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/turing2.dvi

����;� TeX output 2003.09.18:2217���������>����Le�����
@���!������孄����^�ō����>;�)�ō���>;��?���N�ffcmbx12�The�ffMec���hanization�of�Mathematics��"�%��?�K�`y

cmr10�Mic���hael�UUBeeson���^��ٓ�Rcmr7�1����i��?�o���		cmr9�San�TJos������We�State�Univ��9ersit�y��:�,�TSan�Jose,�CA�95192,�USA��:.2����?�t�:		cmbx9�Summary���.���r�4�The�����j��		cmti9�me��chanization��of�mathematics��refers�to�the�use�of�computers�to����?nd,�z�or�to�help�nd,�mathematical�pro�A�ofs.�T��:�uring�sho��9w�ed�z�that�a�complete�reduction����?of��Smathematics�to�computation�is�not�p�A�ossible,�but�nev��9ertheless�the�art�and�science����?of�1�automated�deduction�has�made�progress.�This�pap�A�er�describ�es�some�of�the�history����?and�Tsurv��9eys�the�state�of�the�art.��!���?���N�cmbx12�1��S@In��tro�`duction���F��?�In�Ⱥthe�nineteen���th�cen�tury��*�,�mac�hines�replaced�h�umans�and�animals�as�ph�ys-����?ical��Glab�Gorers.�While�for�the�most�part�this�w���as�a�w�elcome�relief,�there�w�ere����?o�Gccasional��p�o�c���k�ets�of�resistance.�The�folk�song���':

cmti10�John��Henry��commemorates�an����?o�Gccasion��when�a�man�and�a�mac���hine�comp�eted�at�the�task�of�drilling�railroad����?tunnels��7through�moun���tains.�The�\drilling"�w�as�done�b�y�hammering�a�steel����?spik���e.��The�mac�hine�w�as�steam-p�Go�w�ered.�The�man�w�as�an�ex-sla�v�e,�a�banjo����?pla���y�er���with�a�deep�singing�v���oice�and�a�reputation�for�ph�ysical�strength�and����?endurance.�(He�b�Geat�the�mac���hine,�drilling�fourteen�feet�to�its�nine,�but�it�w�as����?a�UUPyrrhic�victory��*�,�as�he�died�after�the�eort.����NEv���en�c�b�Gefore�the�rst�computers�w�ere�dev�elop�Ged,�p�eople�w���ere�sp�eculating����?ab�Gout��the�p�ossibilit���y�that�mac�hines�migh�t�b�Ge�made�to�p�erform�in���tellectual����?as�rmw���ell�as�ph�ysical�tasks.�Alan�T��*�uring�w�as�the�rst�to�mak�e�a�careful�analy-����?sis�r�of�the�p�Goten���tial�capabilities�of�mac�hines,�in�v�en�ting�his�famous�\T��*�uring����?mac���hines"��for�the�purp�Gose.�He�argued�that�if�an�y�mac�hine�could�p�Gerform�a����?computation,�)qthen�some�T��*�uring�mac���hine�could�p�Gerform�it.�The�argumen�t�fo-����?cuses��@on�the�assertion�that�an���y�mac�hine's�op�Gerations�could�b�e�sim���ulated,�one����?step�D�at�a�time,�b���y�certain�simple�op�Gerations,�and�that�T��*�uring�mac�hines�w�ere����?capable���of�those�simple�op�Gerations.�T��*�uring's�rst�fame�resulted�from�applying����?this��analysis�to�a�problem�p�Gosed�earlier�b���y�Hilb�ert,�whic���h�concerned�the�p�os-����?sibilit���y��of�mec�hanizing�mathematics.�T��*�uring�sho�w�ed�that�in�a�certain�sense,����?it���is�imp�Gossible�to�mec���hanize�mathematics:�W��*�e�shall�nev�er�b�Ge�able�to�build����?an�]s\oracle"�mac���hine�that�can�correctly�answ�er�all�mathematical�questions����?presen���ted�k�to�it�with�a�\y�es"�or�\no"�answ�er.�In�another�famous�pap�Ger�[101],����?T��*�uring�Iqw���en�t�on�to�consider�the�somewhat�dieren�t�question,�\Can�mac�hines����?think?".���It�is�a�dieren���t�question,�b�Gecause�p�erhaps�mac���hines�can�think,�but����?they���migh���t�not�b�Ge�an�y�b�Getter�at�mathematics�than�h�umans�are;�or�p�Gerhaps����?they�0�migh���t�b�Ge�b�etter�at�mathematics�than�h���umans�are,�but�not�b�y�thinking,������*�c�ō���>;����?�2���`-	Mic��9hael�TBeeson����)�ō���>;��?�just�9b���y�brute-force�calculation�p�Go�w�er.�These�t�w�o�pap�Gers�of�T��*�uring�lie�near����?the��Aro�Gots�of�the�sub��8jects�to�da���y�kno�wn�as��automate��}'d��dde�duction��A�and��articial����?intel���ligenc��}'e�.����^��1������N�Although�3@T��*�uring�had�already�pro���v�ed�3@there�w���ere�limits�to�what�one�could����?exp�Gect�-Aof�mac���hines,�nev�ertheless�mac�hines�b�Gegan�to�comp�ete�with�h���umans����?at�Klin���tellectual�tasks.�Arithmetic�came�rst,�but�b�y�the�end�of�the�cen�tury����?computers�^�could�pla���y�excellen�t�c�hess,�and�in�1997�a�computer�program�b�Geat����?w���orld��-c�hampion�Garry�Kasparo�v.�The�New�Y��*�ork�Times�describ�Ged�the�matc�h:����?\In��Xa�dazzling�hourlong�game,�the�Deep�Blue�IBM���computer�demolished����?an��yob���viously�o�v�erwhelmed�Garry�Kasparo�v�and�w�on�the�six-game�man-vs.-����?mac���hine�UUc�hess�matc�h."����^��2������N�In�(1956,�Herb�Simon,�one�of�the�\fathers�of�articial�in���telligence",�pre-����?dicted�a�that�within�ten�y���ears,�computers�w�ould�b�Geat�the�w�orld�c�hess�c�hampion,����?comp�Gose��\aesthetically�satisfying"�original�m���usic,�and�pro�v�e�new�mathemati-����?cal��Ntheorems.����^��3�����It�to�Gok�fort���y�y�ears,�not�ten,�but�all�these�goals�w�ere�ac�hiev�ed|����?and�_�within�a�few�y���ears�of�eac�h�other!�The�m�usic�comp�Gosed�b�y�Da�vid�Cop�Ge's����?programs�Q�[33{35]�cannot�b�Ge�distinguished,�ev���en�b�y�professors�of�m�usic,�from����?that�UUcomp�Gosed�b���y�Mozart,�Beetho�v�en,�and�Bac�h.����^��4������N�In��b1976,�a�computer�w���as�used�in�the�pro�Gof�of�the�long-unsolv�ed�\four����?color��Gproblem'.����^��5���n��This�did�not�fulll�Simon's�prediction,�b�Gecause�the�role�of����?the��Jcomputer�w���as�simply�to�c�hec�k�b�y�calculation�the�1476�dieren�t�sp�Gecic����?cases���to�whic���h�the�mathematicians�had�reduced�the�problem�[2,3].�T��*�o�Gda�y�this����?w���ould���not�cause�a�ripple;�but�in�1976�it�created�quite�a�stir,�and�there�w�as����?serious��discussion�ab�Gout�whether�suc���h�a�\pro�of��"�w���as�acceptable!�The�journal����?editors�<�required�an�indep�Genden���t�computer�program�to�b�e�written�to�c���hec�k����?the�4�result.�The�use�of�computer�calculations�to�pro���vide�\empirical"�evidence��?�h��ff8�ϟ
L͍�������-=��Aa�cmr6�1�����	?��One��con��9tro�v�ersy�concerns�the�question�whether�the�limiting�theorems�ab�A�out�T��:�ur-����	?�ing�r;mac��9hines�also�apply�to�h�uman�in�telligence,�or�whether�h�uman�in�telligence�has����	?�some��uqualit��9y�not�imitable�b�y�a�T��:�uring�mac�hine�(a�vital�force,�free�will,�quan�tum����	?�indeterminacy��in�the�synapses?)�These�questions�w��9ere�already�tak�en�up�b�y�T��:�ur-����	?�ing,��and�w��9ere�still�under�discussion�(without�agreemen�t)�b�y�scien�tic�luminaries����	?�at�Tthe�end�of�the�t��9w�en�tieth�Tcen�tury�[79,80].����������-=�2�����	?��After�	�the�game,�IBM�	@retired�Deep�Blue,�\quitting�while�it�w��9as�ahead."�Some����	?�said��that�Kasparo��9v�lost�only�b�A�ecause�he�got�nerv�ous�and�blundered.�No�rematc�h����	?�w��9as�\%held.�In�Octob�A�er,�2002,�another�c�hampion�pla�y�ed�another�computer�program:����	?�This�Ttime�it�w��9as�a�dra�w.����������-=�3�����	?��This���prediction�is�usually�cited�as�ha��9ving�b�A�een�made�in�1957,�but�I���b�eliev��9e�it����	?�w��9as���actually�rst�made�in�1956�at�Simon's�inaugural�address�as�Presiden�t�of�the����	?�Op�A�erations�TResearc��9h�So�ciet��9y�of�America.����������-=�4�����	?��That�h4lev��9el�of�p�A�erformance�w�as�not�demanded�b�y�Simon's�prediction,�and�his����	?�criterion�lof�\aesthetically�satisfying"�m��9usic�w�as�met�m�uc�h�earlier.�It�is�in�teresting����	?�that��Simon�set�a�lo��9w�er��bar�for�m��9usic�than�for�mathematics�and�c�hess,�but�m�usic����	?�turned�Tout�to�b�A�e�easier�to�computerize�than�mathematics.����������-=�5�����	?��This��problem�asks�whether�it�is�p�A�ossible�to�color�an��9y�map�that�can�b�e�dra��9wn�on����	?�a�e�plane�using�at�most�four�colors,�in�suc��9h�a�w�a�y�that�coun�tries�with�a�common����	?�b�A�order�Treceiv��9e�dieren�t�colors.�����
Ϡc�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics����E�3����)�ō���>;��?�for�(mathematical�claims�has�led�to�\exp�Gerimen���tal�mathematics"�and�ev�en�to����?rep�Gorts�$of�the�\death�of�pro�of��"�[53].�As�Mark�Tw���ain�said,�\the�rep�orts�of�m���y����?death�UUare�greatly�exaggerated".����NOn���Decem���b�Ger�10,�1996,�Simon's�prediction�came�true.�The�fron�t�page�of����?the�"7New�Y��*�ork�Times�carried�the�follo���wing�headline:��Computer�PeMath�Pr��}'o�of����?Shows���R��}'e�asoning�Power�.�UUThe�story�b�Gegan:���#��P�Computers�Q�are�whizzes�when�it�comes�to�the�grun��9t�w�ork�of�mathematics.����PBut��for�creativ��9e�and�elegan�t�solutions�to�hard�mathematical�problems,����Pnothing��has�b�A�een�able�to�b�eat�the�h��9uman�mind.�That�is,�p�erhaps,�un��9til����Pno��9w.��YA���computer�program�written�b�y�researc�hers�at�Argonne�National����PLab�A�oratory�-|in�Illinois�has�come�up�with�a�ma���jor�mathematical�pro�of�that����Pw��9ould��~ha�v�e�b�A�een�called�creativ�e�if�a�h�uman�had�though�t�of�it.�In�doing�so,����Pthe�Vcomputer�has,�for�the�rst�time,�got�a�to�A�ehold�in��9to�pure�mathematics,����Pa��ueld�describ�A�ed�b��9y�its�practitioners�as�more�of�an�art�form�than�a�science.������N�The�V(theorem�w���as�pro�v�ed�b�y�the�computer�program�EQP��*�,�written�b�y�Bill����?McCune.��Before�it�w���as�pro�v�ed,�it�w�as�kno�wn�as�the��R��}'obbins�	�Conje�ctur�e�,����?and�i�p�Geople�seem�reluctan���t�to�c�hange�the�name�to�\EQP's�theorem".�It�is����?ab�Gout�� certain�algebras.�An�algebra�is�a�set�with�t���w�o�� op�erations,�written�as����?w���e�cusually�write�addition�and�m�ultiplication,�and�another�op�Geration�called����?\complemen���t"��and�written��
�b>

cmmi10�n�(�x�).�If�an�algebra�satises�certain�nice�equations����?it��is�called�a�Bo�Golean�algebra.�Robbins�exhibited�three�short�simple�equations����?and�f�conjectured�that�these�three�equations�can�b�Ge�used�to�axiomatize�Bo�olean����?algebras;���that�is,�those�three�equations�imply�the�usual�axioms�for�Bo�Golean����?algebras.��A���complete,�precise�statemen���t�of�the�Robbins�conjecture�is�giv�en�in����?Fig.�UU1.����NEQP��solv���ed��this�problem�in�a�computer�run�lasting�eigh�t�da�ys,�and�using����?30��megab���ytes�of�memory��*�.�The�pro�Gof�it�pro�duced,�ho���w�ev�er,��w�as�only�fteen����?lines���long�and�ts�on���to�a�single�page�or�computer�screen.�Y��*�ou�sometimes����?ha���v�e���to�sho���v�el���a�lot�of�dirt�and�gra���v�el���to�nd�a�diamond.����^��6���{d�Since�the�pro�Gof����?w���as�qYeasily�c�hec�k��q�able�b�y�h�umans,�there�w�as�no�
urry�of�discussion�ab�Gout�the����?acceptabilit���y���of�the�pro�Gof,�as�there�had�b�een�ab�out�the�four-color�problem.����?(There���w���as,�ho�w�ev�er,�a�bit�of�discussion�ab�Gout�whether�h�umans�had�really����?giv���en�Vthis�problem�their�b�Gest�shot|�but�indeed,�T��*�arski�studied�it,�and�none����?of�t�the�h���umans�who�w�ere�tempted�to�b�Ge�critical�w�ere�able�to�nd�a�pro�Gof,�so����?these���discussions�w���ere�generally�short-liv�ed.)�An�am�using�sideligh�t:�The�job����?w���as�Cijust�running�in�the�bac�kground�and�its�successful�completion�w�as�not����?noticed�UUun���til�a�da�y�later!��?�59�ff8�ϟ
L͍�������-=�6�����	?��In��71966�(within�ten�y��9ears�of�Simon's�prediction),�a�computer�program�w�as�in-����	?�v��9olv�ed�i�in�the�solution�of�an�op�A�en�problem.�The�user�w��9as�guiding�an�in�teractiv�e����	?�theorem��!pro��9v�er�kno�wn�as�SAM��to�a�pro�A�of�of�a�kno�wn�theorem,�and�noticed�that����	?�an�r�equation�that�had�b�A�een�deriv��9ed�led�directly�to�the�answ�er�to�a�related�op�A�en����	?�question�Q�[47].�This�ev��9en�t�Q�is�\widely�regarded�as�the�rst�case�of�a�new�result�in����	?�mathematics���b�A�eing�found�with�help�from�an�automated�theorem�pro��9v�er",���accord-����	?�ing�Tto�[72],�p.�6.�������c�ō���>;����?�4���`-	Mic��9hael�TBeeson����)�ō��>;�������?�LqC��������Fig.��h1.��ʻWhat�Texactly�is�the�Robbins�Conjecture?����?A��Bo�A�olean��algebra�is�a�set��5��"		cmmi9�A��together�with�binary�op�erations�+�and������		cmsy9���and�a�unary����?op�A�eration�eF���,�and�elemen��9ts�0,�1�of��A��suc�h�that�the�follo�wing�la�ws�hold:�comm�utativ�e����?and���asso�A�ciativ��9e�la�ws�for�addition�and�m�ultiplication,�distributiv�e�la�ws�b�A�oth�for����?m��9ultiplication�Ȉo�v�er�addition�and�for�addition�o�v�er�m�ultiplication,�and�the�follo�wing����?sp�A�ecial��la��9ws:��x��S�+�(�x����y�R��)�މ=��x�,���x��S���(�x��+��y��)�މ=��x�,���x��S�+�(��x�)�މ=�1,��x��S���(��x�)�މ=�0.�This����?denition,��and�other�basic�information�on�the�sub���ject,�can�b�A�e�found�in�[73].�The����?Robbins���conjecture�sa��9ys�that�an�y�algebra�satisfying�the�follo�wing�three�equations����?is�Ta�Bo�A�olean�algebra.������^�x�8�+��y��p�=��ƽy�`�+��x��������^�(�x�8�+��y�R��)�+��z��S�=��ƽx��+�(�y�`�+��z�c��)��������^�n�(�n�(�x�8�+��y�R��)�+��n�(�x��+��n�(�y��)))���=��x������?�Previous�Tw��9ork�had�sho�wn�that�it�is�enough�to�pro�v�e�the��Huntington�N<e��quation�:������^�n�(�n�(�x�)�8+��y�R��)�+��n�(�n�(�x�)�+��n�(�y��))���=��x:������?�That�@�is,�if�this�equation�is�satised,�then�the�algebra�is�Bo�A�olean.�What�EQP�@�actually����?did,�U�then,�is�come�up�with�a�pro�A�of�that�the�three�Robbins�equations�imply�the�Hun��9t-����?ington�w�equation.�T��:�ak��9e�out�y�our�p�A�encil�and�pap�er�and�giv��9e�it�a�try�b�efore�reading�on.����?Y��:�ou��idon't�need�a�Ph.�D.�in�mathematics�to�understand�the�problem:�Just�see�if�the����?three�
�Robbins�equations�imply�the�Hun��9tington�equation.�It�is�imp�A�ortan�t�to�under-����?stand�+Rthe�nature�of�the�game:�Y��:�ou�do�not�need�to�\understand"�the�equations,�or�the����?\meaning"��eof�the�sym��9b�A�ols��n�,�+�and���.�Y��:�ou�migh�t�b�A�e�happier�if�y�ou�could�think�of�+����?as���\or",����as�\and",�and��n��as�\not",�but�it�is�completely�unnecessary��:�,�as�y��9ou�are�not����?allo��9w�ed���to�use�an��9y�prop�A�erties�of�these�sym�b�A�ols�except�those�giv�en�b�y�the�equations.�����?�LqC���� �.��N�It�)seems,�ho���w�ev�er,�)that�the�in���tellectual�triumph�of�the�computer�is�b�y�no����?means�=as�thorough�as�the�ph���ysical�triumph�of�the�steam�drill.�The�computer����?has�W�y���et�to�b�Geat�a�h�uman�c�hess�c�hampion�reliably�and�rep�Geatedly��*�,�and�the����?n���um�b�Ger���of�mathematical�theorems�whose�rst�pro�of�w���as�found�b�y�a�computer����?is�z�still�less�than�100,�though�there�is�some�fuzziness�ab�Gout�what�coun���ts�as�a����?theorem���and�what�coun���ts�as�a�computer�pro�Gof.�No�graduate�studen�t�to�Gda�y����?c���ho�Goses���not�to�b�ecome�a�mathematician�for�fear�that�the�computer�will�pro���v�e����?to�Go���dicult�a�comp�etitor.�The�da���y�when�a�computer�pro�duces�a�v���e�h�undred����?page�UUpro�Gof�that�answ���ers�a�famous�op�en�question�is�not�imminen���t.��
@���NAnother���analogy��*�,�p�Gerhaps�closer�than�the�steam�drill,�is�to�mec���hanizing����?
igh���t.�klWith�regard�to�mec�hanizing�mathematics,�are�w�e�no�w�at�the�stage�of����?Leonardo�jMda�Vinci's�dra���wings�of�men�with�wings,�or�at�the�stage�of�the�W��*�righ�t����?brothers?�<�Can�w���e�exp�Gect�the�analog�of�jetliners�an�ytime�so�Gon?�Airplanes�
y��*�,����?but��not�quite�lik���e�birds�
y;�and�Dijkstra�famously�remark�ed�that�the�question����?whether�ܤmac���hines�can�think�is�lik�e�the�question,�\Can�submarines�swim?".����?Since��`p�Geople�ha���v�e��`no�wings,�the�prosp�ect�of�mac���hines�
ying�did�not�create�the����?anxieties��and�con���tro�v�ersies��that�surround�the�prosp�Gect�of�mac���hines�thinking.�����.T�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics����E�5����)�ō���>;��?�But���mac���hines�do�mathematics�somewhat�in�the�w�a�y�that�submarines�swim:����?p�Gonderously��*�,��kwith�more�p�o���w�er��kand�duration�than�a�sh,�but�with�less�grace����?and�UUb�Geaut���y��*�.����^��7������N�A��}'cknow���le�dgments�.�aYI�aVam�grateful�to�the�follo���wing�p�Geople,�who�read�drafts����?and��	suggested�c���hanges:�Nadia�Ghamra�wi,�Marvin�Ja�y�Green�b�Gerg,�Mik�e�P�alle-����?sen,�UUand�Larry�W��*�os.���Q���?�2��S@Before��T���uring���Q��?�In��this�section�w���e�review�the�ma��8jor�strands�of�though�t�ab�Gout�the�mec�ha-����?nization�s�of�mathematics�up�to�the�time�of�T��*�uring.�The�ma��8jor�gures�in�this����?history��w���ere�Leibniz,�Bo�Gole,�F��*�rege,�Russell,�and�Hilb�ert.�The�ac���hiev�emen�ts����?of��3these�men�ha���v�e��3b�Geen�discussed�in�man���y�other�places,�most�recen�tly�in����?[39],��Wand�t���w�en�t�y��Wy�ears�ago�in�[38].�Therefore�w�e�will�k�eep�this�section�short;����?nev���ertheless,�UUcertain�minor�c�haracters�deserv�e�more�atten�tion.����NGottfried�nLeibniz�(1646-1716)�is�famous�in�this�connection�for�his�slogan����?�Calculemus�,�o�whic���h�means�\Let�us�calculate."�He�en�visioned�a�formal�language����?to�1�reduce�reasoning�to�calculation,�and�said�that�reasonable�men,�faced�with����?a���dicult�question�of�philosoph���y�or�p�Golicy��*�,�w�ould�express�the�question�in����?a��Pprecise�language�and�use�rules�of�calculation�to�carry�out�precise�reason-����?ing.�K*This�is�the�rst�reduction�of�reasoning�to�calculation�ev���er�en�visioned.����?One��imagines�a�ro�Gomful�of�generals�and�p�olitical�leaders�turning�the�crank�of����?Leibniz's��mac���hine�to�decide�whether�to�launc�h�a�military�attac�k.�It�is�in�ter-����?esting�?�that�Leibniz�did�not�restrict�himself�to�theoretical�sp�Geculation�on�this����?sub��8ject|he��yactually�designed�and�built�a�w���orking�calculating�mac�hine,�the����?�Stepp��}'e�d��R�e�ckoner�.���He�w���as�inspired�b�y�the�somewhat�earlier�w�ork�of�P�ascal,����?who���built�a�mac���hine�that�could�add�and�subtract.�Leibniz's�mac�hine�could����?add,���subtract,�divide,�and�m���ultiply��*�,�and�w�as�apparen�tly�the�rst�mac�hine����?with�]all�four�arithmetic�capabilities.����^��8���ف�Tw���o�of�Leibniz's�Stepp�Ged�Rec�k�oners����?ha���v�e�UUsurviv�ed�and�are�on�displa�y�in�m�useums�in�Munic�h�and�Hano�v�er.����NGeorge���Bo�Gole�(1815-1864)�to�ok�up�Leibniz's�idea,�and�wrote�a�b�o�ok�[26]����?called�%��The�h=L��}'aws�of�Thought�.�The�la���ws�he�form�ulated�are�no�w�called�Bo�Golean��?��V�ff8�ϟ
L͍�������-=�7�����	?��This���is�the�ne�prin��9t�con�taining�the�disclaimers.�In�this�pap�A�er,�\mec�hanization����	?�of��xmathematics"�refers�to�getting�computers�to��nd��pro�A�ofs,�rather�than�ha��9ving����	?�them��*�che��ck��pro�A�ofs�that�w��9e�already�knew,�or��stor�e��pro�A�ofs�or�pap�ers�in�a�database����	?�for��sreference,�or��typ��eset��our�pap�A�ers,�or��send��them�con��9v�enien�tly��sto�one�another,�or����	?��display�B��them�on�the�W��:�eb.�All�these�things�are�indeed�mec��9hanizations�of�mathe-����	?�matics,�[in�a�broader�sense,�and�there�are�man��9y�in�teresting�pro���jects�on�all�these����	?�fron��9ts,��but�w�e�shall�limit�the�scop�A�e�of�our�discussions�to�ev�en�ts�in�the�spirit�of����	?�John�ipHenry�and�Big�Blue.�Moreo��9v�er,�ipw�e�do�not�discuss�past�and�presen�t�eorts�to����	?�enable��computer�programs�to�mak��9e�conjectures,�or�to�apply�mec�hanized�reason-����	?�ing�-�to�other�areas�than�mathematics,�suc��9h�as�v�erication�of�computer�programs����	?�or�Tsecurit��9y�proto�A�cols,�etc.����������-=�8�����	?��The��1abacus�do�A�es�not�coun��9t�b�ecause�it�is�not�automatic.�With�Leibniz's�mac��9hine,����	?�the�Th��9uman�only�turned�the�crank.�����=��c�ō���>;����?�6���`-	Mic��9hael�TBeeson����)�ō���>;��?�Algebra{y���es,�:]the�same�la�ws�of�concern�in�the�Robbins�conjecture.�Lik�e�Leib-����?niz,�0�Bo�Gole�seems�to�ha���v�e�0�had�a�grandiose�vision�ab�out�the�applicabilit���y�of����?his� *algebraic�metho�Gds�to�practical�problems{�his�b�o�ok�mak���es�it�clear�that�he����?hop�Ged��these�la���ws�w�ould�b�Ge�used�to�settle�practical�questions.�William�Stanley����?Jev���ons��(1835-1882)�heard�of�Bo�Gole's�w�ork,�and�underto�Gok�to�build�a�mac�hine����?to��7mak���e�calculations�in�Bo�Golean�algebra.�He�successfully�designed�and�built����?suc���h���a�mac�hine,�whic�h�he�called�the��L��}'o�gic�al��pPiano�,���apparen�tly�b�Gecause�it����?w���as�KQab�Gout�the�size�and�shap�e�of�a�small�piano.�This�mac���hine�and�its�creator����?deserv���e�9m�uc�h�more�fanfare�than�they�ha�v�e�so�far�receiv�ed:�This�w�as�the�rst����?mac���hine�w�to�do�mec�hanical�inference.�Its�predecessors,�including�the�Stepp�Ged����?Rec���k�oner,�yonly�did�arithmetic.�The�mac���hine�is�on�displa�y�at�the�Museum����?of���Science�at�Oxford.�The�design�of�the�mac���hine�w�as�describ�Ged�in�a�pap�er,����?�On�`Uthe�Me��}'chanic�al�`UPerformanc�e�of�L�o�gic�al�Infer�enc�e�,�Gread�b�Gefore�the�British����?Ro���y�al�UUSo�Gciet�y�in�1870.����^��9����-��N�Gottlob�q�F��*�rege�(1848-1925)�created�mo�Gdern�logic�including�\for�all",�\there����?exists",��*and�rules�of�pro�Gof.�Leibniz�and�Bo�ole�had�dealt�only�with�what�w���e����?no���w���call�\prop�Gositional�logic"�(that�is,�no�\for�all"�or�\there�exists").�They����?also��did�not�concern�themselv���es�with�rules�of�pro�Gof,�since�their�aim�w�as�to����?reac���h�ʅtruth�b�y�pure�calculation�with�sym�b�Gols�for�the�prop�ositions.�F��*�rege�to�ok����?the��opp�Gosite�tac���k:�instead�of�trying�to�reduce�logic�to�calculation,�he�tried�to����?reduce�J'mathematics�to�logic,�including�the�concept�of�n���um�b�Ger.�J'F��*�or�example,����?he�F�dened�the�n���um�b�Ger�F�2�to�b�e�the�class�of�all�classes�of�the�form��
!",�

cmsy10�f�x;���y�[ٸg��with����?�x�7��6�=��y�[ٲ.���Lo�Gosely�sp�eaking,�2�is�the�class�of�all�classes�with�t���w�o���mem�b�ers;���but����?put��that�w���a�y��*�,��the�denition�sounds�circular,�whic���h�it�is�not.�His�ma��8jor�w�ork,����?the����Be��}'grischrift��[43],�w���as�published�in�1879,�when�F��*�rege�w�as�31�y�ears�old.����?He�|�describ�Ged�it�as�a�sym���b�olic�language�of�pure�though���t,�mo�deled�up�on�that����?of�UUarithmetic.��-��NBertrand��tRussell�(1872-1970)�found�F��*�rege's�famous�error:�F�rege�had�o���v�er-����?lo�Gok���ed�
what�is�no�w�kno�wn�as�the�Russell�parado�x.����^��10�����Namely��*�,�F�rege's�rules����?allo���w�ed�G�one�to�dene�the�class�of��x��suc���h�that��P�c��(�x�)�is�true�for�an�y�\concept"����?�P�c��.�/F��*�rege's�idea�w���as�that�suc�h�a�class�w�as�an�ob��8ject�itself,�the�class�of�ob-����?jects�d�\falling�under�the�concept��P�c��".�Russell�used�this�principle�to�dene�the����?class�rh�R��/�of�concepts�that�do�not�fall�under�themselv���es.�This�concept�leads�to����?a��bcon���tradiction�kno�wn�as�Russell's�P�arado�x.�Here�is�the�argumen�t:�(1)�if��R����?�falls���under�itself�then�it�do�Ges�not�fall�under�itself;�(2)�this�con���tradiction�sho�ws����?that���it��do��}'es���not��fall�under�itself;�(3)therefore�b���y�denition�it��do�es��fall�under����?itself�UUafter�all.��?�Ɖff8�ϟ
L͍�������-=�9�����	?��In���Decem��9b�A�er�2002,�an�original�cop�y�of�this�pap�A�er�w�as�a�v��|railable�for�purc�hase�from����	?�a�Trare�b�A�o�ok�Tdealer�in�New�Y��:�ork�for�a�price�exceeding�$2000.���������UZ��-=�10�����	?��Russell���w��9as�thirt�y�y�ears�old�at�the�time{ab�A�out�the�same�age�that�F��:�rege�had�b�een����	?�when��he�made�the�error.�Russell's�resp�A�ectful�letter�to�F��:�rege�with�the�bad�news�is����	?�reprin��9ted���in�[102],�p.�124,�along�with�F��:�rege's�reply:�\Y�our�disco��9v�ery���of�the�con��9tra-����	?�diction�m�caused�me�the�greatest�surprise�and,�I�m�w��9ould�almost�sa�y��:�,�consternation,����	?�since�Tit�has�shak��9en�the�basis�on�whic�h�I�in�tended�to�build�arithmetic."�����M:�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics����E�7����)�ō���>;��N�Russell��(with�co-author�Whitehead)�wrote��Principia�^�Mathematic��}'a��[91]�to����?sa���v�e���mathematics�from�this�con���tradiction.�They�restricted�the�applicabilit�y�of����?F��*�rege's�{�class-denition�principle,�th���us�blo�Gc�king�Russell's�parado�x,�and�sho�w�ed����?(b���y� )actually�carrying�out�h�undreds�of�pages�of�pro�Gofs)�that�the�main�lines����?of��mathematics�could�still�b�Ge�dev���elop�ed�from�the�restricted�principle.�This����?w���ork���w�as�v�ery�in
uen�tial�and�b�Gecame�the�starting�p�oin���t�for�t�w�en�tieth-cen�tury����?logic;��3thirt���y�y�ears�later,�when�G����odel�needed�a�sp�Gecic�axiom�system�for�use����?in�s�stating�his�incompleteness�theorem,�the�ob���vious�c�hoice�w�as�the�system�of����?�Principia�.����NDa���vid��=Hilb�Gert�(1862-1943)�w�as�one�of�the�foremost�mathematicians�of�the����?early�2t���w�en�tieth�cen�tury��*�.�He�con�tributed�to�the�dev�elopmen�t�of�formal�logic����?(rules���for�reasoning),�and�then�b�Gecame�in���terested�in�a�t�w�o-step�reductionist����?program���that�com���bined�those�of�Leibniz�and�F��*�rege:�he�w�ould�rst�reduce����?mathematics�Q�to�logic,�using�formal�languages,�and��then��reduce�logic�to�com-����?putation.���His�plan�w���as�to�consider�the�pro�Gofs�in�logic�as�ob��8jects�in�their����?o���wn�ҿrigh�t,�and�study�them�as�one�w�ould�study�an�y�nite�structure,�just�as����?mathematicians�9�study�groups�or�graphs.�He�hop�Ged�that�w���e�w�ould�then�b�Ge����?able�AYto�giv���e�algorithms�for�determining�if�a�giv�en�statemen�t�could�b�Ge�pro�v�ed����?from��Agiv���en�axioms,�or�not.�By�consideration�of�this�researc�h�program,�he����?w���as�led�to�form�ulate�the�\decision�problem"�for�logic,�b�Getter�kno�wn�b�y�its����?German�>name,�the�\En���tsc�heidungsproblem".�>This�problem�w���as�published�in����?1928��in�the�in
uen���tial�logic�b�Go�ok��b�y�Hilb�Gert�and�Ac�k�ermann�[51].�This�w�as����?the���problem�whose�negativ���e�solution�made�T��*�uring�famous;�the�next�section����?will�UUexplain�the�problem�and�its�solution.��1���?�3��S@Hilb�`ert��and�the�En��tsc�heidungsproblem��1��?�The���En���tsc�heidungsproblem�asks�whether�there�exists�a�decision�algorithm����?suc���h�UUthat:��ɧ��P��UU�It�tak���es�t�w�o�inputs:�a�nite�set�of�axioms,�and�a�conjecture.���G��P����It�computes�for�a�nite�time�and�outputs�either�a�pro�Gof�of�the����Pconjecture�UUfrom�the�axioms,�or�\no�pro�Gof�exists".����P��UU�The�result�is�alw���a�ys�UUcorrect.���`��NP���art�<�of�the�reason�for�the�historical�imp�Gortance�of�this�problem�is�that�it����?w���as�� a�signican�t�ac�hiev�emen�t�just�to�state�the�problem�precisely��*�.�What�are����?�axioms��(�?�What�is�a��pr��}'o�of�?��(What�is�an��algorithm�?�Progress�on�the�rst�t���w�o��(of����?those��vquestions�had�b�Geen�made�b���y�Russell�and�b�y�Hilb�Gert�himself.�There�w�as����?an�@8imp�Gortan���t�dierence�in�their�approac�hes,�ho�w�ev�er.�Russell�w�ork�ed�with����?pro�Gofs�|Rand�axioms�in�order�to�nd�axioms�that�w���ere�eviden�tly�true,�and�w�ould����?therefore��enable�one�to�deriv���e�true�(and�only�true)�mathematical�theorems.����?He�W had�in�mind�one�xed�in���terpretation�of�his�axioms{that�is,�they�w�ere�ab�Gout����?the��one�true�mathematical�univ���erse�of�classes,�if�they�w�ere�ab�Gout�an�ything����?at�l�all.�In�the�man���y�pages�of��Principia��KMathematic��}'a�,�Russell�and�Whitehead�����^�c�ō���>;����?�8���`-	Mic��9hael�TBeeson����)�ō���>;��?�nev���er��discussed�the�question�of�what�w�e�w�ould�to�Gda�y�call�the�in�terpretations����?of�# their�formal�theory��*�.�Hilb�Gert,�on�the�other�hand,�understo�o�d�v���ery�w�ell�that����?the���same�axioms�could�ha���v�e���more�than�one�in���terpretation.�Hilb�Gert's�most����?w���ell-kno�wn�~�w�ork�on�axiomatization�is�his�b�Go�ok��F��;�oundations��Qof�Ge��}'ometry��[50].����?This�50b�Go�ok�pro���vided�a�careful�axiomatic�rew�orking�of�Euclid�from�21�axioms.����?Hilb�Gert�Nxemphasized�the�distinction�b�et���w�een�Nxcorrect�reasoning�(ab�out�p�oin���ts,����?lines,�TOand�planes)�and�the�facts�ab�Gout�p�oin���ts,�lines,�and�planes,�b�y�sa�ying����?that�7Fif�y���ou�replace�\p�Goin�ts,�lines,�and�planes"�b�y�\tables,�c�hairs,�and�b�Geer����?m���ugs",��Gthe�reasoning�should�still�b�Ge�correct.�This�seems�ob�vious�to�to�Gda�y's����?mathematicians,��Zb�Gecause�the�axiomatic�approac���h�to�mathematics�pro�v�ed�so����?fruitful��cin�the�rest�of�the�t���w�en�tieth��ccen�tury�that�ev�ery�studen�t�of�mathematics����?is��ito�Gda���y�steep�ed�in�this�basic�idea.�But,�at�the�da���wn�of�the�t�w�en�tieth�cen�tury��*�,����?this��idea�seemed�radical.�The�mathematician�P���oincar���Ge��understo�Go�d�Hilb�ert's����?p�Goin���t��Kv�ery�clearly��*�,�as�one�can�see�in�the�follo�wing�quotation�[78],�but�he����?though���t�UUit�an�tithetical�to�the�spirit�of�mathematics:��PI��P�Th��9us���it�will�b�A�e�readily�understo�o�d�that�in�order�to�demonstrate�a�theorem,����Pit��Ois�not�necessary�or�ev��9en�useful�to�kno�w�what�it�means.�W��:�e�migh�t�replace����Pgeometry�|gb��9y�the�reasoning�piano�imagined�b�y�Stanley�Jev�ons,�or���:���:�:���t�a����Pmac��9hine�߲where�w�e�should�put�in�axioms�at�one�end�and�tak�e�out�theorems����Pat��the�other,�lik��9e�that�legendary�mac�hine�in�Chicago�where�pigs�go�in�aliv�e����Pand�Tcome�out�transformed�in��9to�hams�and�sausages.��|6��?�The�x(date�of�that�quotation�is�1908,�almost�a�decade�after��F��;�oundations��vof����?Ge��}'ometry�.��But�the�concept�of�\pro�Gof��"�w���as�still�a�bit�unclear.�The�distinction����?that�=�w���as�still�lac�king�w�as�what�w�e�call�to�Gda�y�the�distinction�b�Get�w�een�a��rst-����?or��}'der���pro�Gof�and�a��se�c�ond-or�der��pro�Gof.�The�axioms�of�geometry�in�Hilb�ert's����?b�Go�ok�^�included�the�\con���tin�uit�y�^�axiom",�whic���h�sa�ys�that�if�y�ou�ha�v�e�t�w�o�subsets����?�A���and��B�[q�of�a�line��L�,�and�all�the�p�Goin���ts�of��A��lie�to�the�left����^��11���
S�of�all�the����?p�Goin���ts�1�of��B��q�,�then�there�exists�a�p�oin���t��P��X�on��L��to�the�righ�t�of�all�p�Goin�ts�of����?�A�L��not�equal�to��P�c��,�and�to�the�left�of�all�p�Goin���ts�of��B��p�not�equal�to��P��.�This����?axiom�*�is�in���tended�to�sa�y�that�there�are�no�\holes"�in�a�line.�F��*�or�example,����?if����L��is�the��x�-axis,�and�if��A��is�the�set�of�p�Goin���ts�with��x���^��2��	O��<���2,�and�if��B�)�is����?the�t�set�of�p�Goin���ts�with��x��>��0�t�and��x���^��2��	"s�>���2,�then�the�axiom�guaran�tees�the����?existence�<*of��x���=������P�p���o���P�fe�E���2����p.�<*But�the�statemen���t�of�the�axiom�men�tions�not�only�p�Goin�ts,����?lines,�ׄand�planes�(the�ob��8jects�of�geometry)�but�also��sets��of�p�Goin���ts.�Remem�b�Ger����?that����F��;�oundations��of�Ge��}'ometry��w���as�written�b�Gefore�the�disco�v�ery�of�Russell's����?parado���x���and��Principia�,�and�apparen�tly�Hilb�Gert�did�not�see�the�necessit�y�of����?careful���atten���tion�to�the�axioms�for�sets�as�w�ell�as�to�the�axioms�for�p�Goin�ts,����?lines,���and�planes.�A���se��}'c�ond-or�der���theory�or�axiomatization�is�one�that,�lik���e����?Hilb�Gert's��Waxiomatization�of�geometry��*�,�uses�v��q�ariables�for�sets�of�ob��8jects�as�w���ell����?as�Ydv��q�ariables�for�ob��8jects.�P���eano's�axioms�for�n�um�b�Ger�theory�are�another�famous��?��v�ff8�ϟ
L͍������UZ��-=�11�����	?��Hilb�A�ert's�_�axioms�use�a�primitiv��9e�relation�\�x��is�b�et��9w�een�_��y��Q�and��z�c��".�W��:�e�can�a��9v�oid����	?�the�Tinformal�term�\lie�to�the�left"�using�this�relation.�����	lG�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics����E�9����)�ō���>;��?�example�~?of�a�second-order�axiomatization.����^��12���
�%�Inciden���tally��*�,�P�eano's�publication����?[75]�2zw���as�a�pamphlet�written�in�Latin,�long�after�Latin�had�b�Geen�displaced�as����?the��3language�of�sc���holarship,�so�that�the�publication�has�b�Geen�view�ed�as�an����?\act���of�roman���ticism".�P�eano,�originally�a�go�Go�d���teac�her,�b�Gecame�an�unp�opular����?teac���her���b�Gecause�he�insisted�on�using�formal�notation�in�elemen�tary�classes;����?nev���ertheless,�A�his�w�ork�ev�en�tually�b�Gecame�in
uen�tial,�and�it�is�his�notation����?that�UUis�used�to�Gda���y�in�logic,�not�F��*�rege's.��(���NIn�D�b�Goth�these�t���w�o�D�famous�examples,�the�theories�ac���hiev�e�D�their�aim:�They����?uniquely��Ydene�the�structures�they�are�trying�to�axiomatize.�Ev���ery�system����?of��~ob��8jects�satisfying�Hilb�Gert's�axioms�for�plane�geometry�is�isomorphic�to�the����?Euclidean��plane.�Ev���en�if�w�e�b�Gegin�b�y�assuming�that�the�system�consists�of����?tables,��c���hairs,�and�b�Geer�m�ugs,�it�turns�out�to�b�Ge�isomorphic�to�the�Euclidean����?plane.�{�Ev���ery�system�of�ob��8jects�satisfying�P�eano's�axioms�is�isomorphic�to����?the�O�natural�n���um�b�Gers.�O�But�the�second-order�nature�of�these�axiom�systems����?is�c�essen���tial�to�this�prop�Gert�y��*�.�The�tec�hnical�term�for�this�prop�Gert�y�is�that�the����?theory���is��c��}'ate�goric�al�.���These�are��se��}'c�ond-or�der���c�ate�goric�al����theories.�The�concept����?of�� second-order�theory�v���ersus�rst-order�theory�is�not�easy�to�grasp,�but�is����?v���ery��Simp�Gortan�t�in�understanding�the�theoretical�basis�of�the�mec�hanization����?of�UUmathematics,�so�here�go�Ges:��(���NIf�Gw���e�require�a�rst-order�v�ersion�of�the�con�tin�uit�y�axiom,�then�instead����?of�HIsa���ying�\for�all�sets��A��and��B��+:���:�:��
��",�the�axiom�will�b�Gecome�man�y�axioms,����?where�Ul�A��and��B��ݲare�replaced�b���y�man�y�dieren�t�rst-order�form�ulas.�In�other����?w���ords,���instead�of�b�Geing�able�to�state�the�axiom�for��al���l��sets�of�p�oin���ts,�w�e����?will�
ha���v�e�to�settle�for��algebr��}'aic�al���ly�<�denable��sets�of�p�Goin���ts.�W��*�e�will�still�b�e����?able��;to�dene������P�p���뒟��P�fe�E���2�����,�but�w���e�will�not�b�Ge�able�to�dene���[ٲ,�b�ecause�����cannot�b�e����?dened�ɥb���y�algebraic�conditions.�Another�w�a�y�of�lo�Goking�at�this�situation�is�to����?consider��systems�of�\p�Goin���ts"�that�satisfy�the�axioms.�Suc�h�systems�are�called����?\mo�Gdels".��|In�the�case�at�hand,�w���e�ha�v�e�the�\real�plane"�consisting�of�all����?p�Goin���ts�T�(�x;���y�[ٲ),�and�on�the�other�hand,�w�e�ha�v�e�the�smaller�\plane"�consisting����?only��Hof�the�n���um�b�Gers��H(�x;���y�[ٲ)�where��x��and��y�+!�are�solutions�of�some�p�olynomial����?equation��with�in���teger�co�Gecien�ts.�Both�these�satisfy�the�rst-order�axioms����?of�_Bgeometry��*�,�but�the�smaller�plane�lac���ks�the�p�Goin�t�(��[�;����0)�and�hence�do�Ges�not����?satisfy�UUthe�second-order�con���tin�uit�y�UUaxiom.��(���NSimilarly��*�,���in�arithmetic,�if�w���e�do�not�use�v��q�ariables�for�sets�in�stating�the����?induction�m�axiom,�w���e�will�b�Ge�able�only�to�\appro�ximate"�the�axiom�b�y�in-����?cluding���its�sp�Gecic�instances,�where�the�inductiv���e�set�is�dened�in�the�xed��?���ff8�ϟ
L͍������UZ��-=�12�����	?��These�K\famous�axioms�c��9haracterize�the�natural�n�um�b�A�ers��N�D@�as�follo�ws:�0�is�in��N���,����	?�and��if��x��is�in��N��ٻthen�the�successor��x���-=�+��	 c�of��x��is�in��N���,�and�0�is�not�the�successor�of����	?�an��9y�C�n�um�b�A�er,�and�if��x���-=�+���K�=��ݽy��R���-=�+��	���then��x��=��y�R��.�(The�successor�of�0�is�1,�the�successor����	?�of��1�is�2,�etc.)�T��:�o�these�axioms�P��9eano�added�the�axiom�of��induction�:�if��X��X�is�an�y����	?�set���satisfying�these�prop�A�erties�with��X��,�instead�of��N���,�then��N�⵻is�a�subset�of��X��[�.����	?�The�	Ainduction�axiom�is�equiv��|ralen��9t�to�the�statemen�t�that�ev�ery�non-empt�y�set����	?�of�M�natural�n��9um�b�A�ers�M�con�tains�a�least�elemen�t,�and�is�also�equiv��|ralen�t�to�the�usual����	?�form��9ulation�8�of�mathematical�induction:�for�sets��X���of�natural�n�um�b�A�ers,�if�0�is�in����	?��X��[�,�Tand�if�whenev��9er��n��is�in��X�ɯ�so�is��n���-=�+��n�,�then��X��con��9tains�all�natural�n�um�b�A�ers.�����
}#�c��>;����?�10���`-	Mic��9hael�TBeeson����)��>;��?�language��kof�arithmetic.�There�are�theorems�that�sa���y�a�certain�equation�has����?no�Aysolution�in�in���tegers,�whose�pro�Gofs�require�pro�ving�a�v�ery�complicated�for-����?m���ula��p�P�9��b�y�induction,�as�a�lemma,�where�the�form�ula��P�9��is�to�Go�complicated����?to��ev���en�b�Ge�stated�in�the�language�of�arithmetic{p�erhaps�it�requires�more����?adv��q�anced��'mathematical�concepts.�Just�as�there�exist�dieren���t�mo�Gdels�of�rst-����?order�:geometry�(in�whic���h�����do�Ges�or�do�es�not�exist),�there�also�exist�dieren���t����?mo�Gdels�
�of�rst-order�n���um�b�er�
�theory��*�,�some�of�whic���h�are�\non-standard",�in����?that���the�\n���um�b�Gers"���of�the�mo�del�are�not�isomorphic�to�the�actual�in���tegers.����?These�jnon-standard�mo�Gdels�are�more�dicult�to�visualize�and�understand����?than�Ya�plane�that�\simply"�omits�n���um�b�Gers�Ywith�complicated�denitions,�b�e-����?cause���these�mo�Gdels�con���tain�\n�um�b�Gers"�that�are�not�really�n�um�b�Gers,�but�are����?\extra".��̍�NUsing���mo�Gdern�language,�w���e�sa�y�that�a�rst-order�theory��*�,�ev�en�one�formed����?b���y�-Arestricting�a�second-order�categorical�theory�to�its�rst-order�instances,����?generally���has�man���y�mo�Gdels,�not�just�one.�This�situation�w�as�not�clearly�un-����?dersto�Go�d�Z<in�the�rst�t���w�o�Z<decades�of�the�t���w�en�tieth�Z<cen�tury��*�,����^��13����"�but�b�y�1928,����?when��Hilb�Gert�and�Ac���k�ermann��published�their�monograph�on�mathematical����?logic��[51],�it�had�b�Gecome�clear�at�least�to�those�authors.�Clarit���y�on�this�p�oin���t����?led�?{directly�to�the�form���ulation�of�the�En�tsc�heidungsproblem:�Since�a�rst-����?order��theory�generally�has�man���y�mo�Gdels,�can�w�e�decide�(giv�en�a�theory)����?whic���h���form�ulas�are�true�in�all�the�mo�Gdels?�It�also�led�directly�to�the�form�u-����?lation���of�the�completeness�problem:�Are�the�form���ulas�true�in�all�the�mo�Gdels����?exactly�RGthose�that�ha���v�e�RGpro�Gofs�from�the�axioms?�The�former�problem�w���as����?solv���ed���b�y�T��*�uring�and�Ch�urc�h,�the�latter�b�y�G����odel,�b�Goth�within�a�few�y�ears����?of�c�the�publication�of�Hilb�Gert-Ac���k�ermann.�c�These�dev���elopmen�ts�c�laid�the�foun-����?dations�xjof�mo�Gdern�mathematical�logic,�whic���h�in�turn�furnished�the�to�ols�for����?the�UUmec���hanization�of�mathematics.��̍�NThe�]�distinction�b�Get���w�een�]�second-order�and�rst-order�confuses�p�eople�b�e-����?cause��#it�has�t���w�o��#asp�Gects:�syn���tax�and�seman�tics.�A��
theory�whic�h�has�v��q�ariables����?for�y�ob��8jects�and�for�sets�of�those�ob�jects�(for�example�in���tegers�and�sets�of����?in���tegers)�f�is�syn�tactically�second-order.�W��*�e�can�write�do�wn�mathematical�in-����?duction�ۀusing�the�set�v��q�ariables.�But�then,�w���e�can�still�consider�this�as�a����?rst-order��theory��*�,�in�whic���h�case�w�e�w�ould�allo�w�mo�Gdels�in�whic�h�the�set�v��q�ari-����?ables�Urange�o���v�er�Ua�suitable�coun���table�collection�of�sets�of�in�tegers,�and�there����?w���ould���also�b�Ge�mo�dels�with�non-standard�in���tegers�in�whic�h�the�set�v��q�ariables����?range��o���v�er�a�collection�of�\subsets�of�in�tegers"�of�the�mo�Gdel.�Or,�w�e�can�con-����?sider�!�it�as�a�second-order�theory��*�,�in�whic���h�case�w�e�do�not�allo�w�suc�h�mo�Gdels,����?but�
�only�allo���w�mo�Gdels�in�whic�h�the�set�v��q�ariables�range�o�v�er��al���l��subsets�of�the����?in���tegers�,of�the�mo�Gdel.�Whether�it�is�second-order�or�rst-order�is�determined����?b���y���what�w�e�allo�w�as�a�\mo�Gdel"�of�the�theory��*�,�not�b�y�the�language�in�whic�h����?w���e�UUexpress�the�theory��*�.��?�k_�ff8�ϟ
L͍������UZ��-=�13�����	?��See��cfor�example�[67],�P��9art�I�A�I�I��7for��cmore�details�on�the�views�of�Hilb�A�ert�and�his����	?�con��9temp�A�oraries.�������c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������11����)�ō���>;���?�4��S@T���uring's���negativ��e�solution�of�the����S@En��tsc�heidungsproblem���T��?�The��?dev���elopmen�ts�describ�Ged�ab�o���v�e�still�left�the�En�tsc�heidungsproblem�some-����?what�
�imprecise,�in�that�the�concept��algorithm��men���tioned�in�the�problem�had����?not���b�Geen�dened.�Apparen���tly�Hilb�ert�hop�ed�for�a�p�ositiv���e�solution�of�the����?problem,��3in�whic���h�case�it�w�ould�not�ha�v�e�b�Geen�necessary�to�dene�\algo-����?rithm",�_4as�the�solution�w���ould�exhibit�a�sp�Gecic�algorithm.�But�a�negativ�e����?solution�k\w���ould�ha�v�e�to�pro�v�e�that�no�algorithm�could�do�the�job,�and�hence����?it�UUw���ould�b�Ge�necessary�to�ha�v�e�a�denition�of�\algorithm".��,D��NAlan��aT��*�uring�(1912-1954),�answ���ered�the�question�\What�is�an�algorithm?"����?in���1936�[100]�b���y�dening�T��*�uring�mac�hines.����^��14���n�He�used�his�denition�to�sho�w����?that�-�there�exist�problems�that�cannot�b�Ge�solv���ed�b�y�an�y�algorithm.�The�most����?w���ell-kno�wn�Tof�these�is�the�halting�problem{there�exists�no�T��*�uring�mac���hine����?that�iAtak���es�as�inputs�a�T��*�uring�mac�hine��M��\�and�an�input��x��for�M,�and�de-����?termines��correctly�whether�M�Vhalts�on�input��x�.�Indeed,�w���e�don't�need�t�w�o����?v��q�ariables�u,here:�no�T��*�uring�mac���hine�can�determine�correctly�whether��M��G�halts����?at�UUinput��M��.��,D��NIn��=that�same�remark��q�able�1936�pap�Ger�[100],�T��*�uring�applied�his�new�T�uring����?mac���hines�K�to�giv�e�a�negativ�e�solution�to�the�En�tsc�heidungsproblem.�His�solu-����?tion�lumak���es�use�of�the�result�just�men�tioned,�that�the�halting�problem�is�not����?solv��q�able��b���y�a�T��*�uring�mac�hine.�W��*�e�shall�describ�Ge�his�solution�to�the�En�tsc�hei-����?dungsproblem��:no���w,�but�not�the�solution�to�the�halting�problem,�whic�h�is����?co���v�ered�n�in�an���y�mo�Gdern�textb�o�ok�on�the�theory�of�computation.�(The�reader����?who��do�Ges�not�already�kno���w�what�a�T��*�uring�mac�hine�is�should�skip�to�the�next����?section.)�UUThe�solution�has�three�steps:���T��P��i��W��*�rite�do���wn�axioms�A�i�to�describ�Ge�the�computations�of�T�uring�ma-����Pc���hines.��,D��P���0�T��*�uring�mac���hine��M��K�halts�at�input��x��if�and�only�if��A��pro�v�es�the����Ptheorem�UU\�M�lp�halts�at�input��x�".����P��5�If�w���e�had�an�algorithm�to�determine�the�consequences�of�axioms����P�A�,��Cit�w���ould�solv�e�the�halting�problem,�con�tradiction.�Hence�no�suc�h����Palgorithm�UUexists.����^��15����?�		=�ff8�ϟ
L͍������UZ��-=�14�����	?��T��:�uring�h�\mac��9hines"�are�conceptual�ob���jects�rather�than�ph�ysical�mac�hines.�They����	?��c��ould��ۻb�A�e�built,�but�in�practice�the��ide�a��of�these�mac��9hines�is�used,�rather�than����	?�ph��9ysical��*examples.�Suc�h�a�mac�hine�can�b�A�e�sp�ecied�b��9y�a�nite�list�of�its�parts����	?�(\states")�I�and�their�connections�(\instructions").�They�w��9ork�on�\inputs"�that�are����	?�represen��9ted��Sb�y�sym�b�A�ols�on�an�input�device,�usually�called�a�\tap�e".�Whenev��9er����	?�the�&tap�A�e�is�ab�out�to�b�e�used�up,�an�attendan��9t�will�attac�h�more,�so�conceptually��:�,����	?�the���tap�A�e�is�innite,�y��9et�the�mac�hine�could�still�b�A�e�built.�T��:�uring's�k�ey�idea�w�as����	?�that���the�descriptions�of�the�mac��9hines�can�b�A�e�giv�en�b�y�sym�b�A�ols,�and�hence�T��:�uring����	?�mac��9hines�Tcan�accept�(descriptions�of��q)�T��:�uring�mac�hines�as�inputs.���������UZ��-=�15�����	?��In��8more�detail�the�argumen��9t�is�this:�Supp�A�ose�some�T��:�uring�mac�hine��K�� �accepts����	?�inputs��describing�axiom�sets��S�w��and�p�A�oten��9tial�theorems��B�r��,�and�outputs�1�or�0�������c�ō���>;����?�12���`-	Mic��9hael�TBeeson����)�ō���>;��N�The���\computations"�referred�to�in�the�rst�step�can�b�Ge�though���t�of�as����?t���w�o-dimensional��tables.�Eac���h�ro�w�of�the�table�corresp�Gonds�to�the�tap�e�of�the����?T��*�uring�ޞmac���hine�at�a�giv�en�stage�in�its�computation.�The�next�ro�w�is�the�next����?stage,�d�after�one�\mo���v�e"�d�of�the�mac���hine.�There�is�an�extra�mark�(y�ou�can����?think�FLof�a�red�color)�in�the�cell�where�the�T��*�uring�mac���hine�head�is�lo�Gcated�at����?that�3Xstage.�When�w���e�refer�to�cell�(�i;���j����)�w�e�mean�the��j����-th�cell�in�the��i�-th�ro�w.����?The���axioms�sa���y�that�suc�h�a�table��T�^h�is�a�computation�b�y�mac�hine��M��if�for�all����?the���en���tries�in��T�c��,�the�con�ten�ts�of�cell�(�i�鐲+�1�;���j����)���are�related�to�the�con�ten�ts�of�the����?three�T�cells�(�i;���j��g��7ܲ1),�(�i;�j����),�and�(�i;�j��g�+�7�1)�according�to�the�program�of�T��*�uring����?mac���hine�3�M��.�Although�this�uses�natural�n�um�b�Gers�(�i;���j����)�to�refer�to�the�cells�of����?�T�c��,���only�a�few�basic�and�easily�axiomatizable�prop�Gerties�of�the�n���um�b�ers���are����?needed���for�suc���h�an�indexing.�Of�course,�it�tak�es�some�pages�to�ll�in�all�the����?details�~�of�the�rst�t���w�o�~�steps,�but�the�basic�idea�is�not�complicated�once�one����?understands�UUthe�concepts�in���v�olv�ed.����NT��*�uring's�Ayresult�sho���w�ed�Ayconclusiv�ely�that�it�will�nev�er�b�Ge�p�ossible�to�com-����?pletely�!mec���hanize�mathematics.�W��*�e�shall�nev�er�b�Ge�able�to�tak�e�all�our�math-����?ematical��questions�to�a�computer�and�get�a�correct�y���es-or-no�answ�er.�T��*�o����?understand�
the�denitiv���eness�of�T��*�uring's�result,�one�needs�G����odel's�complete-����?ness���theorem.�The�completeness�theorem�iden���ties�the�t�w�o�natural�meanings����?of���\logical�consequence":��P�f�is�a�logical�consequence�of��A�,�if��P��is�true�in�all����?systems�G3(mo�Gdels)�that�satisfy�axioms��A�.�On�the�other�hand,��P��²should�hop�e-����?fully���b�Ge�a�logical�consequence�of��A�,�if�and�only�if�there�exists�a�pro�of�of��P����?�from�l�A�.�This�turns�out�to�b�Ge�the�case,�and�is�exactly�the�con���ten�t�l�of�G����odel's����?completeness���theorem.�Therefore,�T��*�uring's�result�means�that�w���e�shall�nev�er����?b�Ge���able�to�tak���e�all�questions�of�the�form,�\do�es�theorem��P�r�follo���w�from�axioms����?�A�?"�UUto�a�computer�and�get�a�guaran���teed�correct�y�es�or�no�answ�er.��ҡ���?�5��S@Ch��urc�h��and�G��@odel��ҡ��?�T��*�uring's��negativ���e�solution�of�the�En�tsc�heidungsproblem�w�as�follo�w�ed�in�the����?1930's��1b���y�other�\negativ�e"�results.�In�1936,�Alonzo�Ch�urc�h�(1903-1995)�in-����?v���en�ted��fthe�lam���b�Gda-calculus�(often�written���-calculus)�and�used�it�to�giv�e�a����?denition�Hof��algorithm��dieren���t�from�T��*�uring's,�and�hence�an�indep�Genden�t�so-����?lution��Lof�the�En���tsc�heidungsproblem��L[29].�He�also�pro���v�ed��Lthe�result�w���e�no�w��?�Ϸ�ff8�ϟ
L͍�	?��according�k�as��S���pro��9v�es�k�B�ޡ�or�do�A�es�not�pro��9v�e�k�B�r��.�T��:�o�solv��9e�the�halting�problem,�whic�h����	?�is��whether�a�giv��9en�T��:�uring�mac�hine��M��ûhalts�at�a�giv�en�input��x�,�w�e�construct�the����	?�set���of�axioms��A��(dep�A�ending�on��M���)�as�in�the�rst�step.�W��:�e�then�construct�the����	?�sequence�vof�sym��9b�A�ols��y�j �expressing�\�M�Z�halts�at�input��x�".�According�to�step�2,��M����	?��halts���at��x��if�and�only�if��A��pro��9v�es���the�theorem��y�R��.�By�h��9yp�A�othesis,�w�e�can�determine����	?�this�3�b��9y�running�T��:�uring�mac�hine��K�ס�at�the�inputs��A��and��y�R��.�If�w�e�get�1,�then��M�,��halts����	?�at�Dt�x�,�and�if�w��9e�get�0,�it�do�A�es�not.�If��K��\�b�eha��9v�es�Dtas�w��9e�ha�v�e�supp�A�osed,�this�algorithm����	?�will��solv��9e�the�halting�problem.�Since�it�in�v�olv�es�only�T��:�uring�mac�hines�connected����	?�b��9y�A�simple�steps,�it�can�b�A�e�done�b�y�another�T��:�uring�mac�hine,�con�tradicting�T��:�uring's����	?�result��2on�the�unsolv��|rabilit��9y�of�the�halting�problem.�Hence�no�suc�h�mac�hine��K�b�can����	?�exist.�����
��c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������13����)�ō���>;��?�summarize�L�in�the�statemen���t,�\Arithmetic�is�undecidable".�Since�P�eano's�ax-����?ioms���are�not�rst-order,�the�En���tsc�heidungsproblem���do�Ges�not�directly�apply����?to���them,�and�one�can�ask�whether�there�could�b�Ge�an�algorithm�that�tak���es�a����?rst-order�#
statemen���t�ab�Gout�the�natural�n�um�b�Gers�as�input,�and�correctly�out-����?puts���\true"�or�\false".�The�En���tsc�heidungsproblem���do�Ges�not�apply��*�,�since�there����?exists��qno�(nite�rst-order)�system�of�axioms��A��whose�logical�consequences����?are�+the�statemen���ts�true�in�the�natural�n�um�b�Gers.�Ch�urc�h�sho�w�ed�that,�nev-����?ertheless,��Ithere�is�no�suc���h�algorithm.�Ch�urc�h's�studen�t�Kleene�pro�v�ed�the����?equiv��q�alence���of�the�T��*�uring-mac���hine�and�the���-calculus�denitions�of��algorithm����?�in�UUhis�Ph.�D.�thesis,�later�published�in�[60].����^��16����
�!��N�In�3.1931,�Kurt�G����odel�[45]�pro���v�ed�3.his�famous�\incompleteness�theorem",����?whic���h��;w�e�can�state�as�follo�ws:�Whatev�er�system�of�axioms�one�writes�do�wn�in����?an�
�attempt�to�axiomatize�the�truths�ab�Gout�the�natural�n���um�b�ers,�
�either�some����?false�m3statemen���t�will�b�Ge�pro�v�ed�from�the�axioms,�or�some�true�statemen�t�will����?not�b�Ge�pro���v�ed.�In�other�w���ords,�if�all�the�axioms�are�true,�then�some�true�fact����?will��,b�Ge�unpro���v��q�able�from�those�axioms.�G����odel�used�neither�T��*�uring�mac�hines����?nor����-calculus�(neither�of�whic���h�w�as�in�v�en�ted�un�til�v�e�y�ears�later),�but�in����?essence��\ga���v�e�a�third�denition�of��algorithm�.����^��17���KB�The�bulk�of�G����odel's�pap�Ger�is����?dev���oted,�'Fnot�to�his�essen�tial�ideas,�but�to�the�details�of�co�Gding�computations����?as�+�in���tegers;�although�he�did�not�use�T��*�uring�mac�hines,�he�still�had�to�co�Gde����?a���dieren���t�kind�of�computation�as�in�tegers.�No�w�ada�ys,�when�\Ascii�co�Gdes"����?used��b���y�computers�routinely�assign�a�n�um�b�Ger�to�eac�h�alphab�Getic�c�haracter,����?and�1�hence�reduce�a�line�of�text�to�a�v���ery�long�n�um�b�Ger,�using�three�digits�p�er����?c���haracter,��Ythis�seems�routine.�F��*�or�example,�`a'�has�the�Ascii�co�Gde�97,�`b'�is����?assigned��98,�`c'�gets�99,�and�so�on.�Th���us�\cat"�gets�the�n�um�b�Ger�099097116.����?Suc���h�7!enco�Gdings�can�also�b�e�used�to�sho���w�that�T��*�uring�mac�hine�computations����?can�UUb�Ge�enco�ded�in�n���um�b�ers.��
�!��NMaking���use�of�T��*�uring�mac���hines,�it�is�not�v�ery�dicult�to�understand����?the�.�main�idea�of�G����odel's�pro�Gof.�The�tec���hnical�details�ab�out�co�ding�can�b�e����?used��Yto�construct�a�n���um�b�Ger-theoretical��Yform�ula��T�c��(�e;���x;�y�[ٲ)�that�expresses�that����?�e��p�is�a�co�Gde�for�a�T��*�uring�mac���hine�(a�nite�set�of�instructions),�and��y��I�is�a����?co�Gde��for�a�complete�(halting)�computation�b���y�mac�hine��e��at�input��x�.�In�other����?w���ords,���\mac�hine��e��halts�at�input��x�"�can�b�Ge�expressed�b�y�\there�exists�a����?�y���suc���h�n
that��T�c��(�e;���x;�y�[ٲ)."�n
No�w�supp�Gose�that�w�e�had�a�correct�and�complete����?axiomatization����A��of�the�true�statemen���ts�of�arithmetic.�W��*�e�could�then�solv�e����?the���halting�problem�b���y�the�follo�wing�algorithm:�w�e�sim�ultaneously�try�to����?pro���v�e���\mac�hine��e��do�Ges�not�halt�at�input��e�"�from�the�axioms��A�,�and�w�e�run��?����ff8�ϟ
L͍������UZ��-=�16�����	?��Kleene��w��9en�t�on�to�b�A�ecome�one�of�the�t�w�en�tieth�cen�tury's�luminaries�of�logic;�his����	?�[61]��is�probably�the�most�in
uen��9tial�logic�textb�A�o�ok��ev�er�written,�and�he�laid�the����	?�foundations�of�\recursion�theory",�whic��9h�includes�the�sub���ject�no�w�kno�wn�as�the����	?�theory�Tof�computation.���������UZ��-=�17�����	?��G��`odel's�rdenition�seemed�at�the�time�rather�sp�A�ecialized,�and�(unlik��9e�T��:�uring�v�e����	?�y��9ears��3later)�he�made�no�claim�that�it�corresp�A�onded�to�the�general�notion�of����	?�\computable",�Tthough�that�turned�out�to�b�A�e�true.������ՠc�ō���>;����?�14���`-	Mic��9hael�TBeeson����)�ō���>;��?�mac���hine����e��at�input��e��to�see�if�it�halts.�Here�\sim�ultaneously"�can�b�Ge�tak�en�to����?mean��P\in�alternating�steps."�A���t�ev�en-n�um�b�Gered�stages,�w�e�run��e��at�input��e��for����?one��	more�step,�and,�at�o�Gdd-n���um�b�ered��	stages,�w���e�mak�e�one�more�deduction����?from�{�the�axioms��A�.�If��e��halts�at�input��e�,�w���e�nd�that�out�at�some�ev�en-����?n���um�b�Gered��stage.�Otherwise,�b���y�the�assumed�completeness�and�correctness�of����?the��Paxioms��A�,�w���e�succeed�at�some�o�Gdd-n�um�b�ered��Pstage�to�nd�a�pro�Gof�that����?�e���do�Ges�not�halt�at�input��e�.�But�since�the�halting�problem�is�unsolv��q�able,�this����?is���a�con���tradiction;�hence�no�suc�h�set�of�axioms��A��can�exist.�That�is�G����odel's����?incompleteness�UUtheorem.���k���?�6��S@The��P��ossible�Lo�`opholes���k��?�The�-Yresults�of�T��*�uring,�Ch���urc�h,�-Yand�G����odel�are�commonly�called�\negativ���e"����?results��in�that�they�sho���w�the�imp�Gossibilit�y�of�a�complete�reduction�of�mathe-����?matics���or�logic�to�computation.�Hilb�Gert's�program�w���as�a�hop�eless�pip�e�dream.����?These��famous�results�seem�to�close�the�do�Gors�on�those�who�w���ould�hop�e�to����?mec���hanize��mathematics.�But�w�e�are�not�completely�trapp�Ged;�there�are�the����?follo���wing�UUp�Gossible�\lo�opholes",�or�a���v�en�ues�UUthat�ma���y�still�pro�v�e�fruitful.���!��P��l#�Ma���yb�Ge�there�exist�in�teresting�axiom�systems��A��suc�h�that,�for�that����P�p��}'articular���axiom�system,�there��do�es��exist�a�\decision�pro�Gcedure",�that����Pp�Germits��Rus�to�compute�whether�a�giv���en�statemen�t��P���follo�ws�from��A����P�or�UUnot.���ۍ�P����Ma���yb�Ge�there�exist�in�teresting�algorithms��f�#A�that�tak�e�an�axiom�sys-����Ptem���A��and�an�input�form���ula��P�H��and,��sometimes�,�tell�us�that��P��follo���ws����Pfrom��A�.�Ev���en�if��f���is�not��guar��}'ante�e�d��to�w�ork�on��al���l��P�c��,�if�it�w�ould�w�ork����Pon�0��some��P��9�for�whic���h�w�e�did�not�kno�w�the�answ�er�b�Gefore,�that�w�ould����Pb�Ge�UUquite�in���teresting.����P���D�Ev���en�if�suc�h�an��f��Ӳw�ork�ed�only�for�a�particular�axiom�system��A��of����Pin���terest,�
it�still�migh�t�b�Ge�able�to�answ�er�mathematical�questions�that����Pw���e�UUcould�not�answ�er�b�Gefore.���F��NThese�>�lo�Gopholes�in�the�negativ���e�results�of�the�thirties�allo�w�the�partial����?mec���hanization���of�mathematics.�It�is�the�pursuit�of�these�p�Gossibilities�that����?o�Gccupies�UUthe�main�business�of�this�pap�er.���k���?�7��S@The��rst�theorem-pro��v�ers���k��?�When�h�the�computer�w���as�still�newb�Gorn,�some�p�eople�tried�to�write�programs����?exploiting��
the�lo�Gopholes�left�b���y�Ch�urc�h�and�G����odel.�The�rst�one�exploited�the����?p�Gossibilit���y���of�decision�pro�cedures.�There�w���as�already�a�kno�wn�decision�pro-����?cedure��Ifor�arithmetic�without�m���ultiplication.�This�is�essen�tially�the�theory�of����?linear�Jequations�with�in���teger�v��q�ariables,�and�\for�all"�and�\there�exists".�This����?theory�?|go�Ges�b���y�the�name�of�\Presburger�arithmetic",�after�M.�Presburger,������{�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������15����)�ō���>;��?�who��urst�ga���v�e��ua�decision�pro�Gcedure�for�it�in�[82].�It�cried�out�for�implemen-����?tation,�&�no���w�that�the�computer�w�as�more�than�a�though�t�exp�Gerimen�t.�Martin����?Da���vis���to�Gok�up�this�c�hallenge�[37],�and�in�1954�his�program�pro�v�ed�that�the����?sum��Rof�t���w�o��Rev�en�n�um�b�Gers�is�ev�en.�This�w�as�p�Gerhaps�the�rst�theorem�ev�er����?pro���v�ed���b�y�a�computer�program.�The�computer�on�whic�h�the�program�ran�w�as����?a��Ev��q�acuum�tub�Ge�computer�kno���wn�as�the�\johnniac",�at�the�Institute�for�Ad-����?v��q�anced���Study�in�Princeton,�whic���h�had�a�memory�of�1024�w�ords.�The�program����?could�UUuse�a�maxim���um�of�96�w�ords�to�hold�the�generated�form�ulas.����NIn�/�1955,�New���ell,�Sha�w,�and�Simon�wrote�a�program�they�called�the��L��}'o�gic����?The��}'orist�[74].�.�This�program�w���en�t�.�through�another�lo�Gophole:�it�tried�to�nd����?pro�Gofs,�Luev���en�though�according�to�T��*�uring�it�m�ust�fail�sometimes.�It�pro�v�ed����?sev���eral�I�prop�Gositional�logic�theorems�in�the�system�of��Principia�t�Mathemat-����?ic��}'a�.��The�authors�w���ere�proud�of�the�fact�that�this�program�w�as�\heuristic",����?b���y�%7whic�h�they�mean�t�not�only�that�it�migh�t�fail,�but�that�there�w�as�some����?analogy���b�Get���w�een�ho�w�it�solv�ed�problems�and�ho�w�a�h�uman�w�ould�solv�e�the����?same� �problems.�They�felt�that�a�heuristic�approac���h�w�as�necessary�b�Gecause����?the���approac���h�of�systematically�searc�hing�for�a�pro�Gof�of�the�desired�theorem����?from�*=the�giv���en�axioms�seemed�hop�Geless.�They�referred�to�the�latter�as�the����?\British�Museum"�algorithm,�comparing�it�to�searc���hing�for�a�desired�item�in����?the�0�British�Museum�b���y�examining�the�en�tire�con�ten�ts�of�the�m�useum.�Ac-����?cording��to�[38],�Alan�New���ell�said�to�Herb�Simon�on�Christmas�1955,�ab�Gout����?their��Pprogram,�\Kind�of�crude,�but�it�w���orks,�b�Go�y��*�,�it�w�orks!".�In�one�of�Simon's����?obituaries��?[66]�(he�died�in�2001�at�age�84),�one�nds�a�con���tin�uation��?of�this����?story:��;���P�The��0follo��9wing�Jan�uary��:�,�Professor�Simon�celebrated�this�disco�v�ery�b�y�w�alk-����Ping���in��9to�a�class�and�announcing�to�his�studen�ts,�\Ov�er�the�Christmas�holi-����Pda��9y��:�,��`Al�New�ell�and�I��Sin�v�en�ted�a�thinking�mac�hine."�A��Ssubsequen�t�letter�to����PLord��%Russell�explaining�his�ac��9hiev�emen�t��%elicited�the�reply�:�\I���am�deligh��9ted����Pto��"kno��9w�that�`Principia�Mathematica'�can�no�w�b�A�e�done�b�y�mac�hinery��:�.�I����Pwish���Whitehead�and�I���had�kno��9wn�of�this�p�A�ossibilit�y�b�A�efore�w�e�w�asted�10����Py��9ears�Tdoing�it�b�y�hand."�����-=�18�������N�In�^�1957,�the�y���ear�of�publication�of�New�ell,�Sha�w,�and�Simon's�rep�Gort�[74],�a����?v���e���w�eek�Summer�Institute�for�Sym�b�Golic�Logic�w�as�held�at�Cornell,�attended����?b���y���man�y�American�logicians�and�some�researc�hers�from�IBM.�A�t�this�meet-����?ing,���Abraham�Robinson�in���tro�Gduced�the�idea�of�Sk�olem�functions�[explained����?b�Gelo���w],���and�shortly�after�the�meeting�a�n�um�b�Ger�of�imp�ortan���t�adv��q�ances�w�ere����?made.�
�Sev���eral�new�programs�w�ere�written�that�searc�hed�more�systematically����?for�ڦpro�Gofs�than�the��L��}'o�gic�#
The�orist�ڦ�had�done.�The�problem�w���as�clearly�seen�as����?\pruning"��|the�searc���h,�i.e.�eliminating�fruitless�deductions�as�early�as�p�Gossible.����?Gelern���ter's�Y�geometry�pro�v�er�[44]�used�a�\diagram"�to�prune�false�goals.�The����?mathematical��logician�Hao�W��*�ang�wrote�a�program�[103]�based�on�a�logical����?system��<kno���wn�as�\natural�deduction".�W��*�ang's�program�pro�v�ed�all�400�pure����?predicate-calculus���theorems�in��Principia���Mathematic��}'a�.�Da���vis�and�Putnam��?��;�ff8�ϟ
L͍������UZ��-=�18�����	?��Russell�Tma��9y�ha�v�e�had�his�tongue�rmly�in�c�heek.�����܅�c�ō���>;����?�16���`-	Mic��9hael�TBeeson����)�ō���>;��?�[40]��3published�a�pap�Ger�that�coupled�the�use�of�Sk���olem�functions�and�conjunc-����?tiv���e�t�normal�form�with�a�b�Getter�algorithm�to�determine�satisabilit�y��*�.�Ov�er����?the���next�sev���eral�y�ears,�these�strands�of�dev�elopmen�t�led�to�the�in�v�en�tion�of����?fundamen���tal���algorithms�that�are�still�in�use.�W��*�e�shall�discuss�three�of�these����?to�Gols:�UUSk���olemization,�resolution,�and�unication.���A��NSk���olem�H�functions�are�used�to�systematically�eliminate�\there�exists".�F��*�or����?instance,�8�\for�ev���ery��x��there�exists��y����suc�h�that��P�c��(�x;���y�[ٲ)"�is�replaced�b�y��P�c��(�x;���g�[ٲ(�x�)),����?where��͵g����is�called�a�\Sk���olem�function".�When�w�e�express�the�la�w�that�ev�ery����?nonzero���x��has�a�m���ultiplicativ�e���in�v�erse�in�the�form��x���6�=�0��!��x����x���^��O!�cmsy7��1��
���=��1,�w���e�are����?using�4�a�Sk���olem�function�(written�as��x���^���1��
�S�instead�of��g�[ٲ(�x�).�T��*�erms�are�built�up,����?using�S-function�and�op�Geration�sym���b�ols,�from�v��q�ariables�and�constan���ts;�usually����?letters�P*near�the�b�Geginning�of�the�alphab�et�are�constan���ts�and�letters�near�the����?end�f�are�v��q�ariables�(a�con���v�en�tion�f�in�tro�Gduced�b�y�Descartes).�Certain�terms�are����?distinguished��as�\prop�Gositions";�in���tuitiv�ely��these�are�the�ones�that�should�b�e����?either��1true�or�false�if�the�v��q�ariables�are�giv���en�sp�Gecic�v�alues.�The�use�of�Sk���olem����?functions��and�elemen���tary�logical�manipulations�enables�us�to�express�ev�ery����?axiom���and�theorem�in�a�certain�standard�form�called�\clausal�form",�whic���h����?w���e��no�w�explain.�A���liter��}'al��is�an�atomic�prop�Gosition�or�its�negation.�A��clause��is����?a���\disjunction�of�literals";�that�is,�a�list�of�literals�separated�b���y�\or".�Giv�en����?some�R�axioms�and�a�conjectured�theorem,�w���e�negate�the�theorem,�and�seek�a����?pro�Gof��b���y�con�tradiction.�W��*�e�use�Sk�olem�functions�and�logical�manipulations����?to���eliminate�\there�exists",�and�then�w���e�use�logical�manipulations�to�bring����?the�1�axioms�and�negated�goal�to�the�form�of�a�list�of�clauses,�where�\and"����?implicitly��joins�the�clauses.�This�pro�Gcess�is�kno���wn�as�\Sk�olemization."�The����?clausal�_�form�con���tains�no�\there�exists",�but�it�do�Ges�con�tain�new�sym�b�Gols�for����?the��(unkno���wn)�Sk�olem�functions.�The�original�question�whether�the�axioms����?imply��the�goal�is�equiv��q�alen���t�to�the�more�con�v�enien�t�question�whether�the����?resulting�UUlist�of�clauses�is�con���tradictory�or�not.���A��NIn���automated�deduction,�it�is�customary�to�use�the�v���ertical�bar�to�mean����?\or",��and�the�min���us�sign�to�mean�\not".�An��infer��}'enc�e�rule��is�a�rule�for����?deducing��"theorems�from�previously-deduced�theorems�or�axioms.�It�therefore����?has�/_\premisses"�and�\conclusions".�As�an�example�of�an�inference�rule�w���e����?men���tion�+athe�rule��mo��}'dus�mNp�onens�,�+awhic�h�is�already�o�v�er�2000�y�ears�old:�from��p����?�and�<\if��p��then��q�[ٲ"�infer��q��.�In�clausal�notation�that�w���ould�b�Ge,�from��p��and���p�j�q����?�infer�2�q�[ٲ.��R��}'esolution��generalizes�this�rule.�In�its�simplest�form�it�sa���ys,�from��p�j�r����?�and�����p�j�q�[ٲ,�infer��r�G�j�q��.�Ev���en�more�generally��*�,��r����and��q�h�can�b�Ge�replaced�with�sev�eral����?prop�Gositions.�`F��*�or�example,�from��p�j�r��j�s��and���p�j�q�[ٸj�t�,�w���e�can�infer��r��j�s�j�q�[ٸj�t�.�The����?rule�>ncan�b�Ge�though���t�of�as�\cancelling"��p��with���p�.�The�cancelled�term��p��do�es����?not��ha���v�e�to�b�Ge�the�rst�one�listed.�If�w�e�deriv�e��p��and�also���p�,�then�resolution����?leads�UUto�the�\empt���y�clause",�whic�h�denotes�a�con�tradiction.���A��NThe�7�third�of�the�three�to�Gols�w���e�men�tioned�is�the��unic��}'ation�d7algorithm�.����?This���w���as�published�b�y�J.�A.�Robinson[89].�Robinson's�publication�(whic�h����?con���tained���more�than�\just"�unication)�app�Geared�in�1965,�but�at�that�time����?unication�v�w���as�already�in�the�air,�ha�ving�b�Geen�implemen�ted�b�y�others�as�early�����쏠c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������17����)�ō���>;��?�as�O1962.�See�[38]�for�this�history��*�.�The�purp�Gose�of�the�unication�algorithm����?is��to�nd�v��q�alues�of�v�ariables�to�mak���e�t�w�o�terms�matc�h.�F��*�or�example:�giv�en����?�f���(�x;���g�[ٲ(�x�))��
and��f��(�g�[ٲ(�c�)�;���z�p��),�w���e�nd��x�OJ�=��g��(�c�),��
�z���=��g��(�g��(�c�))��
b���y�applying�unica-����?tion.�.The�input�to�the�algorithm�is�a�pair�of�terms�to�b�Ge�unied.�The�output����?is���a�substitution;�that�is,�an�assignmen���t�of�terms�to�v��q�ariables.�W��*�e�shall�not����?giv���e��the�details�of�the�unication�algorithm�here;�they�can�b�Ge�found�in�man�y����?b�Go�oks,�UUfor�example�in�[25],�Ch.�17,�or�[5],�pp.�453���.���ˍ�NCom���bining���resolution�and�unication,�w�e�arriv�e�at�the�follo�wing�rule�of����?inference:�Supp�Gose�that��p��and��s��can�b�e�unied.�Let����denote�the�substitu-����?tion���found�b���y�the�unication�algorithm.�Then�from��p�j�q�ߖ�and���s�j�r��ڲinfer��q��[ٟ�^�����j�r��G��^�������?�pro���vided�N��p���^����_��=���s���^������.�This�rule�is�also�commonly�kno�wn�as�\resolution"{in�fact,����?resolution��without�unication�is�only�of�historical�or�p�Gedagogical�in���terest.����?Resolution��is��always��com���bined�with�unication.�J.�A.�Robinson�pro�v�ed�[89]����?that���this�rule�is��r��}'efutation�,�c�omplete�.���That�means�that�if�a�list�of�clauses�is����?con���tradictory��*�,�O�there�exists�a�pro�Gof�of�the�empt�y�clause�from�the�original�list,����?using�UUresolution�as�the�sole�rule�of�inference.����^��19�����ˍ�N�The���basic�paradigm�for�automated�deduction�then�w���as�b�Gorn:�Start�with����?the�ioaxioms�and�negated�goal.�P���erform�resolutions�(using�unication)�un�til�a����?con���tradiction��-is�reac�hed,�or�un�til�y�ou�run�out�of�time�or�memory��*�.�The�mo�Gdern����?era��lin�automated�deduction�could�b�Ge�said�to�ha���v�e��lb�egun�when�this�paradigm����?w���as�Ʒin�place.����^��20���?��One�v�ery�imp�Gortan�t�strand�of�w�ork�in�the�sub��8ject�since�the����?sixties�m�has�b�Geen�dev���oted�to�v��q�arious�attempts�to�prev�en�t�running�out�of�time����?or�(memory��*�.�These�attempts�will�b�Ge�discussed�in�the�section�\Searc���hing�for����?pro�Gofs"�UUb�elo���w.����^��21����?�.k\�ff8�ϟ
L͍������UZ��-=�19�����	?��W��:�e�3`ha��9v�e�o�v�ersimplied�in�the�text.�The�resolution�rule�as�w�e�ha�v�e�giv�en�it�do�A�es����	?�not�)Op�A�ermit�one�to�infer��p�(�z�c��)�from��p�(�x�)�j�p�(�y�R��).�Either�the�resolution�rule�has�to�b�e����	?�stated�nha�bit�more�generally��:�,�as�Robinson�did,�or�w��9e�ha�v�e�to�supplemen�t�it�with����	?�the���rule�called��factoring�,�whic��9h�sa�ys�that�if��A��and��B�k��can�b�A�e�unied,�and����is�the����	?�substitution�Tpro�A�duced�b��9y�the�unication�algorithm,�w�e�can�infer��A��.���������UZ��-=�20�����	?��There�#w��9ere�sev�eral�more�attempts�to�write�programs�that�pro�v�ed�theorems����	?�\heuristically",��to�some�exten��9t�trying�to�imitate�h�uman�though�t,�but�in�the�end����	?�these�Tprograms�could�not�comp�A�ete�with�an�algorithmic�searc��9h.���������UZ��-=�21�����	?��It��Gis�true�that�sev��9eral�other�approac�hes�ha�v�e�b�A�een�dev�elop�A�ed,�and�ha�v�e�succeeded����	?�on���some�problems.�W��:�e�note�in�particular�the�successes�of�A��9CL2�[20]�and�RRL����	?�[59]�]�on�problems�in��9v�olving�]�mathematical�induction,�and�regret�that�our�limited����	?�space���and�scop�A�e�do�not�p�ermit�a�fuller�discussion�of�alternativ��9e�approac�hes.�The����	?�author�̢is�partial�to�approac��9hes�deriv�ed�from�the�branc�h�of�mathematical�logic����	?�kno��9wn��as�\pro�A�of�theory";�in�the�USSR��this�approac�h�w�as�follo�w�ed�early�on,�and����	?�an�o�algorithm�closely�related�to�resolution�w��9as�in�v�en�ted�b�y�Maslo�v�at�ab�A�out�the����	?�same�e�time�as�resolution�w��9as�in�v�en�ted.�A�e�theorem-pro�v�er�based�on�these�principles����	?�w��9as�Tbuilt�in�Leningrad�(1971).�See�[68]�for�further�details�and�references.��������c�ō���>;����?�18���`-	Mic��9hael�TBeeson����)�ō���>;���?�8��W�Kinds��of�Mathematical�Reasoning��ڕ��?�In��Othis�section,�w���e�abandon�the�historical�approac�h�to�the�sub��8ject.�Instead,����?w���e�y�examine�the�mec�hanization�of�mathematics�b�y�taking�in�v�en�tory�of�the����?mathematics�:[to�b�Ge�mec���hanized.�Let�us�mak�e�a�rough�taxonom�y�of�mathe-����?matics.��`Of�course�librarians�and�journal�editors�are�accustomed�to�classifying����?mathematics��b���y�sub��8ject�matter,�but�that�is�not�what�w�e�ha�v�e�in�mind.�In-����?stead,��qw���e�prop�Gose�to�classify�mathematics�b�y�the��kind��of�pr��}'o�ofs��q�that�are�used.����?W��*�e�UUcan�distinguish�at�least�the�follo���wing�categories:��ڕ�����E�������P�Purely�UUlogical��+������E�������P�Simple�UUtheory��*�,�as�in�geometry�(one�kind�of�ob��8ject,�few�relations)�������E�������P�Equational,�UUas�in�the�Robbins�problem,�or�in�group�or�ring�theory��*�.�������E�������P�Uses�UUcalculations,�as�in�algebra�or�calculus�������E�������P�Uses�UUnatural�n���um�b�Gers�UUand�mathematical�induction�������E�������P�Uses�UUdenitions�(p�Gerhaps�lots�of�them)�������E�������P�Uses��a�little�n���um�b�Ger��theory�and�simple�set�theory�(as�in�undergraduate����Palgebra�UUcourses)�������E�������P�Uses�UUinequalities�hea���vily�(as�in�analysis)���ݍ�NPurely��logical�theorems�are�more�in���teresting�than�ma�y�app�Gear�at�rst����?blush.� One�is�not�restricted�to�logical�systems�based�on�resolution�just�b�Ge-����?cause��Gone�is�using�a�theorem-pro���v�er��Gthat�w���orks�that�w�a�y��*�.�There�are�h�undreds����?of�in���teresting�logical�systems,�including�v��q�arious�axiom�systems�for�classical����?prop�Gositional�Zlogic,�m���ulti-v��q�alued�logic,�mo�dal�logic,�in���tuitionistic�logic,�etc.����?All�8~of�these�can�b�Ge�analyzed�using�the�follo���wing�metho�d.�W��*�e�use�a�predicate����?�P�c��(�x�)�7�to�stand�for�\�x��is�pro���v��q�able".�W��*�e�use��i�(�x;���y�[ٲ)�to�mean��x��implies��y��.�Then,����?for���example,�w���e�can�write�do�wn���P�c��(�x�)�j����P��(�i�(�x;���y�[ٲ))�j�P��(�y��)���to�express�\if��x��and����?�i�(�x;���y�[ٲ)�Lare�pro���v��q�able,�so�is��y��."�When�(a�commonly-used�v��q�arian���t�of��)�resolution����?is��tused�with�this�axiom,�it�will�ha���v�e��tthe�same�eect�as�an�inference�rule�called����?\condensed��detac���hmen�t"�that�has�long�b�Geen�used�b�y�logicians.�W��*�e�will�return����?to�1�this�discussion�near�the�end�of�the�pap�Ger,�in�the�section�on�\Searc���hing�for����?pro�Gofs".����NEuclidean�!&geometry�can�b�Ge�form���ulated�in�a�rst-order�theory�with�a�sim-����?ple,�OLnatural�set�of�axioms.�In�fact,�it�can�b�Ge�form���ulated�in�a�theory�all�of����?whose�4�v��q�ariables�stand�for�p�Goin���ts;�direct�references�to�lines�and�planes�can�b�e����?eliminated��[97].�But�that�is�not�imp�Gortan���t|w�e��could�use�unary�predicates����?for��;p�Goin���ts,�lines,�and�planes,�or�w�e�could�use�three�\sorts"�of�v��q�ariables.�What����?w���e�jEcannot�do�in�suc�h�a�theory�is�men�tion�arbitrary�sets�of�p�Goin�ts;�therefore,����?the��4con���tin�uit�y�axiom�(discussed�ab�Go�v�e)�cannot�b�Ge�stated�in�suc�h�a�theory��*�.�W�e����?can�'�state�some�instances�of�the�con���tin�uit�y�'�axiom�(for�example,�that�a�line����?segmen���t���with�one�end�inside�a�circle�and�one�end�outside�the�circle�m�ust�meet����?the��Lcircle);�or�w���e�could�ev�en�consider�a�theory�with�an��axiom��schema��(in-����?nitely�`man���y�axioms�of�a�recognizable�form)�stating�the�con�tin�uit�y�axiom�for����?all��rst-order�denable�sets.�But�if�w���e�are�in�terested�in�Euclid's�prop�Gositions,�����
Ӡc�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������19����)�ō���>;��?�extremely��complex�forms�of�the�con���tin�uit�y��axiom�will�not�b�Ge�necessary{w���e����?can��~consider�a�simple�theory�of�geometry�instead.�It�will�not�pro���v�e��~all�the�the-����?orems���one�could�pro���v�e���with�the�full�rst-order�con���tin�uit�y���axiom,�but�w���ould����?b�Ge��sucien���t�for�Euclid.�On�the�other�hand,�if�w�e�wish�to�pro�v�e�a�theorem����?ab�Gout��all�regular��n�-gons,�the�concept�of�natural�n���um�b�er��will�b�e�required,�and����?pro�Gofs�kb���y�mathematical�induction�will�so�on�arise.�In�rst-order�geometry��*�,�w���e����?w���ould�~�ha�v�e�one�theorem�for�a�square,�another�for�a�p�Gen�tagon,�another�for�a����?hexagon,��and�so�on.�Of�course�not�only�Euclidean,�but�also�non-Euclidean����?geometry��*�,�^�can�b�Ge�form���ulated�in�a�rst-order�theory�.�I�^�kno���w�of�no�w�ork�in����?automated�r�deduction�in�non-Euclidean�geometry��*�,�but�there�exists�at�least����?one���in���teresting�op�Gen�problem�in�h�yp�Gerb�olic���geometry�whose�solution�migh�t����?b�Ge�UUp�ossible�with�automated�deduction.����^��22�������N�Another�hexample�of�a�simple�theory�is�ring�theory��*�.�Ring�theory�is�a�sub��8ject����?commonly��taugh���t�in�the�rst�y�ear�of�abstract�algebra.�The�\ring�axioms"�use����?the�@�sym���b�Gols�+�and���,�and�include�most�of�the�familiar�la�ws�ab�Gout�them,�except����?the�q4\m���ultiplicativ�e�in�v�erse"�la�w�and�the�\comm�utativ�e�la�w�of�m�ultiplication",����?�x��>���y�k^�=���y����x�.�dMan���y�sp�Gecic�systems�of�mathematical�ob��8jects�satisfy�these����?la���ws,��and�ma�y�or�ma�y�not�satisfy�additional�la�ws�suc�h�as��x�����y�EG�=��n�y�۸��x�.��A����?system�p�of�ob��8jects,�with�t���w�o�p�giv�en�(but�p�Gossibly�arbitrarily�dened)�op�erations����?to�Ob�Ge�denoted�b���y�the�sym�b�Gols�+�and���,�is�called�a��ring��if�all�the�ring�axioms����?hold��when�the�v��q�ariables�range�o���v�er��these�ob��8jects�and�+�and����are�in���terpreted����?as�wthe�giv���en�op�Gerations.�In�ring�theory��*�,�one�tries�to�pro�v�e�a�theorem�using����?only��ythe�ring�axioms;�if�one�succeeds,�the�theorem�will�b�Ge�true�in�all�rings.����?Ho���w�ev�er,�P*in�b�Go�oks�P*on�ring�theory�one�nds�man���y�theorems�ab�Gout�rings�that����?are�v�not�form���ulated�purely�in�the�language�of�ring�theory��*�.�These�theorems�ha�v�e����?a���larger�con���text:�they�deal�with�rings�and�subrings,�with�homomorphisms�and����?isomorphisms��Pof�rings,�and�with�matrix�rings.�Homomorphisms�are�functions����?from���one�ring�to�another�that�preserv���e�sums�and�pro�Gducts;�isomorphisms����?are�B�one-to-one�homomorphisms;�subrings�are�subsets�of�a�ring�that�are�rings����?in���their�o���wn�righ�t;�matrix�rings�are�rings�whose�elemen�ts�are�matrices�with����?co�Gecien���ts�R]dra�wn�from�a�giv�en�ring.�Th�us�passing�from�a�ring��R�f$�to�the�ring����?of�[k�n��b���y��n��matrices�with�co�Gecien�ts�in��R�o2�is�a�metho�Gd�of�constructing�one����?ring�F�from�another.�If,�ho���w�ev�er,�F�w�e�wish�to�consider�suc�h�rings�of�matrices����?for��=an���y��n�,�then�the�concept�of�natural�n�um�b�Ger�en�ters�again,�and�w�e�are����?b�Gey���ond��the�simple�theory�lev�el.�Also,�if�w�e�wish�to�form�ulate�theorems�ab�Gout����?arbitrary���subrings�of�a�ring,�again�w���e�ha�v�e�a�theory�that�(at�least�on�the�face����?of�)it)�is�second-order.�A�(�recen���t�master's�thesis�[54]�w�en�t�through�a�t�ypical��?�Xt�ff8�ϟ
L͍������UZ��-=�22�����	?��The��Kop�A�en�problem�is�this:�Giv��9en�a�line��L��and�a�p�oin��9t��P�.�not�on��L�,�pro�v�e�that����	?�there��exist�a�pair�of��limiting�R�p��ar�al�x�lels���to��L��through��P�H�.�The�denition�of�limiting����	?�parallel�|]sa��9ys�that��K� E�and��R����form�a�pair�of�limiting�parallels�to��L��through����P��@�if����	?�one��Fof�the�four�angles�formed�at��P��)�b��9y��K�+.�and��R����do�A�es�not�con�tain�an�y�ra�y�that����	?�do�A�es��"not�meet��L�.�It�is�kno��9wn�that�limiting�parallels�exist,�but�no�rst-order�pro�of����	?�is�
kno��9wn,�and�exp�A�erts�tell�me�that�pro�ducing�a�rst-order�pro�of�w��9ould�b�e�w��9orth����	?�a�TPh.�D.�������c�ō���>;����?�20���`-	Mic��9hael�TBeeson����)�ō���>;��?�algebra�j�textb�Go�ok�[56],�and�found�that�of�ab�out�150�exercises�on�ring�theory��*�,����?14�Ocould�b�Ge�straigh���tforw�ardly�Oformalized�in�rst-order�ring�theory��*�.�One�more����?could�hb�Ge�form���ulated�using�a�single�natural-n�um�b�Ger�v��q�ariable�in�addition�to����?the�yring�axioms.�The�rest�w���ere�more�complex.�The�14�rst-order�exercises,����?ho���w�ev�er,�could�b�Ge�pro���v�ed�b�y�the�theorem-pro�ving�program�Otter.�(Otter�is�a����?w���ell-kno�wn���and�widely�used�mo�Gdern�theorem�pro���v�er,���describ�ed�in�[70],�and����?readily�UUa���v��q�ailable�on�the�W��*�eb.)���A��NA�J�great�J�man���y�mathematical�pro�Gofs�seem�to�dep�end�on�calculations�for����?some��of�the�steps.�In�fact,�t���ypically�a�mathematical�pro�Gof�consists�of�some����?parts�w�that�are�calculations,�and�some�parts�that�are�logical�inferences.�Of����?course,�F�it�is�p�Gossible�to�recast�calculations�as�logical�pro�ofs,�and�it�is�p�os-����?sible�s to�recast�logical�pro�Gofs�as�calculations.�But�there�is�an�in���tuitiv�e�s dis-����?tinction:���a�calculation�pro�Gceeds�in�a�straigh���tforw�ard���manner,�one�step�after����?another,�+}applying�ob���vious�rules�at�eac�h�step,�un�til�the�answ�er�is�obtained.����?While�4�p�Gerforming�a�calculation,�one�needs�to�b�e�careful,�but�one�do�es�not����?need�.to�b�Ge�a�genius,�once�one�has�gured�out�what�calculation�to�mak���e.�It�is����?\merely�LIa�calculation."�When�nding�a�pro�Gof,�one�needs�insigh���t,�exp�erience,����?in���telligence{ev�en��genius{to�succeed,�b�Gecause�the�searc���h�space�is�to�o�large�for����?a�UUsystematic�searc���h�to�succeed.���A��NIt�mis�not�surprising�that�a�go�Go�d�mdeal�of�progress�has�b�Geen�made�in�mec���ha-����?nizing��3those�parts�of�pro�Gof�that�are�calculations.�It�ma���y�b�e�sligh���tly�surprising����?that�EYmetho�Gds�ha���v�e�EYb�een�found�for�automatically�disco���v�ering�EYnew�rules�to�b�e����?used�\�for�calculations.�F��*�urthermore,�the�relations�b�Get���w�een�\�the�computational����?parts���of�pro�Gofs�and�the�logical�parts�ha���v�e���b�een�explored�to�some�exten���t.�Ho�w-����?ev���er,�#7there�is�still�some�w�ork�to�b�Ge�done�b�efore�this�sub��8ject�is�nished,�as�w���e����?will�UUdiscuss�in�more�detail�b�Gelo���w.���A��NOne�~�asp�Gect�of�mathematics�that�has�not�b�een�adequately�mec���hanized�at����?the���presen���t�time�is��denitions�.�Let�me�giv�e�a�few�examples�of�the�use�of����?denitions��bin�mathematics.�The�concept�\�f���is�con���tin�uous��bat��x�",�where��f��is����?a�6{real-v��q�alued�function,�has�a�w���ell-kno�wn�6{denition:�\for�ev���ery���>V>��0�6{there����?exists�h���H>��;�0�suc���h�that�for�all��y��v�with��j�y�����E��x�j��<��`�,�w���e�ha�v�e��j�f���(�x�)�E����f��(�y�[ٲ)�j��;�<��."����?One���imp�Gortan���t�virtue�of�this�denition�is�that�it�sw�eeps�the�quan�tiers�\for����?ev���ery"��and�\there�exists"�under�the�rug:�W��*�e�are�able�to�w�ork�with�con-����?tin���uit�y�d�in�a�quan���tier-free�con�text.�If,�for�example,�w�e�wish�to�pro�v�e�that����?�f���(�x�)��=�(�x��}�+�3)���^��100�����is�Ba�con���tin�uous�Bfunction,�the�\easy�w���a�y"�Bis�to�recognize����?that�d͵f�x\�is�a�comp�Gosition�of�t���w�o�d�con�tin�uous�functions�and�app�Geal�to�the�theorem����?that��&the�comp�Gosition�of�t���w�o��&con�tin�uous�functions�is�con�tin�uous.�That�theo-����?rem,�Jho���w�ev�er,�has�to�b�Ge�pro�v�ed�b�y�expanding�the�denitions�and�using����and����?��`�.��This�kind�of�argumen���t�do�Ges�not�mesh�w�ell�with�the�clausal�form�paradigm����?for��automated�reasoning,�b�Gecause�when�the�denition�is�expanded,�the�result����?in���v�olv�es�[�quan�tiers.�Theorem-pro�ving�programs�usually�require�clausal�form����?at��input,�and�do�not�p�Gerform�dynamic�Sk���olemization.�Theorems�that�ha�v�e����?b�Geen�>pro���v�ed�ab�Gout�con�tin�uit�y�ha�v�e,�therefore,�had�the�denition-expansion����?and���Sk���olemization�p�Gerformed�b�y�hand�b�Gefore�the�automated�deduction�pro-�����-��c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������21����)�ō���>;��?�gram�8�b�Gegan,�or�ha���v�e�8�used�another�paradigm�(Gen���tzen�sequen�ts�or�natural����?deduction),�r�that�do�Ges�not�suer�from�this�problem,�but�is�not�as�w���ell-suited����?to���searc���hing�for�pro�Gofs.�Merely�recognizing��f���(�x�)�WK=�(�x�r��+�3)���^��100��!3�as���a�comp�osi-����?tion�1�of�t���w�o�1�functions�is�b�Gey���ond�the�reac�h�of�curren�t�theorem-pro�v�ers{it�is�an����?application�UUof�the�author's�curren���t�researc�h�in�to�\second-order�unication".����NOne��Dmigh���t�w�ell�lo�Gok,�therefore,�for�the��simplest��example�of�a�denition.����?Consider��7the�denition�of�a�\comm���utator"�in�group�theory��*�.�The�notation����?usually��used�for�a�comm���utator�is�[�x;���y�[ٲ],�but�to�a�v�oid�notational�complexities,����?let��us�use�the�notation��x�U6�
��y�[ٲ.��The�denition�is��x�U6�
��y�iʲ=�
�x���^���1��
�t�y����^���1��M�xy��,��where�as����?usual��w���e�lea�v�e�the�sym�b�Gol����for�the�group�op�eration�un���written,�and�assume����?that�`Vasso�Gciation�is�to�the�righ���t,�i.e.��abc���=��a�(�bc�).�`VW��*�e�can�nd�problems�in�group����?theory�Iqthat�men���tion�comm�utators�but�do�not�need�second-order�concepts�or����?natural��n���um�b�Gers�for�their�form�ulation�or�solution.�Here�w�e�ha�v�e�a�single����?denition�!�added�to�a�simple�theory��*�.�No���w�the�p�Goin�t�is�that�sometimes�w�e�will����?need�&$to�recognize�complicated�expressions�as�b�Geing�actually�\nothing�but"����?a���comm���utator.�Long�expressions�b�Gecome�short�ones�when�written�using�the����?comm���utator���notation.�On�the�other�hand,�sometimes�w�e�will�not�b�Ge�able�to����?solv���e�o)the�problem�without�using�the�denition�of��x�l��
��y���to�o)eliminate�the�sym�b�Gol����?�
�.��That�is,�sometimes�the�denition�of��x��o�
��y�L��will��b�Ge�needed�in�the�left-to-����?righ���t�O;direction,�and�sometimes�in�the�righ�t-to-left�direction.�Existing�theorem-����?pro���v�ers�s?ha�v�e�no�metho�Gd�to�con�trol�equations�with�this�degree�of�subtlet�y��*�.����?Either�>ø
��will��always��b�Ge�eliminated,�or��never�.�This�example�denition�also����?serv���es�A�to�bring�out�another�p�Goin�t:�denitions�can�b�Ge�explicit,�lik�e�the�denition����?of���x���
��y�w��giv���en�ab�Go�v�e,�or�implicit.�Cancellativ�e�semigroups�are�systems�lik�e����?groups�R-except�that�in���v�erse�R-is�replaced�b���y�the�cancellation�la�w,��xy�"�=���xz��IJimplies����?�y�k�=�<�z�p��.���W��*�e�can�dene��x�U��
��y��w�in���the�con���text�of�cancellativ�e�semigroups�b�y�the����?equation�Hn�xy�"�=���y�[�x�(�x��
��y��).�HnThis�is�an�\implicit�denition".�If�the�la���w�holds�in����?a�UUsemigroup��S����,�for�some�op�Geration��
�,�w���e�sa�y�\�S���admits�comm�utators."����NConsider�t�the�follo���wing�three�form�ulas,�tak�en�from�[41],�and�originally�from����?[64].��q덍��X(�x�8�
��y�[ٲ)��
��z������J�=������,�x�8�
��(�y����
��z�p��)��!UX(1)�UU�c��}'ommutator���is�asso�ciative���������Z��(�x�8���y�[ٲ)��
��z������J�=������,(�x�8�
��z�p��)����(�y����
��z��)��
(2)�UU�c��}'ommutator���distributes�over�pr�o�duct���������Z��(�x�8�
��y�[ٲ)����z������J�=������,�z��w��8�(�x��
��y�[ٲ)��!UX(3)�UU�semigr��}'oup���is�nilp�otent�class�2�������?�These��three�prop�Gerties�are�equiv��q�alen���t�in�groups�(in�fact,�in�cancellativ�e�semi-����?groups��xthat�admit�comm���utators).�One�of�the�p�Goin�ts�of�considering�this�exam-����?ple��is�that�it�is�not�clear�(to�the�h���uman�mathematician)�whether�one�ough�t����?to� �eliminate�the�denition�of��x��a�
��y�|p�to� �pro���v�e�these�theorems,�or�not.�Otter����?is�}Iable�to�pro���v�e�}I(1)�implies�(2),�(2)�implies�(3),�and�(3)�implies�(1),�in�three����?separate�m2runs,�in�spite�of�not�ha���ving�a�systematic�w�a�y�to�handle�denitions;����?but�ithe�pro�Gofs�are�not�found�easily��*�,�and�a�lot�of�useless�clauses�are�generated����?along�UUthe�w���a�y��*�.����^��23����?��#�ff8�ϟ
L͍������UZ��-=�23�����	?��An�g�example�of�the�use�of�a�denition�to�help�Otter�nd�a�pro�A�of�that�it�cannot�nd����	?�without��using�a�denition�is�the�pro�A�of�of�the�\HCBK-1�problem"�found�recen��9tly�����>P�c�ō���>;����?�22���`-	Mic��9hael�TBeeson����)�ō���>;��N�Another���in���teresting�problem�in�v�olving�comm�utators�is�often�an�exercise�in����?an���elemen���tary�abstract�algebra�course:�Sho�w�that�in�a�group,�the�comm�utator����?subgroup���(consisting�of�all��x�_��
��y�[ٲ)���is�a�normal�subgroup.�F��*�or�the�part�ab�Gout����?normalit���y��*�,��ew�e�ha�v�e�to�sho�w�that�for�all��a�,�b�,�and��c�,��c���^���1��
�t�(�a�b?�
��b�)�c��has�the�form����?�u�yw�
��v�Qy�for���some��u��and��v�[ٲ.�Otter�can�nd�sev���eral�pro�Gofs�of�this�theorem,�but�the����?�u��βand��v�鱗in�the�rst�few�pro�Gofs�are�not�the�ones�a�h���uman�w�ould�nd|although����?it���do�Ges�ev���en�tually���nd�the�h���uman�pro�of|and�Otter�do�es�a�fairly�large�searc���h,����?while�UUa�h���uman�do�Ges�v�ery�little�searc�hing�on�this�problem.����NIn�Ѩmathematics�up�through�calculus,�if�w���e�do�not�go�deeply�in�to�the�foun-����?dations�8vof�the�sub��8ject�but�consider�only�what�is�actually�taugh���t�to�studen�ts,����?there���is�mostly�calculation.�In�abstract�algebra,�most�of�the�w���ork�in�a�one-����?semester���course�in���v�olv�es���some�rst-order�axioms�(groups,�rings,�etc.),�along����?with��lthe�notions�of�subgroup,�homomorphism,�isomorphism,�and�a�small����?amoun���t��5of�the�theory�of�natural�n�um�b�Gers.�The�latter�is�needed�for�the�con-����?cept�#)of�\nite�group"�and�the�concept�of�\order�of�a�group".�Num���b�Ger�theory����?is�[needed�only�(appro���ximately)�up�to�the�concept�of�\�a��divides��b�"�and�the����?factorization��of�a�n���um�b�Ger��in�to�a�pro�Gduct�of�primes.�One�pro�v�es,�for�example,����?the��structure�theorem�for�a�nite�ab�Gelian�group,�and�then�one�can�use�it�to����?pro���v�e���the�b�Geautiful�theorem�that�the�m���ultiplicativ�e���group�of�a�nite�eld�is����?cyclic.��`These�theorems�are�presen���tly�b�Gey�ond�the�reac�h�of�automated�deduc-����?tion�@'in�an���y�honest�sense,�although�of�course�one�could�prepare�a�sequence�of����?lemmas�UUin�suc���h�a�w�a�y�that�the�pro�Gof�could�ultimately�b�e�found.����NHo���w�ev�er,��there�is�a�natural�family�of�mathematical�theories�that�is�just����?sucien���t��for�expressing�most�undergraduate�mathematics.�Theories�of�this����?kind��*include�a�simple�theory�as�discussed�ab�Go���v�e��*(simple�axioms�ab�out�a�single����?kind�K*of�ob��8ject),�and�in�addition�parameters�for�subsets�(but�not�arbitrary����?quan���tication��[o�v�er�subsets),�v��q�ariables�for�natural�n�um�b�Gers�and�mathematical����?induction,��{and�functions�from�natural�n���um�b�Gers��{in�to�the�ob��8jects�of�the�simple����?theory��*�,�c!so�that�one�can�sp�Geak�ab�out�sequences�of�the�ob��8jects.�These�additional����?features,��,plus�denitions,�will�encompass�most�of�the�pro�Gofs�encoun���tered�in����?the�:]rst�semester�of�abstract�algebra.�If�w���e�add�inequalities�and�calculations����?to��8this�mix,�w���e�will�encompass�undergraduate�analysis,�complex�analysis,�and����?top�Gology�UUas�w���ell.����^��24������N�Of�i�course,�there�exist�branc���hes�of�mathematics�that�go�b�Gey�ond�this�kind����?of��mathematics�(e.g.�Galois�theory�or�algebraic�top�Gology).�W��*�e�prop�ose�to�not����?ev���en�O�think�ab�Gout�automated�deduction�in�these�areas�of�mathematics.�Deal-����?ing��with�the�c���hallenges�of�second-order�v��q�ariables�(without�quan�tication),��?�t��ff8�ϟ
L͍�	?��b��9y��Rob�A�ert�V��:�ero.�Although�it�is�to�o�tec��9hnical�to�discuss�here,�the�problem�is����	?�listed��as�an�op�A�en�problem�(whic��9h�previously�had�a�mo�del-theoretic�pro�of,�but����	?�no���rst-order�pro�A�of��q)�in�App�endix�3�of�[72]�(whic��9h�also�lists�other�c�hallenges�to����	?�theorem-pro��9ving�Tprograms).�The�solution�can�b�A�e�found�on�V��:�ero��q's�w�eb�page.���������UZ��-=�24�����	?��There��>is�an�ob��9vious�ob���jection�to�the�ab�A�o�v�e�taxonom�y:�F��:�unctions�can�b�A�e�reduced����	?�to�c�sets,�and�n��9um�b�A�ers�c�can�b�e�reduced�to�sets,�so�that�all�of�mathematics�can�b�e����	?�formalized��cin�set�theory��:�.�This�ob���jection�will�b�A�e�tak��9en�up�in�the�last�section�of�the����	?�pap�A�er.�����O��c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������23����)�ō���>;��?�denitions,�wcalculations,�incorp�Gorating�natural�n���um�b�ers,�wsequences,�and�in-����?duction,��=should�k���eep�researc�hers�busy�for�at�least�a�generation.�A�t�that�p�Goin�t����?computers�Nbshould�ha���v�e�Nbmore�or�less�the�capabilities�of�an�en���tering�Ph.�D.����?studen���t��Min�mathematics.�No�w,�in�2003,�they�are�at�appro�ximately�freshman����?lev���el.��I��do�not�mean�that�this�progress�is�inevitable|it�will�require�resources����?and�UUeort�that�ma���y�not�b�Ge�forthcoming.�But�it�is��p��}'ossible�.�� ۍ��?�9��S@Computer��Algebra��ۍ�?�\Computer��algebra",�while�a�common�and�descriptiv���e�term,�is�a�bit�mislead-����?ing���since�the�sub��8ject�encompasses�calculus�and�to�some�exten���t�com�binatorics,����?as�>w���ell�as�algebra.�Originally�computers�w�ere�view�ed�as�n�umerical�calcula-����?tors.�sIn�fact,�when�the�rst�c���hec�k�ers-pla�ying�sprogram�w���as�written�in�1948,����?there���w���as�no�prin�ter�at�the�IBM���researc�h�lab�that�could�prin�t�an�ything�but����?n���um�b�Gers,�so�the�output�of�the�c���hec�k�ers�pla�ying�program�had�to�b�Ge�co�ded�n���u-����?merically��*�.�_�But�b���y�the�late�1950s,�realization�w�as�spreading�that�there�w�as�suc�h����?a���thing�as�\sym���b�Golic�computation",�and�algebra�and�calculus�w�ere�among�the����?rst�+�areas�attac���k�ed.�+�Programs�for�elemen���tary�calculus�w�ere�so�Gon�written,�in-����?corp�Gorating�Frules�of�th���um�b�F(\heuristic"�is�the�ten-dollar�w���ord�for�a�rule�of����?th���um�b)�>�for�elemen���tary�in�tegration.�John�McCarth�y�in�v�en�ted�LISP��*�,�a�com-����?puter�lHlanguage�designed�for�programming�sym���b�Golic�tasks,�and�the�eld�of����?computer�Calgebra�b�Gegan�to�dev���elop�so�on�afterw���ards.�The�sub��8ject�had,�and����?still�q�has,�t���w�o�q�sides.�One�side�is�the�theoretical�dev���elopmen�t�q�of�algorithms�to����?solv���e���problems�of�sp�Gecic�classes.�The�other�side�is�the�ecien�t,�practical����?implemen���tation�UUof�those�algorithms�in�useful�programs.��_��NIn�[�the�sixties�and�sev���en�ties,�[�there�w���ere�sev�eral�imp�Gortan�t�theoretical�dev�el-����?opmen���ts�v�in�the�sub��8ject.�Although�w�e�cannot�aord�a�thorough�in�tro�Gduction����?to���computer�algebra,�w���e�will�men�tion�three�dev�elopmen�ts.�Consider�the�fol-����?lo���wing���Mathematica�input:���F��;�actor���!�[�x���^��119��
�[��H�1].�Mathematica�resp�Gonds�instan�tly����?with������^(��1�8�+��x�)(1�+��x��+��x�����2���S�+��x�����3���+��x�����4���+��x�����5���+��x�����6��|s�)(1�+��x��+��x�����2���+��x�����3���+��x�����4���+��x�����5���������^�+�x�����6���S�+�8�x�����7���+��x�����8���+��x�����9���+��x�����10��
�Ʋ+��x�����11���+��x�����12���+��x�����13���+��x�����14���+��x�����15���+��x�����16��x�)��������^(1�8���x��+��x�����7���S���x�����8���+��x�����14��
�Ƹ��x�����15���+��x�����17�����x�����18���+��x�����21�����x�����22���+��x�����24�����x�����25���������^�+�x�����28��
�Ƹ�8�x�����29���+��x�����31�����x�����32���+��x�����34�����x�����36���+��x�����38�����x�����39���+��x�����41�����x�����43���+��x�����45���������^��x�����46��
�Ʋ+�8�x�����48�����x�����50���+��x�����51�����x�����53���+��x�����55�����x�����57���+��x�����58�����x�����60���+��x�����62�����x�����64���������^�+�x�����65��
�Ƹ�8�x�����67���+��x�����68�����x�����71���+��x�����72�����x�����74���+��x�����75�����x�����78���+��x�����79�����x�����81���+��x�����82���������^��x�����88��
�Ʋ+�8�x�����89�����x�����95���+��x�����96��x�)������?This���w���as�not�done�b�y�trial�and�error.�It�uses�algorithms�for�p�Golynomial�fac-����?torization����^��25���|�that���rst�factor�the�p�Golynomial�mo�d��p��for�v��q�arious�small�primes��?�]��ff8�ϟ
L͍������UZ��-=�25�����	?��F��:�or��-readers�unfamiliar�with�mo�A�d��p�,�this�means�that�n��9um�b�ers��-are�alw��9a�ys��-replaced����	?�with��6their�remainders�after�division�b��9y��p�.�F��:�or�example,�3�times�5�is�1�mo�A�d�7,�����`I�c�ō���>;����?�24���`-	Mic��9hael�TBeeson����)�ō���>;��?�p�,�эand�then�put�the�results�together�clev���erly��*�,�using�a�19th-cen�tury�theorem����?kno���wn��Pas�Hensel's�lemma.�The�factorization�mo�Gd��p��can�b�e�done�quic���kly��*�,�using����?an�Շalgorithm�disco���v�ered�Շb�y�Berlek��q�amp�in�1967.�The�application�of�Hensel's����?lemma�0to�this�problem�w���as�disco�v�ered�b�y�Zassenhaus�in�1969.�F��*�ull�historical����?and�UUmathematical�details�can�b�Ge�found�in�[57]�and�[63].����NOur��Ksecond�example�concerns�the�in���tegration�of�elemen�tary�functions.�An����?�elementary�;8function�2�is�one�that�y���ou�migh�t�encoun�ter�in�freshman�calculus:����?it�Vis�dened�using�m���ultiplication,�addition,�subtraction,�division,�trig�func-����?tions,�*'exp�Gonen���ts,�and�logarithms.�Muc�h�eort�in�freshman�calculus�go�Ges�in�to����?rules�D�and�metho�Gds�for�computing�elemen���tary�in�tegrals�of�elemen�tary�func-����?tions.�^>Ho���w�ev�er,�not�ev�ery�elemen�tary�function�has�an�elemen�tary�in�tegral.����?F��*�or�ADexample,���������u

cmex10�R�����e���^��	0e�rcmmi7�x����r���Zcmr5�2���
���dx��cannot�b�Ge�expressed�in�elemen���tary�form.�Risc�h�[95,96]����?disco���v�ered�Ipin�1969�that�the�trial-and-error�metho�Gds�y���ou�ma�y�ha�v�e�studied����?in�x_freshman�calculus,�suc���h�as�in�tegration�b�y�substitution�and�in�tegration�b�y����?parts,��Zcan�b�Ge�replaced�b���y�a�single,�systematic�pro�cedure,�that�alw���a�ys��Zw�orks�if����?the�#�in���tegral�has��any��elemen�tary�answ�er.�A�#�complete�exp�Gosition�of�the�theory����?is�UUin�[21].����NOur�Zthird�example�concerns�sets�of�sim���ultaneous�p�Golynomial�equations.����?Sa���y��*�,�UUfor�example,�that�y�ou�wish�to�solv�e�the�equations��q덍��a��z��w�+�8�x�����4���S���2�x��+�1������EK=������i0��������s�y��[ٟ���2��,�+�8�x�����2���S���1������EK=������i0��������X�x�����5���S��8�6�x�����3���+��x�����2�����1������EK=�����i0�����?If�u6y���ou�ask�Mathematica�to�solv�e�this�set�of�three�equations�in�three�unkno�wns,����?it�l�answ���ers�(immediately)�with�a�list�of�the�ten�solutions.�Since�the�solutions����?do�Ԅnot�ha���v�e�Ԅexpressions�in�terms�of�square�ro�Gots,�they�ha���v�e�Ԅto�b�e�giv���en�in�the����?form���of�algebraic�n���um�b�Gers.���F��*�or�example,�the�rst�one�is��x���=���	z;���y�"�=���i���`�1�;�z�7��=����?��1�
P+�2��	z�,���where����x�is�the�smallest�ro�Got�of���1�
P+������^��2���=���6�����^��3���=�+������^��5��	`�=��00.���This����?problem��lhas�b�Geen�solv���ed�b�y�constructing�what�is�kno�wn�as�a�\Gr����obner�basis"����?of�the�ideal�generated�b���y�the�three�p�Golynomials�in�the�original�problem.�It����?tak���es�Seto�Go�m�uc�h�space,�and�demands�to�Go�m�uc�h�mathematical�bac�kground,�to����?explain�n�this�more�fully;�see�[106],�Chapter�8�for�explanations.�(This�example�is����?Exercise�ߒ4,�p.�201).�Although�metho�Gds�(due�to�Kronec���k�er)�ߒw�ere�kno�wn�in�the����?nineteen���th��Ncen�tury�that�in�principle�could�solv�e�suc�h�problems,�the�concept����?of��3a�Gr����obner�basis�and�the�algorithm�for�nding�one,�kno���wn�as�\Buc�h�b�Gerger's����?algorithm",�6�ha���v�e�pla�y�ed�an�indisp�Gensable�role�in�the�dev�elopmen�t�of�mo�Gdern����?computer�i�algebra.�These�results�w���ere�in�Buc�h�b�Gerger's�Ph.�D.�thesis�in�1965.����?Th���us��Cthe�p�Gerio�d��C1965-70�sa�w�the�theoretical�foundations�of�computer�algebra����?laid.����NIt��to�Gok�some�time�for�implemen���tation�to�catc�h�up�with�theory��*�,�but�as�the����?t���w�en�t�y-rst�}�cen�tury�op�Gened,�there�w�ere�sev�eral�w�ell-kno�wn,�widely�a�v��q�ailable����?programs�:�con���taining�implemen�tations�of�these�imp�Gortan�t�algorithms,�as�w�ell��?��#�ff8�ϟ
L͍�	?��b�A�ecause�ig15�has�remainder�1�after�division�b��9y�7.�So�(�x�FE�+�3)(�x��+�5)��=��x���-=�2��p�+�FE�x��+�1����	?�mo�A�d�T7.�����p�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������25����)�ō���>;��?�as�z�man���y�others.�Sym�b�Golic�mathematics�up�to�and�including�freshman�calculus����?can��Eth���us�b�Ge�regarded�as�completely�mec�hanized�at�this�p�Goin�t.�While�one����?cannot��csa���y�that�the�eld�is�complete{ev�ery�y�ear�there�is�a�large�in�ternational����?conference���dev���oted�to�the�sub��8ject�and�man�y�more�sp�Gecialized�conferences{on����?the���whole�the�mec���hanization�of�computation�has�progressed�m�uc�h�further����?than�UUthe�mec���hanization�of�pro�Gof.��P%��NIn���addition�to�the�w���ell-kno�wn���general-purp�Gose�sym���b�olic�computation�pro-����?grams�u�suc���h�as�Maple,�Mathematica,�and�Macsyma,�there�are�also�a�n�um�b�Ger����?of�e�sp�Gecial-purp�ose�programs�dev���oted�to�particular�branc�hes�of�mathematics.����?These�&'are�programs�suc���h�as�MA�GMA,�P��*�ARI-GP�&(algebraic�n�um�b�Ger�theory),����?SnapP���ea�o�(top�Gology),�GAP�o�(group�theory),�Surface�Ev�olv�er�(dieren�tial�geom-����?etry),�UUetc.�These�are�used�b���y�sp�Gecialists�in�those�elds.��P%��NWhat�1Ois�the�place�of�computer�algebra�in�the�mec���hanization�of�mathe-����?matics?��,Ob���viously�there�are�some�parts�of�mathematics�that�consist�mainly����?of�s�computations.�The�fact�is�that�this�part�of�mathematics�includes�high-����?sc���ho�Gol�u�mathematics�and�rst-y�ear�calculus�as�it�is�usually�taugh�t,�so�that����?p�Geople���who�do�not�study�mathematics�b�ey���ond�that�p�oin���t�ha�v�e�the�(mis)-����?impression�L�that�mathematics�consists�of�calculations,�and�they�imagine�that����?adv��q�anced���mathematics�consists�of�y���et�more�complicated�calculations.�That�is����?not���true.�Beginning�with�the�course�after�calculus,�mathematics�relies�hea���vily����?on���pro�Gofs.�Some�of�the�pro�ofs�con���tain�some�steps�that�can�b�e�justied�b���y�cal-����?culation,��'but�more�emphasis�is�placed�on�precisely�dened,�abstract�concepts,����?and��dthe�study�of�what�prop�Gerties�follo���w�from�more�fundamen�tal�prop�Gerties����?b���y�UUlogical�implication.��!�����?�10��ZDecision��Pro�`cedures�in�Algebra�and�Geometry������?�The��u\rst�lo�Gophole"�allo���ws�the�p�ossibilit���y�that��some��branc�hes�of�mathematics����?can��Jb�Ge�mec���hanized.�An�algorithm�whic�h�can�answ�er�an�y�y�es-no�question�in����?a���giv���en�class�of�mathematical�questions�is�called�a�\decision�pro�Gcedure"�for����?those�
=questions.�W��*�e�will�giv���e�a�simple�example�to�illustrate�the�concept.�Y�ou����?ma���y�S-recall�studying�trigonometry��*�.�In�that�sub��8ject,�one�considers�\trigono-����?metric�3iden���tities"�suc�h�as��cos��h�(2�x�)��=��cos���*���^��2��Qµx�������sin���
߻���*�2��ֵx�.�3The�iden�tities�considered����?in�@�trigonometry�alw���a�ys�@�ha�v�e�only�linear�functions�in�the�argumen�ts�of�the����?trig�Lfunctions;�for�example,�they�nev���er�consider��sin��Zj(�x���^��2��|s�),�although��sin��(2�x��ϲ+�3)����?w���ould�k~b�Ge�allo�w�ed.�Moreo�v�er,�the�co�Gecien�ts�of�those�linear�functions�are�al-����?w���a�ys��Iin�tegers,�or�can�b�Ge�made�so�b�y�a�simple�c�hange�of�v��q�ariable.�The�question����?is,��Tgiv���en�suc�h�an�equation,�determine�whether�or�not�it�holds�for�all�v��q�alues����?of�W�x��(except�p�Gossibly�at�the�p�oin���ts�where�one�side�or�the�other�is�not�de-����?ned,�l�e.g.�b�Gecause�a�denominator�is�zero.)�Y��*�ou�ma���y�b�e�surprised�to�learn����?that��qthere�is�a�decision�metho�Gd�for�this�class,�whic���h�w�e�no�w�giv�e.�First,�use����?kno���wn���iden�tities�to�express�ev�erything�in�terms�of��sin���and��cos���.�If�necessary��*�,����?mak���e��a�c�hange�of�v��q�ariable�so�that�the�linear�functions�in�the�argumen�ts�of�����?sin��O��and��_�cos��"�ha���v�e�_�in�teger�co�Gecien�ts.�Ev�en�though�ev�erything�is�in�no�w�in�������c�ō���>;����?�26���`-	Mic��9hael�TBeeson����)�ō���>;��?�terms�M�of��sin���and��cos���Z,�there�could�still�b�Ge�dieren���t�argumen�ts,�for�example�����?sin��KG(2�x�)�c����sin��
)�x�.�EIf�so,�w���e�next�use�the�iden�tities�for��sin���5(�x�c�+��y�[ٲ)�Eand��cos����(�x�c�+��y��)����?to��fexpress�ev���erything�in�terms�of��sin���,�x��and��cos��Ν�x�.�The�equation�is�no�w�a�ratio-����?nal�Y�function�of��sin��K\�x��and��cos��g͵x�.�No���w�for�the�k�ey�step:�Mak�e�the�\W��*�eierstrass����?substitution"�֛�t����=��tan��W(�x=�2).�Then��sin���a�x��and��cos���ҵx��b�Gecome�rational�functions����?of���t�.�Sp�Gecically��*�,�w���e�ha�v�e��sin��ͫ�x��\�=�2�t=�(1���+��t���^��2��|s�)���and��cos����x��\�=�(1������t���^��2���)�=�(1�+��t���^��2���).����?After��this�substitution,�the�equation�b�Gecomes�a�p�olynomial�iden���tit�y��in�one����?v��q�ariable,���and�w���e�just�ha�v�e�to�simplify�it�to�\standard�form"�and�see�if�the����?t���w�o�2ysides�are�iden���tical�or�not.�All�that�suering�that�y�ou�w�en�t�through�in����?trigonometry�UUclass!�and�a�computer�can�do�the�job�in�an�instan���t.����NThe�Pquestion�is,�then,�exactly�where�the�b�Gorderline�b�et���w�een�Pmec�haniz-����?able�r�theories�and�non-mec���hanizable�theories�lies.�It�is�somewhere�b�Get�w�een����?trig��Biden���tities�and�n�um�b�Ger�theory��*�,�since�b�y�T��*�uring�and�Ch�urc�h's�results,�w�e����?cannot��giv���e�a�decision�pro�Gcedure�for�n�um�b�Ger�theory��*�.�The�b�orderline�is�in����?some��sense�not�v���ery�far�b�Gey�ond�trig�iden�tities,�since�a�result�of�Ric�hardson����?[85]���sho���ws�that�there�is�no�algorithm�that�can�decide�the�truth�of�iden�tities����?in���v�olving��Zp�Golynomials,�trig�functions,�logarithms,�and�exp�onen���tials�(with�the����?constan���t��2����allo�w�ed,�and�the�restriction�that�the�argumen�ts�of�trig�functions����?b�Ge���linear�remo���v�ed).����^��26���u��Nev�ertheless,���there�are�man���y�examples�of�decision�pro-����?cedures��for�signican���t�b�Go�dies��of�mathematics.�P�erhaps�the�most�striking�is����?one�e�rst�explored�b���y�Alfred�T��*�arski�(1902-1983).�The�branc�h�of�mathematics����?in���question�is,�roughly�sp�Geaking,�elemen���tary�algebra.�It�is�really�more�than����?elemen���tary�n�algebra,�b�Gecause�\for�all"�and�\there�exists"�are�also�allo�w�ed,�so����?suc���h�UUquestions�as�the�follo�wing�are�legal:��ҍ����E�������P�Do�Ges�UUthe�equation��x���^��3���S��8�x���^��2���+�1��=�0�UUha���v�e�a�solution�b�Get�w�een�0�and�1?���č����E�������P�F��*�or�T�whic���h�v��q�alues�of��a��and��b��do�Ges�the�equation��x���^��4�����7��ax���^��3���+��b�T��tak���e�on�only����Pp�Gositiv���e�UUv��q�alues�as��x��v�aries?��K��?The�#�rst�question�has�an�implicit�\there�exists�an��x�",�and�the�second�has�an����?implicit�UU\for�all��x�".�W��*�e�will�call�this�part�of�mathematics�\T�arski�algebra."����N\F��*�or��;all"�and�\there�exists"�are�called�\quan���tiers".�A��form�ula�without����?quan���tiers�8�is�called�\quan�tier-free".�F��*�or�example,�`�x���^��2��{��+���2��=��y�[ٲ'�8�is�quan�tier-����?free.�UUA�quan���tier-free�form�ula�migh�t�ha�v�e�the�form��ҍ����f���(�x����1��|s�;����:�:�:����;���x����n��q~�)��=�0�UU&��g�[ٲ(�x����1���;����:�:�:����;���x����n��q~�)�����0�;����?�where��
�f�and��g�6�are�p�Golynomials.�More�generally��*�,�y���ou�migh�t�ha�v�e�sev�eral�in-����?equalities�,instead�of�just�one.�Using�simple�iden���tities,�one�can�sho�w�that�an�y����?quan���tier-free�Oform�ula�is�equiv��q�alen�t�to�one�in�the�form�indicated.�That�is,����?if��Zsuc���h�form�ulas�are�com�bined�with�\not",�\and",�or�\or",�the�result�can�b�Ge����?equiv��q�alen���tly��qexpressed�in�the�standard�form�men�tioned.�T��*�arski's�idea�is�called��?����ff8�ϟ
L͍������UZ��-=�26�����	?��The��Eexact�b�A�orderline�for�classes�of�iden��9tities�still�is�not�kno�wn�v�ery�accurately��:�.����	?�F��:�or��Kexample,�what�if�w��9e�k�eep�the�restriction�that�the�argumen�ts�of�trig�functions����	?�should�-b�A�e�linear�with�in��9teger�co�ecien��9ts,�but�w�e�allo�w�logarithms�and�exp�A�onen-����	?�tials?������o�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������27����)�ō���>;�������?�LqC������
��Fig.��h2.��ʻWhat�Tis�T��:�arski�algebra?����?The���tec��9hnical�name�of�this�branc�h�of�mathematics�is�the�theory�of��r��e�al-close�d��^elds�.����?The��language�for�this�branc��9h�of�mathematics�has�sym�b�A�ols�for�t�w�o�op�A�erations�+����?and�>��,�the�in��9v�erse�>op�A�erations���x��and��x���-=�q�%cmsy6��1��	�,�the�additiv��9e�and�m�ultiplicativ�e�iden�tit�y����?elemen��9ts���0�and�1,�the�ordering�relation��<�,�and�the�equalit�y�relation�=.�The�axioms����?include���the�usual�la��9ws�for�+�and���,�and�axioms�relating��<��to�the�op�A�erations�+�and���.����?Dening�fv0��ƽ<�x��as��P�H�(�x�)�(�P��Y�for�\p�A�ositiv��9e"),�those�axioms�sa�y�that�the�sum�of�p�A�ositiv�e����?elemen��9ts�kis�p�A�ositiv�e�and�the�pro�A�duct�of�p�ositiv��9e�elemen�ts�is�p�A�ositiv�e.�These�are�the����?axioms�t�of��or��der�e�d���elds�.�t�The�axioms�for�real-closed�elds�sp�A�ecify�in�addition�that�all����?p�A�ositiv��9e��Relemen�ts�ha�v�e�a�square�ro�A�ot,�and�all�p�olynomials�of�o�dd�degree�ha��9v�e�a�ro�A�ot.����?One��will,�of�course,�need�innitely�man��9y�axioms�to�express�this�without�men�tioning����?the�A�concept�of�\natural�n��9um�b�A�er",�A�one�axiom�for�eac��9h�o�dd�degree.�The�classical����?theory�R^of�real-closed�elds�is�dev��9elop�A�ed�in�most�algebra�textb�o�oks,�for�example�in����?[65],�Tpp.�273���.����?T��:�arski�z�algebra�escap�A�es�the�negativ��9e�results�of�Ch�urc�h�and�G��`odel�b�A�ecause�it�do�es�not����?ha��9v�e��[v��|rariables�for�natural�n��9um�b�A�ers.��[The�v�ariables�range�o��9v�er��[\real�n��9um�b�A�ers"{these����?are�C�the�n��9um�b�A�ers�C�that�corresp�ond�to�p�oin��9ts�on�a�line�and�are�used�for�co�ordinates.����?Ev��9en�'0though�the�v��|rariables�of�T��:�arski�algebra�are�mean�t�to�stand�for�suc�h�n�um�b�A�ers,����?not�KXall�individual�n��9um�b�A�ers�KXcan�b�e�dened�in�T��:�arski�algebra.�In�this�language,�one����?cannot��8directly�write�in��9tegers�in�decimal�notation�suc�h�as�3.�Instead�of�3,�one����?ocially���has�to�write�1�+�1�+�1.�Aside�from�the�incon��9v�enience,���one�can�in�eect����?write�^"an��9y�rational�n�um�b�A�er;�for�example�2�=�3�is�(1�>�+�1)(1�+�1�+�1)���-=��1��	�.�^"But�one�do�es����?not,�Tfor�example,�ha��9v�e�Ta�name�for���R��.�����?�LqC����#�ō�?�elimination�k�of�quantiers�.�)�He�sho���w�ed�)�in�[97]�that�ev���ery�form�ula�in�T��*�arski�al-����?gebra��is�equiv��q�alen���t�to�one�without�an�y�quan�tiers.�F��*�or�example,�the�question����?whether��E�x���^��2���5�+�:µbx��+��c���=�0�has�a�solution��x��with�0����x��app�Gears�to�in���v�olv�e�\there����?exists�;0an��x�",�but�from�the�quadratic�form���ula�w�e�nd�that�the�answ�er�can�b�Ge����?expressed�~Ob���y�a�condition�in�v�olving�only��b��and��c�,�namely��*�,��b���^��2��J���ײ4�c�����0�~Oand�either����?�b�#ָ��0���or��c�#ָ��0.���The�quan���tier�\there�exists��x�"�has�b�Geen�eliminated.�Sev�eral����?classical�R�results�of�algebra�ha���v�e�R�a�similar�
a���v�or.�R�F��*�or�example,�Sturm's�theorem����?from���the�1830s�[65],�p.�276,�coun���ts�the�n�um�b�Ger�of�ro�ots�of�a�p�olynomial�in�an����?in���terv��q�al�}�in�terms�of�the�alternations�of�signs�in�the�co�Gecien�ts.�Another�classi-����?cal���result�is�the�existence�of�the��r��}'esultant�:�If�w���e�are�giv�en�p�Golynomials��f���(�a;���x�)����?and�Th�g�[ٲ(�a;���x�),�w���e�can�compute�another�p�Golynomial��R�Dz(�a;�b�)�called�the�resultan���t����?of�7ٵf�Kh�and��g�[ٲ,�suc���h�that��R�Dz(�a;���b�)�@�=�0�7�if�and�only�if�a�common�solution��x��can����?b�Ge�dfound�for�the�equations��f���(�a;���x�)�
1=�0�dand��g�[ٲ(�a;�x�)�
1=�0.�dAgain�the�quan���tier����?\there�{eexists��x�"�has�b�Geen�eliminated.�T��*�arski�sho���w�ed�{ethat�algebraic�metho�ds����?can�A�alw���a�ys�b�Ge�applied�to�eliminate�\there�exists"�from�algebraic�form�ulas,����?ev���en��ones�in�v�olving�inequalities.�The�elimination�of�one�quan�tier�dep�Gends����?essen���tially��on�the�fact�that�a�p�Golynomial�has�only�nitely�man�y�ro�Gots,�and�w�e����?can�"Scompute�the�n���um�b�Ger,�"Sthe�maxim���um�size,�and�some�information�ab�out����?the�)�lo�Gcation�of�the�ro�ots�from�the�co�ecien���ts�of�the�p�olynomial.�Applying����?this�]�pro�Gcedure�again�and�again,�w���e�can�strip�o�one�quan�tier�after�another������X�c�ō���>;����?�28���`-	Mic��9hael�TBeeson����)�ō���>;��?�(from���the�inside�out),�eliminating�all�the�quan���tiers�in�a�form�ula�with�nested����?quan���tiers.��W��*�e�need�only�deal�with�\there�exists"�b�Gecause�\for�all"�can�b�e����?expressed��das�\not�there�exists��x��not"�.�T��*�arski's�pro�Gcedure�is�a�decision�pro-����?cedure���for�T��*�arski�algebra,�b�Gecause�if�w���e�start�with�a�form�ula�that�has�only����?quan���tied��Yv��q�ariables�(so�it�mak�es�an�assertion�that�should�b�Ge�true�or�false),����?after�PVw���e�apply�the�pro�Gcedure�w�e�get�a�purely�n�umerical�form�ula�in�v�olving����?equations��_and�inequalities�of�rational�n���um�b�Gers,��_and�w���e�can�simply�compute����?whether�UUit�is�true�or�false.��ߊ��NDescartes���sho���w�ed�sev�eral�cen�turies�earlier�that�geometry�could�b�Ge�reduced����?to�Ӓalgebra,�b���y�the�device�of�co�Gordinates.�This�reduction,�kno�wn�as�analytic����?geometry��*�,��qcoupled�with�T�arski's�reduction�of�algebra�with�quan���tiers�to�com-����?putation,�ɲyields�a�reduction�of�geometry�(with�quan���tiers)�to�computation.����?In�Z�more�tec���hnical�w�ords:�a�decision�pro�Gcedure�for�Euclidean�geometry��*�.�Th�us����?Hilb�Gert's���program,�to�reduce�mathematics�to�computation,�migh���t�seem�to����?b�Ge��:ac���hiev�ed�for�the�mathematics�of�the�classical�era,�algebra�and�geome-����?try��*�.��T�arski's�studen���t�Szmielew�made�it�w�ork�for�non-Euclidean�(h�yp�Gerb�olic)����?geometry���to�Go�[27].�Since�the�W��*�eierstrass�substitution�reduces�trigonometry�to����?algebra,�#
a�decision�metho�Gd�for�real-closed�elds�also�applies�to�trigonometry��*�,����?as�UUlong�as�the�argumen���ts�of�the�trig�functions�are�linear.��ߊ��NT��*�arski's�Mresult�is�regarded�as�v���ery�imp�Gortan�t.�Hundreds�of�researc�hers�ha�v�e����?pursued,�J�and�con���tin�ue�J�to�pursue,�the�lines�of�in���v�estigation�J�he�op�Gened.�There����?are�z\t���w�o�reasons�for�that:�First,�his�results�con�trast�sharply�with�Ch�urc�h's����?and���G����odel's,�and�sho���w�that�the�classical�areas�of�algebra�and�geometry�are����?not��sub��8ject�to�those�limiting�theorems.�Second,�there�are�plen���t�y��of�op�Gen����?and�P�in���teresting�problems�that�can�b�Ge�form�ulated�in�the�theory�of�real-closed����?elds,�!&and�this�has�raised�the�hop�Ge�that�decision�pro�cedures�implemen���ted�on����?a�vzcomputer�migh���t�one�da�y�routinely�answ�er�op�Gen�questions.�Our�purp�ose�in����?this�UUsection�is�to�in���v�estigate�UUthis�p�Gossibilit���y��*�.��ߊ��NFirst,�6�let�us�giv���e�an�example�of�an�op�Gen�problem�one�can�form�ulate�in����?T��*�arski��algebra.�Here�is�an�example�from�the�theory�of�sphere-pac���king.�This����?example,�'Fand�man���y�others,�can�b�Ge�found�in�[32].�The�\kissing�problem"�asks����?ho���w�m`man�y��n�-dimensional�spheres�can�b�Ge�pac�k�ed�disjoin�tly�so�that�they�eac�h����?touc���h�\�the�unit�sphere�cen�tered�at�origin.�F��*�or��n���=�2�\�the�answ�er�is�six�(2-spheres����?are�1]circles).�F��*�or��n���=�3�1]the�answ���er�is�12.�F�or��n���=�4�1]the�answ���er�is�either�24,�or����?25,��Obut�nob�Go�dy��Okno���ws�whic�h!�The�problem�can�b�Ge�form�ulated�in�the�theory����?of�� real-closed�elds,�using�100�v��q�ariables�for�the�co�Gordinates�of�the�cen���ters�of����?the�4uspheres.�W��*�e�simply�ha���v�e�4uto�sa���y�that�eac�h�cen�ter�is�at�distance�2�from����?the�D�origin�and�that�eac���h�of�the�300�pairs�of�p�Goin�ts�are�at�least�2�units�apart.����?Explicitly��*�,���w���e�wish�to�kno�w�if�there�exist��x����1��|s�;����:�:�:����;���x����25��x�;�y����1���;��:�:�:����;�y����25��x�;�z����1���;��:�:�:����;�z����25��x�,����?and�~��w����1��|s�;����:�:�:����;���w����25����suc���h�that��x���^���2��;Z��i���{��+��N�y���^���[ٱ2��;Z��i���ך�+��z���^���p��2��;Z��i����X�+��w���^���D�2��;Z��i���	x��=��,4�~�and�(�x����i��S����N�x����j��6��)���^��2���+�(�y����i��S������?�y����j��6��)���^��2��f��+��M(�z����i��>����z����j���)���^��2��f��+�(�w����i��>����w����j���)���^��2��C�����4�.for��i��6�=��j����.�All�w���e�ha�v�e�to�do�is�run�T��*�arski's����?algorithm�UUon�that�form���ula,�and�the�op�Gen�problem�will�b�e�answ���ered.����^��27����?��W�ff8�ϟ
L͍������UZ��-=�27�����	?��Another�<�in��9teresting�sphere-pac�king�problem�w�as�op�A�en�for�cen�turies,�un�til�it�w�as����	?�solv��9ed��in�1998.�Namely��:�,�what�is�the�densest�pac�king�of�spheres�in�to�a�large�cub�A�e������L�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������29����)�ō���>;��N�W��*�ell�
=then,�wh���y�is�this�problem�still�op�Gen?�The�suspicion�ma�y�b�Ge�da�wning����?on��
y���ou�that�it�isn't�so�easy�to�run�this�pro�Gcedure�on�a�form�ula�with�100����?quan���tied���v��q�ariables!�In�the�half-cen�tury�since�T��*�arski's�w�ork,�researc�hers�ha�v�e����?found��more�ecien���t�algorithms�for�quan�tier�elimination,�but�on�the�other����?hand,��,they�ha���v�e��,also�pro���v�ed��,theorems�sho���wing�that��any��algorithm�for�quan�ti-����?er�R�elimination�m���ust�necessarily�run�slo�wly�when�large�n�um�b�Gers�of�v��q�ariables����?are��Min���v�olv�ed.�These�lines�of�researc�h�no�w�almost�meet:�the�b�Gest�algorithms����?almost�o�ac���hiev�e�the�theoretical�limits.�Ho�w�ev�er,�it�seems�that�the�edge�of�ca-����?pabilit���y� of�algorithms�is�close�to�the�edge�of�h�uman�capabilit�y�as�w�ell,�so�the����?p�Gossibilit���y�@2that�decision�pro�cedures�migh���t�settle�an�op�en�question�cannot�b�e����?denitiv���ely�UUrefuted.�W��*�e�therefore�review�the�situation�carefully�.��Z!��NFirst��"w���e�review�the�w�orst-case�analyses�that�sho�w�quan�tier�elimination����?m���ust��*run�slo�wly��*�.�Fisc�her�and�Rabin�sho�w�ed�[42]�that�an�y�algorithm�for�quan-����?tier��elimination�in�real-closed�elds�will�necessarily�require�exp�Gonen���tial�time����?for��w���orst-case�input�form�ulas;�that�is,�time�of�the�order�of�2���^��dn�����where��n��is�the����?length�}�of�the�input�form���ula,�and��d��is�a�xed�constan�t.�This�is�true�ev�en�for�for-����?m���ulas�W in�v�olving�only�addition�(not�m�ultiplication).�Later�[104,36]�a�stronger����?lo���w�er���b�Gound�w���as�pro�v�ed:�sometimes�quan�tier�elimination�will�require�time����?(and��Nspace)�of�the�order�2���^��2����r�O
�\cmmi5�n����-�(double�exp�Gonen���tial).�(�See�[84]�for�a�surv�ey�of����?results���on�the�complexit���y�of�this�problem.)�T��*�aking��n�2��=�64���w�e�get�a�n�um�b�Ger����?with�hAmore�than�10���^��18���'�decimal�digits.�No�w���onder�the�kissing�n�um�b�Ger�problem����?is�UUstill�op�Gen.��Z!��NT��*�arski's�+�original�algorithm,�whic���h�w�as�nev�er�implemen�ted,�w�as�in�prin-����?ciple�`�m���uc�h�slo�w�er�ev�en�than�double�exp�Gonen�tial.�T��*�arski's�metho�Gd�eliminates����?one�s�quan���tier�at�a�time,�and�the�form�ula�expands�in�length�b�y�a�double�ex-����?p�Gonen���tial��eac�h�time,�so�the�running�time�cannot�b�Ge�b�ounded�b���y�an�y�to�w�er�of����?exp�Gonen���ts.��
Fisc�her�and�Rabin's�result�w�as�obtained�in�the�fall�of�1972,�but����?not��published�un���til�1974.�In�the�in�terim,�not�kno�wing�that�ecien�t�quan�tier����?elimination�{[is�imp�Gossible,�George�Collins�in���v�en�ted�{[an�impro���v�ed�{[quan�tier-����?elimination��Pmetho�Gd�kno���wn�as��cylindric��algebr��}'aic�de�c�omp�osition��P�(CAD)��@[31].����?Actually��*�,��according�to�the�preface�of�[28],�Collins�had�b�Geen�w���orking�on�quan-����?tier�g�elimination�since�1955,�but�the�1973�w���ork�generalized�his�metho�Gd�to��n����?�v��q�ariables�|�and�hence�made�it�a�general�quan���tier�elimination�metho�Gd.�Collins's����?metho�Gd�'Nruns�in�double�exp�onen���tial�time,�m�uc�h�b�Getter�than�T��*�arski's�metho�d,����?and�e�almost�b�Gest-p�ossible�e�[36].�W��*�e�knew�from�Fisc���her�and�Rabin's�lo�w�er�b�Gound����?that��xthere�w���as�no�hop�Ge�of�a�really�ecien�t�quan�tier�elimination�algorithm,����?but�Q�the�CAD�metho�Gd�is�m���uc�h�Q�faster�than�T��*�arski's�or�Cohen's�metho�ds.�The����?w���orst��gcase,�when�the�algorithm�tak�es�time�2���^��2����r�n���	c߲,�arises�only�when�there�are�lots����?of��v��q�ariables.�The�algorithm�is�double�exp�Gonen���tial�in�the�n�um�b�Ger�of�v��q�ariables,��?����ff8�ϟ
L͍�	?��in���3-space?�Kepler�conjectured�that�it�is�the�usual�pac��9king�used�b�y�gro�A�cers�for����	?�stac��9king���oranges,�but�this�w�as�dicult�to�pro�v�e.�It�can,�ho�w�ev�er,�easily�b�A�e�for-����	?�m��9ulated�M�in�the�theory�at�hand,�so�in�principle,�\all�w�e�ha�v�e�to�do"�is�quan�tier����	?�elimination.������t�c�ō���>;����?�30���`-	Mic��9hael�TBeeson����)�ō���>;��?�but��for�a�xed�n���um�b�Ger��of�v��q�ariables,�the�time�increases�only�as�some�p�o���w�er��of����?the�UUlength�of�the�input.����^��28����]��N�`Mo�Gore's�ycla���w"�is�the�observ��q�ation,�made�in�1965�b�y�Gordon�Mo�Gore,�co-����?founder�"qof�In���tel,�that�data�densit�y�in�computers�(bits�p�Ger�square�cen�timeter)����?has�.kb�Geen�gro���wing�exp�onen���tially��*�,�doubling�ev�ery�12-18�mon�ths,�ev�er�since�the����?in���tegrated���circuit�w�as�in�v�en�ted�in�1962.�P�erhaps�incorrectly��*�,�man�y�p�Geople�also����?use��-\Mo�Gore's�la���w"�to�refer�to�the�exp�onen���tial�increase�in�computer�sp�eed.����^��29������?�One��Yshould�clearly�understand�that�Mo�Gore's�la���w�cannot�help�us�m�uc�h�with����?an�s�algorithm�whose�running�time�is�double�exp�Gonen���tial.�If�the�running�time����?is��(2���^��2����r�n�����and�w���e�w�an�t�to�increase��n��b�y�one,�w�e�need�a�computer�that�runs�2���^��2����r�n������?�times��hfaster,�as�a�short�calculation�will�sho���w�y�ou:�tak�e�the�ratio�of�the�new����?running�AWtime,�2���^��2����r�n�+1���,�to�the�old�running�time�2���^��2����r�n���	c߲.�Y��*�ou�will�get�2���^��2����r�n����6�when�y���ou����?simplify�sthat�ratio.�It�tak���es�2���^��n���~�Mo�Gore's�la�w�doubling�p�Gerio�ds�sjust�to�increase��n����?�b���y�{one.�The�imp�Gort�of�the�double-exp�onen���tial�running�time�theorems�ab�out����?quan���tier�/elimination�is�therefore�almost�as�grim�as�the�imp�Gort�of�G����odel's����?theorem.�14It�seems�that�Fisc���her,�Rabin,�W��*�eispfenning,�and�Da�v�enp�Gort�ha�v�e����?destro���y�ed�UUT��*�arski's�dream�as�thoroughly�as�G����odel�destro���y�ed�UUHilb�Gert's.��]��NBut�Itp�Geople�nev���er�giv�e�up!�Ma�yb�Ge�there�is�an�escap�e�route.�It�w���as�disco�v-����?ered��in�1992�b���y�Grigorev�[46]�that�if�w�e�restrict�atten�tion�to�form�ulas�that����?only���ha���v�e�\there�exists",�and�no�\for�all",�then�w�e�can�escap�Ge�the�dreaded����?double-exp�Gonen���tial.�NHe�ga�v�e�a�decision�pro�Gcedure�for�this�class�of�form�ulas����?whic���h�P)is�\only"�exp�Gonen�tial�in�the�n�um�b�Ger�of�v��q�ariables.�This�is�an�imp�or-����?tan���t���dierence,�since�with�an�exp�Gonen�tial�algorithm,�if�it�do�Gesn't�run�to�da���y��*�,����?p�Gerhaps���our�c���hildren�will�b�e�able�to�run�it;�while�with�a�double-exp�onen���tial����?running�QCtime,�our�p�Gosterit���y�is�also�do�omed�to�failure.�F��*�urther�impro���v�emen�ts����?since�UU1992�are�describ�Ged�in�[8].��]��NA��FW��*�eb��qsearc���h�sho�ws�that�dozens,�if�not�h�undreds,�of�researc�hers�are�w�ork-����?ing��Son�quan���tier�elimination�these�da�ys.�Although�w�e�kno�w�that�quan�tier����?elimination���will�tak���e�\forev�er"�on�large�problems,�there�still�migh�t�b�Ge�some����?in���teresting��!op�Gen�problems�within�reac�h|a�tan�talizing�p�Gossibilit�y��*�.�Hong�[52]����?made�j-impro���v�emen�ts�to�the�CAD�i�algorithm,�calling�his�enhanced�v�ersion�\par-����?tial��CAD",�and�implemen���ted�it�in�a�program�called�qep�Gcad�(quan�tier�elimi-����?nation��b���y�partial�CAD).�This�program�has�subsequen�tly�b�Geen�impro�v�ed�up�Gon����?b���y�tuman�y�other�p�Geople,�and�is�publicly�a�v��q�ailable�on�the�W��*�eb�[22].�A�t�least�some����?of�its�functionalit���y�has�b�Geen�included�with��Mathematic��}'a��v�ersions�4.1�and�4.2,����?in��the��Exp��}'erimental��pac���k��q�age.�But�to�the�b�Gest�of�m�y�kno�wledge,�the�algorithms����?in�UU[46]�and�[8]�ha���v�e�UUnot�b�Geen�implemen���ted.��?�7܉ff8�ϟ
L͍������UZ��-=�28�����	?��The�+ in��9terested�reader�should�pursue�the�CAD�+algorithm�in�[28];�w�e�cannot�tak�e����	?�the�Gspace�here�ev��9en�to�correctly�dene�what�a�CAD�is,�let�alone�describ�A�e�the����	?�original�Talgorithm�and�its�recen��9t�impro�v�emen�ts�in�full.���������UZ��-=�29�����	?��Mo�A�ore's�32pap�er�con��9tains�a�prophetic�carto�on�sho��9wing�a�\happ�y�home�computer"����	?�coun��9ter�2�b�A�et�w�een�\notions"�and�\cosmetics".�Bill�Gates�w�as�ten�y�ears�old�at�the����	?�time.�������c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������31����)�ō���>;��N�It� seems�fair�to�ask,�then,�what�are�the�presen���t-da�y� limits�of�quan���tier����?elimination�Din�algebra?�The�rst�in���teresting�example,�often�used�as�a�b�Genc�h-����?mark��`for�quan���tier�elimination,�is�to�nd�the�conditions�on��a;���b;�c;�d��`�suc�h�that����?the��fourth-degree�p�Golynomial��x���^��4��7|�+��	�ax���^��3���+��bx���^��2���+��cx��+��d���is�p�Gositiv���e�for�all��x�.����?That�A�is,�to�eliminate�the�quan���tier�\for�all��x�"�in�that�statemen�t.�The�an-����?sw���er��&is�a�surprisingly�complex�p�Golynomial�in��a�,�b�,�c�,�and��d�.�Qep�cad�do�es�this����?v���e-v��q�ariable��8problem�almost�instan�taneously��*�.�The�curious�reader�can�nd�the����?answ���er�UUat�the�qep�Gcad�examples�w�eb�page�[22].���A��NIt�Sw���as�fairly�easy�to�create�a��Mathematic��}'a��noteb�Go�ok�Swith�functions�de-����?ned��to�facilitate�asking�simple�questions�ab�Gout�sphere�pac���king.I�3dened�����?�TwoSpher��}'es��r��[�M��],��whic���h�uses��Ine��}'qualityInstanc�e���to�ask�whether�there�exist����?t���w�o��disjoin�t�spheres�of�radius�1�in�a�cub�Ge�of�side�2�M��.�This�is�a�sev�en-v��q�ariable����?problem,���coun���ting��M���and�the�co�Gordinates�of�the�cen�ters�of�the�spheres.�T��*�o����?mak���e��0it�a�six-v��q�ariable�problem,�w�e�can�put�in�sp�Gecic�v��q�alues�of��M��:��Mathe-����?matic��}'a�	R�answ���ers�the�query��*�,���TwoSpher�es��7�;�[7�=�4]�with��T��;�rue�;�and�with�a�suitable����?v��q�arian���t���of�the�question,�it�could�ev�en�exhibit�an�example�of�t�w�o�suc�h�spheres.�����?�TwoSpher��}'es��r��[3�=�2]�Ireturns��F��;�alse�.�The�time�required�is�less�than�a�second.�The����?sev���en-v��q�ariable���problem,�with��M���v�ariable,�seems�to�b�Ge�to�o�dicult�for�Mathe-����?matica's��Cylindric��}'alA���lgebr�aicDe�c�omp�osition��function.�I��also�tried�these�prob-����?lems���with�the�v���ersion�of�qep�Gcad�a�v��q�ailable�from�[22];�this�program�w�as�able����?to��}express��TwoSpher��}'es��3��[�M��]�as��M��t���Y�1�&�3�M����^��2��"������6�M���+�2��Y���0.��}The�least�suc���h����?�M��²is�Ƨ1��k+�1�=�����P�p���UW���P�fe�E���3����
UX,�whic���h�is�the�v��q�alue�of��M��²one�nds�with�p�Gencil�and�pap�er�if����?one�sqassumes�that�the�cen���ters�of�the�spheres�are�on�the�main�diagonal�of�the����?cub�Ge.��(But�the�program�did�not�mak���e�that�assumption{it��pr��}'ove�d��(�that�this�is����?the���b�Gest�p�ossible�arrangemen���t.�Similar�queries���Thr��}'e�eSpher�es��=z�[�M��],�for�v��q�arious����?sp�Gecic��Cv��q�alues�of��M��,�nev���er�returned�answ�ers.�After�sev�eral�hours�I��"stopp�Ged����?Mathematica;�ϒin�the�v���ersion�of�qep�Gcad�from�[22],�the�jobs�failed�(after�sev�eral����?hours)�UUb�Gecause�they�ran�out�of�memory��*�.�Nine�v��q�ariables�is�to�o�man���y�in�2003.���A��NUsing��a�double-exp�Gonen���tial�algorithm,�w�e�could�exp�Gect�that�with�running����?time���v��q�arying�as�2���^��2����r�n���	c߲,�if��n�F²=�7���corresp�Gonds�to�one�second,�then��n�F²=�8���should����?corresp�Gond���to�2���^��128��C�seconds,�or�more�than�10���^��35��вy���ears,�so�a�sharp�cuto�is�to����?b�Ge��rexp�ected.�As�calculated�ab�o���v�e,�to�increase��n��from�7�to�8,�w�e�need�2���^��7��|s�,����?or��(128,�doublings�of�computer�sp�Geed.�But�the�kissing�problem�needs�only����?existen���tial�nquan�tiers,�so�as�discussed�ab�Go�v�e,�it�sneaks�under�the�w�all:�w�e����?can�.solv���e�it�in�\only"�exp�Gonen�tial�time.�In�that�case�if�2���^��7�����corresp�Gonds�to�one����?second,�0then�2���^��8���v�is�only�t���w�o�0seconds;�but�to�attac���k�the�kissing�problem�w�e����?need�F�100�v��q�ariables,�and�2���^��100���K�corresp�Gonds�to�2���^��93���زseconds{ab�out�10���^��24���زy���ears.����?[Ph���ysicists��7kno�w�a�con�v�enien�t�coincidence:�that�to�three�signican�t�digits,����?there��are����ݸ�E�10���^��7��	d�seconds�p�Ger�y���ear.]�Ev�en�when�the�w�ork�of�Grigorev�is����?implemen���ted,��Dit�still�w�on't�solv�e�the�kissing�problem.�Nev�ertheless,�there�ma�y����?w���ell�9�b�Ge�op�en�questions�with�few���er�than�fteen�v��q�ariables,�so�it�seems�the�jury����?is�UUstill�out�on�the�p�Goten���tial�usefulness�of�quan�tier�elimination.���A��NQuan���tier��,elimination�has�not�b�Geen�the�only�decision�metho�d�used�in����?geometry��*�.��In�1978,�W�u�W�en-Tsen�pioneered�the�reduction�of�geometry�to����� �p�c�ō���>;����?�32���`-	Mic��9hael�TBeeson����)�ō���>;��?�p�Golynomial��dideal�theory��*�,�in���tro�ducing�\W��*�u's�metho�d"�[115].�The�idea�here�is����?that���an�imp�Gortan���t�class�of�geometric�theorems�can�b�e�stated�in�algebraic����?language���without��using�ine��}'qualities�.�If�the�theorem�can�b�Ge�stated�using�only����?conjunctions�bqof�equations�in�the�h���yp�Gothesis�and�an�equation�in�the�conclusion,����?then��it�reduces�to�asking�if�a�certain�p�Golynomial�lies�in�the�ideal�generated����?b���y��)a�nite�set�of�other�p�Golynomials.�Since�that�time,�other�metho�ds,�based����?on���Gr����obner�bases,�ha���v�e���also�b�Geen�applied�to�geometric�theorem�pro���ving.�This����?w���ork��has�reduced�geometric�theorem-pro�ving�to�computer�algebra,�and�when����?it�`*is�applicable,�it�seems�to�b�Ge�more�ecien���t�than�quan�tier�elimination.�Man�y����?in���teresting��2theorems�in�classical�plane�and�solid�geometry�ha�v�e�b�Geen�pro�v�ed����?this�UUw���a�y��*�.�� ����?�11��ZEqualit��y��Reasoning�����?�The�UUusual�axioms�for�equalit���y��*�,�as�giv�en�in�mathematics�textb�Go�oks,�UUare��u������^õx���=��x�������"K�re
exivit���y���������X�x���=��y��.�&�UU�y�"�=��z�7��!��x��=��z�������"K�transitivit���y���������}�T�x���=��y�"�!��y��=��x�������"K�symmetry���������a_�x���=��y��.�&�UU��(�x�)��!���(�y�[ٲ)������"Ksubstitutivit���y������?In�GJthis�form,�these�axioms�are�useless�in�automated�deduction,�b�Gecause�they����?will�w�(in�fact�ev���en�just�the�rst�three�will)�generate�a�nev�er-ending�stream�of����?useless�3deductions.�The�\righ���t�metho�Gd"�for�dealing�with�equalit�y�w�as�disco�v-����?ered��
three�times�in�the�p�Gerio�d��
1965-1970,�indep�Genden���tly�in�[87],�[47],�and�[62].����?The��approac���hes�had�sligh�tly�dieren�t�emphases,�although�the�k�ernel�of�the����?metho�Gds�UUis�the�same.�W��*�e�will�rst�explain�the�Kn���uth-Bendix�metho�d.��'5��NBy�j�an�\orien���ted�equation"��p���=��q��ղw�e�j�simply�mean�a�pair�of�terms�separated����?b���y�h�an�equalit�y�sign,�so�that��p���=��q�Ğ�is�h�not�considered�the�same�orien�ted�equation����?as�2>�q���=�7E�p�.�The�idea�is�that�an�orien���ted�equation�is�to�b�Ge�used�from�left�to����?righ���t�j�only��*�.�The�orien�ted�equation��x�(�y��۲+�G�z�p��)��n=��xy��+�G�xz���can�j�b�Ge�used�to�c���hange����?3�8���(4�+�5)�T�to�3�8���4�+�3����5,�T�but�not�vice-v���ersa.�The�v��q�ariables�can�b�Ge�matc�hed����?to�complicated�expressions,�although�this�example�sho���ws�them�matc�hed�to����?constan���ts.��}Another�name�for�\orien�ted�equation"�is�\rewrite�rule",�whic�h����?con���v�eys�UUan�in���tuition�ab�Gout�ho�w�rewrite�rules�are�to�b�Ge�used.��'5��NSupp�Gose�� one�is�giv���en�a�set��E�6��of�orien�ted�equations.�Giv�en�an�expression����?�t�,��w���e�can�rewrite��t��or�its�subterms�using�(orien�ted)�equations�from��E�}��un�til����?no��Xmore�rules�can�b�Ge�applied.�If�this�happ�ens,�the�resulting�term�is�called�a����?\normal�#�form"�of��t�.�It�need�not�happ�Gen:�for�example,�if��E����includes�the�equa-����?tion�Z�xy�*�=���y�[�x�,�then�w���e�ha�v�e��ab���=��ba��=��ab��=���:���:�:�����ad��Kinnitum�.�ZIf�one�sequence����?of�k]rewrites�do�Ges�terminate�in�a�normal�form,�it�still�do�es�not�guaran���tee�that����?ev���ery���suc�h�sequence�terminates�(dieren�t�subterms�can�b�Ge�rewritten�at�dif-����?feren���t�ustages).�If,�no�matter�what�subterm�w�e�rewrite�and�no�matter�what����?equation�G�from��E���w���e�use,�the�result�alw�a�ys�terminates�in�a�normal�form,�and�����!���c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������33����)�ō���>;��?�if�@�this�happ�Gens�no�matter�what�term��t��w���e�start�with,�then��E���is�called��termi-����?nating�.��@��NEv���en��)this�do�Ges�not�guaran�tee�the�uniqueness�of�the�normal�form�of��t�.����?That��lw���ould�b�Ge�guaran�teed�b�y�the�follo�wing�desirable�prop�Gert�y�of��E����,�kno�wn�as����?�c��}'on
uenc�e�.���E��y�is�called��c��}'on
uent��if�whenev���er�a�term��t��can�b�Ge�rewritten�(using����?one���or�more�steps)�in�t���w�o���dieren�t�w�a�ys�to��r�7��and��s�,�then�there�exists�another����?term��&�q� ��suc���h�that�b�Goth��r�C�and��s��can�b�e�rewritten�to��q�[ٲ.�This�prop�ert���y�clearly����?ensures���the�uniqueness�of�normal�forms�b�Gecause,�if��r��and��s��w���ere�distinct����?normal�UUforms�of��t�,�it�w���ould�b�Ge�imp�ossible�to�rewrite�them�as��q�[ٲ.��@��NThese��concepts�will�b�Ge�made�clear�b���y�considering�the�example�of�group����?theory��*�.�UUConsider�the�usual�three�axioms�of�group�theory:���Z����n�e�8���x�������Բ=��������x��������b��i�(�x�)�8���x�������Բ=��������e��������X�(�x�8���y�[ٲ)����z�������Բ=��������x�8���(�y������z�p��)������?This���set�is�not�con
uen���t.�F��*�or�example,�the�term�(�i�(�a�)����a�)����b�)���can�b�Ge�rewritten����?to���e��¸��b��and�then�to��b�,�but�it�can�also�b�Ge�rewritten�to��i�(�a�)����(�a����b�),�whic���h����?cannot��%b�Ge�rewritten�further.�Do�es�there�exist�a�terminating�con
uen���t�set�of����?equations�*�E����extending�these�three,�and�suc���h�that�eac�h�of�the�equations�in����?�E��O�is��a�theorem�of�group�theory?�This�is�an�in���teresting�question�b�Gecause�if����?there�iis,�it�w���ould�enable�us�to�solv�e�the��wor��}'d�\�pr�oblem�for�gr�oup�the�ory�:�igiv���en����?an��dequation��t�l1�=��s�,��ddo�Ges�it�follo���w�from�the�three�axioms�of�group�theory?�If����?w���e�d�had�a�complete�con
uen�t�set��E����,�w�e�could�simply�rewrite��t��and��s��to�their����?resp�Gectiv���e���unique�normal�forms,�and�see�if�the�results�are�iden�tical.�If�so,�then����?the�UUequation��t���=��s�UU�is�a�theorem�of�group�theory��*�.�If�not,�it�is�not�a�theorem.��@��NThe���answ���er�for�group�theory�is�a�set�of�ten�equations.�These�are�the����?original�UUthree,�plus�the�follo���wing�sev�en:������X�i�(�x�)�8���(�x����y�[ٲ)��������=������\�y��������z�x�8���e��������=������\�x��������~��i�(�e�)��������=������\�e��������rk�i�(�i�(�x�))��������=������\�x��������n{�x�8���i�(�x�)��������=������\�e��������Xx�8���(�i�(�x�)����y�[ٲ)��������=������\�y��������n�i�(�x�8���y�[ٲ)��������=������\�i�(�y�[ٲ)�8���i�(�x�)������?W��*�e���call�this�set�of�ten�equations�\complete"�b�Gecause�it�pro���v�es���the�same����?equations�&�as�the�original�three�axioms,�i.e.,�all�the�theorems�of�group�theory��*�,����?but�Uyit�can�do�so�b���y�using�the�ten�equations�only�left-to-righ�t,�while�the�original����?three�Tdm���ust�b�Ge�used�in�b�oth�directions�to�pro���v�e�Tdthe�same�theorems.�In�tec���hnical����?language:��the�ten�equations�constitute�a�complete�con
uen���t�set.�That�set����?happ�Gens�ˆto�con���tain�the�original�three�axioms,�but�that�can�b�e�view���ed�as����?acciden���tal.��W��*�e�w�ould�not�ha�v�e�cared�if�the�original�axioms�had�themselv�es�����"*�c�ō���>;����?�34���`-	Mic��9hael�TBeeson����)�ō���>;��?�simplied��somewhat�in�the�nal�ten.�(Of�course,�the�original�axioms�of�group����?theory���w���ere�c�hosen�to�b�Ge�as�simple�as�p�ossible,�so�it�is�not�really�acciden���tal����?that�UUthey�are�among�the�ten.)���A��NThis��solution�of�the�w���ord�problem�for�groups�can�b�Ge�v��q�astly�generalized.����?Donald��Kn���uth�in�v�en�ted�an�algorithm,�whic�h�w�as�implemen�ted�b�y�his�studen�t����?Bendix���(in�F���OR��*�TRAN��cIV�for���the�IBM��c7094),�and�has�b�Gecome�kno�wn�as����?the��VKn���uth-Bendix�algorithm�since�they�w�ere�the�join�t�authors�of�[62].�This����?algorithm���w���as�published�in�1970,�but�the�w�ork�w�as�done�considerably�earlier.����?The��input�is�a�set��E��#�of�(unorien���ted�or�orien�ted)�equations.�The�output�(if����?the��halgorithm�terminates)�is�a�set��Q��of�orien���ted�equations�(rewrite�rules)����?that��{is�con
uen���t�and�terminating,�and�has�the�same�(unorien�ted)�equations����?as�5}logical�consequences�as�the�original�set��E����.�Ho���w�ev�er,�5}in�general�there�is�no����?guaran���tee�K�of�termination.�One�can�run�this�algorithm�with�the�three�axioms����?of�]�group�theory�as�input�and�obtain�the�ten-equation�system�giv���en�ab�Go�v�e�as����?output.���A��NThe��`Kn���uth-Bendix�metho�Gd�is�(or�can�b�e�with�appropriate�commands)����?used�51b���y�most�mo�Gdern�theorem-pro�v�ers.�It�is�in�tegrated�with�the�other�meth-����?o�Gds�$Pused�in�suc���h�theorem-pro�v�ers.�Here�is�an�example�of�an�in�teresting�theo-����?rem���pro���v�ed�b�y�this�metho�Gd:�In�a�ring�supp�ose��x���^��3��C��=���x��for�all��x�.�Then�the�ring����?is�bcomm���utativ�e.�This�is�pro�v�ed�b�y�starting�out�with�the�set��E���con�taining�the����?ring�yaxioms�and�the�axiom��xxx���=��x�.�yThen�the�Kn���uth-Bendix�algorithm�is�run����?un���til�m�it�deduces��xy�Ki�=�y�[�x�.�When�that�happ�Gens,�a�con�tradiction�will�b�Ge�found����?b���y�4hresolution�with�the�negated�goal��ab���6�=��ba�,�4hso�the�Kn�uth-Bendix�algorithm����?will���not�go�o��ad�	innitum��using�the�comm���utativ�e���la�w�(as�it�w�ould�if�running����?b���y��Sitself.)�The�resulting�pro�Gof�is�52�steps�long.�Up�un�til�1988�it�to�Gok�ten�hours����?to��fnd�this�pro�Gof;�then�the�pro���v�er��fRRL��Z[59]�w���as�able�to�reduce�this�time�to����?t���w�o�g�min�utes.�Actually��*�,�the�h�yp�Gothesis��x���^��3��bI�=��ֵx��can�b�e�replaced�b���y��x���^��n��WT�=��ֵx��for����?an���y���natural�n�um�b�Ger��n�Z���2,���and�RRL��scould�also�do�the�cases��n�Z�=�4�;����6�;��8���:�:�:��
UO�,����?and��man���y�other�ev�en�v��q�alues�of��n��[117],�but�it�still�tak�es�a�h�uman�b�Geing�to����?pro���v�e��wit�for�all��n�,�b�Gecause�the�(only�kno���wn)�pro�of�in���v�olv�es��winduction�on��n��and����?the��theory�of�the�Jacobsen�radical�(a�second-order�concept).�The�o�Gdd�cases����?are�UUstill�quite�hard�for�theorem�pro���v�ers.���A��NThe��buse�of�a�set�of�orien���ted�equations�to�rewrite�subterms�of�a�giv�en�term����?is�pcalled�\demo�Gdulation"�in�the�automated�theorem�pro���ving�comm�unit�y��*�,�and����?\rewriting"�M�in�an�almost�separate�group�of�researc���hers�who�study�rewrite�rules����?for�N�other�reasons.�A�N�set�of�orien���ted�equations�can�b�Ge�sp�ecied�b���y�the�user�of����?Otter��-as�\demo�Gdulators".�They�will�b�e�used�to�\reduce"�(rep�eatedly�rewrite)����?all�}�newly-generated�clauses.�What�the�Kn���uth-Bendix�algorithm�do�Ges,�in�ad-����?dition��to�this,�is�to�use�the�existing�demo�Gdulators�at�eac���h�stage�to�generate����?new���demo�Gdulators�dynamically��*�.�The�metho�d�is�simple:�Find�a�subterm�of�one����?of���the�left-hand�sides�that�can�b�Ge�rewritten�in�t���w�o���dieren�t�w�a�ys�b�y�dier-����?en���t�cdemo�Gdulators.�Reduce�the�left-hand�side�as�far�as�p�ossible�after�starting����?in���these�t���w�o���dieren�t�w�a�ys.�Y��*�ou�will�obtain�t�w�o�terms��p��and��q�[ٲ.�If�they�are����?dieren���t,��then�the�set�of�existing�demo�Gdulators�is�manifestly�not�con
uen�t,�����#Ƞc�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������35����)�ō���>;��?�and�vthe�equation��p����=��q���is�va�candidate�for�a�new�demo�Gdulator.�The�pair��p�,��q����?�is���called�a��critic��}'al���p�air�.���Also��q��"�=�xI�p��is�a�candidate,�so�the�dicult���y�is�whic�h����?w���a�y�S�the�new�equation�should�b�Ge�orien���ted.�The�solution�is�to�put�the�\hea�vi-����?est"�� term�on�the�left.�In�the�simplest�case,�\hea���viest"�just�means�\longest",����?but���if�the�algorithm�do�Ges�not�halt�with�that�denition�of�\w���eigh�t",���other����?more��complicated�denitions�migh���t�mak�e�the�algorithm�con�v�erge.�In�short,����?the���Kn���uth-Bendix�algorithm�dep�Gends�on�a�w�a�y�of�orien�ting�new�equations,����?and�UUman���y�pap�Gers�ha�v�e�b�Geen�written�ab�out�the�p�ossible�metho�ds.������NBecause�:�comm���utativit�y�is�imp�Gortan�t�in�man�y�examples,�but�mak�es�the����?Kn���uth-Bendix��algorithm�fail�to�con�v�erge,�some�eort�has�b�Geen�exp�ended�to����?generalize�8�the�algorithm.�If�the�matc���hing�for�rewrites�is�done�using�\asso�Gciativ�e-����?comm���utativ�e�vunication"�instead�of�ordinary�unication,�then�the�algorithm����?still�S�w���orks,�and�one�can�simply�omit�the�comm�utativ�e�and�asso�Gciativ�e�ax-����?ioms��[5].�This�w���as�the�metho�Gd�emplo�y�ed�in�McCune's�theorem-pro�v�er�EQP����?to�UUsettle�the�Robbins�Conjecture�[71].������NReturning�nfto�the�late�1960s,�w���e�no�w�describ�Ge�the�con�tribution�of�George����?Robinson�8�and�Larry�W��*�os.�They�dened�the�inference�rule�they�called��p��}'ar�amo�d-����?ulation�.��This�is�essen���tially�the�rule�used�to�generate�new�critical�pairs�at�eac�h����?step��Lof�the�Kn���uth-Bendix�algorithm.�But�in�[87],�it�w�as�not�restricted�to�theo-����?ries��dwhose�only�relation�sym���b�Gol�is�equalit�y��*�.�Instead,�it�w�as�view�ed�as�a�general����?adjunct��{to�resolution.�One�retains�the�re
exivit���y�axiom��x�{W�=��x��{�and�replaces����?transitivit���y��*�,�n�symmetry�,�and�substitutivit���y�with�the�new�inference�rule.�They����?used�vthis�metho�Gd�to�nd�pro�ofs�of�theorems�that�w���ere�previously�b�ey���ond�the����?reac���h���of�computer�programs.�F��*�or�example,�with�(�x���
��y�[ٲ)���dened�as�the�comm�u-����?tator��of��x��and��y�[ٲ,�they�pro���v�ed��that�in�a�group,�if��x���^��3��C��=��1�then�(�x���
��y��)��
��y�"�=��1.����?Although�Q�this�example�is�purely�equational,�the�rule�of�paramo�Gdulation�is����?generally��Yapplicable,�whatev���er�relation�sym�b�Gols�ma�y�o�Gccur�in�addition�to����?equalit���y��*�.��fRobinson�and�W�os�pro���v�ed��f[88]�the�refutation-completeness�of�this����?metho�Gd,�xji.e.,�an���y�theorem�has�a�pro�of�b���y�con�tradiction�using�resolution�and����?paramo�Gdulation,�with�the�axiom��x���=��x�.�On�the�other�hand,�Robinson�and����?W��*�os��bdid�not�in���tro�Gduce�the�concept�of�con
uence�or�of�a�complete�con
uen�t����?set�2�of�rules,�so,�for�example,�the�deduction�of�the�ten�group-theory�theorems����?giv���en�UUab�Go�v�e�escap�Ged�their�notice.������NF��*�or�o�theories�relying�exclusiv���ely�on�equalit�y��*�,�no�serious�distinction�should����?made�I�b�Get���w�een�the�Kn�uth-Bendix�metho�Gd�and�paramo�dulation.�They�are�es-����?sen���tially���the�same�thing.����^��30���O��Nev�ertheless,�as�men�tioned�b�Gefore,�there�is�a�com-����?m���unit�y��of�researc���hers�in�\rewrite�rules"�and�an�almost�disjoin�t�comm�unit�y����?of��+researc���hers�in�\automated�deduction",�eac�h�with�their�o�wn�conferences����?and�*2journals.�The�c���hallenge�for�to�Gda�y's�w�ork�ers�in�equalit�y�reasoning�is�to����?connect�X�the�v��q�ast�b�Go�dy�X�of�existing�w���ork�with�the�w�ork�that�has�b�Geen�done�in��?�D�ff8�ϟ
L͍������UZ��-=�30�����	?��The�9four�dierences�listed�on�p.�20�of�[72]�are�actually�dierences�in�the�w��9a�y����	?�the�]�tec��9hnique�is�used�in�Otter�and�the�w�a�y�it�w�ould�b�A�e�used�in�a�program�that����	?�implemen��9ted�Tonly�Kn�uth-Bendix.�����$%+�c�ō���>;����?�36���`-	Mic��9hael�TBeeson����)�ō���>;��?�computer��Yalgebra,�so�that�pro�Gofs�in���v�olving��Ycomputation�can�b�egin�to�b�e�done����?b���y�UUcomputer.�This�task�has�hardly�b�Geen�b�egun.��!�����?�12��^�Pro�`ofs��In��v�olving�Computations������?�There��ha���v�e�alw�a�ys�b�Geen�t�w�o�asp�Gects�of�mathematics:�logical�reasoning�and����?computation.�}�These�ha���v�e�}�historical�ro�Gots�as�far�bac���k�as�Greece�and�Bab�ylo-����?nia,��[resp�Gectiv���ely��*�.�Eorts�to�mec�hanize�mathematics�b�Gegan�with�computation,����?and���as�discussed�ab�Go���v�e,���the�mac���hines�of�P�ascal�and�Leibniz�preceded�the�Log-����?ical�EYPiano.�In�our�time,�the�mec���hanization�of�computation�via�computer�has����?b�Geen�4�m���uc�h�more�successful�than�the�mec�hanization�of�logical�reasoning.�The����?mec���hanization��Fof��symb��}'olic��computation�(as�opp�Gosed�to�n�umerical�computa-����?tion)��b�Gegan�in�the�fties,�as�did�the�mec���hanization�of�logic.�What�is�in�ter-����?esting,���and�surprising�to�p�Geople�outside�the�eld,�is�that�the�mec���hanization����?of���logic�and�the�mec���hanization�of�computation�ha�v�e�pro�Gceeded�somewhat����?indep�Genden���tly��*�.��vW�e�no���w�ha�v�e�computer�programs�that�can�carry�out�v�ery����?elab�Gorate�Սcomputations,�and�these�programs�are�used�b���y�mathematicians�\as����?required".�'�W��*�e�also�ha���v�e�'�\theorem-pro�v�ers",�but�for�the�most�part,�these�t�w�o����?capabilities��do�not�o�Gccur�in�the�same�program,�and�these�programs�do�not����?ev���en�UUcomm�unicate�usefully��*�.��P%��NP���art�,of�the�problem�is�that�p�Gopular�sym�b�Golic�computation�soft�w�are�(suc�h����?as��yMathematica,�Maple,�and�Macsyma)�is�logically�incorrect.�F��*�or�example:����?Set��u�a����=�0.�Divide�b�Goth�sides�b���y��a�.�Y��*�ou�get�1�=�0,�b�Gecause�the�soft���w�are����?thinks��a=a��e�=�1�and�0�=a��=�0.�This�kind�of�problem�is�p�Gerv��q�asiv���e�and�is�not����?just��wan�isolated�\bug",�b�Gecause�computation�soft���w�are��wapplies�transformations����?without�c���hec�king�the�assumptions�under�whic�h�they�are�v��q�alid.�Alternately��*�,�if����?transformations�(are�not�applied�unless�the�assumptions�are�all�c���hec�k�ed,�(then����?computations�I�grind�to�a�halt�b�Gecause�the�necessary�assumptions�are�not�v���eri-����?able.���The�author's�soft���w�are���MathXp�Gert�[11],�whic���h�w�as�written�for�education����?rather�9than�for�adv��q�anced�mathematics,�handles�these�matters�correctly��*�,�as����?describ�Ged��Sin�[10].�Later�v���ersions�of�Mathematica�ha�v�e�b�Gegun�attac�king�this����?problem��Sb���y�restricting�the�applicabilit�y�of�transformations�and�allo�wing�the����?user�qBto�sp�Gecify�assumptions�as�extra�argumen���ts�to�transformations,�but�this����?is�Цnot�a�complete�solution.�Buc���h�b�Gerger's�Ц�The��}'or�ema��pro��8ject�[23]�is�the�b�Gest�at-����?tempt�ٻso�far�to�com���bine�logic�and�computation,�but�it�is�not�in�tended�to�b�Ge�a����?pro�Gof-nder,�h�but�rather�a�pro�of-c���hec�k�er,�h�enabling�a�h���uman�to�in�teractiv�ely�de-����?v���elop�'�a�pro�Gof.�The�dicult�y�here�is�that�when�the�underlying�computational����?abilit���y��lof�Mathematica�is�used,�it�is�hard�to�b�Ge�certain�that�all�error�has�b�een����?excluded,��b�Gecause�Mathematica�do�es�not�ha���v�e��a�systematic�w���a�y��of�trac���king����?or�UUv���erifying�the�pre-conditions�and�p�Gost-conditions�for�its�transformations.��P%��NAnother���program�that�w���as�a�pioneer�in�this�area�is��A���nalytic��}'a��[30].�This����?w���as�S�a�theorem-pro�v�er�written�in�the�Mathematica�programming�language.����?\W��*�as"�mYis�the�appropriate�tense,�since�this�program�is�no�longer�in�use�or�under����?dev���elopmen�t.����A���nalytic��}'a��w���as�primarily�useful�for�pro�ving�iden�tities,�and�made�����%5_�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������37����)�ō���>;��?�a�Nosplash�b���y�pro�ving�some�dicult�iden�tities�from�Raman�ujan's�noteb�Go�oks.����?It��Tcould�not�deal�with�quan���tied�form�ulas�and�did�not�ha�v�e�state-of-the-art����?searc���hing�UUabilities.��8�NOn�4�the�theorem-pro���ving�side�of�the�endea�v�or,�eorts�to�incorp�Gorate�com-����?putation�*�in�theorem-pro���v�ers�*�ha�v�e�b�Geen�restricted�to�t�w�o�approac�hes:�using����?rewrite�KHrules�(or�demo�Gdulators),�and�calling�external�decision�pro�cedures�for����?form���ulas��qin�certain�sp�Gecialized�forms.�The�sub��8ject�kno�wn�as�\constrain�t�logic����?programming"�_|(CLP)�_ycan�b�Ge�considered�in�this�latter�category��*�.�T�o�Gda���y�there����?are�h�a�few�exp�Gerimen���ts�in�linking�decision-pro�cedure�mo�dules�to�pro�of-c���hec�k�ers����?(e.g.��Oqep�Gcad�to�PVS),�but�there�is�little�w���ork�in�linking�decision-pro�cedure����?mo�Gdules�UUto�pro�of-nding�programs.����^��31����8�N�The�z�author's�soft���w�are�z�MathXp�Gert�con���tains�computational�co�de�that�prop-����?erly���trac���ks�the�preconditions�for�the�application�of�mathematical�transforma-����?tions.��gAfter�publishing�MathXp�Gert�in�1997,�I��4then�com���bined�some�of�this�co�de����?with���a�simple�theorem�pro���v�er���I���had�written�earlier�[9],�and�w���as�therefore�in����?a���unique�p�Gosition�to�exp�erimen���t�with�the�automated�generation�of�pro�ofs����?in���v�olving��wcomputation.�I��'named�the�com���bined�theorem-pro�v�er�W��*�eierstrass����?b�Gecause�T�the�rst�exp�erimen���ts�I�p�erformed�in���v�olv�ed�T�epsilon{delta�argumen���ts.����?These���are�the�rst�pro�Gofs,�other�than�simple�mathematical�inductions,�to����?whic���h�<�studen�ts�of�mathematics�are�exp�Gosed.�I�<�used�W��*�eierstrass�in�1988-1990����?to��vnd�epsilon-delta�pro�Gofs�of�the�con���tin�uit�y��vof�sp�ecic�functions�suc���h�as�p�o���w-����?ers��hof�x,�square�ro�Got,�log,�sine�and�cosine,�etc.�Before�this,�the�b�est�that�could����?b�Ge��Ydone�w���as�the�con�tin�uit�y�of�a�linear�function�[19].�These�pro�Gofs�in�v�olv�e����?simple���algebraic�la���ws�(or�la�ws�in�v�olving�sine,�cosine,�log,�and�the�lik�e),�but,����?what��is�more,�they�in���v�olv�e��com�bining�those�computations�with�inequalit�y����?reasoning.��8�NI�˴then���mo���v�ed�from�analysis�to�n�um�b�Ger�theory��*�,�and�considered�the�pro�of����?of�a�the�irrationalit���y�of��e�.�W��*�eierstrass�w�as�able,�after�sev�eral�impro�v�emen�ts,����?to�Oautomatically�generate�a�pro�Gof�of�this�theorem�[13].�The�pro�of�in���v�olv�es����?inequalities,�F�b�Gounds�on�innite�series,�t���yp�e�distinctions�(b�et���w�een�F�real�n���um�b�ers����?and���natural�n���um�b�Gers),���a�subpro�of�b���y�mathematical�induction,�and�signican�t��?�1�'�ff8�ϟ
L͍������UZ��-=�31�����	?��P��9ossible���exceptions:�if�the�set�of�demo�A�dulators�is�con
uen�t�and�complete,�then����	?�demo�A�dulation���could�b�e�regarded�as�a�decision�pro�cedure�for�equations�in�that����	?�theory��:�.���Bledso�A�e's�rules�[19],�[25]�(Ch.�8)�for�inequalities�could�b�e�regarded�as�a����	?�decision��pro�A�cedure�for�a�certain�class�of�inequalities.�Theorem�pro��9v�ers��suc�h�as����	?�EQP�#and�#RRL�that�ha��9v�e�A�C-unication�could�b�A�e�regarded�as�ha�ving�a�decision����	?�pro�A�cedure�f/for�linear�equations.��The��or�ema�f/�do�es�con��9tain�decision�pro�cedures,�but�it����	?�is�
Aprimarily�a�pro�A�of-c��9hec�k�er,�
Anot�a�pro�of-nder.�An�extension�of�the�pro��9v�er�
ARRL����	?�called�TT��:�ecton�[1]�has�a�decision�pro�A�cedure�for�Presburger�arithmetic.�����&D��c�ō���>;����?�38���`-	Mic��9hael�TBeeson����)�ō���>;��?�mathematical�ssteps,�including�correct�simplication�of�expressions�in���v�olving����?factorials�UUand�summing�an�innite�geometrical�series.����^��32������;��^�33������N�Inequalities�%pla���y�ed�a�cen�tral�role�in�b�Goth�the�epsilon-delta�pro�ofs�and�the����?pro�Gof���of�the�irrationalit���y�of��e�.�Inequalities�certainly�pla�y�a�cen�tral�role�in�clas-����?sical�analysis.�Bo�Goks�and�journal�articles�ab�out�partial�dieren���tial�equations,����?for�vzexample,�are�full�of�inequalities�kno���wn�as�\estimates"�or�\b�Gounds",�that����?pla���y�>�k�ey�roles�in�existence�pro�Gofs.�Classically��*�,�mathematics�has�b�een�divided����?in���to���algebra�and�analysis.�I��~w�ould�v�en�ture�to�call�algebra�the�mathematics�of����?equalit���y��*�,�UUand�analysis�the�mathematics�of�inequalit�y��*�.����NThe���mec���hanization�of�equalit�y�reasoning�has�made�more�progress�than�the����?mec���hanization��Nof�inequalit�y�reasoning.�W��*�e�ha�v�e�discussed�the�\rst�lo�Gophole"����?ab�Go���v�e,�Dvwhic�h�allo�ws�for�the�complete�mec�hanization�of�certain�subelds�of����?mathematics��?b���y�a�\decision�pro�Gcedure"�that�algorithmically�settles�questions����?in��Ea�sp�Gecic�area.�The�mec���hanization�of�equalit�y�reasoning�has�b�Geneted�from����?the��`disco���v�ery�of�decision�pro�Gcedures�with�surprisingly�wide�applicabilit�y��*�.�In����?particular,�8ka�decision�pro�Gcedure�has�b�een�found�for�a�class�including�what����?are��usually�called��c��}'ombinatorial�)�identities�.�Com���binatorial�iden�tities�are�those����?in���v�olving��(sums�and�binomial�co�Gecien���ts,�often�in�quite�complicated�algebraic����?forms.�UUT��*�o�illustrate�with�a�v���ery�simple�example,��j���������Ѵn�������ެ����X����t���,9�j�g��=0����������^�����<$���W��n��
�卒�|j�����X+���^������ട��۱2���*�=��������^�����<$��
#��2�n��
�卑��n����$'���^��������:���R��?�In�"�1974�it�w���as�recognized�b�y�Gosp�Ger,�who�w�as�at�that�time�in�v�olv�ed�in����?the���creation�of�Macsyma,�that�almost�all�suc���h�iden�tities�are�sp�Gecial�cases����?of�\Ka�few�iden���tities�in�v�olving��hyp��}'er�ge�ometric��Nfunctions�,�\Kan�area�of�mathemat-����?ics�=initiated,�lik���e�so�man�y�others,�b�y�Gauss.�In�1982,�Doron�Zeilb�Gerger�re-����?alized���that�recurrence�relations�for�suc���h�iden�tities�can�b�Ge�generated�auto-����?matically��*�.��,This�realization�is�the�basis�for�\Zeilb�Gerger's�paradigm"�(see�[81],����?p.���23).�This�\paradigm"�is�a�metho�Gd�for�pro���ving�an�iden�tit�y�of�the�form�����?����P���I�;��k���P$s�summand��y�7�(�n;���k�P��)��=���answer��!��(�n�).��/Namely:�(i)�nd�a�recurrence�relation�satis-����?ed�1bb���y�the�sum;�(ii)�sho�w�that�the�prop�Gosed�answ�er�satises�the�same�recur-����?rence;��u(iii)�c���hec�k��uthat�\enough"�initial�v��q�alues�of�b�Goth�sides�are�equal.�Here����?\enough"���dep�Gends�on�the�rational�functions�in���v�olv�ed���in�the�recurrence�rela-����?tion.�ѡThe�k���ey�to�automating�pro�Gofs�of�com�binatorial�iden�tities�is�to�automate��?��<�ff8�ϟ
L͍������UZ��-=�32�����	?��Tw��9o���things�particularly�am�used�me�ab�A�out�this�piece�of�w�ork:�First,�one�of�the����	?�referees���said�\Of�course�it's�a�stun��9t."�Second,�audiences�to�whom�I�ҷlectured�w�ere����	?�quite�ȋready�to�accept�that�next�I��\migh��9t�b�A�e�pro�ving�the�irrationalit�y�of�Euler's����	?�constan��9t�j5�
��ûor�solving�other�op�A�en�problems.�P�eople�to�A�da�y�are�quite�jaded�ab�A�out����	?�the�X�amazing�latest�accomplishmen��9ts�of�computers!�What�the�referee�mean�t�w�as����	?�that�90the�\stun��9t"�w�as�not�going�to�b�A�e�rep�eated�an��9y�time�so�on�with�famous�op�en����	?�problems�Tof�n��9um�b�A�er�Ttheory��:�.���������UZ��-=�33�����	?��It��w��9as�dicult�for�others�to�build�up�A�on�this�w�ork�in�that�the�co�A�de�from�MathXp�ert����	?�could�
�not�b�A�e�shared,�b�ecause�it�is�part�of�a�commercial�pro�duct�no�longer�under����	?�the�-author's�con��9trol.�In�the�future,�similar�features�should�b�A�e�added�to�an�existing,����	?�widely-used�Ttheorem�pro��9v�er,�Twhose�source�co�A�de�is�accessible,�suc��9h�as�Otter.�����'S��c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������39����)�ō���>;��?�the�-ndisco���v�ery�of�the�appropriate�recurrence�relation.�In�[81],�one�can�learn����?ho���w�=this�is�done,�using�metho�Gds�whose�ro�ots�lie�in�Gosp�er's�algorithm�for����?the�]�summation�of�h���yp�Gergeometric�series,�and�in�y�et�earlier�w�ork�b�y�Sister����?Mary�z�Celine�on�recurrence�relations.�The�app�Gendix�of�[81]�con���tains�p�oin���ters����?to��Mathematica�and�Maple�implemen���tations�of�the�algorithms�in�question.����?In�O addition�to�v���erifying�prop�Gosed�iden�tities,�some�of�these�algorithms�can,����?giv���en��only�the�left-hand�sum,�determine�whether�there�exists�an�\answ�er"�in����?a��certain�form,�and�if�so,�nd�it.�The�algorithms�presen���ted�in�[81]�are�note-����?w���orth�y�
�b�Gecause,�unlik���e�either�pro�of-searc���h�or�quan�tier�elimination�for�the����?reals,�lthey�routinely�p�Gerform�at�h���uman�lev�el�or�b�Getter�in�nding�and�pro�ving����?com���binatorial�UUiden�tities.��!�+���?�13��ZSearc��hing��for�Pro�`ofs���+��?�In�N�essence,�automated�deduction�is�a�searc���h�problem.�W��*�e�ha�v�e�a�list�of�axioms,����?a�B�few�\sp�Gecial�h���yp�otheses"�of�the�theorem�to�b�e�pro���v�ed,�B�and�the�negation�of����?its���conclusion,�and�some�inference�rules.�These�inputs�determine�a�large�space����?of�B�p�Gossible�conclusions�that�can�b�e�dra���wn.�W��*�e�m�ust�searc�h�that�space�to�see����?if�/ it�con���tains�a�con�tradiction.�In�some�approac�hes�to�automated�deduction����?(that���do�not�use�pro�Gof�b���y�con�tradiction),�w�e�migh�t�not�put�the�negation�of����?the��&conclusion�in,�and�then�searc���h�the�p�Gossible�deductions�for�the�conclusion,����?instead���of�a�con���tradiction.�Either�w�a�y��*�,�a�searc�h�is�in�v�olv�ed.�T��*�o�the�exten�t�that����?calculation�8is�in���v�olv�ed,�8the�searc���h�can�b�Ge�limited{when�w�e�are�calculating,����?w���e�.�\kno�w�what�w�e�are�doing".�But�the�logical�comp�Gonen�t�of�mathematics����?in���v�olv�es,�w�ev�en�in�tuitiv�ely��*�,�a�searc�h.�W��*�e�\nd"�pro�Gofs,�w�e�do�not�\construct"����?them.��Wo��NThis�T�searc���h�app�Gears�to�b�e�fundamen���tally�infeasible.�Let�us�see�wh�y�b�y����?considering�`>a�straigh���tforw�ard�`>\breadth-rst�searc���h",�as�a�computer�scien�tist����?w���ould�L�call�it.�Supp�Gose�w�e�start�with�just�3�axioms�and�one�rule�of�inference.����?The���three�axioms�w���e�call�\lev�el�0".�Lev�el��n��*�+�1���is�the�set�of�form�ulas�that����?can�Ϥb�Ge�deduced�in�one�step�from�form���ulas�of�lev�el��n��or�less,�at�least�one�of����?whic���h�k_has�lev�el�exactly��n�.�The�\lev�el�saturation�strategy"�is�to�generate�the����?lev���els,��one�b�y�one,�b�y�applying�the�inference�rule�to�all�pairs�of�form�ulas�of����?lo���w�er���lev�els.�It�is�dicult�to�coun�t�the�size�of�the�lev�els�exactly�b�Gecause�w�e����?cannot��tell�in�adv��q�ance�ho���w�man�y�pairs�of�form�ulas�can�function�as�premisses����?of��jthe�inference�rule.�But�for�a�w���orst-case�estimate,�if��L����n��	a�is�the�n�um�b�Ger�of����?form���ulas���in�lev�el��n��or�less,�w�e�w�ould�ha�v�e��L����n�+1���E�=�!ĵL����n��΢�+�]$�L����n��q~�(�L����n�����L����n��1����).���T��*�o����?mak���e��"a�ten�tativ�e�analysis,�assume�that��L����n��1��[�can�b�Ge�neglected�compared�to����?the�d6m���uc�h�larger��L����n��q~�.�Then�the�recursion�is�appro�ximately��L����n�+1��qf�=���L���^���2��፴n���q~�,�whic�h����?is��7solv���ed�b�y��L����n��庲=�t<2���^��2����r�n���	c߲.�When��n��=�7�w���e�ha�v�e�2���^��128��uY�,�a�n�um�b�Ger�that�compares����?with�Ythe�n���um�b�Ger�Yof�electrons�in�the�univ���erse�(said�to�b�e�10���^��44��x�).�Y��*�et�pro�ofs�of����?lev���el�}30�are�often�found�b�y�Otter�(according�to�[109],�p.�225).�Of�course,�w�e����?ha���v�e�1Ygiv�en�a�w�orst-case�estimate,�but�in�practice,�lev�el�saturation�is�not�a����?feasible�UUw���a�y�to�organize�pro�Gof�searc�h.�����(d��c�ō���>;����?�40���`-	Mic��9hael�TBeeson����)�ō���>;��N�In���tuitiv�ely��*�,��the�dicult���y�with�lev�el�saturation�is�this:�what�w�e�are�doing����?with��*lev���el�saturation�(whether�or�not�the�negation�of�the�conclusion�is�thro�wn����?in)��kis�dev���eloping�the�en�tire�theory�from�the�axioms.�Naturally�there�will�b�Ge����?man���y�Yconclusions�that�are�irrelev��q�an�t�to�the�desired�one.�Whole�b�Go�oks�Yma�y����?exist�[�lled�with�in���teresting�deductions�from�these�axioms�that�are�irrelev��q�an�t����?to�Gda���y���in�spite�of�b�eing�in���teresting�on�another�da�y��*�,�and�there�will�of�course�b�Ge����?ev���en�UUmore�unin�teresting�conclusions.�What�w�e�need,�then,�are�tec�hniques�to��W�����E�������P�prev���en�t�UUthe�generation�of�un���w�an�ted�UUdeduced�clauses,���9�����E�������P�discard���un���w�an�ted�clauses�b�Gefore�they�are�used�to�generate�y�et�more�un-����Pw���an�ted�UUclauses,�������E�������P�generate�UUuseful�clauses�so�Goner,�������E�������P�use�UUuseful�clauses�so�Goner�than�they�w���ould�otherwise�b�e�used.��x䍑?Metho�Gds�UUdirected�to���w�ards�UUthese�ob��8jectiv���es�are�called�\strategies".����NIn��1962,�when�none�of�the�strategies�kno���wn�to�Gda�y�had�y�et�b�Geen�in�v�en�ted,����?the�a%follo���wing�problem�w�as�to�Go�dicult�for�automated�theorem�pro�ving:�In�a����?group,�ɂif��x�!;���x���=��e�ɂ�for�ev���ery��x�,�then�the�group�is�comm�utativ�e,�i.e.��z��Ҹ�!;�y�"�=���y�}���z����?�for�Y�ev���ery��y����and��z�p��.�T��*�o�Gda�y�this�is�trivial�(for�b�Goth�h�umans�and�computers).�It����?w���as��consideration�of�this�example�that�led�Larry�W��*�os�to�in�v�en�t�the�\set�of����?supp�Gort"��<strategy�[107],�whic���h�is�to�da���y�basic�to�the�organization�of�a�mo�dern����?theorem-pro���v�er.����NHere��*is�an�explanation�of�(one�v���ersion�of��)�this�strategy��*�.�Divide�the�axioms����?in���to���t�w�o�lists,�usable�and�set�of�supp�Gort�(sos).�Normally��*�,�sos�con�tains�the����?negation�Uof�the�desired�theorem�(that�is,�it�con���tains�the�\sp�Gecial�h�yp�Gothesis"����?of�B�the�theorem�and�the�negation�of�the�conclusion�of�the�theorem).�The�axioms����?of��the�theory�go�in���to�usable.�T��*�o�generate�new�clauses,�use�resolution�(or�a����?v��q�arian���t���of�resolution)�with�one�paren�t�from�sos�and�one�paren�t�from�usable.����?Sp�Gecically��*�,���pic���k�one�\giv�en�clause"�from�sos.�Mo�v�e�the�giv�en�clause�from�sos����?to�G�usable.�Then�mak���e�all�p�Gossible�inferences�using�the�giv�en�clause�as�one����?paren���t,��awith�the�other�paren�t�c�hosen�from�usable.�Add�the�new�conclusions����?(p�Gossibly�z�after�some�p�ost-pro�cessing)�to�the�sos�list.�Con���tin�ue,�z�c�ho�osing�z�a����?new�c giv���en�clause,�un�til�the�set�of�supp�Gort�b�ecomes�empt���y�or�a�con�tradiction����?is�UUderiv���ed.����NThe��lfollo���wing�fragmen�t�of�an�Otter�input�le�illustrate�the�c�hoice�of�sos����?and�Z<usable�in�the�example�men���tioned�ab�Go�v�e.�(Here����<x

cmtt10�f��means�the�group�op�Ger-����?ation,�UUand��g��is�the�in���v�erse.)��W��?�list(usable).����?x�?�=�x.�s�\%�equality����?f(e,x)�?�=�x.�Y?�\%�identity����?f(g(x),x)�?�=�e.�I�\%�inverse����?f(f(x,y),z)�?�=�f(x,f(y,z)).�
�\%associativity����?end_of_list.�����?list(sos).�����)t��c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������41����)�ō���>;��?�f(x,x)�?�=�e.�S��\%�special�hypothesis����?f(a,b)�?�!=�f(b,a).�4�\%�Denial�of�conclusion����?end_of_list.����?�Otter���nds�a�6-step�pro�Gof,�of�lev���el�4,�for�this�problem.�W��*�os,�George�Robin-����?son,��Yand�Carson�pro���v�ed��Y(ac�kno�wledging�in�v��q�aluable�assistance�from�J.�	z�A.����?Robinson)���[107]�that�the�appropriate�use�of�this�strategy�still�preserv���es�the����?refutation-completeness���prop�Gert���y;�that�is,�if�there�exists�a�pro�of�of�con���tradic-����?tion��from�the�form���ulas�in�usable�and�sos�together,�then�in�principle�it�can����?b�Ge�found�b���y�this�strategy��*�,�if�w�e�do�not�run�out�of�space�or�time�rst.�The����?h���yp�Gothesis��of�this�theorem�is�that�the�usable�list�m�ust�itself�b�Ge�satisable,����?i.e.��not�con���tain�a�con�tradiction.�That�will�normally�b�Ge�so�b�ecause�w���e�put�the����?denial�UUof�the�conclusion�in���to�sos.��@@��NAnother��	w���a�y�of�trying�to�generate�useful�form�ulas�so�Goner,�or�to�a�v�oid����?generating�֯useless�form���ulas,�is�to�in�v�en�t�and�use�new�rules�of�inference.�Quite����?a���n���um�b�Ger�of�v��q�ariations�of�resolution�ha�v�e�b�Geen�in�tro�Gduced�and�sho�wn�to����?b�Ge��puseful,�and�v��q�arious�theorems�ha���v�e��pb�een�pro���v�ed��pab�out�whether�refutation����?completeness��Tis�preserv���ed�using�v��q�arious�com�binations�of�the�rules.�F��*�or�an����?o���v�erview��fof�these�matters,�see�[109].�F��*�or�additional�details�and�man���y�exam-����?ples,�bpsee�[108].�No���w�ada�ys,�bpthe�user�of�a�theorem-pro���v�er�bpcan�t���ypically�sp�Gecify����?the��Binference�rules�to�b�Ge�used�on�a�particular�problem,�and�ma���y�try�v��q�arious����?c���hoices;��Qwhile�there�ma�y�b�Ge�a�default�selection�(Otter�has�an�\autonomous����?mo�Gde"),�UUexp�ertise�in�the�selection�of�inference�rules�is�often�helpful.��@@��NAnother��common�w���a�y��of�trying�to�generate�useful�form���ulas�so�Goner�is�to����?simply��+thro���w�out�\useless"�form�ulas�as�so�Gon�as�they�are�generated,�instead����?of���putting�them�in�sos�for�further�pro�Gcessing.�F��*�or�example,�if�a�form���ula�is�a����?substitution�=�instance�of�a�form���ula�already�pro�v�ed,�there�is�no�use�k�eeping�it.����?If��'y���ou�feel�(or�hop�Ge)�that�the�pro�of�will�not�require�form���ulas�longer�than�20����?sym���b�Gols,�� wh�y�not�thro�w�out�longer�form�ulas�as�so�Gon�as�they�are�generated?����?More�j'generally��*�,�w���e�can�assign�\w�eigh�ts"�to�form�ulas.�The�simplest�\w�eigh�t"����?is���just�the�length�(total�n���um�b�Ger���of�sym���b�ols),�but�more�complex�w���eigh�tings����?are�tOp�Gossible.�Then�w���e�can�sp�ecify�the�maxim���um�w�eigh�t�of�form�ulas�to�b�Ge����?retained.�>�Of�course,�doing�so�destro���ys�refutation�completeness,�but�it�ma�y����?also���enable�us�to�nd�a�pro�Gof�that�w���ould�otherwise�nev�er�ha�v�e�b�Geen�pro�duced����?in�
Xour�lifetimes.�If�w���e�do�not�nd�a�pro�Gof,�w�e�can�alw�a�ys�try�again�with�a����?larger�UUmaxim���um�w�eigh�t.��@@��NThe��wdescription�of�the�sos�strategy�ab�Go���v�e��wlea�v�es�sev�eral�things�impre-����?cise:��ho���w�do�w�e�\select"�a�form�ula�from�sos�to�b�Ge�the�next�giv�en�form�ula?����?What�gOis�the�nature�of�the�\p�Gost-pro�cessing"?�gOThese�questions�ha���v�e�gOin�terest-����?ing��nansw���ers,�and�the�answ�ers�are�not�unique.�There�are�dieren�t�strategies����?addressing���these�questions.�Otter�has�man���y�user-con�trollable�parameters�that����?in
uence��these�kinds�of�things.�There�are�so�man���y�parameters�that�running����?Otter��is�more�of�an�art�than�a�science.�F��*�or�a�more�detailed�description�of�the����?basic��$algorithm�of�Otter,�see�[109],�p.�94,�where�the�program's�main�lo�Gop�is����?summarized�UUon�a�single�page.�����*��c��>;����?�42���`-	Mic��9hael�TBeeson����)��>;��N�It���has�no���w�b�Geen�nearly�fort�y�y�ears�since�the�in�v�en�tion�of�the�set�of�supp�Gort����?strategy��*�,��eand�the�general�approac���h�to�theorem�pro�ving�describ�Ged�ab�o���v�e��ehas����?not�s?c���hanged,�nor�has�an�y�comp�Geting�approac�h�met�with�as�m�uc�h�success.����?Ov���er���that�fort�y�y�ears,�the�approac�h�has�b�Geen�rened�b�y�the�dev�elopmen�t�of����?man���y�F�in�teresting�strategies.�The�skillful�application�of�these�strategies�has����?led�sto�the�solution�of�more�and�more�dicult�problems,�some�of�whic���h�w�ere����?previously���unsolv���ed.�An�impressiv�e�list�of�suc�h�problems�solv�ed�just�in�the����?last�=kcouple�of�y���ears�is�giv�en�in�[41].����^��34����Q�These�problems�are�in�highly�tec�hnical����?areas,�p�so�it�is�dicult�to�list�and�explain�them�in�a�surv���ey�article.�T��*�o�giv�e�a����?taste�6~of�this�kind�of�researc���h,�w�e�shall�explain�just�one�of�the�areas�in�v�olv�ed:����?prop�Gositional�
%logic.�Y��*�ou�ma���y�think�that�prop�ositional�logic�is�trivial.�After����?all,��y���ou�kno�w�ho�w�to�decide�the�v��q�alidit�y�of�an�y�prop�Gosition�b�y�the�metho�Gd����?of�R�truth�tables.�Therefore�it�is�rst�necessary�to�con���vince�y�ou�that�this�is�an����?area�p�with�in���teresting�questions.�W��*�e�write��i�(�x;���y�[ٲ)�for�\�x��implies��y��",�and��n�(�x�)����?for���\not��x�".�Since�\and"�and�\or"�can�b�Ge�dened�in�terms�of�implication�and����?negation,�Fw���e�will�restrict�ourselv�es�to�the�connectiv�es��i��and��n�.�The�P�olish����?logician��Jan��WZ ���8L��,�uk��q�asiewicz�(1878-1956)�in���tro�Gduced�the�follo�wing�axioms�for����?prop�Gositional�UUlogic:��������^�i�(�i�(�x;���y�[ٲ)�;�i�(�i�(�y�;�z�p��)�;�i�(�x;�z��)))������(1)��������^�i�(�i�(�n�(�x�)�;���x�)�;�x�)������(2)��������^�i�(�x;���i�(�n�(�x�)�;�y�[ٲ))������(3)������?T��*�o���w���ork�with�these�axioms�in�Otter,�w�e�use�the�predicate��P�c��(�x�)�to�mean�\�x����?�is�UUpro���v��q�able".�W��*�e�then�put�in�to�the�usable�list,����?�P(i(i(x,y),i(i(y,z),i(x,z)))).����?P(i(i(n(x),x),x)).����?P(i(x,i(n(x),y))).����?-P(x)�?�|�-P(i(x,y))�|�P(y).����?�No���w���to�ask,�for�example,�whether��i�(�x;���x�)�is�a�theorem,�w�e�put��-P(i(c,c))����?�in���to���list(sos).�That�is,�w�e�put�in�the�negation�of�the�assertion�that��i�(�c;���c�)�is����?pro���v��q�able.�The�steps�tak�en�b�y�resolution�corresp�Gond�to�the�rule�of�\detac�h-����?men���t"��used�b�y�logicians:�T��*�o�deduce�a�new�form�ula�from��A��and��i�(�B��q;���C���),�mak�e����?a��$substitution����so�that��A���^����QT�=��p�B���q��^����U�.�Then�y���ou�can�deduce��C�����^����P�.����^��35���_
�Wh�y�do�w�e��?�"}�ff8�ϟ
L͍������UZ��-=�34�����	?��If�fothe�non-exp�A�ert�user�lo�oks�at�the�list�giv��9en�in�[41]�of�dicult�problems�solv�ed����	?�using�СOtter,�he�or�she�will�v��9ery�lik�ely�not�b�A�e�able�in�a�straigh�tforw�ard�manner�to����	?�get��Otter�to�pro��9v�e��these�theorems.�He�or�she�will�ha��9v�e��to�go�to�the�appropriate�w��9eb����	?�site��#and�get�the�input�les�prepared�b��9y�the�exp�A�erts,�sp�ecifying�the�inference�rules����	?�and���parameters�con��9trolling�the�searc�h�strategies.�As�the�authors�state,�they�do����	?�not�Mbha��9v�e�a�single�uniform�strategy�that�will�enable�Otter�to�solv�e�all�these�dicult����	?�problems,���and�a�lot�of�h��9uman�trial�and�error�has�gone�in�to�the�construction�of����	?�those�Tinput�les.���������UZ��-=�35�����	?��T��:�ec��9hnically�,���since�a�theorem�pro�v�er�alw�a�ys�uses�the��most���gener��al��unier,�it�corre-����	?�sp�A�onds�bLto�the�rule�kno��9wn�as�\condensed�detac�hmen�t",�in�whic�h�only�most�general����	?�substitutions�Tare�allo��9w�ed.�����+���c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������43����)�ō���>;��?�need�
*the�predicate��P�c��?�Because�w���e�are�in�terested�in�pro�Gofs�from�L1{L3�us-����?ing�h�condensed�detac���hmen�t;�h��P�̉�is�used�to�force�the�theorem�pro���v�er�h�to�imitate����?that���rule.�W��*�e�are�not�just�in���terested�in�v�erifying�tautologies,�but�in�nding����?pro�Gofs��nfrom�the�sp�ecic�axioms�L1{L3.�No���w,�the�reader�is�in�vited�to�try�to����?pro���v�e���i�(�x;���x�)�from�the�axioms�L1{L3.�This�should�b�Ge�enough�to�con���vince�y�ou����?that��the�eld�is�not�trivial.�Other�axiom�systems�for�prop�Gositional�logic�w���ere����?giv���en�8b�y�F��*�rege,�b�y�Hilb�Gert,�and�b�y��~� ���8L��	T9uk��q�asievic�h.�(See�the�w�onderful�app�Gendix����?in�0*[83],�where�these�and�man���y�other�axiom�systems�are�listed.�The�questions����?then�y
arise�ab�Gout�the�equiv��q�alence�of�these�axiom�systems.�W��*�e�w���an�t�y
pro�ofs�of����?eac���h���of�these�axiom�systems�from�eac�h�of�the�others.�The�app�Gendix�of�[109]����?(pp.�Zh554-55)�lists�Otter�pro�Gofs�of�the�some�of�these�systems�from�L1-L3.�F��*�or����?example,��the�rst�axiom�in�F��*�rege's�system�is��i�(�x;���n�(�n�(�x�))).�Go�ahead,�John����?Henry:�UUtry�to�pro���v�e�UUit�from�L1-L3�using�p�Gencil�and�pap�er.��
n���NOne��t���yp�Ge�of�in�teresting�question�studied�b�y�logicians�in�the�1930s�through����?1950s{and�u'resumed�again�to�Gda���y�with�the�aid�of�automated�reasoning{w�as����?this:���giv���en�a�theory��T�XU�dened�b�y�sev�eral�axioms,�can�w�e�nd�a�\single�ax-����?iom"�Nufor��T�c��?�That�is,�a�single�form���ula�from�whic�h�all�the�axioms�of��T���can�b�Ge����?deriv���ed.��If�so,�what�is�the�shortest�p�Gossible�suc�h�axiom?�This�t�yp�Ge�of�question����?has���b�Geen�attac���k�ed���using�Otter�for�a�large�n���um�b�er���of�dieren���t�systems,�includ-����?ing���v��q�arious�logics,�group�theories,�and�recen���tly��*�,�lattice�theory�.�F�or�example,����?\equiv��q�alen���tial�j&calculus"�is�the�logical�theory�of�bi-implication�(if�and�only����?if��).�F�It�can�b�Ge�represen���ted�using��e�(�x;���y�[ٲ)�instead�of��i�(�x;�y�[ٲ),�and�treated�using�a����?\pro���v��q�abilit�y���predicate"��P��d�as�ab�Go���v�e.���See�[113]�for�an�example�of�an�Otter�pro�of����?that��settled�a�long-op�Gen�question�in�equiv��q�alen���tial�calculus,�namely��*�,�whether����?a�j�certain�form���ula�X�CB�j�is�a�single�axiom�for�this�theory��*�.�This�is�p�Gerhaps�the����?most�=�recen���t�example�of�a�theorem�that�has�b�Geen�pro�v�ed�for�the�rst�time�b�y����?a��computer.�Before�it�w���as�pro�v�ed�(in�April,�2002),�p�Geople�w�ere�not�willing�to����?giv���e�UUo�Gdds�either�w�a�y�on�the�question.��')����?�14��ZPro�`ofs��In��v�olving�Sets,�F���unctions,�and�Num�b�`ers��)���?�If���w���e�examine�a�textb�Go�ok���for�an�in�tro�Gductory�course�in�abstract�algebra,����?suc���h��4as�[56],�w�e�nd�that�only�ab�Gout�ten�p�ercen���t�of�the�problems�can�b�e����?form���ulated�&Rin�the�rst-order�languages�of�groups,�rings,�etc.�The�rest�in�v�olv�e����?subgroups,���subrings,�homomorphisms,�and/or�natural�n���um�b�Gers.���F��*�or�example,����?one��of�the�rst�theorems�in�group�theory�is�Lagrange's�theorem:�if��H����is����?a���subgroup�of�a�nite�group��G�,�then�the�order�of��H��˲(the�n���um�b�Ger���of�its����?elemen���ts)�צdivides�the�order�of��G�.�Here�w�e�need�natural�n�um�b�Gers�to�dene�the����?order,��and�a�bit�of�n���um�b�Ger��theory�to�dene�\divides";�w���e�need�the�concept�of����?subgroup,�� and�the�pro�Gof�in���v�olv�es�� constructing�a�function�to�put��H���in�one-one����?corresp�Gondence��with�the�coset��H��a�,�namely��*�,��x���7!��xa�.��A���t�presen�t,�no�theorem-����?pro���ving��eprogram�has�ev�er�generated�a�pro�Gof�of�Lagrange's�theorem,�ev�en����?though�
�the�pro�Gof�is�v���ery�short�and�simple.�The�obstacle�is�the�mingling�of�����,���c��>;����?�44���`-	Mic��9hael�TBeeson����)��>;��?�elemen���ts,���subgroups,�mappings,�and�natural�n�um�b�Gers.����^��36���L��The�presen�t�p�Go�w�er����?of��xautomated�theorem�pro���v�ers��xhas�yielded�results�only�in�theories�based�on����?equalit���y���and�a�few�op�Gerations�or�in�other�v�ery�simple�theories.�A�t�least�half����?of��undergraduate�mathematics�should�come�within�the�scop�Ge�of�automated����?pro�Gof���generation,�if�w���e�are�able�to�add�in�a�relativ�ely�small�abilit�y�to�deal�with����?sets,���n���um�b�Gers,�and�functions.�W��*�e�do�not�(usually)�need�sets�of�sets,�or�sets����?of��sets�of�sets,�and�the�lik���e.�Nor�do�w�e�usually�need�functions�of�functions,����?except��*sp�Gecial�functions�of�functions�lik���e�the�deriv��q�ativ�e�op�Gerator.�If�w�e�add�to����?a�1�rst-order�theory�some�v��q�ariables�for�sets�(of�the�ob��8jects�of�the�theory)�and����?functions��P(from�ob��8jects�to�ob�jects),�w���e�ha�v�e�what�is�kno�wn�as�a�second-order����?theory��*�.�PThe�lam���b�Gda-calculus�can�b�e�used�to�dene�functions,�and�sets�can�b�e����?regarded���as�Bo�Golean-v��q�alued�functions.�The�author's�curren���t�researc�h�in�v�olv�es����?adding�~�capabilities�to�the�existing,�widely-used�theorem�pro���v�er�~�Otter�to�assist����?it���in�handling�second-order�theories,�without�in���terfering�with�its�rst-order����?capabilities.����^��37���
}�Sp�Gecically��*�,���a�new�second-order�unication�algorithm�[14,15],����?has��b�Geen�added�to�Otter,�and�will�b�e�impro���v�ed��and�applied.�Preliminary����?results,�UUand�the�direction�of�the�researc���h�program,�are�describ�Ged�in�[16].���[��NOne��ma���y�ob��8ject�to�the�use�of�second-order�logic,�and�indeed�to�the�whole����?idea��Nof�a�\taxonom���y"�of�mathematics,�on�the�grounds�of�the�univ�ersalit�y�of����?set��theory��*�.�Let�us�b�Gegin�b���y�stating�the�ob��8jection�clearly�.�Set�theory�is�a�\sim-����?ple��utheory",�with�one�relation�sym���b�Gol�for�mem�b�Gership�and�a�small�n�um�b�Ger�of����?axioms.��/T��*�rue,�one�of�the�\axioms"�of�ZF��set�theory�is�an�innite�sc���hema,�with����?one�dOinstance�for�eac���h�form�ula�of�the�language;�but�there�is�another�form�ula-����?tion��of�set�theory��*�,�G����odel{Berna���ys�set�theory�(GB),�whic�h�has�a�small�nite����?n���um�b�Ger��tof�axioms.�In�GB,�v��q�ariables�range�o���v�er��tclasses,�and�sets�are�dened�as����?classes�j=whic���h�b�Gelong�to�some�other�class.�(The�idea�is�that�prop�erties�dene����?classes,�	�but�not�ev���ery�class�is�a�set|w�e�escap�Ge�the�Russell�parado�x�in�GB����?b�Gecause�<the�Russell�class�is�a�class,�but�not�a�set.)�Because�of�the�p�ossibilit���y����?of�%�form���ulating�set�theory�in�this�w�a�y�as�a�simple�theory��*�,�the�taxonom�y�giv�en����?ab�Go���v�e�LIcollapses{all�of�mathematics�is�con���tained�in�a�single�simple�rst-order����?theory��*�.��kNo���w�for�some�relev��q�an�t�history:�a�cen�tury�ago,�this�program�for�the����?foundations�qOof�mathematics�w���as�laid�out,�but�in�the�middle�t�w�en�tieth�cen�tury��*�,����?the��Bourbaki�sc���ho�Gol�prev��q�ailed,�at�least�in�practice,�organizing�mathematics����?according���to�the�\man���y�small�theories"�program.�A�t�presen�t,�most�w�ork�in��?�
՚�ff8�ϟ
L͍������UZ��-=�36�����	?��A�l�pro�A�of�l�of�Lagrange's�theorem�dev��9elop�ed�b��9y�a�h�uman�has�b�A�een�c�hec�k�ed�b�y�the����	?�computer��program�A��9CL2,�see�[116].�That�pro�A�of�is�not�the�ordinary�pro�of,�but����	?�instead��ypro�A�ceeds�b��9y�mathematical�induction�on�the�order�of�the�group.�The�ordi-����	?�nary�.�pro�A�of�has�b�een�pro�of-c��9hec�k�ed�.�using�HOL�.�[58].�That�pap�er�also�presen��9ts�an����	?�in��9teresting�R\theory�hierarc�h�y"�sho�wing�exactly�what�is�needed.�It�has�also�b�A�een����	?�c��9hec�k�ed�Tin�Mizar�[99].���������UZ��-=�37�����	?��Of�_@course,�second-order�and�(wh��9y�not?)�higher-order�theorem�pro�ving�has�b�A�een����	?�in��existence�for�a�long�time,�and�there�are�ev��9en�whole�conferences�dev�oted�to�the����	?�sub���ject,���e.g.�TPHOL�à(Theorem�Pro��9ving�in�Higher-Order�Logic).�It�seems�that����	?�most��wof�this�researc��9h�is�not�directed�to�w�ards�pro�ving�new�theorems,�so�it�has�not����	?�b�A�een�Tdiscussed�in�this�pap�er.�����-���c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������45����)�ō���>;��?�automated�u'deduction�is�based�on�the�\small�theories"�approac���h,�although����?one��<bra���v�e�man,�Belinfan�te,�has�b�Geen�pro�ceeding�for�man���y�y�ears�to�dev�elop����?computerized��Mpro�Gofs�based�on�GB��set�theory�[17,18].�F��*�ollo���wing�this�approac�h,����?he��has�enabled�Otter�to�pro���v�e��more�than�1000�theorems�in�set�theory{but����?he��still�is�not�up�to�Lagrange's�theorem,�or�ev���en�close.�Belinfan�te�built�on����?the�=�pioneering�w���ork�of�Quaife�[94].�Finally:�the�answ�er�to�the�ob��8jection�is����?simply��nthat�it�is�to�Go�complex�to�regard�n���um�b�ers��nand�functions�as�built�out����?of�� sets.�No�mathematician�do�Ges�so�in�ev���eryda�y�� practice,�and�neither�should����?automated�UUdeduction�when�the�aim�is�to�someda���y�pro�v�e�new�theorems.����^��38����W@��N�On��the�other�hand,�one�ma���y�tak�e�the�opp�Gosite�view�and�sa�y�that,�b�Gecause����?of���the�diculties�of�dev���eloping�mathematics�within�set�theory��*�,�one�should�use����?\higher-order��logic".�This�has�b�Geen�the�view�of�man���y�of�the�\pro�of-c���hec�king"����?pro��8jects,��Qand�they�ha���v�e��Qb�Geen�successful�in�c���hec�king��Qpro�ofs�of�man���y�fairly����?complicated��?theorems.�A��)prop�Ger�review�of�this�w���ork�w�ould�double�the�length����?of���this�article,�so�w���e�m�ust�forego�it.�The�in�terested�reader�can�consult�[105]����?for���a�list�of�fteen�pro�Gof�c���hec�k�ers���and�pro�of�nders,�as�w���ell�as�references�to����?further�UUinformation.��!�A���?�15��ZConclusion���A��?�Alan���T��*�uring�wrote�a�seminal�pap�Ger�[101]�in�whic���h�he�raised�the�question����?\Can���mac���hines�think?".����^��39���2˲After�discussing�v��q�arious�examples,�suc�h�as�c�hess,����?m���usical��Mcomp�Gosition,�and�theorem�pro�ving,�he�then�form�ulated�the�\T��*�uring����?test"��as�a�replacemen���t�for�that�imprecise�question.�In�the�T��*�uring�test,�a�com-����?puter�tOtries�to�deceiv���e�a�h�uman�in�to�thinking�that�the�computer�is�h�uman.����^��40������?�Of�	Hcourse�in�the�foreseeable�future�it�will�b�Ge�to�o�dicult�for�a�single�com-����?puter�R�to�b�Ge�able�to�reac���h�h�uman�lev�el�in�man�y�areas�sim�ultaneously;�but�w�e����?migh���t�z�consider�restricted�v�ersions�of�the�T��*�uring�test�for�sp�Gecic�areas�of�en-����?dea���v�or.��5As�men���tioned�in�the�in�tro�Gduction,�the�T��*�uring�test�has�already�b�een����?passed�r�for�m���usical�comp�Gosition:�Da�vid�Cop�Ge�has�written�a�program�EMI�rf(pro-����?nounced�A�\Emm���y",�for�Exp�Gerimen�ts�in�Musical�In�telligence)���whic�h�pro�Gduces����?m���usic�@2that�regularly�fo�Gols�sophisticated�audiences{at�least,�it�did�un�til�Cop�Ge����?stopp�Ged���conducting�\The�T��*�est"�at�his�concerts|stopp�ed�b�ecause�the�exp�erts����?w���ere�vto�Go�em�barrassed.�Since�Cop�Ge�is�a�comp�oser�rather�than�a�computer�sci-����?en���tist,���he�presen�ts�his�results�primarily�at�concerts�rather�than�conferences.��?�	�.�ff8�ϟ
L͍������UZ��-=�38�����	?��F��:�or��[the�record,�Belinfan��9te�agrees�with�this�statemen�t.�His�aim,�ho�w�ev�er,�is�foun-����	?�dational.��As�a�b�A�o��9y��:�,�he�to�ok��Principia�1�Mathematic��a��from�his�ph��9ysicist�father's����	?�b�A�o�okshelf���and�said�to�himself,�\Someda��9y�I'm�going�to�c�hec�k�if�all�these�pro�A�ofs�are����	?�really�Trigh��9t!".�That�spirit�still�animates�his�w�ork.���������UZ��-=�39�����	?��Lik��9e�
�Stanley�Jev�on's�pap�A�er,�an�original�cop�y�of�this�journal�article�no�w�is�priced����	?�at�T$2000.���������UZ��-=�40�����	?��A�� more��/detailed�discussion�of�the�T��:�uring�test�can�b�A�e�found�in�T�uring's�pap�A�er��op.����	?�cit.�ZQ�or�in�an��9y�mo�A�dern�textb�o�ok�on�articial�in��9telligence;�the�idea�of�a�computer����	?�trying�Tto�app�A�ear�h��9uman�is�enough�for�our�purp�oses.�����.�Ϡc�ō���>;����?�46���`-	Mic��9hael�TBeeson����)�ō���>;��?�I�sRheard�s�a�sev���en-piece�c�ham�b�Ger�orc�hestra�p�Gerform�the�Eigh�th�Branden�burg����?Concerto�b�(comp�Gosed�b���y�EMI).�(Bac�h�comp�Gosed�the�rst�sev�en�Branden�burg����?Concertos.)��
�m��NIn�K�theorem-pro���ving,�as�in�articial�in�telligence,�there�w�as�initially�a�divi-����?sion�{8b�Get���w�een�those�who�though�t�computers�should�b�Ge�programmed�to�\think"����?lik���e�a�h�umans�and�those�who�fa�v�ored�a�more�computational�approac�h.�Should����?w���e��*try�to�nd�\heuristics"�(rules�of�th�um�b)�to�guide�a�computer's�eorts�to����?nd�� a�pro�Gof,�or�pla���y�a�game�of�c�hess,�or�comp�Gose�a�piece�of�m�usic?�Or�should����?w���e��ujust�giv�e�the�computer�the�rules�and�a�simple�algorithm�and�rely�on�the����?p�Go���w�er��qof�silicon�c���hips?�It�is�in�teresting�to�compare�computerized�c�hess�and����?computerized�=[theorem-pro���ving�in�this�resp�Gect.�Both�can�b�e�view���ed�as�searc�h����?problems:��c���hess�is�a�searc�h�organized�b�y�\if�I��mak�e�mo�v�e��x����1��f �and�he�mak�es����?mo���v�e�t�y����1���v�and�then�I�s�mak���e�mo�v�e��x����2���v�and�he�mak�es�mo�v�e��y����2���v�and�then���:���:�:����";�w�e����?searc���h�1�the�v��q�arious�p�Gossibilities,�up�to�a�certain�\lev�el"�or�\depth",�and�then,����?for��eac���h�sequence�of�p�Gossible�mo�v�es,�w�e�score�the�situation.�Then�w�e�pic�k�our����?mo���v�e,�K�using�a�\max-min"�algorithm.�As�in�theorem�pro���ving,�the�dicult�y�is����?to�:\\prune�the�searc���h�tree"�to�a�v�oid�getting�sw�amp�Ged�b�y�the�consideration����?of�n�useless�mo���v�es.�n�In�b�Goth�endea���v�ors,�n�theorem�pro���ving�and�c�hess,�one�feels����?that�{exp�Gert�h���uman�b�eings�ha���v�e�{subtle�and�p�o���w�erful�{metho�ds.�Chess�pla���y�ers����?analyze���far��few���er�p�Gossibilities�than�c�hess�programs�do,�and�those�go�Go�d��p�ossi-����?bilities���are�analyzed�deep�Ger.�One�feels�that�the�same�ma���y�b�e�true�of�mathe-����?maticians.��In�c���hess�programs,�\kno�wledge"�ab�Gout�op�enings�and�end�games�is����?stored�U�in�a�database�and�consulted�when�appropriate.�But�in�the�mid-game,����?ev���ery��eort�in�c�hess�programming�to�use�more�sp�Gecialized�c�hess�kno�wledge����?and��less�searc���h�has�failed.�The�computer�time�is�b�Getter�sp�en���t�searc�hing�one����?mo���v�e��deep�Ger.�On�the�other�hand,�the�game�of�go�is�pla���y�ed��at�only�sligh���tly����?ab�Go���v�e��the�b�eginner�lev���el�b�y�computers.�The�b�Goard�is�19�b�y�19�instead�of�8�b�y����?8,��3and�there�are�more�pieces;�the�com���binatorial�explosion�is�to�Go�deadly�for����?computers�UUto�adv��q�ance�m���uc�h�UUb�Gey�ond�this�lev�el�at�presen�t.����^��41����
�m��N�Similarly��*�,��Oin�mathematics,�so�far�at�least,�searc���h�has�pro�v�ed�the�most����?fruitful���general�tec���hnique.�One�can�view�computer�algebra�and�computerized����?decision��pro�Gcedures,�suc���h�as�quan�tier�elimination�or�Wilf�and�Zeilb�Gerger's����?decision���pro�Gcedure�for�com���binatorial�sums,�as�w�a�ys�of�em�b�Gedding�mathemat-����?ical�v�kno���wledge�in�computer�programs.�Where�they�are�applicable,�they�pla�y����?an�u~indisp�Gensable�role,�analogous�to�the�role�of�op�ening�b�o�oks�and�end�game����?databases��$in�c���hess.�In�areas�of�mathematics�in�whic�h�it�is�dicult�to�bring����?kno���wledge��dto�b�Gear�(suc�h�as�elemen�tary�group�theory�or�prop�Gositional�logic)��?���ff8�ϟ
L͍������UZ��-=�41�����	?��The��*game�tree,�searc��9hing��n��mo�v�es�in�the�future,�has�ab�A�out��b���-=�;�cmmi6�n���a�no�des,�where�at����	?�a�9ncrude�appro��9ximation��b���=�8���-=�2��d�for�9nc�hess�and�19���-=�2��d�for�go.�So�the�ratio�is�(19�=�8)���-=�2�n���޻.����	?�T��:�aking�-��n��ƻ=�10�w��9e�get�more�than�2���-=�20���N�,�whic�h�is�ab�A�out�a�million:�go�is�a�million�times����	?�harder�@tthan�c��9hess.�On�the�other�hand,�computer�sp�A�eeds�and�memory�sizes�ha�v�e����	?�historically��&increased�exp�A�onen��9tially��:�,�doubling�ev�ery�18�mon�ths;�so�if�Mo�A�ore's�la�w����	?�con��9tin�ues�`�to�hold,�w��9e�migh�t�hop�A�e�that�go�programs�w�ould�p�A�erform�w�ell�enough����	?�in�T30�y��9ears,�ev�en�without�impro�v�emen�ts�in�the�programs.�����/�}�c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������47����)�ō���>;��?�b�Gecause��the�axioms�are�v���ery�simple�and�to�ols�from�outside�the�sub��8ject�area����?are��"not�applicable,�theorem-pro���ving�programs�can�outp�Gerform�h�uman�b�Geings,����?at�UUleast�sometimes,�just�as�in�c���hess.���A��NHo���w�9�did�the�trade-o�b�Get�w�een�high-sp�Geed�but�simple�computation�and����?heuristics��pla���y�out�in�the�area�of�m�usical�comp�Gosition?�The�answ�er�to�this����?question�V�is�quite�in���teresting,�and�ma�y�ha�v�e�implications�for�the�future�of�re-����?searc���h���in�theorem�pro�ving.�EMI���do�Ges�not�comp�ose�from�a�blank�slate.�T��*�o�use����?EMI,��y���ou�rst�decide�on�a�comp�Goser�to�b�e�imitated;�let's�sa���y�Bac�h.�Then,����?y���ou��,feed�EMI��sev�eral�comp�Gositions�b�y�Bac�h.�EMI��extracts�from�these�data����?a��\grammar"�of�m���usical�structures.�EMI��then�uses�this�grammar�to�gener-����?ate�R3a�new�comp�Gosition,�whic���h�will�b�e�p�erceiv���ed�as�\in�the�st�yle�of�Bac�h".����?The���selection�of�the�initial�database�calls�for�m���usical�exp�Gertise�and�for�exp�er-����?tise�v=with�EMI.�F��*�or�example,�to�comp�Gose�the�Eigh���th�Branden�burg�Concerto,����?Cop�Ge�z\c���hose�some�of�the�original�sev�en�Branden�burg�Concertos�and�a�couple����?of�=violin�concertos.�When�the�database�con���tained�only�the�sev�en�Branden-����?burg��Concertos,�the�resulting�comp�Gosition�seemed�to�o�\deriv��q�ativ���e",�and�ev�en����?con���tained�{�recognizable�phrases.�Y��*�et,�once�the�data�has�b�Geen�digested,�the����?program�QDw���orks�according�to�sp�Gecic,�precise�rules.�There�is�nothing�\heuris-����?tic"�fab�Gout�the�pro�cess.�A�6result�that�\lo�oks�h���uman"�has�b�een�ac���hiev�ed�fb�y����?computational�UUmeans.���A��NThis�Xuis�at�presen���t�not�true�of�most�pro�Gofs�pro�duced�b���y�most�theorem-����?pro���v�ers.��T��*�o�exhibit�a�simple�example,�consider�the�theorem�that�a�group����?in�\pwhic���h��x���^��2����=�}�1�for�all��x��is�comm�utativ�e.�A�\,h�uman�pro�Gof�migh�t�start�b�y����?substituting��!�uv��for��x�,�to�get�(�uv�[ٲ)���^��2���޲=�k1.�Multiplying�on�the�left�b���y��u��and�on����?the�N�righ���t�b�y��v����the�desired�result�is�immediate.�The�pro�Gof�that�a�theorem-����?pro���v�er�;�nds�is�m���uc�h�;�less�clev���er.�In�longer�pro�Gofs,�the�computer's�inh�uman����?st���yle��{stands�out�ev�en�more.�The�most�notable�feature�of�suc�h�pro�Gofs�is�that����?theorem�Ջpro���v�ers�nev�er�in�v�en�t�concepts�or�form�ulate�lemmas.�A��kpap�Ger�written����?b���y��a�mathematician�ma�y�ha�v�e�a�\main�theorem"�and�t�w�en�t�y�supp�Gorting����?lemmas.�k�The�pro�Gof�of�the�main�theorem�ma���y�b�e�quite�short,�but�it�relies�on����?the�a&preceding�lemmas.�Not�only�do�Ges�this�help�understanding,�but�the�lemmas����?ma���y���b�Ge�useful�in�other�con�texts.�The�most�p�Go�w�erful�presen�t-da�y�theorem����?pro���v�ers���nev�er�nd,�organize,�or�presen�t�their�pro�Gofs�in�this�w�a�y�(unless�led�to����?do�UUso�b���y�a�h�uman�after�a�failure�to�nd�the�pro�Gof�in�one�go).���A��NTheorem-pro���v�ers���of�the�future�should�b�Ge�able�to�in���v�en�t���terminology�and����?form��
denitions.�The�basis�of�their�abilit���y�to�do�this�should�b�Ge�an�underlying����?abilit���y�`)to�monitor�and�reason�ab�Gout�their�o�wn�deduction�pro�Gcess.�As�it�is����?no���w,��7h�umans�using�a�theorem�pro�v�er�monitor�the�output,�and�then�c�hange����?parameters��and�restart�the�job.�In�the�future,�this�kind�of�feedbac���k�should����?b�Ge���automated�and�dynamic,�so�that�the�parameters�of�a�run�can�b�e�altered����?(b���y���the�program�itself��)�while�the�run�is�in�progress.�With�this�capabilit�y�in����?hand,��cone�should�then�b�Ge�able�to�detect�candidates�for�\lemma"�status:�short����?form���ulas���that�are�used�sev�eral�times.�It�is�then�a�go�Go�d���idea�to�k�eep�an�ey�e�out����?for���further�deductions�similar�in�form�to�the�form���ulas�in�v�olv�ed�in�the�pro�Gof�����0⠠c�ō���>;����?�48���`-	Mic��9hael�TBeeson����)�ō���>;��?�of��Ethe�lemmas.����^��42���+�Giving�a�program�the�abilit���y�to�form�ulate�its�o�wn�lemmas����?dynamically�<Lmigh���t,�in�conjunction�with�the�abilit�y�to�mo�Gdify�the�criteria�for����?k���eeping��or�using�deduced�form�ulas,�enable�the�program�to�nd�pro�Gofs�that����?migh���t�UUotherwise�b�Ge�b�ey���ond�reac�h.������NSuc���h�7a�pro�v�er�migh�t�pro�Gduce�pro�ofs�that�lo�ok�more�h���uman.�The�in�v�es-����?tigation,��"the�st���yle�of�though�t,�and�the�dev�elopmen�t�of�the�theory�should����?eac���h�lo�Gok�more�lik�e�pro�Gofs�p�eople�pro�duce.�Searc���hing�w�ould�still�b�Ge�the�basis����?(not�v�heuristics),�but�the�result�w���ould�lo�Gok�less�lik�e�it�had�b�Geen�pro�duced�b���y����?P���oincar���Ge's�*&famous�mac���hine�that�tak�es�in�pigs�and�pro�Gduces�sausages.�This����?t���yp�Ge��lof�pro�v�er�w�ould�b�Ge�a�little�more�lik�e�EMI��Vthan�to�Gda�y's�theorem�pro�v�ers.����?One��Qmigh���t�ev�en�b�Ge�able�to�prime�it�with�sev�eral�pro�Gofs�b�y�the�famous�logician����?Meredith�c in�order�to�ha���v�e�c it�pro�Gduce�pro�ofs�in�prop�ositional�logic�in�Mered-����?ith's�P�st���yle,�m�uc�h�as�EMI�P�can�pro�Gduce�m�usic�in�the�st�yle�of�Bac�h�or�the�st�yle����?of���Scott�Joplin.�A���t�presen�t�this�is�rather�farfetc�hed,�as�there�is�nothing�lik�e����?the��\grammar"�of�Meredith's�pro�Gofs.�The�closest�appro���ximation�at�presen�t����?w���ould��b�Ge�to�tell�the�program�to�retain�deduced�clauses�similar�in�form�to�the����?lines�UUof�the�pro�Gofs�used�to�prime�the�program.������NW��*�e�j&do�not�exp�Gect,�ho���w�ev�er,�j&that�all�mac���hine-generated�pro�ofs�will�\lo�ok����?h���uman".�NkF��*�or�example,�there�exists�a�mac�hine-generated�pro�Gof�that�a�certain����?form���ula���is�a�single�axiom�for�groups�satisfying��x���^��19��Ѯ�=�X�1�for�all��x�.�This�pro�Gof����?con���tains�UUa�form�ula�715�sym�b�Gols�long.�No�h�uman�will�nd�that�pro�Gof.����NRemem���b�Ger:��Dthe�question�whether�mac�hines�can�think�is�lik�e�the�ques-����?tion��|whether�submarines�can�swim.�W��*�e�exp�Gect�mac���hine�mathematics�to�b�e����?dieren���t�,�from�h�uman�mathematics{but�it�seems�a�safe�prediction�that�the����?t���w�en�t�y-rst��cen�tury�will�see�some�amazing�ac�hiev�emen�ts�in�mac�hine�mathe-����?matics.��"����?�References��������?�1.���K1�Agarw��9al,���R.,�Kapur,�D.,�Musser,�D.�R.,�and�Nie,�X.,�T��:�ecton�pro�A�of�system.�In:����K1�Bo�A�ok,�U�R.�(ed.)��Pr��o�c.��<F��J�ourth�International�Confer��enc�e��<on�R��ewriting�T�e��chniques����K1�and�N<Applic��ations,�Milan,�Italy,�1991.�T�LNCS��488�,�Springer-V��:�erlag�(1991).��������?2.���K1�App�A�el,���K.,�and�Hak��9en,�W.,�Ev�ery�planar�map�is�four�colorable.�P�art�I.�Disc�harg-����K1�ing,�T�Il�x�linois�N<J.�Math.��21��429{490,�1977.������?3.���K1�App�A�el,��K.,�Hak��9en,�W.�Hak�en,�and�Ko�A�c�h,�J.,�Ev�ery�planar�map�is�four�colorable.����K1�P��9art�TI�A�I.�Reducibilit�y��:�,��Il�x�linois�N<J.�Math.�T�21�491{567,�1977.������?4.���K1�Arnon,�S�D.,�and�Buc��9h�b�A�erger,�S�B.,��A�Îlgorithms���in�R��e�al���A�lgebr�aic���Ge�ometry�,�S�Acad-����K1�emic�\*Press,�London�(1988).�Reprin��9ted�from��J.��Symb��olic�Computation�\*�5��Num�b�A�ers����K1�1�Tand�2,�1988.������?5.���K1�Baader,�TF.,�and�Sn��9yder,�W.,�Unication�theory��:�,�in�[90],�pp.�435-534.��?�
�]�ff8�ϟ
L͍������UZ��-=�42�����	?��There�X�are�v��|rarious�p�A�ossible�notions�of�\similar�in�form"�that�migh��9t�b�e�used.�F��:�or����	?�example,��Ione�idea�is�to�call�form��9ulas�similar�if�they�b�A�ecome�the�same�when�all����	?�v��|rariables���are�replaced�with�the�same�letter.�This�notion�is�b�A�ehind�a�successful����	?�strategy�Tcalled�\resonance".�[109],�p.�457.�����1�ߠc�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������49����)�ō���>;����?6.���K1�Barendregt,�t�H.,��The��-L��amb�da�Calculus:�Its�Syntax�and�Semantics�,�t�Studies�in����K1�Logic�g�and�the�F��:�oundations�of�Mathematics��103�,�Elsevier�Science�Ltd.�Revised����K1�edition�T(Octob�A�er�1984).��������?7.���K1�Barendregt,�/�H.,�and�Geuv��9ers,�H.,�Pro�A�of-Assistan�ts�Using�Dep�A�enden�t�T�yp�A�e�Sys-����K1�tems,�zVin:�Robinson,�A.,�and�V��:�oronk��9o�v,�zVA.�(eds.),��Handb��o�ok��of�A�Îutomate��d�R�e�a-����K1�soning,�N<vol.�II�,�Tpp.�1151-1238.�Elsevier�Science�(2001).������?8.���K1�Basu,�j�S.,�P��9ollac�k,�j�R.,�and�Ro��9y��:�,�M.�F.,�On�the�Com�binatorial�and�Algebraic����K1�Complexit��9y��of�Quan�tier�Elimination,��Journal��sof�the�A�ÎCM���43���(6)�1002{1046,����K1�1996.������?9.���K1�Beeson,�naM.,�Some�applications�of�Gen��9tzen's�pro�A�of�theory�to�automated�deduc-����K1�tion,���in�P��:�.�Sc��9hro�A�eder-Heister�(ed.),��Extensions��Nof�L��o�gic��NPr�o�gr�amming�,���Lecture����K1�Notes�Tin�Computer�Science��475��101-156,�Springer-V��:�erlag�(1991).������?10.���O��Beeson,��IM.,���q[�		cmsl9�Mathp�A�ert�.�:�Computer�supp�ort�for�learning�algebra,�trigonometry��:�,����K1�and��!calculus,�in:�A.�V��:�oronk��9o�v��!(ed.),�Logic�Programming�and�Automated�Rea-����K1�soning,�TLecture�Notes�in�Articial�In��9telligence�624,�Springer-V��:�erlag�(1992).������?11.���O��Beeson,��.M.,��Mathp��ert�	-Calculus�Assistant�.��.This�soft��9w�are��.pro�A�duct�w��9as�published����K1�in��July��:�,�1997�b��9y�Mathp�A�ert�Systems,�San�ta�Clara,�CA.�See�www.mathxp�A�ert.com����K1�to�Tdo��9wnload�a�trial�cop�y��:�.������?12.���O��Beeson,��RM.,�Automatic�generation�of�epsilon-delta�pro�A�ofs�of�con��9tin�uit�y��:�,��Rin:�Cal-����K1�met,��Jacques,�and�Plaza,�Jan�(eds.)��A�Îrticial��Intel�x�ligenc��e�and�Symb�olic�Com-����K1�putation:��International�Confer��enc�e��AISC-98,�Plattsbur��gh,�New�Y��J�ork,�USA,�Sep-����K1�temb��er�N<1998�Pr�o�c�e�e�dings�,�Tpp.�67-83.�Springer-V��:�erlag�(1998).������?13.���O��Beeson,�T�M.,�Automatic�generation�of�a�pro�A�of�of�the�irrationalit��9y�of�e,��Journal����K1�of�N<Symb��olic�Computation�T�32�,�No.�4�(2001),�pp.�333-349.������?14.���O��Beeson,�`�M.,�Unication�in�Lam��9b�A�da�Calculus�with�if-then-else,�in:�Kirc�hner,����K1�C.,��9and�Kirc��9hner,�H.�(eds.),��A�Îutomate��d���De�duction-CADE-15.�15th�International����K1�Confer��enc�e��on�A�Îutomate��d�De�duction,�Lindau,�Germany,�July�1998�Pr�o�c�e�e�dings�,����K1�pp.�fc96-111,�Lecture�Notes�in�Articial�In��9telligence��1421�,�Springer-V��:�erlag�(1998).������?15.���O��Beeson,�w�M.,�A�w�second-order�theorem�pro��9v�er�w�applied�to�circumscription,�in:�Gor,����K1�R.,�N�Leitsc��9h,�A.,�and�Nipk�o�w,�T.�(eds.),��A�Îutomate��d���R�e�asoning,�First�International����K1�Joint�
kConfer��enc�e,�IJCAR�
:2001,�Siena,�Italy,�June�2001,�Pr�o�c�e�e�dings�,��Lecture����K1�Notes�Tin�Articial�In��9telligence��2083�,�Springer-V��:�erlag�(2001).������?16.���O��Beeson,�#M.,�Solving�for�functions,�to�app�A�ear�in��Journal�Z�of�Symb��olic�Computa-����K1�tion�.�@�A�@�preliminary�v��9ersion�app�A�eared�in:��LMCS�v2002,�v)L��o�gic,�Mathematics,�and����K1�Computer�^GScienc��e:�Inter�actions,�Symp�osium�in�Honor�of�Bruno�Buchb�er�ger's����K1�60th��^Birthday�,�Öpp.�24{38,�RISC-Linz�Rep�A�ort�Series�No.�02-60,�Researc��9h�Insti-����K1�tute�Tfor�Sym��9b�A�olic�Computation,�Linz�(2002).������?17.���O��Belinfan��9te,�J.,�Computer�pro�A�ofs�in�G��`odel's�class�theory�with�equational�deni-����K1�tions�N�for�comp�A�osite�and�cross,��J.���A�Îutomate��d�R�e�asoning�N��22�,�No.�3,�pp.�311-339,����K1�1988.������?18.���O��Belinfan��9te,��\J.,�On�computer-assisted�pro�A�ofs�in�ordinal�n�um�b�A�er�theory��:�,��J.�5A�Îuto-����K1�mate��d�N<R�e�asoning�T�22�,�No.�3,�pp.�341-378,�1988.������?19.���O��Bledso�A�e,��W.�W.,�and�Hines,�L.�M.,�V��:�ariable�elimination�and�c��9haining�in�a����K1�resolution-based�Tpro��9v�er�for�inequalities������?20.���O��Bo��9y�er,��jR.�S.,�and�Mo�A�ore,�J.�S.,��A��&Computational��RL��o�gic�Handb�o�ok�,��jAcademic����K1�Press,�TBoston�(1988).������?21.���O��Bronstein,��M.,��Symb��olic�*�Inte�gr�ation�I:�T��J�r�ansc�endental�F��J�unctions�,��Springer-����K1�V��:�erlag,�TBerlin/�Heidelb�A�erg/�New�Y�ork�(1997).�����2�c�ō���>;����?�50���`-	Mic��9hael�TBeeson����)�ō���>;����?22.���O��Bro��9wn,�g�C.,�QEPCAD-B,�a�program�for�computing�with�semi-����K1�algebraic�
K�sets�using�CADs,�to�app�A�ear.�Preprin��9t�a�v��|railable�at����K1�h��9ttp://www.cs.usna.edu/��pw�cbro�wn/researc�h/MOTS2002.2.p�A�df.�The�program����K1�itself�Tis�a��9v��|railable�from��
�
��K1�h��9ttp://www.cs.usna.edu/�UMqep�A�cad/B/QEPCAD.h�tml,�and�v�e�example�problems����K1�can�Tb�A�e�view��9ed�at����K1�h��9ttp://www.cs.usna.edu/�Tqep�A�cad/B/examples/Examples.h�tml.������?23.���O��Buc��9h�b�A�erger,��B.,��et.�>�al.��Theorema:�An�In��9tegrated�System�for�Computation�and����K1�Deduction�_oin�Natural�St��9yle,�in:�Kirc�hner,�C.,�and�Kirc�hner,�H.�(eds.),��Pr��o�c�e�e�d-����K1�ings��of�the�Workshop�on�Inte��gr�ation��of�De��ductive�Systems�at�CADE-15,�Lindau,����K1�Germany,�N<July�1998�,�TLNAI��1421�,�Springer,�Berlin�(1998).������?24.���O��Buc��9h�b�A�erger,���B.,�Collins,�G.,�and�Lo�os,�R.,��Computer���A�Îlgebr��a:�Symb�olic�and�A�Îl-����K1�gebr��aic�N<Manipulation�,�Tsecond�edition,�Springer-V��:�erlag�Wien/�New�Y�ork�(1983).������?25.���O��Bundy��:�,���A.,��The���Computer�Mo��del�x�ling�of�Mathematic�al�R�e�asoning�,���Academic����K1�Press,�TLondon�(1983).������?26.���O��Bo�A�ole,��G.,��The�,�L��aws�of�Thought�,��Do��9v�er,�New�Y��:�ork�(1958).�Original�edition,����K1�MacMillan�T(1854).������?27.���O��Borsuk,�,mK.,�and�Szmielew,�W.��F��J�oundations�cuof�Ge��ometry�,�,mNorth-Holland,�Am-����K1�sterdam,�T(1960).������?28.���O��Ca��9viness,�=VB.F.,�and�Johnson,�J.R.�(eds.)��Quantier���Elimination�and�Cylindric��al����K1�A�Îlgebr��aic�N<De�c�omp�osition�,�TSpringer,�Wien/New�Y��:�ork�(1998).������?29.���O��Ch��9urc�h,�A.,�An�unsolv��|rable�problem�of�elemen��9tary�n�um�b�A�er�theory��:�,��A�Îmeric��an����K1�Journal�N<of�Mathematics�T�58��345{363,�1936.������?30.���O��Clark��9e,��wE.,�and�Zhao,�X.:�Analytica:�A��KTheorem�Pro�v�er�in�Mathematica,�in:����K1�Kapur,��D.�(ed.),��A�Îutomate��d�#CDe�duction:�CADE-11:�Pr�o�c.�of�the�11th�Inter-����K1�national��Confer��enc�e�on�A�Îutomate�d�De�duction�,�зpp.�761{765,�Springer-V��:�erlag,����K1�Berlin/Heidelb�A�erg�T(1992).������?31.���O��Collins,���G.E.,�Quan��9tier�elimination�for�real�closed�elds�b�y�cylindrical�al-����K1�gebraic�({decomp�A�osition,�in:��Pr��o�c.�K2nd�Conf.�on�A�Îutomata�The��ory�and�F��J�ormal����K1�L��anguages�,�TSpringer�LNCS��33��134{183.�Reprin��9ted�in�[28],�pp.�85-121.������?32.���O��Con��9w�a�y��:�,�J�J.,�and�Sloane,�N.,��Spher��e���Packings,�L�attic�es�and�Gr�oups�,�J�Grundlehren����K1�Der���Mathematisc��9hen�Wissensc�haften��290�,�Springer-V��:�erlag,�Berlin/�Heidelb�A�erg/����K1�New�TY��:�ork�(1998).������?33.���O��Cop�A�e,�TD.,��Computers�N<and�Music��al�Style�,�TA-R�Editions,�Madison�(1991).������?34.���O��Cop�A�e,�FD.,��Exp��eriments�J�in�Music�al�Intel�x�ligenc�e�,�FA-R�EEditions,�Madison�(1996).������?35.���O��Cop�A�e,�TD.,��The�N<A�Îlgorithmic�Comp��oser�,�TA-R�Editions,�Madison�(2000).������?36.���O��Da��9v�enp�A�ort,�G�J.,�and�Hein��9tz,�J.,�Real�quan�tier�elimination�is�doubly�exp�A�onen�tial,����K1�in�T[4],�pp.�29{35.������?37.���O��Da��9vis,��*M.,�A��
computer�program�for�Presburger's�algorithm,�in��Summaries���of����K1�T��J�alks�l�Pr��esente�d�at�the�Summer�Institute�for�Symb�olic�L�o�gic,�1957� �Second�edition,����K1�published�Tb��9y�Institute�for�Defense�Analysis,�1960.�Reprin�ted�in�[92],�pp.�41-48.������?38.���O��Da��9vis,��M.�The�prehistory�and�early�history�of�automated�deduction,�in�[92],����K1�pp.�T1{28.������?39.���O��Da��9vis,��ZM.,���The�̰Universal�Computer:�The�R��o�ad�̰fr�om�L�eibniz�to�T��J�uring�,��ZNorton����K1�(2000).�TReprin��9ted�in�2001�under�the�title��Engines�N<of�L��o�gic�.������?40.���O��Da��9vis,�&xM.,�and�Putnam,�H.,�A�&tcomputing�pro�A�cedure�for�quan�tication�theory��:�,����K1��JA�ÎCM�T�7��201-215,�1960.������?41.���O��Ernst,�]yZ.,�Fitelson,�B.,�Harris,�K.,�McCune,�W.,�V��:�ero,�R.,�W�os,�L.,�More�First-����K1�order�LT��:�est�Problems�in�Math�and�Logic,�in:�Sutclie,�G.,�P��9elletier,�J.�and�Suttner,�����3��c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������51����)�ō���>;��K1�C.�E(eds.)��Pro�A�ceedings�of�the�CADE-18�W��:�orkshop{Problems�and�Problem�Sets����K1�for��&A��:�TP,�T�ec��9hnical�Rep�A�ort�02/10,�Departmen�t�of�Computer�Science,�Univ�ersit�y����K1�of���Cop�A�enhagen,�Cop�enhagen�(2002).�The�pap�er�and�asso�ciated�Otter�les�can����K1�b�A�e�Taccessed�from�h��9ttp://www.mcs.anl.go�v/�Tmccune/pap�ers/paps-2002.��
V�����?42.���O��Fisc��9her,�aM.�J.,�and�Rabin,�M.�O.,�Sup�A�er-exp�onen�tial�acomplexit�y�of�Presburger����K1�arithmetic,��U�SIAM-AMS��Pr��o�c�e�e�dings�,�V��:�olume�VI�A�I,�pp.�27{41;�reprin��9ted�in�[28],����K1�pp.�T122{135.������?43.���O��F��:�rege,���G.,��Be��grishrift,��\a�formula�language,�mo�dele�d�up�on�that�of�arithmetic,����K1�for�N<pur��e�thought.�,�TEnglish�translation�in�[102],�pp.�1{82.�Original�date�1879.������?44.���O��Gelern��9ter,��oH.,�Realization�of�a�geometry-theorem�pro�ving�mac�hine,��Pr��o�c.��
In-����K1�ternational�naConfer��enc�e�on�Information�Pr�o�c�essing,�UNESCO�nHouse�N�273{282,����K1�(1959).��Reprin��9ted�in�F��:�eigen�baum�and�F��:�eldman�(eds.),��Computers�Jand�Thought�,����K1�McGra��9w-Hill,�TNew�Y��:�ork�(1963).�Reprin�ted�again�in�[92],�pp.�99-124.������?45.����P��������O��Ub�A�er���formal�unen��9tsc�heidbare���S��`atze�der��Principia���Mathematic��a��und�v��9erw�andter����K1�Systems�*I,�in:�F��:�eferman,�S.,��et.�val.��(eds.),��Kurt�G��Godel:�Col�x�le��cte�d�Works,�V��J�olume�I,����K1�Public��ations�
�1929-1936�,��!pp.�144-195.�(The�translation,�On�formally�undecidable����K1�prop�A�ositions���of��Principia�7�Mathematic��a��and�related�systems�I,�app�ears�on�facing����K1�pages.)�7�Oxford�Univ��9ersit�y�7�Press,�New�Y��:�ork,�and�Clarendon�Press,�Oxford�(1986).������?46.���O��Grigor'ev,�iD.,�and�V��:�orob���jo��9v,�N.,�Solving�systems�of�p�A�olynomial�inequalities�in����K1�sub�A�exp�onen��9tial�Ttime.�J.�Sym�b.�Comput.��5�,�37-64,�1988.������?47.���O��Guard,��@J.,�Oglesb��9y��:�,�F.,�Bennett,�J.,�and�Settle,�L.,�Semi-automated�mathemat-����K1�ics,�T�JA�ÎCM��16�(1)�49{62,�1969.������?48.���O��Harrison,��J.,�and�Th������Wery��:�,�L.:�Extending�the�HOL��theorem�pro��9v�er��with�a�com-����K1�puter��algebra�system�to�reason�ab�A�out�the�reals,�in��Higher�?�Or��der�L�o�gic�The�or�em����K1�Pr��oving�ϊand�its�Applic�ations:�6th�International�Workshop,�HUG��i'93�,��
pp.�174{����K1�184,�TLecture�Notes�in�Computer�Science��780�,�Springer-V��:�erlag�(1993).������?49.���O��Harrison,���J.,��The��or�em���Pr�oving�with�the�R�e�al�Numb�ers�,���Springer-V��:�erlag,����K1�Berlin/Heidelb�A�erg/New�TY��:�ork�(1998).������?50.���O��Hilb�A�ert,�:D.,��Grund�x�lagen�A�der�Ge��ometrie�,�:T��:�eubner�(1899).�English�translation����K1�(of�8�the�10th�edition,�whic��9h�app�A�eared�in�1962):��F��J�oundations�n�of�Ge��ometry�,�8�Op�en����K1�Court,�TLa�Salle,�Illinois�(1987).������?51.���O��Hilb�A�ert,��=D.,�and�Ac��9k�ermann,��=W.,�'Grundzge�der�theoretisc��9hen�Logik',�2nd�edi-����K1�tion���(1938;�rst�edition�1928).�English�translation:��Principles��_of�Mathematic��al����K1�L��o�gic�,�K�translated�b��9y�Lewis�M.�Hammond,�George�G.�Lec�kie�and�F.�Steinhardt;����K1�edited��|with�notes�b��9y�Rob�A�ert�E.�Luce,�Chelsea�(1950).�The�translation�is�to�b�e����K1�reprin��9ted�Tb�y�the�A.�M.�S.�in�2003.������?52.���O��Hong,��~H.,�Simple�solution�form��9ula�construction�in�cylindrical�algebraic�de-����K1�comp�A�osition��based�quan��9tier�elimination,�in:��Pr��o�c.�KInternational�Symp��osium�on����K1�Symb��olic�~�and�A�Îlgebr�aic�Computation�,�`�pp.�177-188,�A��9CM,�1992.�Reprin�ted�in����K1�[28],�Tpp.�210{210.������?53.���O��Horgan,�TJ.,�The�death�of�pro�A�of,��Scientic�N<A�Îmeric��an�,�Octob�er�1993,�74{82.������?54.���O��Huang,���T.,��A�Îutomate��d�"�De�duction�in�R�Îing�The�ory�,���Master's�W��:�riting�Pro���ject,����K1�Departmen��9t�Tof�Computer�Science,�San�Jose�State�Univ�ersit�y�(2002).������?55.���O��G.�jHuet.�fA�junication�algorithm�for�t��9yp�A�ed���-calculus.��The��or�etic�al�� Computer����K1�Scienc��e�T�1��(1975)�27{52.������?56.���O��Jacobsen,�TN.,��Basic�N<A�Îlgebr��a�I�,�T(1985).������?57.���O��Kaltofen,�TE.,�F��:�actorization�of�p�A�olynomials,�in�[24],�pp.�95{114.������?58.���O��Kamm�A���;uller,���F.,�and�P��9aulson,�L.,�A���formal�pro�of�of�Sylo��9w's�theorem:�an�exp�er-����K1�imen��9t���in�abstract�algebra�with�Isab�A�elle�HOL,��J.��$A�Îutomate��d�R�e�asoning����23�,�pp.����K1�235{264,�T1999.�����4$d�c�ō���>;����?�52���`-	Mic��9hael�TBeeson����)�ō���>;����?59.���O��Kapur,��>D.,�and�Zhang,�H.,�An�o��9v�erview��>of�Rewrite�Rule�Lab�A�oratory�(RRL),��J.����K1�of�N<Computer�and�Mathematics�with�Applic��ations�,�T�29��2,�91{114,1995.��
�,����?60.���O��Kleene,�'mS.�C.,���-denabilit��9y�and�recursiv�eness,�Duk�e�Mathematical�Journal��2�,����K1�pp.�T340{353,�1936.������?61.���O��Kleene,�^�S.�C.,��Intr��o�duction�h9to�Metamathematics�,�^�v��|ran�Nostrand,�Princeton����K1�(1952).������?62.���O��Kn��9uth,�)zD.�E.,�and�Bendix,�P��:�.�B.,�Simple�w�ord�problems�in�univ�ersal�algebras,����K1�in:��aLeec��9h,�J.�(ed).,��Computational��Pr��oblems�in�A�Îbstr�act�A�Îlgebr�as��a�pp.�263{297,����K1�P��9ergamon�TPress�(1970).�Reprin�ted�in�[93],�pp.�342{376.������?63.���O��Kn��9uth,��D.�E.,��Seminumeric��al�A�Îlgorithms:�The�A�rt�of�Computer�Pr��o�gr�amming,����K1�V��J�olume�N<2�,�Tsecond�edition,�Addison-W��:�esley�,�TReading,�MA�(1981).������?64.���O��Kurosh,�TA.�G.,��The�N<The��ory�of�Gr�oups,�volume�1�,�TChelsea,�New�Y��:�ork�(1955).������?65.���O��Lang,�TS.��A�Îlgebr��a�,�Addison-W��:�esley�,�TReading,�MA.�(1965).������?66.���O��Lewis,�TP��9aul,�obituary�of�Herb�Simon,�h�ttp://www.cs.osw�ego.edu/�blue/hx/����K1�courses/cogsci1/s2001/section04/subsection01/main.h��9tml.������?67.���O��Mancuso,�_�P��:�.,��F��J�r��om��iBr�ouwer�to�Hilb�ert�,�_�Oxford�Univ��9ersit�y�_�Press,�Oxford�(1998).������?68.���O��Maslo��9v,�I�S.,�Min�ts,�G.,�and�Orevk�o�v,�V.,�Mec�hanical�pro�A�of-searc�h�and�the�theory����K1�of�Tlogical�deduction�in�the�USSR,�in:�[92],�pp.�29{37.������?69.���O��Mo�A�ore,�
~yG.�E.,�Cramming�more�comp�onen��9ts�on�to�in�tegrated�cir-����K1�cuits,���Ele��ctr�onics��38�,�Num��9b�A�er�8,�pp.�April�18,�1965.�Av��|railable�at����K1�h��9ttp://www.in�tel.com/researc�h/silicon/mo�A�orespap�er.p�df.������?70.���O��McCune,�@�W.,�Otter�2.0,�in:�Stic��9k�el,�@�M.�E.�(ed.),��10th�akInternational�Confer-����K1�enc��e���on�A�Îutomate�d�De�duction�Z��pp.�663{664,�Springer-V��:�erlag,�Berlin/Heidelb�A�erg����K1�(1990).������?71.���O��McCune,�'�W.,Solution�of�the�Robbins�problem,��J.�_A�Îutomate��d�R�e�asoning�'��19�(3)����K1�263{276,�T1997.������?72.���O��McCune,�S�W.,�and�P��9admanabhan,�R.,��A�Îutomate��d��QDe�duction�in�Equational�L�o�gic����K1�and�N<Cubic�Curves�,�TLNAI��1095�,�Springer,�Berlin/�Heidelb�A�erg�(1996).������?73.���O��Monk,�J.�D.,�The�mathematics�of�Bo�A�olean�algebra,�in�the��Stanfor��d�O�Dictionary����K1�of�N<Philosophy�,�Th��9ttp://plato.stanford.edu/en�tries/b�A�o�olalg-math.������?74.���O��New��9ell,��
A.,�Sha�w,�J.�C.,�Simon,�H.,�Empirical�explorations�with�the�Logic�The-����K1�ory���Mac��9hine:�a�case�study�in�heuristics,��Pr��o�c�e�e�dings�6�of�the�Western�Joint�Com-����K1�puter�
\Confer��enc�e,�Institute�of�R�adio�Engine�ers�λ�218-230,�1957.�Reprin��9ted�in�[92],����K1�pp.�T49{74.������?75.���O��P��9eano,�!6G.,��A�Îrithmetic��es�m�principia,�nova�metho�do�exp�osita�,�!6Bo�A�cca,�T��:�urin,�(1889).������?76.���O��Pietrzyk��9o�wski,�N�T.,�and�Jensen,�D.,�A�N�complete�mec��9hanization�of�second�order����K1�logic,�T�J.�N<Asso��c.�Comp.�Mach.��20��(2),�1971,�pp.�333-364.������?77.���O��Pietrzyk��9o�wski,�K�T.,�and�Jensen,�D.,�A�K�complete�mec��9hanization�of��!�R��-order�t�yp�A�e����K1�theory��:�,�T�Asso��c.�N<Comp.�Math.�Nat.�Conf.��1972,�V��:�ol.�1,�82-92.������?78.���O��P��9oincar������We,�MH.,��Scienc��e�;Kand�Metho�d�,�Mtranslated�b��9y�Maitland�from�the�original����K1�F��:�renc��9h�Tedition�of�1908,�Do�v�er�(1952).������?79.���O��P��9enrose,�^�R.,��The��OEmp��er�or's�New�Mind:�Conc�erning�Computers,�Minds,�and�the����K1�L��aws� �of�Physics��,��American�Philological�Asso�A�ciation�(1989).�See�also�the�review����K1�b��9y���McCarth�y��:�,�J.,�in��Bul�x�letin�sof�the�A�Îmeric��an�Mathematic�al�So�ciety�,���Octob�A�er,����K1�1990.������?80.���O��P��9enrose,���R.,��Shadows��>of�the�Mind:�A���Se��ar�ch��>for�the�Missing�Scienc��e�of����K1�Consciousness�,���Oxford�Univ��9ersit�y���Press,�Oxford�(1996).�See�also�the�crit-����K1�ical��reviews�b��9y�McCarth�y��:�,�J.,�in�the�electronic�journal��Psyche��2��(11)����K1�(1995),�0<h��9ttp://psyc�he.cs.monash.edu.au/v2/psyc�he-2-11-mccarth�y��:�.h�tml,�and�b�y�����56��c�ō���>;�����IȻThe�TMec��9hanization�of�Mathematics������53����)�ō���>;��K1�F��:�eferman,��S.,��Psyche��2�(7)�(1995),�h��9ttp://psyc�he.cs.monash.edu.au/v2/psyc�he-����K1�2-07-feferman.h��9tml,�qand�P�enrose's�replies�to�these�and�other�critics,�at����K1�h��9ttp://psyc�he.cs.monash.edu.au/v2/psyc�he-2-23-p�A�enrose.h�tml.��
������?81.���O��P��9etk�o�v�����sek,��JM.,�Wilf,�H.,�and�Zeilb�A�erger,�D.,��A=B�,�A.�K.�P��9eters,�W��:�ellesley�,��JMA����K1�(1996).������?82.���O��Presburger,�[�M.,���c������Ub�A�er�die�V��:�ollst��`andigk��9eit�eines�gewissen�Systems�der�Arith-����K1�metik��Gganzer�Zahlen,�in�w��9elc�hem��Gdie�Addition�als�einzige�Op�A�eration�herv��9ortritt,����K1�Sparw��9ozdanie�7�z�I�7�Kongresu�Matemat�yk�o���X�w�Kra���jo���X�w�Slo��9wia�A���;nskic�h�7�W��:�arsza�w�a,�pp.����K1�92-101,�T1929.������?83.���O��Prior,�TA.�N.,��F��J�ormal�N<L��o�gic�,�Tsecond�edition,�Clarendon�Press,�Oxford�(1962).������?84.���O��Renegar,���J.,�Recen��9t�progress�on�the�complexit�y�of�the�decision�problem�for�the����K1�reals,�Tin:�[28],�pp.�220{241.������?85.���O��Ric��9hardson,��D.,�Some�unsolv��|rable�problems�in�v�olving�elemen�tary�functions�of�a����K1�real�Tv��|rariable,�J.�Sym��9b�A�olic�Logic��33��511{520�(1968).������?86.���O��Robinson,��A.,�Pro��9ving�theorems,�as�done�b�y�man,�mac�hine,�and�logician,�in����K1��Summaries��Nof�T��J�alks�Pr��esente�d��Nat�the�Summer�Institute�for�Symb��olic�L�o�gic,�1957����K1ĻSecond�(Hedition,�published�b��9y�Institute�for�Defense�Analysis,�1960.�Reprin�ted�in����K1�[92],�Tpp.�74-76.������?87.���O��Robinson,���G.,�and�W��:�os,�L.,�P��9aramo�A�dulation�and�theorem-pro�ving�in�rst-order����K1�theories�ԇwith�equalit��9y��:�,�in�Meltzer�and�Mic�hie�(eds.),��Machine��Intel�x�ligenc��e��4�,�pp.����K1�135-150,�TAmerican�Elsevier,�New�Y��:�ork�(1969).�Reprin��9ted�in�[110],�pp.�83{99.������?88.���O��Robinson,���G.,�and�W��:�os,�L.,�Completeness�of�paramo�A�dulation,��Journal�2of�Sym-����K1�b��olic�N<L�o�gic�T�34�,�p.�160�(1969).�Reprin��9ted�in�[110],�pp.�102{103.������?89.���O��Robinson,��J.�A.,�A���mac��9hine�orien�ted�logic�based�on�the�resolution�principle,����K1��JA�ÎCM�T�12�,�23-41,�1965.�Reprin��9ted�in�[92],�pp.�397-415.������?90.���O��Robinson,�ˈAlan,�and�V��:�oronk��9o�v,�ˈA.�(eds.),�Handb��o�ok���of�A�Îutomate��d�R�e�asoning,����K1�V��J�olume�N<I�,�TMIT�Press,�Cam��9bridge,�and�North-Holland,�Ammsterdam�(2001).������?91.���O��Russell,�RB.,�and�Whitehead,�A.�N.,��Principia�6�Mathematic��a�,�Cam��9bridge�Uni-����K1�v��9ersit�y�;�Press,�Cam��9bridge,�England.�First�edition�(1910),�second�edition�(1927),����K1�reprin��9ted�T1963.������?92.���O��Siekmann,��ZJ.,�and�W��:�righ��9tson,�G.�(eds),��A�Îutomation���of�R��e�asoning���1:����K1�Classic��al�	�uPap�ers�on�Computational�L�o�gic�1957-1966�,�
7�Springer-V��:�erlag,����K1�Berlin/Heidelb�A�erg/New�TY��:�ork�(1983).������?93.���O��Siekmann,��ZJ.,�and�W��:�righ��9tson,�G.�(eds),��A�Îutomation���of�R��e�asoning���2:����K1�Classic��al�	�uPap�ers�on�Computational�L�o�gic�1967-1970�,�
7�Springer-V��:�erlag,����K1�Berlin/Heidelb�A�erg/New�TY��:�ork�(1983).������?94.���O��Quaife,�	�A.,��A�Îutomate��d�.�Development�of�F��J�undamental�Mathematic�al�The�ories,����K1�A�Îutomate��d�N<R�e�asoning,�V��J�ol.2�,�TKlu��9w�er�Academic�Publishers,�Dordrec�h�t�(1992).������?95.���O��Risc��9h,��iR.,�The�problem�of�in�tegration�in�nite�terms,��T��J�r��ansactions�	dof�the�AMS����K1�139�T�167{189,�1969.������?96.���O��Risc��9h,�v8R.,�The�solution�of�the�problem�of�in�tegration�in�nite�terms,��Bul�x�letin����K1�of�N<the�AMS�T�76��605{608,�1970.������?97.���O��T��:�arski,A.,�q�A�qMdecision�metho�A�d�for�elemen��9tary�algebra�and�geometry�.�Rep�A�ort����K1�R-109,�֒second�revised�edition,�Rand�Corp�A�oration,�San��9ta�Monica,�CA,�1951.����K1�Reprin��9ted�Tin�[28],�pp.�24{84.������?98.���O��T��:�arski,�9�A.,�What�is�elemen��9tary�geometry?�in:�Henkin,�L.,�Supp�A�es,�P�,�and�T�arski,����K1�A.��](eds.),��Pr��o�c�e�e�dings��
of�an�International�Symp��osium�on�the�Axiomatic�Metho�d,����K1�with�@*Sp��e�cial�R�efer�enc�e�to�Ge�ometry�and�Physics��16-29,�North-Holland,�Amster-����K1�dam�T(1959).�����6G��c�ō���>;����?�54���`-	Mic��9hael�TBeeson����)�ō���>;����?99.���O��T��:�rybulec,�n�W.�A.,�Subgroup�and�cosets�of�subgroup,��Journal�Mpof�F��J�or-����K1�malize��d��Mathematics��2�,�1990.�This�is�an�electronic�journal,�published�at����K1�h��9ttp://mizar.u�wb.edu.pl/JFM/index.h�tml.������?100.���Tq�T��:�uring,�PDA.,�On�computable�n��9um�b�A�ers,�PDwith�an�application�to�the�En��9tsc�heidung-����K1�sproblem,���Pr��o�c�e�e�dings��Jof�the�L��ondon�Mathematic�al�So�ciety�,��ser.�2��42��230{265,����K1�1936-7;�Tcorrections,��ibid.��43��(1937)�pp.�544-546.������?101.���Tq�T��:�uring,���A.,�Computing�Mac��9hines�and�In�telligence,�in��MIND,���A���Quarterly�R��e-����K1�view�N<of�Psycholo��gy�and�Philosophy�59�,�TNo.�236,�Octob�A�er,�1950,�pp.�433-460.������?102.���Tq�v��|ran���Heijeno�A�ort,�J.�(ed.)��F��J�r��om��F�r�e�ge��to�G��Godel:�A�}Sour�c�e�Bo�ok�in�Mathematic�al����K1�L��o�gic,�N<1879{1931�,�THarv��|rard�Univ��9ersit�y�TPress,�Cam��9bridge,�MA�(1967).������?103.���Tq�W��:�ang,�ΘH.,�T�o��9w�ard�Θmec�hanical�mathematics,�1960.�Reprin�ted�in�[92],�pp.�244{����K1�267.������?104.���Tq�W��:�eispfenning,�C�V.,�The�complexit��9y�of�linear�problems�in�elds,��J.�d2Symb��olic����K1�Computation�T�5��3{27�(1988).������?105.���Tq�Wiedijk,F.,�TThe�fteen�pro��9v�ers�Tof�the�w��9orld,�to�app�A�ear.������?106.���Tq�Winkler,�OnF.,��Polynomial�n�A�Îlgorithms�in�Computer�A�lgebr��a�,�OnSpringer-V��:�erlag,����K1�Wien/�TNew�Y��:�ork�(1996).������?107.���Tq�W��:�os.,�yhL.,�Robinson,�G.,�and�Carson,�D.,�Eciency�and�completeness�of�the�set����K1�of��supp�A�ort�strategy�in�theorem�pro��9ving,��JA�ÎCM��12�(4)�536{541,�1965.�Reprin�ted����K1�in�T[110],�pp.�29{36.������?108.���Tq�W��:�os,��L.,��The��tA�Îutomation�of�R��e�asoning:��tA�n�Exp��erimenter's�Noteb�o�ok�with����K1�OTTER�N<T��J�utorial�,�TAcademic�Press,�San�Diego�(1996).������?109.���Tq�W��:�os,�4�L.,�and�Piep�A�er,�G.,��A�k-F��J�ascinating�k4Country�in�the�World�of�Computing�,����K1�W��:�orld�TScien��9tic,�Singap�A�ore�(1999).������?110.���Tq�W��:�os,�KvL.,�and�Piep�A�er,�W.,��The��Col�x�le��cte�d�Works�of�L�arry�Wos,�V��J�olume�I:�Ex-����K1�ploring�N<the�Power�of�A�Îutomate��d�R�e�asoning�,�TW��:�orld�Scien��9tic,�Singap�A�ore�(2000).������?111.���Tq�W��:�os,��L.,�and�Piep�A�er,�W.,��The�(Col�x�le��cte�d�Works�of�L�arry�Wos,�V��J�olume�II:�Ap-����K1�plying��|A�Îutomate��d�R�e�asoning�to�Puzzles,�Pr�oblems,�and�Op�en�Questions�,�LW��:�orld����K1�Scien��9tic,�TSingap�A�ore�(2000).������?112.���Tq�W��:�os,�A�L.,�Re
ections�on�Automated�Reasoning�at�Argonne:�A�A�P��9ersonalized����K1�F��:�ragmen��9t�Tof�History�,�on�the�CD-R��9OM�included�with�[109].������?113.���Tq�W��:�os,�!?L.,�Ulric��9h,�D,�and�Fitelson,�B.,�X�CB,�the�last�of�the�shortest�single�axioms����K1�for�Tthe�classical�equiv��|ralen��9tial�calculus,�to�app�A�ear�in��J.�N<A�Îutomate��d�R�e�asoning������?�114.���Tq�W��:�u,�,HW�en-Tsun,�On�the�decision�problem�and�the�mec��9hanization�of�theorem-����K1�pro��9ving���in�elemen�tary�geometry��:�,��Scientia��uSinic��a��21��(2),�1978.�Reprin�ted�in:����K1�Bledso�A�e,�X7W.�W.,�and�Lo��9v�eland,�X7D.�W.�(eds.),��A�Îutomate��d��tThe�or�em�Pr�oving:�After����K1�25�N<Y��J�e��ars�,�TAMS,�Pro��9vidence,�RI�(1983).������?115.���Tq�W��:�u,�$|W�en-Tsun,��Me��chanic�al�p�The�or�em�Pr�oving�in�Ge�ometries:�Basic�Principles�,����K1�Springer-V��:�erlag,�TWien/�New�Y�ork�(1994).������?116.���Tq�Y��:�u,�p�Y.�Computer�Pro�A�ofs�in�Group�Theory�,��Journal��Gof�A�Îutomate��d�R�e�asoning����K1�6�(3)�T251-286,�1990.������?117.���Tq�Zhang,�x}H.,�Automated�pro�A�of�of�ring�comm��9utativit�y�x}problems�b��9y�algebraic����K1�metho�A�ds,�T�J.�N<Symb��olic�Computation��9��423{427,�1990.�����Y���;�c����6��q[�		cmsl9���<x

cmtt10��':

cmti10���N�cmbx12��j��		cmti9�t�:		cmbx9�q�%cmsy6�����		cmsy9�;�cmmi6�5��"		cmmi9��Aa�cmr6�o���		cmr9���N�ffcmbx12�
!",�

cmsy10�O!�cmsy7�
�b>

cmmi10�	0e�rcmmi7�O
�\cmmi5�K�`y

cmr10�ٓ�Rcmr7���Zcmr5���u

cmex10�hG�����

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