This course develops the central ideas of modern model theory through the lens of types and stability. It begins with types as consistent descriptions of possible elements, then studies the Stone spaces they form and the semantic notions of realization, saturation, and homogeneity. From there, it moves into omitting types and the structure of atomic, prime, and countable models, building a bridge between syntactic consistency and the classification of models up to isomorphism. The second half of the course shifts to stability theory, where the focus is on counting types and understanding when theories behave in a controlled, classifiable way. Definability of types and forking appear as key tools for measuring dependence, while strongly minimal sets provide the first major prototype of a tame geometry inside model theory. These ideas culminate in Morley’s Categoricity Theorem and then broaden into Shelah’s classification program, which aims to organize first-order theories by their structural complexity. Each chapter prepares the ground for the next: types lead to topological and semantic viewpoints, those yield model-existence and uniqueness results, and those in turn motivate stability as a dividing line between tame and wild theories. By the end, the course gives a coherent path from basic type theory to the major classification theorems that shape contemporary model theory. # Introduction This course begins where a first course in model theory usually leaves off. Compactness and completeness tell us that finitely consistent syntactic data can be realized in some model; the next question is how to organize all such data at once. The organizing objects are types, and the central theme of the course is that spaces of types convert logical consistency into a usable geometry. The course has two intertwined goals. The first is technical: to build the machinery of complete types, Stone spaces, saturated models, omitting types, and indiscernible sequences. The second is structural: to understand why some complete theories have models that are tightly controlled by their types, while others encode enough combinatorial complexity to resist classification. ## From Compactness to Types A first course proves compactness as a theorem about theories: if every finite subset of a set of sentences has a model, then the whole set has a model. In practice, we often want to prescribe the behaviour of a tuple inside a fixed ambient theory rather than build an entire theory from scratch. This shifts attention from complete theories to consistent descriptions of possible elements. [definition: Partial Type] Let $T$ be a first-order theory in a language $L$, let $M \models T$, let $A \subset M$, and let $x = (x_1, \dots, x_n)$ be a tuple of variables. A partial $n$-type over $A$ with respect to $M$ is a set $p(x)$ of $L(A)$-formulas in the variables $x$ such that there are an elementary extension $N \succeq M$ and a tuple $b \in N^n$ with $N \models \varphi(b)$ for every $\varphi(x) \in p(x)$. [/definition] The point of the definition is that $p(x)$ records a possible local description relative to the intended meanings of the parameters from $A$, not merely relative to the abstract theory $T$. Those parameters must be kept tied to their values in the ambient model; otherwise the same symbols could be reinterpreted in a way that changes the formulas being tested. To use such descriptions, we need a practical test for when infinitely many formula requirements are jointly possible. Compactness supplies exactly that test by reducing the question to finite fragments. [quotetheorem:5051] [citeproof:5051] This theorem explains why types are the natural continuation of compactness. The finite satisfiability hypothesis is essential: if a finite subset already contradicts $T$, compactness has no way to produce a realizing tuple. The role of parameters is also not cosmetic, since formulas over $A$ describe positions relative to named elements and must be interpreted in a compatible ambient structure. Instead of applying compactness separately to many finite diagrams, we package a desired infinite behaviour as a single object and then ask whether it is realized, omitted, isolated, or forced by a formula. [example: Non-Principal Type At Infinity] Let $M=\mathbb N$ be the usual model of arithmetic, and consider $p(x)=\{n<x:n\in\mathbb N\}$ over the named standard natural numbers. If \begin{align*} F=\{n_1<x,\dots,n_k<x\}\subseteq p(x), \end{align*} let $m=\max\{n_1,\dots,n_k\}$ and interpret the new constant $c$ as $m+1$. Then for each $i\le k$, \begin{align*} n_i\le m < m+1=c, \end{align*} so $M\models n_i<c$. Thus every finite part of the elementary diagram of $M$ together with finitely many formulas from $p(c)$ is satisfiable in $M$ itself. By *[Realization of Partial Types by Compactness](/theorems/5051)*, there is an elementary extension $N\succeq M$ and an element $b\in N$ such that \begin{align*} N\models n<b \end{align*} for every standard $n\in\mathbb N$. No element of the original standard model realizes $p(x)$, since if $a\in\mathbb N$, then the formula $a<x$ belongs to $p(x)$ but $M\not\models a<a$. The type is therefore a limiting description: every finite initial demand can be met inside $\mathbb N$, while the whole collection is realized only after passing to a nonstandard elementary extension. [/example] The example also signals a recurring distinction. Some types are forced by a single formula and behave like definable pieces of the structure; others exist only as limiting descriptions and are invisible to any finite observation. ## Spaces of Possible Elements Once types are treated as objects, the next problem is to compare them. Two types may agree on many formulas and differ only on a later decision; a family of types may have finite compatibility properties; a definable set may determine a region of the type space. These questions are topological in nature. [definition: Stone Space Of Types] Let $T$ be a complete first-order theory, let $M \models T$, and let $A \subset M$ be a parameter set. The Stone space $S_n(A)$ is the set of complete $n$-types over $A$ realized in some elementary extension of $M$. For each $L(A)$-formula $\varphi(x)$, define \begin{align*} [\varphi] = \{p \in S_n(A) : \varphi(x) \in p\}. \end{align*} The topology on $S_n(A)$ has the sets $[\varphi]$ as a basis. [/definition] The notation turns formulas into open sets, and negation turns the same basic open sets into closed sets. The next issue is whether this topology has enough compactness and separation to support arguments by finite approximation, since Chapter 2 will cover type spaces by definable conditions and then need finite subcovers. [quotetheorem:5052] [citeproof:5052] This is the first place where the course systematically turns syntax into geometry. Completeness of $T$ matters because complete types must decide every formula relative to a fixed theory; without a fixed complete background, the same syntactic data could split across incompatible completions. The Hausdorff and totally disconnected conclusions depend on complete types rather than partial types, since two complete types can be separated by a formula on which they disagree. A formula is no longer only a statement about tuples; it is also a clopen subset of a [compact space](/page/Compact%20Space). Model-theoretic arguments will often move between these two readings, especially when a logical problem becomes a finite-subcover argument in a type space. [example: Dense Linear Orders] Work in a model $M$ of dense linear orders without endpoints. By *[quantifier elimination for dense linear orders](/theorems/4305)*, a complete $1$-type over $M$ is determined by deciding, for each $a \in M$, exactly one of \begin{align*} x<a,\qquad x=a,\qquad a<x. \end{align*} If the type contains $x=a$ for some $a \in M$, then completeness and linearity force it to be the realized type of $a$: for every $b \in M$, it contains $b<x$ when $b<a$, contains $x=b$ when $b=a$, and contains $x<b$ when $a<b$. If the type is not realized in $M$, define \begin{align*} L_p&=\{a\in M:a<x\in p\},\\ U_p&=\{a\in M:x<a\in p\}. \end{align*} For $a\in L_p$ and $b\in U_p$, the type contains $a<x$ and $x<b$, so it contains $a<b$ by transitivity of the order. Thus $L_p<U_p$. Also, if $a\in L_p$ and $c<a$, then $c<x$ belongs to $p$, since $c<a$ and $a<x$ imply $c<x$; hence $L_p$ is downward closed. Similarly, if $b\in U_p$ and $b<c$, then $x<c$ belongs to $p$, so $U_p$ is upward closed. Therefore a non-realized type determines a cut $(L_p,U_p)$ in $M$, with every element of $L_p$ lying below every element of $U_p$. Conversely, any cut $(L,U)$ with $L<U$, $L$ downward closed, $U$ upward closed, and $M=L\cup U$, determines the partial description \begin{align*} p_{L,U}(x)=\{a<x:a\in L\}\cup\{x<b:b\in U\}. \end{align*} Every finite subset mentions only finitely many lower bounds $a_1,\dots,a_m\in L$ and upper bounds $b_1,\dots,b_n\in U$. Since $a_i<b_j$ for all $i,j$, density gives an element between the largest relevant lower bound and the smallest relevant upper bound whenever both sides are present; if only one side is present, the no-endpoints axiom supplies an element above the listed lower bounds or below the listed upper bounds. Hence the finite fragment is satisfiable, so the description is a type. The two remaining possibilities are the directions \begin{align*} \{a<x:a\in M\} \quad\text{and}\quad \{x<a:a\in M\}, \end{align*} which say that $x$ lies to the right of all of $M$ or to the left of all of $M$. Thus $S_1(M)$ contains the realized points of $M$, the cuts of $M$, and the two endpoint directions. The Stone space is an order-completion-like compactification, but its basic open sets are formulas such as $a<x$, $x<b$, and $a<x<b$, not intervals introduced externally. [/example] ## Saturation And Homogeneity Type spaces describe all locally consistent behaviours, but a particular model may realize only some of them. The next problem is to construct and recognize models that realize enough types to serve as universal domains for the theory. This leads to saturation and homogeneity. [definition: Saturated Model] Let $T$ be a complete theory and let $\kappa$ be an infinite cardinal. A model $M \models T$ is $\kappa$-saturated if for every parameter set $A \subset M$ with $|A| < \kappa$ and every finite tuple of variables $x = (x_1, \dots, x_n)$, every complete $n$-type over $A$ that is realized in some elementary extension of $M$ is realized in $M$. [/definition] Saturation says that the model contains all small behaviours compatible with its own parameter configuration. The reference to elementary extensions is important: a type over $A \subset M$ may mention formulas such as $a < x < b$, so its consistency depends on how $a$ and $b$ sit inside $M$, not only on the axioms of $T$. For this notion to be useful, we need an existence theorem showing that saturated models are not rare accidents but can be built at controlled cardinalities under standard size hypotheses. [quotetheorem:5053] [citeproof:5053] [example: Algebraically Closed Fields] Let $K$ be a $\kappa$-saturated algebraically closed field, let $A\subset K$ have $|A|<\kappa$, and let $F$ be the subfield of $K$ generated by $A$. If $f(X)\in F[X]$ is irreducible and nonconstant, the algebraic condition \begin{align*} f(x)=0 \end{align*} is consistent over $A$, because every algebraically closed elementary extension of $K$ contains a root of $f$. Since $K$ is $\kappa$-saturated and $|F|\le |A|+\aleph_0<\kappa$, this type is realized by some $b\in K$ with \begin{align*} f(b)=0. \end{align*} The transcendental type over $F$ is the set of requirements \begin{align*} p(x)=\{g(x)\ne 0:g(X)\in F[X]\setminus\{0\}\}. \end{align*} For any finite subset $\{g_1(x)\ne 0,\dots,g_m(x)\ne 0\}$, form \begin{align*} h(X)=g_1(X)\cdots g_m(X). \end{align*} Because $F[X]$ is an integral domain and each $g_i$ is nonzero, $h(X)\ne 0$. A nonzero polynomial has only finitely many roots in a field, while an algebraically closed field is infinite, so there is some element $c$ in an elementary extension with \begin{align*} h(c)\ne 0. \end{align*} Then \begin{align*} g_1(c)\cdots g_m(c)\ne 0, \end{align*} so each factor satisfies $g_i(c)\ne 0$. Thus every finite part of $p(x)$ is satisfiable, and saturation gives $t\in K$ such that \begin{align*} g(t)\ne 0 \end{align*} for every nonzero $g(X)\in F[X]$. This means exactly that $t$ is transcendental over $F$. Hence a saturated algebraically closed field realizes both algebraic types, given by roots of irreducible polynomials, and the transcendental type over each small parameter field. [/example] ## Omitting Types Realizing types builds large models, but classification also depends on controlling which behaviours can be avoided. A non-principal type may describe an unwanted limit object, and omitting it can produce a model with a prescribed gap. This gives a second compactness-adjacent tool with a different flavour. [definition: Principal Type] Let $T$ be a complete theory, let $M \models T$, let $A \subset M$, and let $p(x) \in S_n(A)$. The type $p(x)$ is principal, or isolated, if there is an $L(A)$-formula $\varphi(x) \in p(x)$ such that for every $L(A)$-formula $\psi(x)$, if $\psi(x) \in p(x)$ then every elementary extension of $M$ satisfies \begin{align*} \forall x\,(\varphi(x) \to \psi(x)). \end{align*} [/definition] Principal types are controlled by one formula, so any model containing a realization of that isolating formula realizes the type. The natural complementary question is whether non-principal types can be avoided altogether, especially in countable theories where a Henkin construction can meet countably many requirements. [quotetheorem:5054] [citeproof:5054] This theorem shows that non-principal types are not automatically present in every model. Countability is essential to the proof strategy because the Henkin construction must schedule both witness requirements and avoidance requirements in a single sequence of stages. It is also essential to the conclusion: in a language with constants $(c_\alpha)_{\alpha < \omega_1}$, any complete theory containing $c_\alpha \neq c_\beta$ for all $\alpha \neq \beta$ has no countable model, even before any type-omission requirements are added. The hypothesis cannot be replaced by an arbitrary large family: in any countable complete theory with no atomic model, the family of all non-principal complete types over $\varnothing$ cannot be omitted all at once, since a model omitting all of them would be atomic. Such theories exist, so some size restriction on the family is necessary. Non-principality is also essential. If $p$ is isolated by $\varphi$ and $T \models \exists x\,\varphi(x)$, then every model of $T$ contains a tuple satisfying $\varphi$, and that tuple realizes $p$. Thus an isolated type forced to be inhabited by the theory cannot be omitted. The contrast between saturated models, which realize many types, and omitting types constructions, which avoid selected types, is one of the practical tensions of the subject. [example: Omitting The Type At Infinity] Let \begin{align*} p(x)=\{n<x:n\in\mathbb N\} \end{align*} over the named standard natural numbers in the usual ordered model $\mathbb N$. No element of $\mathbb N$ realizes $p(x)$: if $a\in\mathbb N$, then the formula $a<x$ belongs to $p(x)$, but substituting $a$ for $x$ gives $a<a$, and the standard order satisfies \begin{align*} \mathbb N \not\models a<a. \end{align*} Thus every candidate $a\in\mathbb N$ fails at least the requirement indexed by itself. The type is nevertheless finitely satisfiable. If \begin{align*} F=\{n_1<x,\dots,n_k<x\}\subseteq p(x), \end{align*} choose \begin{align*} m=\max\{n_1,\dots,n_k\}. \end{align*} For each $i\le k$, the definition of maximum gives $n_i\le m$, and the successor property in $\mathbb N$ gives $m<m+1$. Hence, by transitivity of $<$, \begin{align*} n_i\le m<m+1, \end{align*} so in particular \begin{align*} \mathbb N\models n_i<m+1. \end{align*} Interpreting the new variable $x$ as $m+1$ satisfies every formula in $F$. By the compactness criterion for types, there is an elementary extension $N\succeq\mathbb N$ and an element $b\in N$ such that \begin{align*} N\models n<b \end{align*} for every standard $n\in\mathbb N$. The same set of formulas is therefore consistent and realizable in a nonstandard elementary extension, but omitted in the original standard model; this is the distinction between consistency of a type and realization in a particular model. [/example] ## Stability As A Classification Problem The machinery of types becomes classification theory when we ask how many types a theory has over models of a given size. If the number of types grows in a controlled way, models can often be analyzed by dimension-like invariants. If the number of types grows too quickly, the theory can encode enough order-like complexity to resist classification. [definition: Stable Theory] Let $T$ be a complete theory and let $\kappa$ be an infinite cardinal. The theory $T$ is stable in $\kappa$ if for every model $M \models T$ with $|M| = \kappa$, the space $S_1(M)$ has cardinality at most $\kappa$. [/definition] Stability is a counting condition, but counting alone does not explain what kind of structure is being excluded. To turn stability into a usable diagnostic, we need a syntactic pattern whose presence creates many types and whose absence predicts controlled behaviour. [definition: Order Property] A formula $\varphi(x,y)$ has the order property with respect to a complete theory $T$ if for every $n \in \mathbb N$ there are a model $M \models T$ and tuples $a_1, \dots, a_n$ and $b_1, \dots, b_n$ in $M$ such that \begin{align*} M \models \varphi(a_i,b_j) \quad \text{if and only if} \quad i < j. \end{align*} [/definition] The order property is the syntactic pattern that produces many incompatible choices. The key theorem is that this pattern is not merely an example of instability: it exactly characterizes instability, connecting cardinal bounds on type spaces to definable finite orders. [quotetheorem:5055] [citeproof:5055] This equivalence motivates the later chapters on stable theories. The distinction between stability in a fixed cardinal and stability across a range of cardinals is important: before the theorem, the definition only gives a cardinal-by-cardinal counting condition, and the cardinal arithmetic clause in the theorem is part of the statement. The theorem does not say that every stable theory is stable in every sufficiently large cardinal without qualification. For example, a theory may be stable but fail to be stable in cardinals below its stability threshold or in cardinals where the relevant exponentiation produces too many parameter-definable choices. The theorem says that the basic obstruction to stability is not a numerical accident at one size but the presence of a formula capable of coding finite linear orders. Rather than treating stability as only a bound on $|S_1(M)|$, we study it as the absence of a particular kind of definable order, which is what later permits ranks and independence notions to behave like dimension. [example: Orders Versus Algebraically Closed Fields] In dense linear orders, the formula $\varphi(x,y)$ given by $x<y$ already has the order property. For each $n\ge 1$, work in $(\mathbb Q,<)$ and choose \begin{align*} a_i&=i,\\ b_j&=j-\frac{1}{2} \end{align*} for $1\le i,j\le n$. Then \begin{align*} \mathbb Q\models a_i<b_j &\Longleftrightarrow i<j-\frac{1}{2}\\ &\Longleftrightarrow i<j, \end{align*} because $i$ and $j$ are integers. Thus $x<y$ codes arbitrary finite linear orders, so by *Stability And The Order Property*, dense linear orders are unstable. Algebraically closed fields behave differently. Let $M$ be an algebraically closed field. By *[quantifier elimination for algebraically closed fields](/theorems/4310)*, a complete $1$-type over $M$ is determined by which polynomial equations $f(x)=0$ with $f(X)\in M[X]$ it contains. Since $M$ is algebraically closed, every nonconstant polynomial in $M[X]$ factors as \begin{align*} f(X)=c(X-a_1)\cdots (X-a_r) \end{align*} with $c\in M^\times$ and $a_1,\dots,a_r\in M$. Therefore \begin{align*} f(x)=0 \end{align*} is equivalent over $M$ to \begin{align*} x=a_1\ \vee\ \cdots\ \vee\ x=a_r. \end{align*} So a non-transcendental $1$-type over $M$ is the realized type $x=a$ for some $a\in M$, while the remaining possibility is the transcendental type \begin{align*} \{f(x)\ne 0:f(X)\in M[X]\setminus\{0\}\}. \end{align*} Thus $S_1(M)$ has exactly the realized points indexed by $M$ together with one generic transcendental point, so $|S_1(M)|=|M|$ for infinite $M$. Dense orders represent the order-coding side of the course, while algebraically closed fields represent the geometry-governed stable side. [/example] ## Morley Rank And Categoricity After stability has been identified, the course moves toward the strongest early classification theorem: Morley's theorem. The problem is to understand when a theory has exactly one model of a given uncountable cardinality and why that condition should force uniqueness in all uncountable cardinalities. [definition: Categoricity In A Cardinal] Let $T$ be a complete theory and let $\kappa$ be an infinite cardinal. The theory $T$ is categorical in $\kappa$ if any two models of $T$ of cardinality $\kappa$ are isomorphic. [/definition] The next question is whether categoricity in one uncountable size is an isolated coincidence or evidence of a deeper structural constraint on the theory. Morley's theorem is needed because it turns a single-cardinal uniqueness hypothesis into a spectrum-wide classification statement, and it gives the course its main destination. [quotetheorem:1492] Morley's theorem is not just a statement about moving from one cardinal to another. It says that uncountable categoricity forces a hidden geometry: after the stability analysis, models are controlled by dimensions of independent sets rather than by arbitrary syntactic presentations. The countability hypothesis keeps the relevant type spaces and prime-model constructions small enough for this analysis, and the restriction to uncountable cardinals avoids the finite-or-countable dimension phenomena that can occur in familiar examples. The algebraically closed field example below is the guiding case: once the characteristic is fixed, transcendence degree is the dimension invariant, and in uncountable cardinalities that dimension is determined by the size of the field. [example: Algebraically Closed Fields Of Fixed Characteristic] [claim]For each fixed characteristic, algebraically closed fields are categorical in every uncountable cardinal.[/claim] [proof]Fix a characteristic and let $P$ be the corresponding prime field, so $P=\mathbb Q$ in characteristic $0$ and $P=\mathbb F_p$ in characteristic $p>0$. In either case, \begin{align*} |P|\le \aleph_0. \end{align*} Let $K$ be an uncountable algebraically closed field of this characteristic, and let $B$ be a transcendence basis of $K$ over $P$. Then $K$ is algebraic over $P(B)$ by the definition of transcendence basis. First compute the size of $P(B)$. Every element of $P(B)$ is a quotient $f(b_1,\dots,b_n)/g(b_1,\dots,b_n)$ with $f,g\in P[X_1,\dots,X_n]$, $g\ne 0$, and $b_1,\dots,b_n\in B$. For an infinite set $B$, \begin{align*} |P[B]| &=\left|\bigcup_{n<\omega} P[X_1,\dots,X_n]\times B^n\right|\\ &\le \sum_{n<\omega} \max(\aleph_0,|B|)^n\\ &=\max(\aleph_0,|B|). \end{align*} Since $B\subseteq P[B]$, we also have $|P[B]|\ge |B|$, and since $P\subseteq P[B]$, we have $|P[B]|\ge |P|$. Hence \begin{align*} |P[B]|=\max(\aleph_0,|B|). \end{align*} The field of fractions $P(B)$ consists of pairs of polynomials modulo the equality relation $f/g=f'/g'$, so \begin{align*} |P(B)|\le |P[B]\times P[B]|=|P[B]| \end{align*} and $P[B]\subseteq P(B)$, giving \begin{align*} |P(B)|=\max(\aleph_0,|B|). \end{align*} Because $K$ is algebraic over $P(B)$, each element of $K$ is a root of some nonzero polynomial in $P(B)[X]$. There are \begin{align*} |P(B)[X]|=\max(\aleph_0,|B|) \end{align*} such polynomials, and each nonzero polynomial has finitely many roots. Therefore \begin{align*} |K| &\le |P(B)[X]|\cdot \aleph_0\\ &=\max(\aleph_0,|B|). \end{align*} Since $B\subseteq K$, we also have $|K|\ge |B|$. If $|B|\le \aleph_0$, then the displayed inequality gives $|K|\le\aleph_0$, contradicting that $K$ is uncountable. Thus $|B|$ is uncountable, and consequently \begin{align*} |K|=|B|. \end{align*} Now let $K$ and $L$ be uncountable algebraically closed fields of the same fixed characteristic and the same cardinality $\kappa$. Choose transcendence bases $B_K$ and $B_L$. The computation above gives \begin{align*} |B_K|=|K|=\kappa=|L|=|B_L|. \end{align*} Choose a bijection $B_K\to B_L$. Since both fields have the same prime field, this bijection extends to an isomorphism \begin{align*} P(B_K)\cong P(B_L). \end{align*} By the standard [extension theorem](/theorems/59) for isomorphisms between algebraic closures, this extends further to an isomorphism \begin{align*} K\cong L. \end{align*} Thus any two uncountable algebraically closed fields of the fixed characteristic and cardinality $\kappa$ are isomorphic.[/proof] So the uncountable model is classified by one invariant, its transcendence degree; in the uncountable case that invariant is exactly the field cardinality, making algebraically closed fields the guiding example behind Morley's categoricity theorem. [/example] ## How The Course Fits Together The course proceeds in a sequence of increasingly structural questions. First we ask what a possible element can look like, leading to partial and complete types. Next we topologize the set of all possibilities, producing Stone spaces. Then we compare models by how many types they realize or omit. The second half of the course asks when this machinery classifies models. Stability begins as a bound on the number of types, becomes a prohibition on definable order, and then supports ranks and independence notions. Morley's theorem is the first major payoff: a uniqueness assumption in one uncountable size propagates through the entire uncountable spectrum. A useful way to read these notes is to keep three translations in mind. A formula is both a syntactic object and a clopen subset of a type space. A model is both a structure and a collection of realized types. A stable theory is both a theory with few types and a theory whose definable sets avoid order-like complexity. The opening chapter framed the course around three parallel translations: syntax, topology, and model-theoretic structure. We now take the first step in that direction by treating types as the basic language for recording consistent local information about tuples inside a theory. # 1. Types as Consistent Descriptions Types are the language in which model theory records the possible local behaviour of elements. In a first course, compactness says that every finitely satisfiable theory has a model; in this chapter we apply the same principle to descriptions of tuples inside a fixed ambient theory. The guiding question is how much information about an element can be specified syntactically before a model is forced either to realize it or to omit it. The chapter begins with partial and complete types over a parameter set, then treats compactness as the existence theorem for finitely satisfiable types. We finish by separating complete, quantifier-free, and $n$-types, with examples from dense linear orders, algebraically closed fields, and arithmetic. ## Descriptions Over Parameters When studying a structure $M$, we often want to describe an element not by naming it, but by listing every formula it should satisfy relative to parameters from some set $A \subseteq M$. The first problem is to make precise what it means for such a list to be internally consistent. [definition: Formula Over A Parameter Set] Let $L$ be a first-order language, let $M$ be an $L$-structure, and let $A \subseteq M$. The language $L(A)$ is obtained from $L$ by adding a constant symbol $c_a$ for each $a \in A$. A formula over $A$ is an $L(A)$-formula. [/definition] The constants from $A$ let formulas speak about positions relative to named parameters. Once parameters are allowed, a formula such as $x < a$ in an ordered structure records a relation between the unknown element $x$ and the fixed element $a$. A raw list of formulas is too permissive: the list may contain $x<a$ and $a<x$, or may contradict the theory after the parameters have been named. We therefore need a consistency notion for sets of such formulas, because a proposed description may contain many conditions without yet deciding every possible question about $x$. [definition: Partial Type] Let $T$ be an $L$-theory, let $M$ be an $L$-structure, let $A \subseteq M$, and let $x=(x_1,\dots,x_n)$ be a tuple of variables. A partial $n$-type over $A$ with respect to $T$ is a set $p(x)$ of $L(A)$-formulas in the free variables among $x$ such that $T \cup \operatorname{Diag}_M(A) \cup p(x)$ is satisfiable. [/definition] Here $\operatorname{Diag}_M(A)$ denotes the atomic diagram of the named parameter set as it sits inside $M$: after adding constants $c_a$ for $a\in A$, it records every atomic and negated atomic $L(A)$-sentence true of those constants in $M$. When the ambient structure is fixed, we often write this as $\operatorname{Diag}(A)$. This matters when the language has function symbols, because terms in elements of $A$ must be interpreted with their values inherited from $M$ rather than treated as unrelated constants. In practice, when $A \subseteq M \models T$ and we work over the elementary diagram of $M$ with parameters named, this definition says that the formulas in $p(x)$ can be realized together in some elementary extension of $M$ after naming $A$. [example: Partial Type Of Being Above A Set] Let $(M,<)$ be a model of dense linear orders without endpoints, let $A\subseteq M$, and consider \begin{align*} p(x)=\{a<x:a\in A\}. \end{align*} A finite subset of $p(x)$ has the form \begin{align*} \{a_1<x,\dots,a_k<x\} \end{align*} for some $a_1,\dots,a_k\in A$. This finite set is realized by an element $b$ exactly when \begin{align*} a_1<b,\quad a_2<b,\quad \dots,\quad a_k<b, \end{align*} which is exactly to say that $b$ is an upper bound for the finite set $\{a_1,\dots,a_k\}$. Thus $p(x)$ is finitely satisfiable precisely when every finite subset of $A$ has an upper bound in a model of the named-parameter theory. If $A$ is finite, say $A=\{a_1,\dots,a_k\}$, then because $<$ is a linear order, one of these elements is maximal; call it $a_i$. Since $M$ has no right endpoint, there is $b\in M$ such that $a_i<b$. For each $j$, maximality gives $a_j\leq a_i$, and transitivity gives \begin{align*} a_j\leq a_i<b, \end{align*} so $a_j<b$. Hence $b$ realizes every formula in $p(x)$ inside $M$. If $A$ is not bounded in $M$ but is bounded in a larger elementary extension, then a realizing element is not an element of $M$; it is a new point whose order position lies strictly to the right of every named parameter from $A$. [/example] Partial types may leave many questions undecided, and this example does not tell us whether $x$ lies below another parameter, equals a parameter, or satisfies unrelated formulas. Without a maximality condition, two genuinely different possible elements may satisfy the same partial description, so the description cannot serve as a complete local state. To represent a full possible behaviour of a tuple over $A$, we require the description to make a choice for every formula. The next definition gives the maximal consistent objects that will form the Stone spaces studied in Chapter 2. [definition: Complete Type] Let $T$ be an $L$-theory and let $A$ be a parameter set. A complete $n$-type over $A$ with respect to $T$ is a partial $n$-type $p(x)$ over $A$ such that for every $L(A)$-formula $\varphi(x)$, exactly one of $\varphi(x)$ and $\neg \varphi(x)$ belongs to $p(x)$. [/definition] A complete type is a maximal consistent description of a tuple over the chosen parameters. It need not name a tuple in the original model; it may describe a tuple that appears only after passing to an elementary extension. [definition: Realized And Omitted Type] Let $M \models T$, let $A \subseteq M$, and let $p(x)$ be a partial $n$-type over $A$. A tuple $b \in M^n$ realizes $p$ in $M$ if $M \models \varphi(b)$ for every $\varphi(x) \in p(x)$. The model $M$ omits $p$ if no tuple in $M^n$ realizes $p$. [/definition] Realization is the bridge from syntax back to the model. A type can be consistent with the theory but still omitted by a particular model, and this distinction becomes central when constructing saturated or omitting models later in the course. [example: A Type Omitted By The Standard Natural Numbers] Work in the language of arithmetic with the usual order, and consider the standard model $\mathbb N$. Let \begin{align*} p(x)=\{n<x:n\in\mathbb N\}. \end{align*} To check finite satisfiability in $\mathbb N$, take a finite subset \begin{align*} \{n_1<x,\dots,n_k<x\}\subseteq p(x). \end{align*} Let $m=\max\{n_1,\dots,n_k\}$. Then $m+1\in\mathbb N$, and for each $i$ we have \begin{align*} n_i\leq m<m+1, \end{align*} so $n_i<m+1$. Hence $m+1$ realizes this finite subset in $\mathbb N$. No element of $\mathbb N$ realizes all of $p(x)$. If $c\in\mathbb N$, then the formula $c<x$ belongs to $p(x)$, so realizing $p$ would require \begin{align*} \mathbb N\models c<c, \end{align*} which is false because $<$ is irreflexive. Thus $p(x)$ is finitely satisfiable but omitted by the standard model. In any elementary extension containing an element $b$ with $n<b$ for every standard $n\in\mathbb N$, that element realizes exactly this type: each formula in $p(x)$ is one of the inequalities $n<x$, and substituting $b$ gives $n<b$. [/example] This example explains why types are not merely a notational convenience. They capture potential elements whose existence is invisible inside a given model but forced in suitable elementary extensions. ## Principal And Non-Principal Types The next question is when an infinite description is really controlled by a single formula. Such types are easier to understand because realizing one formula automatically realizes the whole type. [definition: Principal Type] Let $p(x)$ be a complete $n$-type over $A$ with respect to $T$. The type $p$ is principal, or isolated, if there is an $L(A)$-formula $\theta(x)$ such that $\theta(x) \in p(x)$ and \begin{align*} T \cup \operatorname{Diag}(A) \models \forall x\, (\theta(x) \to \varphi(x)) \end{align*} for every $\varphi(x) \in p(x)$. The formula $\theta(x)$ is called an isolating formula for $p$. [/definition] A principal type is a complete description with a finite syntactic handle: a single formula determines the rest of the description. This is stronger than merely containing a satisfiable formula, since a formula may define a large set containing points of many different complete types. We need a separate name for complete types without such a handle, since these are the types whose realization cannot be forced by realizing one definable set. The next definition names this opposing class, which will be the focus of compactness and omitting-types arguments. [definition: Non-Principal Type] A complete type is non-principal if it is not principal. [/definition] The definition of principality quantifies over every formula in the type, which is often inconvenient to check directly. A more usable test asks whether a single formula decides every formula, placing the type on one side or the other of each possible definable partition. The following criterion is the form in which isolation is used throughout the course. [quotetheorem:5056] [citeproof:5056] The criterion turns the topological idea of an isolated point into syntax: the basic condition $\theta$ contains no other complete type. The membership hypothesis $\theta \in p$ is essential. A satisfiable formula that decides every formula isolates some complete type, but without requiring agreement with $p$ it may isolate the wrong one; for example, in a theory with two named constants $a\ne b$, the formula $x=a$ decides the complete type of $a$, not the complete type of $b$. Completeness of $T$ prevents the ambient theory from changing underneath the test: in the incomplete theory of fields, the quantifier-free formula $x=x$ does not decide the sentence $\operatorname{char}(K)=0$ when this sentence is viewed as a formula with no free variables, so it cannot isolate a complete type uniformly across all fields. The diagram of $A$ is also essential, since formulas over parameters can only be compared after those parameters have been interpreted; if two constants named by $A$ are identified in one interpretation and distinct in another, formulas such as $x=c_a$ and $x=c_b$ no longer determine the same descriptions. This will later reappear as the statement that principal types are isolated points of the Stone space. [example: Principal Types In Dense Linear Orders] Let $T$ be the theory of dense linear orders without endpoints. Over $\varnothing$, dense linear orders without endpoints have quantifier elimination, so every formula $\varphi(x)$ in one free variable is equivalent modulo $T$ to a quantifier-free formula in the variable $x$ alone. The only atomic formulas in one variable over $\varnothing$ are $x=x$ and $x<x$; in every linear order, \begin{align*} T &\models \forall x\, x=x,\\ T &\models \forall x\, \neg(x<x). \end{align*} Thus every quantifier-free formula in one variable is equivalent over $T$ either to $x=x$ or to $x\ne x$. Hence any two elements in any models of $T$ satisfy the same formulas over $\varnothing$, so there is a unique complete $1$-type over $\varnothing$. The formula $x=x$ isolates it, because for every formula $\varphi(x)$ in that type, \begin{align*} T \models \forall x\, (x=x\to \varphi(x)). \end{align*} Now let $M\models T$ and let $a\in M$. The realized type $\operatorname{tp}(a/M)$ is isolated by $x=a$: if $\varphi(x)\in \operatorname{tp}(a/M)$, then $M\models \varphi(a)$, and equality substitution gives \begin{align*} T\cup \operatorname{Diag}(M)\models \forall x\, (x=a\to \varphi(x)). \end{align*} By contrast, let $M=L\cup R$ be a proper cut not filled in $M$, with every element of $L$ below every element of $R$, with $L$ having no maximum and $R$ having no minimum. Consider the cut type \begin{align*} p(x)=\{m<x:m\in L\}\cup\{x<m:m\in R\}. \end{align*} This type is not principal over $M$. Indeed, suppose $\theta(x)\in p(x)$. By quantifier elimination, $\theta(x)$ is equivalent over $T$ to a quantifier-free formula using only finitely many parameters from $M$; call this finite parameter set $F$. Since $L$ has no maximum, choose $l\in L$ above every element of $F\cap L$. Since $R$ lies above $L$, every element of $F\cap R$ is above $l$. By density, choose $c\in M$ with \begin{align*} \max(F\cap L)<c<l \end{align*} when $F\cap L$ is nonempty, and choose any $c<l$ still below all elements of $F\cap R$ when $F\cap L$ is empty. Then $c$ has the same order relation to every parameter in $F$ as a realization of the cut has, so $c\models \theta(x)$. But $c<l$, while $l<x$ belongs to $p(x)$. Therefore $\theta(x)$ does not imply all formulas in $p(x)$, so no formula isolates the cut type. Thus parameters make realized elements principal through equations such as $x=a$, while unfilled cuts remain non-principal because every single formula can mention only finitely many parameters of the cut. [/example] The contrast between realized elements and cuts is one of the main reasons that types are always indexed by a parameter set. Adding parameters refines the available formulas and can turn the description of a named element into an isolated one, while also creating new non-realized cut types over the same model. ## Compactness As Type Existence The [compactness theorem](/theorems/2748) is usually presented as a theorem about theories. For types, the same theorem says that a description is realizable somewhere exactly when all finite parts are realizable somewhere. [definition: Finitely Satisfiable Type Scheme] Let $T$ be an $L$-theory, let $A$ be a parameter set, and let $\Sigma(x)$ be a set of $L(A)$-formulas. The set $\Sigma(x)$ is finitely satisfiable with respect to $T$ if every finite subset $\Sigma_0(x) \subseteq \Sigma(x)$ is realized in some model of $T \cup \operatorname{Diag}(A)$. [/definition] Finite satisfiability is the exact hypothesis compactness can see. It does not require a single model to realize all formulas at once; it only asks that no finite obstruction exists. Without this hypothesis, a contradiction can already appear in a finite fragment, such as the pair of formulas $x=a$ and $x\ne a$, and no model could realize the whole scheme. [quotetheorem:5057] [citeproof:5057] This theorem is the practical engine behind type construction. Instead of building an element directly, we verify all finite fragments and allow compactness to supply a model where the full description is realized. The finite satisfiability hypothesis cannot be weakened to informal plausibility of the infinite list: if $\Sigma(x)$ contains $x\ne x$, or contains both $\varphi(x)$ and $\neg\varphi(x)$, a one-formula or two-formula fragment already blocks realization. The expanded language and the diagram of $A$ are part of the hypothesis, not bookkeeping. For instance, in the language of rings, naming a parameter $a$ from a field $K$ without recording its polynomial relations could allow a model to reinterpret $a$ as an element with a different [minimal polynomial](/page/Minimal%20Polynomial), so the resulting realization would not be a type over the intended parameter set. The conclusion is also external to the original model: compactness supplies some model of the expanded theory, not a guarantee that the starting model realizes the type. [example: Compactness Produces Infinite Elements] Let $T=\operatorname{Th}(\mathbb N,+,\cdot,<,0,1)$, and consider the scheme \begin{align*} \Sigma(x)=\{n<x:n\in\mathbb N\}. \end{align*} To verify finite satisfiability, take a finite subcollection \begin{align*} \Sigma_0(x)=\{n_1<x,\dots,n_k<x\}. \end{align*} Let $m=\max\{n_1,\dots,n_k\}$. Then $m+1\in\mathbb N$, and for each $i=1,\dots,k$, \begin{align*} n_i\leq m,\qquad m<m+1, \end{align*} so transitivity of $<$ gives \begin{align*} n_i<m+1. \end{align*} Thus $m+1$ realizes every formula in $\Sigma_0(x)$ in the standard model, so $\Sigma(x)$ is finitely satisfiable. By *[Type Existence Theorem By Compactness](/theorems/5057)*, there is a model $M^*\models T$ and an element $b\in M^*$ such that \begin{align*} M^*\models n<b \end{align*} for every standard $n\in\mathbb N$. This element cannot lie in the original $\mathbb N$: if $b=c\in\mathbb N$, then the formula $c<x$ is one of the formulas in $\Sigma(x)$, so realization would require \begin{align*} \mathbb N\models c<c, \end{align*} contradicting irreflexivity of $<$. Compactness therefore produces a new element in an elementary extension whose order position is above every standard natural number. [/example] Compactness also explains why omitted types are delicate. A type omitted in one model can become realized in an extension, so omission is not a property of consistency alone. ## Complete Types, Quantifier-Free Types, And $n$-Types Once types are available, we need to track how much syntax they record. The main distinction is between complete types, which remember all formulas, and quantifier-free types, which remember only the atomic shape of a tuple. [definition: Type In Finitely Many Variables] Let $T$ be an $L$-theory and let $A$ be a parameter set. An $n$-type over $A$ is a type in variables $x=(x_1,\dots,x_n)$ over $A$. [/definition] The integer $n$ records the arity of the tuple being described. A $1$-type describes a possible element; a $2$-type can also record relations between two possible elements. Full types can be difficult to compute because quantified formulas may express hidden existence conditions, such as the existence of a point between two elements or a root of a polynomial. A useful workflow is therefore: first compute the quantifier-free data of the tuple over the parameters, then check whether the ambient theory has quantifier elimination, and finally decide whether the resulting complete type is isolated by a finite part of that data or remains non-principal. We now need a weaker version of type that keeps only quantifier-free information, because many examples start by computing the atomic diagram of a tuple before asking whether quantifiers add new information. [definition: Quantifier-Free Type] Let $M$ be an $L$-structure, let $A \subseteq M$, and let $b \in M^n$. The quantifier-free type of $b$ over $A$, denoted $\operatorname{qftp}(b/A)$, is the set of all quantifier-free $L(A)$-formulas $\varphi(x)$ such that $M \models \varphi(b)$. [/definition] The quantifier-free type records the immediate algebraic or relational diagram of the tuple over $A$. In general it may lose information hidden behind quantifiers, such as the existence of roots or points in intervals. The next theorem identifies the important case where this loss does not occur, allowing complete types to be computed from quantifier-free data. [quotetheorem:5058] [citeproof:5058] This theorem justifies many computations of types in familiar theories. Rather than analyzing arbitrary formulas, we reduce to order relations, polynomial equations, or other quantifier-free data. Quantifier elimination is the decisive hypothesis: without it, two tuples can have the same atomic diagram but differ on an existential formula, such as whether a polynomial equation has a solution over the generated substructure. Completeness also keeps the ambient theories synchronized; otherwise the same quantifier-free information may be interpreted inside models satisfying different completions of the theory. The common-parameter hypothesis prevents a separate failure mode: if a named constant $a$ denotes $0$ in one structure and $1$ in another, then agreement on formulas written using the symbol $c_a$ no longer compares types over the same parameter. The theorem therefore does not say that quantifier-free types always determine complete types, only that they do so in a fixed complete theory where quantifiers can be eliminated and the parameters have been identified. [example: Empty-Set Types In Dense Linear Orders] Let $T$ be the theory of dense linear orders without endpoints. For a pair $(a,b)$ over $\varnothing$, the only atomic order comparisons between the variables are \begin{align*} x=y,\qquad x<y,\qquad y<x, \end{align*} together with tautological or contradictory formulas such as $x=x$ and $x<x$. In every linear order, exactly one of the three alternatives \begin{align*} a=b,\qquad a<b,\qquad b<a \end{align*} holds. Since *[quantifier elimination for dense linear orders without endpoints](/theorems/4273)* reduces every formula over $\varnothing$ to a quantifier-free formula, the complete $2$-type of $(a,b)$ is determined by which one of these three alternatives is true. For a single variable over $\varnothing$, the only atomic formulas are $x=x$ and $x<x$. Every element satisfies \begin{align*} x=x \end{align*} and no element satisfies \begin{align*} x<x, \end{align*} so all elements have the same quantifier-free type over $\varnothing$, hence the same complete $1$-type by quantifier elimination. Over a model $M$, the quantifier-free information of a possible element $x$ records, for each $m\in M$, exactly one of \begin{align*} x<m,\qquad x=m,\qquad m<x. \end{align*} If $x=a$ for some $a\in M$, this gives the realized type of $a$. Otherwise the formulas split $M$ into \begin{align*} L=\{m\in M:m<x\},\qquad R=\{m\in M:x<m\}, \end{align*} with every element of $L$ below every element of $R$. Thus a non-realized $1$-type over $M$ is determined by a cut in $M$, while realized types are the special cases isolated by equations $x=a$. [/example] Dense order examples show how types encode order-theoretic location. Algebraically closed fields show the parallel algebraic phenomenon: a type records algebraic equations and inequations over the parameter field. [example: Types In Algebraically Closed Fields] Let $T=\operatorname{ACF}_p$, and let $K\models T$. Since $K$ is algebraically closed, every nonzero polynomial $f(x)\in K[x]$ factors as \begin{align*} f(x)=\lambda (x-c_1)\cdots (x-c_r) \end{align*} with $\lambda\in K^\times$ and $c_1,\dots,c_r\in K$. Thus if an element $a$ in an elementary extension satisfies $f(a)=0$, then \begin{align*} 0=f(a)=\lambda(a-c_1)\cdots(a-c_r). \end{align*} Fields have no zero divisors and $\lambda\ne 0$, so $a-c_i=0$ for some $i$, hence $a=c_i\in K$. Therefore every algebraic $1$-type over the model $K$ is realized in $K$, and the type of $c\in K$ is isolated by the formula $x=c$: if $\varphi(x)\in \operatorname{tp}(c/K)$, then $K\models \varphi(c)$, so equality substitution gives \begin{align*} T\cup \operatorname{Diag}(K)\models \forall x\, (x=c\to \varphi(x)). \end{align*} Now let $t$ be transcendental over $K$ in some elementary extension. Its type over $K$ contains \begin{align*} g(x)\ne 0 \end{align*} for every nonzero $g(x)\in K[x]$, because $g(t)=0$ would make $t$ algebraic over $K$. This transcendental type is not isolated. Suppose $\theta(x)$ belonged to this type and isolated it. By quantifier elimination for algebraically closed fields, $\theta(x)$ is equivalent over $T\cup\operatorname{Diag}(K)$ to a Boolean combination of polynomial equations over $K$. Since $t\models\theta(x)$ and no nonzero polynomial over $K$ vanishes at $t$, one satisfied quantifier-free piece of $\theta$ has the form \begin{align*} h_1(x)\ne 0,\quad \dots,\quad h_m(x)\ne 0 \end{align*} with each $h_j\in K[x]$ nonzero. Each polynomial $h_j$ has only finitely many roots in $K$, and $K$ is infinite, so choose $c\in K$ such that \begin{align*} h_1(c)\ne 0,\quad \dots,\quad h_m(c)\ne 0. \end{align*} Then $c\models\theta(x)$. But the transcendental type also contains the polynomial inequation \begin{align*} x-c\ne 0, \end{align*} while $c$ satisfies \begin{align*} c-c=0. \end{align*} Thus $\theta(x)$ cannot imply every formula in the transcendental type. Algebraic types over $K$ are isolated by equations naming their realized elements, whereas the transcendental type remains non-principal because every single formula over $K$ can exclude only finitely many polynomial conditions. [/example] This algebraic example foreshadows the role of dimension in stability theory. Types over algebraically closed fields behave like generic points of algebraic sets, and the non-principal type of a transcendental element is the first sign of that geometry. [remark: Complete Types As Maximal Consistent Descriptions] A complete type over $A$ can be viewed as a maximal consistent theory in the variables $x$ after all elements of $A$ have been named. This viewpoint is syntactic, but it is equivalent to the semantic picture of a tuple in some elementary extension. The equivalence is exactly the compactness argument from the type existence theorem. [/remark] The remainder of the course develops this idea systematically. Stone spaces organize complete types topologically, saturated models realize many types at once, and stability theory studies when the collection of types over a parameter set is small enough to admit classification. Once types are understood as consistent descriptions, the next question is how to organize all such descriptions into a single geometric object. Stone spaces provide exactly that organization, turning compactness into a topology on complete types over a parameter set. # 2. Stone Spaces of Types This chapter turns the informal idea of a type as a consistent description into a topological object. In Chapter 1, compactness guaranteed that finitely satisfiable descriptions have realizations in elementary extensions; here the same compactness principle becomes compactness of a space. The points of the space are complete types, and its open sets record which formulas a type contains. The guiding question is how much geometry is already present in syntax. We will see that formulas behave like clopen subsets, complete types behave like ultrafilters, and compactness turns finite satisfiability into a covering property. This viewpoint is the bridge from basic type existence to later topics such as saturation, omitting types, and stability. ## Complete Types as Points If types are sets of formulas, the first problem is to decide which types should count as the points of a space. Partial types are too flexible: they can be extended in many directions and do not decide enough formulas to behave like points. As Chapter 1 emphasized, complete types over a parameter set are the right objects because every formula defines a yes-or-no property at such a point. From this point on, it is convenient to work inside a fixed sufficiently large and sufficiently saturated ambient model $\mathfrak{C}\models T$, often called a monster model. Parameter sets $A$ are subsets of this ambient structure, and formulas over $A$ use the elements of $A$ with their interpretations in $\mathfrak{C}$ fixed. [definition: Formula Over A] Let $T$ be a complete first-order theory in a language $L$, let $\bar{x}=(x_1,\dots,x_n)$, and let $A$ be a parameter set in a monster model $\mathfrak{C}\models T$. An $L(A)$-formula in variables $\bar{x}$ is a formula $\varphi(\bar{x},a)$ whose free variables are among $\bar{x}$ and whose parameters $a$ lie in $A$. [/definition] The parameter set is part of the data: changing $A$ changes the formulas that are allowed to describe the tuple. With the allowable formulas fixed, the next task is to specify which maximal consistent collections of those formulas will be treated as individual points. [definition: Complete n-Type Over A] Let $T$ be a complete first-order theory, let $A\subseteq \mathfrak{C}$, and let $n\in \mathbb N$. A complete $n$-type over $A$ is a set $p(\bar{x})$ of $L(A)$-formulas in variables $\bar{x}=(x_1,\dots,x_n)$ such that: 1. $p(\bar{x})\cup T$ is satisfiable; 2. for every $L(A)$-formula $\varphi(\bar{x})$, exactly one of $\varphi(\bar{x})$ and $\neg\varphi(\bar{x})$ belongs to $p$. [/definition] To study all such point-like objects at once, we need a name for the collection of complete $n$-types over the same parameters. This collection is the underlying set on which the Stone topology will be placed. [definition: Stone Space of Complete Types] Let $T$ be a complete first-order theory and let $A\subseteq \mathfrak{C}$. The Stone space $S_n(A)$ is the set of all complete $n$-types over $A$. [/definition] The notation $S_n(A)$ suppresses the ambient theory $T$, which is fixed throughout the discussion. When the theory is not complete, one instead takes complete types consistent with $T$; in this course the complete-theory setting keeps the notation lighter. To make $S_n(A)$ into a space rather than just a set, formulas should determine the basic observable properties of a type. [definition: Basic Clopen Set] For an $L(A)$-formula $\varphi(\bar{x})$, define \begin{align*} [\varphi] := \{p\in S_n(A): \varphi(\bar{x})\in p\}. \end{align*} The Stone topology on $S_n(A)$ is the topology generated by the sets $[\varphi]$ as $\varphi$ ranges over all $L(A)$-formulas in variables $\bar{x}$. [/definition] The notation $[\varphi]$ should be read as a truth set inside the space of complete types. It is not the set of realizations of $\varphi$ in a particular model; it is the set of all complete descriptions that include $\varphi$. [example: Dense Linear Orders Over The Empty Set] Let $T=\operatorname{DLO}$ be the theory of dense linear orders without endpoints, in the language $\{<\}$, and work in one variable over $\varnothing$. By quantifier elimination for dense linear orders without endpoints, every $L$-formula $\varphi(x)$ is equivalent modulo $T$ to a quantifier-free formula in the single variable $x$. With no parameters and no constants, the only atomic formulas in one free variable are \begin{align*} x=x \qquad\text{and}\qquad x<x. \end{align*} The theory $T$ proves $\forall x\, x=x$ and proves $\forall x\, \neg(x<x)$, because $<$ is irreflexive. Therefore every quantifier-free formula in $x$ is equivalent modulo $T$ to either $x=x$ or $\neg(x=x)$. Define \begin{align*} p(x):=\{\varphi(x): T\models \forall x\,\varphi(x)\}. \end{align*} This set is satisfiable, since any element of any model of $\operatorname{DLO}$ realizes all formulas true of every element. If $\varphi(x)$ is any $L$-formula, then the preceding paragraph gives exactly one of \begin{align*} T\models \forall x\,\varphi(x), \qquad T\models \forall x\,\neg\varphi(x), \end{align*} so exactly one of $\varphi(x)$ and $\neg\varphi(x)$ belongs to $p$. Thus $p$ is a complete $1$-type over $\varnothing$, and no second complete type can differ from it, because there is no formula over $\varnothing$ on which to make a different decision. Hence $S_1(\varnothing)=\{p\}$ is a one-point Stone space: a rich order can have only one type over the empty set when the language has no named cuts or constants. [/example] The previous example is small because there are no parameters. If parameters are allowed, formulas such as $x<a$, $a<x<b$, and $x>a$ begin to cut the order into many possible positions, and the Stone space records those cuts as distinct points. [example: Cuts Over A Parameter Set In Dense Linear Orders] Let $T=\operatorname{DLO}$, let $A\subseteq\mathfrak{C}$, and work in one variable over $A$. For each $a\in A$, the formulas \begin{align*} x<a,\qquad x=a,\qquad a<x \end{align*} are pairwise inconsistent, and $T$ proves that exactly one of them holds for any value of $x$: \begin{align*} T\models \forall x\,\bigl((x<a)\vee(x=a)\vee(a<x)\bigr) \end{align*} by trichotomy of linear orders. Hence a complete type $p\in S_1(A)$ contains exactly one of these three formulas for each $a\in A$. If $p$ contains $x=a$ for some $a\in A$, then for every $b\in A$ its decisions are forced by the order relation between $a$ and $b$: \begin{align*} a<b &\implies T\models \forall x\,(x=a\to x<b),\\ a=b &\implies T\models \forall x\,(x=a\to x=b),\\ b<a &\implies T\models \forall x\,(x=a\to b<x). \end{align*} Thus this type is the realized type of $a$. If $p$ contains no formula $x=a$, define \begin{align*} L_p&:=\{a\in A: a<x\in p\},\\ R_p&:=\{a\in A: x<a\in p\}. \end{align*} For every $a\in A$, exactly one of $a\in L_p$ and $a\in R_p$ holds, since the third option $x=a$ has been excluded. Also, if $a\in L_p$ and $b\in R_p$, then $a<b$: if $b\le a$, then $T$ proves \begin{align*} (b=a\vee b<a)\wedge (a<x)\wedge (x<b)\to x<x, \end{align*} contradicting irreflexivity of $<$. Therefore all elements placed on the left side lie below all elements placed on the right side. Conversely, any partition $A=L\cup R$ with $L\cap R=\varnothing$ and $l<r$ for all $l\in L$ and $r\in R$ gives the partial type \begin{align*} \{l<x:l\in L\}\cup\{x<r:r\in R\}. \end{align*} Every finite subset is satisfiable in a dense linear order: for finitely many left bounds $l_1,\dots,l_m$ and right bounds $r_1,\dots,r_k$, choose a largest listed left bound and a smallest listed right bound when they exist; density gives a point strictly between them, while the cases with only left or only right bounds use that the order has no endpoints. By compactness, this partial type extends to a complete $1$-type over $A$. Thus $S_1(A)$ consists of realized points of $A$ together with non-realized cuts of $A$, so allowing parameters turns the one empty-set type of dense linear orders into the space of possible positions relative to $A$. [/example] This example explains why types are often described as generalized points. They include ordinary elements of a model, but they also include ideal positions that become realized only after passing to a larger elementary extension. ## The Stone Topology The next question is whether the topology generated by formulas behaves like the zero-dimensional compact spaces familiar from Boolean algebra. There is a possible obstruction: declaring $[\varphi]$ open does not by itself make its complement open, unless every point has already decided between $\varphi$ and $\neg\varphi$. Completeness of types is exactly what turns this possible failure into a Boolean topological structure. [quotetheorem:5059] [citeproof:5059] This theorem depends on using complete types as points. For example, in $\operatorname{ACF}_0$ the partial type $\{x=x\}$ contains neither $x=0$ nor $x\ne 0$, so in a space of partial types $[x\ne 0]$ would not be the complement of $[x=0]$. It also records a limitation: formula-defined sets account for the compact open pieces, but an arbitrary [open set](/page/Open%20Set) can be an infinite union of such pieces and need not itself be defined by a single formula. The next point is to verify that formula sets are still enough locally, because finite intersections of formula neighbourhoods remain formula neighbourhoods. [quotetheorem:5060] [citeproof:5060] A useful consequence is that local questions about $S_n(A)$ can be translated back into single formulas, but only locally. The theorem does not say that every open set is formula-defined; for instance, a union of infinitely many pairwise distinct isolated algebraic types in $S_1(\varnothing)$ for $\operatorname{ACF}_0$ is open, but it need not be definable by a single formula. In one variable, quantifier elimination identifies formulas over $\varnothing$ with finite Boolean combinations of polynomial equations over $\mathbb Q$, so their definable subsets of the affine line are finite or cofinite. An infinite non-cofinite union of algebraic types therefore gives an open subset of the type space that is not a basic formula set. The completeness hypothesis is also doing real work: if partial types were used as points, the sets $[x=0]$ and $[x\ne 0]$ would no longer be complementary, so formula sets would not form a Boolean clopen basis. The formula-defined basis is likewise essential; a topology enlarged by declaring an arbitrary infinite union of algebraic singletons to be a basic open would have basic opens that are not represented by single formulas. This distinction will matter in Chapters 4 and 5 when isolated types are studied: an isolated type is precisely a type that has a one-formula neighbourhood containing no other type, not merely a neighbourhood obtained by an arbitrary union. [example: Basic Neighbourhoods In Algebraically Closed Fields] Let $T=\operatorname{ACF}_0$ in the language of rings, and work over $\varnothing$. We compute the basic neighbourhood \begin{align*} [x^2+1=0]=\{p\in S_1(\varnothing): x^2+1=0\in p\}. \end{align*} The polynomial $x^2+1$ is irreducible in $\mathbb Q[x]$: since it has degree $2$, reducibility over $\mathbb Q$ would give a rational root $r$, but $r^2+1=0$ has no solution in $\mathbb Q$. Let $i$ and $-i$ be the two roots of $x^2+1$ in an [algebraic closure](/page/Algebraic%20Closure). We show that they determine the same complete type over $\varnothing$. By *quantifier elimination for algebraically closed fields*, it is enough to compare polynomial equations $h(x)=0$ with $h\in\mathbb Q[x]$. Divide $h$ by $x^2+1$: \begin{align*} h(x)=q(x)(x^2+1)+(ax+b) \end{align*} for some $q\in\mathbb Q[x]$ and $a,b\in\mathbb Q$. At $x=i$ and $x=-i$, the term $q(x)(x^2+1)$ vanishes, so \begin{align*} h(i)=0 &\iff ai+b=0,\\ h(-i)=0 &\iff -ai+b=0. \end{align*} If $a=b=0$, both equations hold. If $a=0$ and $b\ne 0$, neither holds. If $a\ne 0$, then $ai+b=0$ would imply $i=-b/a\in\mathbb Q$, and $-ai+b=0$ would imply $-i=-b/a\in\mathbb Q$, both impossible. Hence \begin{align*} h(i)=0\iff h(-i)=0 \end{align*} for every $h\in\mathbb Q[x]$, so $i$ and $-i$ satisfy the same formulas over $\varnothing$. Thus $[x^2+1=0]$ is a singleton: it is the algebraic type associated to the irreducible polynomial $x^2+1\in\mathbb Q[x]$. Its complement is \begin{align*} [x^2+1\ne 0], \end{align*} because every complete type contains exactly one of $x^2+1=0$ and $x^2+1\ne 0$. This complementary clopen set contains the generic type, which contains $g(x)\ne 0$ for every non-zero $g\in\mathbb Q[x]$, and it contains every algebraic type attached to an irreducible polynomial $f\ne x^2+1$: since distinct irreducibles in $\mathbb Q[x]$ are coprime, there are $u,v\in\mathbb Q[x]$ with \begin{align*} u(x)(x^2+1)+v(x)f(x)=1, \end{align*} so no element can satisfy both $x^2+1=0$ and $f(x)=0$. [/example] The example also shows a common feature of type spaces: a formula may describe a finite algebraic condition, while its complement may still contain many different complete types. A clopen set need not be a singleton. ## Compactness and Separation After defining the topology, the central problem is to prove that it has strong global structure. A naive topology on partial descriptions would not have a reason to preserve limits: a family of increasingly restrictive finite conditions might have no designated point unless the whole finitely satisfiable set is completed. First-order compactness is exactly the principle that prevents this failure. The proof below is the topological form of that theorem: finite satisfiability of formulas becomes the finite intersection property for closed sets. [quotetheorem:5061] [citeproof:5061] This theorem is the main reason Stone spaces are useful, and its hypothesis is not cosmetic. Without first-order compactness, a family of formula conditions could be finitely satisfiable while having no complete type extending all of it, so the corresponding closed sets would have the finite intersection property but empty total intersection. Compactness of $S_n(A)$ also does not make the space metrizable, second countable, or discrete; large languages and large parameter sets usually produce spaces with too many basic clopens for those stronger properties. What compactness does provide is the exact tool used later: a failure of a global property can often be converted into a finitely satisfiable family of formulas, and compactness then supplies a type witnessing the failure. Compactness alone does not say that limits are unique or that distinct descriptions can be pulled apart. If two partial descriptions disagree only after different completions are chosen, the topology of partial types would not separate them by a formula. Complete types avoid this obstruction because any actual disagreement is witnessed by a formula that one type accepts and the other rejects. [quotetheorem:5062] [citeproof:5062] Hausdorffness separates pairs of points, but the separation again depends on completeness. If we allowed partial types, the points $\{x=x\}$ and $\{x\ne 0\}$ in $\operatorname{ACF}_0$ would be distinct and compatible; every formula-neighbourhood of $\{x=x\}$ contains all partial types, so no disjoint basic neighbourhoods separate them. The theorem also does not imply discreteness: separating $p$ from each $q\ne p$ by a formula need not give a single formula isolating $p$ from all other types at once. The next question asks whether any connected subset can contain two separated points; clopen separation will show that connected components are singletons. [quotetheorem:5063] [citeproof:5063] Total disconnectedness is a strong zero-dimensional separation statement, but it has a narrow scope. It rules out connected continua inside $S_n(A)$; it does not say that convergent behaviour disappears or that every singleton is open. The generic type in $\operatorname{ACF}_0$ gives the standard warning: it can be separated from each algebraic type by some clopen formula condition, yet no single formula over $\varnothing$ isolates it from all algebraic types at once. [example: Non-Isolated Generic Type In Algebraically Closed Fields] Let $T=\operatorname{ACF}_0$ and work in one variable over $\varnothing$. For each irreducible polynomial $f\in\mathbb Q[x]$, choose a root $\alpha$ of $f$ in an algebraic closure and let $p_f=\operatorname{tp}(\alpha/\varnothing)$. This type is isolated by $f(x)=0$. Indeed, if $h\in\mathbb Q[x]$ and $\beta$ is another root of $f$, then \begin{align*} h(\alpha)=0 &\iff \text{the minimal polynomial } f \text{ of } \alpha \text{ divides } h\\ &\iff h(\beta)=0. \end{align*} Thus all roots of $f$ satisfy the same polynomial equations over $\mathbb Q$, and by *quantifier elimination for algebraically closed fields* they satisfy the same formulas over $\varnothing$. Hence \begin{align*} [f(x)=0]=\{p_f\}. \end{align*} There is also a generic type \begin{align*} p_{\mathrm{gen}}(x):=\{g(x)\ne 0:g\in\mathbb Q[x],\ g\ne 0\}. \end{align*} Every finite subset is satisfiable: if $g_1,\dots,g_m$ are non-zero polynomials, then \begin{align*} G(x):=g_1(x)\cdots g_m(x) \end{align*} is non-zero, has only finitely many roots in an algebraically closed field, and therefore some $a$ satisfies $G(a)\ne 0$; for that $a$, \begin{align*} G(a)\ne 0 \implies g_1(a)\ne 0,\dots,g_m(a)\ne 0. \end{align*} By compactness this partial type extends to a complete type, and quantifier elimination shows the extension is unique because every one-variable formula over $\varnothing$ is equivalent to a Boolean combination of polynomial equations. Now let $[\varphi]$ be a basic neighbourhood of $p_{\mathrm{gen}}$, so $\varphi\in p_{\mathrm{gen}}$. By quantifier elimination in one variable, the set defined by $\neg\varphi$ is a finite Boolean combination of polynomial zero sets over $\mathbb Q$, hence is finite. Equivalently, there are non-zero polynomials $g_1,\dots,g_m\in\mathbb Q[x]$ such that \begin{align*} g_1(x)\ne 0\wedge\cdots\wedge g_m(x)\ne 0 \implies \varphi(x). \end{align*} Choose an irreducible polynomial $f\in\mathbb Q[x]$ distinct from every irreducible factor of $g_1\cdots g_m$. If $\alpha$ is a root of $f$, then no $g_j$ vanishes at $\alpha$, since otherwise $f$ would divide $g_j$. Therefore \begin{align*} g_1(\alpha)\ne 0,\dots,g_m(\alpha)\ne 0, \end{align*} so $\varphi\in p_f$ and $p_f\in[\varphi]$. Thus every basic neighbourhood of $p_{\mathrm{gen}}$ contains some algebraic type $p_f\ne p_{\mathrm{gen}}$, so the generic type is not isolated, while each algebraic type $p_f$ is isolated by its defining equation $f(x)=0$. [/example] This distinction between separated and isolated points reappears throughout stability theory. A Stone space may separate points very well, but still have limit behaviour encoded by non-isolated types. ## Boolean Algebras of Formulas The final question of the chapter is what algebraic object is dual to the Stone space. The first naive answer, "the set of formulas", has a duplication problem: two syntactically different formulas may define the same neighbourhood of every type over $A$. It also has a parameter problem: equivalence must be tested after the parameters from $A$ have their intended interpretation, not in the bare language with no constants for them. Quotienting by equivalence over the theory together with the parameter data removes these false distinctions, and logical connectives then become Boolean operations. [definition: Formula Boolean Algebra] Let $T$ be a complete first-order theory, let $A\subseteq\mathfrak{C}$, and fix variables $\bar{x}=(x_1,\dots,x_n)$. Define an [equivalence relation](/page/Equivalence%20Relation) on $L(A)$-formulas by \begin{align*} \varphi\equiv_{T,A}\psi \end{align*} if $\varphi$ and $\psi$ are equivalent in every elementary extension after the parameters from $A$ are interpreted as the named elements of $\mathfrak{C}$. The Boolean algebra $\mathcal{B}_n(A)$ is the set of equivalence classes of $L(A)$-formulas under $\equiv_{T,A}$, with Boolean operations induced by $\neg$, $\wedge$, and $\vee$. [/definition] A useful test for equality in the quotient is the following: $\varphi\equiv_{T,A}\psi$ exactly when the elementary diagram of $A$ in $\mathfrak{C}$ together with $T$ entails $\forall \bar{x}\,(\varphi(\bar{x})\leftrightarrow\psi(\bar{x}))$. This reformulation is often the most convenient way to check equality in the quotient, while the definition keeps the parameter interpretation explicit. The definition packages formulas into algebraic objects, but it is not automatic that this quotient sees exactly the topological information in the Stone space. Injectivity would fail if two different formula classes determined the same basic open set, while surjectivity would fail if a clopen subset required an infinite open description rather than one formula. The point to check is that compactness turns clopen covers into finite formula data, so the formula algebra and the clopen algebra have the same Boolean content. This is the comparison theorem needed to turn the construction into duality rather than just notation. It asks whether the map sending a formula class to the set of types containing that formula is exactly an isomorphism from the formula Boolean algebra onto the Boolean algebra of clopen subsets of the Stone space. [quotetheorem:5064] [citeproof:5064] This theorem is the precise Stone duality statement used in model theory, and both parts of the quotient matter. If formulas were not quotiented by equivalence over $A$, syntactically different formulas such as $\varphi$ and $\varphi\wedge(a=a)$ would give duplicate algebra elements with the same clopen set; if equivalence ignored the elementary diagram of the parameters, formulas using distinct named parameters could be incorrectly identified or separated. Compactness is the other essential ingredient in the surjectivity argument: without it, a clopen set covered by formula neighbourhoods need not have a finite formula subcover, so it might fail to be represented by a finite disjunction. Conversely, the theorem does not identify arbitrary open subsets with single formulas; it identifies the Boolean algebra of clopen subsets, whose compactness forces finite Boolean descriptions. The Boolean algebra of formulas modulo theory is recovered as the clopen algebra of the type space, and the type space is recovered next as the space of ultrafilters on that Boolean algebra. [definition: Ultrafilter On The Formula Algebra] An ultrafilter on $\mathcal{B}_n(A)$ is a subset $U\subseteq\mathcal{B}_n(A)$ such that: 1. $1\in U$ and $0\notin U$; 2. if $b,c\in U$, then $b\wedge c\in U$; 3. if $b\in U$ and $b\le c$, then $c\in U$; 4. for every $b\in\mathcal{B}_n(A)$, exactly one of $b$ and $\neg b$ belongs to $U$. [/definition] This definition translates completeness of a type into algebraic maximality. A complete type selects exactly one side of every definable Boolean split, and that selection is closed under logical consequence. The remaining step is to make this translation into a bijection. [quotetheorem:5065] [citeproof:5065] The ultrafilter viewpoint is often the most efficient way to remember the topology: basic opens ask whether a Boolean algebra element belongs to the ultrafilter. The maximality condition is necessary here. In $\operatorname{ACF}_0$, the filter generated by the tautology $x=x$ leaves $x=0$ undecided, so it corresponds to a partial description rather than a point of $S_1(\varnothing)$. The correspondence also does not choose realizations in the monster model; an ultrafilter may describe a non-realized type, with realization available only after passing to a sufficiently saturated or elementary extension. Thus $S_n(A)$ is the Stone dual of the Boolean algebra of formulas in $n$ variables over $A$. [example: Algebraic And Generic Types In Algebraically Closed Fields] Let $T=\operatorname{ACF}_0$ and work in one variable over $\varnothing$. By *quantifier elimination for algebraically closed fields*, every formula over $\varnothing$ is equivalent modulo $T$ to a Boolean combination of polynomial equations $h(x)=0$ with $h\in\mathbb Q[x]$. For an irreducible polynomial $f\in\mathbb Q[x]$, choose a root $\alpha$ of $f$ in an algebraic closure and define \begin{align*} p_f:=\operatorname{tp}(\alpha/\varnothing). \end{align*} If $\beta$ is any other root of $f$ and $h\in\mathbb Q[x]$, then \begin{align*} h(\alpha)=0 &\iff \text{the minimal polynomial of }\alpha\text{ over }\mathbb Q\text{ divides }h\\ &\iff f\mid h\\ &\iff h(\beta)=0. \end{align*} The first and last equivalences use that $f$ is the minimal polynomial of both $\alpha$ and $\beta$. Therefore all roots of $f$ satisfy exactly the same polynomial equations over $\mathbb Q$, and quantifier elimination gives that they satisfy exactly the same formulas over $\varnothing$. Hence $p_f$ is well-defined independently of the chosen root, and \begin{align*} [f(x)=0]=\{p_f\}, \end{align*} so $p_f$ is isolated by $f(x)=0$. The generic type is \begin{align*} p_{\mathrm{gen}}(x):=\{g(x)\ne 0:g\in\mathbb Q[x],\ g\ne 0\}. \end{align*} Every finite subset is satisfiable: for non-zero $g_1,\dots,g_m\in\mathbb Q[x]$, the product \begin{align*} G(x):=g_1(x)\cdots g_m(x) \end{align*} is non-zero, so it has finitely many roots in an algebraically closed field. Choose $a$ outside that finite root set. Then \begin{align*} G(a)\ne 0 &\iff g_1(a)\cdots g_m(a)\ne 0\\ &\implies g_j(a)\ne 0\qquad\text{for each }1\le j\le m. \end{align*} Thus the generic conditions are finitely satisfiable and extend to a complete type. Quantifier elimination makes the extension unique: every non-zero polynomial equation $h(x)=0$ is rejected, while $0=0$ is accepted, and Boolean combinations are then decided by their truth value under these polynomial decisions. The ultrafilter corresponding to $p_{\mathrm{gen}}$ consists exactly of the cofinite Zariski-open conditions on the affine line. Indeed, if $g\ne 0$, then \begin{align*} [g(x)\ne 0] \end{align*} removes only the finite zero set of $g$, while finite intersections satisfy \begin{align*} [g_1(x)\ne 0]\cap\cdots\cap[g_m(x)\ne 0] = [(g_1\cdots g_m)(x)\ne 0]. \end{align*} Consequently $S_1(\varnothing)$ consists of the algebraic types $p_f$ for irreducible $f\in\mathbb Q[x]$, together with the single generic type $p_{\mathrm{gen}}$. [/example] This example anticipates the stability-theoretic role of types. In stable theories, type spaces often admit strong structural descriptions, and generic types become algebraic objects rather than merely topological remainders. ## What This Chapter Establishes The chapter has converted complete types into a compact Hausdorff totally disconnected space. Formula-defined sets form a clopen basis, and the Boolean algebra of formulas modulo the theory is isomorphic to the algebra of clopen subsets. Compactness of first-order logic is the engine behind compactness of $S_n(A)$ and behind the identification of types with ultrafilters. For Chapter 3, the key idea is that compactness of $S_n(A)$ lets saturated models realize many small types at once. The later isolated-point interpretation returns in Chapters 4 and 5, where isolated points correspond to principal types and dense open sets encode omitting-type arguments. After introducing the Stone space viewpoint, we can ask when those abstract points are actually realized in models. Saturation and homogeneity answer that question by describing how large models can realize many small types at once and extend partial symmetry to full elementary control. # 3. Realization, Saturation, and Homogeneity This chapter turns the topology of type spaces into a working method for building and controlling large models. The main question is: when a small set of formulas is consistent, can we arrange for a model to contain an element realizing it? Once enough types are realized, elementary maps can be extended by back-and-forth, and types become the same thing as automorphism orbits inside a sufficiently rich ambient structure. ## Realizing Small Type Spaces The first problem is local: given a structure $M$ and a small parameter set $A \subset M$, we want $M$ to realize every complete description over $A$ that is consistent with its theory. Before asking for all small descriptions to be realized, we isolate what realization of a single type means. [definition: Type Realized In A Structure] Let $M$ be an $L$-structure, let $A \subset M$, and let $p(x) \in S_n(A)$. The type $p$ is realized in $M$ if there is $a \in M^n$ such that \begin{align*} M \models \varphi(a) \end{align*} for every formula $\varphi(x)$ in $p$. [/definition] This definition uses the parameter set $A$ to say which formulas are allowed in the description. Larger parameter sets can describe obstructions that no small finite fragment detects: in $(\mathbb Q,<)$, the formulas saying that $x$ lies above every rational below $\sqrt 2$ and below every positive rational above $\sqrt 2$ are finitely satisfiable, but no rational realizes the whole cut. Saturation is the condition that these small finitely consistent descriptions are not omitted. [definition: Kappa Saturated Structure] Let $\kappa$ be an infinite cardinal. An $L$-structure $M$ is $\kappa$-saturated if, for every set $A \subset M$ with $|A| < \kappa$, every complete type $p(x) \in S_n(A)$ with $n \in \mathbb N$ and $n \geq 1$ is realized in $M$. [/definition] Saturation is stated using complete types, but compactness usually gives us partial information first: a list of formulas whose finite sublists can be realized. The following equivalence is needed because it converts the topological and compactness language of the previous chapters into the model-internal condition used here. [quotetheorem:5066] [citeproof:5066] This theorem explains why saturation is the model-theoretic analogue of compactness inside a fixed model. Compactness says that finitely satisfiable descriptions have some realization in some elementary extension; saturation says the realization already lies in the chosen structure. The finite satisfiability hypothesis cannot be weakened to arbitrary syntactic consistency. For a concrete failure, take $M=(\mathbb N,<)$ and name the actual parameter $0\in M$ by a constant $c$. The single formula $x<c$ is consistent with the unparameterised theory $\operatorname{Th}(M)$ after interpreting $c$ as a non-minimal element in another model of that theory, but it is not finitely satisfiable in $M$ with $c$ interpreted as $0$. The theorem therefore uses finite satisfiability in the given parameter structure, not consistency with a reduct in which the named parameters can move. It also does not produce new elements; it is a criterion for when the chosen structure is already large enough to contain the witnesses demanded by its small parameter sets. [example: Dense Order Of The Rationals Is Not Saturated] Consider $(\mathbb Q,<)$ as a model of dense linear order without endpoints, and use the parameter set \begin{align*} A=\{q\in \mathbb Q:q^2<2\}\cup \{q\in \mathbb Q:q^2>2,\ q>0\}. \end{align*} Let $\Sigma(x)$ be the partial type over $A$ consisting of the formulas \begin{align*} q<x &\quad \text{for each } q\in\mathbb Q \text{ with } q^2<2,\\ x<r &\quad \text{for each } r\in\mathbb Q \text{ with } r^2>2 \text{ and } r>0. \end{align*} Every finite subset $\Sigma_0(x)\subset \Sigma(x)$ mentions only finitely many lower parameters $q_1,\dots,q_m$ with $q_i^2<2$ and finitely many upper parameters $r_1,\dots,r_n$ with $r_j^2>2$ and $r_j>0$. If both lists are nonempty, put \begin{align*} q_*=\max(q_1,\dots,q_m),\qquad r_*=\min(r_1,\dots,r_n). \end{align*} For each $i$, $q_i<\sqrt 2$: if $q_i<0$ this is immediate, and if $q_i\geq 0$, then $q_i^2<2=(\sqrt2)^2$ implies $q_i<\sqrt2$. For each $j$, $r_j>\sqrt2$, since $r_j>0$ and $r_j^2>2=(\sqrt2)^2$. Hence \begin{align*} q_*<\sqrt2<r_*, \end{align*} so $q_*<r_*$. The rational number \begin{align*} s=\frac{q_*+r_*}{2} \end{align*} satisfies \begin{align*} q_i\leq q_*<\frac{q_*+r_*}{2}<r_*\leq r_j \end{align*} for all $i$ and $j$, so $s$ realizes $\Sigma_0(x)$. If only lower bounds occur, choose $s=q_*+1$; if only upper bounds occur, choose $s=r_*-1$; if no bounds occur, choose any rational $s$. Thus $\Sigma(x)$ is finitely satisfiable in $(\mathbb Q,<)$. No rational realizes all of $\Sigma(x)$. Suppose $s\in\mathbb Q$ did. Since $0^2<2$, the formula $0<x$ belongs to $\Sigma(x)$, so $s>0$. If $s^2<2$, then $s<x$ belongs to $\Sigma(x)$, forcing $s<s$, impossible. If $s^2>2$, then $x<s$ belongs to $\Sigma(x)$ because $s>0$, again forcing $s<s$. Therefore $s^2=2$, but no rational number has square $2$. Hence $(\mathbb Q,<)$ omits this finitely satisfiable type over the [countable set](/page/Countable%20Set) $A$, so it is not $\aleph_1$-saturated. [/example] The rational order fails because it has a missing cut. Saturation repairs this defect by filling every cut, and more generally every small consistent pattern, at the relevant cardinal scale. [example: Countably Saturated Nonstandard Arithmetic] Let $M \models \operatorname{PA}$ be countably saturated and nonstandard, and let $(a_i)_{i\in\mathbb N}$ be an increasing sequence from $M$. Consider the partial type over the countable parameter set $\{a_i:i\in\mathbb N\}$: \begin{align*} p(x)=\{a_i<x:i\in\mathbb N\}. \end{align*} We show that $p(x)$ is finitely satisfiable in $M$. Let $p_0(x)\subset p(x)$ be finite. Then there are indices $i_1,\dots,i_m\in\mathbb N$ such that \begin{align*} p_0(x)=\{a_{i_1}<x,\dots,a_{i_m}<x\}. \end{align*} By assumption, the finite set $\{a_{i_1},\dots,a_{i_m}\}$ has an upper bound $c\in M$, so \begin{align*} a_{i_j}\leq c \end{align*} for every $j=1,\dots,m$. Since $M\models \operatorname{PA}$, the successor $c+1$ exists in $M$, and the order axioms give \begin{align*} a_{i_j}\leq c<c+1 \end{align*} for every $j=1,\dots,m$. Hence $c+1$ realizes $p_0(x)$ in $M$. Thus every finite subset of $p(x)$ is realized in $M$. Since $p(x)$ uses only countably many parameters and $M$ is countably saturated, there is $b\in M$ such that \begin{align*} M\models a_i<b \end{align*} for every $i\in\mathbb N$. Saturation has compressed the countable family of upper-bound requirements into one element of the model. [/example] ## Saturated Elementary Extensions The next problem is global: if the model we start with is not saturated, can we pass to an elementary extension that is saturated enough for the arguments we want? Compactness gives individual realizations, while a transfinite construction realizes many types in succession. [definition: Elementary Extension] Let $M$ and $N$ be $L$-structures. The structure $N$ is an elementary extension of $M$, written $M \preceq N$, if $M \subset N$ and for every $L(M)$-formula $\varphi(x)$ and every tuple $a$ from $M$, \begin{align*} M \models \varphi(a) \quad \iff \quad N \models \varphi(a). \end{align*} [/definition] Elementary extensions preserve first-order truth over the old model, but they may add realizations of types omitted by the old model. To use this systematically, we need an existence theorem saying that a sufficiently large elementary extension can realize all small types at once. [quotetheorem:5053] [citeproof:5053] The theorem gives saturated extensions when the language and the cardinal arithmetic leave enough room to repeat the compactness construction without exceeding the target size. Each hypothesis prevents a specific failure. The assumption $|M|\leq\kappa$ is forced by containment: if $|M|>\kappa$, no elementary extension $N\succeq M$ can have $|N|\leq\kappa$. The language bound is also substantive. If $L$ contains constants $(c_i)_{i<\lambda}$ with $\lambda\geq\kappa$, then even over the empty set there are $\lambda$ many atomic formulas $x=c_i$ to account for, and the collection of complete $1$-types over small parameter sets need not be listable in $\kappa$ many steps. The cardinal arithmetic assumption controls the repeated bookkeeping: without $\kappa^{<\kappa}=\kappa$, there may be more than $\kappa$ small parameter sets or more than $\kappa$ formulas over them, so a construction that realizes all of them can be forced past size $\kappa$ before saturation is reached. The conclusion is also limited. It does not say that every elementary extension of size $\kappa$ is saturated, and it does not realize types over parameter sets of size $\kappa$ or larger. It gives one carefully built extension that realizes all descriptions over smaller parameter sets. Since later arguments repeatedly introduce new small sets and new types, naming a single sufficiently rich ambient model avoids rebuilding an elementary extension at each step. This motivates the monster model convention: we place all subsequent small configurations inside one preselected structure with enough saturation and symmetry for the argument at hand. [definition: Monster Model Convention] Let $T$ be a complete theory. A monster model for $T$ is a model $\mathfrak C\models T$ chosen to be sufficiently saturated and sufficiently homogeneous for all parameter sets and tuples under discussion. [/definition] The phrase "sufficiently" is a convention rather than a new axiom of first-order logic. During an argument, all sets called small are assumed to have cardinality below the saturation and homogeneity degree of $\mathfrak C$. [remark: Small Sets] Inside a monster model $\mathfrak C$, a set $A\subset \mathfrak C$ is called small when $|A|$ is below the ambient saturation bound. The exact bound is usually suppressed because every construction in the notes uses only sets below that fixed threshold. [/remark] The convention lets us stop rebuilding elementary extensions every time a new type appears. Instead of saying "move to a larger elementary extension realizing this type," we say "work in $\mathfrak C$" and use saturation there. [example: Saturated Algebraically Closed Fields] Let $K\models \operatorname{ACF}_p$ be saturated enough for the small subfield $A\subset K$. The partial type expressing that $x$ is transcendental over $A$ is \begin{align*} \Sigma(x)=\{f(x)\ne 0: f\in A[X]\setminus\{0\}\}. \end{align*} We verify finite satisfiability in $K$. Let $\Sigma_0(x)\subset \Sigma(x)$ be finite, say \begin{align*} \Sigma_0(x)=\{f_1(x)\ne 0,\dots,f_m(x)\ne 0\}, \end{align*} with each $f_i\in A[X]\setminus\{0\}$. Since $A$ is a field, $A[X]$ is an integral domain, so \begin{align*} F(X)=f_1(X)\cdots f_m(X) \end{align*} is a nonzero polynomial in $A[X]$. Viewed as a polynomial over $K$, the polynomial $F$ has only finitely many roots in $K$. The field $K$ is infinite, so choose $a\in K$ outside this finite root set. Then \begin{align*} F(a)\ne 0. \end{align*} But \begin{align*} F(a)=f_1(a)\cdots f_m(a), \end{align*} so no factor $f_i(a)$ can be $0$. Hence \begin{align*} K\models f_i(a)\ne 0 \end{align*} for every $i=1,\dots,m$, and $a$ realizes $\Sigma_0(x)$. Thus $\Sigma(x)$ is finitely satisfiable in $K$. By saturation, there is $b\in K$ such that \begin{align*} K\models f(b)\ne 0 \end{align*} for every nonzero $f\in A[X]$. This exactly says that $b$ is not algebraic over $A$, so $b$ is transcendental over $A$; saturation turns the infinitely many polynomial-avoidance requirements into one element of the field. [/example] ## Saturation And Homogeneity Realizing types gives existence of tuples. Homogeneity strengthens this into symmetry: two tuples with the same type over a small parameter set should be movable to each other by an automorphism fixing that parameter set. [definition: Kappa Homogeneous Structure] Let $\kappa$ be an infinite cardinal. An $L$-structure $M$ is $\kappa$-homogeneous if, whenever $A\subset M$ has $|A|<\kappa$ and $a,b\in M^n$ satisfy \begin{align*} \operatorname{tp}^M(a/A)=\operatorname{tp}^M(b/A), \end{align*} the partial elementary map with domain $A\cup\{a_1,\dots,a_n\}$ that fixes $A$ pointwise and sends each $a_i$ to $b_i$ extends to an automorphism of $M$. [/definition] Homogeneity is what makes types behave geometrically, but it is not built into the definition of saturation. Saturation realizes types over small sets; homogeneity asks for a global automorphism extending a small partial elementary map. The obstruction is that a back-and-forth construction can grow its domain too large unless the cardinal arithmetic keeps every initial stage small. In the regular cardinal range used for monster models, saturation supplies enough realizations at each stage for the back-and-forth to produce the required automorphisms. [quotetheorem:5068] [citeproof:5068] The regularity and size hypotheses are part of the mechanism. Saturation supplies witnesses only over domains of size less than $\kappa$, so the construction must ensure that every proper initial stage still has small domain. Regularity gives exactly that: before stage $\kappa$, fewer than $\kappa$ elements have been added. The older-looking formulation with $|M|<\kappa$ would be misleading for infinite structures, because no infinite $M$ of size below $\kappa$ can be $\kappa$-saturated: the partial type $\{x\ne m:m\in M\}$ over the parameter set $M$ is finitely satisfiable but has no realization in $M$. Thus the useful theorem is not "saturation far above the model size gives homogeneity"; it is that saturation at the model's own regular size is enough for back-and-forth over smaller parameter sets. [example: Back And Forth In Dense Linear Orders] Let $M$ be $\kappa$-saturated and let $A\subset M$ be small. Suppose $a,b\in M^n$ have the same order pattern over $A$: for every $i,j$ and every $\alpha\in A$, \begin{align*} a_i<a_j &\iff b_i<b_j,\\ a_i=\alpha &\iff b_i=\alpha,\\ a_i<\alpha &\iff b_i<\alpha,\\ \alpha<a_i &\iff \alpha<b_i. \end{align*} Then the map fixing $A$ pointwise and sending $a_i$ to $b_i$ is well-defined and preserves all order relations among its domain elements. Since dense linear orders without endpoints eliminate quantifiers by *Quantifier Elimination for Dense Linear Orders*, this map is partial elementary. We extend this partial elementary map by back-and-forth. Suppose at some stage we have a small partial elementary order-isomorphism $f:B\to C$ with $A\subset B,C$. To add an element $c\in M$, form the cut over $C$ \begin{align*} L&=\{f(d):d\in B \text{ and } d<c\},\\ U&=\{f(d):d\in B \text{ and } c<d\}. \end{align*} If $\ell\in L$ and $u\in U$, then $\ell=f(d_1)$ and $u=f(d_2)$ for some $d_1,d_2\in B$ with \begin{align*} d_1<c<d_2. \end{align*} Thus $d_1<d_2$, and because $f$ preserves order, \begin{align*} \ell=f(d_1)<f(d_2)=u. \end{align*} So every finite subset of the partial type \begin{align*} \{\,\ell<x:\ell\in L\,\}\cup \{\,x<u:u\in U\,\} \end{align*} is realized in $M$: finitely many lower and upper bounds reduce to one interval $\ell_*<x<u_*$, which has a point by density; if only lower bounds occur, use no endpoints to choose a point above them, and similarly for only upper bounds. Since $B$ is small, this type is over a small parameter set, so saturation gives $e\in M$ such that \begin{align*} \ell<e<u \end{align*} for all $\ell\in L$ and $u\in U$. Then $f\cup\{(c,e)\}$ again preserves order, hence is partial elementary. The back step is the same argument applied to $f^{-1}$. Enumerating $M$ and alternating forth and back steps puts every element of $M$ into both the domain and the range. The union is therefore an automorphism of $M$ fixing $A$ pointwise and sending $a$ to $b$. In dense linear orders, saturation turns equality of cuts over small sets into actual automorphic symmetry. [/example] ## Automorphism Orbits And Types The final problem of the chapter is to connect syntax and symmetry. A complete type over $A$ records all formulas with parameters from $A$; an automorphism fixing $A$ preserves exactly that information. [definition: Orbit Over A Parameter Set] Let $\mathfrak C$ be a monster model, let $A\subset \mathfrak C$ be small, and let $a\in \mathfrak C^n$. The orbit of $a$ over $A$ is \begin{align*} \operatorname{Orb}(a/A)=\{\sigma(a):\sigma\in \operatorname{Aut}(\mathfrak C/A)\}, \end{align*} where $\operatorname{Aut}(\mathfrak C/A)$ denotes the automorphisms of $\mathfrak C$ fixing every element of $A$. [/definition] The orbit relation is semantic, while equality of types is syntactic. One direction is forced because automorphisms fixing $A$ preserve every formula with parameters from $A$. The reverse direction is the real issue: two tuples may satisfy the same formulas over $A$ without any visible map carrying one to the other unless the ambient model is homogeneous enough to extend the partial elementary correspondence. In a monster model, that obstruction is removed, so syntactic type classes become exactly automorphism orbits. [quotetheorem:5069] [citeproof:5069] This theorem uses the Stone-space viewpoint of Chapter 2 and prepares the stability-theoretic use of types in Chapter 6. Stone spaces describe the possible complete descriptions; saturated homogeneous monsters turn those descriptions into geometric pieces controlled by automorphism groups. The hypotheses are essential. Completeness of $T$ keeps all tuples inside one common elementary universe: if a class is presented by an incomplete theory, two tuples may satisfy the same formulas in their separate models of different completions while there is no single ambient automorphism group relating them. Saturation ensures that the relevant types have representatives in the chosen universe; without it, the standard model $(\mathbb N,+,\cdot,<)$ omits the finitely satisfiable partial type $\{n<x:n\in\mathbb N\}$, so that description has no orbit inside the model. Homogeneity is the reverse direction: take a graph $G$ formed as the disjoint union of two rooted connected graphs $(G_0,a)$ and $(G_1,b)$ that are elementarily equivalent as rooted graphs but not isomorphic as rooted graphs. Then $a$ and $b$ satisfy the same parameter-free formulas in $G$, but no automorphism of $G$ sends $a$ to $b$, since such an automorphism would give an isomorphism between the rooted connected components. The finite-tuple restriction matters because the theorem compares complete $n$-types and automorphisms act coordinatewise on fixed finite arities; infinite tuples require a separate bound on tuple length and a matching strong homogeneity assumption. Smallness of $A$ keeps the parameter set below the ambient saturation and homogeneity bounds; if $A$ has the size of the monster, the type $\{x\ne c:c\in \mathfrak C\}$ over $A=\mathfrak C$ is finitely satisfiable but cannot be realized inside $\mathfrak C$. [remark: Why The Monster Is Useful] When working in $\mathfrak C$, a type $p\in S_n(A)$ can be studied through any realization $a\models p$ and the orbit $\operatorname{Orb}(a/A)$. Definability, invariance, forking, and stability will later be expressed by asking how these orbits behave as $A$ varies. [/remark] The discussion so far has emphasized realization, but model theory also needs a way to exclude unwanted descriptions. Omitting types reverses the previous perspective: instead of forcing types to appear in a model, we study how to build models in which certain consistent nonprincipal types are absent. # 4. Omitting Types This chapter explains how to build models in which specified consistent descriptions fail to be realized. Chapters 2 and 3 treated types as points of Stone spaces and used compactness and saturation to guarantee realization in suitable elementary extensions; omitting types asks for the opposite construction. The main theorem says that in a countable language, a countable collection of sufficiently non-isolated types can be avoided by a countable model. ## Realization, Omission, and Isolation When a type is consistent with a theory, compactness gives a model somewhere realizing it, but it does not say that every model must realize it. The first question is therefore: which types are forced into every model of a theory, and which types can be avoided by choosing the model carefully? [definition: Realized Type] Let $T$ be a first-order theory in a language $L$, let $M \models T$, let $A \subseteq M$, and let $p(x)$ be a partial type over $A$. The model $M$ realizes $p(x)$ if there is a tuple $a$ from $M$ such that $M \models \theta(a)$ for every formula $\theta(x) \in p(x)$, with parameters from $A$ interpreted by their given elements of $M$. [/definition] Realization turns a syntactic list of requirements into an actual element. To use types for model construction, we also need to record the opposite outcome: the theory may be satisfied while the entire description remains absent. [definition: Omitted Type] Let $T$ be a first-order theory in a language $L$, let $M \models T$, let $A \subseteq M$, and let $p(x)$ be a partial type over $A$. The model $M$ omits $p(x)$ if no tuple $a$ from $M$ realizes $p(x)$, with parameters from $A$ interpreted by their given elements of $M$. [/definition] For complete types over the empty set, omission means that every tuple in the model has a different complete theory. Partial types are often the useful objects in applications because they can express an infinite limiting behaviour, such as lying above every named natural number or below every right endpoint candidate. [example: Finitely Satisfiable but Omitted Description] Work in the ordered set $(\mathbb Q,<)$, and fix the irrational real number $\alpha=\sqrt 2$. Let \begin{align*} A=\{q\in\mathbb Q:q<\alpha\}\cup\{r\in\mathbb Q:\alpha<r\}. \end{align*} Over the parameter set $A$, consider the partial type \begin{align*} p(x)=\{q<x:q\in\mathbb Q,\ q<\alpha\}\cup\{x<r:r\in\mathbb Q,\ \alpha<r\}. \end{align*} This type asks for an element filling the Dedekind cut of $\alpha$ inside the rationals. Every finite subset of $p(x)$ is realized in $(\mathbb Q,<)$. Indeed, a finite subset mentions only finitely many lower bounds $q_1,\ldots,q_m<\alpha$ and finitely many upper bounds $r_1,\ldots,r_n>\alpha$. Put \begin{align*} q_*=\max\{q_1,\ldots,q_m\},\qquad r_*=\min\{r_1,\ldots,r_n\}, \end{align*} with the evident omissions if one side is empty. Since $q_*<\alpha<r_*$ and $\mathbb Q$ is dense in $\mathbb R$, there is $s\in\mathbb Q$ such that \begin{align*} q_*<s<r_*. \end{align*} Then $s$ satisfies all the finitely many inequalities under consideration. The whole type is omitted in $(\mathbb Q,<)$. If $s\in\mathbb Q$ and $s<\alpha$, choose a rational $q$ with \begin{align*} s<q<\alpha; \end{align*} then the formula $q<x$ belongs to $p(x)$ and is not satisfied by $s$. If $s>\alpha$, choose a rational $r$ with \begin{align*} \alpha<r<s; \end{align*} then the formula $x<r$ belongs to $p(x)$ and is not satisfied by $s$. Since $\alpha$ is irrational, no rational equals $\alpha$. Thus no element of $\mathbb Q$ realizes all of $p(x)$. The point is that finite satisfiability of a description guarantees only that each finite approximation has room; it does not force the limiting cut to be filled inside the chosen model. [/example] This example shows that omission is not the same as inconsistency. We next need the obstruction to omission: sometimes a single formula already forces the whole type, and then any model containing a solution of that formula realizes the type. [definition: Isolated Type] Let $T$ be a complete theory and let $p(x) \in S_n(T)$. The type $p(x)$ is isolated if there is a formula $\theta(x)$ such that $\theta(x) \in p(x)$ and for every formula $\varphi(x)$, if $\varphi(x) \in p(x)$ then \begin{align*} T \models \forall x(\theta(x) \to \varphi(x)). \end{align*} [/definition] A formula isolating a type is a finite handle on what would otherwise be an infinite description. The next question is whether this finite handle forces realization in every model; the answer identifies the basic obstruction that must be excluded before an omitting theorem can hold. [quotetheorem:5070] [citeproof:5070] This theorem explains why the omitting theorem cannot apply to isolated types, but its hypotheses are doing real work. Completeness ensures that the complete type $p$ is tested against a fixed theory of models; without completeness, a formula may isolate a description relative to one completion but have no uniform force across all models of the weaker theory. The satisfiability assumption $T \models \exists x\,\theta(x)$ is also necessary: if no model of $T$ contains a solution of $\theta$, then $\theta$ can imply every formula in $p$ vacuously while giving no realized tuple. Finally, the result does not say that every isolated-looking partial description is realized; it says that once a satisfiable formula genuinely forces a complete type, every model of the complete theory must contain a realization of that formula and hence of the type. The construction needs a negative version of isolation, formulated for partial types as the absence of any satisfiable formula that forces the whole description. Isolation was defined for complete types of complete theories, but the omitting theorem has to handle partial types over theories that may not decide every sentence. We therefore separate the finite forcing phenomenon from completeness: a type is dangerous for omission exactly when some satisfiable formula already commits every realizing tuple to satisfy all formulas in the type. [definition: Principal and Non-Principal Type] Let $T$ be a theory and let $p(x)$ be a partial type over the empty set. The type $p(x)$ is principal over $T$ if there is a formula $\theta(x)$ such that $T \cup \{\exists x\,\theta(x)\}$ is satisfiable and, for every $\varphi(x) \in p(x)$, \begin{align*} T \models \forall x(\theta(x) \to \varphi(x)). \end{align*} The type $p(x)$ is non-principal over $T$ if it is not principal over $T$. [/definition] For complete types of a complete theory, principality agrees with isolation. For partial types, the definition says that no satisfiable formula can force all members of the partial description. The next example shows the typical non-principal pattern: every finite approximation can be met, but the infinite list describes a cut missing from the chosen model. [example: Endpoint Cuts in Dense Linear Order] Let $T$ be the theory of dense linear orders without endpoints in the language $\{<\}$. In $(\mathbb Q,<)$, choose an increasing sequence $(a_n)_{n\in\mathbb N}$ of rationals converging in $\mathbb R$ to an irrational number $\alpha$. Since the sequence is increasing and converges to $\alpha$, each $a_n$ satisfies $a_n<\alpha$. Consider the partial type \begin{align*} p(x)=\{a_n<x:n\in\mathbb N\}\cup\{x<b:b\in\mathbb Q,\ \alpha<b\}. \end{align*} This type asks for an element strictly above every rational $a_n$ and strictly below every rational number lying above the cut point $\alpha$. Let $\Delta(x)\subseteq p(x)$ be finite. Then there are finitely many lower-bound indices $n_1,\ldots,n_k$ and finitely many rational upper bounds $b_1,\ldots,b_\ell$ with $\alpha<b_i$ such that \begin{align*} \Delta(x)\subseteq \{a_{n_1}<x,\ldots,a_{n_k}<x\}\cup\{x<b_1,\ldots,x<b_\ell\}. \end{align*} If lower-bound formulas occur, let \begin{align*} A=\max\{a_{n_1},\ldots,a_{n_k}\}; \end{align*} otherwise choose any rational $A<\alpha$. If upper-bound formulas occur, let \begin{align*} B=\min\{b_1,\ldots,b_\ell\}; \end{align*} otherwise choose any rational $B>\alpha$. In either case, \begin{align*} A<\alpha<B, \end{align*} so there is a rational $r$ with \begin{align*} A<r<B \end{align*} by the density of $\mathbb Q$ in $\mathbb R$. For each mentioned lower bound, $a_{n_j}\le A<r$, so $\mathbb Q\models a_{n_j}<r$. For each mentioned upper bound, $r<B\le b_i$, so $\mathbb Q\models r<b_i$. Hence $r$ realizes $\Delta(x)$ in $(\mathbb Q,<)$, and every finite part of $p$ is realized in $\mathbb Q$. The full type is not realized in $(\mathbb Q,<)$. Let $q\in\mathbb Q$. Since $\alpha$ is irrational, $q\ne \alpha$. If $q<\alpha$, then $\alpha-q>0$, and convergence $a_n\to\alpha$ gives some $n$ such that \begin{align*} \alpha-a_n<\alpha-q. \end{align*} Subtracting $\alpha$ from both sides gives \begin{align*} -a_n<-q, \end{align*} and multiplying by $-1$ reverses the inequality: \begin{align*} q<a_n. \end{align*} Thus $q$ does not satisfy the formula $a_n<x$. If $q>\alpha$, then the formula $x<q$ belongs to $p(x)$, and substituting $x=q$ would require \begin{align*} q<q, \end{align*} which is false in a linear order. Therefore no rational realizes all formulas in $p(x)$. The example isolates the basic pattern of a missing cut: all finite approximations fit inside rational intervals, but the infinite description asks for the irrational cut point itself. [/example] This endpoint example is the template for the rest of the chapter. The type is finitely satisfiable, since any finite part gives only finitely many lower bounds and finitely many upper bounds with a genuine interval between them. The whole infinite description may still be absent from the model because it asks for an element realizing the irrational cut itself. ## The Omitting Types Theorem The main problem is to turn local avoidability into a global model construction. For a single non-principal type, every basic condition can be strengthened to avoid at least one formula from the type; for countably many types, the construction must meet countably many avoidance requirements at once while still producing a model of the theory. [quotetheorem:5071] [citeproof:5071] The local lemma says that every finite stage of a construction has room to dodge the next type requirement. Its non-principality hypothesis is necessary: if $p$ is principal, witnessed by a satisfiable formula $\theta$, then every realization of $\theta$ must satisfy every formula in $p$, so there is no formula from $p$ whose negation can be added while preserving consistency with $\theta$. The satisfiability of the starting formula is also part of the content, since an inconsistent condition cannot be strengthened into a useful consistent one. Thus the lemma is not a general way to make arbitrary types disappear; it is a local extension principle available exactly when no finite satisfiable condition already forces the type. [quotetheorem:5054] [citeproof:5054] The theorem is a Baire category argument in syntactic form. The Stone space viewpoint says that non-principal types are nowhere dense singleton or closed sets, and countability allows a point to be chosen outside their union while also meeting the open sets needed for a model. Non-principality is necessary because an isolated complete type whose isolating formula is satisfiable is realized in every model, as the previous theorem showed. Countability is also essential to this proof: the Henkin construction only meets obligations that can be enumerated, and in uncountable languages there are versions of omitting types that require stronger set-theoretic or topological hypotheses. The theorem is stated for types over the empty set; parameter types require either naming the parameters in the language or working inside a fixed parameter set with compatible interpretations. It also does not produce a saturated model, a unique model, or a model omitting arbitrary principal types; it only guarantees one countable model avoiding the specified countable family of non-principal types. [remark: Why Countability Matters] The countability of the language and the family of types lets the construction enumerate all obligations. In uncountable languages, omitted type statements need stronger hypotheses, and the plain countable theorem can fail. [/remark] The theorem becomes concrete when the omitted type describes an element that is infinitesimal, infinitely large, or located at a missing cut. Ordered structures provide the most transparent examples because the formulas in the type have a direct order-theoretic meaning. [example: Avoiding Infinitesimals in Ordered Fields] Work in the language of ordered rings expanded by constant symbols for each rational number, and let $T$ be a theory of ordered fields extending the ordered field of rational constants. The partial type \begin{align*} p(x)=\{0<x\}\cup\{x<1/n:n\in\mathbb N,\ n\ge 1\} \end{align*} asks for an element that is positive but smaller than every positive rational number of the form $1/n$. In the standard ordered field $(\mathbb Q,+,\cdot,<)$, no rational realizes $p(x)$. If $q\in\mathbb Q$ and $q\le 0$, then $q$ already fails the formula $0<x$. If $q>0$, then $1/q$ is a positive rational, so choose $n\in\mathbb N$ with \begin{align*} n>1/q. \end{align*} Multiplying both sides by the positive rational $q$ gives \begin{align*} nq>1, \end{align*} and dividing by the positive integer $n$ gives \begin{align*} q>1/n. \end{align*} Thus $q$ fails the formula $x<1/n$ from $p(x)$. Hence every rational number fails at least one formula in $p(x)$, so $(\mathbb Q,+,\cdot,<)$ omits the positive infinitesimal type. The model-building point is that omitting this type enforces Archimedean behaviour for positive elements: instead of adding a single axiom saying “there are no infinitesimals,” we omit the infinite description requiring \begin{align*} 0<x,\quad x<1,\quad x<1/2,\quad x<1/3,\quad \ldots . \end{align*} When this infinitesimal type is non-principal over the background ordered-field theory being used, the [omitting types theorem](/theorems/5054) constructs a countable model in which no element satisfies all these inequalities at once. [/example] The same mechanism is often used in reverse. Instead of naming the intended model first, we describe all unwanted behaviours as non-principal types and then apply the theorem to obtain a model avoiding them. ## Applications to Definability and Special Models The theorem becomes useful when a semantic property can be encoded as the realization of a type. To prove that a property is not forced or not definable, we build a model that satisfies the background theory but omits the type whose realization would witness the property. [quotetheorem:5072] [citeproof:5072] This restricted form captures the model-theoretic idea behind Vaught-style tests: if all countable models are the same, then there is no room for a non-isolated type to appear in one countable model and disappear in another. Completeness matters because types over the empty set are being compared across all models of one fixed complete theory; for an incomplete theory, different completions may realize incompatible descriptions without producing a contradiction to uniqueness inside a completion. Countability is needed twice: omitting types supplies countable models, and the conclusion compares countable models up to isomorphism rather than arbitrary models. The uniqueness hypothesis is the decisive semantic input, since if a complete countable theory has two non-isomorphic countable models, a non-principal type may be realized in one and omitted in another. This restricted test does not classify uncountable models and does not say that every type is isolated merely because it is consistent; it only forces isolation for complete types that are realized somewhere when there is a unique countable model. [example: Non-Definability from Omission] Let $T$ be countable, and suppose the property $P(x)$ is represented over models of $T$ by the partial type $p(x)$ in the sense that, for every $M\models T$ and every $a\in M$, \begin{align*} M\models P(a)\quad\text{exactly when}\quad M\models \varphi(a)\text{ for every }\varphi(x)\in p(x). \end{align*} Assume also that $p$ is non-principal over $T$. By the *Omitting Types Theorem*, there is a countable model $M\models T$ which omits $p$. Since $M$ omits $p$, for every $a\in M$ there is some formula $\varphi_a(x)\in p(x)$ such that \begin{align*} M\models \neg \varphi_a(a). \end{align*} Thus no $a\in M$ satisfies every formula in $p(x)$, and by the displayed equivalence no element of $M$ has property $P$. On the other hand, if $\Delta(x)=\{\varphi_1(x),\ldots,\varphi_k(x)\}$ is any finite subset of $p(x)$, then finite satisfiability of the partial type means \begin{align*} T\cup\{\exists x(\varphi_1(x)\wedge\cdots\wedge\varphi_k(x))\} \end{align*} is satisfiable. So every finite approximation to $P$ can still occur in some model of $T$, even though one countable model of $T$ omits the whole infinite description. The example shows that being supported by all finite pieces of an infinite type is weaker than being definable by one formula modulo $T$. [/example] A common application is the construction of models avoiding unwanted completions. The background theory fixes the finite behaviour, while omitted types remove entire possible complete descriptions. [example: Avoiding Non-Principal Completions] Let $T$ be a countable complete theory, and let $(p_i(x))_{i\in\mathbb N}$ be a countable list of non-principal complete types in $S_1(T)$. Since the language is countable and each $p_i$ is non-principal over $T$, the *Omitting Types Theorem* gives a countable model $M\models T$ such that $M$ omits every $p_i$. Unpacking the word “omits,” this means that for every $i\in\mathbb N$ and every element $a\in M$, the element $a$ does not realize $p_i(x)$. Equivalently, for each such pair $(i,a)$, there is some formula $\varphi_{i,a}(x)\in p_i(x)$ with \begin{align*} M\models \neg\varphi_{i,a}(a). \end{align*} Thus no realized complete $1$-type of $M$ is equal to any listed $p_i$: if $\operatorname{tp}^M(a)=p_i$, then $a$ would satisfy every formula in $p_i(x)$, contradicting the displayed formula. Therefore, if the types $p_i$ are exactly the unwanted complete extensions of some partial description, the model $M$ still satisfies every sentence of $T$, but its realized $1$-types avoid precisely those listed completions. The construction turns a countable list of forbidden non-principal behaviours into a countable model in which none of those behaviours occurs. [/example] This final example is less about a particular algebraic structure and more about a method. Model theorists often define a desired class of models by saying which types should be realized and which should be omitted; the omitting types theorem supplies the countable existence theorem for the omitted side. [remark: Relation to Saturation] Saturated models realize all types over small parameter sets, while omitting type arguments deliberately construct models that avoid selected types. The two techniques are complementary: saturation builds rich universal models, and omission builds sparse or special models that separate properties. [/remark] The chapter therefore completes the first cycle of the course. Types began as consistent descriptions, Stone spaces organized them topologically, saturation explained when many of them are realized, and omitting types now shows how non-isolated descriptions can be kept out of countable models. With omitting types in hand, the course now turns from what can be excluded to what must be present in especially well-behaved models. Atomicity and primeness translate the topological picture of types into algebraic structure, showing how models can be assembled from the simplest local descriptions. # 5. Atomic, Prime, and Countable Models This chapter turns the topology of type spaces into algebraic information about models. In the previous chapters, types measured which finite pieces of information can be made consistent and which complete descriptions are realized in saturated models. We now ask when a model can be built entirely from isolated types, when such a model is prime, and why countable categoricity is governed not by merely countably many types, but by finiteness of the type space in each finite arity. The guiding theme is that isolated types behave like named construction steps. If every finite tuple in a model realizes an isolated type, then elementary embeddings out of the model can be built one tuple at a time. This gives the model a minimality property among models of the same complete theory. ## Isolated Types as Building Blocks A natural way to build a small model is to insist that every tuple in it has a type determined by one formula. The problem is that most complete types are infinite objects, so a tuple whose type is not finitely pinned down carries extra choices that may prevent canonical embeddings into other models. Isolated types are the types for which this obstruction disappears. [definition: Isolated Type] Let $T$ be a complete first-order theory and let $p(x) \in S_n(T)$. The type $p$ is isolated if there is a formula $\varphi(x)$ such that $\varphi(x) \in p$ and, for every formula $\psi(x)$, either \begin{align*} T \models \forall x(\varphi(x) \to \psi(x)) \end{align*} when $\psi(x) \in p$, or \begin{align*} T \models \forall x(\varphi(x) \to \neg \psi(x)) \end{align*} when $\neg \psi(x) \in p$. [/definition] The formula $\varphi(x)$ is called an isolating formula for $p$. In the Stone space $S_n(T)$, this says that the singleton $\{p\}$ is open, because $[\varphi] = \{p\}$. Thus isolated types are exactly the isolated points of the type space. [example: Isolated Order Type In Dlo] Let $T=\operatorname{DLO}$. We show that the unique complete $1$-type over $\varnothing$ is isolated by $x=x$. By *quantifier elimination for DLO*, every parameter-free formula $\psi(x)$ is equivalent modulo $T$ to a quantifier-free formula in the single variable $x$. In the language $\{<\}$, every atomic parameter-free formula in one variable is one of \begin{align*} x=x,\qquad x<x. \end{align*} The theory $T$ proves \begin{align*} T &\models \forall x(x=x),\\ T &\models \forall x\neg(x<x), \end{align*} so every quantifier-free parameter-free formula in one variable is equivalent modulo $T$ to either $x=x$ or $x\ne x$. Let $p(x)\in S_1(T)$. Since every complete type contains the logical formula $x=x$, we have $x=x\in p$. If $\psi(x)\in p$, then $\psi(x)$ cannot be equivalent modulo $T$ to $x\ne x$, because then $p$ would contain an inconsistent formula. Hence \begin{align*} T\models \forall x((x=x)\to \psi(x)). \end{align*} If instead $\neg\psi(x)\in p$, the same argument applied to $\neg\psi(x)$ gives \begin{align*} T\models \forall x((x=x)\to \neg\psi(x)). \end{align*} Thus $x=x$ isolates $p$. Isolation here does not come from named elements; it comes from the fact that, over $\varnothing$, DLO has no nontrivial one-variable parameter-free distinctions. [/example] The preceding example is special because there is only one complete $1$-type over $\varnothing$. To use isolation as a construction method, we need a model-theoretic condition saying that not only single elements, but every finite tuple in the model, has its complete behavior pinned down by one formula. [definition: Atomic Model] Let $T$ be a complete first-order theory. A model $M \models T$ is atomic if, for every $n \ge 1$ and every tuple $a \in M^n$, the complete type $\operatorname{tp}^M(a/\varnothing) \in S_n(T)$ is isolated. [/definition] Atomicity is a global condition on all finite tuples. It does not merely say that individual elements have isolated $1$-types; it says that every finite configuration appearing in the model is syntactically determined by a single formula. [example: The Rational Order Is Atomic] Let $T=\operatorname{DLO}$ and let $M=(\mathbb Q,<)$. For $a=(a_1,\dots,a_n)\in \mathbb Q^n$, define its order-pattern formula by listing every equality and strict inequality true among its coordinates: \begin{align*} \theta_a(x_1,\dots,x_n) = \bigwedge_{a_i=a_j} x_i=x_j \wedge \bigwedge_{a_i<a_j} x_i<x_j. \end{align*} This formula is consistent with $T$, since $M\models \theta_a(a)$. We show that $\theta_a$ isolates $\operatorname{tp}^M(a/\varnothing)$. Let $\psi(x_1,\dots,x_n)$ be any parameter-free formula. By *quantifier elimination for DLO*, there is a quantifier-free formula $\chi(x_1,\dots,x_n)$ such that \begin{align*} T\models \forall x\bigl(\psi(x)\leftrightarrow \chi(x)\bigr). \end{align*} Every atomic parameter-free formula in the variables $x_1,\dots,x_n$ is of the form $x_i=x_j$ or $x_i<x_j$. The formula $\theta_a$ decides each such atomic formula: if $a_i=a_j$, then $\theta_a$ includes $x_i=x_j$; if $a_i<a_j$, then $\theta_a$ includes $x_i<x_j$; if $a_j<a_i$, then $\theta_a$ includes $x_j<x_i$, and the axioms of linear order give \begin{align*} T\models \forall x\bigl(\theta_a(x)\to \neg(x_i<x_j)\bigr). \end{align*} Thus $\theta_a$ fixes the truth value of every atomic formula, and therefore fixes the truth value of every Boolean combination of atomic formulas, including $\chi$. If $M\models \psi(a)$, then $M\models \chi(a)$, so every tuple satisfying $\theta_a$ satisfies $\chi$, and hence \begin{align*} T\models \forall x\bigl(\theta_a(x)\to \psi(x)\bigr). \end{align*} If $M\models \neg\psi(a)$, the same argument applied to $\neg\chi$ gives \begin{align*} T\models \forall x\bigl(\theta_a(x)\to \neg\psi(x)\bigr). \end{align*} Therefore $\theta_a$ isolates $\operatorname{tp}^M(a/\varnothing)$. For example, when $a_1<a_2=a_3$, the formula \begin{align*} x_1<x_2\wedge x_2=x_3 \end{align*} already determines the complete parameter-free type of $(a_1,a_2,a_3)$. Since every finite tuple in $\mathbb Q$ has such a finite order-pattern formula, $(\mathbb Q,<)$ is atomic. [/example] The example suggests that atomicity can be produced when every consistent finite demand can be refined to an isolated complete type. The main existence question is whether this local supply of isolated types is sufficient to build an entire countable model whose finite tuples all come from such refinements. [quotetheorem:5073] [citeproof:5073] The criterion explains why atomic models belong to countable model theory rather than arbitrary model theory. Countability is doing real work: it lets the construction list both the Henkin witness requirements and the finite tuples of constants whose types must be controlled. The theorem does not say that density of isolated types in uncountably many spaces can be met by the same sequential construction, and it does not produce an atomic model of an arbitrary prescribed cardinality. A useful boundary case is the theory of dense linear orders expanded by uncountably many constants naming a strictly increasing chain: even when finite pieces look individually manageable, an elementary construction must respect uncountably many named constraints, so the countable Henkin scheduling argument is no longer available in the same form. ## Prime Models and Uniqueness Atomic models matter because they are the models that can be embedded into any other model of the theory. The question is how a syntactic condition on realized types gives such a semantic universal property. [definition: Prime Model] Let $T$ be a complete theory. A model $M \models T$ is prime if, for every model $N \models T$, there is an elementary embedding $f:M \to N$. [/definition] A prime model is the smallest model of $T$ in the elementary-embeddability ordering. The embedding is not required to be onto; it says that every model of $T$ contains an elementary copy of $M$. [example: The Prime Model Of Dlo] Let $T=\operatorname{DLO}$. We show that $(\mathbb Q,<)$ is prime by building, for every $N\models T$, an elementary embedding $(\mathbb Q,<)\to N$. Enumerate $\mathbb Q$ as $q_0,q_1,q_2,\dots$. We define finite order-preserving injections \begin{align*} f_k:\{q_0,\dots,q_{k-1}\}\to N \end{align*} by induction. Suppose $f_k$ has been defined. Let \begin{align*} L&=\{f_k(q_i):i<k\text{ and }q_i<q_k\},\\ R&=\{f_k(q_i):i<k\text{ and }q_k<q_i\}. \end{align*} If both $L$ and $R$ are nonempty, let $\ell=\max L$ and $r=\min R$. Since $f_k$ preserves order, every element of $L$ is below every element of $R$, so $\ell<r$. By density in $N$, choose $c\in N$ with \begin{align*} \ell<c<r. \end{align*} If only $L$ is nonempty, choose $c>\max L$ using that $N$ has no right endpoint. If only $R$ is nonempty, choose $c<\min R$ using that $N$ has no left endpoint. If both are empty, choose any $c\in N$. Define \begin{align*} f_{k+1}(q_i)&=f_k(q_i)\quad(i<k),\\ f_{k+1}(q_k)&=c. \end{align*} In each case, for every $i<k$, \begin{align*} q_i<q_k &\Longrightarrow f_{k+1}(q_i)<f_{k+1}(q_k),\\ q_k<q_i &\Longrightarrow f_{k+1}(q_k)<f_{k+1}(q_i), \end{align*} so $f_{k+1}$ is again an order-preserving injection. Let \begin{align*} f=\bigcup_{k<\omega}f_k:\mathbb Q\to N. \end{align*} Then $f$ is an order-preserving injection. It remains to check elementarity. Let $\varphi(x_1,\dots,x_m)$ be any parameter-free formula and let $a=(a_1,\dots,a_m)\in\mathbb Q^m$. By *quantifier elimination for DLO*, there is a quantifier-free formula $\chi(x_1,\dots,x_m)$ such that \begin{align*} T\models \forall x\bigl(\varphi(x)\leftrightarrow \chi(x)\bigr). \end{align*} For each atomic formula $x_i=x_j$, \begin{align*} \mathbb Q\models a_i=a_j &\Longleftrightarrow a_i=a_j\\ &\Longleftrightarrow f(a_i)=f(a_j)\\ &\Longleftrightarrow N\models f(a_i)=f(a_j), \end{align*} because $f$ is a function and is injective. For each atomic formula $x_i<x_j$, \begin{align*} \mathbb Q\models a_i<a_j &\Longleftrightarrow a_i<a_j\\ &\Longleftrightarrow f(a_i)<f(a_j)\\ &\Longleftrightarrow N\models f(a_i)<f(a_j), \end{align*} because $f$ preserves order. Therefore $\mathbb Q$ and $N$ give the same truth value to every Boolean combination of these atomic formulas, in particular to $\chi$. Hence \begin{align*} \mathbb Q\models \varphi(a) &\Longleftrightarrow \mathbb Q\models \chi(a)\\ &\Longleftrightarrow N\models \chi(f(a))\\ &\Longleftrightarrow N\models \varphi(f(a)). \end{align*} Thus $f$ is elementary, so $(\mathbb Q,<)$ embeds elementarily into every model of $\operatorname{DLO}$ and is prime. [/example] The example is not an accident: the finite order pattern of a tuple in $\mathbb Q$ is isolated, so it can be reproduced inside any model of DLO. The next theorem abstracts this mechanism, showing that countable atomicity supplies enough finite control to build an elementary embedding into an arbitrary model. [quotetheorem:5074] [citeproof:5074] This proof is the atomic analogue of a back-and-forth construction, but only the forth direction is needed. The hypotheses are not cosmetic. Countability is what allows the embedding to be built by a sequence of finite stages; for an uncountable atomic model, the same proof would require a transfinite construction and additional control at limit stages. Atomicity is also essential: in algebraically closed fields, an algebraically closed subfield of positive transcendence degree realizes a non-isolated transcendental type over $\varnothing$, and it need not embed elementarily into the algebraic closure of the prime field. Thus being a model of the same complete theory is far weaker than being prime; the finite tuple types must be reproducible in every target model. Once prime models exist, the next structural question is whether there can be several different minimal countable cores of the same complete theory. [quotetheorem:5075] [citeproof:5075] The uniqueness theorem justifies speaking of the prime countable model of a complete countable theory when it exists, but it is a uniqueness statement, not an existence theorem. The atomic hypothesis is doing the work in the proof: it lets each finite partial elementary map be extended by a realized isolated type. Elementary embeddings in both directions by themselves would not be enough, since an elementary embedding may have a proper elementary image. Many complete theories have no prime model: for example, if a countable complete theory has a non-isolated type that must be realized in every model, the omitting types obstruction prevents any model from being atomic and therefore prevents a countable prime model. Conversely, when the prime model does exist, it need not be obtained from algebraic closure in a literal field-theoretic sense; DLO has prime model $(\mathbb Q,<)$ even though there is no algebraic closure operator generating it from $\varnothing$. In many algebraic examples the prime model is the definable or algebraic closure of the empty parameter set, but the definition only asks for elementary embeddability into every model. [remark: Atomic Versus Prime] For complete countable theories, [countable atomic models are prime](/theorems/5074), and prime models are atomic. The converse direction uses the fact that a non-isolated type realized in a prime model would be carried by an elementary embedding into every model of $T$, contradicting the omitting types theorem for countable languages. Thus, in the countable setting, prime model existence and atomic model existence are two forms of the same phenomenon. [/remark] The preceding equivalence connects this chapter back to omitting types. Non-isolated types are precisely the types that can be omitted by some countable model, so a model that embeds into every model cannot realize them. ## Countably Many Types Prime models describe the bottom of the model spectrum. The next question is when the entire countable model spectrum collapses to a single isomorphism type. Ryll-Nardzewski's theorem answers this by linking countable categoricity with finiteness of type spaces. [definition: Omega-Categorical Theory] A complete theory $T$ in a countable language is $\omega$-categorical if any two countable models of $T$ are isomorphic. [/definition] The definition concerns only countable models, but it has strong consequences for all finite tuple types over $\varnothing$. To state the corresponding [symmetry condition](/theorems/1360) inside the unique countable model, we record the finiteness property of its automorphism group on finite tuples. [definition: Oligomorphic Automorphism Group] Let $M$ be a countable structure. The automorphism group $\operatorname{Aut}(M)$ is oligomorphic if, for every $n \ge 1$, the componentwise action of $\operatorname{Aut}(M)$ on $M^n$ has finitely many orbits. [/definition] Oligomorphicity is the permutation-group shadow of $\omega$-categoricity. Orbits of tuples correspond to complete types over $\varnothing$ when the theory is sufficiently homogeneous, and Ryll-Nardzewski makes this precise. [quotetheorem:5076] [citeproof:5076] The theorem is the bridge promised by the chapter title. Having countably many types is not enough for categoricity; the decisive condition is that each finite arity has finitely many complete types. Each hypothesis has a role. Completeness is needed because otherwise the phrase "models of $T$" mixes different completions: the incomplete theory of infinite sets has countable models in several completions once additional structure or sentences are added, and type spaces over $T$ no longer describe one fixed elementary universe. The assumption of infinite models removes finite degeneracy: a theory saying "there are exactly $m$ elements" is categorical in every cardinal where it has a model, but it is not the countable Fraisse-style phenomenon measured by oligomorphic automorphism groups on an infinite countable model. Countability of the language is what lets the omitting-types and back-and-forth constructions operate over a countable list of formulas; in uncountable languages, one can name too much structure by predicates or constants, and finiteness conditions must be replaced by cardinal-sensitive versions. In a compact type space, finite means every type is isolated, so the unique countable model is atomic and saturated for finite arities in the appropriate countable sense. [example: Random Graph] Let $T_{RG}$ be the complete theory of the countable random graph in the language with one binary relation symbol $E$. We show that, for each $n\ge 1$, there are only finitely many complete $n$-types over $\varnothing$. By *quantifier elimination for the random graph*, every formula $\psi(x_1,\dots,x_n)$ is equivalent modulo $T_{RG}$ to a Boolean combination of atomic formulas of the forms \begin{align*} x_i=x_j,\qquad E(x_i,x_j) \end{align*} with $1\le i,j\le n$. A complete quantifier-free pattern first chooses which variables are equal. Equivalently, it chooses a partition $\pi$ of $\{1,\dots,n\}$. If $\pi$ has $k$ blocks, then after the equalities are fixed, the remaining graph information is the choice, for each unordered pair of distinct blocks, whether the corresponding edge relation holds or fails. There are \begin{align*} \binom{k}{2} \end{align*} such unordered pairs, so there are \begin{align*} 2^{\binom{k}{2}} \end{align*} possible edge patterns over that equality pattern. Since $\{1,\dots,n\}$ has only finitely many partitions, the total number of complete quantifier-free patterns is \begin{align*} \sum_{\pi\in \operatorname{Part}(n)} 2^{\binom{|\pi|}{2}}, \end{align*} which is finite. Each complete $n$-type is determined by one of these finitely many quantifier-free patterns, because quantifier elimination reduces every formula to a Boolean combination of the displayed atomic formulas. Hence $S_n(T_{RG})$ is finite for every $n\ge 1$. By *[Ryll-Nardzewski Theorem](/theorems/5076)*, $T_{RG}$ is $\omega$-categorical. [/example] This example shows how quantifier elimination often makes Ryll-Nardzewski practical. If every formula is equivalent to a quantifier-free formula, then counting $n$-types becomes a finite combinatorial problem whenever the language is finite and relational with finite arities. [example: Dense Linear Order] Let $T=\operatorname{DLO}$ and fix $n\ge 1$. We show that $S_n(T)$ is finite by describing exactly what a complete $n$-type over $\varnothing$ can decide. By *quantifier elimination for DLO*, every parameter-free formula $\psi(x_1,\dots,x_n)$ is equivalent modulo $T$ to a Boolean combination of atomic formulas of the forms \begin{align*} x_i=x_j,\qquad x_i<x_j \end{align*} with $1\le i,j\le n$. An equality-and-order pattern first partitions $\{1,\dots,n\}$ into blocks, where $i$ and $j$ lie in the same block exactly when $x_i=x_j$. If the partition has $k$ blocks, then the strict order between distinct blocks must be a linear order on those $k$ blocks. Thus, for this fixed partition, there are exactly $k!$ possible orders of the blocks. Since there are only finitely many partitions of $\{1,\dots,n\}$, the total number of possible equality-and-order patterns is \begin{align*} \sum_{\pi\in \operatorname{Part}(n)} |\pi|!, \end{align*} which is finite. For each such pattern $P$, form the quantifier-free formula $\theta_P(x_1,\dots,x_n)$ by conjoining all equalities $x_i=x_j$ true in the pattern and all strict inequalities $x_i<x_j$ true in the pattern. The axioms of linear order imply that $\theta_P$ also decides the negation of every atomic formula not true in $P$: for example, if $x_j<x_i$ is in the pattern, then \begin{align*} T\models \forall x\bigl(\theta_P(x)\to \neg(x_i<x_j)\bigr) \end{align*} by antisymmetry and irreflexivity of $<$. Therefore $\theta_P$ decides every Boolean combination of atomic formulas, and by quantifier elimination it decides every parameter-free formula in the variables $x_1,\dots,x_n$. Hence each complete $n$-type over $\varnothing$ is isolated by exactly one consistent equality-and-order pattern, so $S_n(T)$ is finite for every $n\ge 1$. By *Ryll-Nardzewski Theorem*, $\operatorname{DLO}$ is $\omega$-categorical. Since $(\mathbb Q,<)$ is a countable model of $\operatorname{DLO}$, it is the unique countable model up to isomorphism. [/example] DLO and the random graph are both homogeneous Fraisse-style examples: finite patterns determine complete types. The next example shows that completeness and quantifier elimination do not by themselves imply $\omega$-categoricity. [example: Algebraically Closed Fields Are Not Omega-Categorical] Fix a characteristic. If the characteristic is a prime $p$, let $K_0=\mathbb F_p$; if the characteristic is $0$, let $K_0=\mathbb Q$. For each $d\in \mathbb N\cup\{\aleph_0\}$, choose algebraically independent elements \begin{align*} t_1,t_2,\dots,t_d \end{align*} over $K_0$, where for $d<\aleph_0$ this list has length $d$, and set \begin{align*} K_d=\overline{K_0(t_1,\dots,t_d)}. \end{align*} The field $K_0(t_1,\dots,t_d)$ is countable, because it is built from countably many rational functions with coefficients in the countable field $K_0$. Its algebraic closure is also countable: for each $n\ge 1$, there are only countably many polynomials \begin{align*} a_nX^n+a_{n-1}X^{n-1}+\cdots+a_0 \end{align*} with coefficients in $K_0(t_1,\dots,t_d)$, and each nonzero polynomial has at most $n$ roots. Thus $K_d$ is a countable algebraically closed field. Each $K_d$ has the chosen characteristic and is algebraically closed, so $K_d\models \operatorname{ACF}_p$ in characteristic $p$, or $K_d\models \operatorname{ACF}_0$ in characteristic $0$. These are all models of the same complete theory, since algebraically closed fields of a fixed characteristic are complete. If $d\ne e$, then $K_d\not\cong K_e$. Indeed, any field isomorphism sends the prime subfield to the prime subfield and preserves polynomial equations with coefficients there. Hence a tuple is algebraically independent over $K_0$ exactly when its image is algebraically independent over $K_0$. Therefore an isomorphism preserves the maximum size of an algebraically independent set over $K_0$, namely transcendence degree. Since \begin{align*} \operatorname{trdeg}(K_d/K_0)=d \end{align*} and these values are different, the fields $K_d$ are pairwise non-isomorphic countable models of the fixed complete theory. Hence algebraically closed fields are not $\omega$-categorical. [/example] The failure can also be seen in type spaces. In algebraically closed fields, the distinction between algebraic dependence and transcendence creates infinitely many tuple types in higher arity, and countable models remember transcendence degree. ## Prime Models Inside Countably Categorical Theories The preceding sections leave a natural synthesis problem. Atomicity can give a canonical countable core, while $\omega$-categoricity says that no other countable model exists. We now ask when the finite-type condition from Ryll-Nardzewski forces the unique countable model to be prime. The Ryll-Nardzewski theorem gives the missing input for the atomic-model theorem: in an $\omega$-categorical complete countable theory, each finite type space is finite, so every realized finite type in the unique countable model is isolated. Thus the unique countable model is atomic. Combining this with the following theorem shows that it is prime. [quotetheorem:5074] [citeproof:5074] This explains why the random graph and $(\mathbb Q,<)$ are simultaneously highly homogeneous, atomic, prime, and countably categorical. The converse is false: a complete countable theory may have a prime countable model without being $\omega$-categorical. Algebraically closed fields of a fixed characteristic have a prime model, namely the algebraic closure of the prime field, but they have many non-isomorphic countable models distinguished by transcendence degree. Thus prime means there is a canonical smallest countable core, while $\omega$-categorical means there are no additional countable models beyond that core. The same finite-pattern phenomenon underlies all these features in the random graph and DLO, but prime existence alone does not force finite type spaces. [remark: Countability Of The Language] The countability hypothesis is used in the model-building arguments and in the formulation of $\omega$-categoricity. In uncountable languages, there are related versions using cardinal categoricity and bounds on type spaces, but the clean finite-type statement of Ryll-Nardzewski belongs to countable first-order languages. [/remark] The chapter's conclusion is a useful hierarchy. Isolated types are local syntactic building blocks. Atomic models are structures all of whose finite pieces come from such blocks. Prime models are the semantic expression of atomicity, and Ryll-Nardzewski says that when there are only finitely many finite blocks in each arity, the countable model is unique. Having seen how types control the construction of individual models, we next ask how many types there are to begin with. Stability answers this by counting types over parameter sets and identifying the theories in which the type space remains small enough to support classification. # 6. Stability: Counting Types The previous chapters built types as points of Stone spaces, used saturation to turn finitely satisfiable descriptions into realized elements in larger models, and used omitting types to keep selected non-principal descriptions out. Stability asks a counting question about these spaces: when the parameter set is large, how many complete 1-types can it support? The answer separates theories whose types are controlled by algebraic or linear data from theories that encode long orders and therefore produce many distinct cuts. ## Bounding Type Spaces Over Parameters The first problem is to make precise what it means for a theory to have few types. Since $S_1(A)$ records complete possible behaviours of a single new element over parameters $A$, a tame theory should not create more such behaviours than the number of parameters already present. [definition: Kappa Stability] Let $T$ be a complete first-order theory and let $\kappa$ be an infinite cardinal. The theory $T$ is $\kappa$-stable if for every model $M \models T$ and every set $A \subseteq M$ with $|A| = \kappa$, we have \begin{align*} |S_1(A)| \leq \kappa. \end{align*} [/definition] The restriction to parameter sets inside models is harmless for the uses in this course: by passing to a monster model, every small parameter set can be regarded as a subset of a sufficiently saturated ambient model. The definition asks only about complete 1-types, but this already detects the first dividing line between stable and unstable theories in the examples considered here. [remark: Why Count One-Types] For a fixed language $L$, the number of formulas with parameters from $A$ is at most $|A| + |L| + \aleph_0$. Hence the crude upper bound is \begin{align*} |S_1(A)| \leq 2^{|A| + |L| + \aleph_0}. \end{align*} Stability demands a much sharper bound: over a parameter set of size $\kappa$, there should be only $\kappa$ complete 1-types rather than all possible Boolean choices of formulas. [/remark] This remark explains why stability is a strong smallness condition. The ambient syntax allows exponentially many complete descriptions, so a stable theory must impose enough structure that most formal choices are inconsistent. [example: Pure Equality] Consider the complete theory of an infinite set in the language with equality only, and let $A$ be an infinite parameter set. For each $a\in A$, the formula $x=a$ determines one complete realised type over $A$: it contains $x=a$, contains $x\neq b$ for every $b\in A$ with $b\neq a$, and every equality-only formula with parameters from $A$ is decided by these equalities and inequalities. There is one further complete type, namely the non-realised type \begin{align*} p_\infty(x)=\{x\neq a : a\in A\}. \end{align*} Every finite subset of $p_\infty(x)$ has the form \begin{align*} \{x\neq a_1,\dots,x\neq a_n\}, \end{align*} and since the ambient model is infinite, it has an element distinct from the finitely many listed parameters $a_1,\dots,a_n$. Hence this partial type is finitely satisfiable, so it extends to a complete type over $A$. Thus the complete 1-types over $A$ are exactly \begin{align*} \{ \operatorname{tp}(a/A): a\in A\}\cup\{p_\infty\}. \end{align*} The first set has cardinality $|A|$, and $p_\infty$ is not realised by any $a\in A$ because it contains $x\neq a$ for every $a\in A$. Therefore \begin{align*} |S_1(A)|=|A|+1=|A| \end{align*} for every infinite $A$, so pure equality is $\kappa$-stable for every infinite cardinal $\kappa$. [/example] Pure equality is the baseline example: the only information a type can carry is whether the new element is already named. The next criterion packages this kind of reasoning into a reusable sufficient condition. [quotetheorem:5077] [citeproof:5077] The criterion is deliberately abstract, but its hypotheses matter. The lower bound $\kappa \geq |L|+\aleph_0$ prevents the language itself from producing more formula-schemes than the parameter set can control; without it, even a harmless expansion by many relation symbols can create too many syntactic distinctions over a small $A$. The injectivity requirement is also essential: assigning each type to some coarse invariant only proves stability when different complete types cannot collapse to the same datum. Thus the criterion is a counting device, not a classification theorem; it works after the course has already identified invariants fine enough to recover complete 1-types. In algebraically closed fields the invariant data will be algebraic dependence and a prime ideal over parameters, while in vector spaces it will be linear dependence or a new generic direction over the span of the parameters. ## Orders and the Production of Many Types The opposite problem is to understand how a theory can have too many types. The main mechanism is an interpretable order: a single formula can compare a potential element with many parameters, and each cut in the ordered parameter set gives a different complete type. [definition: Order Property] A formula $\varphi(x;y)$ has the order property in a complete theory $T$ if for every $n \in \mathbb N$ there are a model $M \models T$ and tuples $a_1,\dots,a_n$ and $b_1,\dots,b_n$ in $M$ such that \begin{align*} M \models \varphi(a_i;b_j) \quad \text{if and only if} \quad i < j. \end{align*} The theory $T$ has the order property if some formula has the order property. [/definition] The definition is finite because compactness converts arbitrarily long finite patterns into infinite indiscernible-looking ordered patterns in a sufficiently saturated extension. Once such a pattern exists, a type can choose where a new element sits relative to the parameter sequence. [quotetheorem:5078] [citeproof:5078] [example: Dense Linear Orders Are Unstable] Let $T$ be the theory of dense linear orders without endpoints in the language $\{<\}$, and view $A=\mathbb Q$ as a countable parameter set inside a sufficiently saturated model of $T$. For a Dedekind cut $(L,U)$ of $\mathbb Q$, with every element of $L$ below every element of $U$, consider the partial type \begin{align*} p_{(L,U)}(x) = \{a<x : a\in L\}\cup\{x<b : b\in U\}. \end{align*} To see that this partial type is consistent, take a finite subset \begin{align*} \{a_1<x,\dots,a_m<x\}\cup\{x<b_1,\dots,x<b_n\}, \end{align*} where each $a_i\in L$ and each $b_j\in U$. If $m,n>0$, let \begin{align*} a_0=\max\{a_1,\dots,a_m\}, \qquad b_0=\min\{b_1,\dots,b_n\}. \end{align*} Since $(L,U)$ is a cut, $a_0<b_0$, and by density there is $c\in\mathbb Q$ such that \begin{align*} a_0<c<b_0. \end{align*} Then $a_i\leq a_0<c$ for every $i$ and $c<b_0\leq b_j$ for every $j$, so $c$ realizes the finite subset. If only lower bounds occur, choose $c>a_0$ using that the order has no right endpoint; if only upper bounds occur, choose $c<b_0$ using that the order has no left endpoint. Thus every finite subset of $p_{(L,U)}(x)$ is satisfiable, so $p_{(L,U)}(x)$ extends to a complete 1-type over $\mathbb Q$. Different cuts give different complete types. Indeed, if $(L,U)\neq (L',U')$, choose $q\in L\triangle L'$. If $q\in L\setminus L'$, then $q\in U'$, so $p_{(L,U)}$ contains $q<x$ while $p_{(L',U')}$ contains $x<q$. These two formulas cannot both hold in a linear order. Hence distinct cuts produce distinct complete 1-types. There are $2^{\aleph_0}$ Dedekind cuts of $\mathbb Q$, while $\mathbb Q$ is countable, so \begin{align*} |S_1(\mathbb Q)|\geq 2^{\aleph_0}>\aleph_0. \end{align*} Since the set of formulas in one variable with parameters from $\mathbb Q$ is countable, there are at most $2^{\aleph_0}$ complete choices of truth values for those formulas, and therefore \begin{align*} |S_1(\mathbb Q)|=2^{\aleph_0}. \end{align*} Thus dense linear orders without endpoints are not $\aleph_0$-stable. [/example] Dense orders demonstrate the type explosion in its purest form. The same idea applies to any theory that can define enough order-like patterns, even when the language does not name an order relation outright. [remark: Instability Without a Named Order] The order property is syntactic rather than linguistic: the ordering may be encoded by a formula $\varphi(x;y)$ using the ambient structure. Thus a theory can be unstable without containing a symbol $<$ in its language. What matters is the ability to arrange parameters so that instances of one formula cut the realisations into a long ordered pattern. [/remark] ## Stable Algebraic and Linear Examples The guiding question on the stable side is whether complete types over parameters can be classified by algebraic data of size at most the parameter set. Algebraically closed fields and vector spaces show two common sources of stability: algebraic dependence and linear dependence. This control is not automatic for every algebraic-looking structure. If formulas can distinguish too many cuts, valuations, orderings, or unrelated predicates over the same parameter set, then the generated closure may be too coarse to determine complete types. The examples below work because quantifier elimination reduces formulas to equations whose solution behaviour is governed by a small closure object. [quotetheorem:5079] [citeproof:5079] This proof uses a theorem from earlier algebraic model theory: quantifier elimination reduces arbitrary formulas to Boolean combinations of polynomial equations. Algebraic closedness is essential for this clean dichotomy, because irreducible polynomial data over the generated field then describes all algebraic possibilities without leaving unresolved extension information. Completeness is handled by fixing the characteristic $p$; mixing characteristics would not define a single complete theory and would not give one uniform type-counting argument. The result should not be read as saying that every field theory is stable: extra definable order or valuation structure can create distinctions not controlled by polynomial dependence alone. In algebraic-geometric language, 1-types over $A$ are governed by prime ideals in $k[x]$, and for one variable those prime ideals are either maximal algebraic choices or the zero ideal giving the generic type. [example: Types Over an Algebraically Closed Subfield] Let $K \models \operatorname{ACF}_0$ and let $A\subseteq K$ be an algebraically closed subfield. We compute the complete 1-types over $A$ using *quantifier elimination for algebraically closed fields*: a 1-type over $A$ is determined by the polynomial equations and inequations with coefficients in $A$ that it contains. If $c\in K$ is algebraic over $A$, let $m_c(X)\in A[X]$ be its minimal polynomial. Since $A$ is algebraically closed, every nonconstant polynomial in $A[X]$ splits into linear factors over $A$, so \begin{align*} m_c(X)=(X-a_1)\cdots(X-a_n) \end{align*} for some $a_1,\dots,a_n\in A$. Because $m_c$ is irreducible in $A[X]$, exactly one linear factor can occur, so $m_c(X)=X-a$ for some $a\in A$. Hence $m_c(c)=0$ gives \begin{align*} c-a=0, \end{align*} so $c=a$. Thus every algebraic type over $A$ is already realised by some $a\in A$, namely the type containing $x=a$. There is no additional type coming from an irreducible polynomial of degree greater than $1$: if $f\in A[X]$ were irreducible and $\deg(f)>1$, algebraic closedness of $A$ would factor $f$ as a product of linear factors in $A[X]$, contradicting irreducibility. If $c$ is transcendental over $A$, then for every nonzero $f\in A[X]$ we have \begin{align*} f(c)\neq 0, \end{align*} because $f(c)=0$ would make $c$ algebraic over $A$. Conversely, any two transcendental elements over $A$ satisfy exactly the same polynomial equations over $A$: only $0=0$ holds, and every nonzero polynomial equation fails. By quantifier elimination, they therefore have the same complete type over $A$. So $S_1(A)$ consists exactly of the realised types $x=a$ for $a\in A$, together with one generic transcendental type \begin{align*} \{f(x)\neq 0 : 0\neq f\in A[X]\}. \end{align*} The algebraically closed subfield leaves no new algebraic choices; the only element not already named by $A$ is a generic transcendental one. [/example] Algebraically closed fields hide their stability inside quantifier elimination. Vector spaces make the mechanism more visible because formulas can only test finite linear dependence data. The possible obstruction to stability is that parameters might create many distinct positions for a new vector. In the pure vector-space language, however, every vector is either already in the span of the parameters or is independent from that span, so the analogue of algebraic closure is ordinary linear span. [quotetheorem:5080] [citeproof:5080] The vector-space case mirrors pure equality with equality replaced by membership in the span of parameters. The fixed-field hypothesis is important: the language names scalar multiplication by elements of $F$, so the size of $F$ contributes directly to the number of linear equations available. The infinite-dimensional hypothesis keeps the independent type consistent over every small span; in a finite-dimensional space over the parameter span, there may be no room for a new independent vector. This theorem is also limited to the pure module language. If the [vector space](/page/Vector%20Space) is expanded by a definable order, a [bilinear form](/page/Bilinear%20Form) with complicated extra predicates, or unrelated structure on a basis, then linear span need not determine complete types. The forward lesson is that stability is preserved here because formulas reduce to finite linear conditions, just as algebraically closed fields reduce formulas to finite polynomial conditions. [example: Infinite-Dimensional Vector Spaces] Let $F$ be countable and let $V$ be an infinite-dimensional $F$-vector space in the language of $F$-modules. If $A\subseteq V$ is countable, then every element of \begin{align*} W=\operatorname{span}_F(A) \end{align*} has the form \begin{align*} \lambda_1a_1+\cdots+\lambda_na_n \end{align*} for some $n\in\mathbb N$, some $a_1,\dots,a_n\in A$, and some $\lambda_1,\dots,\lambda_n\in F$. For each fixed $n$, the set of such expressions is contained in \begin{align*} F^n\times A^n, \end{align*} which is countable because both $F$ and $A$ are countable. Since \begin{align*} W=\bigcup_{n\in\mathbb N}\{\lambda_1a_1+\cdots+\lambda_na_n:\lambda_i\in F,\ a_i\in A\}, \end{align*} the subspace $W$ is a [countable union of countable sets](/theorems/755), hence countable. In the pure language of $F$-vector spaces, formulas in one variable over $A$ reduce to finite Boolean combinations of affine linear equations over $W$. If $v\in W$, then its complete type over $A$ is the realised type determined by the equation \begin{align*} x=v. \end{align*} Thus the realised types over $A$ are indexed by the countable set $W$. There is one further possibility: $x$ may lie outside $W$. The corresponding type contains \begin{align*} x\neq w \end{align*} for every $w\in W$. It is consistent because $V$ is infinite-dimensional, while $W$ is generated by the countable set $A$, so in a sufficiently large model there is a vector not in $W$. Any two vectors outside $W$ satisfy the same affine linear equations over $W$: if \begin{align*} \lambda x+w=0 \end{align*} with $\lambda\in F$ and $w\in W$, then for $\lambda=0$ the equation is either always true or always false according as $w=0$ or $w\neq 0$, while for $\lambda\neq 0$ it is equivalent to \begin{align*} x=-\lambda^{-1}w\in W, \end{align*} which fails for every $x\notin W$. Hence all vectors outside $W$ realise the same independent type over $A$. Therefore \begin{align*} S_1(A)=\{\operatorname{tp}(w/A):w\in W\}\cup\{p_{\mathrm{ind}}\}, \end{align*} where $p_{\mathrm{ind}}$ is the type of a vector outside $W$. Since $W$ is countable, this gives \begin{align*} |S_1(A)|\leq |W|+1=\aleph_0. \end{align*} Thus the theory of infinite-dimensional vector spaces over a countable field is $\aleph_0$-stable. [/example] These examples explain why type-counting is a useful first definition of stability. In an ordered theory, parameters can define many cuts; in algebraic and linear theories, parameters generate a closure object of comparable size, and a 1-type is controlled by its relation to that closure. ## The Counting Dividing Line The central comparison in this chapter is between cuts and closures. Orders turn a parameter set into many possible locations for a new element, while stable structures force a new element to be either controlled by a small closure operation or generic over it. [quotetheorem:5081] [citeproof:5081] This summary theorem is deliberately limited to the examples proved in the chapter. It is not the full stability theorem, nor a classification of all stable and unstable theories: many theories require more delicate invariants than algebraic closure, linear span, or visible order cuts. The hypotheses used above also differ from example to example, with quantifier elimination doing the work for algebraically closed fields and linear elimination doing the work for vector spaces. Later chapters refine this first counting test using Morley rank, indiscernibles, forking, and categoricity, but the initial signal is already visible in $S_1(A)$: stable theories have few types over large parameter sets, and unstable theories manufacture many types by encoding order. The first counting test for stability naturally leads to a finer analysis of how types depend on parameters. Definability of types and forking shift the focus from the sheer number of types to the way they vary, giving a first geometric notion of independence. # 7. Definability of Types and Forking Preview This chapter begins the transition from the topology of type spaces to the geometry of dependence. In the preceding chapters, types were treated as points of compact spaces and saturated models supplied enough realisations to make those spaces usable. We now ask when a type varies in a controlled way over a model, and how this control anticipates the independence relation that appears in stable theories. The guiding theme is that a formula $\varphi(x;y)$ can be read as a [test function](/page/Test%20Function) on a type space: a type $p(x)$ decides, for each parameter $b$, whether $\varphi(x;b)$ belongs to $p$. In stable theories these decisions are themselves definable in the parameter $b$. This definability is the first sign that stable type spaces behave less like arbitrary Stone spaces and more like algebraic or linear geometric objects. ## Formulas as Functions on Type Spaces The first problem is to turn syntax into geometry without losing the logical content. A complete type is a maximal consistent description, so every formula with parameters cuts out the set of types containing it. The question is how much of the type can be recovered from these cuts, and how formulas should be regarded as functions on $S_x(A)$. [definition: Basic Clopen Set in a Type Space] Let $T$ be a complete first-order theory, let $A$ be a parameter set, and let $\varphi(x)$ be an $A$-formula. The basic clopen set determined by $\varphi$ is \begin{align*} [\varphi] := \{p \in S_x(A) : \varphi(x) \in p\}. \end{align*} [/definition] These sets were the basis for the Stone topology in Chapter 2, so they already describe the open regions of type space. To compare types using a single formula at a time, it is useful to encode each clopen set as a two-valued observable. [definition: Formula Function] Let $T$ be a complete first-order theory, let $A$ be a parameter set, and let $\varphi(x)$ be an $A$-formula. The formula function associated to $\varphi$ is the map \begin{align*} f_\varphi : S_x(A) \to \{0,1\}, \qquad f_\varphi(p)=1 \iff \varphi(x) \in p, \end{align*} where $\{0,1\}$ has the discrete topology. [/definition] The definition is useful because continuity becomes a reformulation of the fact that membership in a formula is locally constant on type space. Before definability can be discussed, we need this topological translation: formulas do not define arbitrary functions on $S_x(A)$, but locally constant ones. [quotetheorem:5082] [citeproof:5082] This result says that formulas are continuous observables on the type space, and it uses exactly the hypotheses built into the Stone topology: completeness of types ensures that each type chooses one of $\varphi$ and $\neg\varphi$, while the topology declares the corresponding basic sets to be clopen. If completeness is replaced by partial types, the same statement can fail: a partial type may contain neither $\varphi$ nor $\neg\varphi$, so its characteristic function for membership in $\varphi$ is no longer a total two-valued observable on complete decisions. If the topology were weakened so that basic formula sets were not open, the proof would also fail at the preimage step. Conversely, every continuous map $S_x(A)\to\{0,1\}$ has clopen fibre over $1$, and compactness of the Stone topology makes that clopen set a finite Boolean combination of basic clopens; after taking the corresponding Boolean combination of formulas, it is again represented by a single $A$-formula. This topological fact is not yet definability in parameters: in a dense linear order, a cut type may give a continuous test on $S_1(M)$ while its associated left cut in $M$ is not first-order definable over $M$. The next question is therefore different: if the object variable $x$ is fixed and the parameter variable $y$ ranges through a model, can the truth pattern of $\varphi(x;y)$ along a type $p(x)$ be described by a formula in $y$? [example: Dense Linear Orders and Cut Tests] In the theory of dense linear orders without endpoints, let $M\models T$ and let $p(x)\in S_1(M)$ describe a cut. Write \begin{align*} L_p&:=\{b\in M: (b<x)\in p\},\\ R_p&:=\{b\in M: (x<b)\in p\}. \end{align*} For the formula $\varphi(x;y)$ given by $y<x$, the parameter set selected by $p$ is \begin{align*} \{b\in M:\varphi(x;b)\in p\} &=\{b\in M:(b<x)\in p\}\\ &=L_p. \end{align*} Thus the truth pattern of the single test formula $y<x$ recovers the left side of the cut: parameters below the cut are exactly those for which the instance $b<x$ belongs to the type. The definability question is therefore whether this subset $L_p\subseteq M$ is cut out inside $M$ by some formula with parameters from $M$. [/example] The example highlights the central obstruction. A type can determine a subset of a model for each formula, but arbitrary subsets need not be definable. Stable theories are those in which this obstruction disappears for complete types over models. ## Definable Types over Models The basic problem is to decide when a complete type over a model is not merely a set of formulas, but a uniformly definable rule for answering all parameter questions. For each partitioned formula $\varphi(x;y)$, the type tells us which parameters $b$ satisfy $\varphi(x;b)\in p$. Definability asks for that set of parameters to be cut out by a formula over the same base model. [definition: Definability of a Type for a Formula] Let $M\models T$, let $p(x)\in S_x(M)$, and let $\varphi(x;y)$ be a parameter-free partitioned formula; formulas with parameters from $M$ are handled by adding those parameters to the tuple $y$. The type $p$ is definable for $\varphi$ over $M$ if there is an $M$-formula $d_p\varphi(y)$ such that for every $b\in M^{|y|}$, \begin{align*} \varphi(x;b)\in p \iff M\models d_p\varphi(b). \end{align*} [/definition] The notation $d_p\varphi$ reminds us that the defining formula depends on the type $p$ and on the test formula $\varphi$. Since later arguments must extend a whole type rather than one formula at a time, we now require such definitions simultaneously for every test formula. [definition: Definable Type] Let $M\models T$ and let $p(x)\in S_x(M)$. The type $p$ is definable over $M$ if $p$ is definable for every parameter-free partitioned formula $\varphi(x;y)$, allowing parameters from $M$ to be included among the $y$-variables. [/definition] Equivalently, after naming parameters from $M$, definability is a coherent scheme $\varphi\mapsto d_p\varphi$ for all partitioned formulas. Coherence means that the definitions commute with Boolean combinations, for example $d_p(\neg\varphi)$ is equivalent over elementary extensions of $M$ to $\neg d_p\varphi$, and $d_p(\varphi\wedge\psi)$ is equivalent to $d_p\varphi\wedge d_p\psi$ when the parameter variables are paired. This coherence is the part of the definition that lets the scheme be evaluated in the monster model, rather than only on tuples from $M$. Definability is stronger than completeness. Completeness decides every instance $\varphi(x;b)$ with $b\in M$, while definability says that those decisions vary definably as $b$ varies. The natural question is whether stability forces this stronger property, or whether non-definable truth patterns can still occur over models. [example: Principal Types Are Definable] Let $M\models T$, and suppose $p(x)\in S_x(M)$ is isolated by an $M$-formula $\theta(x)$, meaning that $p$ is the unique complete type over $M$ containing $\theta(x)$. For a partitioned formula $\varphi(x;y)$, set \begin{align*} d_p\varphi(y) := \forall x(\theta(x)\to\varphi(x;y)). \end{align*} We show that this formula defines the $\varphi$-decisions of $p$ over $M$. Fix $b\in M^{|y|}$. Since $M\preceq\mathfrak C$ and $d_p\varphi(b)$ has parameters from $M$, \begin{align*} M\models d_p\varphi(b) &\iff \mathfrak C\models \forall x(\theta(x)\to\varphi(x;b))\\ &\iff \text{for every }a\in\mathfrak C^{|x|},\ \mathfrak C\models\theta(a)\text{ implies }\mathfrak C\models\varphi(a;b). \end{align*} If $a\models\theta(x)$, then $\operatorname{tp}(a/M)$ contains $\theta(x)$, so by isolation $\operatorname{tp}(a/M)=p$. Therefore the last condition is equivalent to saying that every realisation of $p$ satisfies $\varphi(x;b)$, which is equivalent to $\varphi(x;b)\in p$. Thus \begin{align*} \varphi(x;b)\in p \iff M\models d_p\varphi(b), \end{align*} so isolated types are definable, with the defining scheme obtained from the single isolating formula $\theta(x)$. [/example] The example is useful but too special, since stability is about non-isolated types as well. It reduces definability to a single isolating formula, while stable theories must control types that have no such finite description. The next theorem is needed because it supplies that control from the absence of the order property. [quotetheorem:5083] [citeproof:5083] This theorem is one of the first places where stability converts a negative combinatorial condition into a positive structural theorem. The stability hypothesis is essential: in dense linear orders without endpoints, a complete type over a model can describe a non-definable cut, and for the formula $y<x$ the set of parameters on the left side of the cut need not be definable in the model. The theorem also has a precise limitation: definability is obtained formula-by-formula, through the scheme $\varphi\mapsto d_p\varphi$, and it does not imply that $p$ is isolated by one formula. What it gives is that the map $b\mapsto \mathbb{1}_{\varphi(x;b)\in p}$ is not just externally visible from the type space; for each fixed test formula it is internally definable in the model. [example: Generic Type of an Irreducible Algebraic Variety] Work in $T=\operatorname{ACF}_k$, let $M\models T$, and let $V\subseteq M^n$ be irreducible with ideal $I(V)\subseteq M[X_1,\dots,X_n]$. The generic type $p_V(x)$ contains $x\in V$ and, for every proper Zariski closed $W\subsetneq V$ defined over $M$, contains $x\notin W$. We compute the parameter set defined by one formula $\varphi(x;y)$. Here $V$ is a Zariski [closed set](/page/Closed%20Set) that cannot be written as the union of two proper Zariski closed subsets. Its ideal is \begin{align*} I(V):=\{f\in M[X_1,\dots,X_n]: f(v)=0\text{ for every }v\in V\}. \end{align*} For a polynomial $g$, the notation $Z(g)$ denotes its zero set on the relevant affine space, and a constructible subset of $V$ is a finite Boolean combination of such polynomial equalities and inequalities inside $V$. When a polynomial $f(x;y)$ has additional parameter variables $y$ and $b\in M^{|y|}$ is fixed, the expression $f(x;b)\in I(V)$ means that the resulting polynomial in the $x$-variables vanishes on all of $V$. By *Quantifier Elimination for Algebraically Closed Fields*, after replacing $\varphi(x;y)$ by an equivalent quantifier-free formula, we may write it as a finite disjunction of conjunctions \begin{align*} \varphi(x;y)\equiv \bigvee_{i=1}^r \left( \bigwedge_{j=1}^{s_i} f_{ij}(x;y)=0 \ \wedge\ \bigwedge_{\ell=1}^{t_i} h_{i\ell}(x;y)\neq 0 \right), \end{align*} where all $f_{ij},h_{i\ell}$ are polynomials over $k$. For $b\in M^{|y|}$, the $i$-th disjunct cuts out inside $V$ the constructible set \begin{align*} C_i(b)=V\cap \bigcap_{j=1}^{s_i} Z(f_{ij}(x;b)) \cap \bigcap_{\ell=1}^{t_i} \left(V\setminus Z(h_{i\ell}(x;b))\right). \end{align*} Because $V$ is irreducible, a finite union is dense in $V$ exactly when one member is dense in $V$. Thus $\varphi(x;b)\in p_V$ holds exactly when some $C_i(b)$ is Zariski dense in $V$. For a fixed $i$, put \begin{align*} H_i(x;y)=\prod_{\ell=1}^{t_i}h_{i\ell}(x;y), \end{align*} with $H_i=1$ if $t_i=0$. The set $C_i(b)$ is dense in $V$ precisely when \begin{align*} f_{ij}(x;b)\in I(V)\quad\text{for every }j \end{align*} and \begin{align*} H_i(x;b)\notin I(V). \end{align*} Indeed, the first condition says every equation in the disjunct holds on all of $V$; the second says the open subset $V\setminus Z(H_i(x;b))$ is nonempty, hence dense in irreducible $V$. Membership in the fixed ideal $I(V)$ is a finite system of polynomial conditions on the coefficients, hence on $b$, and non-membership is the negation of such a system. Therefore the defining formula for the $\varphi$-decisions of $p_V$ is the finite disjunction over $i$ of these conditions: \begin{align*} d_{p_V}\varphi(y) := \bigvee_{i=1}^r \left( \bigwedge_{j=1}^{s_i} f_{ij}(x;y)\in I(V) \ \wedge\ H_i(x;y)\notin I(V) \right). \end{align*} So the generic type is definable: it answers each family $\varphi(x;y)$ by asking whether the corresponding constructible subset of $V$ contains a dense open part of $V$. [/example] The algebraic example gives the intended geometry behind definable types. The generic point of an irreducible variety is not usually in the model, but its incidence with definable families is controlled by formulas over the model. In practice, verifying definability of a type starts with a single test formula $\varphi(x;y)$. First identify, for $b\in M^{|y|}$, the condition under which $\varphi(x;b)$ belongs to the type. Then rewrite that condition using only the parameter variable $y$ and parameters from $M$. Finally check Boolean coherence: the proposed definitions should respect negation and conjunction, since the global extension construction will evaluate the whole scheme in elementary extensions. For isolated types this procedure produces $\forall x(\theta(x)\to\varphi(x;y))$; for algebraic generic types it produces a density condition in a definable family. ## Heirs, Coheirs, and Invariant Extensions The next problem is extension. Given $p(x)\in S_x(M)$ and a larger parameter set $B\supseteq M$, compactness gives extensions of $p$ to $S_x(B)$, but it does not distinguish the extensions that preserve the structure of $p$. Heirs, coheirs, and invariant extensions are three elementary ways to express that preservation. [definition: Heir] Let $M\subseteq B$ be parameter sets and let $q(x)\in S_x(B)$ extend $p(x)\in S_x(M)$. The type $q$ is an heir of $p$ over $B$ if, for every formula $\varphi(x;y;z)$, every tuple $b\in B^{|y|}$, and every tuple $m\in M^{|z|}$, whenever $\varphi(x;b;m)\in q$, there exists $a\in M^{|y|}$ such that \begin{align*} \varphi(x;a;m)\in p. \end{align*} [/definition] The heir condition says that every formula chosen by the extension is already foreshadowed over the base model. This controls formulas by pulling their parameter behaviour back to $M$. A second kind of control goes in the object direction: instead of replacing new parameters by old ones, ask whether finite pieces of the extension are realised in the base. [definition: Coheir] Let $M\subseteq B$ be parameter sets and let $q(x)\in S_x(B)$ extend $p(x)\in S_x(M)$. The type $q$ is a coheir of $p$ over $B$ if every finite subset of $q$ is realised in $M$. [/definition] Coheirs are governed by finite satisfiability rather than definability. They are particularly concrete: every finite demand made by the extension can already be met by an element of the base model. [example: A Coheir from a Dense Order] Let $M$ be a dense linear order without endpoints, let $B\supseteq M$, and let $q(x)\in S_1(B)$ determine a cut over $B$. Write \begin{align*} L_q&:=\{b\in B:(b<x)\in q\},\\ R_q&:=\{b\in B:(x<b)\in q\}. \end{align*} Assume this cut is approximated by $M$: whenever $l_1,\dots,l_m\in L_q$ and $r_1,\dots,r_n\in R_q$, there is some $c\in M$ such that \begin{align*} l_i<c\quad\text{for all }1\leq i\leq m, \qquad c<r_j\quad\text{for all }1\leq j\leq n. \end{align*} We show that every finite subset of $q$ is realised in $M$. Let $\Delta(x)\subseteq q$ be finite. In one variable in a dense linear order, each formula over $B$ is equivalent over the theory of dense linear orders to a Boolean combination of inequalities $x<b$, $b<x$, and equalities $x=b$ with $b\in B$. Since $q$ is a cut type, no equality $x=b$ belongs to $q$. Thus the finite information in $\Delta$ is implied by finitely many inequalities \begin{align*} l_1<x,\dots,l_m<x,\qquad x<r_1,\dots,x<r_n \end{align*} with each $l_i\in L_q$ and each $r_j\in R_q$. By the approximation assumption, choose $c\in M$ with \begin{align*} l_i<c\quad\text{for all }i, \qquad c<r_j\quad\text{for all }j. \end{align*} Then $c$ satisfies all these inequalities, hence satisfies every formula in $\Delta(x)$. Therefore every finite subset of $q$ is realised in $M$, so $q$ is a coheir over $M$. This example shows that coheir extensions record finite approximation by elements of the base model. [/example] Heirs and coheirs are not the same definition, but over saturated models they often coincide with a more conceptual condition. The preceding notions are phrased by formulas and finite realisations; the next one says directly that the extension cannot distinguish tuples with the same structure over the base. This automorphism formulation is the version that will connect most directly to forking. [definition: Invariant Type] Let $\mathfrak C\models T$ be a monster model, let $A\subseteq \mathfrak C$, and let $q(x)\in S_x(B)$ for some $B\supseteq A$. The type $q$ is $A$-invariant if for every automorphism $\sigma\in\operatorname{Aut}(\mathfrak C/A)$ and every formula $\varphi(x;b)$ with $b\in B$ and $\sigma(b)\in B$, \begin{align*} \varphi(x;b)\in q \iff \varphi(x;\sigma(b))\in q. \end{align*} [/definition] The clean standard form is the global one: a global type $q\in S_x(\mathfrak C)$ is $A$-invariant when the displayed equivalence holds for all parameters $b\in\mathfrak C$. The relative formulation above is the restriction of that notion to a parameter domain $B$ that may not be fixed setwise by every automorphism over $A$. Invariance packages the idea that a type depends only on information visible over the base. To extend a type from $M$ to the monster, one must decide formulas $\varphi(x;b)$ with external parameters $b\in\mathfrak C$; the obstruction is that arbitrary choices may depend on accidental representatives of $\operatorname{tp}(b/M)$ rather than on $M$-information. A definable type avoids this by providing, for each $\varphi$, an $M$-definable test $d_p\varphi(y)$ that decides membership uniformly for all parameters $b$. That uniform decision rule is exactly what should produce an $M$-invariant global extension. [quotetheorem:5084] [citeproof:5084] The construction is canonical once the definitions $d_p\varphi$ are fixed, and this is exactly where definability over $M$ is used: automorphisms fixing $M$ also fix the formulas $d_p\varphi(y)$, so they preserve membership in the global extension. If $p$ is not definable, the same recipe has no formula deciding which external parameters $b$ should satisfy $\varphi(x;b)\in q$; for instance, a non-definable cut in a dense linear order has no $M$-formula selecting its left side. Coherence is also a real hypothesis, not bookkeeping: without compatibility with negation and conjunction, the declaration could choose $\varphi(x;b)$ and $\neg\varphi(x;b)$ simultaneously, or choose several formulas whose conjunction has not been declared consistently. The theorem also does not assert uniqueness of global extensions, nor does it say that the extension is finitely satisfiable in $M$. It only supplies an invariant extension built from a defining scheme, which is the kind of controlled extension needed in stable theories. To recognize the invariant extension in computations, use the defining scheme as a decision table. Given two parameter tuples $b,c\in\mathfrak C$ with the same type over $M$, an automorphism over $M$ sends $b$ to $c$ in the monster model. Since $d_p\varphi$ has parameters from $M$, the truth values of $d_p\varphi(b)$ and $d_p\varphi(c)$ agree, so the extension makes the same decision about $\varphi(x;b)$ and $\varphi(x;c)$. This is the practical test for invariance: membership in the extension depends on $\operatorname{tp}(b/M)$, not on the particular representative $b$. ## Extension over Saturated Models The extension question becomes cleaner when the base model is saturated. A $\kappa$-saturated model realises every complete type over each subset of the model of size less than $\kappa$; with the parameter domains below restricted to $M\cup A$ for $|A|<\kappa$, this is the precise sense in which the base contains enough realisations for small configurations. Under these size bounds, finite satisfiability and definability interact more tightly. The goal is to state the extension principle in the form used later for forking. [definition: Saturated Base for a Type] Let $\kappa$ be an infinite cardinal. A model $M\models T$ is a saturated base for types over sets of size less than $\kappa$ if $M$ is $\kappa$-saturated and every larger parameter domain is of the form $M\cup A$ with $|A|<\kappa$. [/definition] This terminology records the size assumptions needed when extending types to larger parameter sets. The preceding invariant-extension theorem gives a global extension when a type is already definable, but for stable theories we want a uniform statement that starts only with an arbitrary type over a saturated model. The next theorem is needed to package definability, saturation, and restriction to a larger parameter set into the extension principle used later for independence. [quotetheorem:5085] [citeproof:5085] This extension theorem is the technical reason stable theories admit a robust independence relation. Stability is essential because, in an unstable theory such as dense linear orders, non-definable cuts can extend over new parameters in ways that depend on the new order information rather than only on the base. The saturation hypothesis should also be read as a base-size condition for the later calculus: without enough saturation, small types over $M$ need not be represented inside $M$, so heirs and coheirs need not line up as cleanly. The conclusion is still limited: it gives at least one controlled extension to $B$, not a unique extension and not automatically finite satisfiability in $M$. What matters for the next chapter is that types can be moved to larger parameter sets without acquiring uncontrolled dependence on the new parameters. [example: Vector Spaces and Linear Independence] Let $T$ be the theory of infinite-dimensional vector spaces over a fixed field $F$, let $M\models T$, and let $B\supseteq M$. Consider the type $p(x)\in S_1(M)$ saying that $x$ is not in the span of any finite tuple from $M$: \begin{align*} p(x)\supseteq \left\{ x\neq \lambda_1m_1+\cdots+\lambda_rm_r: r<\omega,\ m_i\in M,\ \lambda_i\in F \right\}. \end{align*} For a finite tuple $b_1,\dots,b_n\in B$, the corresponding nonforking extension requires \begin{align*} x\notin \operatorname{Span}_F(M,b_1,\dots,b_n). \end{align*} Written out, this means that for every $m\in M$ and every $\lambda_1,\dots,\lambda_n\in F$, \begin{align*} x\neq m+\lambda_1b_1+\cdots+\lambda_nb_n. \end{align*} If this requirement fails for a realisation $v$, then for some $m\in M$ and $\lambda_i\in F$ we have \begin{align*} v=m+\lambda_1b_1+\cdots+\lambda_nb_n, \end{align*} so \begin{align*} v-\lambda_1b_1-\cdots-\lambda_nb_n=m\in M. \end{align*} Thus the new parameters have imposed a linear dependence of $v$ over the base subspace $M$. Conversely, if such an equation holds, then $v\in\operatorname{Span}_F(M,b_1,\dots,b_n)$, so the extension has acquired exactly that new dependence. Hence, in vector spaces, the model-theoretic independence relation is ordinary [linear independence](/page/Linear%20Independence) over $M$: adding parameters forks precisely when they create a new linear equation over the base. [/example] The vector-space case is the cleanest model for what the abstract theory tries to imitate. Independence is extension without new algebraic constraint over the base. ## Forking as a Preview of Dependence The final problem of the chapter is conceptual rather than technical. We want an independence relation, written provisionally as $a\ind_M B$, saying that the type of $a$ over $B$ is a good extension of its type over $M$. In stable theories, definability and invariant extension supply the first approximation to what “good” means. [definition: Nonforking Extension Preview] Let $T$ be stable, let $M\models T$, let $B\supseteq M$, and let $q(x)\in S_x(B)$ extend $p(x)\in S_x(M)$. In this preview, $q$ is called a nonforking extension of $p$ over $M$ if it is the restriction to $B$ of an $M$-invariant global extension of $p$. [/definition] This is not yet the full definition of forking, because the general theory uses dividing, finite character, symmetry, and transitivity. It is the right provisional definition for the present course stage because it captures the extension intuition developed in this chapter. [remark: What the Preview Omits] The full forking calculus proves that nonforking is invariant, monotone, has finite character, satisfies extension, and in stable theories is symmetric and transitive over models. These properties require a careful analysis of dividing formulas and indiscernible sequences. This chapter only prepares the definability and extension tools that make the later proofs possible. [/remark] The preview should be read together with the examples. In algebraically closed fields, nonforking corresponds to algebraic independence of generic points; in vector spaces, it corresponds to linear independence. The common pattern is that a type extends well when the new parameters do not create a hidden definable constraint over the base. [example: Algebraic Independence as the Geometric Model] Let $T=\operatorname{ACF}$, let $M$ be an algebraically closed subfield, and let $V\subseteq M^n$ be irreducible with prime ideal $I(V)\subseteq M[X_1,\dots,X_n]$. If $a=(a_1,\dots,a_n)$ is generic in $V$ over $M$, then the polynomial relations of $a$ over $M$ are exactly \begin{align*} \{f\in M[X_1,\dots,X_n]: f(a)=0\}=I(V). \end{align*} For a larger parameter field $B\supseteq M$, the base change $V_B$ is defined by the extended ideal \begin{align*} I(V)_B:=I(V)\cdot B[X_1,\dots,X_n]. \end{align*} The extension of $\operatorname{tp}(a/M)$ to $B$ is the geometric nonforking extension precisely when the only polynomial equations over $B$ satisfied by $a$ are those forced by $I(V)_B$, that is, \begin{align*} \{g\in B[X_1,\dots,X_n]: g(a)=0\}=I(V)_B. \end{align*} Equivalently, write $K=M(a_1,\dots,a_n)$ for the function field generated by the generic point. Since $a$ is generic in $V$ over $M$, \begin{align*} \operatorname{trdeg}(K/M)=\dim V. \end{align*} After adjoining $B$, the point $a$ remains generic in $V_B$ exactly when \begin{align*} \operatorname{trdeg}(B(a_1,\dots,a_n)/B)=\dim V. \end{align*} Combining the two displayed equalities gives \begin{align*} \operatorname{trdeg}(B(a)/B) = \operatorname{trdeg}(M(a)/M), \end{align*} so passing from $M$ to $B$ has introduced no new algebraic relation among the coordinates of $a$ beyond the equations defining $V$ over $M$. Thus, in algebraically closed fields, the invariant-extension picture says exactly that nonforking is algebraic independence over the base: the generic point stays generic after base change. [/example] The chapter closes with the main lesson: stability makes types definable over models, definability gives invariant extensions, and invariant extensions provide the first usable notion of independence. The next chapter develops forking itself and proves that this preview agrees with the formal definition in stable theories. At this point stability has become a theory of controlled dependence rather than just a counting condition. Strongly minimal sets provide the cleanest setting for that geometry, where definable closure and dimension behave much like algebraic dependence in linear algebra. # 8. Strongly Minimal Sets This chapter studies the point at which stability becomes geometry. In Chapters 1 through 3, types and saturated models gave a way to describe and realize possible elements over a parameter set; strongly minimal sets give a setting where algebraic dependence among elements behaves like linear dependence in a vector space. The guiding question is how the finite-or-cofinite behaviour of definable subsets in one variable can force a dimension theory for entire models. ## Strong Minimality for Theories and Definable Sets The first problem is to isolate a model-theoretic analogue of an irreducible one-dimensional object. We want a definable set whose definable subsets do not admit intermediate infinite complexity: every definable subset should be either small or almost everything. [definition: Strongly Minimal Theory] A complete first-order theory $T$ is strongly minimal if every model of $T$ is infinite and, for every model $M \models T$, every definable subset of $M$ with parameters from $M$ is finite or cofinite in $M$. [/definition] This definition is demanding because it is not enough to inspect parameter-free formulas. Parameters are allowed, so the theory must remain finite-or-cofinite after naming arbitrary elements of a model. [example: Pure Infinite Sets] Let $T$ be the theory of an infinite set with equality and no additional structure, and let $M\models T$. Fix parameters $a_1,\dots,a_n\in M$. Since the only atomic formulas in the variable $x$ with these parameters are equations of the form $x=a_i$ and equalities among the parameters, every formula $\theta(x,a_1,\dots,a_n)$ is, modulo $T$, a Boolean combination of the conditions $x=a_i$. For each $b\in M\setminus\{a_1,\dots,a_n\}$, the map fixing every $a_i$ and sending $b$ to any other element of $M\setminus\{a_1,\dots,a_n\}$ is a partial symmetry of the pure equality structure. Thus $\theta(x,a_1,\dots,a_n)$ cannot distinguish between two elements outside $\{a_1,\dots,a_n\}$. Therefore either every element of $M\setminus\{a_1,\dots,a_n\}$ satisfies $\theta$, or none of them does. On the named finite set itself, $\theta$ may select some subset \begin{align*} F\subseteq \{a_1,\dots,a_n\}. \end{align*} Hence $\theta(M,a_1,\dots,a_n)$ is either $F$, which is finite, or \begin{align*} (M\setminus\{a_1,\dots,a_n\})\cup F = M\setminus(\{a_1,\dots,a_n\}\setminus F), \end{align*} which is cofinite. Since this holds for every model $M\models T$ and every parameter tuple from $M$, the theory $T$ is strongly minimal. [/example] Pure equality is the basic case: algebraic information is exhausted by equality with named parameters. The example also exposes a limitation of defining strong minimality only for whole theories, since many important theories have a distinguished one-dimensional definable set with finite-or-cofinite behaviour even though higher-dimensional definable sets are richer. We therefore need a version of the definition that applies to a definable set and remains stable under elementary extension. [definition: Strongly Minimal Definable Set] Let $T$ be a complete theory, let $M \models T$, and let $D \subseteq M^n$ be definable, possibly with parameters. The definable set $D$ is strongly minimal if, in every elementary extension $N \succeq M$, the set $D(N)$ is infinite and every definable subset of $D(N)$ with parameters from $N$ is finite or cofinite in $D(N)$. [/definition] The reference to elementary extensions is essential: strong minimality is a property of the definable set in the theory, not just of its trace inside one model. The condition rules out the possibility that a new parameter in a larger model cuts $D$ into two infinite definable pieces. [example: Algebraically Closed Fields in One Variable] Let $K \models \operatorname{ACF}$ and let $N\succeq K$. We show that the definable set $N$ is infinite and that every definable subset of $N$ with parameters from $N$ is finite or cofinite. By *quantifier elimination for algebraically closed fields*, any subset of $N$ definable in one variable with parameters from $N$ is defined by a Boolean combination of polynomial equations \begin{align*} p(x)=0 \end{align*} with $p(x)\in N[x]$. For such a polynomial $p$, there are two cases. If $p=0$, then \begin{align*} \{x\in N:p(x)=0\}=N. \end{align*} If $p\neq 0$ and $\deg(p)=d$, then $p$ has at most $d$ roots, so \begin{align*} \{x\in N:p(x)=0\} \end{align*} is finite. Hence each basic polynomial equation defines either a finite subset of $N$ or all of $N$, and its negation defines either a cofinite subset of $N$ or the empty set. It remains only to check that Boolean combinations preserve the finite-or-cofinite property. If $A,B\subseteq N$ are finite or cofinite, then \begin{align*} N\setminus A \end{align*} is again finite or cofinite. Also, $A\cup B$ is finite if both $A$ and $B$ are finite; otherwise one of them is cofinite, say $A$, and then \begin{align*} N\setminus(A\cup B) =(N\setminus A)\cap(N\setminus B) \subseteq N\setminus A, \end{align*} so $A\cup B$ is cofinite. Since intersections can be written as \begin{align*} A\cap B = N\setminus\bigl((N\setminus A)\cup(N\setminus B)\bigr), \end{align*} intersections also preserve the finite-or-cofinite property. Therefore every one-variable definable subset of $N$ is finite or cofinite in $N$. Since this holds in every elementary extension $N\succeq K$, the affine line is a strongly minimal definable set. [/example] This example explains why strongly minimal sets are often described as model-theoretic curves. It also raises a technical question: the definition quantifies over all elementary extensions, while in practice we often work inside a large saturated model. The next result justifies checking strong minimality in that ambient model. [quotetheorem:5086] [citeproof:5086] The hypotheses are doing real work. Completeness prevents different completions from giving different one-variable behaviour; for instance, a disjoint union of two complete theories may have a formula that is strongly minimal in one component and not in another. The saturation bound is what realizes the infinite-and-coinfinite pattern encoded by all finite lower bounds; a small model can miss parameters witnessing a cut that appears in an elementary extension. The fixed formula and fixed parameter type are also necessary: changing parameters can change the definable set being tested, as $x=e$ and $x\neq e$ have different sizes inside an infinite pure set. The theorem does not say that every definable family has uniformly strongly minimal fibres, nor that checking a single non-saturated model is enough. Its use here is narrower and more powerful: once the formula and parameter type are fixed, we may work inside a large ambient model and then pass the conclusion to all elementary extensions containing corresponding parameters. ## Algebraic Closure as Geometry The natural closure operation in model theory is algebraic closure. In a strongly minimal setting it behaves less like arbitrary definability and more like the span operation in linear algebra. [definition: Algebraic Closure] Let $M$ be a structure. The algebraic closure operator in $M$ is the map \begin{align*} \operatorname{acl}_M:\mathcal P(M)\to\mathcal P(M) \end{align*} defined, for each $A \subseteq M$, by \begin{align*} \operatorname{acl}_M(A)=\{b \in M : b \text{ belongs to a finite } A\text{-definable subset of } M\}. \end{align*} [/definition] This closure operator records exactly which elements are forced up to finitely many choices by the parameters in $A$. To turn this into geometry, we need an abstract list of closure properties that captures the behaviour of span in a vector space and algebraic closure in a field. The decisive extra property will be exchange, because it is what makes bases and dimension independent of choices. [definition: Pregeometry] A pregeometry on a set $X$ is a closure operator $\operatorname{cl}:\mathcal P(X)\to\mathcal P(X)$ satisfying, for all $A,B\subseteq X$ and $a,b\in X$: 1. $A\subseteq \operatorname{cl}(A)$. 2. If $A\subseteq B$, then $\operatorname{cl}(A)\subseteq \operatorname{cl}(B)$. 3. $\operatorname{cl}(\operatorname{cl}(A))=\operatorname{cl}(A)$. 4. If $a\in \operatorname{cl}(A)$, then $a\in \operatorname{cl}(A_0)$ for some finite $A_0\subseteq A$. 5. If $a\in \operatorname{cl}(Ab)\setminus \operatorname{cl}(A)$, then $b\in \operatorname{cl}(Aa)$. [/definition] The final axiom is exchange. It says that if $b$ is genuinely needed to make $a$ algebraic over $A$, then $a$ can replace $b$ as a generator over $A$. Since ordinary algebraic closure in an arbitrary theory need not satisfy exchange, the next theorem is the first major structural consequence of strong minimality. [quotetheorem:5087] [citeproof:5087] The exchange conclusion is special to the strongly minimal situation. Reflexivity, monotonicity, idempotence, and finite character hold for algebraic closure in arbitrary first-order structures, but exchange can fail. For a concrete failure, take a structure with universe $U\cup V$, where $U$ and $V$ are infinite disjoint definable sets, and a binary relation $R(x,y)$ such that each fibre $R(U,y)$ has exactly two elements while each fibre $R(x,V)$ is infinite. If $R(a,b)$ with $a\in U$ and $b\in V$, then $a\in\operatorname{acl}(b)$, since $a$ lies in the two-element set $R(U,b)$, but $b\notin\operatorname{acl}(a)$ because the $a$-fibre $R(a,V)$ remains infinite and the language has no additional structure selecting a finite subset of that fibre. Strong minimality rules out that asymmetric finite dependence by forcing every relevant one-variable definable set to be finite or cofinite. The theorem does not identify the geometry with vector-space geometry; it supplies the pregeometry axioms, and different strongly minimal sets can have different closed-set structures. This is the point where the model-theoretic notion of algebraicity becomes suitable for bases, independence, and dimension. [example: Vector Spaces over a Finite Field] Let $V$ be an infinite vector space over a fixed finite field $F$, in the language with addition and scalar multiplication by elements of $F$. Fix parameters $a_1,\dots,a_n\in V$. By *quantifier elimination for vector spaces over a fixed finite field*, every one-variable formula over these parameters is equivalent to a Boolean combination of atomic affine linear equations. Each such equation has the form \begin{align*} \lambda x+u=v, \end{align*} where $\lambda\in F$ and $u,v\in \operatorname{span}_F(a_1,\dots,a_n)$. If $\lambda=0$, then the equation is either $u=v$, which defines all of $V$, or $u\neq v$, which defines the empty set. If $\lambda\neq 0$, then \begin{align*} \lambda x+u=v &\Longleftrightarrow \lambda x=v-u\\ &\Longleftrightarrow x=\lambda^{-1}(v-u), \end{align*} so it defines a singleton. Thus every one-variable definable subset of $V$ is a Boolean combination of finitely many singletons, hence is finite or cofinite. Therefore the theory is strongly minimal. For $A\subseteq V$, we compute $\operatorname{acl}(A)$. If $b\in \operatorname{span}_F(A)$, then for some $a_1,\dots,a_m\in A$ and $\lambda_1,\dots,\lambda_m\in F$, \begin{align*} b=\lambda_1a_1+\cdots+\lambda_m a_m. \end{align*} The formula \begin{align*} x=\lambda_1a_1+\cdots+\lambda_m a_m \end{align*} defines the singleton $\{b\}$ over $A$, so $b\in\operatorname{acl}(A)$. Conversely, suppose $b\notin \operatorname{span}_F(A)$. Any formula over $A$ uses only finitely many parameters $a_1,\dots,a_m\in A$. Since $b\notin \operatorname{span}_F(a_1,\dots,a_m)$, a linear automorphism fixing $\operatorname{span}_F(a_1,\dots,a_m)$ pointwise can send $b$ to any chosen vector outside that finite-dimensional span. There are infinitely many such choices, so no formula over those parameters that contains $b$ can define a finite set. Hence $b\notin\operatorname{acl}(A)$, and therefore \begin{align*} \operatorname{acl}(A)=\operatorname{span}_F(A). \end{align*} With this identification, the exchange axiom says exactly the linear replacement property: if \begin{align*} a\in \operatorname{span}_F(A\cup\{b\})\setminus \operatorname{span}_F(A), \end{align*} then $a=u+\lambda b$ for some $u\in\operatorname{span}_F(A)$ and some nonzero $\lambda\in F$, so \begin{align*} b=\lambda^{-1}(a-u)\in\operatorname{span}_F(A\cup\{a\}). \end{align*} Thus algebraic closure in this strongly minimal theory is ordinary linear span, and its dimension theory is the usual basis theory of vector spaces. [/example] This example is the template for the abstract theory. Strongly minimal sets may not be vector spaces, but their algebraic closure supports the same [basis and dimension](/page/Basis%20and%20Dimension) formalism. To use that formalism, we next name the corresponding notions of independence and basis over a parameter set. [definition: Independence and Basis] Let $T$ be strongly minimal, let $M\models T$, and let $A\subseteq M$. A set $B\subseteq M$ is independent over $A$ if, for every $b\in B$, $b\notin\operatorname{acl}(A\cup(B\setminus\{b\}))$. A basis of $M$ over $A$ is a maximal subset of $M$ independent over $A$. [/definition] The definition mirrors linear algebra with $\operatorname{acl}$ replacing span. The next issue is whether the definition depends on arbitrary choices: a maximal independent set might appear to depend on the order in which elements are added. The pregeometry axioms rule out that ambiguity and give a well-defined size. [quotetheorem:5088] [citeproof:5088] The theorem depends on exchange, not merely on having some closure operator. Without exchange, maximal independent generating sets can have different sizes. For example, on $X=\{a,b,c\}$ define a finite-character closure by \begin{align*} \operatorname{cl}(\varnothing)=\varnothing,\quad \operatorname{cl}(\{a\})=X,\quad \operatorname{cl}(\{b\})=\{b\},\quad \operatorname{cl}(\{c\})=\{c\}, \end{align*} declare $\operatorname{cl}(\{b,c\})=X$, and close all other larger sets by inclusion and idempotence. Then $\{a\}$ is independent and generates $X$, because $\operatorname{cl}(\{a\})=X$. The set $\{b,c\}$ is also independent and generates $X$, because neither $b$ nor $c$ lies in the closure of the other and $\operatorname{cl}(\{b,c\})=X$. Thus two maximal independent generating sets have sizes $1$ and $2$. Strong minimality is used through the previous theorem to exclude this kind of asymmetric dependence. The result does not say that a basis is canonical, nor that there is a definable procedure selecting one. It says only that all maximal independent generating sets have the same cardinality, which makes the next definition of model dimension meaningful. ## Dimension and Models The next problem is to determine how much information is needed to recover a model of a strongly minimal theory. The answer is that, once the algebraic closure of the empty set is fixed, a model is controlled by the size of a basis. [definition: Dimension of a Model] Let $T$ be strongly minimal and let $M\models T$. The dimension of $M$ is the cardinal \begin{align*} \dim(M)=|B|, \end{align*} where $B$ is any basis of $M$ over $\operatorname{acl}(\varnothing)$. [/definition] Dimension measures the number of mutually independent generic elements in the model. Elements outside the basis are algebraic over finitely many basis elements and the empty-set algebraic part. In practice, computations follow a three-step pattern: first identify $\operatorname{acl}(A)$ in the concrete structure, then test independence by checking that no listed element lies in the closure of the others over the base, and finally compare models by counting a maximal independent family. In pure equality this says $\operatorname{acl}(A)=A$, so independence is pairwise distinctness over $A$; in algebraically closed fields it says $\operatorname{acl}(A)$ is field-theoretic algebraic closure of the generated field, so independence is algebraic independence. The next classification theorem says that this numerical invariant is not merely descriptive: it determines the model up to isomorphism. [quotetheorem:5089] [citeproof:5089] The explicit map on $\operatorname{acl}(\varnothing)$ is necessary because the algebraic part is part of the structure, not part of the basis. In algebraically closed fields, a transcendence-basis bijection says nothing by itself about which conjugate of an algebraic element such as a square root of $-1$ should be chosen; an elementary map on the algebraic closure of the prime field supplies that missing compatibility. There is no extra generation hypothesis hidden in the theorem: maximality of a basis already implies $M=\operatorname{acl}^M(A_M\cup B)$, and similarly for $N$. What the theorem gives is the main reduction: after the algebraic base is fixed, the non-algebraic part of a strongly minimal model has no additional structure beyond independence. [example: Dimensions in Pure Infinite Sets] In the pure equality theory, every formula $\theta(x,a_1,\dots,a_n)$ with parameters from $A$ is equivalent to a Boolean combination of equations $x=a_i$, using only finitely many parameters $a_1,\dots,a_n\in A$. If $b\in A$, then the formula $x=b$ defines the singleton $\{b\}$ over $A$, so $b\in\operatorname{acl}(A)$. Conversely, if $b\notin A$, then for every finite tuple $a_1,\dots,a_n\in A$ the elements of \begin{align*} M\setminus\{a_1,\dots,a_n\} \end{align*} all satisfy the same equality conditions with the named parameters. Hence any $A$-definable set containing $b$ contains all elements outside some finite subset of $A$, so it is infinite. Therefore $b\notin\operatorname{acl}(A)$, and \begin{align*} \operatorname{acl}(A)=A. \end{align*} A set $B\subseteq M$ is independent over $\operatorname{acl}(\varnothing)=\varnothing$ exactly when, for each $b\in B$, \begin{align*} b\notin \operatorname{acl}(B\setminus\{b\}) = B\setminus\{b\}, \end{align*} which is automatic because $b\notin B\setminus\{b\}$. A maximal independent set must contain every element of $M$: if $c\in M\setminus B$, then \begin{align*} c\notin \operatorname{acl}(B)=B, \end{align*} so $B\cup\{c\}$ would still be independent. Thus the only basis is $M$ itself, and \begin{align*} \dim(M)=|M|. \end{align*} So in pure equality, the dimension invariant is just cardinality: two infinite pure sets are isomorphic exactly when their underlying sets have the same size. [/example] The pure-set case is the classification with no algebraic closure beyond the chosen elements. Algebraically closed fields show the same principle with a more substantial closure operation. [example: Transcendence Degree in Algebraically Closed Fields] Let $K\models \operatorname{ACF}_p$, where $p$ is either $0$ or a prime, and let $k_0$ be the prime field: $k_0=\mathbb{Q}$ if $p=0$, and $k_0=\mathbb{F}_p$ if $p>0$. We compute the model-theoretic dimension of $K$ by identifying $\operatorname{acl}(\varnothing)$ and then comparing bases with transcendence bases. First, $\operatorname{acl}(\varnothing)$ is the field-theoretic algebraic closure of $k_0$ inside $K$. Indeed, if $a$ is algebraic over $k_0$, then there is a nonzero polynomial \begin{align*} q(X)=c_0+c_1X+\cdots+c_dX^d\in k_0[X] \end{align*} with $q(a)=0$. The parameter-free formula $q(x)=0$ defines the finite set of roots of $q$ in $K$, so $a\in\operatorname{acl}(\varnothing)$. Conversely, if $a\in\operatorname{acl}(\varnothing)$, then $a$ lies in a finite parameter-free definable subset of $K$. By quantifier elimination for algebraically closed fields, that finite set is definable by a Boolean combination of polynomial equations with coefficients in $k_0$. Since every infinite Boolean combination piece in one variable is cofinite, finiteness forces $a$ to satisfy some nonzero polynomial over $k_0$. Hence $a$ is field-theoretically algebraic over $k_0$. More generally, for any $A\subseteq K$, \begin{align*} \operatorname{acl}(A) = \{b\in K:b\text{ is algebraic over the subfield }k_0(A)\}. \end{align*} The same two inclusions prove this: a polynomial equation over $k_0(A)$ gives a finite $A$-definable root set, while membership in a finite $A$-definable set gives, by one-variable quantifier elimination, a nonzero polynomial over $k_0(A)$ vanishing at the element. Therefore a set $B\subseteq K$ is independent over $\operatorname{acl}(\varnothing)$ exactly when no $b\in B$ is algebraic over \begin{align*} k_0\bigl(B\setminus\{b\}\bigr). \end{align*} That is precisely algebraic independence over the prime field. A maximal such independent set is therefore a transcendence basis of $K$ over $k_0$. Hence \begin{align*} \dim(K)=|B|=\operatorname{trdeg}_{k_0}(K). \end{align*} So, in fixed characteristic $p$, the strongly minimal dimension of an algebraically closed field is exactly its transcendence degree over the prime field, and the classification by dimension is the usual classification of algebraically closed fields by characteristic and transcendence degree. [/example] This example is the bridge to categoricity because, in uncountable cardinalities, the size of an algebraically closed field determines its transcendence degree. The same cardinal arithmetic works for any countable strongly minimal theory: algebraic closure of finite data is countable, so an uncountable model must have a basis of the same size as the model. The next theorem turns the dimension classification into the advertised categoricity result. [quotetheorem:5090] [citeproof:5090] Each hypothesis marks a boundary of the argument. Strong minimality gives the dimension classification; without it, a stable theory may have several non-isomorphic models of the same uncountable size. Countability of the language is used to make algebraic closures of finite sets countable; in an uncountable language with many named constants, the algebraic part can already have the size of the language and can affect small uncountable cardinalities. The restriction to uncountable cardinalities is also essential: algebraically closed fields of fixed characteristic have countable models of different finite transcendence degrees and one of countably infinite transcendence degree, so the theorem does not imply countable categoricity. The forward connection is that strong minimality supplies a complete test case for Morley-style categoricity: control algebraic closure, compute basis size, and compare generated models. [remark: Role in Morley Theory] Strongly minimal theories form the geometric core of the stability hierarchy. They are the setting in which non-algebraic types behave like generic points of an irreducible geometry, and their dimension theory supplies the prototype for ranks and independence in later chapters. Morley's theorem begins with categoricity in one uncountable cardinal and extracts strong minimality-like behaviour from the resulting stable structure. [/remark] The geometry of strongly minimal sets sets up the final classification result of the course. Morley's categoricity theorem shows how categoricity in one uncountable cardinal forces strong structural consequences, linking the abstract rank-and-dimension picture back to model uniqueness. # 9. Morley's Categoricity Theorem The chapter is organised as a guided proof rather than as a catalogue of lemmas. We isolate the model-theoretic ingredients already built in the course: stability, saturated models, prime and atomic constructions, strongly minimal sets, and dimension. The central idea is that an uncountably categorical theory has a definable geometry whose dimension determines its uncountable models. ## The Categoricity Problem Suppose $T$ is a complete first-order theory in a countable language. If $T$ has exactly one model of size $\beth_1$, why should this say anything about models of size $\beth_{17}$, or of size $\aleph_{423}$? The compactness and Löwenheim-Skolem theorems produce models in many cardinalities, but they do not by themselves explain why uniqueness should transfer between cardinalities. [definition: Categoricity In A Cardinal] Let $T$ be a complete first-order theory and let $\kappa$ be an infinite cardinal. The theory $T$ is categorical in $\kappa$ if any two models $M,N \models T$ with $|M|=|N|=\kappa$ are isomorphic. [/definition] Categoricity is a statement about the entire class of models of a fixed size. For finite structures it can be accidental, while for infinite structures it usually reflects strong internal organisation of definable sets and types. [example: Algebraically Closed Fields] Fix a characteristic $p$, where $p=0$ or $p$ is prime, and let $T=\operatorname{ACF}_p$. Let $F_p$ denote the prime field: $F_p=\mathbb Q$ if $p=0$, and $F_p=\mathbb F_p$ if $p>0$. If $K \models T$, choose a transcendence basis $B$ of $K$ over $F_p$. By definition, $B$ is algebraically independent over $F_p$, and $K$ is algebraic over the purely transcendental field $F_p(B)$. We show that the isomorphism type of $K$ is determined by $|B|$. If $K,L \models \operatorname{ACF}_p$ have transcendence bases $B,C$ over the same prime field and $|B|=|C|$, choose a bijection $f:B \to C$. This extends uniquely to an isomorphism of rational function fields \begin{align*} F_p(B) &\longrightarrow F_p(C),\\ \frac{P(b_1,\ldots,b_n)}{Q(b_1,\ldots,b_n)} &\longmapsto \frac{P(f(b_1),\ldots,f(b_n))}{Q(f(b_1),\ldots,f(b_n))}, \end{align*} where $P,Q \in F_p[X_1,\ldots,X_n]$ and $Q(b_1,\ldots,b_n)\neq 0$. Since $K$ and $L$ are algebraic closures of $F_p(B)$ and $F_p(C)$ inside algebraically closed fields, the field isomorphism extends to an isomorphism $K \cong L$. Thus two models of $\operatorname{ACF}_p$ are isomorphic exactly when their transcendence degrees over the prime field are equal. Now suppose $K$ is uncountable and let $\lambda=|B|$. Every element of $K$ is algebraic over some finitely generated field $F_p(B_0)$ with $B_0 \subset B$ finite, because algebraicity over $F_p(B)$ uses only finitely many coefficients from $F_p(B)$. For each finite $B_0$, the field $F_p(B_0)$ has cardinality at most $\max(\aleph_0,|F_p|)=\aleph_0$, and it has only countably many algebraic elements. Hence \begin{align*} |K| &\leq \left|\bigcup_{\substack{B_0 \subset B\\ B_0 \text{ finite}}} \operatorname{acl}(F_p(B_0))\right|\\ &\leq |\{B_0 \subset B : B_0 \text{ finite}\}|\cdot \aleph_0\\ &= \max(\lambda,\aleph_0). \end{align*} Since $B \subset K$, we also have $\lambda \leq |K|$, so \begin{align*} |K|=\max(\lambda,\aleph_0). \end{align*} If $|K|$ is uncountable, this forces $|K|=\lambda=\operatorname{trdeg}(K/F_p)$. Therefore, in each uncountable cardinal, all models of $\operatorname{ACF}_p$ have the same transcendence degree and are isomorphic. In countable cardinality, the dimensions $0,1,2,\ldots,\aleph_0$ all give countable algebraically closed fields, and different dimensions give non-isomorphic fields, so $\operatorname{ACF}_p$ is not $\aleph_0$-categorical. [/example] This example already displays the eventual mechanism in Morley's theorem: classification by dimension. The difficult part is proving that every countable uncountably categorical theory admits an analogue of this dimension argument. [example: Dense Linear Orders] The theory $\operatorname{DLO}$ is $\aleph_0$-categorical because any two countable dense linear orders without endpoints are isomorphic by the usual back-and-forth construction. If $A$ and $B$ are countable models, enumerate them as $A=\{a_0,a_1,\ldots\}$ and $B=\{b_0,b_1,\ldots\}$. Starting from the empty partial isomorphism, suppose a finite order-preserving bijection has already been built. To add the next element $a_n$, its already-mapped neighbours determine a cut in the finite image inside $B$; density and the absence of endpoints provide an element of $B$ in exactly that cut. The same argument adds the next unused $b_n$ on the reverse step. The union of the finite partial isomorphisms is an order-preserving bijection $A \cong B$. In uncountable cardinalities, uniqueness fails. Fix an uncountable cardinal $\kappa$. There is a dense linear order of size $\kappa$ with countable cofinality, for example one obtained by arranging $\kappa$ many points inside each interval of a countable dense skeleton. There is also a dense linear order of size $\kappa$ whose cofinality is uncountable, for example a lexicographic sum of $\operatorname{cf}(\kappa)$ many dense blocks of total size $\kappa$. Cofinality is preserved by order isomorphism: if $f:L\to L'$ is an isomorphism and $C\subseteq L$ is cofinal, then $f(C)$ is cofinal in $L'$, and applying the same argument to $f^{-1}$ gives equality of the least possible sizes of cofinal subsets. Hence these two dense linear orders without endpoints have the same cardinality but are not isomorphic. Thus countable categoricity and uncountable categoricity are genuinely separate phenomena. [/example] Dense linear orders also warn us that quantifier elimination alone is not enough. Their types are transparent, but the resulting order geometry has too many cuts in uncountable size. [example: Real Closed Fields] Fix an uncountable cardinal $\kappa$. We construct two real closed fields of size $\kappa$ whose underlying orders have different cofinalities, so they cannot be isomorphic as ordered fields. First choose an ordered field \begin{align*} F_\omega=\mathbb Q(t_0,t_1,t_2,\ldots,u_\alpha:\alpha<\kappa) \end{align*} with \begin{align*} 1<t_0<t_1<t_2<\cdots \end{align*} and with every $u_\alpha$ bounded between $0$ and $t_0$, while the sequence $(t_n)_{n<\omega}$ is cofinal in $F_\omega$. Its real closure $K_\omega$ is a model of $\operatorname{RCF}$. Since $F_\omega$ has $\kappa$ generators over the countable field $\mathbb Q$, \begin{align*} |F_\omega|=\max(\kappa,\aleph_0)=\kappa. \end{align*} The real closure is algebraic over $F_\omega$, so \begin{align*} |K_\omega|\leq |F_\omega|\cdot \aleph_0=\kappa\cdot \aleph_0=\kappa, \end{align*} and because $F_\omega\subseteq K_\omega$, we get $|K_\omega|=\kappa$. Algebraic ordered extensions do not change cofinality: if $a$ is algebraic over $F_\omega$, then some element of $F_\omega$ bounds $a$ in absolute value, by applying the defining polynomial of $a$ and taking a bound larger than the absolute values of its coefficients. Hence $(t_n)_{n<\omega}$ is still cofinal in $K_\omega$, so $\operatorname{cf}(K_\omega)=\aleph_0$. Now choose an ordered field \begin{align*} F_\kappa=\mathbb Q(s_\alpha:\alpha<\kappa) \end{align*} with \begin{align*} 1<s_\alpha<s_\beta \quad \text{whenever } \alpha<\beta, \end{align*} and with $(s_\alpha)_{\alpha<\kappa}$ cofinal. Let $K_\kappa$ be its real closure. The same cardinality calculation gives \begin{align*} |K_\kappa|=\kappa. \end{align*} Every subset of $F_\kappa$ of size $<\operatorname{cf}(\kappa)$ is bounded by some $s_\alpha$, because each of its elements involves only finitely many generators and the union of fewer than $\operatorname{cf}(\kappa)$ finite supports is bounded below $\kappa$. Algebraicity again preserves cofinality, so $\operatorname{cf}(K_\kappa)=\operatorname{cf}(\kappa)$. Thus, whenever $\operatorname{cf}(\kappa)>\aleph_0$, the two real closed fields $K_\omega$ and $K_\kappa$ have the same cardinality $\kappa$ but different order cofinalities. Cofinality is preserved by order isomorphism, so they are not isomorphic. For singular $\kappa$ of countable cofinality, one instead varies the chain of proper convex valuation rings in the real closure; this convex valuation structure is also preserved by ordered-field isomorphism. Therefore $\operatorname{RCF}$ is not categorical in any uncountable cardinal, even though it has quantifier elimination. [/example] The contrast with algebraically closed fields identifies the role played by stability. Algebraic closure in algebraically closed fields behaves as a pregeometry, while definable order in dense linear orders and real closed fields creates too many independent cuts. ## Morley's Theorem The theorem proved in this chapter is a transfer theorem. Its hypothesis mentions only one uncountable cardinal, but its conclusion covers every uncountable cardinal. The countability of the language is part of the statement and is used throughout the proof through countable type spaces, prime model constructions, and the omitting types theorem. [quotetheorem:1492] [citeproof:1492] The theorem is a transfer statement with sharp hypotheses. The countability hypothesis keeps the relevant type spaces and prime constructions small enough for the stability analysis; without it, categoricity transfer needs stronger machinery and is not a formal consequence of the statement written here. The uncountability hypothesis is also essential: countable categoricity alone permits unstable examples such as dense linear orders, where back-and-forth controls only the countable model. The theorem does not classify the countable models of $T$, nor does it say that all uncountably categorical theories are fields; it says that every uncountable model is eventually measured by a stable independence geometry. The next sections explain how the argument eliminates order-like behaviour before extracting that geometry. ## From Categoricity To Stability Why should uniqueness of one large model force a bound on types over arbitrary parameter sets? The answer is that instability gives a formula capable of encoding long linear orders, and long linear orders carry enough combinatorial variation to build many models of the same cardinality. [definition: Unstable Formula] Let $T$ be a complete first-order theory. A formula $\varphi(x;y)$ has the order property in $T$ if there are a model $M \models T$ and tuples $(a_i)_{i \in \mathbb N}$, $(b_j)_{j \in \mathbb N}$ from $M$ such that \begin{align*} M \models \varphi(a_i;b_j) \quad \text{if and only if} \quad i<j. \end{align*} The theory $T$ is unstable if some formula has the order property in $T$. [/definition] The order property is the syntactic trace of an order-like configuration. A formula with this pattern can encode cuts in arbitrarily long linear orders, and different cut structures can survive as non-isomorphic global models. The obstruction to categoricity is therefore not merely that many types appear over one parameter set, but that the order pattern can be expanded into genuinely different models of the same uncountable size. [quotetheorem:5091] [citeproof:5091] This lemma is the first place where the proof converts a combinatorial failure into many models. The countability of the language keeps the EM templates and omitted-type bookkeeping within the standard first-order spectrum theorem used here. The conclusion is deliberately weaker than the full unstable spectrum theory: it only supplies two models in each uncountable cardinal, because that is enough to contradict categoricity. Its limitation is also its strength for Morley's proof: any definable order pattern, even a local one coming from a single formula, is fatal to uncountable uniqueness. The next theorem records the contrapositive in exactly the form used in Morley's argument. [quotetheorem:5092] [citeproof:5092] Stability is the entry point to the rest of the proof. The hypothesis of uncountable categoricity is used only once in this reduction, namely to rule out the two EM models in the categorical cardinal. The result does not yet give a classification of models; a stable countable theory can still have many non-isomorphic uncountable models. What it gives is a controlled type theory, so the later arguments can replace order cuts by nonforking, ranks, and eventually a geometric dimension. ## Saturated Models And Prime Extensions Once stability is known, the proof shifts from counting models to controlling their internal generation. The central question becomes: can a large model be rebuilt from a small base and a well-chosen independent set? [definition: Saturated Model] Let $T$ be a complete theory and let $\kappa$ be an infinite cardinal. A model $M \models T$ is $\kappa$-saturated if every complete type $p(x) \in S_n(A)$ over every parameter set $A \subset M$ with $|A|<\kappa$ is realised in $M$, for every $n \ge 1$. [/definition] Saturation turns type existence into elements of the model. For Morley's proof, the obstruction is that a categorical model could a priori omit small types, leaving too little internal realization for back-and-forth or prime-over-basis arguments. Uniqueness at an uncountable cardinal rules out this defect: if the model failed to realize the relevant small types, one could build competing models at the same size by adding or omitting those realizations. [quotetheorem:5093] [citeproof:5093] The role of saturation is not only existential. It makes back-and-forth arguments possible over parameter sets of size below the model, which is how type-theoretic information becomes an isomorphism theorem. The proof has to use an invariant such as dimension of a stationary type; merely building an extension that realises a type over $A$ would not contradict abstract uniqueness, because an arbitrary isomorphism need not fix $A$. Saturation is also limited to the uncountable categorical size at this point, so later transfer work is still needed. The next construction explains how models are generated over bases once enough types are realised. [definition: Prime Model Over A Set] Let $T$ be a complete theory, let $A$ be a parameter set in a monster model, and let $M \models T$ contain $A$. The model $M$ is prime over $A$ if for every model $N \models T$ containing $A$, there is an elementary embedding $f:M \to N$ such that $f(a)=a$ for all $a \in A$. [/definition] Prime models over sets are minimal models generated over prescribed data, up to elementary embedding. To transfer categoricity, it is not enough to know that two models have bases of the same size; the models must be recoverable from those bases in a canonical way. The possible obstruction is extra structure outside the chosen basis that is invisible to dimension. Prime-over-basis constructions remove that freedom by forcing the rest of the model to be the minimal elementary structure generated over the base. [quotetheorem:5094] [citeproof:5094] This theorem explains why dimension is enough once the right geometry has been found. Its hypotheses are stronger than bare stability: the prime models must exist over the chosen bases, and the strongly minimal set must make finite independent tuples homogeneous over the same parameter set. These requirements are exactly what Morley's rank analysis supplies before the final transfer step. The limitation is important for using the theorem correctly: equality of cardinalities of arbitrary parameter sets is not enough; the sets must be independent bases for the same geometry over the same base. The hard work is locating that geometry inside an arbitrary uncountably categorical theory. ## Strongly Minimal Sets And Geometry The next problem is to identify the replacement for transcendence bases in algebraically closed fields. Strong minimality is the model-theoretic condition that every definable subset of a universe is either small because it is finite, or large because its complement is finite. [definition: Strongly Minimal Set] Let $M$ be a structure and let $D \subset M^n$ be a definable set. The set $D$ is strongly minimal if, in every elementary extension $N \succ M$, every definable subset of $D(N)$, with parameters from $N$, is finite or cofinite in $D(N)$. [/definition] The quantification over elementary extensions is essential because strong minimality is a property of all definable subsets that may appear after adding parameters. It is stronger than saying that definable subsets in one model are finite or cofinite. [example: The Field Universe In Algebraically Closed Fields] Let $K \models \operatorname{ACF}_p$, and let $N \succ K$ be any elementary extension. By quantifier elimination for algebraically closed fields, every definable subset of $N$ in one variable, with parameters from $N$, is defined by a Boolean combination of polynomial equations \begin{align*} f(x)=0 \end{align*} with $f \in N[X]$. If $f=0$, then $f(x)=0$ defines all of $N$. If $f\neq 0$ and $\deg(f)=d$, then $f$ has at most $d$ roots in the field $N$, so $f(x)=0$ defines a finite subset of $N$. Therefore each atomic polynomial condition defines either a finite set or all of $N$. Complements of finite sets are cofinite, finite unions of finite sets are finite, and finite intersections of cofinite sets are cofinite. Hence every Boolean combination of one-variable polynomial equations defines either a finite or cofinite subset of $N$. Since the same argument works in every elementary extension and with arbitrary parameters from that extension, the field universe is strongly minimal. Thus on the field sort, model-theoretic algebraic closure gives the corresponding geometric closure operation; in this case it is exactly field-theoretic algebraic dependence over the chosen parameters and the prime field. [/example] Strong minimality is useful because it produces a dependence relation from algebraic closure. For this relation to support dimension, algebraic closure must behave like linear span or field-theoretic algebraic closure: independent generating sets should be exchangeable and all bases should have the same size. The obstruction is that algebraic closure in an arbitrary definable set may fail exchange, making dimension depend on the chosen generating set. Strong minimality supplies the missing exchange property. [quotetheorem:5095] [citeproof:5095] A pregeometry gives bases and dimension exactly as in linear algebra or field theory. Strong minimality is essential here: algebraic closure in an arbitrary definable set need not satisfy exchange, so the resulting "dimension" could depend on the chosen generating set. The theorem is also local to $D$; it does not assert that algebraic closure on the whole monster model is a pregeometry. Its forward role is to turn the strongly minimal set found later into an operational invariant: choose independent elements, extend to a basis, and compare models by the size of that basis. The next definition is needed so that this invariant has a precise meaning. [definition: Dimension In A Strongly Minimal Set] Let $D$ be a strongly minimal set and let $\operatorname{cl}: \mathcal{P}(D) \to \mathcal{P}(D)$ be the closure operator $\operatorname{cl}(A)=D \cap \operatorname{acl}(A)$. A subset $I \subset D$ is independent if $a \notin \operatorname{cl}(I \setminus \{a\})$ for every $a \in I$. A basis of $X \subset D$ is a maximal independent subset of $X$. The dimension of $X$ is the cardinality of any basis of $X$. [/definition] The exchange axiom ensures that all bases of the same set have the same cardinality. Thus the definition gives a genuine invariant rather than a choice-dependent number. [example: Transcendence Degree As Strongly Minimal Dimension] Let $K \models \operatorname{ACF}_p$ and let $D=K$. For $A \subset K$, let $F_p(A)$ be the smallest subfield of $K$ containing the prime field $F_p$ and $A$. Then \begin{align*} \operatorname{cl}(A) &=D \cap \operatorname{acl}(A)\\ &=K \cap \{x : x \text{ is algebraic over } F_p(A)\}\\ &=\{x \in K : x \text{ is algebraic over } F_p(A)\}. \end{align*} The middle equality is the usual identification, in algebraically closed fields, of model-theoretic algebraicity over $A$ with field-theoretic algebraicity over the subfield generated by $A$ and the prime field. Thus an element $a \in K$ lies in $\operatorname{cl}(A)$ exactly when there are $c_0,\ldots,c_{n-1} \in F_p(A)$ such that \begin{align*} a^n+c_{n-1}a^{n-1}+\cdots+c_1a+c_0=0. \end{align*} So a set $I \subset K$ is independent for $\operatorname{cl}$ exactly when no element $a \in I$ is algebraic over $F_p(I\setminus\{a\})$, which is precisely algebraic independence over $F_p$. A basis for $K$ in this pregeometry is therefore a maximal algebraically independent subset of $K$ over $F_p$, namely a transcendence basis. Hence the strongly minimal dimension of $K$ is \begin{align*} \dim(K)=|I|=\operatorname{trdeg}(K/F_p), \end{align*} where $I$ is any transcendence basis of $K$ over the prime field. [/example] This example is the model for the general proof. Morley's theorem says that an arbitrary countable theory categorical in one uncountable cardinal is forced into a comparable geometric classification. ## The Guided Proof Of The Transfer The remaining obstacle is to show that the geometry extracted from one categorical cardinal classifies arbitrary uncountable models, not just saturated models in the original size. The proof has two directions of motion: categoricity gives stability and strong minimal geometry, then geometry gives categoricity in all uncountable sizes. The next results make precise what the phrase "classified by dimension" must mean. [explanation: Advanced Terms in the Transfer Argument] The final transfer uses a few stability-theoretic terms that belong to the standard proof of Morley's theorem. A **stationary type** is a complete type whose non-forking extension to larger parameter sets is unique, so it behaves like a generic point with a well-defined independence relation. A **regular type** is a stationary type that cannot be decomposed into simpler forking directions; it is the model-theoretic analogue of an irreducible geometric direction. Two types are **orthogonal** when their realizations remain independent in every common extension, so they describe unrelated geometries. They are **nonorthogonal** when the two geometries interact through definable or algebraic dependence. With this language, **unidimensionality** means that all non-algebraic regular types relevant to the models belong to one nonorthogonality class. There is then only one geometric dimension to measure, rather than several unrelated dimensions. **Dimension-transfer** is the step saying that once an uncountable model is prime over a base together with an independent basis for that one geometry, its isomorphism type is controlled by the size of that basis. These notions are not extra hypotheses invented for the exposition; they name the technical bridges between strong minimality and the final categoricity transfer. [/explanation] The first bridge is an existence theorem. Up to this point, strong minimality has been illustrated by algebraically closed fields, where the geometric set is already visible. In an arbitrary uncountably categorical countable theory, the corresponding geometry has to be found. The next theorem says that the rank and categoricity analysis produces such a strongly minimal set and that it is not an isolated ornament: the regular types that can influence uncountable models are tied to its generic type. [quotetheorem:5096] [citeproof:5096] This theorem is the geometric extraction step, and its hypotheses are doing real work. Countability gives the rank calculus and prime models used in the argument, while uncountable categoricity collapses competing regular types that would otherwise produce independent dimensions. Both assumptions have concrete failure modes. In an uncountable language with unary predicates $(P_i)_{i \in I}$ naming many independent pieces, categoricity in a chosen large cardinal can reflect the language already recording the pieces rather than a countable rank analysis; the countable proof has no way to keep the corresponding type spaces small. Without uncountable categoricity, even a stable countable theory can have several unrelated geometries: the disjoint union of two independent algebraically closed fields of the same characteristic has two strongly minimal field sorts, and a basis in one field says nothing about the transcendence degree of the other. The theorem therefore does not say that every definable set is strongly minimal; it identifies one strongly minimal geometry whose generic type is connected, by nonorthogonality, to the non-algebraic regular types that can affect models. The next theorem is needed because finding such a set is useful for categoricity only if every uncountable model is governed by its dimension in that set. [quotetheorem:5097] [citeproof:5097] This is the point at which dimension replaces case-by-case construction. The extra hypotheses prevent a common misuse of the statement: a strongly minimal set may classify only its own sort. For a concrete failure, take the disjoint union of two independent copies of $\operatorname{ACF}_0$, with separate field languages and no relations between the two sorts. Each field sort is strongly minimal, but a basis for the first sort records only the first transcendence degree. Two models can have the same first-sort dimension and the same total cardinality while having different second-sort transcendence degrees, so they are not isomorphic. What fails is precisely unidimensionality, or equivalently the assertion that every regular type relevant to the model is nonorthogonal to the chosen generic type. Under Morley's hypotheses the extracted $D$ is not arbitrary; it is the geometry through which all non-algebraic regular behaviour is measured. The theorem also explains how to use the invariant in practice: find the strongly minimal set $D$, compute a basis of $D(M)$ under algebraic closure over $A$, and read the uncountable size from that basis. The next theorem is needed to finish the categoricity transfer by applying dimension equality to two arbitrary models of the same uncountable size. [quotetheorem:1492] [citeproof:1492] Together with the earlier reduction from categoricity to stability and strong minimality, this proves Morley's Categoricity Theorem. The transfer step uses countability twice: prime models over $A \cup I$ remain of size $\max(|I|,\aleph_0)$, and the rank analysis that produced $D$ was carried out in a countable language. The restriction to uncountable $\lambda$ avoids the countable-model phenomena seen in algebraically closed fields, where finite and countable dimensions can all lead to countable structures. The theorem therefore transfers uncountable categoricity, but it does not claim countable categoricity or describe the full countable spectrum. Its forward connection is the central lesson of stability theory: once all uncountable models are prime over bases in one geometry, categoricity becomes equality of dimension. ## Consequences And Limitations What changes after Morley's theorem? Before it, categoricity could look like a collection of isolated examples. After it, categoricity becomes evidence that a theory belongs to a highly organised stability-theoretic regime. [quotetheorem:5098] [citeproof:5098] This consequence explains why dense linear orders and real closed fields are excluded. Both contain definable order, and the order supplies cuts that cannot be measured by a single geometric dimension. The hypothesis is stronger than the conclusion: many stable theories have no order property but are still far from categorical. The result is therefore best used as an obstruction test; finding a definable order pattern immediately rules out uncountable categoricity for countable theories. In the positive direction, the absence of such a pattern permits the rank and independence tools used earlier in the chapter. [remark: Countability Is A Real Hypothesis] Morley's theorem is stated for countable languages. In larger languages, categoricity transfer requires additional hypotheses and belongs to the broader setting of classification theory and abstract elementary classes. The countable proof uses countable syntax at several points, including the management of type spaces and the construction of prime models. [/remark] The theorem also does not say that all categorical theories look like algebraically closed fields. It says that uncountable categoricity forces a geometric analysis with strong restrictions on types and independence. [remark: The Theorem As A Turning Point] Morley's theorem initiated the classification-theoretic programme: classify theories by the behaviour of types, independence, ranks, and the number of models in each cardinal. Its proof showed that model-theoretic dividing lines are not bookkeeping devices; they control the global spectrum of models. Shelah's later work vastly extended this perspective by studying stable, superstable, and unstable theories through their model-counting behaviour. [/remark] The chapter closes the course by connecting the technical tools of types and stability to a structural theorem about all uncountable models of a theory. The slogan is that in the stable categorical world, models are built over independent bases, and cardinality becomes dimension. The course now moves from a single classification theorem to Shelah's broader program for understanding all complete theories. The tools developed so far, from types and saturation to stability and categoricity, become the starting point for a much wider structure theory of models across cardinals. # 10. Shelah's Classification Program Shelah's classification program asks what counts as a usable structure theory for all models of a complete first-order theory. Earlier chapters built types, Stone spaces, saturation, omitting types, stability, strongly minimal geometry, and Morley's theorem; this chapter places those tools into a wider organizing scheme. The central theme is that model theory does not classify theories one by one, but separates them by dividing lines: stable behaviour supports ranks, independence, and controlled model construction, while unstable behaviour tends to encode order-like complexity. ## From Morley to Dividing Lines What remains after Morley's theorem is not only the result that categoricity transfers for countable theories, but the method: count types, analyze how they vary over parameter sets, and use that information to build models. Shelah's insight was that this method can be turned into a taxonomy of theories. The first dividing line is stability, which says that over a set of parameters there are no more complete types than the size of the parameter set forces. [definition: Order Property] A complete theory $T$ has the order property if there is a formula $\varphi(x,y)$, a model $M \models T$, and sequences $(a_i)_{i\in\mathbb N}$ and $(b_j)_{j\in\mathbb N}$ from $M$ such that \begin{align*} M \models \varphi(a_i,b_j) \quad \text{if and only if} \quad i<j. \end{align*} [/definition] The order property is the basic source of too many types: the formula $\varphi(x,y)$ can distinguish initial segments of a long ordered sequence of parameters. To turn this obstruction into a dividing line, we measure the size of type spaces over parameter sets. Stability is the cardinal-sensitive condition saying that, once the parameter set has size $\kappa$, the space of complete finite-tuple types over it has no more than $\kappa$ elements. [definition: Stable Theory] A complete theory $T$ is stable in an infinite cardinal $\kappa$ if for every parameter set $A$ with $|A|=\kappa$ and every $n\geq 1$, the set $S_n(A)$ of complete $n$-types over $A$ has cardinality at most $\kappa$. The theory $T$ is stable if it is stable in some infinite cardinal. [/definition] Equivalently, it is enough to count $1$-types over models of size $\kappa$: the finite-tuple and arbitrary-parameter-set formulation follows by the usual coding and extension arguments. The definition is stated in the broader form because later tools, especially forking and stationarity, are used over parameter sets as well as models. Stability is the negation of the order-producing explosion of types. Once a theory is stable, the next question is whether this type bound persists uniformly rather than appearing only in isolated cardinals; this motivates the stronger condition of superstability. [definition: Superstable Theory] A complete theory $T$ is superstable if it is stable in all sufficiently large infinite cardinals. [/definition] Superstability is stronger than stability because it says that type-counting eventually behaves uniformly. The next refinement asks what happens at the smallest infinite scale. For countable theories this is the scale directly tied to Morley's theorem, since countability of the language makes countable parameter sets the first serious test of whether types are being controlled by syntax rather than by uncontrolled combinatorics. [definition: Omega-Stable Theory] A complete countable theory $T$ is $\omega$-stable if for every countable model $M \models T$, the type space $S_1(M)$ is countable. [/definition] The notation $\omega$-stable records stability at the first infinite cardinal. For many classification arguments, even this type-counting condition is used through a rank that always terminates on definable sets, which motivates total transcendence. [definition: Totally Transcendental Theory] A complete theory $T$ is totally transcendental if every formula $\varphi(x,a)$ with parameters $a$ from a monster model has ordinal-valued Morley rank. [/definition] The point of total transcendence is that definable sets can be measured by a rank that always terminates. At this stage there are three related controls on complexity: stability bounds the number of types over large sets, $\omega$-stability bounds types over countable models, and Morley rank asks for a terminating rank on definable sets. The obstruction is that these conditions need not coincide in arbitrary languages or cardinals. In the countable setting, however, the syntax is small enough for the rank and type-counting formulations to line up in the expected hierarchy. [explanation: Stability Hierarchy Overview] For a complete countable theory $T$, total transcendence is equivalent to $\omega$-stability, and $\omega$-stability implies superstability, which implies stability. Each implication marks a genuine strengthening in the general stability hierarchy. This is best read here as a map of available tools rather than as a new classification theorem: total transcendence gives terminating Morley ranks, $\omega$-stability gives countable control of type spaces, superstability gives stronger forking and rank behavior, and stability gives the basic independence calculus. [/explanation] The countability hypothesis is essential in the stated equivalence between total transcendence and $\omega$-stability: for uncountable languages, the right counting bound is no longer simply countability of the type spaces over countable sets. Outside the countable setting, one must replace this statement by cardinal-sensitive versions involving $|T|$ and stability in cardinals at least $|T|$. The theorem also does not say that every stable theory has Morley rank, prime models over arbitrary sets, or a geometric decomposition; it only locates the first implications in the hierarchy. In the sequel, this hierarchy tells us which tools are available: total transcendence gives terminating ranks and atomic constructions, superstability gives stronger rank and forking control, and stability gives the basic independence calculus. This hierarchy should therefore be read as a first map, not as the whole classification program. The same model-theoretic objects recur at each level, but with different degrees of control: types may be countable, ranked, finitely based, or merely bounded by the size of the parameter set. To see why stability is weaker than the strongly minimal geometry from Morley's theorem, we need an example where types are controlled but definable subsets are not forced to be finite or cofinite. [example: Stable But Not Strongly Minimal] Let $T$ be the complete theory in the language $\{E\}$ saying that $E$ is an equivalence relation with infinitely many equivalence classes and that every equivalence class is infinite. We show that $T$ is stable by counting types over an infinite parameter set $A$ with $|A|=\kappa$. A complete $1$-type over $A$ has exactly one of the following forms: \begin{align*} &x=a \quad \text{for some } a\in A,\\ &x\neq a \text{ for all } a\in A,\quad E(x,a_0) \text{ for some } a_0\in A,\\ &x\neq a \text{ for all } a\in A,\quad \neg E(x,a) \text{ for every } a\in A. \end{align*} The first family has $\kappa$ possibilities, the second has at most one possibility for each $E$-class meeting $A$, hence at most $\kappa$ possibilities, and the third is a single type. Thus \begin{align*} |S_1(A)|\leq \kappa+\kappa+1=\kappa. \end{align*} For an $n$-type, the same description is finite: one records which variables are equal, which unequal variables are $E$-equivalent, which variables equal named parameters, and which remaining $E$-classes meet $A$. There are only finitely many equality and $E$-patterns among $x_1,\dots,x_n$, and for each fixed pattern at most $\kappa^n=\kappa$ choices of parameter classes from $A$. Hence $|S_n(A)|\leq \kappa$ for every finite $n\geq 1$, so $T$ is stable. The theory is not strongly minimal. If $a$ is any parameter, then $E(x,a)$ defines the $E$-class of $a$, which is infinite. Its complement is also infinite, because there are infinitely many other $E$-classes and each contains at least one element. Thus $E(x,a)$ is neither finite nor cofinite. Stability controls the number of types, but it does not force every definable subset of the universe to be finite or cofinite. [/example] The example shows that stable theories need not look like algebraically closed fields or pure sets. Stability controls the number and behaviour of types, while strong minimality imposes a much sharper geometry on definable subsets of the universe. The contrast case is a theory where a formula can encode long finite order patterns, so dense orders supply the first nonstructure prototype. [example: Dense Orders Produce Instability] In dense linear orders without endpoints, take the formula $x<y$. Fix $n\geq 1$ and choose \begin{align*} a_1< a_2<\cdots<a_n. \end{align*} For each $m$ with $0\leq m\leq n$, choose $b_m$ as follows: choose $b_0<a_1$, choose $b_n>a_n$, and for $1\leq m<n$ choose \begin{align*} a_m<b_m<a_{m+1}, \end{align*} which is possible by density and by the absence of endpoints. Then for each $1\leq i\leq n$, \begin{align*} a_i<b_m &\quad\Longleftrightarrow\quad i\leq m. \end{align*} Indeed, if $i\leq m<n$, then $a_i\leq a_m<b_m$; if $m=n$, then $a_i<a_n<b_n$ for $i<n$ and $a_n<b_n$ for $i=n$. If $i>m$ and $m\geq 1$, then $b_m<a_{m+1}\leq a_i$, so $a_i<b_m$ fails; if $m=0$, then $b_0<a_1\leq a_i$, so $a_i<b_0$ fails. These finite initial-segment patterns give the order property for $x<y$: after reindexing $b_m$ as $b_{j}$ with the cut placed just before $a_j$, the truth value of $a_i<b_j$ records exactly whether $i<j$. Over a countable dense parameter set $A$, each Dedekind cut $C=(L,R)$ determines a partial type \begin{align*} p_C(x)=\{a<x:a\in L\}\cup\{x<a:a\in R\}. \end{align*} Every finite subset of $p_C(x)$ is realized between the largest finitely mentioned element of $L$ and the smallest finitely mentioned element of $R$, or beyond all finitely mentioned parameters if one side is empty. Hence compactness extends $p_C$ to a complete type over $A$. Distinct cuts give distinct complete types, because some parameter lies on different sides of the two cuts and the formulas $x<a$ and $a<x$ separate the resulting types. Since a countable dense order has continuum many cuts, $S_1(A)$ is uncountable for countable $A$, so dense orders are unstable. [/example] Dense orders illustrate why instability is treated as a dividing line rather than a mild complication. Once a theory can encode long linear orders, spaces of types become too large for the stable toolkit to classify models by rank and independence alone. ## Forking, Ranks, and Prime Models Over Sets How does stability turn type-counting into geometry? The answer is forking. Earlier chapters used types as consistent descriptions; stability theory adds an independence relation that tells us when a type over a larger set carries no extra dependence beyond a smaller base. [definition: Forking Independence] Let $T$ be a complete stable theory, let $A \subset B$ be parameter sets, and let $p(x) \in S(B)$ be a complete type. A formula $\varphi(x,b) \in p$ with $b \in B$ divides over $A$ if there is an $A$-indiscernible sequence $(b_i)_{i\in\mathbb N}$ with $b_1=b$ and $b_i \equiv_A b$ for all $i$ such that the set $\{\varphi(x,b_i):i\in\mathbb N\}$ is inconsistent. The type $p$ forks over $A$ if it implies a finite disjunction of formulas, each of which divides over $A$. The type $p$ does not fork over $A$ if it does not fork over $A$. [/definition] The formal definition hides several equivalent formulations in stable theories. What matters in these notes is the role it plays: non-forking extension is the correct analogue of linear independence or algebraic independence. For this analogy to support model construction, forking must satisfy a robust calculus, so the next result records the properties used throughout stability theory. [quotetheorem:5100] [citeproof:5100] Stability is needed here because the proof repeatedly uses the failure of the order property to control indiscernible sequences. In dense linear orders, for instance, cuts over a parameter set produce many extensions whose behaviour is governed by order position rather than a symmetric dimension calculus; the formula $x<y$ gives the concrete pattern behind this failure. The base conditions are also substantive: over a set that is not sufficiently closed, an algebraic choice can split into several extensions after new parameters are named, so uniqueness over that base is not expected. In algebraically closed fields, a type over a non-algebraically closed subfield can have different conjugate extensions after adjoining a root of a polynomial, while over an algebraically closed base the generic extensions are stationary. The uniqueness clause is therefore deliberately restricted: stationarity over a model, or over an algebraically closed base in common stable contexts, is what prevents two distinct non-forking extensions from making different choices over the larger parameter set. Thus forking is not merely a relation on arbitrary triples; it becomes a workable calculus only after the base has enough closure and the theory has enough stability. These properties make forking usable as a dimension theory. The next question is how to measure the depth of dependence: if every genuine forking extension lowers complexity, then a rank can count how far a type is from being generic. [definition: Forking Rank] A forking rank assigns to types or definable sets an ordinal or infinity, decreasing when one passes to a proper forking extension. [/definition] Different ranks appear in different parts of stability theory: Morley rank in totally transcendental theories, Lascar rank in superstable settings, and related ranks in more general stable theories. The common idea is that rank records the depth of possible dependence. In practice, a rank drop is a test for dependence: if extending a type by a formula lowers its rank, then the new formula has imposed a genuine constraint rather than merely naming harmless extra parameters. The geometric example to keep in mind is the generic point of a variety, where rank recovers dimension. [example: Generic Type of an Irreducible Variety] Let $\mathfrak K \models \mathrm{ACF}$ be a monster algebraically closed field containing $k$, and write \begin{align*} I(V)=\{f\in k[X_1,\dots,X_n]: f(v)=0 \text{ for every } v\in V\}. \end{align*} Since $V$ is irreducible, $I(V)$ is prime by *[Hilbert's Nullstellensatz](/theorems/2124)*. The generic type of $V$ over $k$ is \begin{align*} p_V(x)=\{f(x)=0:f\in I(V)\}\cup\{g(x)\neq 0:g\in k[X_1,\dots,X_n]\setminus I(V)\}. \end{align*} This exactly says that $x$ lies on $V$ and avoids every proper $k$-definable Zariski closed subset of $V$: if $W\subsetneq V$ is closed over $k$, then some $g\notin I(V)$ vanishes on $W$, so $g(x)\neq 0$ excludes $W$. To see that the displayed partial type is consistent, take finitely many conditions \begin{align*} f_1(x)=0,\dots,f_r(x)=0,\qquad g_1(x)\neq 0,\dots,g_s(x)\neq 0 \end{align*} with each $f_i\in I(V)$ and each $g_j\notin I(V)$. The equations impose only membership in $V$. If no point of $V$ satisfied all inequalities, then \begin{align*} V\subseteq Z(g_1)\cup\cdots\cup Z(g_s)=Z(g_1\cdots g_s), \end{align*} so $g_1\cdots g_s\in I(V)$. Since $I(V)$ is prime, this would imply $g_j\in I(V)$ for some $j$, contradicting the choice of the $g_j$. Thus every finite part is realized in some algebraically closed [field extension](/page/Field%20Extension), and compactness gives a complete type extending it. In algebraically closed fields, Morley rank of an irreducible constructible set agrees with its Zariski dimension. Hence \begin{align*} \operatorname{RM}(p_V)=\operatorname{RM}(V)=\dim V. \end{align*} If $a\models p_V$ and $B\supseteq k$, then $a$ remains generic over $B$ exactly when \begin{align*} \operatorname{trdeg}(B(a)/B)=\dim V; \end{align*} forking occurs exactly when the transcendence degree drops, which is algebraic dependence over the base field. Thus the model-theoretic generic point is the usual geometric generic point: it satisfies precisely the equations forced by membership in $V$, and no extra algebraic condition over the base. [/example] The variety example shows ranks acting like dimensions of independent choices. To turn this geometry into an actual classification of models, we also need canonical ways to build models over a chosen base; this is the role of prime models over sets. [definition: Regular Type] In a stable theory, a stationary type $p$ is regular if every forking extension of $p$ is orthogonal to $p$. [/definition] Regular types are the indecomposable pieces of stable geometry. Their dimensions count the size of maximal independent sets of realizations, much as the dimension of a vector space counts the size of a basis. This is the invariant that later obstructions such as DOP try to make order-dependent. [definition: Prime Model Over A Set] Let $T$ be a complete theory and let $A$ be a parameter set. A model $M \models T$ containing $A$ is prime over $A$ if for every model $N \models T$ containing $A$, there is an elementary embedding $f:M \to N$ fixing $A$ pointwise. [/definition] Prime models over sets are the model-construction counterpart of isolated and stationary types. They are rarely available in arbitrary theories, but in stable and especially totally transcendental contexts they become part of the controlled architecture of models. The next theorem gives the basic existence result that makes this architecture available in the ranked countable setting. [quotetheorem:5101] [citeproof:5101] Countability matters because the construction enumerates the relevant formulas and existential requirements in the expanded language. Total transcendence matters because it supplies enough isolated types: without ordinal-valued rank, the attempted atomic construction may encounter persistent non-isolated type spaces. The hypothesis on $A$ is also not cosmetic; after naming $A$, the expanded theory must remain in the countable totally transcendental setting where the isolation argument applies. The theorem does not assert that arbitrary stable theories have prime models over arbitrary sets, and it does not replace saturation or homogeneity as general construction methods. Its use here is narrower and more powerful: in the ranked countable setting, models can be built canonically from a base by realizing isolated requirements. This theorem explains why rank, isolation, and primeness belong together. A good rank theory turns the task of constructing models into a controlled bookkeeping process over types. ## Classification and Nonstructure When should a theory be considered classified? The answer in Shelah's program is not merely that we can list its models in a few small cardinals. Classification means that, in every relevant uncountable cardinal, models are organized by structural invariants arising from forking geometry, decomposition, and prime extensions. [definition: Classifiable Theory] A complete countable first-order theory $T$ is classifiable if it is superstable, has no dimensional order property, and has no omitting types order property. [/definition] This is the technical sense of classifiability used in the countable first-order main gap. It is not a vague synonym for being stable, nor does it mean that the models can be listed by a short formula. It means that after superstability is available, the two standard order-like obstructions to decomposition have also been removed. [explanation: Reading DOP and OTOP] The definitions of DOP and OTOP are included as orientation to the main gap, not as tools the course will prove from first principles. The independence relation in the DOP definition is forking independence. A model prime over $M_1\cup M_2$ is the canonical minimal model generated over the union, when such prime models exist. A stationary regular type over that prime model represents one indecomposable geometric direction. DOP says that such directions can be arranged so that dimensions remember a hidden order. OTOP says that, even when dimensions are not the coding device, the pattern of omitted types can still remember an order. Both properties are therefore order-coding obstructions inside the superstable world. [/explanation] The first obstruction arises when dimensions of regular types can be arranged along a hidden order. A miniature picture is a chain of components indexed by a linear order $I$, where each cut or interval in $I$ changes the dimension of a regular type over the corresponding part of a decomposition. If two different orders produce different dimension patterns, then the proposed structural invariant is secretly remembering the order rather than reducing the model to independent geometric pieces. [definition: Dimensional Order Property] A superstable theory $T$ has the dimensional order property if there are models $M_0 \preceq M_1$ and $M_0 \preceq M_2$ with $M_1$ independent from $M_2$ over $M_0$, a prime model $M_3$ over $M_1 \cup M_2$, and a stationary regular type $p \in S(M_3)$ such that $p$ is orthogonal to both $M_1$ and $M_2$. [/definition] The absence of DOP is what lets decomposition trees behave like structural invariants rather than hidden encodings of arbitrary orders. There is a second obstruction, OTOP, where the order coding is not through numerical dimensions but through which types are omitted. In a typical schematic picture, one builds components $M_i$ along an ordered index set and arranges a type $q_{ij}$ to be realized or omitted according to the relation $i<j$. The model then remembers the order through the pattern of omissions. [definition: Omitting Types Order Property] A superstable theory $T$ has the omitting types order property if there is a partial type $r(x;y,z)$ over $\varnothing$ such that, for every linear order $I$, there is a model $M \models T$ and tuples $(a_i)_{i\in I}$ and $(b_j)_{j\in I}$ from $M$ with $r(x;a_i,b_j)$ omitted in $M$ exactly for the pairs satisfying $i<j$. [/definition] DOP and OTOP explain why superstability alone is not the final structural condition. Superstability supplies ranks and forking, but these obstructions show that the resulting decomposition can still carry an arbitrary order inside its dimensions or omission patterns. Removing both obstructions is what makes the structure side of the program a genuine decomposition theory rather than a repackaged coding theorem. [explanation: Structure Side of the Program] On the structure side, models behave like objects built from independent components. Forking gives the independence relation, ranks measure the complexity of pieces, and prime models over sets allow the pieces to be assembled without arbitrary choices. The resulting classification is not a list of models by presentation, but a description by invariants such as dimensions of regular types and decomposition trees. This perspective generalizes the lessons of Morley's theorem. Categoricity is an extreme form of structure: in a given cardinal there is only one model up to isomorphism. Shelah's program asks for a broader theory of when there may be many models, but still many in a controlled and analyzable way. [/explanation] The opposite side needs its own name because failure of classification is also a mathematical phenomenon with recognizable mechanisms. When a theory can code orders, trees, or graphs into its models, the resulting diversity of models is called nonstructure. [definition: Nonstructure] Nonstructure is the side of a classification theorem on which a complete theory has a dividing-line configuration that codes independent combinatorial data into its models. [/definition] The term names the mechanism and theorem scheme rather than a single cardinal-by-cardinal definition. The usual mechanism is coding independent combinatorial data such as orders, trees, or graphs into models; the precise model-count conclusion depends on the language size, the cardinal range, and the dividing line being used. The guiding slogan is that unstable theories tend to have nonstructure because the order property allows models to remember complicated order patterns. Dense orders are the basic example, while graphs provide a useful comparison with later dividing lines beyond stability. This makes the random graph a useful warning: not every unstable theory is chaotic in the same way. [example: Random Graph as Motivation Beyond Stability] Let $R(x,y)$ denote adjacency in the countable random graph. Its extension property says that for any finite disjoint sets $U,V$ there is a vertex adjacent to every element of $U$ and to no element of $V$. Choose distinct vertices \begin{align*} a_1,a_2,\dots . \end{align*} For each finite set $S\subseteq \{1,\dots,n\}$, apply the extension property with \begin{align*} U=\{a_i:i\in S\},\qquad V=\{a_i:1\leq i\leq n,\ i\notin S\}. \end{align*} Then there is a vertex $b_S$ such that, for each $1\leq i\leq n$, \begin{align*} R(a_i,b_S) \quad\Longleftrightarrow\quad i\in S. \end{align*} Thus one formula, namely $R(x,y)$, realizes all Boolean adjacency patterns on finite parameter sets. Now let $A=\{a_i:i\in\mathbb N\}$. For every subset $X\subseteq \mathbb N$, consider the partial type \begin{align*} p_X(y)=\{R(a_i,y):i\in X\}\cup\{\neg R(a_i,y):i\notin X\}. \end{align*} Every finite subset of $p_X$ mentions only $a_1,\dots,a_n$ for some $n$, and it is realized by the vertex $b_{X\cap\{1,\dots,n\}}$ constructed above. Hence, by the *Compactness Theorem*, $p_X$ extends to a complete type over $A$. If $X\neq Y$, choose $i\in X\triangle Y$; then one of the two complete types contains $R(a_i,y)$ and the other contains $\neg R(a_i,y)$, so they are distinct. Since there are $2^{\aleph_0}$ subsets of $\mathbb N$, the space $S_1(A)$ has size at least $2^{\aleph_0}$ over the countable set $A$, so the random graph is unstable. The same graph is still highly homogeneous: its finite extension property gives the usual Fraisse-limit back-and-forth construction, so finite partial graph isomorphisms extend step by step to automorphisms. The lesson is that instability is not the end of classification; it is the first major boundary after which one needs new independence notions, such as those developed in simple theories. [/example] The random graph therefore plays a double role. It warns us that stable tools cannot classify every homogeneous structure, while also suggesting that some unstable theories retain enough independence theory to be studied by later refinements. This motivates the broad organizing principle called the main gap. [explanation: Main Gap Orientation] Shelah's main gap says, informally, that complete countable first-order theories split into a structure side and a nonstructure side. On the structure side, classifiable theories have models in uncountable cardinals described by invariants arising from stable decomposition. On the nonstructure side, non-classifiable theories satisfy precise spectrum theorems giving the maximum possible number of pairwise non-isomorphic models in the relevant uncountable cardinals. This is an orientation rather than a theorem statement for this course. The full main gap has technical hypotheses and cardinal-specific formulations, and it is not proved here. Its role is to mark the forward connection: stability, forking, ranks, regular types, and prime constructions are the tools that make the structure side possible, while order-like coding properties drive the nonstructure side. [/explanation] The countability assumption keeps the classical first-order form of the program visible; for uncountable languages the spectrum statements must be adjusted to the size of the language and the relevant cardinals. Completeness is also part of the setup, since the program classifies models of one fixed complete theory rather than a disjunction of unrelated completions. The closing example returns to dense orders to make the nonstructure slogan concrete. [example: Dense Orders as a Nonstructure Prototype] Dense linear orders without endpoints are locally tame in two separate senses. First, every formula in one free variable is equivalent, in the theory of dense linear orders without endpoints, to a Boolean combination of inequalities against parameters; this is the usual *quantifier elimination for dense linear orders*. Second, any two countable dense linear orders without endpoints are isomorphic: enumerate them as $(a_i)_{i\in\mathbb N}$ and $(b_i)_{i\in\mathbb N}$, and build a finite partial order-isomorphism back and forth. At each step, if $a_m$ is not yet in the domain, its already-mapped neighbours determine a cut \begin{align*} L&=\{f(a_i):a_i<a_m,\ a_i\in\operatorname{dom}(f)\},\\ R&=\{f(a_i):a_m<a_i,\ a_i\in\operatorname{dom}(f)\}. \end{align*} Every element of $L$ is below every element of $R$, so density and the absence of endpoints give some $b$ with \begin{align*} \ell<b<r \quad \text{for all finitely many } \ell\in L,\ r\in R, \end{align*} with the evident endpoint modifications if $L$ or $R$ is empty. Sending $a_m$ to such a $b$ preserves all order relations already mentioned. The symmetric step adds the next unused $b_m$ to the range. The union of the finite maps is therefore an isomorphism. This countable uniqueness does not persist in uncountable size. For an uncountable cardinal $\kappa$, the lexicographic order $\kappa\times\mathbb Q$ is dense and has no endpoints: if $(\alpha,q)<(\beta,r)$ and $\alpha=\beta$, choose $s\in\mathbb Q$ with $q<s<r$; if $\alpha<\beta$, choose $s\in\mathbb Q$ with $q<s$, so \begin{align*} (\alpha,q)<(\alpha,s)<(\beta,r). \end{align*} Its cofinality is $\operatorname{cf}(\kappa)$, because a cofinal subset must be unbounded in the first coordinate. Other dense orders of the same cardinality can have different cofinalities or different cut structures, and those invariants are preserved by order-isomorphism. Thus first-order tameness and countable categoricity do not prevent uncountable models from remembering global order data; this is why dense orders are a prototype for nonstructure. [/example] This final example also explains why Morley's theorem is surprising. Categoricity in one uncountable cardinal rules out precisely the kind of uncontrolled order coding seen in dense orders, forcing a theory into the stable part of the landscape. Shelah's classification program begins from that phenomenon and expands it into a systematic geography of first-order theories. ## Beyond These Notes The next natural step is to compare stability with the weaker dividing lines that keep some independence theory after the order property appears. The random graph points toward simple theories, where forking is replaced by a broader independence relation and algebraic closure no longer behaves as rigidly as it does in strongly minimal settings. Algebraically closed fields point in a different direction: stable theories can carry geometric structure rich enough to recover dimension, generic points, and independence from algebraic data. Several themes in these notes also connect directly to other Androma topics. Compactness underlies the construction of type spaces and omitted-type arguments; compare the topological viewpoint in [Ultrafilter Characterisation of Compactness](/theorems/1049) and [Sequential Compactness](/page/Sequential%20Compactness). Boolean algebras and Stone duality explain why spaces of complete types are compact totally disconnected spaces. Algebraic geometry supplies the prototype for Morley rank and generic types; see [Algebraic Geometry Overview](/page/Algebraic%20Geometry%20Overview) and [Cambridge II Algebraic Geometry](/page/Cambridge%20II%20Algebraic%20Geometry) for the geometric background behind irreducible varieties, generic points, and dimension. Finally, [Morley's Categoricity Theorem](/theorems/1492) is the public theorem-level anchor for the classification direction developed here, while combinatorial order coding explains why unstable theories have large type spaces. The common thread is that a model is studied not only by its individual formulas, but by the global geometry of all consistent descriptions over parameter sets. For further study, one can follow three paths. The classification path develops superstability, regular types, DOP, OTOP, and the main gap in more detail. The geometric path studies strongly minimal sets, pregeometries, and Zilber-style questions about when model-theoretic geometry resembles algebraic geometry. The modern dividing-lines path continues beyond stability to simple, NIP, and NTP theories, asking which fragments of independence and dimension survive in less rigid settings. ## References These course notes are self-contained and use internal theorem citations throughout; no external bibliography is required for the present page.