,��15��
�c].��The�in���tegral�has�b�Geen�treated�as�a�v��q�ariable-binding����?op�Gerator�UUb���y�Keisler[�18��
].����NThe��problem�of�com���bining�a�theorem-pro�v�er�with�a�computer-algebra�system����?has�Ȅb�Geen�addressed�exp�erimen���tally�b�y�Harrison�and�Th���Gery[�17��
]�(who�link�ed�HOL����?and��Maple),��and�Clark���e�and�Zhao[�12��
],�who�wrote�a�theorem�pro���v�er���A���nalytic��}'a����?�in��%the��Mathematic��}'a��language.�RbThe�former�pap�Ger�p�oin���ts�out�that�it�is�safe�to�use����?the���result�of�a�sym���b�Golic�in�tegration�p�Gerformed�b�y�a�p�Gossibly�unreliable�computer����?algebra�6�system,�<�if�y���ou�can�c�hec�k�the�result�b�y�dieren�tiation�within�the�pro�v�er.����?The�%psecond�is�a�direct�attempt�to�k���eep�trac�k�of�conditions�of�op�Gerations.��A�t����?the��)end�of�the�pap�Ger,���the�authors�call�for�building�an�in���tegrated�system�for����?computer�Y�algebra�and�logic.�<This�is�what�has�b�Geen�done�in��Mathp��}'ert�,�Z�although����?on�?�the�logical�side�a�full-
edged�theorem�pro���v�er�?�has�not�b�Geen�pro���vided;���only����?enough�UUto�supp�Gort�the�correctness�of�sym���b�olic�computation.����?�Ac��9kno�wledgemen�ts����N�The��[nal�form�of�this�pap�Ger�w���as�greatly�in
uenced�b�y�the�careful�reading,����?helpful�qcriticisms,�xand�suggestions�made�b���y�sev�eral�anon�ymous�referees�and�b�y����?Carolyn�UUT��*�alcott,�for�whic���h�I�am�v�ery�grateful.��!č�?�References�������D�[1]���S�<M.���Ballan���t�yne�and�W.�W.�Bledso�Ge,���1977�JA�CM��spap�Ger,���Automatic�Pro�ofs����S�<in�0"Analysis�Using�Non-Standard�T��*�ec���hniques,�7��JA���CM�0�24�(3)�(1977)�353-371.������D[2]���S�<M.�2Ballan���t�yne,�&The�Metatheorist:�T6Automatic�Pro�Gofs�of�Theorems�in�Anal-����S�<ysis�q�Using�Non-Standard�T��*�ec���hniques,�x�P�art�q�I�GI,�in:���R.�S.�Bo���y�er�q�(ed.),��A���uto-����S�<mate��}'d���R�e�asoning�,�UUKlu���w�er�(1991).������D[3]���S�<M.���Beeson,���F��;�oundations��yof�Constructive�Mathematics�,�Springer-V��*�erlag����S�<(1985).������D[4]���S�<M.���Beeson,��Pro���ving�programs�and�programming�pro�Gofs,�in:��Barcan,�Mar-����S�<cus,�'�Dorn,�and�lW��*�eingartner�(eds.),�'��ALogic,�Metho�Gdology�,�and�lPhilosoph���y�of����S�<Science��CVI�GI�,�pro�ceedings�of�the�In���ternational�Congress,��GSalzburg,�1983,�pp.����S�<51-81,�UUNorth-Holland,�Amsterdam�(1986).�������40����)=�l�����'��������D�[5]���S�<M.���Beeson,�j�Logic�and�computation�in��Mathp��}'ert��ز:���an�exp�Gert�system�for����S�<learning�ǁmathematics,��in:�VKaltofen,�E.,�and�W��*�att,�S.�M.,��Computers���and����S�<Mathematics�,�UUpp.�202-214,�Springer-V��*�erlag�(1989).������D[6]���S�<M.�[\Beeson,����Mathp��}'ert��ز:�}�a�computerized�en���vironmen�t�[\for�learning�algebra,����S�<trig,���and�~{calculus,��J.��GA���rticial�Intel���ligenc��}'e�and�Educ�ation�~{�2��(1990),���pp.����S�<1-11.������D[7]���S�<M.��.Beeson,��$�AMathp�Gert�7��:�yyComputer�supp�ort�for�learning�algebra,��$trigonom-����S�<etry��*�,��uand�K�calculus,�in:�^�A.�V��*�oronk���o�v�(ed.),�ALogic�Programming�and�Auto-����S�<mated�~Reasoning�,��)Lecture�Notes�in�Computer�Science��624�,�Springer-V��*�erlag����S�<(1992).������D[8]���S�<M.�-�Beeson,���Some�applications�of�Gen���tzen's�pro�Gof�theory�in�automated����S�<deduction,��fin:�Y�Sc���hro�Geder-Heister,�P��*�.,��Extension�_�of�L��}'o�gic�_�Pr�o�gr�amming�,����S�<Springer��Lecture�Notes�in�Computer�Science��475�,�kpp.�101-156,�Springer-����S�<V��*�erlag�UU(1991)������D[9]���S�<W.���W.�Bledso�Ge,�@Some�automatic�pro�ofs�in�analysis,�@pp.�89-118�in�W.�Bled-����S�<so�Ge,��!and�y�D.�Lo���v�eland�y�(eds):���A���utomate��}'d���The�or�em�Pr�oving:��After�25�ye�ars�,����S�<v���olume�}S29�in�the��Contemp��}'or�ary��.Mathematics�}S�series,���AMS,�Pro�vidence,���R.�I.����S�<(1984).������?[10]���S�<W.��{W.�Bledso�Ge,��Non-resolution�theorem�pro���ving,��A���rticial��)Intel���ligenc��}'e��9�:����S�<1-36�UU(1977).������?[11]���S�<R.�UUS.�Bo���y�er�UU(ed.),��A���utomate��}'d���R�e�asoning�,�UUKlu���w�er�(1991).������?[12]���S�<Clark���e,��E.,�and��VZhao,�X.�[1992],�Analytica{an�exp�Gerimen���t�in�com�bining����S�<theorem��3pro���ving�and�sym�b�Golic�manipulation,��*T��*�ec�hnical�Rep�Gort�CMU-CS-����S�<92-1�]E7,���Sc���ho�Gol�of�Computer�Science,�Carnegie�Mellon�Univ���ersit�y��*�,�Pittsburgh����S�<P��*�A�UU15213.������?[13]���S�<Chou,���Shang-Ching,�Pro���ving���elemen�tary�geometry�theorems�using�W��*�u's�al-����S�<gorithm,���pp.��O243-286�in�Bledso�Ge,�W.,�and�Lo���v�eland,�D.��O(eds):�M��A���utomate��}'d����S�<The��}'or�em��VPr�oving:��uAfter�25�ye�ars�,�k�v���olume�g229�in�the��Contemp��}'or�ary�Mathe-����S�<matics�UU�series,�AMS,�Pro���vidence,�R.�I.�(1984).������?[14]���S�<W.�UUM.�F��*�armer,�Theory�of�T���yp�Ges,��J.���Symb��}'olic�L�o�gic�UU�55��(1990)�1269-91.������?[15]���S�<W.��M.�F��*�armer,��Guttman,�and��Tha���y�er,��IMPS:�an�In�teractiv�e�Mathematical����S�<Pro�Gof�UUSystem,��J.���A���utomate��}'d�R�e�asoning�UU�11��(1993)�213-248.������?[16]���S�<S.��UF��*�eferman,���Logics�for�termination�and�correctness�of�functional�programs,����S�<in:�+~Y.�20Mosc���ho�v��q�akis�(ed.),�ig�L��}'o�gic�_and�Computer�Scienc��}'e:�/�Pr�o�c.�_of�the�'89����S�<MSRI���Confer��}'enc�e�,�UUSpringer-V��*�erlag�(1989).�������41����*I l�����'��������?�[17]���S�<J.�0Harrison�and�L.�Th�����Gery��*�,��Extending�the�HOL�theorem�pro�v�er�with�a�com-����S�<puter���algebra�system�to�reason�ab�Gout�the�reals,�fWin��Higher�AOr��}'der�L�o�gic����S�<The��}'or�em��VPr�oving�and�its�Applic�ations:��t6th�International�Workshop,��rHUG����S�<'93�,��pp.��{174{184,�Lecture�Notes�in�Computer�Science��780�,�Springer-V��*�erlag����S�<(1993).������?[18]���S�<H.��J.�Keisler,��CProbabilit���y�quan�tiers,��Cin:��DJ.Barwise�and�S.F��*�eferman�(eds.)����S�<�Mo��}'del-the�or�etic���L�o�gics�,�B�pp.509-556,�P���ersp�Gectiv�es���in�Mathematical�Logic,����S�<Springer-V��*�erlag�UU(1985).������?[19]���S�<S.��"Kleene,��T�Intr��}'o�duction���to�Metamathematics�,�v��q�an��"Nostrand-Reinhold,����S�<Princeton,�UUNew�Jersey�(1952).������?[20]���S�<G.��Kreisel,�#[Axiomatizations�of�non-standard�analysis�that�are�conserv��q�ativ���e����S�<extensions��of�formal�systems�for�classical�standard�analysis,�+in:��RW.�A.�J.����S�<Luxem���b�Gourg��(ed.),���Applic��}'ations�8�of�Mo�del�The�ory�to�A���lgebr�a,�J�A�nalysis,�and����S�<Pr��}'ob�ability�,�UUpp.�93-106.�Holt,�Rinehart,�and�Winston�(1967).������?[21]���S�<E.�j-Nelson,�ocIn���ternal�Set�Theory:��xA�j(new�approac�h�to�nonstandard�analysis,����S�<�Bul���l.���A.�M.�S.�UU�83��(1977)�1165-1198.������?[22]���S�<E.���Nelson,��The�syn���tax�of�nonstandard�analysis,��A���nnals��\of�Pur��}'e�and�Applie�d����S�<L��}'o�gic�UU�38��(1988)�123-134.������?[23]���S�<A.�UURobinson,��Non-standar��}'d���A���nalysis�,�North-Holland,�Amsterdam�(1974).������?[24]���S�<A.�`Robinson�and�E.�Zak���on,�=�A�0set-theoretical�c�haracterization�of�enlarge-����S�<men���ts,��4in��W.�A.�J.�Luxem�b�Gourg�(ed.),��4�Applic��}'ations���of�Mo�del�The�ory�to����S�<A���lgebr��}'a,���A�nalysis,�and�Pr��}'ob�ability�,�UUHolt,�New�Y��*�ork�(1969).������?[25]���S�<C.�H�L.�Siegel,�~��T��;�opics��in�Complex�F�unction�The��}'ory,��fvol.�I�,�H�Wiley��*�,�~�In���terscience,����S�<New�UUY��*�ork�(1969).������?[26]���S�<R.�'AStoutemey���er,�0yCrimes�and�misdemeanors�in�the�computer�algebra�trade,����S�<�Notic��}'es���of�the�A.M.S.�UU�38�(7)�779-785,�Septem���b�Ger�1991.�������42����U����;�l�*�Ap�0J

cmsl10�;�':

cmti10�8��<x

cmtt10�2�"V

cmbx10�1��N�ffcmbx12�/�j��		cmti9�.t�:		cmbx9�-o���		cmr9�"#�f�cmti8��C�scmtt8�2�@�cmbx8���N�cmbx12�q�%cmsy6��K�cmsy8�;�cmmi6��2cmmi8��Aa�cmr6�|{Ycmr8�X�Qcmr12�D��tG�G�cmr17�
!",�

cmsy10�O!�cmsy7�
�b>

cmmi10�	0e�rcmmi7�K�`y

cmr10�ٓ�Rcmr7���Zcmr5���u

cmex10�^������

Sindbad File Manager Version 1.0, Coded By Sindbad EG ~ The Terrorists