n-DEPENDENT CONTINUOUS THEORIES AND

HYPERDEFINABLE SETS
arXiv:2405.19830v1 [math.LO] 30 May 2024

ADRIAN PORTILLO FERNÁNDEZ

Abstract. We define the continuous modeling property for first-order struc-

tures and show that a first-order structure has the continuous modeling prop-
erty if and only if its age has the embedding Ramsey property. We use gen-
eralized indiscernible sequences in continuous logic to study and characterize
n-dependence for continuous theories and first-order hyperdefinable sets in
terms of the collapse of indiscernible sequences.

1. Introduction
Continuous model theory is a growing area of model theory that has been de-
veloping very fast in recent years. Many of the most important dividing lines for
first-order theories have been also defined for continuous theories (Stability [BYU10;
BY10], NIP [BY09], Distality [And23]). One invaluable tool for the characteriza-
tion of dividing lines in first-order theories is (generalized) indiscernible sequences
(See [She82, Chaper VII], [Sco12], [CPT14], [GH19]). We present natural contin-
uous counterparts of generalized indiscernibles and the modeling property (where
the index structures are still first-order) and show that a first-order structure has
the continuous modeling property (see Definition 3.4) if and only if its age has the
embedding Ramsey property (Theorem 3.10). Several notions around this topic
have been also defined in positive logic. Dobrowolski and Kamsma (see [DK21])
proved that s-trees have the (positive logic version of) modeling property in thick
theories, later in [Kam23, Theorems 1.2, 1.2 and 1.3] it was shown that str-trees,
str0 -trees (the reduct of str-trees that forgets the length comparison relation) and
arrays also have the modeling property in positive thick theories.
The notion of a dependent theory was first introduced by Shelah in [She82].
In later work [She05; She07], Shelah introduced the more general notion of n-
dependence. This notion was studied in depth in [CPT14], where the authors give
a characterization of n-dependent theories in terms of the collapse of indiscernible
sequences (See [CPT14, Theorem 5.4]). In the continuous context, the definition
of n-dependence was introduced in [CT20] using a generalization of the V Cn di-
mension. Section 10 of the aforementioned paper is dedicated to several operations
which preserve n-dependence. The proof of [CPT14, Theorem 5.4] contains a mis-
take 1 in the implication (3) =⇒ (2); we provide a counterexample to the key
claim (see Counterexample 4.14). Using the tools developed in Section 3 we give

Date: January 2024.

2020 Mathematics Subject Classification. 03C45.
Key words and phrases. n-dependence, continuous logic, hyperimaginary.
The author was supported by the Narodowe Centrum Nauki grant no. 2016/22/E/ST1/00450.
1After sending the manuscript to the authors of [CPT14], they acknowledged that there is a
mistake and proposed a short correction that we discuss at the end of Section 4
1
2 ADRIAN PORTILLO FERNÁNDEZ

an alternative proof of the theorem, obtaining generalizations of [CPT14, Theorem

5.4] to continuous logic theories (Theorem 1.1) and hyperdefinable sets (Theorem
1.3).
Let Gn+1,p be the Fraı̈ssé limit of the class of ordered (n + 1)-partite (n + 1)-
uniform hypergraphs and Gn+1 be the Fraı̈ssé limit of the class of ordered (n + 1)-
uniform hypergraphs. By a Gn+1,p -indiscernible sequence (ag )g∈ Gn+1,p we mean
that for any W, W ′ ⊆ Gn+1,p if the quantifier free types of W and W ′ coincide (in
the language Lopg defined in Section 4), then the types of the tuples (ag )g∈W and
(ag )g∈W ′ also coincide (similarly for Gn+1 -indiscernibility, see Definition 3.1). We
say that the sequence (ag )g∈ Gn+1,p is Lop -indiscernible if for any W, W ′ ⊆ Gn+1,p
with the same quantifier free type in the language Lop (see Section 4), the types of
the tuples (ag )g∈W and (ag )g∈W ′ also coincide.
Theorem 1.1. Let T be a complete continuous logic theory. The following are
equivalent:
(1) T is n-dependent.
(2) Every Gn+1,p -indiscernible is Lop -indiscernible.
(3) Every Gn+1 -indiscernible is order-indiscernible.
We introduce the following definition of n-dependent hyperdefinable sets:
Definition 1.2. The hyperdefinable set X/E has the n-independence property,
IPn for short, if for some m < ω there exist two distinct complete types p, q ∈
SX/E×Cm (∅) and a sequence (a0,i , . . . , an−1,i )i<ω such that for every finite w ⊂ ω n
there exists bw ∈ X/E such that
tp(bw , a0,i0 , . . . , an−1,in−1 ) = p ⇐⇒ (i0 , . . . , in−1 ) ∈ w

tp(bw , a0,i0 , . . . , an−1,in−1 ) = q ⇐⇒ (i0 , . . . , in−1 ) ∈

/ w.
Using the tools developed in Sections 3 and 4, we prove the theorem below,
which is a counterpart for hyperdefinable sets of Theorem 1.1. Here, Ψn+1
FX/E is a
special family of functions (see Notation 3 in Section 5).
Theorem 1.3. The following are equivalent:
(1) X/E is n-dependent.
(2) Every Gn+1,p -indiscernible (ag )g∈Gn+1,p , where for every g ∈ P0 (Gn+1,p )
we have ag ∈ X/E, is Lop -indiscernible.
(3) For every m ∈ N, every Gn+1 -indiscernible with respect to Ψn+1
FX/E sequence
of elements of Cm × X is order-indiscernible with respect to Ψn+1
FX/E .

Here P0 (Gn+1,p ) is the first part of the partition of Gn+1,p , and by a Gn+1 -
indiscernible sequence (ag )g∈ Gn+1 with respect to Ψn+1FX/E we mean that for any
W, W ⊆ Gn+1,p if the quantifier free types of W and W ′ coincide (in the language
′

Log defined in Section 4), then the Ψn+1

FX/E -types of the tuples (ag )g∈W and (ag )g∈W ′
also coincide.
Item (3) of Theorem 1.3 cannot be improved by replacing indiscernibility with
respect to Ψn+1
FX/E by indiscernibility with respect to more general families of func-
′
n +1
tions from F(Cm ×X/E)n+1 or by replacing F(Cm ×X/E)n+1 by ΨFX/E
with n′ > n as
showed by Example 5.11.
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 3

2. Preliminaries
2.1. Continuous logic. We present some basic definitions and facts about con-
tinuous logic needed for this paper, we refer the reader to [BYU10] and [BY10] for
a more detailed exposition.
Definition 2.1. A continuous (metric) signature consists of:
• A collection of function symbols f, together with their arity nf < ω.
• A collection of predicate symbols P, together with their arity nP < ω.
• A binary predicate symbol, denoted d, specified as the distinguished distance
symbol.
• For each n-ary symbol s and i < n a continuity modulus δs,i , called the
uniform continuity modulus of s with respect to the ith argument.
Given a continuous signature L, the collection of L-terms and atomic L-formulas
are constructed as usual. In the continuous context, the quantifiers supx and inf x
play the roles of ∀ and ∃ respectively. The issue of connectives is a bit more delicate
and we refer the reader to [BYU10] for an in depth treatment. Depending on the
context, we will consider all uniformly continuous functions u : [0, 1]n → [0, 1] for
all n < ω as connectives or just some finite subset of such functions.
A condition is an expression of the form φ = 0 where φ is a formula. Note
that expressions of the form φ ≥ r and φ ≤ r can be expressed as conditions.
A continuous L-theory T is a consistent set of L-conditions φ = 0 where φ is a
sentence.
A complete n-type in variables x is a maximal satisfiable set of conditions with
free variables x1 , . . . , xn ⊆ x for some n. The space of all types over the empty set
in variables x is denoted by Sx . If x = (x1 , . . . , xn ), we denote Sx by Sn . The space
Sx is a compact Hausdorff space when equipped with the finest topology for which
all continuous formulas φ are continuous functions φ : Sx → [0, 1].
Definition 2.2. A definable predicate f in variables x is a continuous function
f : Sx → [0, 1].
The following was proven in [BYU10, Proposition 3.4] for a finite number of
variables but the proof also applies for an infinite x.
Fact 2.3. Definable predicates in variables x can be uniformly approximated by
continuous formulae in variables contained in x.
By an abuse of notation, when results apply to both, we will usually also refer
to definable predicates as formulas. We need to allow the domain of definable
predicate to be an infinite Cartesian power of C to deal with hyperdefinable sets
X/E which are contained in an infinite product of sorts.

2.2. The embedding Ramsey property. Let L′ be a first-order language. Given

′ B


L -structures A ⊆ B, we write A for the set of all embeddings from A into B.
Definition 2.4. Let A ⊆ B ⊆ C be L′ -structures and let r ∈ N. We write
C → (B)A
r
C C



if for each coloring χ : A → r there exists some f ∈ B such that χ ↾f ◦(B) is
A
constant.
4 ADRIAN PORTILLO FERNÁNDEZ

Definition 2.5. Let L′ be a first-order language and C be a class of finite L′ -

structures. We say that C has the embedding Ramsey property, ERP for short, if
for every A ⊆ B ∈ C and r < ω there is C ∈ C such that C → (B)A
r .

Note that the following holds:

Fact 2.6. If a class of finite L′ -structures has ERP , then all the structures of C
are rigid (i.e. they have no nontrivial automorphisms).
Remark 2.7. Sometimes, the symbol B


A is used to denote the set of all isomorphic
copies of A in B. If the class C consists of finite L′ -structures which are rigid, then
coloring embeddings from A into B is equivalent to coloring substructures A ⊆ B.

3. Generalized indiscernibles
Let L′ be a first-order language and L be a continuous logic language. Unless
specified otherwise, T is a complete continuous L-theory with C |= T a monster
model (i.e. κ-saturated and strongly κ-homogeneous for a strong limit cardinal
> |T |) and I, J are L′ -structures.
In this section, we present natural adaptations of the concepts of generalized
indiscernibles and the modeling property to continuous logic and give a charac-
terization of the continuous modeling property in the form of a continuous logic
counterpart of [Sco21, Theorem 2.10].
The following idea first appeared in [She82, Definition VIII.2.4].
Definition 3.1. Let I = (ai : i ∈ I) be an I-indexed sequence, and let A ⊂ C be a
small set of parameters. We say that I is an I-indexed indiscernible sequence over
A if for all n ∈ ω and all sequences i1 , . . . , in , j1 , . . . , jn from I we have that
tpqf (i1 , . . . , in ) = tpqf (j1 , . . . , jn ) =⇒ tp(ai1 , . . . , ain /A) = tp(aj1 , . . . , ajn /A).
We will refer to I-indexed indiscernible sequences as I-indiscernibles.
Next, we adapt the definition of locally based on given in [Sco15]. The first
reference to this concept can be found in [Zie88].
Definition 3.2 (Locally based on). Let I = (ai : i ∈ I) be an I-indexed sequence.
We say that a J -indexed sequence (bj : j ∈ J ) is locally based on I if for any finite
set of L formulas ∆, any finite tuple j ⊆ J and ε > 0 there is i ⊆ I such that:
(1) tpqf (i) = tpqf (j).
(2) |φ(bj ) − φ(ai )| ≤ ε for all φ ∈ ∆.
The original definition presented in [Sco15, Definition 2.5] is the following:
Definition 3.3 (Classical definition of Locally based on). Let I = (ai : i ∈ I) be
an I-indexed sequence. We say that a J -indexed sequence (bj : j ∈ J ) is locally
based on I if for any finite set of L formulas ∆, any finite tuple j ⊆ J and ε > 0
there is i ⊆ I such that:
(1) tpqf (i) = tpqf (j).
(2) tp∆ (bj ) = tp∆ (ai ).
Note that if we tried to use this stronger version of the property, it is easy to show
that even for I = (N, <) we can find a sequence for which there are no indiscernible
sequences locally based on it. Consider for example the theory Th([0, 1], d) where
d is the distance predicate and the sequence (1/n)n<ω .
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 5

The next definition is then the natural continuous counterpart of [Sco12, Defini-
tion 2.17].
Definition 3.4 (Continuous Modeling property). Given a continuous theory T ,
we say that I-indexed indiscernibles have the continuous modeling property in T if
given any I-indexed sequence I = (ai : i ∈ I) in a monster model C of T there exists
an I-indiscernible sequence (bi : i ∈ I) in C locally based on I. We say that I has
the continuous modeling property if I-indexed indiscernibles have the continuous
modeling property in every continuous theory.
As it is natural, if a first-order structure I has the continuous modeling property
then it has the modeling property. More precisely:
Proposition 3.5. Let T be a first-order theory, and let T ′ be its continuous logic
counterpart (i.e., T and T ′ have the same models). Then I has the continuous
modeling property in T ′ if and only if I has the modeling property in T .
Proof. Clearly, If I has the continuous modeling property in T ′ then it has the
modeling property in T since classical formulas are a subset of the {0, 1}-valued
continuous logic formulas.
Assume now that I has the modeling property in T . Let I = (ai : i ∈ I) be any
sequence in C |= T . Since I has the modeling property, there is an I-indiscernible
sequence (bi : i ∈ I) locally based on I (in the classical sense). We show that the
sequence (bi : i ∈ I) is locally based on I in our continuous logic sense. Since
first-order formulas generate a dense subalgebra A of the set of all continuous logic
formulas, for each continuous logic formula f (x) and ε > 0 there is φ(x) ∈ A such
that |f (x) − φ(x)| ≤ ε/2. Thus, for any tuples i, j ⊆ I and tuples aj , bi we have
|f (aj ) − f (bi )| ≤ |f (x) − φ(x)| + |φ(aj ) − φ(bi )| + |f (x) − φ(x)| ≤ |φ(aj ) − φ(bi )| + ε.
Finally, note that by the definition of being locally based on (in the classical sense)
for any bi and finite Σ ⊂ A, there is j with the same quantifier free type as i such
that φ(bi ) = φ(aj ) for every φ ∈ Σ. Therefore, the sequence (bi : i ∈ I) is locally
based on I in the continuous sense. □

Next, we define two partial types that will be useful during this section. The
first one is a generalization of the classical Ehrenfeucht-Mostowski type (EM-type
for short). The second is a type whose realizations are exactly the I-indiscernible
sequences. They are based on [Sco12, Definitions 2.6 and 2.10] respectively.
Definition 3.6. Let I = (ai : i ∈ I) be an I-indexed sequence. The EM -type of
I is the set of all conditions φ(xi1 . . . , xin ) = 0 such that φ(aj1 . . . , ajn ) = 0 holds
for every j1 , . . . , jn ∈ I with the same quantifier free type as i1 , . . . , in . That is
EMtp(I)(xi : i ∈ I) = {φ(xi1 , . . . , xin ) = 0 : φ ∈ L, i1 , . . . , in ∈ I
and for any j1 , . . . , jn ∈ I such that
tpqf (j1 , . . . , jn ) = tpqf (i1 , . . . , in ), |= φ(aj1 . . . , ajn ) = 0}.
Definition 3.7. We define Ind(I, L) as the following partial type:
Ind(I, L)(xi : i ∈ I) := {φ(xi1 , . . . , xin ) = φ(xj1 , . . . , xjn ) :
n < ω, i, j ⊆ I, tpqf (i) = tpqf (j), φ(xi1 , . . . , xin ) ∈ L}.
6 ADRIAN PORTILLO FERNÁNDEZ

Finally, we define what it means for a partial type to be finitely satisfiable in a

sequence.
Definition 3.8 (Finitely satisfiable). Let Γ(xi : i ∈ I) be an L-type, and let
I = (ai : i ∈ I) be an I-indexed sequence. We say that Γ is finitely satisfiable in I
if for every finite Γ0 ⊆ Γ+ and for every finite A ⊆ I, there is B ⊆ I, a bijection
f : A → B, and an enumeration i of A such that:
tpqf (i) = tpqf (f [i]) and (af (i) : i ∈ A) |= Γ0 ↾ {xi : i ∈ A}.
Where Γ+ := {φ ≤ 1/n : n < ω; φ = 0 ∈ Γ}.
The following result gives a sufficient condition for the existence of an I-indiscernible
sequence locally based on I = (ai : i ∈ I).
Lemma 3.9. Let J ⊇ I be L′ -structures with the same age and let I = (ai : i ∈ I)
be an I-indexed sequence.
(1) A J -indexed sequence J = (bj : j ∈ J ) is locally based on I if and only if
EMtp(J) ⊇ EMtp(I).
(2) If Ind(I, L) is finitely satisfiable in I, then there is an I-indexed indis-
cernible sequence Ĩ := (bi : i ∈ I) locally based on I.
Proof. (1) Suppose J is locally based on I. Fix φ(xi1 , . . . , xin ) = 0 ∈ EMtpL′ (I)
and let i = (i1 , . . . , in ). If φ(xi ) = 0 is not in EMtp(J), then φ(bj ) ≥ ε
for some ε > 0 and j ⊆ J with the same quantifier free type as i. By
′
assumption, there is i ⊆ I satisfying the same quantifier free type as j and
such that φ(ai′ ) ≥ ε/2, which contradicts φ(xi ) ∈ EMtp(I).
Suppose now that EMtp(J) ⊇ EMtp(I). For a contradiction, assume
that J = (bj : j ∈ J ) is not locally based on I. That is, there is ∆ ⊆
L, bj := (bj1 , . . . , bjn ) from J and ε > 0 such that there is no i ⊆ I
satisfying tpqf (i) = tpqf (j) and |φ(bj ) − φ(ai )| ≤ ε for all φ ∈ ∆. Let
ψ(x) := max{|φ(x) − φ(bj )| : φ ∈ ∆}, ψ is a continuous logic formula
and ψ(bj ) = 0. By assumption, for any i ⊆ I with the same quantifier
free type as j, ψ(ai ) ≥ ε. Thus, ψ(x) ≥ ε ∈ EMtp(I), which contradicts
EMtp(J) ⊇ EMtp(I).
(2) Observe that if the type Ind(I, L)(xi : i ∈ I) is finitely satisfiable in I, then
Ind(I, L) ∪ EMtp(I) is satisfiable. Let J |= Ind(I, L) ∪ EMtp(I). J is an
I-indiscernible sequence and is locally based on I by (1).
□

We now prove the main result of this section. It is an extension of [Sco21,

Theorem 2.10] to continuous logic.
Theorem 3.10. Let L′ be a first-order language and let I be an infinite locally
finite L′ -structure. Then, the following are equivalent:
(1) Age(I) has ERP.
(2) I-indiscernibles have the continuous modeling property.
Proof. (1) =⇒ (2). Assume Age(I) has ERP and let I = (ai )i∈I be any I-indexed
sequence. Our goal is to prove that there exists J = (bi )i∈I locally based on I. By
Lemma 3.9, it is enough to show that Ind(I, L) is finitely satisfiable in I.
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 7

Let Γ0 ⊂ Ind(I, L)+ be any finite subset. For some K, M < ω

1
Γ0 = {|φ(xip ) − φ(xj p )| < : tpqf (ip ) = tpqf (j p ), φ ∈ ∆, p < K, m < M }.
nm

Γ0 involves finitely many formulas ∆ := {φ0 , . . . , φm }, finitely many tuples ip , j p

and finitely many rationals n1m . Without loss of generality, we may assume that the
formulas φ ∈ ∆ (and their tuples of variables) are of the form φ((xg )g∈A ) for some
A ∈ Age(I). Let B ∈ Age(I) be the structure generated by all the coordinates of
the tuples tuples ip , j p involved in Γ0 . It is enough to prove the existence of a copy
B ′ of B such that for any φ((xg )g∈A ) ∈ ∆ and A′ , A′′ ⊆ B copies of A,
1
|φ((ag )g∈A′ ) − φ((ag )g∈A′′ )| ≤
n
for some n < ω such that 1/n is smaller than any rational involved in Γ0 .
If ∆ involves only one formula φ((xg )g∈A ), we proceed in the following manner:
Linearly order the set of intervals {[ ni , i+1
n ] : i < n} and define an n-coloring of the
copies A′ of A by coloring each A′ with the first interval that contains φ((ag )g∈A′ ).
Since Age(I) is Ramsey, we can find a copy B ′ of B homogeneous with respect to
the coloring. Then, (ag )g∈B ′ witnesses that Γ0 is satisfied in I. If ∆ involves k < ω
formulas {φi ((xg )g∈Ai ) : i < k} and all the sets Ai involved are isomorphic we can
apply a similar trick, using as colors the hypercubes
i1 i1 + 1 ik ik + 1
{[ , ] × ··· × [ , ] : i1 , . . . , ik < n}.
n n n n
We claim that the latter is the only case we need to check. The proof is a
standard argument in Ramsey theory which we sketch here for completeness
Let A1 . . . , Am be structures in Age(I) and let B ∈ Age(I) embed every Ai for
i < m. Let k1 , . . . , kn be natural numbers and let Zn ∈ Age(I) be such that

Zn → (Zn−1 )A
kn
n

for every n < m. We construct by induction a sequence of structures Yn ∈ Age(I)

for 0 ≤ n ≤ m.
Case n = 0: Y0 = Zm
Case 0 < n < m: By induction we have Yn−1 ∈ Age(I) isomorphic to Zm−n+1 .
Color the copies of Am−n+1 inside Yn−1 with kn−1 colors. By definition of Zm−n+1 ,
there is a copy Yn of Zm−n inside Yn−1 such that all of the copies of Am−n+1
contained in Yn have the same color.
Note that since Yn ⊆ Yn−1 for 0 ≤ n ≤ m and all copies of Am−n+1 inside
Yn are of the same color, we have that Yn is homogeneous for copies of Aj for all
m − n + 1 ≤ j ≤ m. Therefore, Ym is homogeneous for all copies of A1 , . . . , Am and
so it is the B ′ we were looking for in the proof.
(2) =⇒ (1). Let A ⊆ B ∈ Age(I) be arbitrary finite substructures of I and
let χ be a k-coloring of the embeddings of A into I. We expand I by adding a
predicate Ri for each fiber of the coloring. Let us denote this expanded structure
by I ′ and this new language by L. Let T be the L-theory of of I ′ . Since I has the
modeling property in T , there is an I-indiscernible sequence (bi )i∈I locally based
on (i)i∈I . Using the definition of locally based on for ∆ := {R1 , . . . , Rk } we can
8 ADRIAN PORTILLO FERNÁNDEZ

find and embedding f from B into I such that

tpqf L′ (B) = tpqf L′ (f [B])
and
tp ((bg )g∈B ) = tp∆ ((f (g))g∈B ).
∆

This implies that χ ↾f ◦(B) is constant. □

A

In light of the previous theorem we will not make a distinction between contin-
uous or classical modeling property from now on.

4. Characterizing n-dependence through collapse of indiscernibles

Let T be a complete continuous L-theory with C |= T a monster model (i.e.
κ-saturated and strongly κ-homogeneous for a strong limit cardinal > |T |).
In this section, we study n-dependent continuous formulae and give an anal-
ogous result to [CPT14, Theorem 5.4]. We present an alternative proof of the
aforementioned result, which corrects a mistake made in the original source.
The next few paragraphs contain basic facts about hypergraphs taken almost
verbatim from [CPT14].
We work with three families of languages
Lnop = {<, P0 (x), . . . , Pn−1 (x)},
Lnog = {<, R(x0 , . . . , xn−1 )},
Lnopg = {<, R(x0 , . . . , xn−1 ), P0 (x), . . . , Pn−1 (x)}.
When n < ω is clear we will simply omit it. We consider the Ramsey classes of
ordered n-uniform hypergraphs and ordered n-partite n-uniform hypergraphs.
An Lnog -structure (M, <, R) is an ordered n-uniform hypergraph if
• (M, <) |= DLO
• R(a0 , . . . , an−1 ) implies that a0 , . . . , an−1 are different,
• the relation R is symmetric.
An Lnopg -structure (M, <, R, P0 , . . . , Pn−1 ) is an ordered n-partite n-uniform hy-
pergraph if
• M is the disjoint union P0 ⊔ · · · ⊔ Pn−1 such that if R(a0 , . . . , an−1 ) then
Pi ∩ {a0 , . . . , an−1 } is a singleton for every i < n,
• the relation R is symmetric,
• M is linearly ordered by < and P0 < · · · < Pn−1 .
The following fact was proven in [NR77], [NR83] and independently in [AH78]
for the case of nonpartite hypergraphs and in [CPT14] for the case of partite hy-
pergraphs.
Fact 4.1. Let K be the set of all finite ordered n-partite n-uniform hypergraphs
and K̃ be the set of all finite ordered n-uniform hypergraphs. The classes K and K̃
have the embedding Ramsey property.
We will denote by Gn,p the Fraı̈ssé limit of K and by Gn the Fraı̈ssé limit of K̃.
Remark 4.2. The theories of Gn and Gn,p can be axiomatized in the following
way:
1. (M, <, R) |= Th(Gn ) if and only if
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 9

• (M, <) |= DLO,

• (M, <, R) is an ordered n-uniform hypergraph,
• For every finite disjoint sets A0 , A1 ⊂ M n−1 such that A0 consists of
tuples with pairwise distinct coordinates and b0 < b1 ∈ M , there is
b ∈ M such that b0 < b < b1 and R(b, ai,1 , . . . , ai,n−1 ) holds for ev-
ery (a0,1 , . . . , a0,n−1 ) ∈ A0 and ¬R(b, a1,1 , . . . , a1,n−1 ) holds for every
(a1,1 , . . . , a1,n−1 ) ∈ A1 .
2. (M, <, R, P0 , . . . , Pn ) |= Th(Gn,p ) if and only if
• For every i < n Pi (M ) |= DLO,
• (M, <, R, P0 , . . . , Pn ) is an ordered n-partite n-uniform
Q hypergraph,
• for every j < n, finite disjoint sets A0 , A1 ⊂ i̸=j Pi (M ) and b0 <
b1 ∈ Pj (M ) there is b ∈ Pj (M ) such that b0 < b < b1 and R(b, ai,1 , . . . , ai,n−1 )
holds for every (a0,1 , . . . , a0,n−1 ) ∈ A0 and ¬R(b, a1,1 , . . . , a1,n−1 ) holds
for every (a1,1 , . . . , a1,n−1 ) ∈ A1 .
Next, we define the n-independence property for continuous formulas. An equiv-
alent definition was first formulated in [CT20] using the V Cn dimension.
Definition 4.3 (n-independent formula). We say that a formula f (x, y0 , . . . , yn−1 )
has the n-independence property, IPn for short, if there exist r < s ∈ R and a
sequence (a0,i , . . . , an−1,i )i<ω such that for every finite w ⊆ ω n there exists bw such
that
f (bw , a0,i0 , . . . , an−1,in−1 ) ≤ r ⇐⇒ (i0 , . . . , in−1 ) ∈ w
and
f (bw , a0,i0 , . . . , an−1,in−1 ) ≥ s ⇐⇒ (i0 , . . . , in−1 ) ∈
/ w.
We say that the L-theory T is n-dependent, or N IPn , if no L-formula has IPn .
The following remark is a collection of basic properties of n-dependent formulas.
Remark 4.4. (1) Naming parameters preserves n-dependence. If the L(A)-
formula f (x, y0 , . . . , yn−1 , A) has IPn witnessed by (a0,i , . . . , an−1,i )i<ω ,
then the L-formula g(x, z0 , . . . , zn−1 ) has IPn witnessed by (a0,i A, . . . , an−1,i A)i<ω
where zi = yi w and g(x, z0 , . . . , zn−1 ) = f (x, y0 , . . . , yn−1 , w).
(2) Adding dummy variables preserves n-dependence. Namely, let x ⊂ w and
yi ⊂ zi for all i < n. If g(x, z0 , . . . , zn−1 ) := f (x, y0 , . . . , yn−1 ) has IPn ,
then so does f (x, y0 , . . . , yn−1 ).
(3) Every n-dependent theory is (n + 1)-dependent.
Next, we define what it means for a continuous logic formula to encode a partite
and a nonpartite hypergraph.
Definition 4.5 (Encoding partite hypergraphs). We say that a formula f (x0 , . . . , xn−1 )
encodes an n-partite n-uniform hypergraph (G, R, P0 , . . . , Pn−1 ) if there is a G-
indexed sequence (ag )g∈G and r < s ∈ R satisfying
f (ag0 , . . . , agn−1 ) ≤ r ⇐⇒ R(g0 , . . . , gn−1 )
and
f (ag0 , . . . , agn−1 ) ≥ s ⇐⇒ ¬R(g0 , . . . , gn−1 )
for all g0 , . . . , gn−1 ∈ P0 × · · · × Pn−1 . We say that a formula f (x0 , . . . , xn−1 ) en-
codes n-partite n-uniform hypergraphs if there exist r < s ∈ R such that f (x0 , . . . , xn−1 )
encodes every finite n-partite n-uniform hypergraph using the same r and s.
10 ADRIAN PORTILLO FERNÁNDEZ

Definition 4.6 (Encoding hypergraphs). We say that a formula f (x0 , . . . , xn−1 )

encodes an n-uniform hypergraph (G, R) if there is a G-indexed sequence (ag )g∈G
and r < s ∈ R satisfying
f (ag0 , . . . , agn−1 ) ≤ r ⇐⇒ R(g0 , . . . , gn−1 )
and
f (ag0 , . . . , agn−1 ) ≥ s ⇐⇒ ¬R(g0 , . . . , gn−1 )
for all g0 , . . . , gn−1 ∈ G. We say that a formula f (x0 , . . . , xn−1 ) encodes n-uniform
hypergraphs if there exist r < s ∈ R such that f (x0 , . . . , xn−1 ) encodes every finite
n-uniform hypergraph using the same r and s.
Note that if a formula encodes n-uniform hypergraphs, then it encodes n-partite
n-uniform hypergraphs.
It is no surprise that the n-independence property and encoding (n + 1)-partite
(n + 1)-uniform hypergraphs are also equivalent in our continuous context.
Proposition 4.7. Let f (x, y0 , . . . , yn−1 ) be a formula. Then, the following are
equivalent:
(1) f has IPn .
(2) f encodes (n + 1)-partite (n + 1)-uniform hypergraphs.
(3) f encodes Gn+1,p as a partite hypergraph.
(4) f encodes Gn+1,p as a partite hypergraph witnessed by a Gn+1,p -indiscernible
sequence.
Proof. (1) =⇒ (2). Let r < s ∈ R, (a0,i , . . . , an−1,i )i<ω and (bw )w⊆ωn witness that
f (x, y0 , . . . , yn−1 ) has IPn . Let a finite n+1-uniform n+1-partite hypergraph G be
given. Without loss of generality, we may assume that |P0 (G)| = · · · = |Pn (G)| = k.
For every g ∈ P0 (G) consider the set wg := {(g1 , . . . , gn ) : G |= R(g, g1 , . . . , gn )}.
By identifying Pm (G) with {(m, i) : i < k} and the definition of IPn , we can find
bwg satisfying
f (bwg , ag1 , . . . , agn ) ≤ r ⇐⇒ (g1 , . . . , gn ) ∈ wg
and
f (bwg , ag1 , . . . , agn ) ≥ s ⇐⇒ (g1 , . . . , gn ) ∈
/ wg .
Then, (ag )g∈G witnesses that f encodes G, where ag := bwg for every g ∈ P0 (G).
(2) =⇒ (3). Follows from compactness.
(3) =⇒ (4). Let I = (ag )g∈Gn+1,p witness that f (x, y0 , . . . , yn−1 ) encodes
Gn+1,p as a partite hypergraph. By Theorem 3.10 and Fact 4.1, there exists a
Gn+1,p -indiscernible sequence (bg )g∈Gn+1,p locally based on I. It is easy to see
that (bg )g∈Gn+1,p also witnesses that f (x, y0 , . . . , yn−1 ) encodes Gn+1,p as a partite
hypergraph.
(4) =⇒ (1). Let (ag )g∈Gn+1,p witness that f (x, y0 , . . . , yn−1 ) encodes Gn+1,p as
a partite hypergraph. We write
Gn+1,p = {(j, m) : j ≤ n; m < ω}.
Then, by randomness of Gn+1,p and compactness, for any finite w ⊆ ω n we can
find bw such that
f (bw , a1,i1 , . . . , an,in ) ≤ r ⇐⇒ (i1 , . . . , in ) ∈ w;
f (bw , a1,i1 , . . . , an,in ) ≥ s ⇐⇒ (i1 , . . . , in ) ∈
/ w.
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 11

As in [CPT14, Corollary 5.3], from the fact that any permutation of the parts
of the partition of Gn+1,p is induced by an automorphism of Gn+1.p treated as a
pure hypergraph, we obtain the following as an easy corollary:
Corollary 4.8. Let f (x, y0 , . . . , yn−1 ) be a formula and (w, z0 , . . . , zn−1 ) be any
permutation of (x, y0 , . . . , yn−1 ). Then g(w, z0 , . . . , zn−1 ) := f (x, y0 , . . . , yn−1 ) is
n-dependent if and only if f (x, y0 , . . . , yn−1 ) is n-dependent.
A more involved proof is given in [CT20, Proposition 10.6].
We cannot guarantee that an n-independent formula will encode (n + 1)-uniform
hypergraphs. However, it is true that for continuous theories having IPn and the
existence of a formula encoding (n + 1)-uniform hypergraphs are equivalent. This
generalizes the result in [LS03, Lemma 2.2] and allows is to give an alternative
proof of [CPT14, Theorem 5.4] that avoids the mistake mentioned in the introduc-
tion. In the proof of the next result, we will write f (y0 , . . . , yn−1 , x) instead of
f (x, y0 , . . . , yn−1 ) for convenience.
Proposition 4.9. Let T be a continuous logic theory. The following are equivalent:
(1) T has IPn .
(2) There is a continuous logic formula encoding (n + 1)-uniform hypergraphs.
(3) There is a continuous logic formula encoding Gn+1 as a hypergraph.
(4) There is a continuous logic formula encoding Gn+1 as a hypergraph wit-
nessed by a Gn+1 -indiscernible sequence.
Proof. (1) =⇒ (2). Let f (y0 , . . . , yn−1 , x) be a formula with IPn . We show that
the symmetric formula
ψ(y00 y01 · · · y0n−1 x0 , . . . , yn0 yn1 · · · ynn−1 xn ) = min 0
{f (yσ(0) n−1
, . . . , yσ(n−1) , xσ(n) )}
σ∈Sym(n)

encodes every finite (n + 1)-uniform hypergraph. By Proposition 4.7, f encodes

Gn+1,p as a partite hypergraph, which is witnessed by a Gn+1,p -indiscernible se-
quence (ag )g∈Gn+1,p and some r < s ∈ R. We enumerate the elements of Gn+1,p
as
i
{gm : i ≤ n; m < ω},
where the superscript indicates which part of the partition the element belongs to.
Let an (n + 1)-uniform finite hypergraph H = (H, RH ) be given, we write H :=
{hi : i < k} for some k < ω. We construct H̃ = (H̃, RH̃ ) an isomorphic copy of H
consisting of elements h̃i for i < k of the form (gi0 , . . . , gin−1 , gc(i)
n
) and show that
the formula ψ encodes H̃ (and hence encodes H), where the relation RH̃ and the
function c : ω → ω are to be defined.
We start by defining the function c. For every i < ω, let c(i) be the smallest
m < ω such that for any (j0 , . . . , jn−1 ) ∈ [k]n
RGn+1,p (gj00 , . . . , gjn−1
n−1
n
, gm ) ⇐⇒ j0 < · · · < jn−1 < i ∧ RH (hj0 , . . . , hjn−1 , hi ).
Note that the existence of such m is guaranteed by randomness of Gn+1,p .
The relation RH̃ is defined in the following manner: for i0 < · · · < in < k we set
RH̃ (h̃i0 , . . . , h̃in ) ⇐⇒ RGn+1,p (gi00 , gi11 , . . . , gin−1
n−1
n
, gc(in)
),
12 ADRIAN PORTILLO FERNÁNDEZ

the rest of the cases are defined by symmetry of RH̃ and by declaring ¬RH̃ (h̃i0 , . . . , h̃in )
whenever im1 = im2 for some m1 ̸= m2 . Note that by construction, we have
(H, RH ) ∼= (H̃, RH̃ ).
Claim. The elements bh̃i := (agi0 , . . . , agn−1 , agc(i)
n ) for i < k witness that ψ encodes
i

H̃ with the above r < s.

Proof of claim. To ease the notation, for each σ ∈ Sym(n) we write
0 n−1
fσ := f (yσ(0) , . . . , yσ(n−1) , xσ(n) ).
Our goal is the following:
ψ(bh̃i , . . . bh̃i ) ≤ r ⇐⇒ RH̃ (h̃i0 , . . . , h̃in )
0 n

ψ(bh̃i , . . . bh̃i ) ≥ s ⇐⇒ ¬RH̃ (h̃i0 , . . . , h̃in )

0 n

First, note that if im1 = im2 for some m1 , m2 < n then we have ¬RH̃ (h̃i0 , . . . , h̃in )
by definition of RH̃ and ψ(bh̃i , . . . bh̃i ) ≥ s by construction of the function c.
0 n
Hence, we only need to prove the equivalences above in the case where all the
i’s are pairwise distinct. We prove the first equivalence; the second one is easily
deduced from it.
Assume that RH̃ (h̃i0 , . . . , h̃in ) holds. By symmetry, without loss of generality
we may assume that i0 < · · · < in . Then, by the definition of RH̃ , this implies
that RGn+1,p (gi00 , . . . , gin−1
n−1
n
, gc(in)
) holds. Since the formula f encodes Gn+1,p , this
is equivalent to f (agi0 , . . . , agn−1 , agc(in
)
) ≤ r . Hence, ψ(bh̃i , . . . bh̃i ) ≤ r.
0 in−1 n 0 n

Assume ψ(bh̃i , . . . bh̃i ) ≤ r, again by symmetry, we may assume without loss

0 n
of generality that i0 < · · · < in . This implies that for some σ ∈ Sym(n) we
have fσ ≤ r. However, by construction of the function c, the only possibility is that
fId ≤ r, that is, f (agi0 , . . . , agn−1 , agc(i
n
)
) ≤ r. Since the formula f encodes Gn+1,p ,
0 in−1 n

this is equivalent to RGn+1,p (gi00 , . . . , gin−1

n−1
n
, gc(in)
), which implies RH̃ (h̃i0 , . . . , h̃in )
by definition of the relation RH̃ . □
(2) =⇒ (3). Follows from compactness.
(3) =⇒ (4). Let I = (ag )g∈Gn+1 witness that ψ(x0 , x1 , . . . , xn ) encodes Gn+1 as
a hypergraph. By Theorem 3.10 and Fact 4.1, there exists a Gn+1,p -indiscernible
sequence (bg )g∈Gn+1 locally based on I. It is easy to see that (bg )g∈Gn+1 also
witnesses that ψ(x0 , x1 , . . . , xn ) encodes Gn+1 as a hypergraph.
(4) =⇒ (1). If a formula encodes Gn+1 as a hypergraph, then it also en-
codes Gn+1,p as a partite hypergraph, which implies that the formula has IPn by
Proposition 4.7. □
To prove the main theorem of this section we need the next two facts about
hypergraphs from [CPT14].
Let (G∗ , Lo∗ , Lg∗ ) be either (Gn , {<}, {<, R}) or (Gn,p , Lnop , Lnopg ).
Let V ⊂ G∗ be a finite set and g0 , . . . , gn−1 , g0′ , . . . , gn−1
′
∈ G∗ \ V such that
R(g0 , . . . , gn−1 ) ⇐⇒
̸ R(g0′ , . . . , gn−1
′
).
By W = g0 . . . gn−1 V we mean the set {g0 , . . . , gn−1 } ∪ V with the inherited struc-
ture from G∗ . Let L∗ be Lo∗ or Lg∗ , and let W = g0 . . . gn−1 V , W ′ = g0′ . . . gn−1
′
V,
∼
we write W =L∗ W if the map acting as the identity in V and sending gi to gi′ for
′

i < n is an L∗ isomorphism.
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 13

Definition 4.10. Let V ⊂ G∗ be a finite set and g0 , . . . , gn−1 , g0′ , . . . , gn−1

′
∈ G∗ \V
′ ′ ′
be as above. W = g0 . . . gn−1 V is V -adjacent to W = g0 . . . gn−1 V if
• W ∼=Lo∗ W ′ ,
• for every nonempty v ∈ V with |v| = k and i0 , . . . , in−k−1 < n
R(gi0 , . . . gin−k−1 , v) ⇐⇒ R(gi′0 , . . . gi′n−k−1 , v)
W is said to be adjacent to W ′ if there is V ⊂ W ∩ W ′ such that W is V -adjacent
to W ′ .
Fact 4.11. Let W, W ′ ⊂ G∗ be subsets such that W ∼ =Lo∗ W ′ . Then there is a
sequence W = W0 , W1 , . . . , Wk such that Wi+1 is adjacent to Wi for every i < k
and Wk ∼=Lg∗ W ′
Fact 4.12. Let V ⊂ G∗ be a finite set and g0 < · · · < gn−1 ∈ G∗ \ V with
R(g0 , . . . , gn−1 ). Then there are infinite sets X0 < · · · < Xn−1 ⊆ G∗ such that
• (G′ ; <; R) ∼ = (Gn,p ; <; R) where G′ = X0 . . . Xn−1 (i.e. each Xi correspond
to the part Pi of the partition),
• for any gi′ ∈ Xi (i < n), either W ∼ =Lopg W ′ or W is V -adjacent to W ′ ,
where W = g0 . . . gn−1 V and W ′ = g0′ . . . gn−1
′
V
We are now ready to prove the main theorem of the section.
Theorem 4.13. Let T be a complete continuous logic theory. The following are
equivalent:
(1) T is n-dependent.
(2) Every Gn+1,p -indiscernible is Lop -indiscernible.
(3) Every Gn+1 -indiscernible is order-indiscernible.
Proof. (1) =⇒ (2). Let (ag )g∈Gn+1,p be a Gn+1,p -indiscernible sequence which
is not Lop -indiscernible. Then there are Lop -isomorphic W, W ′ ⊂ Gn+1 subsets
of size m, a formula f (x0 , . . . , xm−1 ) and r < s ∈ R such that f ((ag )g∈W ) ≤
r and f ((ag )g∈W ′ ) ≥ s (where the elements ag are substituted for the variables
x0 , . . . , xm−1 according to the ordering on W and W ′ ). Without loss of generality,
by Fact 4.11, we may assume that W is V -adjacent to W ′ for some subset V such
that W = g0 . . . gn V , W ′ = g0′ . . . gn′ V , R(g0 , . . . gn ) and ¬R(g0′ , . . . gn′ ).
′
Now we apply Fact 4.12 Q to V and g0 , . . . , gn . This yields G ⊆ Gn+1,p such that
for every (h0 , . . . , hn ) ∈ i≤n Pi (G′ )
R(h0 , . . . , hn ) ⇐⇒ h0 . . . hn V ∼
=Lopg W
and
¬R(h0 , . . . , hn ) ⇐⇒ h0 . . . hn V ∼
=Lopg W ′ .
Recall that the sequence (ag )g∈Gn+1,p is Gn+1,p -indiscernible and let f ′ be the
formula defined by permuting the variables (x0 , . . . , xm−1 ) of f in such a way that
the first n + 1-variables are the ones corresponding to h0 , . . . , hn according to the
ordering on the set {h0 , . . . , hn } ∪ V . Then,
f ′ (ah0 , . . . , ahn , A) ≤ r ⇐⇒ R(h0 , . . . , hn )
and
f ′ (ah0 , . . . , ahn , A) ≥ s ⇐⇒ ¬R(h0 , . . . , hn ),
14 ADRIAN PORTILLO FERNÁNDEZ

where A = (ag )g∈V . Since G′ is isomorphic to Gn+1,p , by Proposition 4.7, the for-
mula f ′ (x, y0 , . . . , yn−1 , A) has IPn and hence, by Remark 4.4 there is a continuous
logic formula g(x, z0 , . . . , zn−1 ) which has IPn .
(1) =⇒ (3). The proof is exactly as the proof of (1) =⇒ (2).
(2) =⇒ (1). It follows from Proposition 4.7. If the formula f encodes Gn+1,p
as a partite hypergraph witnessed by a Gn+1,p -indiscernible sequence (ag )g∈Gn+1,p ,
then (ag )g∈Gn+1,p cannot be Lop -indiscernible.
(3) =⇒ (1) Follows from Proposition 4.9. If T has IPn , there is a continuous
logic formula encoding Gn+1 as a hypergraph witnessed by a Gn+1 -indiscernible
sequence(ag )g∈Gn+1 . Then (ag )g∈Gn+1 cannot be order-indiscernible. □
We know explain the error in the proof of [CPT14, Theorem 5.4 (3) =⇒ (2)]
Without loss of generality we write
Gn+1,p = {gqi : i < n; q ∈ Q},
where gqi ∈ Pi (Gn+1,p ) and gqi < gpi for all q < p ∈ Q. We define the ordered
(n + 1)-uniform hypergraph G∗n+1 as follows:
• G∗n+1 = {hq : hq = (gq0 , . . . , gqn ), q ∈ Q},
• RG∗n+1 (hq0 , . . . , hqn ) ⇐⇒ RGn+1,p (gq00 , . . . , gqnn ) for q0 < · · · < qn ,
• hq < hp ⇐⇒ q < p.
Clearly, G∗n+1 embeds every finite ordered (n + 1)-uniform hypergraph.
Let (ag )g∈Gn+1,p be a Gn+1,p -indiscernible sequence which is not Lop -indiscernible.
For hq ∈ G∗n+1 , let bhq = (agq0 , . . . , agqn ) and consider the G∗n+1 -indexed sequence
(bh )h∈G∗n+1 . The following claim is made in the proof:
Claim. Whenever X ≡<G∗ ,RG∗ Y ⊆ G∗n+1 , we have
n+1 n+1

tp((bh )h∈X ) = tp((bh )h∈Y ).

However, this claim is not true as shown by the following counterexample.
Counterexample 4.14. Consider the theory T = Th(Gn+1,p ) and the sequence
(g)g∈Gn+1,p (This sequence corresponds to the sequence (ag )g∈Gn+1,p of the claim,
which is clearly Gn+1,p -indiscernible but not order-indiscernible). Then, for hq =
(gq0 , . . . , gqn ) we have bhq := (gq0 , . . . , gqn ). Let X := {hq0 , hq1 } for some q0 < q1 ∈ Q,
by randomness, there is q0 < q̃ ∈ Q and hq̃ = (gq̃0 , . . . , gq̃n ) such that
RGn+1,p (gq00 , . . . , , gqn−1
0
, gqn1 ) ⇐⇒ ¬RGn+1,p (gq00 , gq10 , . . . , gqn−1
0
, gq̃n ).
Thus, hq0 , hq1 ≡<G∗ ,RG∗ hq0 , hq̃ since:
n+1 n+1

• Th(G∗n+1 ) has quantifier elimination,

• hq0 < hq1 and hq0 < hq̃ ,
• RG∗n+1 has arity n + 1.
We also have tp(bhq0 , bhq1 ) ̸= tp(bhq0 , bhq̃ ) since, by our choice of q̃,
RGn+1,p (gq00 , . . . , , gqn−1
0
, gqn1 ) ⇐⇒ ¬RGn+1,p (gq00 , gq10 , . . . , gqn−1
0
, gq̃n ).
The following is due to Artem Chernikov, Daniel Palacı́n and Kota Takeuchi.
If we substitute the claim above in the original proof by the following: Let A ⊂
Gn+1,p be any finite substructure. Without loss of generality, we may assume
that if gqi , gpj ∈ A and i < j then q < p. Let A∗ = {hq : gqi ∈ A} and let
φ∗ ((x0q , . . . , xn−1
q )hq ∈A∗ ) := φ((xiq )gqi ∈A ) for each formula φ((xiq )gqi ∈A ).
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 15

Claim. Whenever A∗ ≡<G∗ ,RG∗ X ⊆ G∗n+1 , we have

n+1 n+1

tpφ∗ ((bh )h∈A∗ ) = tpφ∗ ((bh )h∈X ).

Then, the proof goes through. Our counterexample shows that one has to work
with a restricted family of formulas φ∗ .
The formula ψ that we constructed in the proof of Proposition 4.9 will be im-
portant throughout the remainder of the chapter.
Notation 1. Given a continuous logic formula f (y0 , . . . , yn−1 , x) we denote the
symmetric formula constructed in the proof of Proposition 4.9 as ψf . Namely:
ψf (y00 y01 · · · y0n−1 x0 , . . . , yn0 yn1 · · · ynn−1 xn ) = min 0
{f (yσ(0) n−1
, . . . , yσ(n−1) , xσ(n) )}.
σ∈Sym(n)

It turns out that IPn of the formulas f and ψf are equivalent. The implication
(2) =⇒ (1) in the next result can also be deduced from [CT20, Proposition
10.4 and Proposition 10.6] since the set of connectives {¬, 21 , −̇} is full (we can
approximate every continuous formula uniformly by formulas constructed using
only that set of connectives). However, we provide a direct proof with ideas that
will be useful in the next section.
Lemma 4.15. Let f (y0 , . . . , yn−1 , x) be a continuous logic formula. Then, the
following are equivalent:
(1) f has IPn .
(2) ψf has IPn .
Proof. (1) =⇒ (2). Follows from the proof of (1) =⇒ (2) of Proposition 4.9 and
Proposition 4.7.
(2) =⇒ (1). First, recall that if a formula is n-dependent and we add dummy
variables, then the new formula is still n-dependent and that n-dependence is pre-
served under permutations of variables (by Remark 4.4 and Corollary 4.8).
We consider the formula
fσ (y00 y01 · · · y0n−1 x0 , . . . , yn0 yn1 · · · ynn−1 xn ) = f (yσ(0)
0 n−1
, . . . , yσ(n−1) , xσ(n) ).
We have ψf = minσ∈Sym(n) {fσ }. Let (bg )g∈Gn+1,p be a witness that ψf encodes
Gn+1,p as a partite hypergraph with some r < s, which exists by Proposition
4.7 and the assumption that ψf has IPn . Fix a linear ordering of Sym(n) and
color the edges of Gn+1,p according to the first σ such that fσ (bg0 , . . . , bgn ) ≤ r
whenever ψf (bg0 , . . . , bgn ) ≤ r. Let H be any finite ordered (n + 1)-partite (n + 1)-
uniform hypergraph. Since Age(Gn+1,p ) has ERP, we can find a monochromatic
isomorphic copy of H inside Gn+1,p . This implies that fσ encodes H as a partite
hypergraph with the same r < s as above (where σ is the color of the edges of
this monochromatic copy of H). Since each finite (n + 1)-partite (n + 1)-uniform
hypergraph is encoded by some fσ , by compactness there is fσ encoding Gn+1,p as
a partite hypergraph. Thus, fσ has IPn which, by the previous paragraph, implies
that f (y0 , . . . , yn−1 , x) has IPn . □

5. n-dependence for hyperdefinable sets

Let T be a complete, first-order theory, and C |= T a monster model (i.e. κ-
saturated and strongly κ-homogeneous for a strong limit cardinal > |T |). Let E be
a ∅-type-definable equivalence relation on a ∅-type-definable subset X of Cλ (or a
product of sorts), where λ < κ.
16 ADRIAN PORTILLO FERNÁNDEZ

We recall the family FX/E defined in [KP22, Section 2]. FX/E is the family of all
functions f : X × Cm → R which factor through X/E × Cm and can be extended to
a continuous logic formula Cλ × Cm → R over ∅ (i.e. factors through a continuous
function Sλ+m (∅)), where m ranges over ω.
Let A ⊂ C (be small). Recall that the complete types over A of elements of X/E
can be defined as the Aut(C/A)-orbits on X/E, or the preimages of these orbits
under the quotient map, or the partial types defining these preimages. The space
of all such types is denoted by SX/E (A).
In this section we apply the results obtained in continuous logic to the context
of hyperdefinable sets to obtain a counterpart to Theorem 4.13.
In [KP22, Proposition 2.1], we showed that the family of functions FX/E sepa-
rates points in SX/E×Cm (∅). Namely,
Fact 5.1. For any a1 = a′1 /E, a2 = a′2 /E in X/E and b1 , b2 ∈ Cm
tp(a1 , b1 ) ̸= tp(a2 , b2 ) ⇐⇒ (∃f ∈ FX/E )(f (a′1 , b1 ) ̸= f (a′2 , b2 ))
This allows us to work with elements of X/E as real elements if we restrict
ourselves to functions from the family FX/E . Hence, we introduce the following
notation:
Notation 2. Let ∆ be a set of (continuous) formulas in variables (xi )i<λ all from
the same product of sorts. We say that a sequence (ai )i∈I of elements from the
appropriate product of sorts is I-indiscernible with respect to ∆ if for any tuples
i1 , . . . , in , j1 , . . . , jn ∈ I we have that
tpqf (i1 , . . . , in ) = tpqf (j1 , . . . , jn ) =⇒ tp∆ (ai1 , . . . , ain ) = tp∆ (aj1 , . . . , ajn ),
where the tuples ai are substituted for the variables of the formulas from ∆.
We define generalised indiscernible sequences of hyperimaginaries exactly as we
did in Definition 3.1.
Definition 5.2. Let I = (ai : i ∈ I) be an I-indexed sequence of hyperimaginaries
(maybe of different sorts), and let A ⊂ C be a small set of parameters. We say that
I is an I-indexed indiscernible sequence over A if for all n ∈ ω and all sequences
i1 , . . . , in , j1 , . . . , jn from I we have that
tpqf (i1 , . . . , in ) = tpqf (j1 , . . . , jn ) =⇒ tp(ai1 , . . . , ain /A) = tp(aj1 , . . . , ajn /A).
By Fact 5.1, a sequence of hyperimaginaries (ai /E)i∈I is I-indiscernible if and
only if the sequence (ai )i∈I is I-indiscernible with respect to the family of functions
f : X n → R that factor through (X/E)n and can be extended to a continuous
formula f : Cnλ → R over ∅ where n ranges over ω.
Next, we define the n-independence property for hyperdefinable sets.
Definition 5.3. A hyperdefinable set X/E has the n-independence property, IPn
for short, if for some m < ω there exist two distinct complete types p, q ∈ SX/E×Cm (∅)
and a sequence (a0,i , . . . , an−1,i )i<ω such that for every finite w ⊂ ω n there exists
bw ∈ X/E satisfying
tp(bw , a0,i0 , . . . , an−1,in−1 ) = p ⇐⇒ (i0 , . . . , in−1 ) ∈ w

tp(bw , a0,i0 , . . . , an−1,in−1 ) = q ⇐⇒ (i0 , . . . , in−1 ) ∈

/ w.
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 17

Note that 1-dependent hyperdefinable sets are exactly the hyperdefinable sets
with N IP (see the definition of hyperdefinable set with NIP in [HP18, Remark
2.3]) by [KP22, Lemma 2.12].
We can easily modify our definition of n-independence to suit functions in FX/E .
Definition 5.4. We say that f (x, y0 , . . . , yn−1 ) ∈ FX/E has the n-independence
property, IPn for short, if there exist r < s ∈ R and a sequence (a0,i , . . . , an−1,i )i<ω
from anywhere such that for every finite w ⊆ ω n there exists bw ∈ X satisfying
f (bw , a0,i0 , . . . , an−1,in−1 ) ≤ r ⇐⇒ (i0 , . . . , in−1 ) ∈ w
and
f (bw , a0,i0 , . . . , an−1,in−1 ) ≥ s ⇐⇒ (i0 , . . . , in−1 ) ∈
/ w.
Similarly, we can define what it means for a function in FX/E to encode (n-
partite) n-uniform hypergraphs.
Definition 5.5. We say that f (x, y1 , . . . , yn−1 ) ∈ FX/E encodes a n-partite n-
uniform hypergraph (G, R, P0 , . . . , Pn−1 ) if there are a G-indexed sequence (ag )g∈G
with ag ∈ X for every g ∈ P0 (G) and r < s ∈ R satisfying
f (ag0 , . . . , agn−1 ) ≤ r ⇐⇒ R(g0 , . . . , gn−1 )
and
f (ag0 , . . . , agn−1 ) ≥ s ⇐⇒ ¬R(g0 , . . . , gn−1 )
for all g0 , . . . , gn−1 ∈ P0 ×· · ·×Pn−1 . We say that f (x0 , . . . , xn−1 ) ∈ FX/E encodes
n-partite n-uniform hypergraphs if there exist r < s ∈ R such that f (x, y1 , . . . , yn−1 )
encodes every finite n-partite n-uniform hypergraph using the same r and s.
The proof of the following fact is exactly as in the case of a general continuous
formula.
Proposition 5.6. Let f (x, y0 , . . . , yn−1 ) ∈ FX/E . Then, the following are equiva-
lent:
(1) f has IPn .
(2) f encodes (n + 1)-partite (n + 1)-uniform hypergraphs.
(3) f encodes Gn+1,p as a partite hypergraph.
(4) f encodes Gn+1,p as a partite hypergraph witnessed by a Gn+1,p -indiscernible
sequence.
The following lemma allows us to understand IPn of a hyperdefinable set X/E
through the family of functions FX/E .
Lemma 5.7. X/E has IPn if and only if some f (x, y0 , . . . , yn−1 ) ∈ FX/E has IPn .
Proof. Assume that X/E has IPn . Take witnesses p and q from Definition 5.3.
Then, by Fact 5.1, there exists f (x, y0 , . . . , yn−1 ) ∈ FX/E and r < s such that
f (x, y0 , . . . , yn−1 ) ≤ r ∈ p and f (x, y0 , . . . , yn−1 ) ≥ s ∈ q. The elements witnessing
IPn for X/E also witness that f (x, y0 , . . . , yn−1 ) has IPn .
Assume now that some f (x, y0 , . . . , yn−1 ) ∈ FX/E has IPn . By Proposition 5.6,
the function f encodes Gn+1,p as a partite hypergraph witnessed by a Gn+1,p -
indiscernible sequence (ag )g∈Gn+1,p and some r < s ∈ R. We write
Gn+1,p = {(j, m) : j ≤ n; m < ω}.
18 ADRIAN PORTILLO FERNÁNDEZ

Then, by randomness of Gn+1,p , for any finite disjoint s0 , s1 ⊆ ω n we can find

bs0 ,s1 ∈ X (in fact, we can choose it to be some ag for g ∈ P0 (Gn+1,p )) such that
(i1 , . . . , in ) ∈ s0 =⇒ f (bs0 ,s1 , a1,i1 , . . . , an,in ) ≤ r,
(i1 , . . . , in ) ∈ s1 =⇒ f (bs0 ,s1 , a1,i1 , . . . , an,in ) ≥ s.
Moreover, by Gn+1,p -indiscernibility, there exist two distinct complete types p, q ∈
SX/E×Cm (∅) such that
(i1 , . . . , in ) ∈ s0 =⇒ tp(bs0 ,s1 /E, a1,i1 , . . . , an,in ) = p;
(i1 , . . . , in ) ∈ s1 =⇒ tp(bs0 ,s1 /E, a1,i1 , . . . , an,in ) = q.
By compactness, it follows that X/E has IPn . □
′
Notation 3. For each f (x, y0 , . . . , yn ) ∈ FX/E , let f (y0 , . . . , yn , x) := f (x, y0 , . . . , yn ).
We denote by Ψn+1
FX/E the set containing all functions ψf ′ for f (x, y0 , . . . , yn ) ∈
FX/E , where ψf ′ is constructed as in Notation 1. ΨFX/E is the union of all Ψn+1
FX/E
for n < ω.
The next definition is the natural counterpart of Definition 4.6 for functions of
the family ΨFX/E .
Definition 5.8. We say that ψ(x0 , . . . , xn−1 ) ∈ ΨFX/E encodes an n-uniform hy-
pergraph (G, R) if there is a G-indexed sequence (ag )g∈G in Cm × X (for some fixed
m < ω) and r < s ∈ R satisfying
ψ(ag0 , . . . , agn−1 ) ≤ r ⇐⇒ R(g0 , . . . , gn−1 )
and
ψ(ag0 , . . . , agn−1 ) ≥ s ⇐⇒ ¬R(g0 , . . . , gn−1 )
for all g0 , . . . , gn−1 ∈ G. We say that ψ(x0 , . . . , xn−1 ) encodes n-uniform hyper-
graphs if there exist r < s ∈ R such that ψ(x0 , . . . , xn−1 ) encodes every finite
n-uniform hypergraph using the same r and s.
As in the general continuous case, we have an equivalence between IPn of the
hyperdefinable set X/E and the existence of some function coding (n + 1)-uniform
hypergraphs.
Proposition 5.9. The following are equivalent:
(1) X/E has IPn .
(2) There is a function in ΨFX/E encoding (n + 1)-uniform hypergraphs.
(3) There is a function in ΨFX/E encoding Gn+1 as a hypergraph.
(4) There is a function in ΨFX/E encoding Gn+1 as a hypergraph witnessed by
a Gn+1 -indiscernible sequence.
Proof. (1) =⇒ (2). By Lemma 5.7 and Proposition 5.9, there exists a function
f (x, y0 , . . . , yn−1 ) ∈ FX/E encoding Gn+1,p as a partite hypergraph. Consider the
function f ′ (y0 , . . . , yn−1 , x) := f (x, y0 , . . . , yn−1 ), following the proof of Proposition
4.9, we see that ψf ′ ∈ ΨFX/E encodes Gn+1 .
(2) =⇒ (3) follows by compactness.
(3) =⇒ (4). Follows from the proof of (3) =⇒ (4) of Proposition 4.9.
(4) =⇒ (1) We slightly modify the proof of Lemma 4.15. Let (bg )g∈Gn+1 be
a witness that ψf encodes Gn+1 as a hypergraph. Note that for every g ∈ Gn+1 ,
 n-DEPENDENT CONTINUOUS THEORIES AND HYPERDEFINABLE SETS 19

bg = (a0g , . . . , ang ) with ang ∈ X. Fix a linear ordering of Sym(n) and color the
edges of Gn+1 according to the first σ such that f (a0gσ(0) , . . . , angσ(n) ) ≤ r when-
ever ψf (bg0 , . . . , bgn ) ≤ r. Let H = {him : i ≤ n, m ≤ k} be any finite ordered
(n + 1)-partite (n + 1)-uniform hypergraph. Since Age(Gn+1 ) has ERP, we can
find a monochromatic isomorphic copy of H inside Gn+1 (as a non partite hyper-
graph). This implies that the function f (y0 , . . . , yn−1 , x) encodes H as a partite
hypergraph, witnessed by the elements {ai σ(i) : i ≤ n, m ≤ k}, where σ is the
hm
color of the monochromatic copy of H. By compactness, f ′ (x, y0 , . . . , yn−1 ) :=
f (y0 , . . . , yn−1 , x) encodes Gn+1,p as a partite hypergraph. Therefore, since the
function f ′ (x, y0 , . . . , yn−1 ) is in FX/E , by Proposition 5.6, X/E has IPn . □
We finish the section with a characterization of n-dependent hyperdefinable sets
analogous to the one in Theorem 4.13. Recall the definition of the family Ψn+1FX/E
from Notation 3.
Theorem 5.10. The following are equivalent:
(1) X/E is n-dependent.
(2) Every Gn+1,p -indiscernible (ag )g∈Gn+1,p where for every g ∈ P0 (Gn+1,p ) we
have ag ∈ X/E is Lop -indiscernible.
(3) For every m ∈ N, every Gn+1 -indiscernible with respect to Ψn+1
FX/E sequence
of elements of Cm × X is order-indiscernible with respect to Ψn+1
FX/E .

Proof. (1) =⇒ (2). Let (ag )g∈Gn+1,p be a Gn+1,p -indiscernible sequence where
for every g ∈ P0 (Gn+1,p ) we have ag ∈ X/E which is not Lop -indiscernible and
let (a′g )g∈Gn+1,p be a sequence of representatives. Then, by Lemma 5.1, there are
Lop -isomorphic W, W ′ ⊂ Gn+1 subsets of size m, a function f (x0 , . . . , xm−1 ) and
r < s ∈ R such that f ((a′g )g∈W ) ≤ r and f ((a′g )g∈W ′ ) ≥ s (where the elements a′g
are substituted for the variables x0 , . . . , xm−1 according to the ordering on W and
W ′ ). Without loss of generality, by Fact 4.11, we may assume that W is V -adjacent
to W ′ for some subset V such that W = g0 . . . gn V , W ′ = g0′ . . . gn′ V , R(g0 , . . . gn )
and ¬R(g0′ , . . . gn′ ). Following as in the proof of (1) =⇒ (2) of Theorem 4.13,
we arrive at the conclusion that f ′ (x, y0 , . . . , yn , A) has IPn , where A = (a′g )g∈V .
The tuple A is contained in some product (with repetition) of X and C. Hence,
the function g obtained at the end of the proof of this implication in Theorem 4.13
possibly has several infinite tuples of variables each of which corresponds to X.
Thus, when performing the change of variables done in Remark 4.4 (1) we might
end with infinite tuples of variables. However, since this new function g(x, z1 . . . , zn )
is the uniform limit of functions from the family FX/E , we might find a suitable
formula g ′ ∈ FX/E witnessing IPn .
(1) =⇒ (3). Let (ag )g∈Gn+1 be a Gn+1 -indiscernible with respect to Ψn+1 FX/E
sequence which is not order-indiscernible with respect to Ψn+1 FX/E . Then there are
′
W, W ⊂ Gn+1 subsets of size n + 1, a function ψf (x0 , . . . , xn ) and r < s ∈ R such
that ψf ((ag )g∈W ) ≤ r and ψf ((ag )g∈W ′ ) ≥ s. By Fact 4.11 and the fact that Gn+1
is self-complementary, we may assume W = g0 . . . gn , W ′ = g0′ . . . gn′ , R(g0 , . . . gn )
and ¬R(g0′ , . . . gn′ ).
By Gn+1 -indiscernibility of (ag )g∈Gn+1 with respect to Ψn+1
FX/E and by symmetry
of the relation R and ψf , this implies that
ψf (ah0 , . . . , ahn ) ≤ r ⇐⇒ R(h0 , . . . , hn )
20 ADRIAN PORTILLO FERNÁNDEZ

and
ψf (ah0 , . . . , ahn ) ≥ s ⇐⇒ ¬R(h0 , . . . , hn ).
By Proposition 5.9, the set X/E has IPn .
(2) =⇒ (1) It follows from Proposition 5.6. If the function f ∈ FX/E
encodes Gn+1,p witnessed by a Gn+1,p -indiscernible sequence (ag )g∈Gn+1,p , then
(ag )g∈Gn+1,p cannot be Lop -indiscernible.
(3) =⇒ (1) It follows from Proposition 5.9. If T has IPn , there is a function
ψf ∈ Ψn+1
FX/E encoding Gn+1 witnessed by a Gn+1 -indiscernible sequence (ag )g∈Gn+1 .
Then, (ag )g∈Gn+1 cannot be order-indiscernible respect to Ψn+1
FX/E . □
Note that the results of this section easily generalise to deal with n-dependence
of imaginary sorts in continuous logic. We finish the chapter with an example
illustrating that, in general, the theorem above is optimal. Namely, for an n-
dependent hyperdefinable set X/E and n′ > n there might be a Gn+1 -indiscernible
sequence of elements of Cm × X for some m < ω which are not order-indiscernible
′
with respect to ΨnFX/E+1
or with respect to more general families of functions from
F(Cm ×X/E)n+1 .
Example 5.11. Let N be a monster model of a NIP theory and R a monster model
of the theory of random ordered graphs. We consider the structure N ⊔ R i.e. the
structure with disjoint sorts for N and R with no interaction between the sorts. Let
X = N and E be the equality relation. Clearly, X/E has NIP.
m
z }| {
Claim. Let n ≥ 1. The sequence (ag )g∈G2 := (g, . . . , g, n0 )g∈G2 is G2 -indiscernible
but it is not order-indiscernible with respect to the family of formulas f (x0 , . . . , xn )
where each xi is a tuple (x0i , . . . , xm
i ) of variables of length m + 1 whose last coor-
dinate corresponds to X.
Proof of the first claim. Let g0′ < g0 < g1 < · · · < gn ∈ G2 be such that R(g0 , g1 )
and ¬R(g0′ , g1 ) and let f (x0 , . . . , xn ) := R(x00 , x01 ). Clearly, the tuples (g0 , g1 , . . . , gn )
and (g0′ , g1 , . . . , gn ) have the same order type but f (ag0 , . . . , agn ) ̸= f (ag0′ , . . . , agn ).
□
n′
′
z }| {
Claim. For n > 1, the sequence (ag )g∈G2 := (g, . . . , g, n0 )g∈G2 is G2 -indiscernible
′
but it is not order-indiscernible with respect to the family of functions ΨnFX/E+1
.

Proof of the second claim. We show it for n′ = 2. Let f (y, z, x) := R(y, z). Then
the formula
ψf (y0 z0 x0 , y1 z1 x1 , y2 z2 y2 ) := min f (yσ(0) , zσ(1) , xσ(2) )
σ∈Sym(2)

belongs to Ψ3FX/E . However, taking g0′ < g0 < g1 < g2 such that
• R(g0 , g1 ), R(g0 , g2 ), R(g1 , g2 )
• R(g0′ , g2 ) and ¬R(g0′ , g1 )
we have that the tuples (g0 , g1 , g2 ) and (g0′ , g1 , g2 ) are order-isomorphic and
ψf (g0 g0 n0 , g1 g1 n0 , g2 g2 n0 ) ̸= ψf (g0′ g0′ n0 , g1 g1 n0 , g2 g2 n0 ).
□
 REFERENCES 21

References
[AH78] Fred G. Abramson and Leo A. Harrington. “Models Without Indis-
cernibles”. In: The Journal of Symbolic Logic 43.3 (1978), pp. 572–600.
issn: 00224812.
[And23] Aaron Anderson. Fuzzy VC Combinatorics and Distality in Continuous
Logic. 2023. arXiv: 2310.04393 [math.LO].
[BY09] Itaı̈ Ben Yaacov. “Continuous and random Vapnik-Chervonenkis classes”.
In: Israel Journal of Mathematics 179.1 (2009), pp. 309–333.
[BY10] Itaı̈ Ben Yaacov. “Stability and stable groups in continuous logic”. In:
J. Symbolic Logic 75.3 (2010), pp. 1111–1136. doi: 10 . 2178 / jsl /
1278682220.
[BYU10] Itaı̈ Ben Yaacov and Alexander Usvyatsov. “Continuous first order logic
and local stability”. In: Trans. Amer. Math. Soc. 362.10 (2010), pp. 5213–
5259. doi: 10.1090/S0002-9947-10-04837-3.
[CPT14] Artem Chernikov, Daniel Palacı́n, and Kota Takeuchi. “On n-Dependence”.
In: Notre Dame J. Formal Log. 60 (2014), pp. 195–214.
[CT20] Artem Chernikov and Henry Towsner. “Hypergraph regularity and higher
arity VC-dimension”. In: ArXiv abs/2010.00726 (2020).
[DK21] Jan Cz. Dobrowolski and Mark Kamsma. “Kim-independence in positive
logic”. In: Model Theory (2021). url: https://api.semanticscholar.
org/CorpusID:234741810.
[GH19] Vincent Guingona and Cameron Donnay Hill. “On positive local com-
binatorial dividing-lines in model theory”. In: Archive for Mathematical
Logic 58.3-4 (June 2019), pp. 289–323. doi: 10 . 1007 / s00153 - 018 -
0635-2.
[HP18] Mike Haskel and Anand Pillay. “On maximal stable quotients of defin-
able groups in NIP theories”. In: The Journal of Symbolic Logic 83.1
(2018), pp. 117–122. issn: 00224812, 19435886. url: https : / / www .
jstor.org/stable/26600311 (visited on 02/29/2024).
[Kam23] Mark Kamsma. Positive indiscernibles. 2023. arXiv: 2305.14127 [math.LO].
[KP22] Krzysztof Krupiński and Adrián Portillo. “On Stable Quotients”. In:
Notre Dame Journal of Formal Logic 63.3 (2022), pp. 373 –394. doi:
10.1215/00294527-2022-0023.
[LS03] M.C. Laskowski and S. Shelah. “Karp complexity and classes with the
independence property”. In: Annals of Pure and Applied Logic 120.1
(2003), pp. 263–283. issn: 0168-0072. doi: https://doi.org/10.1016/
S0168-0072(02)00080-5.
[NR77] Jaroslav Nešetřil and Vojtěch Rödl. “Partitions of finite relational and
set systems”. In: Journal of Combinatorial Theory, Series A 22.3 (1977),
pp. 289–312. issn: 0097-3165. doi: https://doi.org/10.1016/0097-
3165(77)90004-8. url: https://www.sciencedirect.com/science/
article/pii/0097316577900048.
[NR83] Jaroslav Nešetřil and Vojtěch Rödl. “Ramsey classes of set systems”.
In: Journal of Combinatorial Theory, Series A 34.2 (1983), pp. 183–201.
issn: 0097-3165. doi: https : / / doi . org / 10 . 1016 / 0097 - 3165(83 )
90055-9. url: https://www.sciencedirect.com/science/article/
pii/0097316583900559.
22 REFERENCES

[Sco12] Lynn Scow. “Characterization of NIP theories by ordered graph-indiscernibles”.

In: Annals of Pure and Applied Logic 163.11 (2012), pp. 1624–1641. doi:
https://doi.org/10.1016/j.apal.2011.12.013.
[Sco15] Lynn Scow. “Indiscernibles, EM-Types, and Ramsey Classes of Trees”.
In: Notre Dame Journal of Formal Logic 56.3 (2015), pp. 429 –447. doi:
10.1215/00294527-3132797.
[Sco21] Lynn Scow. “Ramsey transfer to semi-retractions”. In: Annals of Pure
and Applied Logic 172.3 (2021), p. 102891. issn: 0168-0072. doi: https:
//doi.org/10.1016/j.apal.2020.102891.
[She05] Saharon Shelah. “Strongly dependent theories”. In: Israel Journal of
Mathematics 204 (2005), pp. 1–83.
[She07] Saharon Shelah. “Definable groups for dependent and 2-dependent the-
ories”. In: arXiv: Logic (2007).
[She82] S. Shelah. “Classification Theory and the Number of Nonisomorphic
Models”. In: Journal of Symbolic Logic 47.3 (1982), pp. 694–696. doi:
10.2307/2273599.
[Zie88] M. Ziegler. Stabilitätstheorie. 1988. url: https://home.mathematik.
uni-freiburg.de/ziegler/skripte/stabilit.pdf.
Adrián Portillo https://orcid.org/0000-0001-9354-8574
Email address: Adrian.Portillo-Fernandez@math.uni.wroc.pl

Instytut Matematyczny Uniwersytetu Wroclawskiego, pl. Grunwaldzki 2, 50-384

Wroclaw, Poland