Monday, October 3, 2016

Generalizing Silver's Theorem Part 2

In this post, I will be proving Silver’s theorem the forcing way. Recall the statement.

Theorem 1   
Suppose that E is a lightface co-analytic equivalence relation on the reals (in this case taken to be Cantor space: infinite strings of 0’s and 1’s). Then either E has countably many equivalence classes, or there is a perfect set of E-inequivalent elements.
Proof.
From here on out I will be referring to lightface analytic sets as Σ11, co-analytic sets as Π11, and sets which are both as Δ11.

As in the topological proof, I will be using lightface analytic sets, and exploiting their properties. The first two of these properties are reasonably straightforward:
  • There are only countably many Σ11 sets, Π11 sets, and Δ11 sets.
  • Disjoint Σ11 sets can be separated by a Δ11 set. That is whenever A and B are disjoint Σ11 sets, there is a Δ11 set C so that AC and BC = ∅. This is equivalent to the reduction property for Π11. For any Π11 sets A and B, we can find disjoint Π11 sets A′ and B′ so that AA′, BB′ and AB = A′ ∪ B′.
The next property will require a little bit of set-up. In descriptive set theory, we try to code everything we possibly can with real numbers. In the case of countable collections, we try to do even better, and code them by natural numbers. In particular, we can obtain a set U × ω × ℝ so that for every Π11 set B, there is an n ∈ ω so that B = U ∩ ({n} × 2ω). That is, every Π11 set shows up as a vertical section of U. I am presenting these facts without proof because they are pretty standard and this post is going to be long as it is.

This proof really consists of test, the results of which provide construction principles. In words, the test is as follows: for every real x, is there a Δ11 set A so that xA ⊆ [x]E? If the answer is yes, then the fact that Δ11 is countable implies that E has countably many equivalence classes. We certainly cannot have more equivalence classes than elements of Δ11 in this case. So what happens if the answer is no? This is where we will use forcing to create a copy of 2ω which consists of pairwise E-inequivalent elements.

Let Q be the collection of non-empty Σ11 sets. For A,BQ, we say that AB (A extends B) iff AB. I will refer to elements of Q as conditions. This is because they include some reals and exclude others. Intuitively, extending conditions in this partial order corresponds to filtering out reals. Now let S = {x ∈ ℝ : (∃ A ∈ Δ11) [xA ⊆ [x]E} and set W to be the complement of S. If the answer to the question is no, W is non-empty. What I want now is for W to be a Σ11 set. That way I can force below W. The information that W imparts will guarantee that the forcing produces pairwise E-inequivalent reals.

Claim 1 
W is Σ11.
Reason
It suffices to show that S is Π11. We create a coding of the Δ11 sets in order to make this computation. Let U ⊆ ω × ℝ be universal for Π11. Define Π11 sets V1 and V2 by
  • V1(⟨ n,m ⟩,x)  ⇐⇒ U(n,x) and
  • V2(⟨ n,m ⟩,x)  ⇐⇒ U(m,x).
Using the reduction property for Π11, we can find disjoint Π11 sets V1′ and V2′ so that ViVi′, V1′ ∩ V2′ = ∅, and V1V2 = V1′ ∪ V2′. Then ⟨ n,m ⟩ codes a Δ11 set iff
(∀ y ∈ ℝ)[V1′(⟨ n,m ⟩,y) ∨ V2′(⟨ n,m ⟩,y)],
that is ⟨ n,m ⟩ finds a section of U which codes complementary Π11 sets. Call the set of such codes C. C is Π11. Now
x ∈ S  ⇐⇒ (∃ n,m ∈ ω)[⟨ n,m ⟩ ∈ C ∧ V1′(⟨ n,m ⟩,x) ∧ (∀ y ∈ ℝ)[¬ V2′(⟨ n,m ⟩,x) ⇒ x E y]].
This shows that S is Π11.


With this knowledge of W in hand, it makes sense to speak of the conditions in Q below W. Call this poset P. I will use this poset to produce pairwise E-inequivalent reals. For the uninitiated, forcing with a poset involves three major pieces.
  • Dense subsets of P(these are dense in the intuitive way; every condition can be extended to be inside the set). Topologically these correspond roughly to comeager sets.
  • Generic filters for P: G is a generic filter if it is closed upwards in the ≤ order, is closed under "meets" in the sense that if p,qG, then there is an rG so that rp,q, and G has non-empty intersection with every dense subset of P.
  • Names for objects which can be described by P. These are essentially collections of pairs (A,τ) where A is a condition in P and τ is another name. The construction of names is inherently recursive.
These three objects come together in a synthesis. G builds actual objects using the names as a guide: it pulls objects into a set it is building according to the conditions in G and the name that condition is tagged with. This process is recursive. In notation, the interpretation of the name τ by the generic G is written as τG. The dense sets are how the poset controls what G can build. The creation of clever dense sets forces G to make decisions about desired properties of the sets it constructs. If we call V our initial universe of objects, V[G] is the universe of objects which can be created from G in this way. Standard forcing theorems show that V is properly contained in V[G] and G in an element of V[G]. Syntactically, this means there are names for elements of V and for G which get interpreted appropriately for all generics G. If xV, the standard name for x is denoted ẋ.

For sentences ϕ (which are allowed to reference names) and conditions A, the notation A ⊩ ϕ means that whenever G is generic, if AG, then ϕ is true in V[G]. This symbol means that A forces ϕ. It is an important theorem in forcing that if a statement ϕ is true about V[G], then it was forced to be true by some condition in the poset. To work with the forcing in this context, I will restrict down to a countable structure M for which enough of ZFC is true for the proofs to go through. This allows for the explicit construction of generics. Actually, what I want is a little bit stronger. By first taking a sub-universe of enough of ZFC, and then taking a countable elementary substructure M, I can assume that there is an embedding ϕ:MV which preserves all of the first order statements which come up in this proof. This small universe does not have Q or P, but M has ϕ−1(P) and ϕ−1(Q), which I will refer to as P and Q. As general notation, the bar will be used to denotes the ϕ-inverse of a set.
Now the poset Q I am considering has a particularly nice interplay with its generics.

Claim 2   
Suppose G is generic filter on Q. Then G gives rise to a real xG so that xGA iff AG.
Reason
I use cylinder sets to construct the name. For a finite string s of 0’s and 1’s, N(s) is those reals which have s as an initial segment. Consider the name ẋ = {(ṡ,A) : AN(s)}. Then if G is a generic filter for Q over M, this name at least guarantees that ẋG ∈ ℝ. It suffices to show that A ⊩ ẋ ∈ A for every AP. Since A is Σ11, I can write A = p[T], where T is a recursive tree on ω × ω. Then
A ⊩ ẋ ∈ A  ⇐⇒ A ⊩ (∃ y ∈ ℝ)(∀ n ∈ ω)[(ẋ|n,y|n) ∈ T].
I will first check that this does not depend on our choice of T. A could have many representations, and this computation should not depend on the representation. So Suppose T0 and T1 are recursive trees on ω × ω so that A = p[T0] = p[T1]. Suppose G is generic for P over M. I need to see that
ẋ ∈ p[T0]  ⇐⇒ ẋ ∈ p[T1].
In M, it is true that (∀ x ∈ ℝ)[xp[T0]  ⇐⇒ xp[T1]]. Since this is a simple enough statement (Π21), Shoenfield absoluteness declares it true in M[H] as well no matter which generic H is chosen. So the forcing statement is true.

Now suppose towards a contradiction that A ⊮ẋ ∈ A. It is part of the formal definition of ⊩ that there then must be a condition BA so that B ⊩ ẋ ∉ A. Then B ⊩ ẋ ∉ B. Thus WLOG A ⊩ ẋ ∉ A. Call this element A0. This is not quite a contradiction, but it is almost there. Write A0 = p[T0]. I will use A0 to build a generic extension G so that in M[G], ẋGA and ẋGp[T0]. In V, let {Dn : n ∈ ω} be the dense subsets of Q from M. I want to fix the first coordinate of the reals in A0, fix a valid first coordinate for a witness that those reals are in A0 and extend A0 to meet D1. Let s0 ∈ ω and t00 ∈ ω be so that
A0′ = {x : (∃ y ∈ ℝ)[(x(0),y(0)) = (s,t00) ∧ (x,y) ∈ [T0]]} ⊆ A0.
This fixes the first digit of reals inside A0 and the first digit of potential witness y which prove that those reals are in A0. Let A1A0′ ∩ N(s0) be so that A1D1.

Continue this way inductively, defining An, Tn, sn, tnm, and An′ for n ∈ ω and mn so that
  • An = p[Tn],
  • sn,tnm ∈ ωn,
  • snsn+1,
  • tnmtn+1m,
  • An′ = {x : snx ∧ (∃ y ∈ ∏mnN(tnm))(∀ mn)[(x,ym) ∈ [Tm]] },
  • AnDn, and
  • AnAn′ ≥ An+1.
We can then define a filter G by taking the ≤-upward closure of the An. Explicitly, define a filter G on P by
A ∈ G  ⇐⇒ (∃ n ∈ ω)[An ≤ A].
This is a generic filter for Q. By construction, ẋG =limn → ∞sn. I also assumed that A0 ⊩ ẋ ∉ A0. Thus ẋGA0. Let y = limn → ∞tn0. Then (ẋG,y) ∈ [T0]. So ẋGp[T0] = A0. This is a contradiction.


With this in mind, I will freely identify a real x as Q-generic (or P-generic). At this point, to meaningfully interact with the equivalence relation, I need to be working in ℝ2. Unfortunately, there are two ways to lift Q and P up to ℝ2. The first way is the simpler, and will be denoted Q × Q. Conditions in Q × Q are simply ordered pairs (A,B), where A,BQ. The second way is slightly more complicated. Q2 is the version of Q defined off of ℝ2. That is AQ2 iff A ⊆ ℝ2, A is Σ11 and A ≠ ∅. Syntactically, this means there is a recursive tree T on ω × ω × ω so that (x,y) ∈ A iff there is a z ∈ ℝ so that (x,y,z) ∈ [T].

Let P × P be the product poset of P, and P2 = {AQ2 : AW × W}. Note that if G is Q2 generic (or P2-generic), then G defines a pair of reals (xG,yG). With the name (x,y) = {(ṡ,ṫ,A) : AN(s) × N(t)}, the same proof as for one dimension shows that
(x,y)G ∈ A  ⇐⇒ A ∈ G.

Now the question is how do these two-dimensional posets interact with the one-dimensional version. The answer is quite well! I just have to define some projection operations, and then there is a smooth translation between the posets. For AQ2, we let
πL(A) = {x ∈ ℝ : (∃ y ∈ ℝ)[(x,y) ∈ A]}  and  πR(A) = {y ∈ ℝ : (∃ x ∈ ℝ)[(x,y) ∈ A]}.
Then πL(A),πR(A) ∈ Q, as Σ11 is closed under projections. If G is Q2-generic, let
πL(G) = {A ∈ 
Q
 : (∃ B ∈ G2)[πL(B) ⊆ A]},
and πR(G) is similarly defined.

One particular nice interaction is that generic reals pass back and forth. I am proving this for Q2 and the barred version for notational simplicity and because I can.

Claim 3 
If (x,y) is Q2 generic, then x and y are generic for Q.
Reason
Suppose (x,y) is Q2 generic. Let G2 be the associated Q2 generic. I will show that x is Q-generic, and the argument for y is symmetric. Let G = πL(G2). I will show that G is generic. Let D be dense for Q. Let A2Q2, and set A = πL[A2]. Then as AQ, there is a BA in Q so that BD. Define B2 by
B2 = (B × ℝ) ⋂ A2.
Then B2 is Σ11 and B2 ≠ ∅. So B2Q2. Note that πL[B2] = BD. Thus {XQ2 : πL[X] ∈ D} is dense in Q2. So there is an XG so that πL[X] ∈ D. Thus GD ≠ ∅, and so G is generic. Moreover,
x ∈ A  ⇐⇒ (x,y) ∈ A × ℝ  ⇐⇒ A × ℝ ∈ G  ⇐⇒ A ∈ G.
So G is the the generic associated to x.


Putting this all together, it finally comes out that forcing below W produces reals which are pairwise E-inequivalent.

Claim 4   
If (x,y) are P × P-generic, then (x,y) ∉ E.
Reason
BWOC suppose that (A,B) ∈ P × P are so that (A,B) ⊩P2 ((x,y) ∈ E). Consider (A,A,B) ∈ P2 × P. Note that A2 ∩ (ℝ2E) is in Σ11 and non-empty. Thus A2 ∩ (ℝ2E) ∈ P2. Let G2 be P2-generic over V so that AG2. Let G be P-generic over V[G2] so that BG. Then in V[G2][G], G2 × G defines a generic triple (x,y,z). Now
  • in V[G2], (x,y) ∉ E (as A2G2),
  • in VL(G2)][G], x E z (as BG, (x,z) is P × P-generic, and A × B ∈ πL(G2) × G) and
  • in VR(G2)][G], y E z (as BG, (y,z) is P × P-generic, and A × B ∈ πR(G2) × G).
Thus,
in V[G2][G], [(x E z) ∧ (y E z) ∧ ((x,y) ∉ E)].
Therefore In V[G2][G], E is not an equivalence relation. But to say E is an equivalence relation is a Π21 statement. So by Shoenfield absoluteness again, in V[G2][G], E is an equivalence relation. This is a contradiction.


Now I will build a perfect set of generics for P. To do so, I have to return to the countable sufficiently elementary submodel M. Note that by elementarity,
M ⊨ [∅ ⊩P_2 (x,y) ∉ E].
The construction is an inductive construction. I will create Σ11 sets As for all finite strings s in such a way that incompatible strings correspond to disjoint Σ11 sets and each descending sequence of A’s produces a generic. If I can follow this through, the set of such generics will be perfect and they msut also be pairwise E-inequivalent.

Let {Dn : n ∈ ω} enumerate the dense sets in P × P. Let (A0,A1) ∈ P2 be so that A0P ẋ(i0) = a0 and A1P ẋ(i0) = b0 with a0b0. In P × P there are A00, A01, A10, A11 so that A00A01 = ∅ and A10A11 = ∅. I can then find AijAij so that AijD0. I simply continue in this way, defining As,AsP for s ∈ 2 so that
  • for all s,t ∈ 2, if st, then AtAs,
  • for all s ∈ 2, AsAs,
  • for all s ∈ 2, As ⁀ 0As ⁀ 1 = ∅, and
  • for all s ∈ 2, AsD|s|−1.
The easiest way to check that this construction has worked is through the language of embeddings. Define ψ:2ω→ ℝ by ψ(x) = ∩n Ax|n. If x ∈ 2ω, then the filter generated by x (AGx  ⇐⇒ ψ(x) ∈ A) is a generic filter for P over M. In fact, if xy ∈ 2ω, then (Gx,Gy) is P × P-generic over M. Thus in M[(Gx,Gy)], (ψ(x),ψ(y)) ∉ E. By Σ11-absoluteness, it must then be true in V that (ψ(x),ψ(y)) ∉ E. This verifies the construction and finishes the proof


This version is reasonably slick, and doesn’t use the axiom of choice in ways which are incompatible with AD. However, it doesn’t actually generalize all that well. If instead of being co-analytic, the equivalence relation is merely co-κ-Suslin (a set A is κ-Suslin if it can be written as A = p[T], where T is a tree on ω × κ), the same proof will not got through as I will lose features of Σ11 which made some of these complexity computations come out just so. So what’s the solution? Infinitary logic and admissible ordinals. Or, in my case, infinity Borel codes and admissible ordinals. Look for that next time.
This document was translated from LATEX by HEVEA. (How did it work?)

No comments:

Post a Comment