Indiscernible Extraction in Theories with Bounded k-splitting

Apply Lemma 2.4 with r=k and take the first \lambda^+ tuples of the resulting subsequence. This is a k-end-homogeneous sequence of length \lambda^+. Lemma 2.5 applies since \operatorname{cf}(\lambda^+)=\lambda^+>\lambda. A final segment of \lambda^+ has order type \lambda^+.

By Lemma 2.4, the given sequence (a_i)_{i<\beth_{k-1}(\lambda)^+} has a (k-1)-end-homogeneous subsequence (b_j)_{j<\lambda^++(k-1)} over B, where b_j=a_{\xi_j} and the indices \xi_j are strictly increasing. Lemma 2.3 shows that the hypothesis of Lemma 2.7 is satisfied, so we may apply it to the first \lambda^+ tuples gives a B-indiscernible subsequence indexed by a cofinal Y\subseteq\lambda^+ of order type \lambda^+, say (c_i)_{i<\lambda^+}. Then the sequence (c_i)_{i<\lambda^+}\;^\frown (b_{\lambda^+},\ldots,b_{\lambda^++k-2}) is a B-indiscernible sequence.

Let Q be the set of extensions, and assume Q\ne\varnothing. By Fact 2.2, every q\in Q does not k-split over some C\subseteq A of size at most {|T|}. There are at most (|A|+{|T|})^{{|T|}} possible choices for C.

Fix C and a realization a_0,\ldots,a_{k-1} of a member of Q. For each m<\omega, define an equivalence relation \sim_C on A^m by setting

There are at most 2^{|T|} classes for each m. For each m, choose one tuple from each equivalence class in A^m, and let D_C consist of C and all coordinates of the chosen tuples. Then |D_C|\leq2^{{|T|}}.

If q\in Q does not k-split over C, and \bar b' represents \bar b, non-k-splitting gives

Thus q is determined by q\mathbin{\upharpoonright} D_C. There are at most 2^{{|T|}+|D_C|}\leq2^{2^{{|T|}}} such restrictions, and so summing over C proves the bound.

For r=1, put \kappa=(2^\lambda)^+ and write the given sequence as (a_i)_{i<\kappa}. Recursively define families \mathcal P_\alpha of nonempty subsets of \kappa, for \alpha\leq\lambda^+. Let \mathcal P_0 partition \kappa by i\sim j if and only if a_i\equiv_Ba_j. For X\in\mathcal P_\alpha, let D_X consist of B and the coordinates of a_{\min(X_\xi)} for \xi<\alpha, where X_\xi\in\mathcal P_\xi is the unique set containing X. For \alpha<\lambda^+, form \mathcal P_{\alpha+1} by partitioning X\setminus\{\min(X)\}, for each X\in\mathcal P_\alpha, under i\sim_Xj\Longleftrightarrow a_i\equiv_{D_Xa_{\min(X)}}a_j. At a nonzero limit \alpha, take all nonempty intersections \bigcap_{\xi<\alpha}X_\xi of decreasing chains with X_\xi\in\mathcal P_\xi.

Inductively, \mathcal P_\alpha partitions the indices not selected as an earlier \min(X), and a_i\equiv_{D_X}a_j for i,j\in X. For \alpha<\lambda^+, |\mathcal P_\alpha|\leq2^\lambda: at stage zero and at successors the relevant parameter sets and the variable blocks have size at most \lambda, and at limits there are at most (2^\lambda)^{|\alpha|}=2^\lambda chains. Thus at most \lambda^+\cdot2^\lambda=2^\lambda indices are selected before stage \lambda^+, so \mathcal P_{\lambda^+}\ne\varnothing. Choose X_{\lambda^+}\in\mathcal P_{\lambda^+} and its containing sets X_\alpha\in\mathcal P_\alpha for \alpha<\lambda^+. Their minima i_\alpha=\min(X_\alpha), \alpha\leq\lambda^+, strictly increase. Since i_\xi,i_\zeta\in X_\alpha for \alpha\leq\xi,\zeta\leq\lambda^+, equality of types over D_{X_\alpha} shows that (a_{i_\alpha})_{\alpha\leq\lambda^+} is end-homogeneous over B.

Now suppose r > 1. Put \mu = \beth_{r-1}(\lambda), and put \rho = \lambda^+ + (r-1). The case r=1 gives an end-homogeneous subsequence (b_i)_{i<\mu^+} ^\frown (d) over B. Since {|T|} + |Bd| \leq \lambda, induction over Bd gives a (r-1)-end-homogeneous subsequence (e_\xi)_{\xi < \rho} of (b_i)_{i < \mu^+}. Then the sequence (e_\xi)_{\xi \leq \rho} (setting e_\rho = d) is r-end-homogeneous over B and has order type \lambda^+ + r.

Let (a_i)_{i<\theta} be the sequence and put A_\alpha=Ba_{<\alpha}. For each \alpha<\theta, define

End-homogeneity makes p_\alpha independent of the choice of I and gives p_\alpha\subseteq p_\beta for \alpha<\beta. Thus p=\bigcup_{\alpha<\theta}p_\alpha is a complete consistent partitioned type over Ba_{<\theta}. By Fact 2.2, choose C\subseteq Ba_{<\theta} of size at most {|T|}+|\gamma|\leq\lambda over which p does not k-split. Choose \beta<\theta with C\subseteq A_\beta, using \operatorname{cf}(\theta)>\lambda. For every I\in[\theta\setminus\beta]^k,

does not k-split over A_\beta, which is also true of the realization.

Enumerate the final segment (a_i)_{\beta\leq i<\theta} in order as (e_i)_{i<\eta}, and put D=A_\beta. It is k-end-homogeneous over D, hence k-indiscernible over D. Suppose it is m-indiscernible, where m\geq k. Write any increasing (m+1)-tuple of its blocks as uv, with v=(v_0,\ldots,v_{k-1}) its last k blocks, and put u_0=e_0\cdots e_{m-k}. For each j<k, m-indiscernibility gives

Non-k-splitting of the type of v over its preceding segment gives uv\equiv_Du_0v, and k-end-homogeneity gives u_0v\equiv_De_0\cdots e_m. For infinite blocks, each formula uses only finite coordinate subtuples of u,u_0. This proves (m+1)-indiscernibility. The final segment is therefore D-indiscernible, hence B-indiscernible.

Let (a_i)_{i<\theta} be the given sequence, where \theta=\lambda^+. Put A_\alpha=B\cup a_{<\alpha} and A=B\cup a_{<\theta}. Adjoin tuples c_1,\ldots,c_{k-1} in an elementary extension so that (a_i)_{i<\theta}\mathbin{{}^\frown}(c_1,\ldots,c_{k-1}) is still (k-1)-end-homogeneous over B. To do this, realize the increasing union

Put c=c_1\cdots c_{k-1}. Then we have, for every \alpha<\theta and 1\leq j<k,

By Fact 2.2, choose C_\alpha\subseteq A_\alpha, |C_\alpha|\leq{|T|}, such that \operatorname{tp}(a_\alpha;c_1;\ldots;c_{k-1}/A_\alpha) does not k-split over C_\alpha. We claim that

Indeed, if \bar b,\bar b'\in A_\alpha^{<\omega} satisfy \bar b\equiv_{C_\alpha c}\bar b', then (?) transfers this equality to the parameters obtained by omitting any one block from a_\alpha,c_1,\ldots,c_{k-1}. Non-k-splitting therefore gives \bar b\equiv_{C_\alpha a_\alpha c}\bar b'. For tuples from A_\alpha c, equality over C_\alpha c forces coordinates belonging to c to match. Delete those coordinates, apply the preceding argument, and reinsert them.

On the stationary set \{\alpha<\theta:\operatorname{cf}(\alpha)={|T|}^+\}, the coordinates of C_\alpha\setminus B occur below some index \beta_\alpha<\alpha. Here {|T|}^+<\theta, since 2^{{|T|}}\leq\lambda^{{|T|}}=\lambda. Fodor’s lemma gives a stationary set on which \beta_\alpha is constant, say with value \beta. Since |A_\beta|\leq\lambda and \lambda^{{|T|}}=\lambda, we may thin further to a stationary set S on which C_\alpha is fixed, say with value C.

For each m<\omega, let R_m\subseteq A^m contain exactly one representative of each equivalence class for the equivalence relation defined by letting \bar b\sim_m\bar b'\Longleftrightarrow\bar b\equiv_{Cc}\bar b'. Let D consist of C and all coordinates of tuples in \bigcup_{m<\omega}R_m. Since |Cc|\leq{|T|}, we have |R_m|\leq2^{{|T|}} and hence |D|\leq2^{{|T|}}\leq\lambda. Choose \delta<\theta with D\subseteq A_\delta, and discard indices at most \delta from S. For \alpha\in S, put q_\alpha=\operatorname{tp}(a_\alpha,c_1,\ldots,c_{k-1}/A_\delta). Let p_i(x_0,\ldots,\widehat{x_i},\ldots,x_{k-1}) be \operatorname{tp}(c/A_\delta) with its variable blocks renamed in order (i<k). By (?), q_\alpha\supseteq\bigcup_{i<k}p_i. The hypothesis on common complete extensions, applied to these types over A_\delta, where |A_\delta|\leq\lambda, gives at most \lambda possible q_\alpha. Thin to a stationary S'\subseteq S on which they agree. Thus the a_\alpha, \alpha\in S', have the same type over A_\delta c.

If \alpha<\alpha' belong to S' and \bar b\in A_\alpha^m, choose \bar d\in R_m\subseteq D^m with \bar b\equiv_{Cc}\bar d. Applying (?) at both indices gives, for every formula \varphi,

Consequently a_\alpha\equiv_{A_\alpha c}a_{\alpha'}. The sequence (a_\alpha)_{\alpha\in S'} is end-homogeneous over A_\delta c. By (?), restriction and enlargement of the base show that each successive type does not split over A_\delta c. Lemma 1.4 therefore gives indiscernibility over A_\delta c, hence over B.

Set \lambda=2^\mu. Then \lambda\geq2^{2^{|T|}}, \lambda^{|T|}=\lambda, and \beth_{k-1}(\lambda)=\beth_k(\mu). Apply Theorem 2.6.

Choose an infinite {|T|}+|B|+|\gamma|\leq\lambda<\kappa. We first prove, by induction on r\geq1, that every sequence of \gamma-tuples of length \kappa has an r-end-homogeneous subsequence of order type \kappa.

For r=1, carry out the partition construction in Lemma 2.4 on the given index set \kappa, at every stage \alpha<\kappa. Its levels satisfy

The strict inequality uses the strong inaccessibility of \kappa. Indeed, the right-hand side bounds the number of types over the parameters on a branch of length \alpha, and at a limit it also bounds the number of preceding chains. Every \mathcal P_\alpha is nonempty by the regularity of \kappa. Order the nodes (\alpha,X), X\in\mathcal P_\alpha, by

Each node has exactly one predecessor at every previous level. After adjoining a root, this is a tree to which the tree property applies. The downward closure of a cofinal branch gives classes X_\alpha\in\mathcal P_\alpha for all \alpha<\kappa. Their minima strictly increase and index an end-homogeneous subsequence of length \kappa, by the same type equalities as in Lemma 2.4.

For r>1, first obtain an end-homogeneous subsequence (b_i)_{i<\kappa}. In an elementary extension choose a \gamma-tuple d realizing

The types in this union are well-defined and increasing, so the union is consistent and complete. Consequently (b_i)_{i<\kappa}\mathbin{{}^\frown}(d) is end-homogeneous over B. Apply induction over Bd to (b_i)_{i<\kappa}. Replacing the last block in each compared r-tuple by d, exactly as in Lemma 2.4, shows that the selected (r-1)-end-homogeneous sequence over Bd is r-end-homogeneous over B.

Take r=k. Lemma 2.5 applies at \theta=\kappa, since \operatorname{cf}(\kappa)=\kappa>\lambda. Its indiscernible final segment has order type \kappa.

Let \mathcal T_\mu be the tree of finitely supported functions f:[\beta,\mu)\to\omega, where \beta<\mu is zero or a successor ordinal, ordered by extension. The tree construction in [HansonBU] gives an s-indiscernible, bu-spread-out family (t_f)_{f\in\mathcal T_\mu} over B, with t_f\equiv_Ba for every f.

For each \xi<\mu, let f_\xi be identically zero on [\xi+2,\mu), and put d_\xi=t_{f_\xi}. For i<\omega, set g_i^\xi=f_\xi\cup\{(\xi+1,i)\} and let c_i^\xi be the concatenation of the tuples t_g for g\in\mathcal T_\mu with g\supseteq g_i^\xi. Use enumerations matched by the map that changes only the value at \xi+1. The domains of f_\xi and g_i^\xi start at successor ordinals, so both are nodes of \mathcal T_\mu. Then (c_i^\xi)_{i<\omega} is bu-Morley over B.

If \eta<\xi, then f_\eta\supseteq g_0^\xi, so every coordinate of d_\eta occurs in c_0^\xi. If \eta>\xi, then f_\eta is a restriction of all the subtrees rooted at g_i^\xi. The s-indiscernibility then(c_i^\xi)_{i<\omega} indiscernible over Bd_{>\xi}. This also holds for the whole subtree tuples, since each formula uses only finitely many coordinates.

Choose an infinite \lambda\geq{|T|}+|B|+|a| and put \mu=\beth_k(\lambda)^+. Theorem 2.1 gives extraction of order type \lambda^+, hence of order type \omega+\omega, over B for tuples of length |a|.

Choose the tuples supplied by Lemma 4.3 and extract a B-indiscernible subsequence (d_{\xi_j})_{j<\omega+\omega}, where the \xi_j strictly increase. Put

Then I+J is B-indiscernible. Every coordinate of I occurs in c_0^\rho, and (c_i^\rho)_{i<\omega} is indiscernible over BJ. Fact 4.1, applied with one tuple enumerating the whole of J, gives c_0^\rho\mathrel{\raise0.2ex\hbox{\ooalign{\hidewidth\hidewidth\cr\raise-0.9ex\hbox{}}}}^{\mathrm{bu}}_BJ. Monotonicity gives I\mathrel{\raise0.2ex\hbox{\ooalign{\hidewidth\hidewidth\cr\raise-0.9ex\hbox{}}}}^{\mathrm{bu}}_BJ, and Fact 4.2 shows that I is total bu-Morley over B. Finally, d_{\xi_0}\equiv_Ba: an automorphism fixing B sends d_{\xi_0} to a and sends I to the required sequence.