• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 1
  • 1
  • 1
  • Tagged with
  • 4
  • 4
  • 3
  • 2
  • 2
  • 2
  • 2
  • 2
  • 2
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
1

Type-alpha DPLs

Arkoudas, Konstantine 05 October 2001 (has links)
This paper introduces Denotational Proof Languages (DPLs). DPLs are languages for presenting, discovering, and checking formal proofs. In particular, in this paper we discus type-alpha DPLs---a simple class of DPLs for which termination is guaranteed and proof checking can be performed in time linear in the size of the proof. Type-alpha DPLs allow for lucid proof presentation and for efficient proof checking, but not for proof search. Type-omega DPLs allow for search as well as simple presentation and checking, but termination is no longer guaranteed and proof checking may diverge. We do not study type-omega DPLs here. We start by listing some common characteristics of DPLs. We then illustrate with a particularly simple example: a toy type-alpha DPL called PAR, for deducing parities. We present the abstract syntax of PAR, followed by two different kinds of formal semantics: evaluation and denotational. We then relate the two semantics and show how proof checking becomes tantamount to evaluation. We proceed to develop the proof theory of PAR, formulating and studying certain key notions such as observational equivalence that pervade all DPLs. We then present NDL, a type-alpha DPL for classical zero-order natural deduction. Our presentation of NDL mirrors that of PAR, showing how every basic concept that was introduced in PAR resurfaces in NDL. We present sample proofs of several well-known tautologies of propositional logic that demonstrate our thesis that DPL proofs are readable, writable, and concise. Next we contrast DPLs to typed logics based on the Curry-Howard isomorphism, and discuss the distinction between pure and augmented DPLs. Finally we consider the issue of implementing DPLs, presenting an implementation of PAR in SML and one in Athena, and end with some concluding remarks.
2

A self-verifying theorem prover

Davis, Jared Curran 24 August 2010 (has links)
Programs have precise semantics, so we can use mathematical proof to establish their properties. These proofs are often too large to validate with the usual "social process" of mathematics, so instead we create and check them with theorem-proving software. This software must be advanced enough to make the proof process tractable, but this very sophistication casts doubt upon the whole enterprise: who verifies the verifier? We begin with a simple proof checker, Level 1, that only accepts proofs composed of the most primitive steps, like Instantiation and Cut. This program is so straightforward the ordinary, social process can establish its soundness and the consistency of the logical theory it implements (so we know theorems are "always true"). Next, we develop a series of increasingly capable proof checkers, Level 2, Level 3, etc. Each new proof checker accepts new kinds of proof steps which were not accepted in the previous levels. By taking advantage of these new proof steps, higher-level proofs can be written more concisely than lower-level proofs, and can take less time to construct and check. Our highest-level proof checker, Level 11, can be thought of as a simplified version of the ACL2 or NQTHM theorem provers. One contribution of this work is to show how such systems can be verified. To establish that the Level 11 proof checker can be trusted, we first use it, without trusting it, to prove the fidelity of every Level n to Level 1: whenever Level n accepts a proof of some phi, there exists a Level 1 proof of phi. We then mechanically translate the Level 11 proof for each Level n into a Level n - 1 proof---that is, we create a Level 1 proof of Level 2's fidelity, a Level 2 proof of Level 3's fidelity, and so on. This layering shows that each level can be trusted, and allows us to manage the sizes of these proofs. In this way, our system proves its own fidelity, and trusting Level 11 only requires us to trust Level 1. / text
3

A Generic Proof Checker

Watson, Geoffrey Norman Unknown Date (has links)
The use of formal methods in software development seeks to increase our confidence in the resultant system. Their use often requires tool support, so the integrity of a development using formal methods is dependent on the integrity of the tool-set used. Specifically its integrity depends on the theorem prover, since in a typical formal development system the theorem prover is used to establish the validity of the proof obligations incurred by all the steps in the design and refinement process. In this thesis we are concerned with tool-based formal development systems that are used to develop high-integrity software. Since the theorem prover program is a critical part of such a system, it should ideally have been itself formally verified. Unfortunately, most theorem provers are too complex to be verified formally using currently available techniques. An alternative approach, which has many advantages, is to include a proof checker as an extra component in the system, and to certify this. A proof checker is a program which reads and checks the proofs produced by a theorem prover. Proof checkers are inherently simpler than theorem provers, since they only process actual proofs, whereas much of the code of a theorem prover is concerned with searching the space of possible proofs to find the required one. They are also free from all but the simplest user interface concerns, since their input is a proof produced by another program, and their output may be as simple as a `yes/no' reply to the question: Is this a valid proof? plus a list of assumptions on which this judgement is based. When included in a formal development system a stand-alone proof checker is, in one sense, superfluous, since it does not produce any proofs -- the theorem prover does this. Instead its importance is in establishing the integrity of the results of the system -- it provides extra assurance. A proof checker provides extra assurance simply by checking the proofs, since all proofs have then been validated by two independent programs. However a proof checker can provide an extra, and higher, level of assurance if it has been formally verified. In order for formal verification to be feasible the proof checker must be as simple as possible. In turn the simplicity of a proof checker is dependent on the complexity of the data which it processes, that is, the representation of the proofs that it checks. This thesis develops a representation of proofs that is simple and generic. The aim is to produce a generic representation that is applicable to the proofs produced by a variety of theorem provers. Simplicity facilitates verification, while genericity maximises the return on the effort of verification. Using a generic representation places obligations on the theorem provers to produce a proof record in this format. A flexible recorder/translator architecture is proposed which allows proofs to be recorded by existing theorem provers with minimal changes to the original code. The prover is extended with a recorder module whose output is then, if necessary, converted to the generic format by a separate translator program. A formal specification of a checker for proofs recorded in this representation is given. The specification could be used to formally develop a proof-checker, although this step is not taken in this thesis. In addition the characteristics of both the specification and possible implementations are investigated. This is done to assess the size and feasibility of the verification task, and also to confirm that the design is not over-sensitive to the size of proofs. This investigation shows that a checker developed from the specification will be scalable to handle large proofs. To investigate the feasibility of a system based on this architecture, prototype proof recorders were developed for the Ergo 5 and Isabelle 98 theorem provers. In addition a prototype checker was written to check proofs in the proposed format. This prototype is compatible with the formal specification. The combined system was tested successfully using existing proofs for both the Ergo 5 and Isabelle 98 theorem provers.
4

Développement du système MathNat pour la formalisation automatique des textes mathématiques / Developing System MathNat for Automatic Formalization of Mathematical texts

Muhammad, Humayoun 18 January 2012 (has links)
Le langage mathématique courant et les langages mathématiques formelssont très éloignés. Par <<langage mathématique courant>> nousentendons la prose que le mathématicien utilise tous les jours dansses articles et ses livres. C'est une langue naturelle avec desexpressions symboliques et des notations spécifiques. Cette langue està la fois flexible et structurée mais reste sémantiquementintelligible par tous les mathématiciens.Cependant, il est très difficile de formaliser automatiquement cettelangue. Les raisons principales sont: la complexité et l'ambiguïté deslangues naturelles en général, le mélange inhabituel entre languenaturelle et notations symboliques tout aussi ambiguë et les sautsdans le raisonnement qui sont pour l'instant bien au-delà descapacités des prouveurs de théorèmes automatiques ou interactifs.Pour contourner ce problème, les assistants de preuves actuelsutilisent des langages formels précis dans un système logique biendéterminé, imposant ainsi de fortes restrictions par rapport auxlangues naturelles. En général ces langages ressemblent à des langagesde programmation avec un nombre limité de constructions possibles etune absence d'ambiguïté.Ainsi, le monde des mathématiques est séparé en deux, la vastemajorité qui utilise la langue naturelle et un petit nombre utilisantaussi des méthodes formelles. Cette seconde communauté est elle-mêmesubdivisée en autant de groupes qu'il y a d'assistants de preuves. Onperd alors l'intelligibilité des preuves pour tous les mathématiciens.Pour résoudre ce problème, on peut se demander:est-il possible d'écrire un programme qui comprend la langue naturellemathématique et qui la traduit vers un langage formel afin depermettre sa validation?Ce problème se subdivise naturellement en deux sous-problèmes tous lesdeux très difficiles:1. l'analyse grammaticale des textes mathématiques et leur traductiondans un langage formel,2. la validation des preuves écrites dans ce langage formel.Le but du projet MathNat (Mathematics in controlled Natural languages)est de faire un premier pas pour répondre à cette question trèsdifficile, en se concentrant essentiellement sur la première question.Pour cela, nous développons CLM (Controlled Language for Mathematics)qui est un sous-ensemble de l'anglais avec une grammaire et un lexiquerestreint, mais qui inclut tout de même quelques ingrédientsimportants des langues naturelles comme les pronoms anaphoriques, lesréférences, la possibilité d'écrire la même chose de plusieursmanières, des adjectifs distributifs ou non, ...Le second composant de MathNath est MathAbs (Mathematical Abstractlanguage). C'est un langage formel, indépendant du choix d'un systèmelogique permettant de représenter la sémantique des textes enpréservant leur structure et le fil du raisonnement. MathAbs est conçucomme un langage intermédiaire entre CLM et un système logique formelpermettant la vérification des preuves.Nous proposons un système qui permet de traduire CLM vers MathAbsdonnant ainsi une sémantique précise à CLM. Nous considèrons que cetravail est déjà un progrès notable, même si pour l'instant on estloin de pouvoir vérifier formellement toutes les preuves en MathAbsainsi générées.Pour le second problème, nous avons réalisé une petite expérience entraduisant MathAbs vers une liste de formules en logique du premierordre dont la validité garantit la correction de la preuve. Nous avonsensuite essayé de vérifier ces formules avec des prouveurs dethéorèmes automatiques validant ainsi quelques exemples. / There is a wide gap between the language of mathematics and itsformalized versions. The term "language of mathematics" or"mathematical language" refers to prose that the mathematician uses inauthoring textbooks and publications. It mainly consists of naturallanguage, symbolic expressions and notations. It is flexible,structured and semantically well-understood by mathematicians.However, it is very difficult to formalize it automatically. Some ofthe main reasons are: complex and rich linguistic features of naturallanguage and its inherent ambiguity; intermixing of natural languagewith symbolic mathematics causing problems which are unique of itskind, and therefore, posing more ambiguity; and the possibility ofcontaining reasoning gaps, which are hard to fill using the currentstate of art theorem provers (both automated and interactive).One way to work around this problem is to abandon the use of thelanguage of mathematics. Therefore in current state of art of theoremproving, mathematics is formalized manually in very precise, specificand well-defined logical systems. The languages supported by thesesystems impose strong restrictions. For instance, these languages havenon-ambiguous syntax with a limited number of possible syntacticconstructions.This enterprise divides the world of mathematics in two groups. Thefirst group consists of a vast majority of mathematicians whose relyon the language of mathematics only. In contrast, the second groupconsists of a minority of mathematicians. They use formal systems suchas theorem provers (interactive ones mostly) in addition to thelanguage of mathematics.To bridge the gap between the language of mathematics and itsformalized versions, we may ask the following gigantic question:Can we build a program that understands the language of mathematicsused by mathematicians and can we mechanically verify its correctness?This problem can naturally be divided in two sub-problems, both very hard:1. Parsing mathematical texts (mainly proofs) and translating thoseparse trees to a formal language after resolving linguistic issues.2. Verification of this formal version of mathematics.The project MathNat (Mathematics in controlled Natural language) aimsat being the first step towards solving this problem, focusing mainlyon the first question.First, we develop a Controlled Language for Mathematics (CLM) which isa precisely defined subset of English with restricted grammar anddictionary. To make CLM natural and expressive, we support some richlinguistic features such as anaphoric pronouns and references,rephrasing of a sentence in multiple ways and the proper handling ofdistributive and collective readings.Second, we automatically translate CLM to a system independent formaldescription language (MathAbs), with a hope to make MathNat accessibleto any proof checking system. Currently, we translate MathAbs intoequivalent first order formulas for verification.

Page generated in 0.0927 seconds