Mathematics of Infinity
Mathematics of Infinity
Mathematics of Infinity
Per M a r t i n - L ~ f
HOCBHmaeTcH na~t~TM
25.4.1903 -- 20.10.1987
t47
Intuitive considerations
(ao~(al,(a2,..o))),
( a o , ( a l , ( a 2 ..... ( a n _ l , n i l ) . . . ) ) ) ,
o(](o(...))),
ral numbers to the two element set. An even simpler example, the
s(s~s(...))),
I
Introduced by J. Wallis, De Sectionibus Conicis, Nova
Methodo Expositis, Tractatus, Oxford, 1655, in the laconic paren-
thesis (esto enim ~ nota humeri infiniti;), apparently without
worrying about its meaningfulness.
t48
of sets. But we may also let sets be infinite, not in the usual
A X(A X(A X . . . ) ) ,
The latter set is the disjoint union of two sets, each of which
union of two sets, etc. in infinitum. It looks very much like the
Cantor set.
it seem clear whether this particular one ought to come out true
aforementioned ones, you soon realize that they are maybe not so
novel creatures after all. We have also old examples, like in-
b0
a0 +
bI
aI +
b2
a2 +
a 3 + ...,
2
P. Aczel, Non-Well-Founded Sets, CSLI Lecture Notes,
Number 14, Stanford University, 1988.
150
and
I I I
' 2' " ' "
mo l~o
n=o@ n=l
(1 + --~)~ .
Only during the period of the last thirty years has there been
for dealing with streams and other infinite objects, like the
pings in Scott's sense, etc. Now, since all the set theoretical
tion, even absurdity, come out true° The reason is that every
ical notation, the formal laws for the fixed point operator that
(x ~ A) (x e A)
f(x) e A f(x) ~ A
They say that every approximable mapping from a domain into it-
l
id(x) ~ A (x 6 A),
id(x) = x ~A (x E A )
formal laws for the fixed point operator are adjoined to standard
fix(id) E A
In particular,
fix(id) E N o = S..
Thus type theory becomes inconsistent when the formal laws for
The previous laws for the fixed point operator then get modi-
fied into
154
(x ~ A) (x ~ A)
aEA f(x) 6 A a E A f(x) E A
These laws, in turn, are readily seen to have the same effect as
= s(~) ~ N.
infinity by putting
= fix(O,s) ~ N,
fix(a,f) = r e c ( ~ , a , ( x , y ) f ( y ) ) E A.
That the second rule for the modified fixed point operator be-
fix(a,f) = rec(~,a,(x,y)f(y))
= rec(s(~),a,(x,y)f(y))
= f(rec(~,a,(x,y)f(y)))
= f(fix(a,f)) E A.
This seems fine, but, although we no longer get the fixed point
NI(N,x,s(x)) t r u e (x E N)
N i(N,~,S(~)) true
hand,
I(N,~gs(~)) true
sible.
n= (((...??},
by relaxing the axioms of standard set theory, namely, by giving
metic rather than set theory, this would mean giving up the prin-
metic and analysis. Also, the idea of getting the fixed point
=,o = s ( s ( s ( . . . ) ) )
greater than
O( = fo(f](f2(...)))
process
~I = f I ( ~ 2 )'
~2 = f2(~3 ),
has been concerned with the case when the functions, from which
Now, think of
= s(s(s(...)))
definitional process
= n 0 = s(~ 1 ),
° ° I : s(~2)'
~2 : s(~3)'
• ~ 6
the axioms
{~i
~i
~ N,
= s(~i+1) ~ N
way all the time that seduced us into making the circular defini-
fixi(a,f) = rsc(~i,a~(x,Y)f(Y)) E A,
(x ~ A) (x ~ A)
a E A f(x) ~ A a E A f(x) E A
~i = fixi(0's) ~ N
and thereby satisfy the axioms for ~ i ' i = 0, I, etc. Thus the
159
rules for fixi(a,f) and the axioms for ~ i are formally equi-
point operator.
induction.
for x G N, then
lim A(x) = A ( ~ )
X=
x E N, then
mute with every other operation, this means that the limit oper-
I
pd(O) = 0 E N,
pd(s(a)) = a E N,
I ~ - 0 = a E N,
-- s ( b ) = pd(a - b ) E N,
pd(a) = rec(a,O,(x,y)x) E N,
a -- b : rec(b,a,(x,y)pd(y)) E N.
i = s ( ~ i + 1 ) E N,
we get inversely
~i+I = Pd(~i) ~ N.
161
equalities
it.
0 ~ x true (x E N),
0 ~ . . true
i+3
sJ(o) ~ s J ( ~ i + j) true.
~. = S j (c~i+ j ) E N.
Therefore we have
sJ(o)~o0 i true
I
A i set,
ai ~ Ai ,
ested in is when
I
A i = N,
N( N< ...< N~ N(
S S S 8 8
163
In the most general case, each A i need not be just a single set
but a whole context
Ail set,
Xini_ I ~ Aini_1(Xil,..-,Xini_2))-
Then
xi = x i I'''" ,Xin i
as a sequence of elements
ail E Ail,
With this amount of vector notation, the general case can be re-
that is,
I
~i E A i,
i = fi ( ~ i+I ) E A i
xj E Aj.
be strictly greater.
~j e A j,
{
o~j_ I E Aj_ I ,
O(j_ I : f j _ 1 ( ~ j ) E Aj_ I,
11
S° Albeverio, J. E. Fenstad, R. Hoegh-Krohn, and T. Lind-
strum, Nonstandard Methods in Stochastic Analysis and Mathemat-
ical Physics, Academic Press, New York, 1986, p. 46.
166
{ ~0
~0
g AO'
= fO(~l ) ~ AO'
But for the crucial first of these axioms, this is nothing but
a sequence of explicit definitions, successively defining ~j-1'
Cgj_2, etc., ultimately 6 0 in terms of ~j. Indeed, we have
~i = fi(fi+1('''fj-1(~j )'.')) E Aj
~ j e A jo
This being the only axiom which governs the constant ~j, we may
as well replace it by a variable xj, thereby transforming the
axiom ~ j ~ Aj into the assumption xj E Aj. This proves the ne-
cessity of the condition.
The sufficiency of the condition is clear, because Tj is an
extension of T, and, in Tj, we have access to the axiom
~ j e Aj
that we have used in the course of the proof are the substitution
axioms of T are.
By means of the lemma, the following proof theoretic version
A true
aEA
a = b ( ~ j ) ~ A,
where
aj E Aj
b(aj) E A.
168
A true
is provable in T already.
itself and not for its meaning. For example, among the objects
case.
rA i set, •
i fi(xi+1)
a i ~ Ai
~ A i (xi+ I E Ai+ I),
Mi = M/Ai .
Ai, say
o(.6A
1 1
simply as
x i E A i (x i ~ A i),
Here f~i denotes composition with fi" Thus f~i takes an object in
A(x i) set (x i E A i)
a(x i) E A(x i) (x i E A i)
Let
C(Xk,X) (x k E A k, x e A(fik(Xk))),
a(x i) ~ N (x i ~ A i)
and
I
~ i E A i,
~i = fi(~i+1 ) E A i
x i E A i (x i E Ai).
~i = fi ( ~ i+I ) ~ Ai'
A(x o) = A (x o E A o)
and
B(x o) = B (x o e A o)
A(foi(Xi)) = B(foi(Xi)) (x i E A i)
A(foi(ai)) = B(foi(ai)).
A(foi(ai)) = A
and
B(foi(ai)) = B,
176
A=B.
commutative diagram
subtheory
embedding
T-- >T
intended
interpretation interpretation
M ~ ~ M~ = !ira M/A i,
star
embedding
A(x O) : A (x 0 E A O)
true
But
178
f00(x0) = x 0 E A 0 (x 0 E A0),
A(f0i(xi)) true (x i 6 A i)
A(f0i(xi)) true (x i E A i)
A(f0i(ai)) true
A(f0i(ai)) = A.
we can conclude
A true
as desired.
If you compare this proof with the earlier proof of the
proof theoretic version of the transfer principle, you will see
may wonder how we have been able to circumvent this in the pre-
the following.
180
general.
false.
introduction,
A true B true
A & B true
A true B true
A(x i) (xi ~ A i)
and
181
A(fil(X I ) true (x ! E A I)
and
B(fjm(X m ) true ( x m E Am )
and
fil(f!n(Xn)) = f~_(x
)± n m = fik(fkn(Xn )) E A i ~fXn E A n )
and
conjunction introduction,
A(fil(Xl)) true (x I E A l)
with 1 ~ i , and
B(fjl(Xl)) true (x I E A l)
a E A B(a) true
( 3 x E A)B(x) t r u e '
Let
a(x k) E A(fik(Xk)) (x k E A k)
B(fjl(Xl),a(fkl(Xl))) (x I E AI),
c(x k) E ( ~ x E A(fik(Xk)))B(fjk(Xk),X) (x k E A k)
and
element a 0 E A O. Let
? = ?o = id(?1)'
?1 = id(?2),
?2 = id(?3)'
( ~ x o ~ A0)(A(x o) ~ B(x0))
( ( ~ x 0 E A0)A(x O) ~ ( V x 0 ~ A0)B(Xo)) true
x 0 E Ao, that is, a family of sets over AO, and B(Xo,X) a pro-
rem is finished.
I87
Miscellaneous examples
cepts.
I
Even(a) = (7~y E N)I(N,a,2.y),
nor do we have
standard truth entails the nonstandard truth of A(a) for all in-
finite a ~ N.
To see this, assume
A(~) true.
A(si(x)) true (x ~ N)
I~ +b =b~N,
(a) + b = s(a + b) ~ N,
we h a v e
si(x) = si(o) + x ~ N (x ~ N ) ,
A(si(O) + x) true (x E N)
189
by preservation of truth under definitional equality. From this,
we get
But
by d e f i n i t i o n , so t h a t , a g a i n by p r e s e r v a t i o n of tr~tfl under
definitional equality,
sJ(0) ~ a true
si(0) ~ a true.
A(a) true
190
a = f(~i ) ~ N
f ( x ) mN (x ~ N )
sJ(o) ~ f ( ~ i ) true
f
P(o) = J_,
e(s(x)) = ~P(x).
P(x) = T(rec(x,J.,(x,y)~y))
p(si(x)) true (x ~ N)
(O,(s(O),(s(s(O)),...)))
192
function
f(1,n) ~ List(N) (I ~ N, ~ e N )
I
f(O,n) = nil @ List(N),
f(l,n) = app(g(1),n),
g(O) = ( A n ) n i l e N--~List(N),
g(1) = r e c ( l , ( ~ n ) n i ! , ( x , y ) ( k n ) ( n , a p P ( y , s ( n ) ) ) )
N-~List(N)
derive
f(~,O) 6 List(N)
f(~,o) = f(s(~]),o)
= (O,f(~l,S(O))) = (O,f(s(~2),s(O)))
= (O,(s(O),f(~2,s(s(o))))) . . . .
193
I
A(n) set (n ~ N),
a(n) ~ A(n) (n E N)
~i e A(si(O))
by the equations
where
l
h(0) = ( ~ n ) a ( n ) ~ (]Tn ~ N)A(n),
h(1) = rec(1,(~n)a(n),(x,y)(kn)f(n,app(y,s(n))))
(Tin N)A(n).
i = g ( ~ i 'si(O)) e A(si(O))
~i = g ( ~ i 'si(O))
= g(s(~i+1),si(O))
= f(si(O),g(~i+1,s(si(O))))
= f(si(O),g(~i+1,si+1(O)))
B(0) = NI,
tions
B(~) = 3(s(~I))
= B(~]) + B ( ~ I) = B(s(oo2)) + B ( s ( ~ 2 ) )
integral of a function
f E B(c,o)-..~ Q,
S(n,f) E Q (n E N, f E B ( n ) - ~ Q )
S(0,f) : a p p ( f , 0 1 ) ,
S(n,f) = app(F(n),f),
where
F(n) E (B(n)--~Q)--~Q (n ~ N)
I F(O) = ( k f ) a p p ( f , 0 1 ) ,
F(s(n)) = (kf)(app(F(n),(kx)app(f,i(x)))
+ app(F(n),(~y)app(f,j(y)))).
I(n,f) E Q (n E N, f E B(n)~-~Q)
is defined by putting
I(n,f) : s(n~f)
2n '
the set B(n). We can now express the searched for integral of
f E B(~)--~Q simply as
197
I(~,f) = s,(~f)
2~
where
g(x) E B(si(x))--~Q (x E N)
numbers
l(si(x),g(x)) E Q (x E N).
l(sJ(0),g(sJ-i(0))) E q.