Computer Understanding of Mathematical Proofs
Proceedings of the Fifth International Joint Conference on Artificial Intelligence. Specialized Systems. Pages 851-856.
© 1977 International Joint Conferences on Artificial Intelligence
Abstract
Mathematical proofs constitute a mixture of formulas with a subset of natural language. They can be represented as a sequence of lines expressible in the symbolism of predicate calculus. The transition from step to step may depend on a series of logical manipulations and/or on intricate mathematical knowledge associated with the domain of the proof. The organization of the proof may depend on different conventions adopted by mathematicians in communication with each other. This paper deals with problems involved in following the mathematical argument along those lines. Some of . the ideas were implemented as a part of a system for teaching axiomatic set theory to college students. The most powerful and frequently used rules of inference utilize a resolution theorem prover. To the best of our knowledge this is the only resolution theorem prover, perhaps the only general purpose theorem prover used in actual production.
- Abrahams, P., "Machine Verification of Mathe- matical Proof", 1963, Doctoral Dissertation, MIT.
- Goldberg, A., "A Generalized Instructional System for Teaching Elementary Mathematical Logic", Tech.Rep.No.l79, 1971, IMSSS, Stanford University.
- Marinov, V., "Maximal Clause Length Resolu- tion", 1973, Doctoral Dissertation, University of Texas at Austin.
- Marinov, V., "Replace Formula", 1974, IMSSS internal memo, Stanford University.
- McCarthy, J., "Computer Programs for Checking Mathematical Proofs", 1961, Proc.Amer.Math. Soc. on Recursive Function Theory, New York.
- Smith, R., Graves, H., Blaine, L. and Marinov, V. 1975. "Computer Assisted Axiomatic Mathe- matics: Informal Rigor." In: Computers in Education, 0. Lecarne and R.-rewis, eds. IFIP Second World Conference on Computer Education. North Holland, Amsterdam. pp.803-809.
- Smith, R., "BOOLE--A Decision Method", 1974, IMSSS internal memo, Stanford University.
- Suppes, P., Introduction to Logic, 1957, Van Nostrand, Princeton, N.J. 9. Suppes, P., Axiomatic Set Theory, 1972~ Dover, New York.
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
Introduction
A mathematician writes down a proof of a theorem on a piece of paper and hands it to a colleague. The latter reads through it, usually understands it, and all too frequently finds an error, which means that at least in its present form it is not a proof at all. Sometimes it takes several readers until an existing error is de tected and on occasion it has taken many years to discover errors in some difficult proofs. It would be very useful if our mathematician could type in the computer whatever he wrote and receive some for of the response he received from the hu man reader. The paper examines problems encoun tered during initial efforts toward such a goal.
Furthermore, if one attempts to teach higher level mathematics by computer, as it is currently being done for set theory at Stanford, it is es-
- Present address: Department of Computer Science, Oregon State University
guage processing will be spent in correct trans lation of the steps in the argument, although it will also affect the structure of the proof and the choice of how to go about making a step in the proof. The rema1n1ng problem, to which this paper is devoted, is how to follow the logical steps in the proof.
Formal and Informal Proofs
In communications among mathematicians or between a teacher of mathematics and his students the steps in the proofs are as large as it is op timal for understanding. They are large enough to avoid obvious but tedious details which would usually make the proof less clear. On the other hand, they are small enough to make it possible for the reader to follow the proof without having to discover it himself. The size of the steps varies widely with different proofs and in some cases only a very general outline is given, thus requiring a greater effort on the part of the reader. In any event, that is what we call an in formal proof. In addition, the concepts of being essential, unfamiliar, or trivial in a proof are not precise at all. This very vagueness of the criteria governing informal proofs is a primary justification for a precise definition of a formal proof [8].
Formal proofs require precise rules for jus tifying each step. In that respect they are some what algorithmic in character. It is no wonder that automatic proof checking is almost as old as artificial intelligence itself [5]. One of the first computer proof checkers was that of Abrahams [1], which implemented an extension of the logical system of Suppes [8]. IMSSS at Stanford has of fered a CAl course in symbolic logic using a proof checker based on Suppes' system since 1963. A later implementation is described in [2].
Size of the Steps
The problem with such formal proofs is that their steps are quite small. A proof of somewhat greater difficulty in set theory (or other mathe matical theory) could easily go up into several hundred steps. Such proofs cannot be used for teaching axiomatic mathematics and hardly for any other purpose. This problem has been resolved to a large degree in the proof checker of QUIP.
Before discussing the techniques used in accepting larger steps we shall illustrate some of the points made above with a comparative exam ple. First is the proof of Theorem 53 on page 112 in [9]. Then follows the proof of the same theo rem as it was accepted by QUIP. K(A) is the car dinal number of a set A, while ~ means equi pollent. Axioms, definitions, and theorems ref erenced in example proofs are given in the appen dix. The command to the proof checker precedes each line. Some of the commands may seem long but are not entirely typed by the user, since the sys tem has a recognition feature (see [6]}. The QUIP proof below is presented as listed by the command "Review". In doing the first three steps a bit of additional typing is required where the user in dicates that he wants to specify variables upon which the system asks him what values to assign. In the last two steps the lines have to be typed after the command.
A proof from "Axiomatic Set Theory" by P. Suppes:
- Theorem 53. There are sets A and B such that (i) AflB = 0, (ii) K(A) = m, (iii) K(B) = n.
Proof. In view of Definition 7 (Cardinal Number), there are sets Al and Bl such that K(Al}= m and K(Bl} = n, and by virtue of Theorem 9 there are sets A and B such that Ai"B = 0, A~Al, and B-::::Bl. By the axiom for cardinals, then, K(A) = m and K(B) = n. Q.E.D.
Proof of the same theorem as accepted by the proof checker:
Theorem 5:1.1
Derive: (lA,B) (Af\B = OAK(A) mi\K(B) n)
Df. of Cardinal Number
(1} K(Al} = m
Df. of Cardinal Number
(2) K(Bl} = n
Theorem 3.8.9
(3) Al~AABl~BAAI\B = 0
1, 2, 3, Axiom for Cardinals VERIFY
(4) K(A) = mAK(B) = n
3, 4 VERIFY
(5} (3 A,B)(AAB = 01\K(A) mAK(B))= n Q.E.D.
Notice that the steps of the QUIP proof fol low exactly the informal proof from the book. In the first three steps a definition and a theorem are invoked as the appropriate instantiations are made. The system has sorted variables and the letters m and n are among those used for cardinal~ This enables us for example to get Line 1 directly from the definition of cardinal number. The QUIP proof requires Line 5 to obtain a formula which is a variant in terms of bound variables (in this case identical} to the statement of the theorem. The informal proof quits at Step 4. This is one of the many shortcuts used in informal proofs, some of which will be discussed below. They will eventually present the greatest problem in accept ing informal text proofs and will require a large amount of additional information to be stored with the internal representation of a line.
Step 4 is the most interesting in this proof. It is achieved by a call to a resolution theorem prover giving it the lines 1, 2, and 3 plus the axiom for cardinals as references. Below is the continuation of the proof in the formal system of Suppes [8] as implemented in [2]. Preceding each line to the left are the rules of inference as typed by the user. In the brackets to the right their names are expanded.
Axiom for Cardinals (4) K(Al} = K(A) ~ Al~A 3 LC [Left Conjunct] (5) Al-=A
4,5 RQ [Replace Equivalence]
(6) K(Al) = K(A)
1,6 RE [Replace Equals]
(7) K(A) = m Axiom for Cardinals (8) K(Bl) = K(B) - Bl~B 3 RC [Right Conjunct] (9) Bl-;::::;BI'Af'\B = 0
9 LC [Left Conjunct]
(10) Bl~B
9,10 RQ [Replace Equivalence]
(ll) K(Bl) = K(B)
2,ll RE [Replace Equals]
(12) K(B) = n
7,12 FC [Form Conjunction]
- K(A) =m~'K(B)=n
As shown above, Step 4 requires 10 elementary steps. Yet the informal proof only says it fol lows "by the axiom for cardinals". For the mathe matician this is an "obvious" step. In [9] one frequently finds lines in proofs that follow by sentential logic (p. 29, p. 52) or by quantifier logic (p. 45, p.60). When a step is justified by sentential logic it is decidable and one can safe ly lean on a tautology checker. It is our belief that the most efficient mechanical way to confirm a tautology is to check the truth table and thus QUIP has a TAUTOLOGY rule based on that principle. Steps justified by logical manipulations involving quantifiers are undecidable in general and conse quently represent a much harder problem to be ver ified mechanically. For this purpose we employ a resolution theorem prover with equality replace ment.
Resolution Theorem Prover in Action
The experience with QUIP shows that a well organized resolution theorem prover gets most of the inferences seen intuitively by the user while working on a proof. This is not quite true for proofs involving equality, where some improvements to the prover are needed·. The user has no inter action with the theorem prover except for asking a formula to be verified and supplying the refer ences, from which he thinks the formula follows. An idea about the power of the theorem prover can be gained from the following theorems from the chapter on finite and infinite sets in QUIP's cur riculum, which were proved in one step.
Theorem 4.1.24 Finite(A)ABCA + B..{A Follows from:
- Df. of Proper Subset
(V'A,B)(AC:B ~ A~BAMB)
- Th. 3.9.2 (V'A,B)(A B +A B)
- Th. 3.9.16 (~A,B)(A B ~A B A B)
- Th. 4.1.23 (~· A)(Finite (A)+ Dfinite(A))
- Df. of Dedekind finite
('r/ A) (Dfini te (A) ~ I (J C) (C CA 1\ C ~Al
The above theorem was proved in 7.4 seconds of CPU time.
Theorem 4.3.2
(3 A, B) (Infinite (A) 1\ Infinite(B) AA ·#a)
Follows from:
- Df. of Infinite Set (V'A)(Infinite (A) ~ -, Finite (A))
- Th. 4.3.1 (3A) Infinite (A)
- Th. 3.9.17 ('i A) A-<P(A)
- Th. 3.9.13 (\>' A,B)(A~B + B ~A)
- Th. 3.9.1 ('VA,B)(A~B +A~ B)
- Th. 3.8.2 ('<fA,B)(A~B+B-;::::.A)
- Th. 4.1.17 (V' A) (Finite(P(A))+Finite(A))
It was proved in 3.7 seconds of CPU times.
Theorem 4.1.22 Finite(A) A IFinite(B) + A<B Follows from:
- Th. 3.9.16 (VA,B)(A~B-A~BvA...(B)
- Th. 4.1.19 N A, B) (Finite (A) I'.A~B ~Finite (B))
- Th. 4.1.20
("1 A, B) (Finite (A) 1\ B-(A-+ Finite (B))
- Th. 4.1.21
NA,B)(Finite (A)+ A-<.BvA~BvB-<.A)
A student proof of the last theorem usually looks more like the following:
Derive: Finite (A) 1\ !Finite (B) + A-<B Hypothesis
(1) Finite (A)/\ !Finite (B)
lCl, Th. 4.1.21
IMPLIES [lCl: Line 1,
- A-(BVA~BVB-(A
Assume
A~B (3)
1,3, Th. 4.1.19 CONTRADICTION
(4) A-:/!.B
Assume
(5) B<_A
5, Th. 3.9.16 VERIFY
(6) B~A
1,6, Th. 4.1.20 CONTRADICTION
(7) B /.,A
2,4, 7 TAUTOLOGY
(8) A..(B
1,8 Conditional Proof
- Finite (A)I\!Finite (B) + A.c(B
Q.E.D
This proof should be transparent even to those unfamiliar with Suppes' logical system [8]. The theorem can be proved in one step, yet the user usually sees the exact references on which this step depends only after developing the proof in the latter form. Furthermore, this form is more like the way mathematicians prove theorems. Still notice that the resolution prover was called upon for Steps 4, 6, and 7.
The prover is used primarily for the rules VERIFY and CONTRADICTION. While using ·VERIFY the user has to type the line, whose negation together with the references is passed to the prover. If the prover is able to confirm the inference it signals the proof checker to accept the line. For the CONTRADICTION rule the user merely points to the references which he believes form an incon sistency. If such is detected by the prover, the proof checker returns the negation of the last assumption on which the references depend. (Pre sumably there must be an incorrect assumption in order to reach a contradiction).
Conjunct 1]
Implementation and Strategy
The main reason for selecting a resolution theorem prover was our belief that for the same generality and the same power it can be designed in a much more compact way than a heuristic theo rem prover. For the purposes we are using it, simply a mechanical tool is needed and resolution seems to be exactly that. Later we shall discuss the possibility of a heuristic coupler to the proof checker, which would make the prover serve better the needs of understanding informal proofs. The prover was written in UCI-LISP. Together with the converter of the formulas into clausal form it is about 10 pages of pretty-printed code.
The prover employs the MU strategy. It con sists mainly of keeping only resolvents containing merge literals or having a unit parent. It is shown in (3] that if in a refutation there are re solvents not satisfying the above restriction, there always exists another refutation from the same input set where such resolvents are obtained first. With this in view the strategy occasion ally allows for a round of general resolution after which the restriction is imposed.
One thing that has plagued work on resolution in the past has been preoccupation with complete ness. Recognizing that a prover is working in an undecidable domain it is obvious that completeness is going to be restricted by the real factors of time and space. The main objective in choosing a strategy and tuning a prover's parameters is op t1m1z1ng the number of inferences it gets. It is the author's conviction that in this context in completeness is a feature, rather than a drawback. Thus, completeness in our prover is restricted severely in many different ways.
Experiments with different strategies for resolution, carried out earlier by the author at the University of Texas have shown the MU strategy to be quite efficient in the set-theoretical do main. One property of the MU strategy, coupled with a limit on the depth of functional nesting in the resolvents, is that it usually runs quickly out of possibilities to resolve when given a sat isfiable set of clauses (i.e., insufficient ref erences). This is very important because one very frequent error of the student users has been to supply insufficient or incorrect references. In such a case it is very desirable that the prover detects this fact as soon as possible, rather than grind until the time limit is reached. This pro perty has strongly influenced the selection of the
MU strategy.
Deficiencies
Probably the most frustrating property of the prover from a CAl point of view has been the fact that thus far it has been impossible to charac terize the class of theorems accepted by the prover despite substantial effort on the part of members of the IMSSS staff. This is most likely a consequence of the unnatural way in which reso lution works. Sometimes the prover is able to verify steps much larger than the user can see, while other times it fails at steps which the user expects to be accepted. It would have been nice if one could give the users some more accurate idea what to expect from the prover.
Inferences wh~ch are missed by the prover, while being obvious to the user, are largely due to the fact that resolution breaks down the form ulas to the atomic level before it can find a proof. For example, the inference
('lfm,n,p,A,B){A(IB=O 1\ K(A}=m 1\ K(B)=n AK(AVB)=p ~ m + n = p}
from
cannot be verified by the prover. The human user, though, with his well-developed abstraction capa bilities quickly sees that the left side of the implication can be looked at as ~(m,n,p,A,B).
Hence the inference becomes:
from
(V'm,n,p)[(3A,B)~(m,n,p,A,B) ~ m + n = p].
Of course, seen this way, the inference is immediate for the prover, too. There is a pro vision in the program for the user to pass to the prover the abstracted form of the formulas. What would be desirable here is to have a heuristic coupler between the prover and the proof checker which looks at the possibilities of abstraction in the set of formulas.
Implication Rule
Proofs in axiomatic systems frequently in clude transformations of formulas justified by the application of definitions, axioms, or theorems. Almost always these inferences include quantifiers and would fall within the domain of the theorem prover. However, such inferences are determin istic and one potentially loses power in taking a chance with the prover. Many definitions are by equivalence, which is a simple connective for the human, but the worst one for resolution. Further more, for the rules VERIFY and TAUTOLOGY the formula must be typed in, which the users are not very enthusiastic about.
All these considerations led to the develop ment of the IMPLIES rule. Variables bound by an universal quantifier containing the main connec tive of the justifying formula within its scope are matched to any terms in the formula to be transformed. For example, let
- f''C s;_ f"D be a line in a proof (f"C means the image of the set C under the function f). From it we can infer the line
- C<rJ x) (x £ f''C -+- x £ f"D) just by referring to the definition of subset: r<rJ A, B)(AS B ._ (V x)(x £ A -+- x £ B)).
In a case like this, which happens most frequently, the rule is equivalent to universal spec ification and modus ponens. The rule also allows for (1) to be inferred from (2) by the definition of subset. In some cases the matching is much more intricate than in the example above. The matching algorithm utilizes a subset of unifica tion. The same algorithm is used in matching terms for replacements by equality in the proof. More details can be found in [4].
Matching of Schemas
A more sophisticated application of the matching algorithm is the matching of theorem schemas. Sometimes axioms and theorems contain formula schemas which can be matched to any form ula satisfying some variable constraints. For example, the theorem
- ( 3 B)("t x)(x £ B - FM (x)) + ('d y)(y £ {x:FM(x)} - FM(y)) will match the proof line
- (3 C)("f z)(z £ C - z.!: D) and return
- (V'y) (y £ {x:x~D} ~ y ~D). Here, FM is matched to zSD, while the match be tween the designated variable x of the schema FM(x) and z is carried along.
Occasionally it is not straightforward to determine which variable to match with the desig nated variable of the schema. Then one has to look at which quantifier matched the innermost quantifier over the designated variable. Since the latter is always quantified, the algorithm would succeed if there is a possible match. Then, when the rule is returning, the current designated variable is substituted for the value of the vari able of the same schema when the match was first made. (Of course, any subsequent conflict would have failed the match). Thus at the point of re covering the value of FM(y) from (3), y is substituted for z in z ~ D (the match to FM) and ySD is returned as seen in (S). Recall that x was associated with z when z !:OD was matched to FM(x) and now y corresponds to x.
Embodiment of Mathematical Knowledge
Some lines in proofs are produced by mathe maticians without references to the justifying definitions or theorems. They usually express well-understood and well-organized domains of knowledge shared by the workers in a particular field. Exact references in such cases would ob struct rather than facilitate communication. A very simple example from set theory is the line
- ASC 1\ B5:C + AUB~C. To justify this line one should really refer to the definitions of subset and union, although no student of set theory does so once he has grasped the Boolean operations and relations between sets. More complicated examples could easily be found in proofs on set theory, for example, in treating functions, Cartesian products, etc.
It would be desirable if "bags" of such know ledge were embodied in decision methods. A good example is the BOOLE rule, which is a part of QUIP
[7]. This rule uses a decision method by Quine and converts the Boolean operations and relations between sets into propositional connectives, after which a tautology check is applied.
This rule could possibly be extended to cover more of the set-theoretical knowledge, and other similar rules could be developed. However, it is not clear how much of set theory could be embodied in decision methods. There certainly will be need for some heuristic knowledge, organized as well as possible. In fact, such embodiments would be a very important part of any completed system han dling some domain of mathematics.
Goal Hierarchy
Informal proofs employ many conventions, assumptions, and shortcuts not explicitly mention ed in the proof. They largely depend on the do main of the theory in which the theorem is being proven, although there are many conventions which apply generally to mathematics. For example, it is customary when one wants to show
- A~ B to assume
- X £ A and eventually derive
- X £ B.
Thus the proof (or part of proof) is terminated, but the implicit assumptions are
- ("/ x) (x £ A + x £ B), and that (7) follows from (10) and the definition of subset.
We propose that this be dealt with in the following way. If the user wants to prove some thing in the form of (10), while making the as sumption (8) he also states the goal (9). When the goal is legitimately achieved, the proof moni tor generates formula (10) and associates it with the line (9). (At the present time a lot of addi tional information is being associated with each line, i.e., free variables, dependencies, etc.). Then if a reference to that line is made there will be no ambiguity, since the form (9) depended on (8) and is no longer available. (The working premise (8) was discarded -- see [8] -- and all lines depending on it are not available).
In more complicated proofs the user will have to state goals on different levels in the proof. Thus the monitor will deal with a hierarchy of embodied goals. It would be better if the goals were declared automatically, but at the present time we do not foresee any reasonable way of im plementing the goal hierarchy handling without the user stating the goals. Top-level goals would not need to be explicitly mentioned.
An Informal Proof
We try to illustrate some of the above dis cussion with the informal proof of a simple theo rem. The definition of image in the proof below is:
(VR,A,y)(y E R"A + (3x)(x E AI\ <x,y> E R)).
Theorem C ~ D + f''C !::f"D Proof. Assume (1) y E f"C By the definition of image (2) (3x) (x E C 1\ <x,y> E f) From (2), the hypothesis of the theorem, and the definition of subset it follows that (3) ( 3x) (x E D A <x,y> E f) Whence by the definition of image (4) y E f''D Q.E.D
The proof checker handles the steps as they are given here. Of course, accepting this proof would require stronger natural language processing capabilities than there currently are available in the program. The semantic analysis of the text should show that steps (2) and (4) are applica tions of the IMPLIES rule, while step (3) requires a call to the prover via VERIFY .
Future Directions
The problems facing future work on under standing informal proofs may be best illustrated by considering another comparative example. A proof of Theorem 37 on page 104 in [9] is present ed below as it is accepted by QUIP. (To fully benefit from the comparison the reader should look at the proof in the book).
Derive: Finite (A) 1\ ( 3 f) f A SURJ B + Finite (B)
Hypothesis
(1) Finite (A) l\(3f)f : A SURJ B 1C2 ES [Line 1, Conjunct 2, Exist. Specific.] (2) f : A SURJ B Axiom of Separation (3) rv x) (x E K - X E P(A) "Finite(f''x)) Th. 1.6, Th. 1.7, Th. Image VERIFY (4) f"O = 0 4, Th. 4.1.1 VERIFY (5) Finite(f''O) 3,5 Th. 1.48 VERIFY (6) 0 E K Assume (7) x E AI\C=.AAC E K 7, Th. Singleton, Df. Subset VERIFY (8) {x} £A 7C3, Th. Concretion IMPLIES (9) C~A 1\ Finite(f''C) BOOLE UG (10) ('VA,B,C)(A~C 1\B-=C + AUB :c) 8,9,10 VERIFY (11) CV{x} ~ A 2,7, Th. 3.2.35, Df. Surjection, Df. Map VERIFY (12) (3y)f"({x}) = {y} 12, Th. 4.1.2 VERIFY (13) Finite(f''{x}) 9,13, Th. 4.1.6 VERIFY (14) Finite(f''C V f"{x}) 14, Th. 3.2.29 RE [Replace by Equality] (15) Finite(C V {x})
3,11,15, Th. Power Set VERIFY (16) C \) {x} E K 7,16 CP UG [Cond. Proof and Univ. Generaliz.] (17) ("VC,x) (x E A 1\ C E K + CV{x} E K) 1Cl,6,17, Th. 4.1.11 IMPLIES (18) A E K 18,4 IMPLIES (19) A E P (A) A Finite (f''A) 2, Df. Surjection, Df. Map. Th. 3.2.20 VERIFY (20) f"A = B 19, 20 VERIFY (21) Finite(B) 1,21 CP (22) Finite(A) 1\ ( 3f) (f :A SURJ B) + Finite(B) Q.E.D.
The proof in the book is given in a very com pact fo:m and requires a lot of thinking from the reader 1n order to follow it in detail. Many :eaders would probably agree with it after glanc1ng through, but when asked for the justification of a certain step they might have difficulties in explaining it. Even a simple thing as defining th~ set K by abstraction really appeals to the ax1om of separation to make sure that the set exists.
Yet in view of the applications we have in mind the proof checker of QUIP is more like the kind of system one needs. First, it is an im portant part of the task of a proof reader to make s~r~ that all steps in a proof were properly jus t1f1ed. Second, in teaching axiomatic mathemat ics it is important to insure that the student understands the exact justification of a proof he might read in a book.
The knowledge necessary for understanding a proof as the one above will be distributed over the.machinery for making large steps, the goal mon1tor, and the semantics of the mathematicai l~nguage processor. Procedural knowledge about d1fferent ways of organizing proofs will be needed. '
Most notably the QUIP proof in the last exam ple seems to be organized in reverse order of the book proof. The latter first states that induc tion is needed to show f''A E K, whence B is finite, since f"A = B, and then proceeds to develop the induction conditions. The QUIP proof works out the induction by Theorem 4.1.11 (Step 18) and then shows the finiteness of B. In fact the proof includes several occurrences of the s~e phenomen on on a smaller scale.
Processing the wide variety of ways of ex pression seen in the mathematical literature is a formidable problem. It will require continuous work. for time to come. But by putting some re s~ra1nts on the form of presenting proofs one m1ght soon be able to make good use of a computer system for the purposes outlined in the beginning of this paper.
Acknowledgements
Many of the presented ideas were formed in conversations with P. Suppes, L. Blaine, H. �rave~ and R. Smith. Also, W. Bledsoe has made many helpful suggestions. The research has been sup ported by the U.S. National Science Foundation under Grant EC43997 to Stanford University.
@inproceedings{ijcai1977p201,
title = {Computer Understanding of Mathematical Proofs},
author = {Marinov, Vesko},
booktitle = {Proceedings of the Fifth International Joint Conference on
Artificial Intelligence, {IJCAI-77}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Raj Reddy},
pages = {851--856},
year = {1977},
note = {Specialized Systems},
url = {https://ijcai.org/proceedings/1977/201},
}
TY - CPAPER TI - Computer Understanding of Mathematical Proofs AU - Marinov, Vesko T2 - Proceedings of the Fifth International Joint Conference on Artificial Intelligence, IJCAI-77 PB - International Joint Conferences on Artificial Intelligence PY - 1977 SP - 851 EP - 856 UR - https://ijcai.org/proceedings/1977/201 ER -Copied
Machine-readable versions of this paper, for tools and AI agents: